Molecular Modeling of Proteins PDF

Methods in
Molecular Biology 1215
Andreas Kukol Editor
Molecular
Modeling
of Proteins
Second Edition
METHODS
IN
M O L E C U L A R B I O LO G Y
Series Editor
John M. Walker
School of Life Sciences
University of Hertfordshire
Hatfield, Hertfordshire, AL10 9AB, UK
For further volumes:

http://www.springer.com/series/7651
Molecular Modeling of Proteins

Second Edition
Edited by
Andreas Kukol
University of Hertfordshire, Hatfield, Hertfordshire, UK
Editor
Andreas Kukol
University of Hertfordshire
Hatfield
Hertfordshire, UK
Additional material to this book can be downloaded from http://extras.springer.com

ISSN 1064-3745
ISSN 1940-6029 (electronic)
ISBN 978-1-4939-1464-7
ISBN 978-1-4939-1465-4 (eBook)
DOI 10.1007/978-1-4939-1465-4
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2014945209
Springer Science+Business Media New York 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction
on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this
legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for
the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publishers location, in its current version, and permission for use must always be obtained from Springer. Permissions
for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither
the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be
made. The publisher makes no warranty, express or implied, with respect to the material contained herein.
Printed on acid-free paper
Humana Press is a brand of Springer
Springer is part of Springer Science+Business Media (www.springer.com)
Preface
Over the years, molecular modeling and simulation of biomolecules has become an important tool in the molecular biosciences. Initially situated in the realm of specialists with indepth knowledge of physics and computer science and access to supercomputers, molecular
modeling is used increasingly by bioscientists who are mainly interested in investigating
biological problems. This development has been supported by improved hardware, such as
multi-core processors or graphic processing units, on the one hand, and accelerated sampling algorithms on the other hand that increase the timescale without increasing the
demands on the hardware or the calculation time. The purpose of Molecular Modeling of
Proteins is to provide a theoretical background of various methods available and to enable
nonspecialists to apply methods to their problems. Most chapters contain, in addition to a
thorough introduction, step-by-step instructions and notes on troubleshooting and how to
avoid common pitfalls.
The current second edition of Molecular Modeling of Proteins provides some updated
chapters and new material not covered in the first edition. The first part describes classical
and advanced simulation methods as well as methods to set up complex systems such as
lipid membranes and membrane proteins. The second part is devoted to the simulation and
analysis of conformational changes of proteins, while Part III covers computational methods for protein structure prediction as well as using experimental data in combination with
computational techniques. The final part contains chapters concerning proteinligand
interactions, which are relevant in the drug design process.
The topics cover some long established methods together with the latest developments
in the field. The chapters are written by internationally renowned investigators: they include
leading developers of popular simulation packages or force fields.
The second edition of Molecular Modeling of Proteins is directed at researchers in the
physical-, chemical-, and biosciences working in industry and academia, who are interested
in applying the methods in their own research. Additionally, the book forms a valuable
resource for educators who wish to teach courses about molecular modeling.
Hertfordshire, UK
Andreas Kukol
Contents
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
PART I
SIMULATION METHODS
1 Molecular Dynamics Simulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Erik Lindahl
2 Transition Path Sampling with Quantum/Classical Mechanics
for Reaction Rates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Frauke Grter and Wenjin Li
3 Current Status of Protein Force Fields for Molecular
Dynamics Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pedro E.M. Lopes, Olgun Guvench, and Alexander D. MacKerell Jr.
4 Lipid Membranes for Membrane Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Andreas Kukol
5 Molecular Dynamics Simulations of Membrane Proteins . . . . . . . . . . . . . . . . .
Philip C. Biggin and Peter J. Bond
6 Membrane-Associated Proteins and Peptides . . . . . . . . . . . . . . . . . . . . . . . . . .
Marc F. Lensink
7 Coarse-Grained Force Fields for Molecular Simulations . . . . . . . . . . . . . . . . . .
Jonathan Barnoud and Luca Monticelli
8 Tackling Sampling Challenges in Biomolecular Simulations . . . . . . . . . . . . . . .
Alessandro Barducci, Jim Pfaendtner, and Massimiliano Bonomi
9 Calculation of Binding Free Energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Vytautas Gapsys, Servaas Michielssens, Jan Henning Peters,
Bert L. de Groot, and Hadas Leonov
PART II
v
ix
27
47
73
91
109
125
151
173
CONFORMATIONAL CHANGE
10 The Use of Experimental Structures to Model Protein Dynamics. . . . . . . . . . .

Ataur R. Katebi, Kannan Sankar, Kejue Jia, and Robert L. Jernigan
11 Computing Ensembles of Transitions with Molecular
Dynamics Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Juan R. Perilla and Thomas B. Woolf
12 Accelerated Molecular Dynamics and Protein Conformational
Change: A Theoretical and Practical Guide Using a Membrane
Embedded Model Neurotransmitter Transporter. . . . . . . . . . . . . . . . . . . . . . .
Patrick C. Gedeon, James R. Thomas, and Jeffry D. Madura
13 Simulations and Experiments in Protein Folding . . . . . . . . . . . . . . . . . . . . . . .
Giovanni Settanni
vii
213
237
253
289
viii
Contents
PART III
PROTEIN STRUCTURE DETERMINATION
14 Comparative Modeling of Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Gerald H. Lushington
15 De Novo Membrane Protein Structure Prediction. . . . . . . . . . . . . . . . . . . . . .
Timothy Nugent
16 NMR-Based Modeling and Refinement of Protein 3D Structures . . . . . . . . . .
Wim F. Vranken, Geerten W. Vuister, and Alexandre M.J.J. Bonvin
PART IV
309
331
351
PROTEINLIGAND INTERACTIONS
17 Methods for Predicting ProteinLigand Binding Sites . . . . . . . . . . . . . . . . . . .

Zhong-Ru Xie and Ming-Jing Hwang
18 Information-Driven Structural Modelling of ProteinProtein Interactions . . . .
Joo P.G.L.M. Rodrigues, Ezgi Karaca, and Alexandre M.J.J. Bonvin
19 Identifying Putative Drug Targets and Potential Drug Leads:
Starting Points for Virtual Screening and Docking . . . . . . . . . . . . . . . . . . . . .
David S. Wishart
20 Molecular Docking to Flexible Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Jesper Srensen, zlem Demir, Robert V. Swift, Victoria A. Feher,
and Rommie E. Amaro
383
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
471
399
425
445
Contributors
ROMMIE E. AMARO Department of Chemistry and Biochemistry, University of California,
San Diego, CA, USA
ALESSANDRO BARDUCCI Laboratory of Statistical Biophysics, Ecole Polytechnique Fdrale
de Lausanne, Lausanne, Switzerland
JONATHAN BARNOUD INSERM, Lyon, France
PHILIP C. BIGGIN Department of Biochemistry, University of Oxford, Oxford, UK
PETER J. BOND Department of Chemistry, The Unilever Centre for Molecular Science
Informatics, Cambridge, USA; Department of Biological Sciences, National University of
Singapore, Singapore
MASSIMILIANO BONOMI Department of Bioengineering and Therapeutic Sciences
and California Institute of Quantitative Biosciences, University of California,
San Francisco, CA, USA
ALEXANDRE M.J.J. BONVIN Computational Structural Biology Group, Bijvoet Center
for Biomolecular Research, Faculty of Science, Utrecht University, Utrecht,
The Netherlands
ZLEM DEMIR Department of Chemistry and Biochemistry, University of California,
San Diego, CA, USA
VICTORIA A. FEHER Department of Chemistry and Biochemistry, University of California,
San Diego, CA, USA
VYTAUTAS GAPSYS Max Planck Institute for Biophysical Chemistry, Gttingen, Germany
PATRICK C. GEDEON Department of Biomedical Engineering, Duke University, Durham,
NC, USA
FRAUKE GRTER Heidelberg Institute for Theoretical Studies, Heidelberg, Germany
BERT L. DE GROOT Max Planck Institute for Biophysical Chemistry, Gttingen, Germany
OLGUN GUVENCH Department of Pharmaceutical Sciences, University of New England,
Portland, ME, USA
MING-JING HWANG Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan
ROBERT L. JERNIGAN National Cancer Institute, National Institute of Health, Bethesda,
MD, USA; Interdepartmental Program for Bioinformatics and Computational Biology,
L.H. Baker Center for Bioinformatics and Biological Statistics, Iowa State University,
Ames, IA, USA
KEJUE JIA National Cancer Institute, National Institute of Health, Bethesda, MD, USA;
Interdepartmental Program for Bioinformatics and Computational Biology, L.H. Baker
Center for Bioinformatics and Biological Statistics, Iowa State University, Ames, IA, USA
EZGI KARACA Computational Structural Biology Group, Bijvoet Center for Biomolecular
Research, Faculty of Science, Utrecht University, Utrecht, The Netherlands
ATAUR R. KATEBI National Cancer Institute, National Institute of Health, Bethesda, MD,
USA; Interdepartmental Program for Bioinformatics and Computational Biology, L.H.
Baker Center for Bioinformatics and Biological Statistics, Iowa State University, Ames,
IA, USA
ANDREAS KUKOL School of Life and Medical Sciences, University of Hertfordshire,
Hatfield, UK
ix
Contributors
MARC F. LENSINK Interdisciplinary Research Institute, CNRS USR3078, University Lille1,

Villeneuve dAscq, France
HADAS LEONOV Max Planck Institute for Biophysical Chemistry, Gttingen, Germany
WENJIN LI Department of Bioengineering, University of Illinois at Chicago, Chicago,
IL, USA
ERIK LINDAHL Department of Biochemistry and Biophysics, Stockholm University,
Stockholm, Sweden
PEDRO E.M. LOPES Department of Pharmaceutical Sciences, University of Maryland,
Baltimore, MD, USA
GERALD H. LUSHINGTON LiS Consulting, Lawrence, KS, USA
ALEXANDER D. MACKERELL JR. Department of Pharmaceutical Sciences,
University of Maryland, Baltimore, MD, USA
JEFFRY D. MADURA Department of Chemistry and Biochemistry and Center
for Computational Sciences, Duquesne University, Pittsburgh, PA, USA
SERVAAS MICHIELSSENS Max Planck Institute for Biophysical Chemistry, Gttingen, Germany
LUCA MONTICELLI INSERM, Lyon, France
TIMOTHY NUGENT Department of Computer Science, University College London, London, UK
JUAN R. PERILLA Beckman Institute, University of Illinois, Urbana at UrbanaChampaign, IL, USA
JAN HENNING PETERS Max Planck Institute for Biophysical Chemistry, Gttingen, Germany
JIM PFAENDTNER Department of Chemical Engineering, University of Washington, Seattle,
WA, USA
JOO P.G.L.M. RODRIGUES Computational Structural Biology Group,
Bijvoet Center for Biomolecular Research, Faculty of Science, Utrecht University,
Utrecht, The Netherlands
KANNAN SANKAR National Cancer Institute, National Institute of Health, Bethesda, MD,
USA; Interdepartmental Program for Bioinformatics and Computational Biology, L.H.
Baker Center for Bioinformatics and Biological Statistics, Iowa State University, Ames,
IA, USA
GIOVANNI SETTANNI Physics Department, Johannes Gutenberg Universitt, Mainz, Germany
JESPER SRENSEN Department of Chemistry and Biochemistry, University of California,
San Diego, CA, USA
ROBERT V. SWIFT Department of Chemistry and Biochemistry, University of California,
San Diego, CA, USA
JAMES R. THOMAS Department of Chemistry and Biochemistry and Center for Computational
Sciences, Duquesne University, Pittsburgh, PA, USA
WIM F. VRANKEN Department of Structural Biology, Structural Biology Brussels,
Vrije Universiteit Brussel, Brussels, Belgium
GEERTEN W. VUISTER Department of Biochemistry, University of Leicester, Leicester, UK
DAVID S. WISHART Department of Computing Science, University of Alberta, Edmonton,
AB, Canada; Department of Biological Sciences, University of Alberta, Edmonton,
AB, Canada
THOMAS B. WOOLF Department of Physiology, John Hopkins University, Baltimore,
MD, USA
ZHONG-RU XIE Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan
Part I
Simulation Methods
Chapter 1
Molecular Dynamics Simulations
ErikLindahl
Abstract
Molecular dynamics has evolved from a niche method mainly applicable to model systems into a
cornerstone in molecular biology. It provides us with a powerful toolbox that enables us to follow and
understand structure and dynamics with extreme detailliterally on scales where individual atoms can
be tracked. However, with great power comes great responsibility: Simulations will not magically
provide valid results, but it requires a skilled researcher. This chapter introduces you to this, and
makes you aware of some potential pitfalls. We focus on the two basic and most used methods; optimizing a structure with energy minimization and simulating motion with molecular dynamics. The
statistical mechanics theory is covered briefly as well as limitations, for instance the lack of quantum
effects and short timescales. As a practical example, we show each step of a simulation of a small protein, including examples of hardware and software, how to obtain a starting structure, immersing it in
water, and choosing good simulation parameters. You will learn how to analyze simulations in terms
of structure, fluctuations, geometrical features, and how to create ray-traced movies for presentations.
With modern GPU acceleration, a desktop can perform s-scale simulations of small proteins in a
dayonly 15 years ago this took months on the largest supercomputer in the world. As a final exercise, we show you how to set up, perform, and interpret such a folding simulation.
Key words Molecular dynamics, Simulation, Force field, Protein, Solvent, Energy minimization,
Position restraints, Equilibration, Trajectory analysis, Secondary structure
1 Introduction
Biomolecular dynamics occur over a wide range of scales in both
time and space, and the choice of approach to study them depends
on the question asked. In many cases the best alternative is an
experimental technique, for instance spectroscopy to study bond
vibrations or electrophysiology to study ion channels opening and
closing. However, theoretical methods have made huge advances
the last few decades, and there are now large domains where modeling and simulation either provide more detail or are more efficient to use compared to setting up a new experiment.
Molecular dynamics simulation is far from the only theoretical
method; when the aim is to predict for example the structure
Andreas Kukol (ed.), Molecular Modeling of Proteins, Methods in Molecular Biology, vol. 1215,
DOI 10.1007/978-1-4939-1465-4_1, Springer Science+Business Media New York 2015
Erik Lindahl
and/or function of proteins (rather than studying the dynamics of

a protein) the best tool is normally bioinformatics that detect
related proteins from amino acid sequence similarity. Similarly, for
computational drug design often it is much more productive to
use less accurate but exceptionally fast statistical methods like
QSAR (Quantitative StructureActivity Relationship) instead of
spending billions of CPU hours to simulate binding of thousands
of compounds.
Traditionally, the role of simulations has been to test if simple
theoretical models can predict experimental observations. For
example, simulations of ion channels have been useful to explain
why some ions pass while others are blocked, although the conductivity itself was already known from experiments. Similarly, simulations can provide detail not accessible through experiments, for
instance pressure distributions inside membranes. However, this is
changing rapidlysimulations have moved far beyond confirming
experiments, and today they frequently make predictions about
properties such as binding or folding dynamics that are later confirmed in the lab. With ever-increasing computational power this
development will not only continue, but it is likely to accelerate
significantly the next few years.
From an ideal physics point-of-view, the time-dependent
Schrdinger equation should be able to predict all properties of
any molecule with arbitrary precision ab initio. However, as soon
as more than a handful of particles are involved it is necessary to
introduce approximations. In quantum chemistry, one common
approximation is to assume that atomic nuclei do not move, and
using an implicit representation of solvent. This is obviously not
realistic for large biomolecules if we are interested in understanding their motion and sampling of lots of different states, so for
most biomolecular systems we instead choose to work with
empirical parameterizations of models, for instance classical
Coulomb interactions between pointlike atomic charges. The
conceptual difference is that quantum chemistry is excellent at
describing the electronic structure and enthalpy (potential) of the
system, while classical molecular dynamics instead excels at sampling the billions of states a macromolecule will adaptin particular this means they properly include the entropy part of free
energy. These models are not only orders of magnitude faster,
but since they have been parameterized from experiments they
also perform better when it comes to reproducing observations
on microsecond scale (Fig.1), rather than extrapolating quantum
models 10 orders of magnitude. The first molecular dynamics
simulation was performed as late as 1957 [1], although it was not
until the 1970s that it was possible to simulate water [2] and
biomolecules [3].

bond length
vibration
lipid
lipid
normal protein
diffusion
rotation
rotation
folding
rapid
ribosome
transport in
around bonds water
synthesis
relaxation ion channel protein folding
10-15s
10-12s
10-9s
10-6s
10-3s
1s
"biology"
membrane
protein fodling
103s
Accessible to atomic-detail simulation today (2013)
Fig. 1 Range of time scales for dynamics in biomolecular systems. While the individual time steps of molecular
dynamics is 12fs, parallel computers make it possible to simulate on microsecond scale, and distributed
computing techniques can sample even slower processes, almost reaching milliseconds
2 Theory
Macroscopic properties measured in an experiment are not direct
observations, but averages over billions of molecules representing
a statistical mechanics ensemble. This has deep theoretical implications that are covered in great detail in the literature [4, 5], but
even from a practical point of view there are important consequences: (1) It is not sufficient to work with individual structures,
but systems have to be expanded to generate a representative
ensemble of structures (see Note 1) at the given experimental conditions, e.g., temperature and pressurethis is one thing that sets
classical molecular dynamics apart from quantum chemistry. (2)
Thermodynamic equilibrium properties related to free energy,
such as binding constant, solubilities, and relative stability cannot
be calculated directly from individual simulations, but require
more elaborate techniques covered in later chaptersthese all rely
on entropy. (3) For equilibrium properties (in contrast to kinetic)
the aim is to examine the ensemble of structures, and not necessarily to reproduce individual atomic trajectories!
The two most common ways to generate statistically faithful
equilibrium ensembles are Monte Carlo and Molecular Dynamics
simulations, where the latter also has the advantage of accurately
reproducing kinetics of non-equilibrium properties such as diffusion or folding times. However, these methods cannot handle the
case where a structure is very far from equilibrium, for instance if
two atoms are almost overlapping after building a new side chain.
To remove this type of clashes prior to simulation, we typically start
with an Energy Minimization. This type of minimization is also
commonly used to refine low-resolution experimental structures.
All classical simulation methods rely on more or less empirical
sets of parameters called Force fields [69] to calculate interactions
and evaluate the potential energy of the system as a function of
pointlike atomic coordinates. A force field consists of both the set
of equations used to calculate the potential energy and forces from
particle coordinates, as well as a collection of parameters used in
Erik Lindahl
Fig. 2 Examples of interaction functions in modern force fields. Bonded interactions include covalent bond-stretching, angle-bending, torsion rotation around
bonds, and out-of-plane or improper torsions (not shown). Nonbonded interactions are based on neighborlists and consist of LennardJones attraction and
repulsion, as well as Coulomb electrostatics. Even a small amino acid residue
contains a large number of interactions, and for a protein there are thousands
the equations. For most purposes these approximations work great,

but they cannot reproduce quantum effects such as bond formation or breaking (see Note 2).
All common force fields subdivide potential functions in two
classes. Bonded interactions cover stretching of covalent bonds,
angle-bending, torsion potentials when rotating around bonds,
and out-of-plane improper torsion potentials, all which are normally fixed throughout a simulation (Fig.2). The remaining nonbonded interactions between atoms that are merely close in space
consist of LennardJones repulsion and dispersion as well as
Coulomb electrostatic. These are typically computed from neighborlists updated periodically.
Given the force on all atoms, the coordinates are updated for
the next step. For energy minimization, the steepest descent algorithm simply moves each atom a short distance in direction of
decreasing energy (force is the negative gradient of energy), while
molecular dynamics is performed by integrating Newtons equations of motion [10]:
Fi = -
V ( r1 ,,rN )
ri
= Fi
t 2
ri
mi
The updated coordinates are then used to evaluate the potential

energy again, as shown in the flowchart of Fig.3.
Initial input data:

Interaction function V(r) - "force field"
coordinates r, velocities v
Repeat for millions of steps
Compute potential V(r) and

forces F i = iV(r) on atoms
Update coordinates &

velocities according to
equations of motion
Collect statistics and write

energy/coordinates to
trajectory files
Yes
More steps?
No
Done!
Fig. 3 Simplified flowchart of a typical molecular dynamics simulation. The basic

idea is to generate structures from a natural ensemble by calculating potential
functions and integrating Newtons equations of motion, structures which are
then used to evaluate equilibrium properties of the system. A typical time step is
in the order of 1 or 2 femtoseconds, unless special techniques are used
Even the smallest chemical sample we can imagine is far too

large to include completely in a simulation. Instead, biomolecular
simulations normally uses periodic boundary conditions to avoid
surface artifacts, so that a water molecule that exits to the right
reappears on the left in the system; if the box is sufficiently large
the molecules will not interact significantly with their periodic copies. This is intimately related to the nonbonded interactions, which
ideally should be summed over all neighbors in the resulting infinite periodic system. Simple cutoffs can work for LennardJones
interactions that decay very rapidly, but for Coulomb interactions
a sudden cutoff can lead to large errors. In the early days of simulation is was common to switch off the electrostatic interaction
before the cutoff as shown in Fig.4, but this too has severe artifacts
the current method of choice is to use Particle-Mesh-Ewald summation (PME) to calculate the infinite electrostatic interactions
bysplitting the summation into short- and long-range parts [11].
Erik Lindahl
4
3
relative potential
2
1
0
-1
4
3
2
1
0
-1
0.1
0.2
0.3
0.4
0.5
0.7
0.6
r (nm)
0.8
0.9
1.1
Fig. 4 Alternatives to a sharp cutoff for nonbonded coulomb interactions. Top: By switching off the interaction
(dashed) before the cutoff the force will be the exact derivative of potential, but the derivative (and thus force)
will unnaturally increase just before the cutoff. Bottom: Particle-Mesh-Ewald is an amazing algorithm where
the coulomb interaction (solid) is divided into a short-range term that is evaluated within a cutoff (dashed) and
a long-range term which can be solved exactly in reciprocal space with Fourier transforms (dot-dash)
For PME, the cutoff is not really a cutoff; it only determines the
balance between the two parts, and the long-range part is treated
by assigning charges to a grid that is solved in reciprocal space
through Fourier transforms.
Cutoffs and rounding errors can lead to drifts in energy, which
will cause the system to heat up during the simulation. Even with
a theoretically perfect simulation we would run into problems
since we typically start from an imperfect structure. As the potential energy of this structure decreases during the simulation, the
kinetic energy (i.e., temperature) would increase if the total system
energy was constant. To control this, the system is normally coupled to a thermostat that scales velocities during the integration to
maintain room temperature. Similarly, the total pressure in the system can be adjusted through scaling the simulation box size, either
isotropically or separately in x/y/z dimensions.
The single most demanding part of simulations is the computation of nonbonded interactions, since millions of pairs have to be
evaluated for each time step. Extending the time step is thus an
important way to improve simulation performance, but unfortunately errors are introduced in bond vibrations already at 1fs.
However, in most simulations these bond vibrations are not of
interest per se, and can be removed entirely by introducing bond
constraint algorithms such as SHAKE [12] or LINCS [13].
Constraints make it possible to extend time steps to 2fs, and fixed-
length bonds are likely better approximations of the quantum
mechanical oscillators than harmonic springs (see Note 3)and in
the final section we will show you how to go even further.
3 Methods
With the basic theory covered, this section will describe how to (1)
choose and obtain a starting structure, (2) prepare it for a simulation, (3) create a simulation box, (4) add solvent water, (5) p
erform
energy minimization, (6) equilibrate the structure with simulation,
(7) perform the production simulation, and (8) analyze the trajectory data. To reproduce it, you will need access to a Unix/Linux
machine (see Note 4) with a molecular dynamics package installed.
While the options and files below refer to the GROMACS program [14], the description should be reasonably straightforward to
follow with other programs like AMBER [15], CHARMM [16],
or NAMD [17]. It will also be useful to have the molecular viewer
PyMOL [18] and Unix graph program Grace installed (see Note5).
3.1 Obtaining
aStarting Structure
The Bovine Pancreatic Trypsin Inhibitor (BPTI) is a small 58-residue

water-soluble protein that inhibits several serine proteases [19].
It was one of the first proteins to be simulated [3], and has often
been referred to as a hydrogen atom of protein simulation.
There are several high-resolution X-ray structures of BPTI [20] in
the Protein Data Bank (http://www.pdb.org), and also NMR
structures. It was actually the early simulations of BPTI [3] that
lead experimentalists to realize that X-ray temperature factors can
be used to study the local dynamics of a protein [21]. Choose the
entry 6PTI with 1.7 resolution [20], and download it as 6PTI.
pdb (see Note 6). Figure5 shows a cartoon representation of this
structure; the small crosses are crystal water oxygen atoms visible
in the X-ray experiment (see Note 7).
3.2 Preparation
ofInput Data
In addition to the coordinates/velocities that change each step,

simulations also need a static description of all atoms and interactions in the system, called topology. In GROMACS, this is created
from the PDB structure by the program pdb2gmx, which also adds
all the hydrogen atoms that are not present in most X-ray structures. For this example we will work with the Amber99SB-ILDN
force field, the TIP3P [22] water model (see Note 8), and accept
the default choices for all residue protonation states, termini, disulfide bridges, etc. If you just try the command below right away you
will get an error due to the issues with the structure mentioned in
Note 6. This is common, so it is something you need to learn how
to fix. Open the PDB file in an editor and scroll down to residue
57 at the end of the chain. Remove the single nitrogen atom line
just before the line starting with TERwe simply skip the missing residues. Just after the TER line there are also some lines for
a phosphate ion (residue PO4). To avoid problems with finding
parameters for this, remove these five lines too. Now you should be
good to go. The command to use is then
10
Erik Lindahl
Fig. 5 Cartoon representation of the BPTI structure 6PTI from Protein Data Bank,
with side chains shown as sticks. Including hydrogens, the protein contains
roughly 800 atoms. Ray-traced image generated with PyMOL
pdb2gmx f 6PTI.pdb water tip3p

You will be prompted for the force field (select 6 for
Amber99SB-ILDN), and the command will produce three files:
conf.gro contains coordinates with hydrogens, topol.top is
the topology, and posre.itp contains a list of position restraints
that will be used in Subheading3.7. For all these programs, you can
use the h flag for help and a detailed list of options (see Note 9).
3.3 Creating
aSimulation Box
The default box is taken from the PDB crystal cell, but a simulation in water requires something larger. The box size is a trade-off,
though: volume is proportional to the box side cubed, and more
water means the simulation is slower. The easiest option it to place
the solute in the center of a cube, with for example 0.75nm to the
box sides. We will show up some more advanced alternatives later,
but for now this will suffice:
editconf f conf.gro d 0.75 o box.gro
where the distance (-d) flag automatically centers the protein in
the box, and the new conformation is written to the file box.gro
(see Note 10).
3.4 Adding
SolventWater
The last step before the simulation is to add water in the box to
solvate the protein. This is done by using a small pre-equilibrated
system of water coordinates that is repeated over the box, and
11
Fig. 6 BPTI solvated in water in a cubic box. Note that there is quite a lot of water,
in particular in the box corners
overlapping water molecules removed. The BPTI system will

require roughly 3,400 water molecules, which increases the number of atoms significantly. GROMACS does not use a special
pre-equilibrated system for TIP3P water since water coordinates
can be used with any modelthe actual parameters are stored in
the topology and force field. In GROMACS, a suitable command
to solvate the new box would be
genbox cp box.gro cs spc216.gro \
p topol.top o solvated.gro
The backslash means the entire command should be written
on a single line. Solvent coordinates (-cs) are taken from an SPC
water system [23], and the p flag adds the new water to the topology file. The resulting system is illustrated in Fig.6.
3.5 Adding Ions
In principle you could use the system as is, but the net charge on
the protein is unphysical in an infinite system, and many proteins
interact with counterions. There is a GROMACS program to help
us with this, but we first need an input file. GROMACS uses a
separate preprocessing program grompp to collect parameters,
topology, and coordinates into a single run input file (em.tpr)
from which the simulation is then started (this makes it easier to
move it to a another computer). Here we are not really going to
run anything, so just create an empty file called ions.mdp and
prepare an input file as:
12
Erik Lindahl
grompp f ions.mdp p topol.top c solvated.gro \

o ions.tpr
To neutralize the system and add 100mM NaCl to the output
file ions.gro, use the command
genion s ions.tpr neutral conc 0.1 \
p topol.top o ions.gro
3.6 Energy
Minimization
The added hydrogens and broken hydrogen bond network in water

would lead to quite large forces and structure distortion if molecular
dynamics was started immediately. To remove these forces it is necessary to first run a short energy minimization. The aim is not to
reach any local energy minimum, so 500 steps of steepest descent (as
mentioned in the theory section) works very well as a stable rather
than maximally efficient minimization. Nonbonded interactions and
other settings are specified in a parameter file (em.mdp); it is only
necessary to specify parameters where we deviate from the default
value (this is why we could use an empty file above), for example:
------em.mdp-----integrator=steep
nsteps=500
nstlist=10
rlist=1.0
coulombtype=pme
rcoulomb=1.0
vdw-type=cut-off
rvdw=1.0
nstenergy=10
-----------------See Note 11 contains a more detailed description of these settings. Then prepare the input file and run the energy minimization:
grompp f em.mdp p topol.top c ions.gro \
o em.tpr
<lots of output>
mdrun v deffnm em
The deffnm is a smart shortcut that uses em as the base
filename for all options, but with different extensions. The minimization will complete in a few seconds (see Note 12).
3.7 Position
Restrained
Equilibration
To avoid unnecessary distortion of the protein when the molecular

dynamics simulation is started, we first perform a 100ps equilibration run where all heavy protein atoms are restrained to their starting positions (using the file posre.itp generated earlier) while
the water is relaxing around the structure. As covered in the theory
section, bonds will be constrained to enable 2fs time steps. Other
settings are identical to energy minimization, but for molecular
13
dynamics we also control the temperature with the Bussi thermostat

[24] (see Note 13). The settings used are (see Note 14):
------pr.mdp-----integrator=md
nsteps=50000
dt=0.002
nstenergy=1000
nstlist=10
rlist=1.0
coulombtype=pme
rcoulomb=1.0
vdw-type=cut-off
rvdw=1.0
tcoupl=v-rescale
tc-grps=protein water_and_ions
tau-t=0.5 0.5
ref-t=300 300
pcoupl=parrinello-rahman
pcoupltype=isotropic
tau-p=2.0
compressibility=4.5e-5
ref-p=1.0
cutoff-scheme=Verlet
define=-DPOSRES
refcoord_scaling=com
constraints=all-bonds
-----------------For a small protein like BPTI it should be more than enough with
100ps (50,000 steps) for the water to equilibrate around it, but in a
large membrane system the slow lipid motions can require several
nanoseconds of relaxation. The only way to know for certain is to
watch the potential energy, and extend the equilibration until it has
converged. Running this equilibration in GROMACS you execute
grompp f pr.mdp p topol.top c em.gro \
o pr.tpr
mdrun v deffnm pr
This simulation will finish in a few minutes on a GPU-equipped
workstation.
3.8 Production Runs
The difference between equilibration and production run is minimal: the position restraints and pressure coupling are turned off
(see Note 15), we decide how often to write output coordinates to
analyze (say, every 5,000 steps), and start a significantly longer
simulation. How long depends on what you are studying, and that
should be decided before starting any simulations. For decent sampling the simulation should be at least ten times longer than the
phenomena you are studying, which unfortunately sometimes
14
Erik Lindahl
conflicts with reality and available computer resources. We will

perform a 10ns simulation (5 million steps), which takes about an
hour on a GPU workstation. If you are not that patient (or have a
slow machine) you can choose a shorter simulation just to get an
idea of the concepts, and the analysis programs in the next section
can read the simulation output trajectory as it is being produced.
------run.mdp-----integrator=md
nsteps=5000000
dt=0.002
nstlist=10
rlist=1.0
coulombtype=pme
rcoulomb=1.0
vdw-type=cut-off
rvdw=1.0
tcoupl=v-rescale
tc-grps=protein water_and_ions
tau-t=0.5 0.5
ref-t=300 300
cutoff-scheme=Verlet
nstxtcout=5000
nstenergy=5000
-----------------Prepare and perform the production run as (the extra option
to mdrun avoids spending too much time on writing out the
current step to frequently):
grompp f run.mdp p topol.top c pr.gro o run.tpr
<output>
mdrun v deffnm run stepout 10000
3.9 Trajectory
Analysis
3.9.1 Deviation
fromX-Ray Structure
One of the most important fundamental properties to analyze is

whether the protein is stable and close to the experimental structure. The standard way to measure this is the root-mean-square
displacement (RMSD) of all heavy atoms with respect to the X-ray
structure. GROMACS has a program to do this, as
g_rms s em.tpr f run.xtc
Note that the reference structure here is taken from the input
before energy minimization. The program will prompt both for a
fit group, and the group to calculate RMSD forchoose
Protein-H (protein except hydrogens) for both. The output
will be written to rmsd.xvg, and if you installed the Grace program you will directly get a finished graph with
xmgrace rmsd.xvg
The RMSD is also illustrated in Fig.7. It increases pretty rapidly in the first part of the simulation, but stabilizes around 0.14nm,
15
RMSD (nm)
0.2
0.15
0.1
0.05
0
Time (ns)
10
Fig. 7 Instantaneous Root-mean-square displacement (RMSD) of all heavy atoms in Lysozyme during the
simulation (solid), relative to the crystal structure. To a large extent atoms are vibrating around an equilibrium,
so the RMSD of a 1-ns running average structure (dashed gray) is a better measure
roughly the resolution of the X-ray structure. The difference is

partly caused by limitations in the force field, but also because atoms
in the simulation are moving and vibrating around an equilibrium
structure. A better measure can be obtained by first creating a running average structure (see Note 16) from the simulation and comparing the running average to the X-ray structure, which gives a
more realistic RMSD around 0.12nm (see Note17).
3.9.2 Comparing
Fluctuations
withTemperature Factors
Vibrations around the equilibrium are not random, but depend on

local structure flexibility. The root-mean-square-fluctuation
(RMSF) of each residue is straightforward to calculate over the
trajectory, but more important they can be converted to temperature factors that are also present for each atom in a PDB file. Once
again there is a program that will do the entire job:
g_rmsf s run.tpr f run.xtc o rmsf.xvg \
oq bfac.pdb
You can use the group C-alpha to get one value per residue.
Figure 8 displays both the residue RMSF from the simulation
(xmgrace rmsf.xvg), as well as the calculated and experimental
temperature factors. The overall agreement is reasonable for a protein this small and a short simulation. Longer simulations of larger
proteins can fit almost perfectly.
3.9.3 Secondary
Structure
Another measure of stability is the protein secondary structure. This

can be calculated for each frame with a program such as DSSP [25]. If
the DSSP program is installed and the environment variable DSSP
points to the binary (see Note 18), the GROMACS program do_dssp
can create time-resolved secondary structure plots. Since the program
writes output in a special xpm (X pixmap) format you probably also
need the GROMACS program xpm2ps to convert it to postscript:
do_dssp s run.tpr f run.xtc dt 50
xpm2ps f ss.xpm o ss.eps
Use the group protein for the calculation. Figure9 shows
the resulting output in grayscale, with some unused formatting
16
Erik Lindahl
RMSF(nm)
0.08
0.07
0.06
0.05
0.04
0.03
30
B-factor
25
20
15
10
5
0
10
20
30
Residue
40
50
60
Fig. 8 Top: Root-mean-square fluctuations of residue coordinates in the simulation. Bottom: The fluctuations
can be converted to X-ray temperature factors (solid), which agree quite well with the experimental B-factors
from the PDB file (dashed)
Coil
B-Sheet
B-Bridge
2000
4000
Bend
Turn
A-Helix
3-Helix
Residue
50
40
30
20
10
0
6000
8000
10000
Time (ps)
Fig. 9 Local secondary structure in BPTI as a function of time during the simulation, according to the DSSP
definition. Note how some elements periodically lose a bit of structure, but it rapidly reforms and the overall
structure is quite stable over 10ns
removed. The DSSP secondary structure definition is pretty tight,

so it is quite normal for residues to fluctuate around the well-
defined state, in particular at the ends of helices or sheets. For a
(long) protein folding simulation, a DSSP plot would show how
the secondary structures form during the simulation.
3.9.4 Radius ofGyration
andHydrogenBonds
There are two more very basic properties that are useful to analyze:
The size of the protein defined by the radius of gyration and the
number of hydrogen bonds. To calculate the radius of gyration,
use the command:
g_gyrate s run.tpr f run.xtc
17
Rgyr (nm)
1.2
1.15
1.1
# H-bonds
1.05
35
30
25
20
2000
4000
Time (ps)
6000
8000
10000
Fig. 10 Top: Radius of gyration of BPTI during 10ns simulation. This is a good measure of how compact a
structure is. Bottom: Number of hydrogen bonds inside the protein
The result will be written to the file gyrate.xvg, which includes

both the overall radius and the radii around the three axes. Similarly,
you get the hydrogen bonds with
g_hbond s run.tpr f run.xtcnum hbnum.xvg
Select the group protein twice to getall hydrogen bonds
between the protein and the protein itself. Figure10 shows the
results for both these analyses (see Note 19).
3.9.5 Making aMovie
A normal movie uses roughly 30 frames/second, so a 10-s movie

requires 300 simulation trajectory frames. To make a smooth
movie the frames should not be more than 12ps apart, or it will
just appear to shake nervously (see Note 20). In many cases it
makes sense to rerun a shorter trajectory just for the movie, but
here we just export a short trajectory from the first 500ps in PDB
format (readable by PyMOL) as
trjconv s run.tpr f run.xtc \
e 2500.0 o movie.pdb
Choose the protein group for output rather than the entire system (see Note 21). If you open this trajectory in PyMOL as PyMOL
movie.pdb you can immediately play it using the VCR-style controls on the bottom right, adjust visual settings in the menus, and
even use photorealistic ray-tracing for all images in the movie. With
MacPyMOL you can directly save the movie as a quicktime file, and
on Linux you can save it as a sequence of PNG images for assembly
in another program. Rendering a movie only takes a few minutes,
and the final product bpti.mov is included with the reference files.
18
Erik Lindahl
4 Speeding Things uptoSolve aReal Problem

Once you have gotten your feet wet you will likely want to approach
more realistic problems. One important such problem is folding of
small proteins, for instance the Villin headpiece where some
mutants fold within a microsecond [30]. This mutant contains a
special residue (norleucine), so you will need to copy the entire
amber force field directory from the installation directory of
Gromacs to your current working directory, place the file
norleucine.rtp in the amber subdirectory, and also put the file
residuetypes.dat in your current directory (so GROMACS
recognizes the new residue as a protein residue).
The first challenge you are likely to hit is that you need your
simulations to run faster to be able to reach relevant time scales.
Here we will briefly go through a couple of recommendations that
will help you achieve this.
You have probably seen that the problem is the number of steps
we must take. One way to improve performance is to make each
time step longer, but this is limited by the vibrations in the angles
involving hydrogens (try a longer timestep in your files above and
see what happens). GROMACS has a feature to remove these vibrations by replacing individual hydrogens with virtual sites. This
retains the rotation of for example CH3-groups, but removed the
fast vibrations. To enable this, use the vsite option when you
run pdb2gmx (or skip it if you want to stick to 2fs steps):
pdb2gmx f protein.pdb water tip3p \
vsite hydrogen
This will instantly enable us to take time steps up to 5fs, which
will improve your performance by 150%. Second, if you look at the
box you used for BPTI you will likely see that it would better match
the shape of the protein as a more spherical shape. Unfortunately
spheres are not periodical, but we can ask GROMACS to use a rhombic dodecahedron box instead, which is at least more spherical than a
cube and has only 71% of the cube volume. That reduces the number
of water molecules required to solvate the protein. This is difficult to
visualize in three dimensions, but Fig.11 illustrates in two dimensions how a hexagonal cell is more efficient than a square (very useful
for membrane simulations). The hexagonal box achieves the same
distance between periodic copies as a rectangular box at 86% of the
volume (see Note 22). Since we really want to push performance, we
also accept a very small margin to the box side (-d option):
editconf f conf.gro bt dodecahedron \
d 0.3 o box.gro
Finally, although PME does provide state-of-the-art electrostatics, the Villin headpiece is quite well-behaved and there are no
large mobile charges in this system. To get a bit of extra performance, this is a case where we can decide to forego PME and use
19
Fig. 11 Two-dimensional example of how a hexagonal box leads to lower volume

than a square one, with the same separation distance. In three dimensions, the
shape most similar to a sphere is the rhombic dodecahedron
reaction-field electrostatics instead. This difference is easyjust

write reaction-field instead of PME for electrostatics in your
mdp files. We also use very short cutoffs (8). The exact files used,
including an input structure from the paper [30], are included in a
separate directory.
With these settings you should be able to work energy minimization and equilibration of Villin exactly the same way as we did for
BPTI, but it will be even fasteron a single Core i7 desktop with
a GPU we getalmost a microsecond a day. Study what happens
with the protein by using the analysis tools you learned above
you should see that it starts to compact and form more hydrogen
bonds, but that it takes quite a while for secondary structure to
form. To see how far away you are from the native state you can
prepare a second TPR file using the native state as reference.
However, after a bit of fluctuation, and possible 23 days of simulation, you should also be able to reach the native state of Villin.
5 Conclusions
This chapter should hopefully provide a basic introduction to general simulations. An important lesson is that high-quality simulations require a lot of care from the userjust as with experimental
techniques the entire result can be ruined by a single sloppy step.
Further, recent techniques based on distributed computing and
markovian state models have been able to probe dynamics in the
millisecond range without extending individual simulations to
those scales [31]this will be covered in much more detail in subsequent chapters presenting metadynamics (Chapter 8) and
accelerated MD (Chapter 12). While simulations are advancing
rapidly due to the continuous development of faster computers,
the field has also been plagued by (published) simulations that
20
Erik Lindahl
have not advanced our knowledge either of simulation methods or

biomolecules. Instead of just starting a simulation and hoping for
something to happen, you should decide beforehand what you
want to study, estimate the timescales necessary or see if it can be
accomplished with more advanced methods (e.g., free energy calculations), and not start simulations until you are fairly confident
both about sampling, analysis required, and the force field accuracy. Used with caution, molecular dynamics is an amazingly powerful tool, and a great complement to experiments.
6 Notes
1. Most simulations rely on systems being ergodic, that is, the
time average of the properties of a single molecule on a long
simulation should be the same as the instantaneous ensemble
average over all molecules in an experimental measurement.
This is often (but not always) true, although it assumes our
single simulation is sufficiently long, which can be very inefficient to achieve.
2. The standard harmonic bond potentials in molecular simulations will never allow atoms to separate. However, the alternative Morse potential is supported in many programs (including
GROMACS) and will allow atoms to separate. Still, this is not
used very frequentlyif your problem involves breaking and
forming bonds it is likely a better solution to use a QM/MM
simulation.
3. The classical representations can be corrected in a number of
ways to make sure that they are faithful representations of the
real system. This is discussed in great detail in the first chapter
of the GROMACS manual, to which we refer the interested
reader. However, the really important thing in modeling is to
understand your system and decide in each case what approximations are reasonable. It is easy to add more detail (e.g., by
using quantum chemistry), but that automatically means you
lose in the other end by not getting as much sampling. The
challenge is to strike the right balance for each problem!
4. In general, most computational chemistry programs behave best
with the Linux operating system, although it is possible to run
GROMACS on Windows. When starting out, you want a standard AMD or Intel desktop. Currently (2013), you will get the
best priceperformance ratio by investing in a single-
socket
machine with fastest consumer processor you can buy, for
instance Intel Core i7 4770. You can get this for well under
$1000. GROMACS and some other codes support GPU acceleration for NVIDIA cards, so to improve performance significantly it is a good idea to add a high-end graphics card such as
21
GTX780 or GTX TITAN.Beware that this development is even

faster than the CPUs, so consult the internet to find up-to-date
hardware. Commercial Linux distribution are not requiredwe
typically use the free Ubuntu (http://www.ubuntu.com). Ifyou
are hesitant about installing Linux, get an Mac instead.
5. GROMACS is freely available from http://www.gromacs.org.
It should be quite easy to install using the step-by-step instructions, and for most common platforms there are finished binary
packages (installation might require root access, though).
PyMOL is distributed from http://www.PyMOL.org, with
binaries for Windows, Linux, and Mac OS X.The MacPyMOL
version requires a license after a trial period, but is very much
recommended for the better movie export capabilities.
Unfortunately, the Grace package is not quite as trivial to
install. The distribution site http://plasma-gate.weizmann.
ac.il/Grace/ only provides source code, so you might want to
perform an internet search for a binary for your platform.
Linux RPMs can often be found at http://www.rpmfind.net.
Grace uses Motif X11 library, but it compiles fine with the
open source clone LessTif, http://www.lesstif.org.
6. For this tutorial pretty much any structures would have been
fine too, but some of the pdb-files contain organic molecules
that are difficult to model automatically, both in GROMACS
and other programs. The key issue is to obtain and validate a
topology for your organic molecule before proceeding with
the simulation. It is often a good idea to both have a look at
the structure in a viewer, and read the text information at the
top of the PDB file to see if there are any special issues. For
6PTI, the header mentions that the last residue was not visible
at all, and only the nitrogen atom in the second last. If large
parts of the protein are inaccurate it might be better to choose
a different structure.
7. Sometimes people remove the crystal water to replace it with
their own solvent later, but this is usually a bad idea. The reason why they are visible is that these waters are tightly bound
to the structure and often form salt bridges, so if they are discarded the structure might distort before new solvent has a
chance to equilibrate in these positions. Keep the crystal water!
8. Water is a very special liquid, and actually quite difficult to
model accurately. However, biomolecular simulations are usually focusing on the protein/DNA/etc., and thus normally
prefer cheap and simple approximate solvent models to the
most accurate one. The most common such models are SPC
[23] (used with the GROMOS96 force field) and TIP3P [22]
(OPLS and Amber force fields), which both represent the
water as an entirely rigid molecule with three sites (oxygen &
two hydrogens). There are a couple of modified models such
22
Erik Lindahl
as SPC/E that improve bulk properties, but the standard

models are often preferred for interface systems like membranes. TIP4P [26] is a smart model with a fourth interaction
site offset from the oxygen, and still reasonably cheap computationally (recommended), while TIP5P [27] with five interaction sites is too expensive for most simulations.
9. pdb2gmx can be somewhat picky with the input structures,
but that is usually a good thingit will for instance not accept
proteins with missing heavy atoms. If that happens, the best
option is to find a better structure, and if that is not possible
you can try to build the missing parts with a program like
Modeller (http://salilab.org/modeller/). However, if you
have to build more than a handful of residues it is doubtful if
the resulting structure is accurate enough to simulate. For
6PTI, pdb2gmx will also issue a warning about net charge, but
that is fine. In general, all GROMACS program try to do both
double- and triple-checking of your input, so if you do not get
any warning you can be pretty confident about the correctness
of your input.
10. All GROMACS programs that write coordinates support a
number of different output formats. The default one is .gro,
simply because it has support for velocities too, but if you want
a PDB file to view for example in PyMOL you simply change
the output file extension to .pdb, when using a gromacs
program.
11. We choose a standard cutoff of 1.0nm, both for the neighborlist generation and the coulomb and LennardJones interactions. nstlist=10 means it is updated at least every 10 steps,
but for energy minimization it will usually be every step.
Energies and other statistical data are stored every 10 steps
(nstenergy), and we have chosen the more expensive
Particle-Mesh-Ewald summation for electrostatic interactions.
The treatment of nonbonded interactions is frequently bordering to religion. One camp advocates standard cutoffs are fine,
another swears by switched-off interactions, while the third
would not even consider anything but PME.One argument in
this context is that true interactions should conserve energy,
which is violated by sharp cutoffs since the force is no longer
the exact derivative of the potential. On the other hand, just
because an interaction conserves energy does not mean it
describes nature accurately. In practice, the difference is most
pronounced for systems that are very small or with large charges,
but the key lesson is really that it is a trade-off. PME is great,
but also clearly slower than cutoffs. Longer cutoffs are always
better than short ones (but slower), and while switched interactions improve energy conservation they introduce artificially
large forces. Using PME is the safe option, but if that is not fast
23
enough it is worth investigating reaction-field electrostatics,

but you should never use a plain cutoff for electrostatics. It is
also a good idea to check and follow the recommended settings
for the force field used.
12. Mdrun will write several output files: em.edr is an "energy
file" with statistical data (energies, temperature, pressure, etc.).
em.trr is a trajectory with full coordinates/velocities of the
system during the run, and em.log a log file. Depending on
the parameters (disabled here), it might also write a compressed
trajectory with low-precision coordinates only, em.xtc.
13. The Bussi thermostat is a great advance for simulations. It is

both efficient and avoids excessive fluctuations, and maintains
a correct statistical mechanics ensemble. We strongly prefer it
over the Nose-Hoover thermostats [28]. For pressure coupling we use the similar ParrinelloRahman barostat [29].
When your only goal is to get the system to a specific temperature or pressure as quickly as possible without fluctuations, you
can also consider the Berendsen weak coupling thermostat/
barostat, but these do not provide correct ensembles. For the
Bussi thermostat we can use relatively slow coupling times
(0.5ps), and the pressure coupling should be clearly slower
than this (25ps).
14. For molecular dynamics simulations the integrator is "md."
Temperature coupling has been enabled for protein and water
separately (to avoid heating the water more than the protein or
vice versa), with a 300K reference temperature. The compressibility is really a symmetric tensor, and by setting the last
three elements (off-diagonal) to 0 we disable any box shear
deformation. The last line causes grompp to include the position restraint file posre.itp generated by pdb2gmx, which
turns on position restraints. Since we are scaling the box with
pressure coupling, we also need to adjust the center-of-mass of
the reference coordinates for the position restraints with the
refcoord_scaling option. Finally, the Verlet cutoff-scheme is a
more accurate setting that also enables us to use GPU accelerators in GROMACS.
15. The easiest way to create a running average in GROMACS is
to use the g_filter program. The command g_filter nf
50 -all s run.tpr f run.xtc ol lowpass.
xtc will create a lowpass version of the trajectory (cosine
averaging over 50 frames), which then can be used as modified
input file to the g_rms program.
16. In this particular case we just used pressure coupling to get the
right density, while the production simulation is performed in
a so-called NVT ensemble (constant number of particles,
volume, and temperature). For some systems, in particular
24
Erik Lindahl
membranes and membrane proteins, it is common to enable

pressure coupling during the entire simulation in a so-called
NPT ensemble.
17. If the RMSD is significantly higher than this, or continuously
increasing, there is likely something very wrong. Start over
with the PDB file, read the headers carefully and make sure the
starting structure is accurate. In the next step, check the different energy terms and RMSD change both during minimization and position restraints. You can also use the posrefc
flag with pdb2gmx to increase the strength of the position
restraints, and extend the equilibration run.
18. The DSSP program can be obtained from http://swift.cmbi.
ru.nl/gv/dssp/. The latest version is now freely available for
everybody, but it also has a new output format. This new output format is supported by Gromacs version 4.6 and later.
Download a precompiled binary if you find a suitable one, or
compile the program and install it, e.g., in/usr/local/bin. Set
the environment variable with a command like export
DSSP=/usr/local/bin/dssp (bash shell).
19. Modern force fields no longer use special hydrogen bond
interactions, partly because it is not necessary and partly
because it is difficult to track formation/breaking of hydrogen bonds separately. Hydrogen bonds are therefore
defined from geometric criteria, typically that the distance
between the donor and acceptor atoms should be smaller
than 0.35nm, and the angle donoracceptorhydrogen
should be below 30 degrees.
20. To visualize slower phenomena such as protein folding, you

can use g_filter to smooth out motions in longer trajectories. In some cases this can lead to strange artifacts, e.g., when
averaging torsion rotation around a bond, but it is usually better than just taking raw trajectory frames with too large
spacing.
21. PyMOL loads all frames of the trajectory into memory, so if
the water molecules are included it will likely run out of memory when creating graphical representations for over 20,000
atoms repeated in 250 frames. Trajectories restricted to the
protein part can thus be much longer.
22. The volume of a rhombic dodecahedron is roughly 71% of a
cube with the same spacing, for a truncated octahedron it is
77%, and a hexagonal box is 86% of a rectangular one.
These difference can appear small, but 30% is quite significant when simulations use weeks of supercomputer time,
and it is a free lunch after all! However, not all programs
support all box shapes.
25
References
1. Alder BJ, Wainwright TE (1957) Phase transition for a hard sphere system. J Chem Phys
27:12081209
2. Rahman A, Stillinger FH (1971) Molecular
dynamics study of liquid water. J Chem Phys
55:33363359
3. McCammon JA, Gelin BR, Karplus M (1977)
Dynamics of folded proteins. Nature 267:
585590
4. Allen MP, Tildesley DJ (1989) Computer simulation of liquids. Clarendon, NewYork, NY
5. Frenkel D, Smit B (2001) Understanding
molecular simulation. Academic, NewYork, NY
6. Kaminski GA, Friesner RA, Tirado-Rives J,
Jorgensen WL (2001) Evaluation and reparametrization of the OPLS-AA force field for
proteins via comparison with accurate quantum chemical calculations on peptides. J Phys
Chem B 105:64746487
7. MacKerell AD Jr etal (1998) All-atom empirical potential for molecular modeling and
dynamics Studies of proteins. J Phys Chem B
102:35863616
8. Oostenbrink C, Villa A, Mark AE, van
Gunsteren WF (2004) A biomolecular force
field based on the free enthalpy of hydration
and solvation: the GROMOS force-field
parameter sets 53A5 and 53A6. J Comput
Chem 25:16561676
9. Wang J, Cieplak P, Kollman PA (2000) How
well does a restrained electrostatic potential
(RESP) model perform in calculating conformational energies of organic and biological
molecules? J Comput Chem 21:10491074
10. Chandler D (1987) Introduction to modern
statistical mechanics. Oxford University Press,
NewYork, NY
11. Essman U, Perera L, Berkowitz M, Darden T,
Lee H, Pedersen LG (1995) A smooth particle mesh Ewald method. J Chem Phys 103:
85778593

12. Ryckaert JP, Ciccotti G, Berendsen HJC
(1977) Numerical integration of the cartesian
equations of motion of a system with constraints; molecular dynamics of n-alkanes.
JComp Phys 23:327341
13. Hess B, Bekker H, Berendsen HJC, Fraaije
JGEM (1998) LINCS: a linear constraint
solver for molecular simulation. J Comput
Chem 18:14631472
14. Lindahl E, Hess B, van der Spoel D (2001)
GROMACS 3.0: A package for molecular simulation and trajectory analysis. J Mol Model
7:306317
15. Case DA etal (2005) The Amber biomolecular

simulation programs. J Comput Chem 26:
16681688
16. Brooks BR etal (1983) CHARMM: a program
for macromolecular energy, minmimization,
and dynamics calculations. J Comput Chem
4:187217

17. Phillips JC etal (2005) Scalable molecular
dynamics with NAMD.J Comput Chem
26:17811802
18. DeLano WL (2002) The PyMOL molecular
graphics system. DeLano Scientific, San Carlos,
CA, http://www.PyMOL.org
19. Ascenzi P etal (2003) The bovine basic pancreatic trypsin inhibitor (kunitz inhibitor): a
milestone protein. Curr Protein Pept Sci
4:231251
20. Wlodawer A etal (1987) Structure of form III
crystals of bovine pancreatic trypsin inhibitor. J
Mol Biol 198:469480

21. Frauenfelder H, Petsko GA, Tsernoglou D
(1979) Temperature-dependent X-ray diffraction as a probe of protein structural dynamics.
Nature 280:558563
22. Jorgensen WL, Chandrasekhar J, Madura JD,
Impey RW, Klein ML (1983) Comparison of
simple potential functions for simulating liquid
water. J Chem Phys 79:926935
23. Berendsen HJC, Postma JPM, van Gunsteren
WF (1981) Interaction models for water in
relation to protein hydration. In: Pullman B
(ed) Intermolecular forces. D.Reidel Publishing
Company, Dordrecht, Germany, pp331342

24. Bussi G, Donadio D, Parrinello M (2007)
Canonical sampling through velocity-rescaling.
J Chem Phys 126:014101
25. Kabsch W, Sanders C (1983) Dictionary of
protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 22:25772637
26. Jorgensen WL, Madura JD (1985) Temperature
and size dependence for monte carlo simulations of TIP4P water. Mol Phys 56:13811392
27. Mahoney MW, Jorgensen WL (2000) A five-
site model for liquid water and the
reproduction of the density anomaly by rigid,
nonpolarizable potential fuynctions. J Chem
Phys 112:89108922
28. Nos S (1984) A molecular dynamics method
for simulations in the canonical ensemble. Mol
Phys 52:255268
29. Parrinello M, Rahman A (1981) Polymorphic
transitions in single crystals: a new molecular
dynamics method. J Appl Phys 52:71827190
26
Erik Lindahl

30. Ensign DL, Kasson P, Pande V (2007)
Heterogeneity even at the speed limit of folding: large-scale molecular dynamics study of a
fast-folding variant of the villin headpiece.
JMol Biol 374:806816
31. Voelz VA, Bowman GR, Beauchamp K,

Pande VS (2010) Molecular simulation of
abinitio protein folding for a millisecond
folder NTL9 (1-39). J Am Chem Soc 132:
15261528
Chapter 2
Transition Path Sampling withQuantum/Classical
Mechanics forReaction Rates
FraukeGrter andWenjinLi
Abstract
Predicting rates of biochemical reactions through molecular simulations poses a particular challenge for
two reasons. First, the process involves bond formation and/or cleavage and thus requires a quantum
mechanical (QM) treatment of the reaction center, which can be combined with a more efficient molecular
mechanical (MM) description for the remainder of the system, resulting in a QM/MM approach. Second,
reaction time scales are typically many orders of magnitude larger than the (sub-)nanosecond scale accessible by QM/MM simulations. Transition path sampling (TPS) allows to efficiently sample the space of
dynamic trajectories from the reactant to the product state without an additional biasing potential. We
outline here the application of TPS and QM/MM to calculate rates for biochemical reactions, by means
of a simple toy system. In a step-by-step protocol, we specifically refer to our implementation within the
MD suite Gromacs, which we have made available to the research community, and include practical advice
on the choice of parameters.
Key words Protein folding, Biochemical reactions, QM/MM, Reactive paths, Rate calculations
1 Introduction
Processes such as biochemical reactions or conformational changes of
biomolecules typically occur on timescales beyond those accessible by
Molecular Dynamics (MD) simulations at atomistic detail. In many
cases, reducing the resolution of the simulation by coarse-graining
the biomolecule is not an option, as critical players such as hydrogen
bonds or hydrophobic effects involved in the reaction under investigation might be lost or are described at insufficient accuracy.
Purely classical MD simulations at atomistic resolution routinely can reach microsecond time scales. In a few recent cases,
millisecond scales were achieved, which allowed the prediction of
quantitative rates for the folding of proteins, either by highly parallel distributed computing of many short trajectories or by special
purpose high-performance computing to obtain a small number of
ultralong trajectories [13]. However, the conventionally reached
27
28
microsecond time scale is mostly insufficient to sample the process

of interest such as a conformational change or (un)folding frequently enough to compute transition rates.
The problem of too short simulation time scales is even larger
for the case for chemical reactions, in which covalent bonds are
broken or formed. Here, a classical molecular mechanical (MM)
description is not sufficient, as it relies on a harmonic potential for
a covalent bond, which does not allow dissociation of the bonded
atoms. Instead, a quantum mechanical (QM) description is
required to treat the change in bonds within the biomolecule accurately. Taking the electronic degrees of freedom into account,
however, entails substantially higher computational costs and
restricts time scales typically to picoseconds or nanoseconds, very
much depending on the theory and basis set of choice. In turn,
chemical reactions typically feature high barriers, i.e., rates at the
microsecond to millisecond scale.
From a computational point of view, the most interesting
quantity for such processes often is the reaction rate. The reason is
that, in contrast to a free energy barrier, rates are experimentally
directly accessible, and thus a straightforward comparison is possible. Also, the rate is the quantity which is physiologically most
relevant, as kinetics determine most of the biological processes.
Rates can be obtained from free energy barriers using the Arrhenius
or Eyring equations, which requires, however, the assumption of
an attempt frequency, the value of which is debated and varies with
the nature of the process and the solvent [4, 5]. An elaborate
method to directly compute reaction rates is Transition Path
Sampling (TPS) [6, 7]. TPS is an algorithm which efficiently
searches the space of transition paths between two states. From the
ensemble of sampled paths obtained from MD simulations combined with a Monte Carlo sampling scheme, reaction rates can be
obtained, without the detour of free energy barriers. TPS can be
straightforwardly used for chemical reactions treated with QM or
combined QM/MM.It has been proven to be a useful method to
obtain quantitative insight into the mechanism of, among others,
the reactions catalyzed by lactate dehydrogenase [8] and human
purine nucleotide phosphorylase [9]. We have employed QM/
MM and TPS to obtain force-dependent rates for a redox reaction,
namely, the reduction of a disulfide bond by a small reducing
agent, dithiothreitol [10], and for peptide hydrolysis [11].
In this chapter, we outline the basics of TPS, and in particular
the calculation of reaction rates based on TPS.We illustrate the
methodological details by way of a toy model, namely, three argon
atoms in a box of water. While our toy model is, for simplicity,
described solely by MM, the same strategy can be employed to a
chemical or enzymatic reaction treated by combined QM/
MM.For the reader interested in the details of a QM/MM setup
for Molecular Dynamics simulations and eventually for TPS, we
Transition Path Sampling withQuantum/Classical Mechanics forReaction Rates
29
refer to recent reviews on this subject [1214], and in combination

with TPS to our own work [10]. TPS does not restrict the QM
MM interface in any way. However, as the TPS and rate calculation
scheme presented here is based on our implementation into
Gromacs [15], only QM/MM features available within Gromacs
can be employed along with our implementation. The recent
review by Groenhof [14] is the most comprehensive introduction
into the current QM/MM implementation within Gromacs, and
reviews the available schemes to treat the interactions between the
regions described by QM and MM, and to cap the QM region in
case of covalent bonds at the QMMM interface.
2 Theory
Many processes such as chemical reactions or protein folding can
be simplified to processes with two stable states that are separated
by a single high energy barrier. In Fig.1a, regions A and B are the
two stable states, and the energy barrier is highlighted in between.
For chemical reactions, regions A and B represent the reactant and
product states, respectively. In this example, the multidimensional
space of the system is projected onto two order parameters, R1 and
R2, both of which change during the reaction. Examples for order
parameters, often distances, angles, or collective coordinates, are
given further below. A reactive trajectory (shown as a black solid
line) leads to the rare but crucial transition between A and B.The
system spends considerably longer times in the two free energy
wells of the reactant and product than in the high free energy states
between the two. Thus, while the transition of interest might only
take a few 100fs, the dwell time of the system in A or B might be
in the microsecond to second time scale. Transition path sampling
(TPS) has been developed to enhance the sampling of the rare
reactive trajectories, which are otherwise hardly harvested by conventional simulations [6, 7, 1618].
2.1 Sampling
theTransition
Path Ensemble
The idea of transition path sampling (TPS) is to sample a new transition path based on an existing (old) one (a transition path refers
to a reactive trajectory) with a Monte Carlo procedure, and the
new path is made sure to be equally weighted with the old one in
the transition path ensemble. In principle, there are many strategies to do this. For illustrating the concept of TPS, we here use the
shooting move in a deterministic simulation as an example.
(a) Defining the probability of a reactive path. In molecular simulations, the time evolution of a system is represented by an
ordered sequence of states, X(T){X0,Xt,X2t,,XT} (see
Fig.1a, black solid line). Here, t is the time increment. X(T)
consists of L=T/t+1 states, and its starting point is X0.
30
Fig. 1 Schematic description of the free energy landscape of a system and the
shooting and shifting moves in TPS. (a) A typical free energy landscape of a process is shown with two stable states (labelled with A and B) and a barrier in the
middle. R1 and R2 are two arbitrary coordinates. A transition pathway (black solid
line) connecting states A and B is given as well. The transition path is represented
by an ordered sequence of states X(T){X0,Xt,X2t,,XT}. (b) An example of
shooting moves. The two filled grey areas represent the states A and B mentioned
o
o
above. A state {qit,pit} is randomly chosen from an old transition path (solid line).
o
The momentum pit is perturbed to be pitn, where pitn=pito+p, while the
o
coordinate is unchanged with qit=qitn. From the newly generated state
n
n
{qit ,pit }, a new transition path (dashed line) is obtained by evolving the system
backward in time to zero and forward in time to T. (c) An example of forward shifting moves. A new path is generated by removing a small segment from the beginning of the old path (the starting frame, shown as a black dot, moves forward to a
new start) and evolving the system forward from the last frame to create a new
part with the same length as the removed one (the dashed line). Figure adopted
from Hierarchical Methods for Dynamics in Complex Molecular Systems, Lecture
Notes, Eds. Grotendorst et al, Juelich, 2012 with permission
Fordeterministic dynamics, the probability of a trajectory

equals to the probability of the initial state in a given ensemble,
(X0). Therefore, the probability of trajectory X(T) to be a
reactive trajectory is given below:

PAB ( X (T ) ) = h A ( X 0 ) ( X 0 ) h B ( X T ) / Z AB (T )
(1)
31
Here, hA(X)(hB(X)) is the characteristic function of region A(B).

hA(X) equals 1 if state X lies in A, and it equals zero otherwise.
ZAB(T) is the normalizing factor, the sum of all the possible reactive trajectories with length T in a given ensemble.

Z AB (T ) dX 0h A ( X 0 ) ( X 0 ) h B ( X T )
(2)
(b) Sampling the transition path ensemble by shooting. In a transition path ensemble, the distribution of transition paths is given
in Eq.1. To make sure that the correctly weighted transition
paths are sampled, the following two probabilities should
equal: the probability to generate a new transition path from a
old one Pgen(Xo(T)Xn(T)), and the probability to generate
the old transition path from the new one Pgen(Xn(T)Xo(T)).
In a shooting move, a state Xito,i[0,L], is randomly chosen.
Then, a new state Xitn is generated by adding a small perturbation to Xito. Here, the superscript o and n refer to the old
path and the new path, respectively. Note that a state X consists of the coordinate q and the momentum p, X=
{q, p}, the perturbation can be added to q or/and p. In practice, it is convenient to keep q untouched and change p by p.
As illustrated in Fig.1b, the selected state Xito={qito,pito} in
an old transition path (the solid line in Fig.1b) is changed to
Xitn={qitn,pitn}, where pitn=pito+p. Starting with Xitn,
one can evolve the system backward in time to 0 and forward
in time to T, then a new transition path is generated if it initials
from region A and ends in region B (the dashed line in Fig.1b).
The probability to generate a new transition path from an old
one is the product of four parts, the probability of the old path
in the given ensemble, the probability to generate Xitn from
Xito (Pgen(XitoXitn)), the probability of that the new path
is reactive, and the probability to accept the new transition
path Pacc(Xn(T)Xo(T)).
Pgen ( X o (T ) X n (T ) ) = PAB ( X o (T ) ) Pgen ( X iot X int ) h A ( X 0n ) hB ( X Tn )

Pacc ( X o (T ) X n (T ) )
(3)

Similarly, for generating the old path from the new one, we have
Pgen ( X n (T ) X o (T ) ) = PAB ( X n (T ) ) Pgen ( X int X iot ) h A ( X 0o ) hB ( X To )

Pacc ( X n (T ) X o (T ) )
(4)

The detailed balance of moves in trajectory space requires Pgen

(Xo(T)Xn(T))=Pgen(Xn(T)Xo(T)), which gives
32
Pacc ( X o (T ) X n (T ) )

Pacc ( X n (T ) X o (T ) )
PAB ( X n (T ) ) Pgen ( X int X iott ) h A ( X 0o ) hB ( X To )
PAB ( X o (T ) ) Pgen ( X iot X int ) h A ( X 0n ) hB ( X Tn )
(5)

This condition can be satisfied using a Metropolis criterion [19]

PAB ( X n (T ) ) Pgen ( X int X iot ) h A ( X 0o ) hB ( X To )
(6)
Pacc ( X (T ) X (T ) ) = min 1,
PAB ( X o (T ) ) Pgen ( X iot X int ) h A ( X 0n ) hB ( X Tn )

o
Note that the old path is reactive, i.e., hA(X0o)=1 and hB(XTo)=1.
Equation6 can be simplified as
( X int ) Pgen ( X int X iot )
Pacc ( X (T ) X (T ) ) = h A ( X ) hB ( X ) min 1,
( X iot ) Pgen ( X iot X int )

o
n
0
n
T
(7)
Here, we apply Eq.1 and the fact that the probabilities of the states
on the same path in deterministic dynamics are the same. Although
Eq.7 is obtained based on deterministic dynamics, it can be also
inferred based on a general dynamics [18]. In the implementation
of shooting moves, a symmetric generation probability is normally
ensured, and thus Pgen(XitoXitn)=Pgen(XitnXito). Specific
strategies are always applied to ensure that states Xito and Xitn are
within the same microcanonical ensemble, i.e., (Xito)=(Xitn).
Thus, the acceptance probability becomes

Pacc ( X o (T ) X n (T ) ) = h A ( X 0n ) hB ( X Tn )
(8)
This equation states that any new trajectory will be accepted if it

initiates from region A and ends in region B.
A new path can be generated by evolving forward from the last
frame (forward shifting move, see Fig.1c) or backward from the
starting frame (backward shifting move) of the old path to grow a
new path segment with a certain length and then deleting a path
segment with the same length from the other end to maintain a
fixed total length. Such shifting moves can be combined with
shooting moves to improve the sampling efficiency.
2.2 Computing
Rate Constants
In this section, we explain how to obtain rate constants from the

transition path ensemble [17]. Given a system with two stable states
A and B, which are separated by a single high energy barrier, molecules transit from one state to the other at equilibrium, while the
populations of states remain unchanged. Since such transitions
arerare, the time correlation function, C(t), relates to the reaction
time of the system (rxn(kAB+kBA)1) via the following formula [20]
C (t ) hB (1 exp {t / rxn })
(9)
33
If the time required for a system to cross the energy barrier and
commit to the other stable state (mol) is far smaller than the reaction time of the system (i.e., mol<<rxn), C(t) scales linearly in the
intermediate time region, and we have
C (t ) kABt ,
mol < t << rxn
(10)
For a system at equilibrium, C(t) characterizes the conditional

probability to find the system in state B at time t, if it was in state
A at time zero, and is defined as follows
h A ( X 0 ) hB ( X t )
C (t )
(11)
hA ( X 0 )
Here, is the ensemble average of all initial states. In deterministic dynamics, C(t) can be written in terms of the probability of all
initial states (X0):
C (t ) =
dX ( X ) h ( X ) h ( X )
dX ( X ) h ( X )
0
(12)
Equations10 and 12 together provide a way to calculate the forward reaction rate constant kAB by molecular simulations. One can
simply run a large set of simulations that start in region A and are
of the same time length t, and then count the probability of the
end state to be in region B, which gives the value of C(t). The
derivative of C(t) over time gives the rate constant. However, this
apparently involves numerous computational efforts.
If region B can be defined by an order parameter (X), and the
distribution of the end states, i.e., X(t), along the order parameter
P(, t) is known, C(t) is simply the integral of P(, t) along over
region B.
C (t ) =
_ max
_ min
dP ( ,t ) .
(13)

Here, _min and _max are the lower and upper bound of region
B along . P(, t) is given by
P ( ,t ) =

dX ( X ) h ( X ) ( X (t ) ) ,
dX ( X ) h ( X )

0
(14)
where (X) is Diracs delta function. P(, t) can be divided into

several overlapped windows, and its distribution in each window
can be estimated separately. The distribution of P(, t) over the
34
whole range of is then obtained by connecting all windows.

Ineach window, transition path sampling can be applied to enhance
the sampling of paths that connect region A and the window
region. Therefore, computational efforts to compute C(t) are
dramatically reduced.
The above-mentioned method can only compute C(t) in time
t at a time, and the evaluation of kAB requires C(t) at different times
to be evaluated. Therefore, it is laborious. Fortunately, C(t) can be
factorized to be written as [18]
C (t ) =

hB (t )
hB (t
AB
C (t ) ,
AB
0 <t <T
(15)

where AB denotes an average on the ensemble of the reactive

paths, which start in region A and visit region B within the time
length of T.T is the time length of the transition path. hB(t)AB is
then the proportion of reactive paths whose configuration at time
t belongs to region B, and can be estimated by a single transition
path sampling run. Only C(t), the C(t) at time t(t<T), is needed
to be evaluated.
Combining Eqs.10 and 15, the rate constant is given by
kAB =
d hB (t )
hB (t
AB
/ dt
AB
C (t ) ,
mol < t << rxn
(16)

dhB(t)AB/dt should show a plateau in the intermediate time

range.
3 Materials
A GROMACS-4.0.7 package [15] with a TPS implementation can
be downloaded from http://wenjin.people.uic.edu/download/
Gromacs4_tps_patch.tar.gz, which is implemented by Dr. Wenjin
Li and currently maintained by him as well (see Note 1). The package can be installed by following the installation instructions of the
original GROMACS-4.0.7 version at http://www.gromacs.org. A
Linux or Unix system is required for compilation, as well as FFTW
libraries.
4 Methods
In this section, we will describe how to (1) establish a toy system, (2)
define the stable basins, (3) obtain the hB(t)AB curve, (4) obtain the
P(, t) distribution, (5) calculate rate constants, and (6) monitor
TPS.All the files necessary to complete this tutorial are available at
http://wenjin.people.uic.edu/download/example_3_Ar.tar.gz.
35
Fig. 2 Simulation setup of the toy system. Black spheres: Ar atoms. Grey lines: water molecules
To complete this tutorial, we assume the reader to have a basic

knowledge of GROMACS and experience in the use of a Linux or
Unix operation system.
4.1 A Toy System
We here will illustrate how to use TPS to calculate the rate constant
of a rare event with a toy system, which consists of three Ar atoms
in a water box (see Fig.2). All three Ar atoms are lying in a line
along the Z-axis. Atoms 1 and 3 are held by position restraints
along the X-, Y-, and Z-axis, while atom 2 is restrained along the
X- and Y-axis, but free to move along the Z-axis. Position restrains
were switched on by setting define=-DPOSRES in the .mdp file,
with parameters for position restraints given in posre.itp. Atoms 1
and 3 are separated by approximately 1.0nm. Due to the van der
Waals interaction with the other two Ar atoms, atom 2 has two
preferred positions (or stable basins). One position is about 0.2nm,
the other 0.8nm away from atom 1. There is a relatively high barrier between the two minima. Atom 2 can overcome the attraction
of one Ar atom and transit from one stable basin to the other.
Here, we will estimate the rate of these transitions with TPS.The
parameters for van der Waals interaction between two Ar atoms
have been modified to unrealistic values (see file ffoplsaanb.itp) to
increase the barrier between the two minima to make sure that the
transition is a rare event (see Note 2). Therefore, we here are
looking at an unphysical toy model to solely focus on the procedure to run TPS with the modified GROMACS package.
36
4.1.1 Definition
ofStableBasins
TPS requires reasonable definitions of the stable basins in terms of

one or multiple order parameters. The chosen order parameter
should be able to distinguish the two stable basins, but must not
necessarily be a good reaction coordinate. Here, the order parameter we choose to distinguish the two minima is d=d12d23, where
d12 is the distance between atoms 1 and 2, and d23 is the distance
between atoms 2 and 3. Then, region A is defined as 1<d<0.5
and region B as 0.5<d<1. The modified Gromacs version includes
a section to define these TPS parameters. To define the stable states
mentioned above, the parameters are (see Note 3),
=========================
tps_npost
=4
tps_grps1
= a_1 a_2
tps_grps2
= a_2 a_3
tps_dimension
= one
tps_weight_dim
=1
tps_initial_max
= -0.5
tps_initial_min
= -1
tps_final_max
=1
tps_final_min
= 0.5
-1
=========================
Here, tps_grps1 and tps_grps2 define the groups to build the

order parameters. Currently, the order parameters consist of only distances between two atoms (or two groups of atoms). tps_grps1 specifies the first group, while tps_grps2 specifies the second group. The
coordinate is the distance between the ith group in tps_grps1 and
tps_grps2. For example, d12 is calculated by the distance between the
first group in tps_grps1 and tps_grps2. a_1, a_2, and a_3 are the name
of atom 1, atom 2, and atom 3in the index file. tps_dimension=one
means the two coordinates specified by tps_grps1 and tps_grps2
are combined into one parameter using the weights in tps_
weight_dim. tps_weight_dim
=
1 -1 means the order parameter
d=1d12+(1)d23 or d12d23 (see Note 4). The values of the
upper bound and lower bound of region A and region B are given by
tps_initial_max, tps_initial_min, tps_final_max, and tps_final_min.
The number of groups in tps_grps1 and tps_grps2 should be the same.
tps_npost is the total number of groups in tps_grps1 and tps_grps2.
4.2 Obtaining
anInitial Transition
Path
In TPS, the shooting and shifting moves are based on an initial

path, which is not necessary to be physically meaningful, as the subsequent TPS will allow to relax towards more representative paths
(see Note 5). There are many ways to get an initial path. Here, we
generate the first path by shooting forward and backward from a
structure near the transition state to ensure that we can get an initial
reactive path with high probability. Velocities are adapted from a
Boltzmann distribution at the given temperature. The setting of the
37
parameters in the TPS section are (the complete parameter file is

available in tps_ini.mdp provided in the tutorial package),
=====
Part of tps_ini.mdp =====
tps_npost
=4
tps_grps1
= a_1 a_2
tps_grps2
= a_2 a_3
tps_dimension
= one
tps_weight_dim
=1
tps_initial_max
= -0.5
tps_initial_min
= -1
tps_final_max
=1
tps_final_min
= 0.5
-1
tps
= rand_ini
tps_maxcycle
=5
tps_maxshoot
= 10
tps_endpoint
= yes
tps_kin_ref
= 100
tps_Temperature
= 300
tps_forward_steps
= 400
tps_backward_steps
= 400
tps_maxframe
=1
tps_ntrrout
=1
=========================
tps=rand_ini defines the attempt to get an initial transition

path. tps_maxcycle and tps_maxshoot specifies the number of sampling circles and the number of samples in each circles. Here, we try
510=50 times to get an initial path. The search will stop immediately if we find one. tps_endpoint=yes means the endpoint of the
path should be in region B. tps_forward_steps and tps_backward_
steps specify the number of frames saved for forward and backward
shooting, respectively. The total frames of the trajectory generated
will be the sum of them. Here, the number of frames in the transition path will be 800. The frequency of saving frames in the transition path is defined by nstxout=10 and the integration step is 2fs
(see Note 6). Therefore, we will obtain a reactive path with 800
frames, with an interval between each frame of 20fs and a total
length of t=16 ps. tps_Temperature=300 produces initial atomic
velocities from a Boltzmann distribution at a temperature of 300K.
tps_kin_ref specifies the amount of perturbation to the momenta of
the atoms in the frame to which a shooting move is applied. The
value will affect the acceptance ratio of shooting moves, with larger
perturbations leading to smaller acceptance ratios. tps_maxframe
isthe number of frames in the input trajectory. In this case, tps_
maxframe=1 because the input is a .gro file which contains only
one frame. tps_ntrrout=1 makes the MD code saving every reactive
trajectories.
38
The following input commands initiate the search for an

initialpath:
tar -zxvf example_3_Ar.tar.gz
cd example_3_Ar && mkdir -p tps/initial
grompp -f tps_ini.mdp -c 3_Ar.gro -n index.ndx -p topol.top -o tps/initial/tps.tpr
cd tps/initial && mdrun -s tps.tpr -rerun ../../TS.gro -deffnm tps_output
The input structure is read via option -rerun. TS.gro is the

structure near the transition region. 3_Ar.gro is a structure at
region A. index.ndx and topol.top define the index of groups used
in the .mdp file and the topology of the system, respectively.
Tutorials to prepare these files can be found elsewhere [21] and is
out of the scope of this chapter. We therefore provide them in the
tutorial package. This run will take about 10min on a single processor to generate an initial transition path saved as traj_0.trr. We
rename it as tps.trr by executing
mv traj_0.trr ../tps.trr
4.3 Obtaining
thehB(t)ABCurve
A requisite to compute the rate constant using TPS is the flux versus time, or the hB(t)AB curve, and the probability distribution
along an order parameter P(, t), or specifically P(d) in this case,
which is then used to calculate the value of C(t) at a specific time t
(see Theory). With an initial path at hand, we can start TPS to
obtain these ingredients for the rate constant calculations. The settings for this purpose are:
=====Part of tps.mdp =====
tps_npost
=4
tps_grps1
= a_1 a_2
tps_grps2
= a_2 a_3
tps_dimension
= one
tps_weight_dim
= 1 -1
tps_initial_max
= -0.5
tps_initial_min
= -1
tps_final_max
=1
tps_final_min
= 0.5
tps
= normal
tps_maxcycle
= 150
= 10
tps_maxshoot
tps_maxshift
= 10
tps_endpoint
= no
tps_kin_ref
= 100
tps_reput_length
= 300
tps_maxframe
= 800
tps_ntrrout
=0
=========================
39
Fig. 3 Results for the hB(t)AB curve. (a) Black curve: the averaged hB(t)AB curve. Grey curves: the five
hB(t)AB curves obtained from five independent samplings. (b) The derivative of the hB(t)AB curve shows a
plateau, indicating a length of 16ps to be sufficient. Grey: the derivative of the black curve in a. Black: the
smoothed curve of the grey one by averaging over five nearby points
tps=normal means that we now perform TPS with an initial

path already at hand. tps_maxcycle specifies the number of TPS
cycles. In each cycle, we perform several shooting moves and shifting moves, and the number of these moves is specified by tps_maxshoot and tps_maxshift, respectively. tps_endpoint=no means the
end structure of a reactive trajectory, given the trajectory starts in
A, may not be in B, but the structure should reach B at some point
within the trajectory (see Theory). tps_reput_length specifies the
maximum shifting length in shifting moves. tps_ntrrout =0 means
no intermediate reactive trajectories are saved.
Performing the TPS requires executing the following
commands:
cd ../../ && grompp -f tps.mdp -c 3_Ar.gro -n index.ndx -p topol.top -o tps/tps.tpr
cd tps && mdrun -s tps.tpr -rerun tps.trr -deffnm tps_output
We read the initial reactive path again via the option -rerun.
Here, we run 150 cycles of TPS, with 10 shooting moves and 10
shifting moves in each cycle. In total there are 3,000 TPS runs.
This will take about 4 days to complete on a single standard processor. The results of the hB(t)AB curve is saved in hahb.dat.
To obtain an accurate hB(t)AB curve, we recommend the reader to
run five independent simulations (see Note 7), and to then combine the resulting five hahb.dat files into one by simple averaging
(Fig.3a). Here, the derivative of hB(t)AB reaches a plateau at 13ps
with dhB(t)AB/dt=0.1 ps1 as shown in Fig.3b (see Note8).
4.4 Obtaining
theP(, t) Distribution
In the next step, we run TPS in different windows to obtain the

distribution of the end points of transition paths (the P(d) distribution in Eq.14). Here, we set the length of the trajectory t to
be t=6ps, with each path containing 300 frames, which is much
40
shorter than 16ps or 800 frames and saves computational cost

(see Note 9). Therefore, we read hB(t)AB=0.14 from Fig. 3a, as
t=6ps. In order to get the P(d) distribution, we divide the
configuration space into five windows, which are defined as window 1: 1<d<0.45, window 2: 0.55<d<0.15, window
3: 0.25<d<0.25, window 4: 0.15<d<0.55, and window 5:
0.45<d<1. A small overlap between adjacent windows is necessary to merge the distributions of adjacent windows into one.
By deleting the segments from the termini of the transition path
obtained above (16ps long), one can easily get an initial path of
6ps long (see Note 10). The sampling procedures in each window are similar. For each window, we define the correct region B
and adjust tps_kin_ref and tps_reput_length to maintain a reasonable acceptance ratio. The parameter files for all windows are provided in the tutorial package and are named as tps_win1.mdp to
tps_win5.mdp. We here show the TPS parameters for window 5 as
an example to illustrate the procedure.
=====Part of tps_win5.mdp =====
tps_npost
=4
tps_grps1
= a_1 a_2
tps_grps2
= a_2 a_3
tps_dimension
= one
tps_weight_dim
=1
-1
tps_initial_max
= -0.5
tps_initial_min
= -1
tps_final_max
=1
tps_final_min
= 0.45
= normal
tps
tps_maxcycle
= 300
= 10
tps_maxshoot
tps_maxshift
= 10
tps_endpoint
= yes
= 200
tps_kin_ref
tps_reput_length
= 100
tps_maxframe
= 300
tps_ntrrout
=0
===========================
To obtain the endpoint distribution, we need to make sure the

endpoints of the transition path to be within the defined region B
by setting tps_endpoint=yes. The simulation is started as follows:
cd ../ && mkdir win5
grompp -f tps_win5.mdp -c 3_Ar.gro -n index.ndx -p topol.top -o win5/tps_win5.tpr
cd win5 && mdrun -s tps_win5.tpr -rerun tps_win5.trr -deffnm tps_output
Here, tps_win5.trr is the constructed initial transition path.

The endpoints of each transition path are saved in endpoint.dat.
41
Fig. 4 Calculation of P(d) through TPS in windows. (a) Distribution of P(d) in different windows. (b) The
connected distribution of P(d) over the whole configuration space. Dashed grey lines: the boundaries
between regions A and B and the transition region
We recommend the reader to run three independent simulations

for each window and then collect all the endpoints into a single file
of endpoint.dat. From the endpoint.dat file, the distribution along
d can be easily obtained. The distributions of in the overlapped
regions are the same but weighted differently. Therefore, we can
connect all windows by re-weighting them properly. The connected window is then normalized. The distributions in different
windows and the normalized distribution are shown in Fig.4. By
integrating the distribution in the range of 0.5<d<1, we obtain
a value of C(t) of 0.00056 (see Theory).
4.5 Calculating
theRate Constant
Given C(t) is 0.00056, hB(t)AB is 0.14, and dhB(t)AB/dt is

0.1ps1, we get a rate constant k of 4.0104ps using Eq.16
(see Note 11).
4.6 Monitoring TPS
The mdrun command will generate four output files that help to
monitor the progress of the sampling: acc.dat, endpoint.dat, hahb.
dat, and summary.dat. They are explained below:
acc.dat: It summarizes the number of shooting trials, the number
of successful shooting trials, the number of shifting trials, and the
number of successful shifting trials at each frame. It also includes
the acceptance ratio for shooting and shifting.
endpoint.dat: It gives the endpoints of the transition paths in the
value of the order parameter, which is used to calculate P(, t)
when tps_endpoint=yes.
hahb.dat: It gives the hB(t)AB curve when tps_endpoint=no.
summary.dat: It summarizes the overall number of shooting and
shifting cycles and their acceptance ratio. An example is given
below:
42
=============== summary.dat ==================
The totol TPS cycle is -----------------------------3000
The totol shooting cycle is ------------------------1524
The totol leftshift cycle is -----------------------747
The totol rightshift cycle is ---------------------729
The totol acceptance is ---------------------------0.3076667
The acceptance for shooting is --------------------0.0577428
The acceptance for leftshift is ------------------0.5689424
The acceptance for rightshift is ------------------0.5624143
===================================
Here, the acceptance for shooting is quite low, around 0.06.

Adjusting tps_kin_ref and tps_reput_length allows to tune the probability of generating a reactive trajectory. A higher acceptance ratio
can be achieved by shortening the length of the transition path
(see Note 12). To achieve a better sampling efficiency, the acceptance for shooting it recommended to be about 0.4 [22].
5 Notes
1. The modified GROMACS package supports simulations on
only a single CPU and not in parallel, as neither domain
decomposition nor particle decomposition are supported
inthe current implementation.
2. Equation10 is based on the assumption that the barrier is so
high that the time of the actual transition is much smaller than
the inverse of the rate constant. Therefore, Eq.10 is only applicable to systems with high energy barriers, i.e., of several kBT.
3. For many systems, the choice of an order parameter is trivial.
One can run a relatively long simulation at the two stable states,
and then find an order parameter to distinguish the stable states
by inspection of the coordinate spaces that the two simulations
sampled at both basins. Usually, an inspection by eye is enough.
If not, principle component analysis [23] can assist in identifying an order parameter. Once an order parameter is found, one
defines the two stable states according to their distribution of
the sampled configuration along the order parameter. Make
sure that the two basins are separated and cover the major part
of the sampled configurations in that state.
4. One can use multiple coordinates to define regions A and B if
the interest is to investigate the mechanism of the transition
process rather than the rate constant. If one want to get the
rate constant, region A can be defined with multiple coordinates, while region B is preferably defined with a single coordinate, as this reduces the computational expense. If defining
region B by multiple coordinates is nevertheless essential, the
distribution of P(, t) is required in the multidimensional
43
space, which might be feasible but is not recommended. It is

generally beneficial to invest some efforts to find a single coordinate to define region B.The coordinates to define regions A
and B are not required to be identical.
5. One can generate an initial transition path in many ways,
depending on the system under investigation. In general, one
can apply a bias to the system to enforce the transition to happen with high probability and short transition times. The bias
can be from for example high temperature [24], replica
exchange [24], position restraints [10], steering forces [26],
conformational flooding [27], or metadynamics [28]. The
resulting transition path is a biased transition path, but the bias
can be removed gradually [26], or by generating an unbiased
path by TPS with shooting moves starting from one frame
(e.g., a frame within or close to the transition state ensemble)
of the biased path.
6. The number of steps for a simulation defined by nsteps in the
.mdp file should be larger than the total length of the TPS
trajectory. Here, nsteps should be no less than 8,000 given the
integral timestep of 2fs.
7. The TPS simulation will generate files with predefined names
in the working directory. If one runs multiple independent
simulations, it is recommended to start them from separate
directories to avoid overwriting output files.
8. In addition to the hB(t)AB curve, the simulation will harvest
an ensemble of transition paths (one can save the transition
paths by setting tps_ntrrout to a positive integer to specify the
frequency of saving the transition paths). Based on the transition path ensemble, the mechanism of the studied rare events
can be elucidated at the atomistic level. Applications include
(just to name a few) a reaction catalyzed by lactate dehydrogenase [8], the -hairpin folding [25], and the chorismatemutase-
catalyzed conversion of chorismate into prephenate
[29]. If committor probabilities of the frames in the transition
path ensemble are estimated, transition states [10, 29] and
reaction coordinates [30] can be identified as well.
9. Choosing a small t reduces the computational costs of the
TPS, but increases the uncertainty of hB(t)AB. As a compromise, we recommend to choose t such that hB(t)AB is about
0.2. Usually, the cost to calculate C(t) is several times higher
than the cost to obtain hB(t)AB. For this reason, it is worth to
invest more efforts into an accurate hB(t)AB curve, and to
then use a smaller t to calculate C(t).
10. The initial path for each window to obtain the hB(t)AB curve
can be obtained following the same procedure as the one for
the initial path for the TPS.Alternatively, the initial path for
44
window 5 should have 300 frames, which can be constructed

by shortening the transition paths sampled in the section of
obtaining the hB(t)AB curve, which comprises 800 frames.
Then, the initial path for window 4 can be obtained from one
of transition paths sampled in window 5. The initial path for
other windows can be taken from one of transition paths from
the subsequent window.
11. As a way of validating the computed rates, one can vary t and/
or t as well as the definitions of regions A and/or B to test if
the results are quantitatively consistent.
12. It is not necessary to randomly select a frame from the entire
transition path to do shooting moves. Specifying the range from
which shooting frames are selected allows to increase the acceptance ratio. One possibility is to choose points near the previous
shooting point from which a reactive trajectory has been generated. Alternatively, one can choose points only from the barrier
region to improve the acceptance ratio. In this case the range to
choose shooting points is variable, and the probability to accept
a reactive trajectory needs to be modified accordingly [25].
Acknowledgment
We are grateful to the Klaus Tschira Foundation for financial
support.
References
1. Lane TJ, Bowman GR, Beauchamp K etal
(2011) Markov state model reveals folding and
functional dynamics in ultra-long MD trajectories. J Am Chem Soc 133:1841318419
2. Bowman GR, Pande VS (2010) Protein folded
states are kinetic hubs. Proc Natl Acad Sci U S A
107:1089010895
3. van der Spoel D, Seibert MM (2006) Protein
folding kinetics and thermodynamics from
atomistic simulations. Phys Rev Lett 96:238102
4. Best RB, Hummer G (2006) Diffusive model
of protein folding dynamics with Kramers turnover in rate. Phys Rev Lett 96:228104
5. Popa I, Fernndez JM, Garcia-Manyes S (2011)
Direct quantification of the attempt frequency
determining the mechanical unfolding of ubiquitin protein. J Biol Chem 286:3107231079
6. Dellago C, Bolhuis PG, Csajka FS etal (1998)
Transition path sampling and the calculation of
rate constants. J Chem Phys 108:1964
7. Dellago C, Bolhuis PG, Chandler D (1998)
Efficient transition path sampling: application
to Lennard-Jones cluster rearrangements.

JChem Phys 108:9236
8. Quaytman SL, Schwartz SD (2007) Reaction
coordinate of an enzymatic reaction revealed
by transition path sampling. Proc Natl Acad Sci
U S A 104:1225312258
9. Saen-Oon S, Quaytman-Machleder S, Schramm
VL etal (2008) Atomic detail of chemical transformation at the transition state of an enzymatic
reaction. Proc Natl Acad Sci U S A
105:1654316548
10. Li W, Grter F (2010) Atomistic evidence of
how force dynamically regulates thiol/disulfide
exchange. J Am Chem Soc 132:1679016795
11. Xia F, Bronowska AK, Cheng S etal (2011)
Base-catalyzed peptide hydrolysis is insensitive to mechanical stress. J Phys Chem B
115:1012610132

12. van der Kamp MW, Mulholland AJ (2013)
Combined quantum mechanics/molecular
mechanics (QM/MM) methods in computational enzymology. Biochemistry 52:27082728

13. Steinbrecher T, Elstner M (2013) QM and
QM/MM simulations of proteins. In: Monticelli
L, Salonen E (eds) Biomolecular simulations.
Humana Press, NewYork, pp91124
14. Groenhof G (2013) Introduction to QM/MM
simulations. In: Monticelli L, Salonen E (eds)
Biomolecular simulations. Humana Press,
NewYork, pp4366
15. Hess B, Kutzner C, Van Der Spoel D etal (2008)
GROMACS 4: algorithms for highly efficient,
load-balanced, and scalable molecular simulation. J Chem Theory Comput 4:435447
16. Bolhuis PG, Dellago C, Chandler D (1998)
Sampling ensembles of deterministic transition
pathways. Faraday Discuss 110:421436
17. Dellago C, Bolhuis PG, Chandler D (1999) On
the calculation of reaction rate constants in the
transition path ensemble. J Chem Phys 110:6617
18. Dellago C, Bolhuis PG, Geissler PL (2002)
Transition path sampling. Adv Chem Phys
123:178

19. Metropolis N, Rosenbluth AW, Rosenbluth
MN etal (1953) Equation of state calculations
by fast computing machines. J Chem Phys 21:
1087

20. Chandler D (1978) Statistical mechanics of
isomerization dynamics in liquids and the transition state approximation. J Chem Phys 68:2959
21. Lindahl EL (2008) Molecular dynamics simulations. In: Kukol A (ed) Molecular modeling of
proteins. Humana Press, Totowa, NJ, pp323
45

22. Bolhuis PG, Chandler D, Dellago C etal
(2002) Transition path sampling: throwing
ropes over rough mountain passes, in the dark.
Annu Rev Phys Chem 53:291318
23. Jolliffe I (2005) Principal component analysis.
Wiley Online Library
24. Dellago C, Bolhuis PG, Geissler PL (2006)
Transition path sampling methods. In:
Computer simulations in condensed matter
systems: from materials to chemical biology,
vol 1. Springer, Berlin, pp349391
25. Bolhuis PG (2003) Transition-path sampling
of -hairpin folding. Proc Natl Acad Sci U S A
100:1212912134
26. Hu J, Ma A, Dinner AR (2006) Bias annealing:
a method for obtaining transition paths de
novo. J Chem Phys 125:114101
27. Grubmller H (1995) Predicting slow structural transitions in macromolecular systems:
conformational flooding. Phys Rev E 52:2893
28. Laio A, Gervasio FL (2008) Metadynamics: a
method to simulate rare events and reconstruct
the free energy in biophysics, chemistry and
material science. Rep Prog Phys 71:126601
29. Crehuet R, Field MJ (2007) A transition path
sampling study of the reaction catalyzed by the
enzyme chorismate mutase. J Phys Chem B
111:57085718
30. Ma A, Dinner AR (2005) Automatic method
for identifying reaction coordinates in complex
systems. J Phys Chem B 109:67696779
Chapter 3
Current Status of Protein Force Fields for Molecular
Dynamics Simulations
Pedro E.M. Lopes, Olgun Guvench, and Alexander D. MacKerell Jr.
Abstract
The current status of classical force fields for proteins is reviewed. These include additive force fields as well
as the latest developments in the Drude and AMOEBA polarizable force fields. Parametrization strategies
developed specifically for the Drude force field are described and compared with the additive CHARMM36
force field. Results from molecular simulations of proteins and small peptides are summarized to illustrate
the performance of the Drude and AMOEBA force fields.
Key words Force field, Molecular dynamics, Drude polarizable force field, CHARMM, AMOEBA,
AMBER, GROMOS, OPLS, NAMD, Electronic polarization
Introduction
Classical molecular dynamics (MD) simulations of proteins using
empirical force fields have reached a mature state after 35 years of
development and are now widely used as tools to investigate their
structure and dynamics under a wide variety of conditions. These
include studies of ligand binding, enzymatic-reaction mechanisms,
protein folding and unfolding, and proteinprotein interactions.
Fundamental to such simulations is determination of the time
evolution of the systems energy (protein for example) as a function of its atomic coordinates. An accurate description of the
energy is thus required, since the lower energy states are expected
to be populated. The gradient of the energy function, which is
differentiable, is related to the forces acting on individual atoms.
In chemistry the set of potential energy functions from which the
forces are derived is commonly referred to as a force field (FF).
As a result of many years of careful refinement, current additive
protein energy functions are of sufficient quality that they may be
used predicatively for studying protein dynamics and protein
protein interactions and in pharmacological applications [1]. It is
clear that the next major step in advancing protein force field accuracy
47
48
Pedro E.M. Lopes et al.
requires a different representation of the molecular energy surface.

Specifically, the effects of charge polarization must be included, as
fields induced by ions, solvent, other macromolecules, and the protein itself will affect electrostatic interactions [26].
Our goal here is to provide an update of the newest developments that have occurred in the field of FF-based MD simulations
of proteins since our last review was published [7]. Previously, we
focused on the functional forms of additive FFs, strategies for
parameter optimization, methodologies to perform MD simulations such as pressure and temperature control, and software packages available for MD simulations. Also briefly mentioned were
efforts to extend additive FFs to other biomolecules. The FFs
detailed were the Amber, CHARMM, GROMOS, and OPLS-AA
additive protein FFs, with particular emphasis on CHARMM
because of our continuing role in its development. While the
present review begins with a brief update on the status of additive
protein FFs, here we primarily focus on the latest developments in
the inclusion of electronic polarizability into protein FFs. Emphasis
is placed on the CHARMM Drude polarizable FF and the polarizable AMOEBA FF, and we direct interested readers to other recent
reviews [5, 6, 8, 9]. Also of interest may be new improvements in
the Amber family of FFs [10], and, to our knowledge, no new
reviews on the OPLS-AA or GROMOS protein FFs have appeared
since our previous review in this series.
A general familiarity with molecular mechanics and dynamics
and their applications to proteins is assumed. Simulation methods
for proteins are well established, with many good textbooks and
monographs covering the basics [1117]. The reader is also
referred to Chapter 1 of this volume.
Current Status of Additive Force Fields

Since the last review in this series [7], some notable developments
have been made to additive FFs for proteins. Below, brief descriptions of the improvements introduced to two of the major additive
FFs for proteins, CHARMM and Amber, are given.
2.1 CHARMM
Force Field
The CHARMM additive all-atom FF has been in development since

the early 1980s [18] and has achieved a substantial degree of completeness with regard to coverage of chemical space. Apart from proteins [19, 20], it supports nucleic acids [2123], lipids [2426], and
carbohydrates [2730], allowing simulations on all commonly
encountered motifs in biological systems. It has also been extended
to cover the wide range of the chemical space required to study compounds common in medicinal chemistry through the CHARMM
General FF (CGenFF) [31]. The CHARMM additive FF for proteins recently underwent a significant update that culminated in the
C36 version of the FF, as detailed below [20].
Current Status of Protein Force Fields for Molecular Dynamics Simulations
49
Long simulations with the additive C22/CMAP FF [19, 32, 33]

had shown that certain fast-folding proteins would reach the
native state, when started from a completely unfolded configuration
(e.g., Villin headpiece subdomain) [34]. However, significant deficiencies were also found. Examples of problems included misfolding
encountered in long simulations of the pin WW domain and differences in the Villin folding mechanism from the experimental results
[34, 35]. In the case of the WW domain, free energy calculations
showed that the misfolded states had lower free energies than the
folded state, confirming that the energy function could be further
improved [36]. A number of studies had suggested that such differences could be the result of small inaccuracies in the energy of
the backbone, resulting in one structure being favored, and that
this behavior could be corrected with minor adjustments to the
backbone potential [3740]. Best et al. reported a revised set of
CHARMM all-atom protein FF parameters (C36) that represents
a significant improvement in the potential energy surface, while
keeping the same functional form [20]. The improvements that
were introduced included (1) a new backbone CMAP potential,
optimized against experimental data on small peptides and larger,
folded proteins and (2) new side-chain dihedral parameters optimized using QM energies for dipeptides, conformational sampling
in the model system (Ala)4-X-(Ala)4 [41] and NMR data from
unfolded proteins. Other improvements relative to the previous
C22/CMAP protein FF included LennardJones (LJ) parameters
for aliphatic hydrogens [42], internal parameters for the guanidinium ion [43], and improved parameters for tryptophan [44].
Changes of the backbone and side-chains were done simultaneously, ensuring that in the new FF their contribution to protein
structure and dynamics is balanced.
2.2 Amber
Force Field
Amber FFs for proteins have been continually improved in recent

years and a detailed discussion of the various changes is beyond the
scope of this review. Significant revisions have been published, with
particular emphasis on important dihedral angles. Simmerling and
coworkers [45] introduced changes to the backbone potential in
the original Amber ff99 FF by fitting to additional quantum-level
data to produce the improved Amber ff99SB FF. Best and Hummer
continued along the same line, modifying the backbone potential
of the ff99SB and ff03 FFs to obtain a better balance between
sampling of helix and coil conformations. The new FFs were named
ff99SB* and ff03*, respectively [38]. Modifications of the sidechain torsion potential for four amino acid types in ff99SB was
introduced by Lindorff-Larsen et al. originating the ff99SB-ILDN
FF [46]. Further enhancements were produced by Li and
Bruschweiler based on experimental NMR data, originating the
ff99SB-ILDN-NMR FF [47]. To our knowledge, the latest update
in the Amber FFs was introduced recently by Neremberg and
50
Head-Gordon, who included a perturbation to the backbone

dihedral potential to shift the betaPPII equilibrium. This resulted
in improved sampling in water (TIP3P and TIP4P-Ew). Their
updates were designated ff99SB-ILDN-Phi [48]. In addition to
proteins, the Amber FFs support most common biomolecules. The
ff10 FF collection includes the most commonly used variants: the
ff99SB protein parameters [45], the BSC0 DNA parameters [49],
the Cheatham et al. ion parameters [50, 51], and updated RNA
parameters [52, 53]. Carbohydrates are supported through the
Glycam FFs [5456], and phospholipids are supported through
the CHARMM FF and the recent Lipid11 FF [57].
Polarizable Force Fields for Biomolecules: Current Status
3.1 Drude
Polarizable Force Field
Development of the Drude polarizable FF in CHARMM [58]

started in 2001 and the capability to simulate the Drude model is
now included in NAMD [59], ChemShell QM/MM [60], and the
OpenMM suite of utilities for GPUs [61]. Development of the
force field first involved implementation of the appropriate integrators to allow computationally efficient extended Lagrangian MD
simulations [62]. This was followed by optimization of the first
water model, in which a positive charge was assigned to the Drude
particle (SWM4-DP) [63]. The SWM4-DP model was reoptimized with a negative charged assigned to the Drude particles,
consistent with their representation of the electronic degrees of
freedom. The new model, called SWM4-NDP, is the standard
polarizable water model of the Drude polarizable FF [64]. It was
calibrated to reproduce important properties of the neat liquid at
room temperature and pressure such as enthalpy of vaporization,
density, static dielectric constant and self-diffusion constant, free
energy of hydration and shear viscosity. Concurrently with development of the water model, methodologies to determine electrostatic parameters for the Drude FF were advanced [65].
An early test of the feasibility of molecular dynamics simulations with the Drude polarizable FF was a successful simulation of
a DNA octamer in a box of water with sodium counterions [66].
Development of the Drude polarizable FF continued with parametrization of small molecules covering the functional groups commonly found in biomolecules. In 2005, the alkane FF was
developed, followed by parametrization of alcohols and aromatic
compounds in 2007 [67, 68]. Harder et al. published the first generation of N-methyl acetamide (NMA) parameters in 2008 [69].
Noteworthy is the proper treatment of the dielectric constant by
the polarizable FF in all systems, a property considered essential for
the accurate treatment of, for example, hydrophobic solvation in
biomolecules. The Drude polarizable FF was extended to the
nitrogen-containing heteroaromatic compounds in 2009 [5]. FF
51
parameters were refitted for ethers by Baker and MacKerell [70],

with significant improvements in the reproduction of liquid phase
dielectric constants, while maintaining the good agreement of the
previous model with all other experimental and quantum mechanical target data [71]. Sulfur containing model compounds were
parametrized in 2010 [72]. Other classes of molecules for which
Drude empirical FF parameters have been developed are nucleic
acid bases [73] and acyclic polyalcohols [74]. Early simulations of
dipalmitoylphosphatidylcholine (DPPC) bilayers and monolayers
were reported [75], followed by completion of a refined model for
DPPC [76].
Significant progress has been made in extending the Drude
polarizable FF from small compounds representative of the building blocks encountered in biological polymers to the polymers
themselves. The Drude empirical FF applicable to MD simulation
studies of peptides and proteins, termed Drude-2013, is covered in
Subheading 5, which includes a full account of the results. The
optimization of the polypeptide backbone parameters is discussed
in detail in Subheading 4.2 and the optimization of side-chain torsions is discussed in Subheading 4.3.
3.2 Amoeba
Force Field
AMOEBA is another classical polarizable FF that has achieved the

goal of producing a fully functional FF model for proteins [77].
Development of the AMOEBA polarizable FF has been ongoing since
1995 [78] and is based on modeling the electrostatic energy using
permanent and induced contributions. Permanent electrostatics originate in atomic multipolemultipole interactions with moments up to
the quadrupole located on each atom. The induced contribution is
modeled iteratively by generating an induced dipole originated by
permanent multipoles and other induced dipoles. Self-consistency is
obtained using an iterative scheme, and the Thole model [79] is used
to dampen electrostatic interactions at short range.
The AMOEBA FF was initially developed for water [80, 81].
Testing included reproduction of a variety of experimental data
and quantum calculations for small clusters, liquid water, and ice.
Several liquid phase properties including bulk thermodynamic,
transport, and structural measures were tested. These included
density, heat of vaporization, self-diffusion coefficient, heat capacity, dielectric constant, and radial distribution functions. Overall,
excellent agreement with reference values was obtained, and the
model was demonstrated to be applicable to structural properties
of two ice forms [81].
Treatment of ions in AMOEBA is described in ref. 82. Absolute
solvation free energies for potassium, sodium, and chloride ions in
liquid water and formamide have been computed. Simulation
results accurately reproduced vacuum QM results, experimental
ion-cluster solvation enthalpies, and experimental solvation free
energies for whole salts.
52
The AMOEBA FF has been extended to organic molecules,

including alkanes, alcohols, amines, sulfides, aldehydes, carboxylic
acids, amides, aromatics, and other small organic molecules [83].
As a validation, the hydrogen bonding energies and structures of
gas phase heterodimers with water were evaluated. Liquid self diffusion and static dielectric constants computed from MD simulations with AMOEBA are consistent with experimental values.
The FF was further tested by computing the solvation free energy
of 27 compounds not included in the parametrization process.
It performed well across different environments and phases, yielding an RMS error of 0.69 kcal/mol. Analysis of the dependence of
computed hydration free energies for seven small organic molecules with the QM level of theory used to derive atomic multipoles
was presented recently [84]. It was concluded that inclusion of
diffuse functions in the QM calculation of the atomic multipoles is
important. More comprehensive descriptions of the AMOEBA FF
have been presented previously and the reader is referred to those
publications for additional details [8587].
Parametrization of Polarizable Force Fields
4.1 Generic
Parametrization
Strategies
for the Drude
Polarizable
Force Field
The quality of FFs is heavily dependent on the quality of the underlying parameters. To obtain parameters of sufficient quality that are
capable of producing accurate simulation results, procedures have
been developed to target properties such as molecular geometries
and vibrations, pure solvent properties, and free energies of solvation, among others during the parametrization. In this section we
will describe parametrization of the polarizable Drude FF implemented in CHARMM. Reference to the well-established protocol
used to derive CHARMM additive FF parameters will be done
whenever a parallel is useful. The general outline of the parametrization process has been described for the CHARMM additive FF
in several publications (see refs. 1 and 19 for more details). Note
that parameter optimization remains an iterative process in the
polarizable FF and several rounds of parametrization are typically
performed until a satisfactory level of agreement with target data is
obtained.
A common strategy in parameter optimization of biological macromolecules is that parameters are developed for small, representative
model compounds and then transferred to the larger macromolecules. The advantages of this approach are: (1) smaller models are
easier to treat using both MM and QM methods and (2) more
experimental data are available for the smaller systems, including
thermodynamic properties of condensed phases, such as heats of
vaporization or sublimation and free energies of aqueous solvation.
It is crucial to include such data in the parameter optimization process to get an accurate description of the non-bond portion of the FF.
53
This strategy was also attempted in the development of Drude FF

parameters for the protein backbone, but ultimately a more involved
procedure was required as detailed in Subheading 4.2.
Parametrization of CHARMM FFs relies on obtaining appropriate intramolecular (bond, angle, dihedral, Urey-Bradley, and
improper terms), van der Waals (vdW), and electrostatic parameters that adequately reproduce selected target data. Determination
of the electrostatic parameters differs between the additive and the
Drude polarizable FFs. In the Drude FFs, in addition to optimization of point charges, which is also required in the additive FF,
polarizabilities and Thole factors must also be determined. In the
additive CHARMM FF, optimization of point charges is based on
a supramolecular approach where the charges are adjusted to
reproduce QM HF/6-31G* interaction energies and geometries
of the model compound with, typically, individual water molecules.
Placement of water molecules at different orientations around the
molecule enforces that local electronic polarization is accounted
for implicitly, an important feature for accurate reproduction of
condensed-phase properties. Additional data often include QM
results on dimers and dipole moments of the models. It is wellknown that in additive force fields, dipole moments must be overestimated to reproduce condensed phase properties [19, 88].
Other additive biomolecular FFs, most notably Amber, determine atomic partial charges based on reproduction of the QM
Electrostatic Potential (ESP), evaluated on grids surrounding
model compounds [8991]. These methods are convenient
because charges can be developed quickly for any compound for
which the QM ESP can be determined. An extension of ESP methods is inclusion of restraints during fitting, referred to as the
restrained ESP (RESP) approach [92]. This overcomes limitations
on the determination of charges on buried atoms [93]. It is important to note that partial charges from both supramolecular and
ESP approaches are conformation-dependent, requiring care in
the selection of appropriate conformations when performing the
charge optimization.
Electrostatic parameters of model compounds in the Drude
polarizable FF are obtained from restrained fitting to perturbed
QM ESP maps on grid points located on concentric Connolly surfaces surrounding the molecule. Often fitting is supplemented with
reproduction of the molecular dipole moment and diagonal elements of the polarization tensor [65, 94]. The determination of the
atomic polarizabilities and Thole factors [79] requires multiple perturbed ESPs typically calculated at the B3LYP/aug-cc-pVDZ level,
with each giving the electronic response of the molecule to a point
charge. Perturbing ions are placed mainly along chemical bonds
and lone pairs (LPs). This protocol was later extended to incorporate additional lone pair parameters and polarizability anisotropy,
and has become the standard in developing electrostatic parameters
54
for small molecules [95]. LPs typically carry the charge of the atom
(e.g., N, O, S in proteins) to which they are attached. The associated polarizability and Thole factor are both assigned to the parent
atom. Anisotropic polarizability of hydrogen bond acceptors was
found to be required to reproduce interactions with ions as a function of orientation. Initial values for the partial atomic charges are
taken from the C22 additive all-atom FF, and those for the polarizabilities are based on adjusted Millers atomic hybrid polarizability
(ahp) values [96].
Although gas-phase properties (e.g., dipole moments) are
easily reproduced with full atomic polarizabilities, scaling of the
polarizabilities has been shown to be necessary to reproduce condensed-phase properties [64]. A scaling factor of approximately
0.7 was found appropriate for the SWM4-DP and SWM4-NDP
water models while for other classes of molecules scaling factors
range from 0.6 to 1.0, with 1.0 being full polarizability. For
instance, scaling factors are 0.7 for primary and secondary alcohols
[67], 0.85 for aromatics [68], N-containing heterocycles [94],
nucleic acid bases [73] and ethers [97], and 1.0 for alkanes [42].
Other scaling factors are 0.7 for thiols, 0.85 for dimethyl disulfide
and 0.6 for ethylmethyl sulfide [72]. A value of 0.724 was recently
used in ion parameters [98]. Final optimization of the electrostatic parameters consists of testing the model for reproduction of
the pure solvent dielectric constants and adjusting the polarizability scaling if necessary.
Development of parameters to model vdW forces in the Drude
FF, which are treated using the LennardJones (LJ) 612 term,
follows closely the protocol established for the additive FF and will
only be briefly outlined here. Jorgensen and coworkers [99, 100]
pioneered the use of condensed-phase simulations, usually pure
liquids, as the basis for optimization of LennardJones (LJ) parameters that account for both vdW attraction and interatomic repulsion. Typically, once electrostatic parameters are determined, the
LJ parameters for a model compound can be adjusted to reproduce
experimental pure solvent properties such as heat of vaporization,
density, isothermal compressibility, heat capacity, heat of sublimation, lattice geometry, and free energy of aqueous solvation, as
available. Although this is an effective method for the fine-tuning
of the parameters, there are important issues. One is parameter
correlation, such that LJ parameters for different atoms in a molecule and/or the magnitudes of ij and Rmin on the same atom, can
compensate for individual unbalanced values, making it difficult to
gauge whether they are balanced relative to one another [101].
To overcome this problem, a method has been developed to determine the relative value of the LJ parameters based on high level
QM data [102] with the absolute values being based on scans of ij
and Rmin that reproduce experimental data [103, 104]. This
approach requires supramolecular interactions between rare gases
55
and the model compound. Importantly, once satisfactory LJ

parameters are obtained for atoms in a class of functional groups,
they can often be directly transferred to other molecules carrying
those functional groups without additional optimization.
Reproduction of experimental hydration free energies reflects
how well the electrostatic and vdW parameters model interactions
with bulk water. Recently, in the context of the polarizable Drude
FF, it was shown that atom-pair-specific LJ parameters (termed
NBFix in the context of CHARMM) needed to be used in order
to minimize discrepancies between calculated and experimental
hydration free energies while simultaneously reproducing pure solvent heats of vaporization and molecular volumes [105].
Optimization of internal parameters is usually done relative to
target data that include geometries, vibrational spectra and conformational energies. Geometries are typically optimized at the
MP2/6-31G* level (or MP2/6-31 + G* in the case of anions), and
vibrational spectra are obtained at the MP2/6-31G* level.
Frequencies are scaled using correction factors prescribed by
Radom and coworkers [106], and a symbolic potential energy distribution (PED) analysis is performed as proposed by Pulay et al.
[107] using the MOLVIB module in CHARMM. This approach
has been shown to yield good agreement with experimental geometries for model compounds of complex systems such as proteins,
nucleic acid bases, and sugars [22, 108, 109], while being computationally feasible.
4.2 Parametrization
of the Polypeptide
Backbone in the Drude
Force Field
Parameterization of the polypeptide backbone was initially assumed

to follow the general rules in use for CHARMM FFs, namely that
parameters would be transferable from smaller model compounds.
The prototype of the protein backbone, for all residues except glycine and proline, was based on alanine polypeptides. The initial
electrostatic model, identified as Drude-NMA, was derived from a
combination of electrostatic parameters that included N-methyl
acetamide (NMA) and ethane, and LJ parameters were also transferred from NMA and ethane. Several rounds of optimization were
previously done on NMA: initial parameters were published by
Harder et al. [69] and a final set by Lin et al. [110] In the latest
model LJ parameters were selected to give acceptable intramolecular hydrogen bond distances in -helix conformations of alanine
polypeptides in addition to allowing reproduction of NMA experimental condensed phase properties [110]. CMAP corrections for
alanine dipeptide were also used to allow the (, ) Ramachandran
map to reproduce a high-level QM (RIMP2/CBS//RIMP2/ccpVDZ) surface, where the CBS (complete basis set) extrapolation
was obtained from RIMP2/cc-pVTZ and RIMP2/cc-pVQZ single point energies following the prescription of Halkier et al. [111].
The Drude-NMA model was tested by calculating gas phase
molecular properties of alanine dipeptide and (Ala)5 in different
56
Table 1
Gas phase dipole moments of alanine dipeptide and (Ala)5a, molecular polarizability of alanine
dipeptide, and relative energies of (Ala)5
Molecular dipole moment of alanine dipeptide (Debye)
R
QMb
C5
DrudeNMA
DrudeALA
Drude-2013
QM
DrudeNMA
DrudeALA
Drude-2013
6.2
5.0
6.4
6.7
4.7
5.8
2.3
2.6
1.3
0.1
3.1
3.0
4.4
5.6
1.8
2.3
1.6
0.9
1.7
1.5
1.0
0.9
0.6
0.3
5.9
4.9
5.4
5.8
1.2
0.9
1.3
1.3
Molecular dipole moment of (Ala)5 (Debye)

R
C5
QMc
DrudeNMA
DrudeALA
Drude-2013
QM
DrudeNMA
22.0
13.5
22.4
20.8
11.6
24.4
DrudeALA
4.5
Drude-2013
9.3
Molecular polarizability of alanine dipeptide ( )

R
C5
QMb
DrudeNMA
DrudeALA
Drude-2013
QM
DrudeNMA
DrudeALA
Drude-2013
x x
13.57
13.40
16.18
15.30
15.49
16.02
19.89
16.07
y y
12.72
12.60
14.29
14.36
12.06
11.87
13.39
12.78
zz
11.71
11.03
12.68
9.94
10.35
9.78
11.05
10.39
Relative energies of (Ala)5 (kcal/mol)

QMd
R-C5
R-PPII
Drude-NMA
Drude-ALA
Drude-2013
6.59
6.21
5.31
3.89
14.83
5.77
0.42
10.17
(Ala)5 is acetyl-(Ala)5-N-methylamide
QM dipole moments and polarizabilities of alanine dipeptide obtained at the B3LYP/aug-cc-pVDZ level with the
polarizabilities scaled by 0.85
c
QM dipole moments for (Ala)5 obtained at the B3LYP/6-31G* level
d
Single point energies were calculated at the RIMP2/cc-pVTZ//RIMP2/cc-pVDZ level
b
conformations, such as dipole moments, relative energies, and

molecular polarizabilities, and through MD simulations of (Ala)5
in solution [112]. Testing the behavior of (Ala)5 in solution has
become common practice in the validation of protein force fields,
being used in the development of the C36 additive FF [20] and
the AMOEBA polarizable FF [77] (see Subheading 5.1) for details.
As alluded to above, direct transfer of Drude-NMA parameters
to polypeptides did not yield acceptably accurate results (Table 1).
57
Fig. 1 Illustration of induced dipoles on dipeptide moieties of alanine dipeptide and (Ala)5. Values in parenthesis
are for alanine dipeptide
Using transferred parameters, the agreement of the computed

dipole moments, polarizabilities, and relative energies with target
values was poor, in particular for the extended conformations.
Tests of NMR J-coupling also indicated poor agreement with
experiment due to a (, ) distribution that predominantly
populated extended C5 conformations.
Included in Fig. 1 are representative orientations and magnitudes of the induced dipoles and separations (pm) of the Drude
particle and main atom in a dipeptide section of the alanine dipeptide (values in parenthesis) and (Ala)5. All calculations were done
enforcing the C5 conformation for both systems. The separation
between the Drude particle and the main atom is a direct measurement of the magnitude of the induced dipole. Using Drude-NMA
electrostatic parameters (Fig. 1a), displacement of the Drude particles in (Ala)5 relative to the parent atom are similar on all carbonyl
C (labeled Ci1) atoms (~14 pm), and are substantially larger than
the displacement for Ci1 in alanine dipeptide. Ci1 in alanine dipeptide is bound to a methyl group and the polarizing field is much
58
weaker than in the longer polypeptide where Ci1 feels the electric
field originating from the same amino acids NH group. The case
is similar for the N atoms, with Ni+1 showing a much stronger
induced local dipole in (Ala)5 as compared to the alanine dipeptide.
The induced dipole on C is smaller on (Ala)5, enhancing the
dipole interaction between Ni and Ci. This results in two effects.
First, local dipoles associated with the peptide bonds interact with
each other enhancing the local dipole moments associated with
each peptide bond and, second, the larger dipole strengthens electrostatic interactions with water leading to overstabilization of the
C5 conformation. Indeed, a comparison of the dipole moments of
acetyl-(Ala)5-N-methylamide for the NMA based model with QM
data indicated the overall dipole moment of the C5 conformation
to be significantly overestimated (Table 1). It was, therefore,
hypothesized that the overestimation, which would lead to even
more favorable interactions with aqueous solvent, was due to the
electrostatic parameter optimization procedure based on NMA
alone not defining balanced electrostatic interactions between the
individual peptide bonds. Based on this analysis it was concluded
that use of larger model compounds allowing communication
between adjacent peptide bonds was required in the determination
of electrostatic parameters, with the initial candidate being the alanine dipeptide.
Electrostatic parameters based on the alanine dipeptide were
determined by averaging the components over five independent
sets of parameters obtained from electrostatic potential (ESP) fitting
corresponding to the R, L, C5, PPII and C7eq conformations.
This model is referred to as Drude-ALA in the text below. For
each conformation the electrostatic parameter optimization, which
included the partial atomic charges, atomic polarizabilities, and
atom-based Thole factors, was performed using the FITCHARGE
module of CHARMM by fitting to the QM ESP maps as described
above. The outcome is electrostatic parameters that better reproduce the change in the ESP associated with electrostatic interactions
between the peptides bonds in the different relative orientations.
The resulting Drude-ALA model yielded a smaller dipole moment
for the C5 conformation for acetyl-(Ala)5-N-methylamide (Table 1).
Simulations of (Ala)5 in aqueous solution were also performed
and compared to Drude-NMA, and while the Drude-ALA model
showed improved agreement with experiment, the agreement was
still poor as compared to the additive C36 FF. It was found that the
PPII region started to be populated, though the C5 conformation
still dominated, indicating that the inclusion of electrostatic interactions between the peptide bonds during parameter optimization did
improve the quality of the FF. However, those improvements were
clearly insufficient, indicating that different target data were needed
to obtain a more accurate electrostatic model for the polypeptide
backbone.
59
The inability of Drude-ALA electrostatics to provide a

reasonable description of properties of alanine polypeptides in gasphase and solution prompted development of a third parametrization strategy. The rationale of the new methodology has its roots
in the fundamental physics of the Drude model. In the presence of
an electric field the position of the Drude particles are optimized
while the main atom remains fixed following the Born
Oppenheimer principle. This creates on each atom a local dipole
that, although small, is able to interact with neighboring dipoles.
The magnitude of each dipole can be controlled by two factors: (1)
the atomic polarizability, and (2) the damping of 12 and 13
interactions through the individual Thole factors. Thus, for polymeric structures, control of the behavior of each dipole is extremely
complex. Boundaries cannot be explicitly imposed locally for each
atom or groups of atoms as in other polarizable methods since the
overall properties are the result of many cross interactions spanning wide regions of the system. As a consequence, an effective
methodology of parametrization needed to include enough information on the whole molecule, thus allowing for a balanced set of
electrostatic parameters. Furthermore, it was necessary to include
information not only in gas phase but also from interactions with
water molecules, since water is the preferred medium where most
of the MD simulations with Drude oscillators are anticipated to
take place. The third optimization method of backbone electrostatic parameters used a Simulated Annealing (SA) protocol [113],
yielding the final model, Drude-2013. The target data consisted of
an array of QM observables determined for the alanine dipeptide
and larger alanine polypeptides. Target data included the polarizability of the alanine dipeptide, relative energies of (Ala)5, dipole
moments of alanine dipeptide and (Ala)5, and energetic and structural data for the interaction of the alanine dipeptide with individual water molecules along specific directions. Several conformations
of the alanine models were used: R, C5, and PPII for the relative
energies of (Ala)5; C5 and PPII for the interactions of the alanine
dipeptide with water; and R, C5, PPII, and C7eq conformations
of the alanine dipeptide for molecular polarizabilities and dipole
moments. In addition to the electrostatic parameters, during the
SA internal parameters were allowed to vary within a limited range
to keep the alanine dipeptide optimized geometries close to the
targeted values. SA started with a temperature of 150 K with individual parameters randomly adjusted followed by accepting or
rejecting the new parameter set based on the Metropolis criterion,
resulting in Monte Carlo Simulated Annealing (MCSA) [114]. The
temperature was gradually reduced to near 0 K yielding a selected
parameter set for testing in (Ala)5 solution simulations. The error
function was the weighted sum of all differences between MM and
QM data for all properties mentioned above with various weighting
factors. During MCSA fitting, a new CMAP that reproduces the
60
QM alanine dipeptide (, ) RIMP2/CBS//MP2/6-311G(d,p)

surface was generated at each iteration. In addition, empirical
adjustments of the CMAP were added to the QM-based surface to
improve agreement with conformational sampling of the peptide
backbone in peptides and proteins, resulting in the final Drude2013 model.
While both C36 and Drude-2013 (, ) surfaces have
undergone some empirical adjustments, the underlying energy
surfaces was based on quantum mechanics, and therefore the
overall landscape of the surfaces is similar. Adjustments in the
C36 CMAP, which was obtained at the LMP2/cc-pVQZ level,
included local optimization of the helical and sheet regions to
reproduce subtle features observed in crystallographic survey
data [32] followed by subsequent shifting of the helical region
to decrease the tendency for the C22/CMAP model to overpopulate that conformation [20]. For the Drude-2013 model,
the overall sheet region was lowered and the areas between the
sheet and helical regions and from = 90 to 180 and = 60
to 105 were raised.
4.3 Side Chain 1, 2
Dihedral Parameter
Optimization
in the Drude Force
Field
Different side chains impact the conformational distribution of the

polypeptide backbone, as observed in experimental studies
[115117]. The peptide (Ala)4-X-(Ala)4 has been used before as a
model system for 1, 2 parameter optimization [46], where X is
the amino acid of interest and the backbone conformation is constrained to fully extended, C7eq, PPII, or R conformations [41].
Those studies indicated that (Ala)4-X-(Ala)4 with either the C7eq or
PPII backbone conformation yields aqueous phase conformational
properties that mimic those occurring in full proteins. Based on this
analysis, 1, 2 parameter optimization was performed by initially
targeting QM data for the respective side chain dipeptides, with the
backbone in the , R, and L conformations. These parameters
were then used in Hamiltonian Replica Exchange MD (H-REMD)
simulations [118] of (Ala)4-X-(Ala)4 in solution, with 1, 2 sampling compared with PDB survey data. Overlap coefficients (OC)
[41] for 1 and 2 distributions from (Ala)4-X-(Ala)4 in the C7eq
conformation and those from a survey of the PDB [119] were computed, with an OC of 1 indicating exact agreement and an OC of 0
indicating no agreement. The extent of overlap for some of the
amino acids based on optimization only targeting the QM data was
found to be quite good. For example, values of 0.87, 0.88, and 0.87
were obtained for 1 for Cys, Leu, and Val, respectively, while the
OC was 0.92 for 2 with Leu. Based on the quality of the fit for
these residues, additional optimization was not performed.
Additional optimization for the remaining residues involved comparison of the computed and target 1 and 2 populations of the
gauche+, gauche-, and trans rotamers and manually adjusting the
corresponding dihedral parameters to improve the level of agreement. After the optimization, significant agreement with the PDB
61
target data was obtained for a number of amino acids, notable examples being Ile, Lys, and Thr. Overall, the final OC values are typically
0.7 or higher, though lower values were also found including Asn 2,
Asp 1, Gln 2, and Glu 1. The final parameters were used for the
reported polypeptide and protein simulations. In ref. 120 we present
detailed descriptions of the optimization protocol and final results.
4.4 The AMOEBA
Force Field
and Parametrization
of Proteins
Detailed methodology for deriving electrostatic parameters for

AMOEBA to allow incorporation of novel molecules has been
published [83], and therefore what follows is a brief overview.
Determination of permanent atomic multipoles for glycine, alanine, and proline residues was done based on capped acetyl-XN-methylamide dipeptides with X = Gly, Ala, and Pro. The first
step is definition of intramolecular direct polarization groups,
which is important because atoms belonging to one group can
only polarize atoms outside that group. The group definitions
for alanine dipeptide are show in Fig. 2 of ref. 77. For side
chains, groups are also selected. The optimization proceeds with
assignment of the initial multipole parameters from Distributed
Multipole Analysis (DMA) at the MP2/6-311G** level. Initial
parameters are then iteratively optimized against the MP2/augcc-pVTZ electrostatic potential computed on a set of grid points
around the dipeptide compounds. Converged Permanent Atomic
Multipoles (PAMs) were determined simultaneously for five
local minima: L, , C5, C7a, and C7e conformers.
Application of Polarizable Force Fields to Protein Simulations

The year 2013 marked important milestones in the development of
polarizable FFs. After years of development, polarizable FFs for
peptides and proteins suitable for MD simulations based on classical
Drude oscillators (Drude-2013) and the AMOEBA model
(AMOEBA-2013) were published. Here, we summarize results of
MD simulations with the two FFs.
5.1 Peptide
Simulations with C36
Additive,
AMOEBA-2013,
and Drude-2013 Force
Fields
With advances in computing capacity, it has become common to

use simulations of oligopeptides in solution to calibrate FF torsional parameters [20, 37, 45, 112, 121123], since results can be
directly compared to experimental nuclear magnetic resonance
(NMR) data for corresponding peptides. Conformational distributions in an NMR experiment are reflected in NMR-derived spin
spin coupling (J-coupling) constants. Using Karplus relations,
J-coupling values can be computed from peptide conformations
from MD simulations, and the ability to achieve ever-increasing
timescales via MD allows for the computational generation of conformational ensembles of a size that can be meaningfully compared
with experiment [37, 124].
62
As an example, simulations of small polypeptides of (Ala)3,

(Ala)5, (Ala)7, (Val)3, and (Gly)3 were used by Best et al. to validate
the improved CHARMM36 additive FF (C36) [20]. Using this
approach, C36 introduced small but significant changes relative to
its predecessor, C22/CMAP. In alanine- and valine-based peptides, minima occur at PPII, with C5 and R being only slightly
higher in energy. The additional minima at L and C7ax are
approximately 23 kcal/mol higher than the PPII conformation.
And while there is only a small difference between (Ala)3, (Ala)5,
and (Ala)7, sampling for (Val)3 was significantly different because
of the presence of the bulky hydrophobic side chain. Compared
with other FFs, AMBER ff99SB9 and ff99SB* are closest to C36,
while OPLS/AA [125] is qualitatively different with a minimum at
C7eq and Gromos 53a6 FF [126] has two minima near R and a
low-energy transition region between R and C7ax.
C36 (, ) sampling has also been compared with experimental NMR J-coupling. Agreement was very good for the alaninebased peptides and for (Gly)3, and reasonable for (Val)3. The new
C36 FF significantly improves over the previous C22/CMAP FF,
with improvement coming from decreased sampling of R conformations and increased sampling of PPII, which is reflected in the
J-couplings. In ref. 20 C36 was also compared with other FFs outside the CHARMM family (AMBER ff99SB [45], OPLS/AA
[125], Gromos 53a6 [126]), showing significantly lower 2 values.
No direct comparison of C36 with the latest improved Amber FFs
has been published, although it is anticipated that C36 will compare very favorably to experimental data based on published results.
Sampling for the unblocked, protonated (Ala)5 peptide has
been tested using the AMOEBA-2013 FF [77]. Sampling is similar
to C36, with a distinct global minimum located around the PII
conformation and two other basins approximately 0.5 kcal/mol
higher in free energy in the -sheet and -helix regions. Barriers
separating the global basin from the two local minima are 12 kcal/
mol. (, ) sampling was compared with experimental J-coupling
constants, and values from MD simulations are in excellent agreement with those probed by experiment with a 2 value of 1.0
Simulations of (Ala)5 polypeptides were not used to validate
the newly developed Drude polarizable FF in CHARMM but
rather were explicitly part of the optimization process, particularly
for the fine-tuning of the CMAP potential so as to yield acceptable
sampling patterns in the tested protein systems. In addition, the
GB1 hairpin [4156], [127, 128] and a dimeric coiled coil (1UOI)
[129] were also used as target data for optimization of the
Drude-2013 model. Due to the use of multiple small peptides, as
well as larger proteins, as target data, sampling of (Ala)5 had to be
slightly compromised, yielding 2 values larger than 1.0.
Explicit solvent simulations of the GB1 hairpin of 100 ns
yielded RMS differences with the Drude model more similar to
63
the crystal structure of the full GB1 protein as compared to the

C36 where the RMS difference fluctuated between 2.5 and 3 ,
indicating drift away from the crystal structure. With the dimeric
coiled coil (1UOI) [129] RMS analysis showed the overall structure of the coiled coil to deviate more from the crystal structure
with the Drude model as compared to C36 (see ref. 120 for
details). The individual helices in the dimer move relative to each
other, while the conformations of the individual helices are well
preserved, suggesting the ability of the Drude-2013 model to
properly treat the helical secondary structure of the individual
monomers. (, ) probability distributions from the simulations
supported this conclusion. Thus, the Drude model satisfactorily
reproduces the conformational properties of small peptides on the
100 ns time scale, though longer simulations will be required to
more rigorously challenge the model.
5.2
Full Proteins
Further validation of the Drude-2013 force field involved explicit

solvent MD simulations on ten proteins: 1EJG (crambin), 1P7E
(protein GB1 domain), 1MJC (cold-shock protein A), 1UBQ
(ubiquitin), 3ZZP (circular permutant of ribosomal protein), 4IEJ
(DNA methyl transferase associated protein), 135L (lysozyme),
1IFC (fatty acid binding protein), 3VQF (PDZ domain from tight
junction regulatory protein), and 1BYI (dethiobiotin synthase).
The proteins are relatively small, typically less than 100 residues,
and cover a range of secondary structures.
The stability of each protein was characterized by the value
of its backbone RMS deviation (RMSD) relative to the crystal
structure. The results summarized in Table 7 and Fig. S2 of the
Supplementary Information of ref. 120 showed the RMS differences are typically larger with the Drude model versus C36
additive force field as are the RMS fluctuations. The Drude2013 model shows additional flexibility compared to the additive model with only one exception, namely ubiquitin (1UBQ).
While the Drude model generally appears to have more flexibility than the additive C36 model, NMR analysis indicated that
for selected residues with high mobility in C36, the Drude
model gave improved agreement with experiment, as shown in
Fig. 7 of Lopes et al. [120].
Results with AMOEBA-2013 protein FF [77] have been
reported for three of the proteins studied with the Drude-2013
force field. These include crambin (1EJG), ubiquitin (1UBQ), and
lysozyme (135L). AMOEBA MD simulations were performed for
30 ns yielding backbone RMSDs in the vicinity of 1 , 2 , and
2 for the three proteins, respectively. At 30 ns of the Drude
simulations the corresponding values were 1.1/1.1, 1.9, and
1.9/2.2 , where two values are from duplicate simulations.
64
Fig. 2 100-ns snapshots from Drude-2013 simulations (red ) of lysozyme (135L) and dethiobiotin synthase
1BYI superimposed on the starting crystallographic structures (blue)
Although the Drude model showed additional flexibility over

the additive C36 model, the overall structures of the proteins are well
maintained. Snapshots taken at 100 ns for lysozyme (135L) and
dethiobiotin synthase (1BYI) superimposed on the corresponding
crystal structures are shown in Fig. 2, showing the overall maintenance of the structures. Consistent with this was the (, ) sampling with the Drude model over all the simulated proteins being
similar to a survey of the PDB as well as sampling occurring with
C36 (Fig. 6 of Lopes et al.). In addition, the NHO=C distance
distributions in secondary structures were reasonably reproduced
by the Drude model, though there is a tendency towards the distances being slightly longer than distributions from PDB crystal
structures (Fig. 5 of Lopes et al.).
Additional analysis involved dipole moments of selected moieties during the MD simulations. These included the peptide bonds
in the GB1 hairpin and ubiquitin and tryptophan residues in lysozyme (Fig. 8 of Lopes et al.). In all cases the Drude dipole moments
are systematically larger than with the additive model. This indicates
that, while partial atomic charges in the additive model are adjusted
to overestimate molecular dipole moments, the extent of overestimation is not enough for the protein environment. In addition, the
dipole moments of the peptide bonds with the Drude model in
sheets are systematically larger than in helices. Finally, significant
variations in the dipole moments were observed in the Drude simulations (e.g., >1.5 D for a Trp in lysozyme). Thus, even though the
additive models were optimized to yield enhanced dipole moments
appropriate for the condensed phase, it does not appear that the
overestimation was sufficient based on these initial polarizable
calculations. That, as well as the large variation in the dipoles occurring in the Drude model, suggests that the underlying physical
forces dictating the overall properties of the peptides and proteins
are significantly different in the Drude versus the additive model.
65
Indeed, the additional flexibility in the Drude model may be due to

the inclusion of electronic polarization in the model, allowing for
the variability of the local molecular dipoles.
Summary
The field of empirical FF based simulations of proteins continues
to develop. Since the last publication of a similar review great progress has been made, including the publication of two polarizable
force fields for proteins as well as improvements in the AMBER
and CHARMM additive protein force fields. Work on other classes
of biopolymers has also made significant progress allowing for simulations of heterogeneous systems. As other researchers start using
the recently published force fields, in particular the polarizable
force fields, limitations will certainly be found and corrections and
improvements are expected.
As was emphasized in this review, development of electrostatic
parameters in the Drude force field is very complex. It is expected
that new optimization algorithms together with more sophisticated target data will lead to significant progress. Polarizable models for other classes of biomolecules based on the Drude oscillator
will be published soon for DNA and carbohydrates as well as a
wider range of lipids.
While polarizable MD simulations will make a significant
contribution to our understanding of protein structure and
function it should be emphasized that these models are more
sensitive to initial conditions than with an additive FF, and can
have polarization catastrophes that will cause simulations to fail.
To overcome this it is suggested that systems initially be set up
and equilibrated with an additive FF and then converted to the
polarizable model. To facilitate this procedure the CHARMMGUI [130] has been extended to include a new utility, the
Drude Prepper. The Drude Prepper reads equilibrated
CHARMM PSF and coordinate files and converts them to Drude
format files. This includes the production of inputs for MD
simulations using CHARMM or NAMD. This utility will greatly
facilitate the application of the Drude model to a range of
proteins as well as other systems.
Concerning computational efficiency, the Drude model typically requires the use of a 1 fs integration time step during MD
simulations. In addition, there is an approximately twofold overhead associated with the calculation of the polarization contribution to the electrostatics. Thus, the model is approximately fourfold
slower than corresponding additive simulations performed with a
2 fs integration time step. However, the NAMD implementation is
highly parallelizable [59], which will facilitate simulations of large
systems using the Drude model.
66
Acknowledgement
Financial support from the NIH (GM072558) and computational
support from the University of Maryland Computer-Aided Drug
Design Center, and the Extreme Science and Engineering Discovery
Environment (XSEDE), which is supported by National Science
Foundation grant number OCI-1053575, are acknowledged.
References
1. MacKerell AD (2004) Empirical force fields
for biological macromolecules: overview and
issues. J Comput Chem 25(13):15841604
2. Stone AJ (2008) Intermolecular potentials.
Science 321(5890):787789
3. Freddolino PL, Harrison CB, Liu YX, Schulten
K (2010) Challenges in protein-folding simulations. Nat Phys 6(10):751758
4. Warshel A, Kato M, Pisliakov AV (2007)
Polarizable force fields: history, test cases, and
prospects. J Chem Theory Comput
3(6):20342045
5. Lopes PEM, Roux B, MacKerell AD (2009)
Molecular modeling and dynamics studies
with explicit inclusion of electronic polarizability: theory and applications. Theor Chem
Acc 124(12):1128
6. Zhu X, Lopes PEM, MacKerell AD (2012)
Recent developments and applications of the
CHARMM force fields. Wiley Interdiscip Rev
Comput Mol Sci 2(1):167185
7. Guvench O, MacKerell AD (2008)
Comparison of protein force fields for molecular dynamics simulations. In: Kukol A (ed)
Molecular modeling of proteins. Humana
Press, Totowa, NJ, pp 6388
8. Lopes PEM, Harder E, Roux B, MacKerell AD
(2009) Formalisms for the explicit inclusion of
electronic polarizability in molecular modeling
and dynamics studies. In: York DM, Lee T-S
(eds) Multi-scale quantum models for biocatalysis. Springer, Netherlands, pp 219257
9. Salomon-Ferrer R, Case DA, Walker RC
(2013) An overview of the Amber biomolecular simulation package. Wiley Interdiscip
Rev Comput Mol Sci 3(2):198210
10. Beauchamp K, Lin Y-S, Das R, Pande V
(2012) Are protein force fields getting better?
A systematic benchmark on 524 diverse NMR
measurements. J Chem Theory Comput 8(4):
14091414
11. Burkert U, Allinger N (1982) Molecular
mechanics. American Chemical Society,
Washington, DC
12. McCammon JA, Harvey SC (1987) Dynamics

of proteins and nucleic acids. Cambridge
University Press, New York
13. Leach AR (2001) Molecular modelling: principles and applications. Prentice Hall, Harlow,
England
14. Becker OM (2001) Computational biochemistry and biophysics. M. Dekker, New York
15. Rapaport DC (2004) The art of molecular
dynamics simulation. Cambridge University
Press, Cambridge, UK
16. Schlick T (2002) Molecular modeling and
simulation: an interdisciplinary guide.
Springer, New York
17. Satoh A. Introduction to practice of molecular
simulation molecular dynamics, Monte Carlo,
Brownian dynamics, Lattice Boltzmann, dissipative particle dynamics. http://site.ebrary.
com/id/10440534
18. Brooks BR, Bruccoleri RE, Olafson BD,
States DJ, Swaminathan S, Karplus M (1983)
CHARMM: a program for macromolecular
energy, minimization, and dynamics calculations. J Comput Chem 4(2):187217
19. MacKerell AD, Bashford D, Bellott M,
Dunbrack RL, Evanseck JD, Field MJ et al
(1998) All-atom empirical potential for molecular modeling and dynamics studies of proteins. J Phys Chem B 102(18):35863616
20. Best RB, Zhu X, Shim J, Lopes PEM, Mittal J,
Feig M et al (2012) Optimization of the additive CHARMM all-atom protein force field
targeting improved sampling of the backbone
, and side-chain 1 and 2 dihedral angles. J
Chem Theory Comput 8(9):32573273
21. MacKerell AD, Wiorkiewicz-Kuczera J,
Karplus M (1995) An all-atom empirical
energy function for the simulation of nucleic
acids. J Am Chem Soc 117(48):1194611975
22. Foloppe N, MacKerell AD (2000) All-atom
empirical force field for nucleic acids:
I. Parameter optimization based on small molecule and condensed phase macromolecular
target data. J Comput Chem 21(2):86104

23. MacKerell AD, Banavali NK (2000) All-atom
empirical force field for nucleic acids:
II. Application to molecular dynamics simulations of DNA and RNA in solution. J Comput
Chem 21(2):105120
24. Feller SE, MacKerell AD (2000) An improved
empirical potential energy function for molecular simulations of phospholipids. J Phys
Chem B 104(31):75107515
25. Feller SE, Gawrisch K, MacKerell AD (2001)
Polyunsaturated fatty acids in lipid bilayers:
intrinsic and environmental contributions to
their unique physical properties. J Am Chem
Soc 124(2):318326
26. Klauda JB, Venable RM, Freites JA,
OConnor JW, Tobias DJ, MondragonRamirez C et al (2010) Update of the
CHARMM all-atom additive force field for
lipids: validation on six lipid types. J Phys
Chem B 114(23):78307843
27. Kuttel M, Brady JW, Naidoo KJ (2002)
Carbohydrate solution simulations: producing a force field with experimentally consistent primary alcohol rotational frequencies
and populations. J Comput Chem 23(13):
12361243
28. Guvench O, Greene SN, Kamath G, Brady
JW, Venable RM, Pastor RW et al (2008)
Additive empirical force field for hexopyranose
monosaccharides. J Comput Chem 29(15):
25432564
29. Hatcher ER, Guvench O, MacKerell AD
(2009) CHARMM additive all-atom force
field for acyclic polyalcohols, acyclic carbohydrates, and inositol. J Chem Theory Comput
5(5):13151327
30. Guvench O, Hatcher E, Venable RM, Pastor
RW, MacKerell AD (2009) CHARMM additive all-atom force field for glycosidic linkages
between hexopyranoses. J Chem Theory
Comput 5(9):23532370
31. Vanommeslaeghe K, Hatcher E, Acharya C,
Kundu S, Zhong S, Shim J et al (2010)
CHARMM general force field: a force field
for drug-like molecules compatible with the
CHARMM all-atom additive biological force
fields. J Comput Chem 31(4):671690
32. MacKerell AD, Feig M, Brooks CL (2004)
Extending the treatment of backbone energetics in protein force fields: limitations of
gas-phase quantum mechanics in reproducing
protein conformational distributions in
molecular dynamics simulations. J Comput
Chem 25(11):14001415
33. MacKerell AD, Feig M, Brooks CL (2004)
Improved treatment of the protein backbone
in empirical force fields. J Am Chem Soc
126(3):698699
67
34. Freddolino PL, Schulten K (2009) Common

structural transitions in explicit-solvent simulations of villin headpiece folding. Biophys J
97(8):23382347
35. Freddolino PL, Liu F, Gruebele M, Schulten
K (2008) Ten-microsecond molecular dynamics simulation of a fast-folding WW domain.
Biophys J 94(10):L75L77
36. Freddolino PL, Park S, Roux B, Schulten K
(2009) Force field bias in protein folding simulations. Biophys J 96(9):37723780
37. Best R, Buchete N-V, Hummer G (2008) Are
current molecular dynamics force fields too
helical? Biophys J 95(1):L07L09
38. Best RB, Hummer G (2009) Optimized
molecular dynamics force fields applied to the
helix-coil transition of polypeptides. J Phys
Chem B 113(26):90049015
39. Best RB, Mittal J (2010) Balance between
and structures in ab initio protein folding.
J Phys Chem B 114(26):87908798
40. Mittal J, Best RB (2010) Tackling force-field
bias in protein folding simulations: folding of
villin HP35 and Pin WW domains in explicit
water. Biophys J 99(3):L26L28
41. Shim J, Zhu X, Best RB, MacKerell AD
(2013) Ala4-X-Ala4 as a model system for the
optimization of the 1 and 2 amino acid sidechain dihedral empirical force field parameters. J Comput Chem 34(7):593603
42. Vorobyov IV, Anisimov VM, MacKerell AD
(2005) Polarizable empirical force field
for alkanes based on the classical drude
oscillator model. J Phys Chem B 109(40):
1898818999
43. Mason PE, Neilson GW, Enderby JE,
Saboungi ML, Dempsey CE, MacKerell AD
et al (2004) The structure of aqueous guanidinium chloride solutions. J Am Chem Soc
126(37):1146211470
44. Macias AT, MacKerell AD (2005) CH/pi
interactions involving aromatic amino acids:
refinement of the CHARMM tryptophan force
field. J Comput Chem 26(14):14521463
45. Hornak V, Abel R, Okur A, Strockbine B,
Roitberg A, Simmerling C (2006) Comparison
of multiple Amber force fields and development of improved protein backbone parameters. Proteins 65(3):712725
46. Lindorff-Larsen K, Piana S, Palmo K,
Maragakis P, Klepeis J, Dror R et al (2010)
Improved side-chain torsion potentials for the
Amber ff99SB protein force field. Proteins
78(8):19501958
47. Li D-W, Bruschweiler R (2011) NMR-based
protein potentials. Angew Chem 122(38):
69306932
68
48. Nerenberg P, Head-Gordon T (2011)

Optimizing proteinsolvent force fields to
reproduce intrinsic conformational preferences of model peptides. J Chem Theory
Comput 7(4):12201230
49. Perez A, Marchan I, Svozil D, Sponer J,
Cheatham TE, Laughton CA et al (2007)
Refinenement of the AMBER force field for
nucleic acids: improving the description of
alpha/gamma conformers. Biophys J 92(11):
38173829
50. Joung
IS,
Cheatham
TE
(2008)
Determination of alkali and halide monovalent ion parameters for use in explicitly solvated biomolecular simulations. J Phys Chem
B 112(30):90209041
51. Joung IS, Cheatham TE (2009) Molecular
dynamics simulations of the dynamic and
energetic properties of alkali and halide ions
using water-model-specific ion parameters.
J Phys Chem B 113(40):1327913290
52. Banas P, Hollas D, Zgarbova M, Jurecka P,
Orozco M, Cheatham TE III et al (2010)
Performance of molecular mechanics force
fields for RNA simulations: stability of UUCG
and GNRA hairpins. J Chem Theory Comput
6(12):38363849
53. Zgarbova M, Otyepka M, Sponer J, Mladek
A, Banas P, Cheatham TE III et al (2011)
Refinement of the Cornell et al. Nucleic acids
force field based on reference quantum chemical calculations of glycosidic torsion profiles.
J Chem Theory Comput 7(9):28862902
54. Kirschner KN, Woods RJ (2001) Solvent
interactions determine carbohydrate conformation. Proc Natl Acad Sci U S A 98(19):
1054110545
55. Woods RJ, Dwek RA, Edge CJ, Fraser-Reid B
(1995) Molecular mechanical and molecular
dynamic simulations of glycoproteins and oligosaccharides. 1. GLYCAM_93 parameter development. J Phys Chem 99(11):38323846
56. Kirschner KN, Yongye AB, Tschampel SM,
Gonzlez-Outeirio J, Daniels CR, Foley BL
et al (2008) GLYCAM06: a generalizable
biomolecular force field. Carbohydrates.
J Comput Chem 29(4):622655
57. Skjevik GA, Madej BD, Walker RC, Teigen
K (2012) LIPID11: a modular framework for
lipid simulations using Amber. J Phys Chem B
116(36):1112411136
58. Brooks BR, Brooks CL III, MacKerell AD Jr,
Nilsson L, Petrella RJ, Roux B et al (2009)
CHARMM: the biomolecular simulation program. J Comput Chem 30(10):15451614
59. Jiang W, Hardy DJ, Phillips JC, Mackerell AD
Jr, Schulten K, Roux B (2011) Highperformance scalable molecular dynamics
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
simulations of a polarizable force field based

on classical Drude oscillators in NAMD.
J Phys Chem Lett 2(2):8792
Boulanger E, Thiel W (2012) Solvent boundary potentials for hybrid QM/MM computations using classical drude oscillators: a fully
polarizable model. J Chem Theory Comput
8:45274538
Eastman P, Friedrichs MS, Chodera JD,
Radmer RJ, Bruns CM, Ku JP et al (2012)
OpenMM 4: a reusable, extensible, hardware
independent library for high performance
molecular simulation. J Chem Theory Comput
8:461469
Lamoureux G, Roux B (2003) Modelling
induced polarizability with drude oscillators:
theory and molecular dynamics simulation
algorithm. J Chem Phys 119:51855197
Lamoureux G, MacKerell AD, Roux B (2003)
A simple polarizable model of water based
on classical Drude oscillators. J Chem Phys
119(10):51855197
Lamoureux G, Harder E, Vorobyov IV, Roux
B, MacKerell AD (2006) A polarizable model
of water for molecular dynamics simulations
of biomolecules. Chem Phys Lett 418(13):
245249
Anisimov VM, Lamoureux G, Vorobyov IV,
Huang N, Roux B, MacKerell AD (2005)
Determination of electrostatic parameters for
a polarizable force field based on the classical
Drude oscillator. J Chem Theory Comput
1(1):153168
Anisimov VM, Vorobyov IV, Lamoureux G,
Noskov S, Roux B, MacKerell AD (2004)
CHARMM all-atom polarizable force field
parameter development for nucleic acids.
Biophys J 86(1):415A
Anisimov VM, Vorobyov IV, Roux B,
MacKerell AD (2007) Polarizable empirical
force field for the primary and secondary alcohol series based on the classical drude model.
Lopes PEM, Lamoureux G, Roux B,
MacKerell AD (2007) Polarizable empirical
force field for aromatic compounds based on
the classical drude oscillator. J Phys Chem B
111(11):28732885
Harder E, Anisimov VM, Whitfield TW,
MacKerell AD, Roux B (2008) Understanding
the dielectric properties of liquid amides from
a polarizable force field. J Phys Chem B
112(11):35093521
Baker
CM,
MacKerell
AD
(2010)
Polarizability rescaling and atom-based Thole
scaling in the CHARMM Drude polarizable
force field for ethers. J Mol Model 16(3):
567576

71. Vorobyov I, Anisimov VM, Greene S, Venable
RM, Moser A, Pastor RW et al (2007)
Additive and classical drude polarizable force
fields for linear and cyclic ethers. J Chem
Theory Comput 3(3):11201133
72. Zhu X, MacKerell AD (2010) Polarizable
empirical force field for sulfur-containing compounds based on the classical drude oscillator
model. J Comput Chem 31(12):23302341
73. Baker CM, Anisimov VM, MacKerell AD
(2011) Development of CHARMM polarizable force field for nucleic acid bases based on
the classical drude oscillator model. J Phys
Chem B 115(3):580596
74. He X, Lopes PEM, MacKerell AD (2013)
Polarizable empirical force field for acyclic
polyalcohols based on the classical drude
oscillator. Biopolymers 99(10):724738
75. Harder E, MacKerell AD, Roux B (2009)
Many-body polarization effects and the membrane dipole potential. J Am Chem Soc
131(8):27602761
76. Chowdhary J, Harder E, Lopes PEM, Huang
L, MacKerell AD, Roux B (2013) A polarizable force field of dipalmitoylphosphatidylcholine based on the classical drude model for
molecular dynamics simulations of lipids.
J Phys Chem B 117(31):91429160
77. Shi Y, Xia Z, Zhang JH, Best RB, Wu C, Ponder
JW et al (2013) Polarizable atomic multipolebased AMOEBA force field for proteins.
78. Dudek MJ, Ponder JW (1995) Accurate
modeling of the intramolecular electrostatic
energy of proteins. J Comput Chem 16(7):
791816
79. Thole B (1981) Molecular polarizabilities calculated with a modified dipole interaction.
Chem Phys 59(3):341350
80. Ren PY, Ponder JW (2003) Polarizable atomic
multipole water model for molecular mechanics simulation. J Phys Chem B 107(24):
59335947
81. Ren PY, Ponder JW (2004) Temperature and
pressure dependence of the AMOEBA water
model. J Phys Chem B 108(35):1342713437
82. Grossfield A, Ren PY, Ponder JW (2003) Ion
solvation thermodynamics from simulation
with a polarizable force field. J Am Chem Soc
125(50):1567115682
83. Ren P, Wu C, Ponder JW (2011) Polarizable
atomic multipole-based molecular mechanics
for organic molecules. J Chem Theory
Comput 7(10):31433161
84. Shi Y, Wu C, Ponder JW, Ren P (2011)
Multipole electrostatics in hydration free
85.
86.
87.
88.
89.
90.
91.
92.
93.
94.
95.
96.
97.
69
energy calculations. J Comput Chem 32(5):

967977
Ponder JW, Case DA (2003) Force fields for
protein simulations, Protein simulations.
Academic, San Diego, pp 2785
Ponder JW, Wu C, Ren P, Pande VS, Chodera
JD, Schnieders MJ et al (2010) Current status
of the AMOEBA polarizable force field.
J Phys Chem B 114(8):25492564
Ren PY, Ponder JW (2002) Consistent treatment of inter- and intramolecular polarization
in
molecular
mechanics
calculations.
J Comput Chem 23(16):14971506
Jorgensen WL, Tirado-Rives J (1988) The
OPLS potential function for proteins. Energy
minimizations for crystals of cyclic peptides and
crambin. J Am Chem Soc 110:16571666
Singh UC, Kollman PA (1984) An approach
to computing electrostatic charges for molecules. J Comput Chem 5(2):129145
Chirlian LE, Francl MM (1987) Atomic
charges derived from electrostatic potentials:
a detailed study. J Comput Chem 8(6):
894905
Merz KM (1992) Analysis of a large data base
of electrostatic potential derived atomic
charges. J Comput Chem 13(6):749767
Bayly CI, Cieplak P, Cornell WD, Kollman
PA (1993) A well-behaved electrostatic
potential based method using charge restraints
for deriving atomic charges: the RESP model.
J Phys Chem 97(40):1026910280
Francl M, Carey C, Chirlian L, Gange D
(1996) Charges fit to electrostatic potentials.
II. Can atomic charges be unambiguously fit
to electrostatic potentials? J Comput Chem
17(3):367383
Lopes PEM, Lamoureux G, Mackerell AD
(2009) Polarizable empirical force field for
nitrogen-containing heteroaromatic compounds based on the classical Drude oscillator. J Comput Chem 30(12):18211838
Harder E, Anisimov VM, Vorobyov IV, Lopes
PEM, Noskov SY, MacKerell AD et al (2006)
Atomic level anisotropy in the electrostatic
modeling of lone pairs for a polarizable force
field based on the classical Drude oscillator.
Miller KJ (1990) Additivity methods in molecular polarizability. J Am Chem Soc 112(23):
85338542
Baker
CM,
MacKerell
AD
(2009)
Polarizability rescaling and atom-based Thole
scaling in the CHARMM Drude polarizable
force field for ethers. J Mol Model 16(3):
567576
70
98. Yu HA, Whitfield TW, Harder E, Lamoureux

G, Vorobyov I, Anisimov VM et al (2010)
Simulating monovalent and divalent ions in
aqueous solution using a drude polarizable
force field. J Chem Theory Comput 6(3):
774786
99. Jorgensen WL, Madura JD, Swenson CJ
(1984) Optimized intermolecular potential
functions for liquid hydrocarbons. J Am
Chem Soc 106(22):66386646
100. Jorgensen WL (1986) Optimized intermolecular potential functions for liquid alcohols.
J Phys Chem 90(7):12761284
101. MacKerell AD (2001) Atomistic models and
force fields. In: Becker O et al (eds)
Computational biochemistry and biophysics.
Marcel Dekker, Inc., New York, pp 738
102. Yin D, MacKerell AD (1996) Ab initio calculations on the use of helium and neon as
probes of the van der Waals surfaces of molecules. J Phys Chem 100(7):25882596
103. Yin DX, MacKerell AD (1998) Combined ab
initio empirical approach for optimization of
Lennard-Jones parameters. J Comput Chem
19(3):334348
104. Chen IJ, Yin D, MacKerell AD (2002)
Combined ab initio/empirical approach for
optimization of Lennard-Jones parameters for
polar-neutral compounds. J Comput Chem
23(2):199213
105. Baker CM, Lopes PEM, Zhu X, Roux B,
MacKerell AD (2010) Accurate calculation of
hydration free energies using pair-specific
Lennard-Jones parameters in the CHARMM
drude polarizable force field. J Chem Theory
Comput 6(4):11811198
106. Scott AP, Radom L (1996) Harmonic vibrational frequencies: an evaluation of HartreeFock, Mller-Plesset, quadratic configuration
interaction, density functional theory, and
semiempirical scale factors. J Phys Chem
100(41):1650216513
107. Pulay P, Fogarasi G, Pang F, Boggs JE (1979)
Systematic ab initio gradient calculation of
molecular geometries, force constants, and
dipole moment derivatives. J Am Chem Soc
101(10):25502560
108. Foloppe N, Hartmann B, Nilsson L,
MacKerell AD (2002) Intrinsic conformational energetics associated with the glycosyl
torsion in DNA: a quantum mechanical study.
Biophys J 82(3):15541569
109. Foloppe N, Nilsson L, MacKerell AD (2001)
Ab initio conformational analysis of nucleic
acid components: intrinsic energetic contributions to nucleic acid structure and dynamics. Biopolymers 61(1):6176
110. Lin B, Lopes PEM, Roux B, MacKerell AD

(2013) Kirkwood-Buff analysis of aqueous
N-methylacetamide and acetamide solutions
modeled by the CHARMM additive and
Drude polarizable force fields. J Chem Phys
139(8):084509
111. Halkier A, Helgaker T, Jrgensen P, Klopper
W, Koch H, Olsen J et al (1998) Basis-set convergence in correlated calculations on Ne, N2,
and H2O. Chem Phys Lett 286(34):243252
112. Graf J, Nguyen PH, Stock G, Schwalbe H
(2007) Structure and dynamics of the homologous series of alanine peptides: a joint
molecular dynamics/NMR study. J Am Chem
Soc 129(5):11791189
113. Kirkpatrick S, Gelatt CD, Vecchi MP (1983)
Optimization by simulated annealing. Science
220(4598):671680
114. Metropolis N, Rosenbluth AW, Rosenbluth
MN, Teller AH, Teller E (1953) Equation of
state calculations by fast computing machines.
J Chem Phys 21(6):10871092
115. Shoemaker KR, Kim PS, York EJ, Stewart JM,
Baldwin RL (1987) Tests of the helix dipole
model for stabilization of -helices. Nature
326(6113):563567
116. Shoemaker KR, Kim PS, Brems DN,
Marqusee S, York EJ, Chaiken IM et al (1985)
Nature of the charged-group effect on the
stability of the C-peptide helix. Proc Natl
Acad Sci 82(8):23492353
117. Padmanabhan S, Marqusee S, Ridgeway T,
Laue TM, Baldwin RL (1990) Relative helixforming tendencies of nonpolar amino acids.
Nature 344(6263):268270
118. Fukunishi H, Watanabe O, Takada S (2002)
On the Hamiltonian replica exchange method
for efficient sampling of biomolecular systems: application to protein structure prediction. J Chem Phys 116(20):90589067
119. Zhu X, Lopes PEM, Shim J, MacKerell AD
(2012) Intrinsic energy landscapes of amino
acid side-chains. J Chem Inf Model 52(6):
15591572
120. Lopes PEM, Huang J, Shim J, Luo Y, Hui L,
Roux B et al (2013) Polarizable force field for
peptides and proteins based on the classical
drude oscillator. J Chem Theory Comput.
doi:10.1021/ct400781b
121. Hegefeld WA, Chen S-E, DeLeon KY,
Kuczera K, Jas GS (2010) Helix formation in
a pentapeptide: experiment and force-field
dependent dynamics. J Phys Chem A 114(47):
1239112402
122. Best RB, Mittal J, Feig M, MacKerell AD
(2012) Inclusion of many-body effects in the
additive CHARMM protein CMAP potential
123.
124.
125.
126.
results in enhanced cooperativity of -helix

and -hairpin formation. Biophys J 103(5):
10451051
Lindorff-Larsen K, Maragakis P, Piana S,
Eastwood MP, Dror RO, Shaw DE (2012)
Systematic validation of protein force fields
against experimental data. PLoS One 7(2):
e32131
Karplus M (1959) Contact electron-spin coupling of nuclear magnetic moments. J Chem
Phys 30(1):1115
Kaminski GA, Friesner RA, Tirado-Rives J,
Jorgensen WL (2001) Evaluation and reparametrization of the OPLS-AA force field for
proteins via comparison with accurate quantum chemical calculations on peptides. J Phys
Chem B 105(28):64746487
Oostenbrink C, Villa A, Mark AE, Van
Gunsteren WF (2004) A biomolecular force
127.
128.
129.
130.
71
field based on the free enthalpy of hydration

and solvation: the GROMOS force-field
parameter sets 53A5 and 53A6. J Comput
Chem 25(13):16561676
Blanco FJ, Rivas G, Serrano L (1994) A short
linear peptide that folds into a native stable
[beta]-hairpin in aqueous solution. Nat Struct
Mol Biol 1(9):584590
Muoz V, Thompson PA, Hofrichter J, Eaton
WA (1997) Folding dynamics and mechanism
of -hairpin formation. Nature 390(6656):
196199
Schuler B, Eaton WA (2008) Protein folding
studied by single-molecule FRET. Curr Opin
Struct Biol 18(1):1626
Jo S, Kim T, Iyer VG, Im W (2008)
CHARMM-GUI: a web-based graphical user
interface for CHARMM. J Comput Chem
29(11):18591865
Chapter 4
Lipid Membranes forMembrane Proteins
AndreasKukol
Abstract
The molecular dynamics (MD) simulation of membrane proteins requires the setup of an accurate
representation of lipid bilayers. This chapter describes the setup of a lipid bilayer system from scratch
using generally available tools, starting with a definition of the lipid molecule POPE, generation of a
lipid bilayer, energy minimization, MD simulation, and data analysis. The data analysis includes the
calculation of area and volume per lipid, deuterium order parameters, self-diffusion constant, and the
electron density profile.
Key words Lipid bilayer, Molecular dynamics, Simulation, Trajectory analysis, Area per lipid, Volume
per lipid, Deuterium order parameter, Self-diffusion constant, Electron density profile
1 Introduction
Molecular simulations of membrane proteins require consideration
of the lipid membrane environment. While molecular dynamics
(MD) simulations with implicit membrane models have been used
successfully [1], for higher accuracy explicit representation of the
lipid bilayer is desirable. Furthermore, dependent on the research
question, if lipidprotein interactions are a subject of the study, an
explicit representation of lipid molecules in unavoidable. Having
decided on an explicit representation of lipids, further choice exists
between coarse-grained, united-atom, and all-atom lipid models
and force fields. In coarse-grained forcefields (covered in Chapter
7 of this book) several atoms are subsumed into one particle, for
example in the MARTINI force field [2, 3] four carbon atoms of
the aliphatic chain are subsumed into one particle. All-atom force
fields usually provide the highest accuracy for the description of
lipids and proteins. United-atom force fields subsume nonpolar
hydrogen atoms into their adjacent carbon-atoms resulting in a
moderate reduction of the number of particles, e.g., for a
1,2-dipalmitoyl-glycero-3-phosphocholine (DPPC) from 130 particles for the all-atom model to 50 particles for the united-atom
73
74
Andreas Kukol
model. United-atom models of several lipids have been shown to

reproduce the experimentally measured properties of lipid bilayers
to reasonable accuracy [4, 5]. Another consideration is the choice
of force field for the protein, which is covered in Chapter 3 of this
book. As a general principle different force fields should not be
mixed, but protein, lipid, and possibly small organic molecules
should be represented by the same force field. For practical reasons
the choice of the force field is often based on the availability of the
desired lipid model or topology for that particular force field.
Once the decision for a particular force field and lipid model
for the use in MD simulations of a membrane protein has been
made, a simulation of the lipid bilayer should be made and compared with experimental data. The membrane protein simulation
may require a mixture of lipids in order to better represent the
invivo environment of a cell membrane, but unfortunately very
few experimental data of mixed lipid membrane systems are available and virtually none of lipid membranes with proteins. The
computational validation is, therefore, restricted to the individual
components of the complete system to be investigated.
In this chapter the setup and MD simulation of a 1-palmitoyl-
2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) bilayer with
the united-atom force field GROMOS96 54a7 [6] is described followed by analysis of the simulation and comparison with experimental data. The MD simulation uses the software Gromacs
version 4.5 [7, 8]. Similar procedures apply to other MD simulation software and/or force fields.
2 Methods
2.1 Materials
GROMACS version 4.5.3 or higher [8, 9]

CHARMM-GUI (http://www.charmm-gui.org/) via a web
browser [10, 11]
GROMOS96 54a7 force field [6, 12] obtained from the Automated
Topology Builder and Repository web site (http://compbio.
chemistry.uq.edu.au/atb/) [13], move the folder gromos54a7.ff directly into share/top and Gromacs will
recognize the new force field automatically.
Perl interpreter
LINUX operating system
8-core computer workstation
2.2 Lipid Topology
The first step is to obtain or develop a molecular topology for POPE

using the desired force field, in this case GROMOS96 54a7. The
molecular topology specifies the atom types, atom charges, bonds
and angles between atoms. Predefined parameters from the force
75
field are assigned to those atoms, bonds, and angles, for example a
predefined bond length and force constant. A good place to obtain
lipid topologies for various force fields is Lipidbook (http://
lipidbook.bioch.ox.ac.uk/) [14]. For this chapter, we will use a
new topology of POPE shown in Fig.1 based on topologies developed in earlier work for the GROMOS96 53a6 force field [4].
2.3 Lipid
BilayerSetup
The starting coordinates of a lipid bilayer can be conveniently

generated with the Membrane Builder of the CHARMM-GUI
[10, 11].
1.
Using a web browser go to http://www.charmm-gui.
org/?doc=input/membrane_only&step=1
2. Under point 3. Length of XY based on select Numbers of
lipid components
3. Into the boxes next to POPE enter 64 for the number of lipids
in upper leaflet and 64 for the lower leaflet.
4. Click on Next Step: Determine the system size.
5. An output box appears, with the progress of the calculation.
On the next page counter ions can be added. Select Add
neutralizing ions, which are zero numbers because POPE is
neutral. Click on Next Step.
6. Once the calculation has finished click on Next Step: Assemble
components.
7. When the calculation has finished, you are taken to the next
page, which allows to download the coordinates of the water/
lipid bilayer. Right click on step5_assembly.pdb and save it on
your local computer for further equilibration with GROMACS
(see Note 1).
2.4 Lipid Bilayer

Equilibration
The first step is the construction of a system topology (Fig.2) and

a parameter (mdp-) file for energy minimization (Fig.3).
Additionally the molecular topology (itp-file) for POPE, pope_
gromos54a7.itp shown in Fig.1 is required. The lipid bilayer
obtained from the CHARMM GUI contains a large number of
near atomic clashes, which would cause energy minimization to fail.
Therefore, we need to adopt a strategy that reduces the almost infinite energy from atomic clashes, by reducing the non-bonded interactions to a very low value as well as setting the emstep value (in
Fig.3) very low. At the same time the number of energy minimization steps is kept low (15 steps in Fig.3) in order to avoid unusual
distortions to the molecules, e.g., due to electrostatic attractions
which are not compensated by van der Waals repulsions.
1. A simulation box is defined around the system:
ediconf f step5_assembly.pdb o step5_assembly.gro d 0.0
76
Andreas Kukol
[ moleculetype ]
nrexcl
; Name
POPE
3
[ atoms ]
;
nr
type
1
H
2
H
3
H
4
NL
5
CH2
6
CH2
7
OA
8
P
9
OM
10
OM
11
OA
12
CH2
13
CH1
14
OE
15
C
16
O
17
CH2
18
CH2
19
CH2
20
CH2
21
CH2
22
CH2
23
CH2
24
CR1
25
CR1
26
CH2
27
CH2
28
CH2
29
CH2
30
CH2
31
CH2
32
CH2
33
OE
34
C
35
O
36
CH2
37
CH2
38
CH2
39
CH2
40
CH2
41
CH2
42
CH2
43
CH2
44
CH2
45
CH2
46
CH2
47
CH2
48
CH2
49
CH2
50
CH3
51
CH2
52
CH3
[ bonds ]
;
ai
4
5
6
7
8
8
8
11
12
13
13
14
15
15
17
18
19
20
21
22
23
24
resnr
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
aj
5
6
7
8
9
10
11
12
13
14
32
15
16
17
18
19
20
21
22
23
24
25
residu
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
POPE
funct
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
atom
H1
H2
H3
N4
C5
C6
O7
P8
O9
O10
O11
C12
C13
O14
C15
O16
C17
C18
C19
C20
C21
C22
C23
C24
C25
C26
C27
C28
C29
C30
C31
C32
O33
C34
O35
C36
C37
C38
C39
C40
C41
C42
C43
C44
C45
C46
C47
C48
C49
C50
CA1
CA2
cgnr
0
0
0
0
0
1
1
1
1
1
1
2
2
2
2
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
18
18
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
charge
0.3000
0.3000
0.3000
-0.2
0.3
0.4
-0.8
1.7
-0.8
-0.8
-0.7
0.4
0.3
-0.7
0.7
-0.7
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.5
-0.7
0.8
-0.6
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
gb_21
gb_27
gb_18
gb_28
gb_24
gb_24
gb_28
gb_18
gb_27
gb_18
gb_27
gb_10
gb_5
gb_23
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_10
Fig. 1 The complete topology of POPE in the GROMOS96 54a7 force field
mass
1.0080 ; qtot:0.3
1.0080 ; qtot:0.6
1.0008 ; qtot:0.9
14.0067 ; qtot:0.7
14.0270 ; qtot:1.0
14.0270 ; qtot:1.0
15.9994 ; qtot:0.54
30.9738 ; qtot:2.3
15.9994 ; qtot:1.5
15.9994 ; qtot:0.7
15.9994 ; qtot:0
14.0270 ; qtot:0.08
13.0190 ; qtot:0.52
15.9994 ; qtot:-0.14
12.0110 ; qtot:0.56
15.9994 ; qtot:0.0
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
13.0190 ; qtot:
13.0190 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
15.9994 ; qtot:
12.0110 ; qtot:
15.9994 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
14.0270 ; qtot:
15.0350 ; qtot:
14.0270 ; tail2
15.0350; tail2

25
26
27
28
29
30
31
51
32
33
34
34
36
37
38
39
40
41
42
43
44
45
46
47
48
49
1
2
3
26
27
28
29
30
31
51
52
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
4
4
4
[ pairs ]
; ai
aj funct
1
6
1
2
6
1
3
6
1
4
7
1
5
8
1
6
9
1
6
10
1
6
11
1
7
12
1
8
13
1
9
12
1
10
12
1
11
14
1
11
32
1
12
15
1
12
33
1
13
16
1
13
17
1
13
34
1
14
18
1
14
33
1
15
19
1
15
32
1
16
18
1
22
25
1
24
27
1
32
35
1
32
36
1
33
37
1
34
38
1
35
37
1
[ angles ]
; ai
aj
4
5
6
7
7
7
8
9
9
10
11
12
12
13
13
14
14
Fig. 1(continued)
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_18
gb_10
gb_5
gb_23
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_27
gb_2
gb_2
gb_2
;H-N bond type
; pair around double bond

; pair around double bond
ak funct
5
6
6
7
7
8
8
9
8
10
8
11
11
12
8
10
8
11
8
11
12
13
13
14
13
32
14
15
32
33
13
32
15
16
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
ga_15
ga_15
ga_26
ga_14
ga_14
ga_5
ga_26
ga_29
ga_14
ga_14
ga_15
ga_13
ga_13
ga_22
ga_15
ga_13
ga_31
77
78
Andreas Kukol
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
1
2
3
1
2
3
[ dihedrals ]
; ai
aj
1
4
4
5
4
5
5
6
6
7
6
7
7
8
7
8
8
11
11
12
11
12
11
12
12
13
12
13
12
13
13
32
13
14
14
13
14
15
15
17
17
18
18
19
19
20
20
21
21
22
21
22
21
22
22
23
24
25
25
26
25
26
25
26
26
27
27
28
28
29
29
30
30
31
13
32
Fig. 1(continued)
15
17
15
18
19
20
21
22
23
24
25
26
27
28
29
30
31
51
33
34
34
36
34
37
38
39
40
41
42
43
44
45
46
47
48
49
4
4
4
4
4
4
ak
5
6
6
7
8
8
11
11
12
13
13
13
32
32
14
33
15
32
17
18
19
20
21
22
23
23
23
24
26
27
27
27
28
29
30
31
51
33
17
18
17
19
20
21
22
23
24
25
26
27
28
29
30
31
51
52
34
35
36
37
36
38
39
40
41
42
43
44
45
46
47
48
49
50
2
3
1
5
5
5
al funct
6
1
7
1
7
1
8
1
11
1
11
1
12
1
12
1
13
1
14
1
32
1
32
1
33
1
33
1
15
1
34
1
17
1
33
1
18
1
19
1
20
1
21
1
22
1
23
1
24
1
24
1
24
1
25
3
27
3
28
1
28
1
28
1
29
1
30
1
31
1
51
1
52
1
34
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
ga_16
ga_15
ga_35
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_27 ;double bond
ga_27 ; double bond
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_22
ga_31
ga_16
ga_15
ga_35
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_15
ga_10
ga_10
ga_10
ga_11
ga_11
ga_11
phi0
gd_29
gd_4
gd_36
gd_29
gd_20
gd_27
gd_20
gd_27
gd_29
gd_34
gd_34
gd_17
gd_34
gd_17
gd_29
gd_29
gd_13
gd_18
gd_40
gd_34
gd_34
gd_34
gd_34
gd_34
0
180
0
2.885
2.885
0
180
0
gd_34
gd_34
gd_34
gd_34
gd_34
gd_29
cp
mult
3.350
1
1.660
2
7.333
3
4.17 7.8 4.4 0.0 0.0
4.17 7.8 4.4 0.0 0.0
3.350
1
1.660
2
7.333
3

32
33
34
36
37
38
39
40
41
42
43
44
45
46
47
33
34
36
37
38
39
40
41
42
43
44
45
46
47
48
34
36
37
38
39
40
41
42
43
44
45
46
47
48
49
[ dihedrals ]
; ai
aj
ak
13
14
15
14
34
33
23
24
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
al funct
32
17
36
25
79
gd_13
gd_40
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
gd_34
12
16
35
26
2
2
2
2
gi_2
gi_1
gi_1
gi_1 ; double bond
#ifdef POSRES_LIPID
#include "lipid_posre.itp"
#endif
Fig. 1(continued)
Fig. 2 The topology of the whole lipid bilayer system composed of 128 POPE molecules and 2560 water molecules.
The molecular topologies are read from itp-files via the #include commands
80
Andreas Kukol
Fig. 3 The parameters for initial energy minimization. Note that the values for nsteps (normally thousands) and
emstep (normally 0.1) are set very low in this example
81
2. The coordinate file with simulation box, topology files, and

mdp-file are put together into a binary simulation run file (tpr-
file) (see Note 2):
grompp -f em1.mdp c step5_assembly.gro p
system.top pp processed.top o run.tpr
The file processed.top contains a complete topology of

the system.
3. The van der Waals interactions in the topology are then drastically reduced:
RescaleNonBond.pl processed.top 0.005 > em1.
top
The Perl script RescaleNonBond.pl is shown in Fig.4.
4. The new topology em1.top is used to run the energy

minimization.
grompp -f em1.mdp c step5_assembly.pdb p
em1.top o run.tpr
mdrun s em1.tpr c em1.gro -v
5. Using em1.gro as the coordinate file instead of step5_

assembly.gro (-c option) the steps 24 above are repeated
with increasing scale factors of 0.05, 0.1, and 0.5. At the same
time the number of energy minimization steps in em1.mdp is
increased to up to 40 steps. The exact value of the scale factors
and number of energy minimization steps is a matter of trial
and error.
6. For the final energy minimization no rescaling of van der Waals
interactions is used, the parameter emstep in Fig.3 is set to 0.01
and the nsteps to 2000. This yields the file final_em.gro.
Further equilibration could be performed, such as performing
a short MD simulation with position restraints on the phosphorus
atoms of the lipid molecules or MD simulations at constant v olume
(NVT-ensemble). However, we did not find that necessary as lipids
tend to self-assemble into their equilibrium position during the
MD simulation at constant pressure (NPT-ensemble). For analysis
of the equilibrium properties the first few 10th of nanoseconds of
an NPT simulation can be discarded.
2.5 MD
SimulationRun
The coordinate file of the final energy minimization can be directly

used to run a long MD simulation over 100ns in this example.
The parameter file for this simulation is shown in Fig.5.
grompp -f md_100ns.mdp c final_em.gro p system.top o run_100ns.tpr
mdrun s run_100ns.tpr x run_100ns.xtc e
run_100ns.edr g run_100ns.log c after_100ns.
gro v stepout 2000 (see Note 3)
The most important output files for further data analysis are
the trajectory run_100ns.xtc and the energy file run_100ns.edr,
which also contains the data about the system dimension.
82
Andreas Kukol
Fig. 4 The Perl-script ScaleNonBond.pl to rescale the non-bonding interactions of a GROMACS processed
topology
Fig. 5 The parameters that were used to run the 100ns MD simulation of the lipid bilayer system
83
84
Andreas Kukol
Fig. 5(continued)
85
2.6 Data Analysis
Since this chapter is aimed at providing lipid membranes for membrane proteins, rather than investigating the properties of lipid
bilayers in detail, the aim of the data analysis is to establish, if the
simulation reproduces experimentally known parameters to reasonable accuracy. Typically the area per lipid and volume per lipid
are compared with experiment; for POPE this is available from a
study by Rappolt etal. [15]. Deuterium order parameters and self-
diffusion constants are available for some lipids. If they are not
available for the particular type of lipid, they may be compared
with values from other lipids in order to check for errors in the
topology or simulation.
2.6.1 Area andVolume

per Lipid
These observables are calculated from the size of the simulation

box, which is usually arranged in a way that the z-coordinate is
perpendicular to the lipid bilayer plane. The area per lipid AL is
then obtained by the area of the simulation box divided by the
number of lipid molecules in one leaflet, i.e., AL=sizex sizey/64.
The volume per lipid VL is obtained as:
VL = ( Vbox Vwater ) / 128 = ( sizex size y sizez Vwater ) / 128
The volume of water Vwater can be obtained by performing an MD

simulation of a box of water molecules using the same MD parameters that were used for the lipid/water simulation. This was done
and the volume of a water molecule was determined as
(0.030580.00008) nm3, leading to a total volume for 2560 water
molecules (see Fig.2) of Vwater=78.28 nm3.
The size of the simulation box along the trajectory is c omputed
with g_energy:
g_energy f run_100ns.edr s run_100ns.tpr
o box_XYZ.txt
In the dialogue select Box-X, Box-Y, and Box-Y by typing
the corresponding numbers. The output file can be imported in a
spreadsheet program for calculating the area and volume per lipid
as described above. For POPE the change of area and volume per
lipid over the 100ns trajectory is shown in Fig.6. Averages are
calculated for the last 70ns of the trajectory (Table1) and show
less than 5% deviation from experimental data.
2.6.2 Lipid Acyl Chain
Order Parameters
Order parameters of the lipid acyl chains can be measured from

solid-state deuterium NMR spectra. The order parameter for a
C-D bond is directly calculated from the measured quadrupolar
splittings. It is indicative of the average orientation of the C-D
vector with respect to the external magnetic field, normally the
z-axis in aligned lipid bilayers. In order to calculate the order
parameter for a united-atom force field, the C-D bond vector is
86
Andreas Kukol
Fig. 6 Area/lipid (black curve) and volume/lipid (grey curve) over the course of the 100ns simulation
Table 1
Properties of the POPE lipid bilayer from simulations compared with experimental data
Simulation
Experiment
Area/lipid
(0.6200.008) nm2
0.6025nm2 [15]
Volume/lipid
(1.1350.003) nm3
1.175nm3 [15]
Self-diffusion coefficient
(6.420.02) 108cm2/s
8.87 108cm2/s (for POPC) [16]
reconstructed automatically by the g_order tool of the Gromacs

software.
1. Prepare an index-file sn1.ndx for the sn1 lipid acyl chain that
contains the atom numbers of each carbon at equivalent
positions:
[C34]
34 86 138
[C36]
36 88 140
[C37]
37 81 141
[C50]
For the sn2 lipid acyl chain, the order parameters must be
calculated separately for the saturated and unsaturated carbons.
The index file for the saturated carbons sn2.ndx contains all
atoms:
[C15]
[C17]
[C18]
87
[C31]
[CA1]
[CA2]
The index file for the unsaturated carbons sn2_unsat.ndx
contains index groups for the unsaturated carbons and the two
neighbors on each side:
[C23], [C24], [C25], [C26]
The index groups are made interactively with:
make_ndx f after_100ns.gro o sn1.ndx
This opens an interactive session, in which you create
index groups for each acyl chain atom using the add command: a c34, a c36, a c37, and so on. Finally
you delete the default groups: del 0-5 and quit.
2. Calculate the order parameters over the last 70ns of the simulation from the trajectory:
g_order f run_100ns.txt s run_100ns.tpr n
sn1.ndx od deuter_sn1.xvg
b 30000
sn2.ndx od deuter_sn2.xvg
b 30000
sn2_unsat.ndx
od deuter_sn2unsat.xvg b 30000
3. Using a text editor replace the order parameters for the unsaturated carbons in deuter_sn2.xvg by the corresponding values
from deuter_sn2unsat.xvg (see Note 4).
2.6.3 Lateral Self-
Diffusion Coefficient
1. An index file need to be prepared that contains all atoms numbers belonging to lipid molecules:
make_ndx f after_100ns.gro o lipids.ndx
One of the default index group should correspond to the
lipid, e.g., 2 POPE. Then you type keep 2 and q for save
and quit.
2. The self-diffusion coefficient can then be calculated with the
g_msd tool.
g_msd f run_100ns.xtc s run_100ns.tpr n
lipids.ndx
lateral z mol diffusion.xvg o msd.xvg b
50000
A value of (6.420.02) 108cm2/s is reported, which is in
the right region for lipid diffusion. Note that only the last 50ns
of the trajectory were analyzed in the example above due to
computer memory limitations.
2.6.4 Electron Density
The electron density distribution in z-direction (perpendicular to

the membrane plane) is one of the best indicators of the experimental accuracy of a lipid membrane simulation, since it can be directly
88
Andreas Kukol
55
H1 = 1
H2 = 1
H3 = 1
N4 = 7
C5 = 8
C6 = 8
O7 = 8
P8 = 15
O9 = 8
O10 = 8
...
...
...
OW = 8
HW1 = 1
HW2 = 1
Fig. 7 An extract of the file electrons.dat that specifies the number of

electrons for each atom. The first line specifies the number of input lines, which
are the sum of 52 lines for lipid atoms and 3 lines for water atoms
obtained from the inverse Fourier-transformation of the small angle

X-ray scattering (SAXS) curve. Other parameters derived from
SAXS experiments, such as area per lipid (see Subheading2.6.1)
require fitting of the SAXS curves to an electron density model
of the lipid bilayer, in which the area per lipid is one of many
free parameters. It is, therefore, preferable to compare the electron
density distribution of the MD simulation directly with the
electron density obtained from SAXS curves.
1. A text file (electrons.dat) is required that contains the
number of electrons associated with each atom in the structure, an extract of the file is shown in Fig.7.
2. The electron density averaged over the trajectory from 30 to
100ns is then calculated with the g_density tool:
g_density f run_100ns.xtc s run_100ns.tpr
ei electrons.dat b 30000 dens electron d
Z symm o e_density.xvg
The option -d specifies the axis normal to the membrane
and -symm symmetrizes the density around the axis; -symm
should not be used for asymmetric bilayers.
3. The electron density is contained in the file e_density.xvg
and plotted in Fig.8.
3 Notes
1. Clicking further on Next Step provides us with the required
input file to run CHARMM MD simulations to equilibrate the
bilayer on the local computer. Since we want to use a different
force field, we will continue the equilibration with Gromacs.
89
Fig. 8 The electron density in electrons/nm3 along the z-coordinate of the simulation box
2. The output of grompp usually contains many warnings about

non-matching atom names between coordinate file and topology.
Grompp terminates with the fatal error of too many warnings.
The warnings about non-matching atom names can be safely
ignored, but others should be inspected carefully. Then rerun
grompp by setting the maxwarn option.
3. On a standard LINUX workstation the mdrun command
would be run in the background: nohup mdrun &
The output produced by the mdrun command is then available in nohup.out.
In a high-performance cluster environment you need to
consult the documentation about how to schedule and run
jobs on the cluster.
4. Note that if you specified N atoms in the index file, the order
parameters are calculated for N-2 atoms, no order parameter is
calculated for the first and the last atom in the index file.
Acknowledgements
This work was supported by the School of Life and Medical
Sciences, University of Hertfordshire and has made use of the
University of Hertfordshire Science and Technology Research
Institute high-performance computing facility. I thank all research
groups that made their tools and programs available to the research
community.
90
Andreas Kukol
References
1. Tanizaki S, Feig M (2006) Molecular dynamics
simulations of large integral membrane proteins with an implicit membrane model. J Phys
Chem B 110(1):548556
2. Monticelli L, Kandasamy SK, Periole X, Larson
RG, Tieleman DP, Marrink SJ (2008) The
MARTINI
coarse-grained
force
field:
Extension to proteins. J Chem Theory Comput
4(5):819834
3. Marrink SJ, Risselada HJ, Yefimov S, Tieleman
DP, de Vries AH (2007) The MARTINI
force field: Coarse grained model for biomolecular simulations. J Phys Chem B 111(27):
78127824
4. Kukol A (2009) Lipid Models for United-Atom
Molecular Dynamics Simulations of Proteins.
JChem Theory Comput 5(3):615626
5. Ulmschneider JP, Ulmschneider MB (2009)
United
Atom
Lipid
Parameters
for
Combination with the Optimized Potentials
for Liquid Simulations All-Atom Force Field.
JChem Theory Comput 5(7):18031813
6. Schmid N, Eichenberger AP, Choutko A,
Riniker S, Winger M, Mark AE etal (2011)
Definition and testing of the GROMOS force-
field versions 54A7 and 54B7. Eur Biophys J
Biophys Lett 40(7):843856
7. Pronk S, Pall S, Schulz R, Larsson P, Bjelkmar P,
Apostolov R etal (2013) GROMACS 4.5: a
high-throughput and highly parallel open source
molecular simulation toolkit. Bioinformatics
29(7):845854
8. Hess B, Kutzner C, van der Spoel D, Lindahl E
(2008) GROMACS 4: Algorithms for highly efficient, load-balanced, and scalable molecular simulation. J Chem Theory Comput 4(3):435447
9. Van der Spoel D, Lindahl E, Hess B, Groenhof

G, Mark AE, Berendsen HJC (2005)
GROMACS: Fast, flexible, and free. J Comput
Chem 26(16):17011718
10. Jo S, Kim T, Im W (2007) Automated Builder
and Database of Protein/Membrane Complexes
for Molecular Dynamics Simulations. Plos
One2(9)

11. Jo S, Lim JB, Klauda JB, Im W (2009)
CHARMM-GUI Membrane Builder for Mixed
Bilayers and Its Application to Yeast
Membranes. Biophys J 97(1):5058
12. Wang DQ, Freitag F, Gattin Z, Haberkern H,
Jaun B, Siwko M etal (2012) Validation of the
GROMOS 54A7 Force Field Regarding Mixed
alpha/beta-Peptide
Molecules.
Helvetica
Chimica Acta 95(12):25622577
13. Malde AK, Zuo L, Breeze M, Stroet M, Poger
D, Nair PC etal (2011) An Automated Force
Field Topology Builder (ATB) and Repository:
Version 1.0. J Chem Theory Comput 7(12):
40264037

14. Domanski J, Stansfeld PJ, Sansom MSP,
Beckstein O (2010) Lipidbook: A Public
Repository for Force-Field Parameters Used in
Membrane Simulations. J Membr Biol 236(3):
255258
15. Rappolt M, Hickel A, Bringezu F, Lohner K
(2003) Mechanism of the lamellar/inverse
hexagonal phase transition examined by high
resolution x-ray diffraction. Biophys J 84(5):
31113122

16. Filippov A, Oradd G, Lindblom G (2003)
Influence of cholesterol and water content on
phospholipid lateral diffusion in bilayers.
Langmuir 19(16):63976400
Chapter 5
Molecular Dynamics Simulations of Membrane Proteins
Abstract
Membrane protein structures are underrepresented in the Protein Data Bank (PDB) due to difficulties
associated with expression and crystallization. As such, it is one area where computational studies, particularly Molecular Dynamics (MD) simulations, can provide useful additional information. Recently, there has
been substantial progress in the simulation of lipid bilayers and membrane proteins embedded within
them. Initial efforts at simulating membrane proteins embedded within a lipid bilayer were relatively slow
and interactive processes, but recent advances now mean that the setup and running of membrane protein
simulations is somewhat more straightforward, though not without its problems. In this chapter, we outline practical methods for setting up and running MD simulations of a membrane protein embedded
within a lipid bilayer and discuss methodologies that are likely to contribute future improvements.
Key words Molecular dynamics, Simulation, Computational, Membrane proteins, Ion channels
Introduction
Membrane proteins are thought to constitute approximately 30 %
of genomes [1]. Furthermore it has been estimated that over half
of all drug targets are membrane proteins [2]. However, due to
problems associated with expression and crystallization, the number of high-resolution crystal structures is less than 1 % of the total
number of structures (see http://blanco.biomol.uci.edu/mpstruc/
for a maintained list of membrane protein structures). The situation is further complicated by the fact that many membrane proteins undergo quite large conformational changes in order to
complete their function (for example transporter proteins [35]
which cycle between at least two distinct states). Crystallography
will at best only be able to capture a time and space averaged snapshot of these states. Computer simulations on the other hand, and
in particular Molecular Dynamics (MD) simulations, are useful
tools that in addition to providing information on the stability of a
membrane protein can also provide insight into the manner in
which these conformational changes can proceed. Thus there has
91
92
been a large increase in the application of computer simulation

methods to membrane proteins [6, 7] ranging from ion channels
[810] to outer-membrane proteins [11]. MD can also be used to
test hypotheses in both idealized systems where one can explore
underlying biophysical principals governing a process [12] through to
systems which represent the in vivo systems as close as possible [13].
The field of membrane-protein simulation has matured over
recent years and there are now many groups worldwide performing simulations. One reason for the recent increase in the number
of research groups is that the computational facilities required to
perform membrane-protein simulations are now accessible to more
people. As the most common computational method used is MD
simulation, we will discuss in this chapter, how to set up and run,
a membrane-protein simulation focussing on the more practical
aspects. Until recently the setup was rather complicated and
required a large amount of interactive input from the researcher.
Now, principally due to increases in computational power, the
setup and running of such simulations is much simpler. We divide
the process up into 4 distinct steps; the preparation of the protein
itself, the preparation of the lipid (though only briefly as the main
thrust of this chapter is on the setup and running of a membrane
protein simulation), the actual insertion and establishing of a stable
system and finally, running the simulation. Of these, it is perhaps
the preparation of the protein itself which requires the most care
and interactive input from the researcher.
Theory
The underlying theory for molecular dynamics simulations is
covered in Chapters 1 and 3, and therefore in this section we briefly
discuss some specific considerations that researchers should bear in
mind when performing simulations of membrane proteins. Perhaps
the most important of these is the timescale of the problem that
is under consideration and the resource that is available. The many
different aspects of membrane dynamics span a large timescale ranging from a few picoseconds (for a protein side-chain to rotate)
though to minutes and longer (for flip-flop motion of lipids). Indeed,
where resources are minimal and only an approximate representation
of the bilayer is required, one may be content with using a slab of
octane to represent the hydrophobic core of the bilayer [14] or even
a hybrid model [15]. Substantial recent efforts have been made to
approximate the lipid molecules in a different way by using a coarsegrain approach, where typically 4 atoms are represented by one particle [16]. These methods have become popular as they allow much
longer timescale events to be explored, but of course they are less
detailed than a fully atomistic simulation.
93
Typically for simulations of membrane proteins a stable

simulation is required, usually reflecting some sort of equilibrium
of the system. In these simulations, we have two components to
worry about: the protein and the lipid bilayer. Thus some metrics
of stability are required. For the protein, the most common of
these is the root mean square deviation (RMSD) of C atoms from
the initial (usually X-ray) starting structure. In the case of the lipid,
a good indicator is the mean surface area per lipid [17].
The basic theory underlying the insertion process below is
very simple. We place the protein in the bilayer and remove overlapping atoms. We then allow the whole system to relax and equilibrate as the lipids adjust conformation around the protein. The
positioning of the protein is still somewhat subjective, especially
with respect to its displacement along the membrane normal axis.
However, structural bioinformatics analysis [18] has demonstrated that nearly all membrane proteins have two aromatic girdles (although not Phenylalanine) separated by about 30 .
These girdles are thought to interact with the interface region of
the lipid bilayer and so provide an approximate indication of how
to position the protein. Another approach is to treat the protein as
a rigid body and optimize the transfer free energy between water
and a hydrophobic slab that represents the location of the lipid
bilayer [1921]. Some of these methods have been developed into
web servers where one can obtain a pre-orientated inserted protein (see for example Subheading 2.6 in Chapter 17). We discuss
towards the end of the chapter how this issue may also be addressed
via a coarse-grained simulation approach. If one is not interested
in the interaction of the protein with the lipid at the particle level
at all, then a more suitable approach may be to use an implicit
membrane.
Methods
There are obviously two main components to a membrane protein
simulation: the actual protein and the lipid bilayer it is to be embedded in. Although we focus on the issues concerning the whole
system, it is worth briefly reviewing practical considerations for
these individual components.
3.1 Preparation
of the Protein
Typically the starting point for the protein will be a structure

deposited in the protein data bank (PDB; www.rcsb.org). Often
however, these structures will need a certain amount of preparation before production level molecular dynamics (MD) can be
run. The most severe of these considerations might be missing
atoms, which can range from entire loops to a couple of side-chain
atoms. How one deals with this problem depends upon the
94
question one is trying to address with the simulation. For the case
where only a few atoms are missing from a small number of
side-chains one can manually build in the missing atoms using an
interactive modelling program such PyMOL [22] or What-If [23]
(see Note 1). For the more complicated case where whole loops are
missing, typically one has to resort to programs which can build
random structures which are geometrically correct such as Modeller
[24, 25]. Indeed, in some cases, it may be that construction of an
entire homology model is required (covered in Chapters 15, 16
and 17). Another related consideration is how to deal with the
termini in the structure. Frequently, the structure is not the whole
sequence of the protein, and therefore charged termini may not
be appropriate. One common procedure has been to build on capping groups that help to best mimic the continuing protein chain
(see Note 2). A simpler approach involves simply protonating the
C-terminus and deprotonating the N-terminus.
In all but the very high-resolution structures, one will still have
to add hydrogen atoms, as these will not be present in the PDB
file. Although this is a very simple process, there are decisions to be
made even for this process: (1) First, the choice of force-field is
importantin particular, whether it is an all atom, such as the
CHARMM parameter sets [26, 27], or a united-atom model in
which only polar hydrogens are explicit, as in the GROMOS [2830]
and Berger lipid [31] sets. It is worth bearing in mind that lipid
force fields are continually under refinement to improve agreement with experimental data [29, 32, 33]. There are many force
fields available for simulating membranes [34], and recent efforts
towards systematically comparing them and assessing their relative
strengths and weaknesses have been reported [35]. (2) Secondly,
the protonation states of ionizable side-chains in proteins must be
considered. United-atom force fields will give the benefit of
reduced computational effort due to reduction in the number of
particles, but all-atom models might be preferred in some cases
where greater accuracy is required. Various programs exist to calculate the pKa of ionisable side-chains (PROPKA [36, 37], H++ [38,
39], WHAT IF [23]), several of which also exist as online servers
(see Note 1). A Graphical User Interface (GUI) has recently been
developed as a plug-in for VMD [40] to help interpret the results
of PROPKA-based pKa predictions [41]. Most of the programs
rely upon calculating an estimate of the free energy (via the thermodynamic cycle) of protonating the residue within its proteinaceous environment. It may be the case that the protonation state
is not important, in which case default ionization states at pH 7 are
assumed. However, there are examples where the protonation state
may be critical as exemplified by the protonation state of Glu71 in
KcsA [4245]. The position of the hydrogens on histidine residues
should also be considered carefully, usually by simple visual inspection to optimize local hydrogen bonding.
95
Finally, although solvation of the system is generally automated,

oxygen atoms from water molecules are often included in protein
crystal structures. These reflect low-energy minima for a water molecule and thus it is usual to include these prior to bulk solvation.
Deciding whether a water molecule belongs to the subunit of
interest from the PDB file is a problem and usually one simply
chooses an arbitrary cutoff within which to include these crystallographic waters in the simulation. A cutoff that we have employed in
the past has been 4 [4648]. The remainder of the simulation
box is solvated with pre-equilibrated water boxes. Although we are
typically interested in the proteinlipid interactions it is important
to have adequate solvation of the entire protein (see Note 3).
3.2 Preparation
of the Lipid Bilayer
Not all of the methods below rely on a preformed lipid bilayer.

However, one invariably will need a lipid-only system for control
purposes so simulation of the pure system should be done at some
point. The simulation of lipid bilayers has matured over the past
2025 years and it is beyond the scope of this chapter to cover this
in depth. The interested reader is referred to several excellent
reviews [34, 4954] and Chapter 4 of the current volume. Some
groups have generously made equilibrated conformations of some
lipid bilayer systems freely available (see Note 4), which provide a
good starting point for the main procedure we outline below.
Sometimes it will be necessary to generate a new lipid bilayer
system from scratch and one then needs a measure of how stable/
good that pure system is before proceeding to insert a protein into
it. The most commonly used measure of equilibration/stability is
to analyze the mean area per lipid; a quantity for which there
frequently exists experimental data for which to make a direct comparison to. Furthermore, if this is incorrect then it is likely that
most other properties will also be inaccurate [17].
3.3 Setup
of the Protein
in the Membrane
We focus here on methods currently used in our laboratory but it

is worthwhile briefly mentioning alternative methods. Earlier setup
methods were developed with the limitations imposed by available
computer power at the time. The approach adopted by Woolf and
Roux was to build up lipids around the protein, by placing isolated
lipid molecules randomly selected from a library of 2000 conformations. The system was adjusted to remove as many overlaps as
possible followed by a constrained minimization procedure [55, 56].
Although one does not start from a preformed lipid-bilayer in this
case, the conformations in the library will be derived from a simulation of pure lipids.
A different approach was proposed by Faraldo-Gmez and
colleagues [57], who used preformed lipid bilayer as the starting
point. Their method relies on creating a cavity in pre-equilibrated
lipid bilayer using the solvent-accessible surface area of the protein
as a template. Lipid molecules whose head groups fall within the
96
volume are removed whilst remaining lipids are subjected to an

ever-increasing force acting perpendicular to the surface of the cavity template until the cavity is empty. The protein itself can then
simply be inserted into the cavity. This method has the advantage
that a preexisting lipid bilayer can be used and in such a way that
non-interfacial lipid molecules are not significantly perturbed,
which results in a faster equilibration time. The approach was
recently updated to incorporate chemical specificity at the protein
lipid interface, with additional features that enable it to handle
complex membrane protein topologies [58]. A key advantage of
the new method, named GRIFFIN (GRId-based Force Field
INput), is that it is a stand-alone tool, enabling it to work alongside
any existing molecular simulation package.
More recently, Wompil Ims group have developed a Webbased membrane-builder, which is tightly coupled to the CHARMM
force-field [59]. The interface offers great flexibility and provides
users with the option to build either a pure lipid membrane or proteinlipid mixture. The user can also select between two insertion
protocols. A similar approach by the same group has been applied
to aid researchers build micelles [60].
Another approach within the GROMACS simulation suite, is
the so-called g_membed program, which allows the efficient insertion of a membrane protein into a membrane with minimal perturbation [61]. In this procedure the protein (positioned in an
equilibrated bilayer) is first shrunk in the xy plane and overlapping
lipids are then removed. The protein is then grown back within a
short MD simulation pushing the lipids away as it does so. The
result is a system that is relatively unperturbed and thus should
require shorter equilibration times. Also developed to work with
GROMACS, is the pair of packages LAMBADA and InflateGRO2
[62], which work together to generate a proteinlipid system in an
efficient and automated manner.
With the advent of more powerful computers, a more direct
approach to the setup has become possible and we will focus on
this. We have implemented this procedure using GROMACS [63]
in combination with VMD [40], but there is also a plugin available
http://www.ks.uiuc.edu/Research/vmd/plugins/membrane/
to do all the whole process in VMD. This plugin however is more
useful if you intend to use NAMD [64] as your molecular dynamics program. Both methodologies exploit the fact that enough
simulation time is available to adequately equilibrate the system.
The procedure is quite interactive and can be summarized by the
following steps:
1. Obtain a pre-equilibrated lipid bilayer (see Note 4).
2. Align protein in the lipid bilayer.
3. Remove overlapping lipid molecules.
4. Equilibrate new system.
97
Fig. 1 (a) shows the protein BtuB (dark molecular surface), embedded in the bilayer after the removal of overlapping lipids (only protein and lipid are shown in this figure for clarity). Lipid atoms are shown as van der
Waals spheres. During the course of the equilibration phase, lipid molecules will move in around the protein as
shown in (b) which is an equilibrated system
The alignment of the protein with the pre-equilibrated lipid

bilayer in this process is essentially something that is done by eye.
As mentioned in the introduction, some guidance is afforded by
the presence of the tryptophan/tyrosine girdles that are associated
with membrane proteins. However, a more objective approach is
to make use of the Orientations of Proteins in Membranes (OPM)
database [21, 65]. OPM contains coordinates of transmembrane
and peripheral proteins and peptides of known structure and their
predicted spatial positions with respect to a lipid bilayer (http://
opm.phar.umich.edu), based on the Positioning of Proteins in
Membrane (PPM) algorithm, which optimizes the transfer energy
from water to the membrane. The PPM server (http://opm.phar.
umich.edu/server.php) may also be used to position new membrane protein structures [65]. Removal of whole overlapping lipid
molecules means that the resulting system will have a vacuum in
between the protein and the lipid molecule (see Fig. 1a). During
the first stage of the equilibration this will be removed as the lipid
molecules relax around the protein. Typically for this stage, it is
important to keep the protein conformation as close to the starting
coordinates as possible. Thus, it is common for positional restraints
to be imposed on the protein atoms (or a subset thereof) during
this stage. An NPT ensemble (see Note 5) MD simulation is then
performed to allow the lipid molecules to equilibrate around the
newly inserted membrane protein (Fig. 1b). During this stage water
may penetrate slightly into the vacuum between lipids and the protein. These will be expelled during the course of the simulation as
98
the lipids move towards the protein and the system equilibrates.
The length of this equilibration phase is usually determined by
monitoring the area per lipid as a function of time. After a period
of time (typically between 1 and 3 ns for systems with 512 lipids)
one should see this plateau off. This value can be checked against
experimental data, although this can be difficult to come by for
exactly the same system. Before unconstrained production or further dynamics can be performed it is best to allow the protein to
relax in stages. There are many different approaches reported in
the literature, which can appear to be rather subjective, but the
underlying philosophy is to work back from the backbone of the
protein (see Note 6).
3.4 An Alternative
CoarseGrained Method
A more recent approach has been to simulate the whole system

de novo from a random arrangement of molecules in the system.
Such an approach is made possible via the use of the coarse-grained
(CG) methods [16] where small groups of atoms (typically 4) are
treated as single particles. Because of the associated reduction in
the number of particles, much longer time and length scales can be
addressed. This presents the opportunity to simulate large-scale
changes in proteinlipid interactions, such as membrane protein
bilayer insertion. Thus, the user can, at that point, decide whether
the problem requires a switch to a fully atomistic description or
whether the CG description is adequate. There are currently ongoing efforts to integrate multi-scale methods into a self-consistent representation [6669], including hybrid Martini-based
approaches [70, 71]. A full discussion of these methods is not possible here, but it seems likely that these methods offer the greatest
potential in terms of flexibility across time and length scales for
membrane systems.
CG simulations have previously proven useful in modelling the
dynamics of lipids and detergents [7277], DNA [78], proteins
[79], and toy peptides [16, 80, 81]. A semiquantitative model
for lipid systems was devised based on thermodynamic data as well
as structural and dynamic properties of atomistic simulations [82].
This model was adapted for application to membrane proteins [83]
and similar models have since been developed for related systems by
several groups [84]. Marrink and coworkers continue to refine their
Martini library of CG parameters for proteins, lipids, and other biomolecules [85], along with Martini-compatible models for polarizable CG water particles [86]. Additionally, improved parameters
for more realistic peptide dynamics [87] have also been introduced.
A common feature of the basic CG models is that instead of representing every atom in a protein or lipid molecule, approximately
four atoms are grouped together into one particle, and are parameterized to capture the hydrophobicity/hydrophilicity, charge and
hydrogen-bonding properties of their constituent atoms. Bonds,
angles, and the overall backbone secondary/tertiary structure
99
Fig. 2 Illustration of the how the atomistic model translates into the coarse-grained model for the KcsA
potassium channel. Aromatic particles are shown as black van der Waals spheres, hydrophobic or backbone
particles are shown in light grey, and polar/charged particles are shown in dark grey
are preserved through either soft harmonic potentials between

hydrogen-bonding atoms, through weak dihedral angles, or for
larger proteins, elastic network model approaches.
The initial CG model for a protein is normally derived by
extracting the coordinates for all C atoms and selecting side-chain
atoms from the corresponding all-atom protein file. The overall
shape and surface area of a lipid or protein molecule is preserved in
the model (Fig. 2). Subsequently, lipid molecules taken from a
library (derived from for example a pure lipid simulation) are
randomly placed in a box containing the protein, before solvation
with a pre-equilibrated box of water particles, and neutralizing
ions (Fig. 3a). A number of factors must be considered when solvating the protein. Firstly, as with atomistic models, the number of
lipid molecules should be such that the bilayer formed is sufficiently large enough to allow plenty of space between the embedded protein and its periodic image within the membrane plane
(see Note 3). This number may be estimated by considering the
equilibrium area per lipid of interest. Secondly, a ratio of CG water
particles to lipid molecules of ~1025 should be used to favor the
formation of a bilayer, rather than for example hexagonal, micellar,
or other nonlamellar phases [82].
Once the system has been prepared, a short energy minimization procedure is carried out, before one or more production runs
are performed. A typical such CG simulation will normally be
about 2 orders of magnitude faster than the corresponding atomistic simulation, as a consequence of the reduced number of
particles, softer potentials (and thus longer MD integration timestep),
100
Fig. 3 (a) Shows the random starting configuration of the coarse-grained simulation of KcsA with dipalmitoylphosphatidylcholine (DPPC). KcsA is drawn as a black backbone trace. Lipid acyl chain particles are drawn as
light grey van der Waals spheres, glycerol backbone particles are shown in dark grey, and lipid headgroups
(including the phosphates) are drawn as black spheres. Water molecules are not shown for clarity. (b) is the
configuration after 200 ns, which clearly shows that the system has evolved into a bilayer arrangement with
KcsA embedded within it
and consideration of only short-range non-bonded interactions.

Our experience with ~40 different membrane proteins suggests
that a period of ~0.10.2 s is normally sufficient to allow the selfassembly of a phospholipid bilayer around the protein of interest.
For a typical system of ~6,000 CG particles, this translates to a
CPU time of ~1 day on a typical workstation computer [83]. The
simulation proceeds via an initial assembly of lipids into a continuous lamellar phase, with a lipid stalk which bridges between the
bilayer and its periodic image. Eventually this stalk is broken to
form a defect-free bilayer with the membrane protein correctly
inserted (Fig. 3b).
This CG method for bilayer insertion has been successfully
applied to a number of membrane proteins [88], and shown to
agree in terms of lipidprotein interactions with extended atomistic simulations of an eight-stranded -barrel, OmpA, a transmembrane -helical dimer, Glycophorin A [83], and a 12-transmembrane
-helix bundle, LacY [83]. It has also been extensively tested
against a number of -helical membrane peptides and proteins and
shown to be in good agreement with experimental data in terms of
orientation within the bilayer [83]. Recent successful applications
have been reported for a wide range of membrane proteins using
the Martini model [89], including large-scale assembly of multiple
proteins in membranes, such as for G-protein coupled receptors
(GPCRs) [90, 91], and even realistic membrane protein structural
101
transitions, such as mechanosensitive channels [9294]. Moreover,

several groups continue to expand the applicability of CG models, by
introducing biologically realistic lipid/protein membrane mixtures
[95] that are often crowded [9698] and sometimes feature functionally important raft assemblies or domains [49, 99].
There is now also a database of all membrane proteins [100]
that have been pumped through this procedure, which provides a
good initial start point for anyone wanting to begin a membrane
protein simulation. The primary advantage of the CG approach to
membrane insertion is the elimination of the need for user input,
e.g., using the aromatic girdles to guide placement. This is particularly useful for proteins that might be tilted with respect to the
bilayer normal, such as the Vpu -helical fragment from HIV-1
[83]. This is also true of proteins which are nonuniform in their
transmembrane distribution, e.g., the coat protein from fd phage
which contains an amphipathic in-plane helix thought to reside at
the membranewater interface [83]; monotopic proteins which sit
on the surface of the bilayer; or proteins with large extracellular
regions such as the multi-domain ABC transporter family. Finally,
self-assembly simulations of complex, atypical membrane proteins
such as the highly charged voltage sensor domain from a potassium
channel reveal that considerable local bilayer deformation may be
necessary for insertion, rather than a bilayer of fixed and uniform
thickness surrounding the protein [83].
Once the protein has stably inserted within the membrane, it
may be desirable to convert from the CG representation back to an
atomistic level of detail. The most obvious method for achieving
this is to use the CG results as a rough guide for positioning the
atomistic protein into a bilayer via the method detailed above
involving placement into a pre-equilibrated bilayer before removal
of overlapping lipid molecules. For example, the peaks in densities
of the headgroups along the membrane normal in the atomistic
preformed bilayer may be matched up with the CG system, before
least-squares fitting the atomistic protein C atoms onto the backbone of the CG model, or alternatively using homology modelling
techniques. For more complex proteins, such as those which are
significantly tilted or which induce local bilayer deformation, a
more direct matching of atomistic and CG coordinates may be
necessary. Recently, a method has been developed [101] whereby
atomistic lipid structures from a pure lipid simulation library are
iteratively least-squares fitted onto their CG lipid counterparts,
with each of the best matching molecules being retained. Thus, an
atomistic bilayer resembling the CG system is gradually built
up around the atomistic protein model, which is again obtained
by fitting the C atoms onto the backbone of the CG protein.
Alternatively, Marrink and coworkers presented a promising
approach in which atomistic and CG representations of a system
102
are initially coupled via harmonic restraints [102]. Following a

simulated annealing protocol, the coupling is gradually removed to
achieve the final, relaxed atomistic model.
3.5 Running
the Simulation
The last step is to actually run your atomistic simulation. The primary
emphasis has been on using parameters and ensembles that best
reproduce the properties of lipid bilayers in the absence of proteins. A full review of these considerations is beyond the scope of
this chapter, but the interested reader is referred to several articles
that discuss sources of error and the best choice of parameters in
membrane simulations [17, 103105].
There are many properties that one could check in the simulation, but probably the most useful is the area per lipid, which gives
an indication of molecular packing and the membrane fluidity. It is
also a property that is sensitive to simulation set up whilst also
being a reasonably reliable indicator that other properties will
also be correct. It is important to remember here what your question islarge undulations across large membrane patches will
require much longer simulation time than a study of waterheadgroup interactions for example.
Finally, there are practical considerations such as disk-space
and storage of very large trajectories (see Note 7), a problem that
is presumably going to parallel the increase in computer power.
Conclusions
We have discussed two approaches that can be used to set up and
perform molecular dynamics simulations of membrane proteins.
The advantage of the first atomistic approach is that it is easy to use
and generally applicable. A disadvantage of this approach is that to
some extent it depends on a subjective positioning of the protein
within the bilayer in terms of its overall tilt and its disposition along
the bilayer normal. The second approach, via the use of coarsegrain methodologies, allows one to circumvent these problems.
The combination of both of these methodologies allows one
to explore a wide range of time and length scales with respect to
membrane proteins and should provide valuable information on
their structure and function.
Notes
1. There is also an online server version of the What-If program
(http://swift.cmbi.ru.nl/servers/html/index.html) that provides useful tools features to rebuild missing atoms in sidechains. Stereochemical checking tools are also available at this
site (useful if you are starting from a model). Similarly, online
servers now exist for pKa calculations of ionisable side-chains,
103
Table 1
Lipid configurations available for download
PI
URL
Lipids
Scott Feller
http://www.lipid.wabash.edu/
POPC,DOPC,
DPPC, SDPC
Helmut
Heller
http://heller.userweb.mwn.de/membrane/membrane.html
POPC
Wonpil Im
http://www.charmm-gui.org/?doc = input/membrane
Many
combinations
possible
Mikko
Karttunen
http://www.softsimu.net/downloads.shtml
DMTAP, DMP,
DPPC
Peter
Tieleman
http://wcm.ucalgary.ca/tieleman/downloads
DPC micelles,
POPC, DMPC,
DPPC, PLPC
Alexander
Lyubartsev
http://people.su.se/~jjm/Stockholm_Lipids/Downloads.html
Various
Stockholm
lipids
Jochen Hub
http://cmb.bio.uni-goettingen.de/downloads.html
Lipid patches
with cholesterol
Oliver
Beckstein
http://lipidbook.bioch.ox.ac.uk/
Various
including H++ (http://biophysics.cs.vt.edu/H++) and PROPKA

(http://propka.ki.ku.dk/).
2. Typical capping groups are an acetyl on the N-terminus (to
give CH3-CO-NH2protein or amidation at the C-terminus
to give protein-CO-NH2. These can be added with a
molecule-building program such as Pymol [22]. These additional groups are either treated as separated residues (as is
case for GROMACS [63]) or as patches (as is the case for
CHARMM [106]).
3. In periodic systems (nearly all lipid bilayer simulations will be
periodic), it is important to make sure that the parts of the
protein that are not in the bilayer are adequately solvated to
ensure that the protein near one edge of the box does not
see itself in the nearest periodic image. To avoid such problems, we have typically set up the system such that there is a
10 of water between the protein and its nearest box edge.
4. Some groups have made freely available their coordinates of preequilibrated lipid bilayers and these provide a useful start point.
Some that are available at the time of writing are summarized in
Table 1. Although some of these sites will also contain various
parameter sets for download, Beckstein and colleagues [107]
104
have developed an online database for different lipid types and

different force-fields. It is available at http://lipidbook.bioch.
ox.ac.uk.
5. NPT refers to the thermodynamic ensemble used. In this case,
constant number of particles (N), constant pressure (P) and
constant temperature (T). This allows the volume of the system
to change and hence the surface area of the lipid which can
then be compared back to experiment as a measure of simulation quality.
6. Typically, the protein is relaxed in steps. For example there may
be a period during which backbone atoms only are constrained,
followed by just C atoms, following by no constraints during
the production phase of the simulation.
7. An issue that requires constant revisiting is how often one
writes simulation frames to the trajectory file. The problem is
compounded by two factors: (a) the ever increasing size of
system that can be reasonably addressed (currently routinely
up to 200,000 atoms) and (b) the length of simulation time
(of the order of tens of nanoseconds). A reasonable value for
atomistic simulations is to write to disk every 5 ps, but again
this will depend on what question you are trying to address.
Acknowledgements
We thank the Leverhulme Trust and Unilever for support and
Dr Jorge Pikunic for the BtuB coordinates and useful discussions.
References
1. Wallin E, von Heijne G (1998) Genome-wide
analysis of integral membrane proteins from
eubacterial, archean, and eukaryotic organisms. Protein Sci 7:10291038
2. Terstappen GC, Reggiani A (2001) In silico
research in drug discovery. Trends Pharmacol
Sci 22:2326
3. Lemieux MJ, Huang Y, Wang DN (2004)
The structural basis of substrate translocation
by the Escherichia coli glycerol-3-phosphate
transporter: a member of the major facilitator superfamily. Curr Opin Struct Biol 14:
405412
4. Guan L, Kaback HR (2006) Lessons from
lactose permease. Annu Rev Biophys Biomol
Struct 35:6791
5. Gether U, Andersen PH, Larsson OM,
Schousboe A (2006) Neurotransmitter transporters: molecular function of important
drug targets. Trends Pharmacol Sci 27:
375383
6. Ash WL, Zlomislic MR, Oloo EO, Tieleman

DP (2004) Computer simulations of membrane proteins. Biochem Biophys Acta
1666:158189
7. Sperotto MM, May S, Baumgaertner A
(2006) Modelling of proteins in membranes.
Chem Phys Lipids 141:229
8. Beckstein O, Biggin PC, Bond P, Bright JN,
Domene C, Grottesi A et al (2003) Ion channel gating: insights via molecular simulations.
FEBS Lett 555:8590
9. Gumbart J, Wang Y, Aksimentiev A,
Tajkhorshid E, Schulten K (2005) Molecular
dynamics simulations of proteins in lipid bilayers. Curr Opin Struct Biol 15:423431
10. Roux B (2005) Ion conduction and selectivity
in K(+) channels. Annu Rev Biophys Biomol
Struct 34:153171
11. Bond PJ, Sansom MSP (2004) The simulation approach to bacterial outer membrane
proteins. Mol Memb Biol 21:151161

12. Beckstein O, Biggin PC, Sansom MSP (2001)
A hydrophobic gating mechanism for nanopores. J Phys Chem B 105:1290212905
13. Beckstein O, Sansom MSP (2006) A hydrophobic gate in an ion channel: the closed state
of the nicotinic acetylcholine receptor. Phys
Biol 3:147159
14. Arinaminpathy Y, Biggin PC, Shrivastava IH,
Sansom MSP (2003) A prokaryotic glutamate
receptor: homology modelling and molecular
dynamics simulations of GluR0. FEBS Lett
553:321327
15. Ohkubo YZ, Pogorelov TV, Arcario MJ,
Christensen GA, Tajkhorshid E (2012) Accelerating membrane insertion of peripheral
proteins with a novel membrane mimetic
model. Biophys J 102(9):21302139
16. Nielsen SO, Lopez CF, Ivanov I, Moore PB,
Shelley JC, Klein ML (2004) Transmembrane
peptide-induced lipid sorting and mechanism
of L-to-inverted phase transition using
coarse-grain molecular dynamics. Biophys J
87(4): 21072115
17. Anzo C, de Vries AH, Hoeltje H-D,
Tieleman
DP,
Marrink
SJ
(2003)
Methodological issues in lipid bilayer simulations. J Phys Chem B 107:94249433
18. Ulmschneider MB, Sansom MSP, Di Nola A
(2005) Properties of integral membrane protein structures: derivation of an implicit membrane potential. Proteins 59:252265
19. Basyn F, Charloteaux B, Thomas A, Brasseur
R (2001) Prediction of membrane protein
orientation in lipid bilayers: a theoretical
approach. J Mol Graph Model 20:235244
20. Tusnady GE, Dosztanyi Z, Simon I (2005)
PDB_TM: selection and membrane localization of transmembrane proteins in the protein data bank. Nucleic Acids Res 33:
D275D278
21. Lomize MA, Lomize AL, Pogozheva ID,
Mosberg HI (2006) OPM: orientations of
protetins
in
membranes
database.
Bioinformatics 22:623625
22. DeLano WL (2004) The PyMOL molecular
graphics system. DeLano Scientific LLC, San
Carlos, CA
23. Vriend G (1990) A molecular modelling and
drug design program. J Mol Graph 8:5256
24. Fiser A, Sali A (2003) Modeller: generation
and refinement of homology-based protein
structure
models.
Methods
Enzymol
374:461491
25. Eswar N, Webb B, Marti-Renom MA,
Madhusudhan MS, Eramian D, Shen M-Y,
et al. (2006) Comparative protein structure
modeling using Modeller. Curr Protoc
Bioinformatics, Chapter 5:Unit 5.6
26. Feller SE, MacKerell AD (2000) An improved
empirical potential energy function for molecu-
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
105
lar simulations of phospholipids. J Phys Chem

B 104(31):75107515
Klauda JB, Brooks BR, MacKerell AD Jr,
Venable RM, Pastor RW (2005) An ab initio
study on the torsional surface of alkanes and
its effect on molecular simulations of alkanes
and a DPPC bilayer. J Phys Chem B 109(11):
53005311
Chandrasekhar I, Kastenholz M, Lins RD,
Oostenbrink C, Schuler LD, Tieleman DP
et al (2003) A consistent potential energy
parameter set for lipids: dipalmitoylphosphatidylcholine as a benchmark of the GROMOS96
45A3 force field. Eur Biophys J 32(1):6777
Klauda JB, Venable RM, Freites JA, OConnor
JW, Tobias DJ, Mondragon-Ramirez C et al
(2010) Update of the CHARMM all-atom
additive force field for lipids: validation on
six lipid types. J Phys Chem B 114(23):
78307843
Schuler LD, Daura X, van Gunsteren WF
(2001) An improved GROMOS96 force field
for aliphatic hydrocarbons in the condensed
phase. J Comput Chem 22(11):12051218
Berger O, Edholm O, Jahnig F (1997) Molecular dynamics simulations of a fluid bilayer of
dipalmitoylphosphatidycholine at full hydration, constant pressure and constant temperature. Biophys J 72:20022013
Jojart B, Martinek TA (2007) Performance of
the general amber force field in modeling
aqueous POPC membrane bilayers. J Comput
Chem 28(12):20512058
Poger D, Van Gunsteren WF, Mark AE
(2010) A new force field for simulating phosphatidylcholine bilayers. J Comput Chem
31(6): 11171125
Lyubartsev AP, Rabinovich AL (2011) Recent
development in computer simulations of lipid
bilayers. Soft Matter 7:2539
Piggot TJ, Pineiro A, Khalid S (2012)
Molecular dynamics simulations of phosphatidylcholine membranes: a comparative force
field study. J Chem Theory Comput 8:
45934609
Li H, Robertson AD, Jensen JH (2005) Very
fast empirical prediction and interpretation of
protein pKa values. Proteins 61:704721
Olsson MHM, Sondergaard CR, Rostkowski
M, Jensen JH (2011) PROPKA3: consistent
treatment of internal and surface residues
in empirical pK(a) predictions. J Chem
Gordon JC, Myers JB, Folta T, Shoja V,
Heath LS, Onufriev A (2005) H++: a server
for estimating pKas and adding missing
hydrogens to macromolecules. Nucleic Acids
Res 33: W368W371
Anandakrishnan R, Aguilar B, Onufriev AV
(2012) H++ 3.0: automating pK prediction
106
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.

and the preparation of biomolecular structures for atomistic molecular modeling and
simulations. Nucleic Acids Res 40(Web Server
issue), W537W541.
Humphrey W, Dalke A, Schulten K (1996)
VMDvisual molecular dynamics. J Mol
Graph 14:3338
Rostkowski M, Olsson MH, Sondergaard
CR, Jensen JH (2011) Graphical analysis of
pH-dependent properties of proteins predicted using PROPKA. BMC Struct Biol 11:6
Luzhkov VB, qvist J (2000) A computational study of ion binding and protonation
states in the KcsA potassium channel. Biochim
Biophys Acta 1481:360370
Ranatunga KM, Shrivastava IH, Smith GR,
Sansom MSP (2001) Side-chain ionization
states in a potassium channel. Biophys J 80:
12101219
Bernche S, Roux B (2002) The ionizastion
state and the conformation of Glu-71 in the
KcsA K(+) channel. Biophys J 82:772780
Cordero-Morales JF, Cuello LG, Zhao Y,
Jogini V, Cortes DM, Roux B et al (2006)
Molecular determinants of gating at the
potassium channel selectivity filter. Nat Struct
Biol 13:319322
Arinaminpathy Y, Sansom MSP, Biggin PC
(2002) Molecular dynamics simulations of
the ligand-binding domain of the ionotropic
glutamate receptor GluR2. Biophys J
82:676683
Arinaminpathy Y, Sansom MSP, Biggin PC
(2006) Binding site flexibility: molecular simulation of partial and full agonists within a
glutamate receptor. Mol Pharm 69:1118
Kaye LS, Sansom MSP, Biggin PC (2006)
Molecular dynamics simulations of an NMDA
receptor. J Biol Chem 281:1273612742
Bennett WF, Tieleman DP (2013) Computer
simulations of lipid membrane domains.
Biochim Biophys Acta 1828(8):17651776
Berkowitz ML (2009) Detailed molecular
dynamics simulations of model biological
membranes containing cholesterol. Biochim
Biophys Acta Biomembr 1788(1):8696
Feller SE (2000) Molecular dynamics simulations of lipid bilayers. Curr Opinion Coll
Interface Sci 5:217223
Mouritsen OG, Jorgensen K (1997) Smallscale lipid-membrane structure: simulation
versus experiment. Curr Opin Struct Biol
7:518527
Scott HL (2002) Modeling the lipid component of membranes. Curr Opin Struct Biol
12:495502
Tieleman DP, Marrink SJ, Berendsen HJC
(1997) A computer perspective of membranes: molecular dynamics studies of lipid
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
bilayer systems. Biochim Biophys Acta

1331:235270
Belohorcova K, Davis JH, Woolf TB, Roux B
(1997) Structure and dynamics of an amphiphilic peptide in a lipid bilayer: a molecular
dynamics study. Biophys J 73:30393055
Woolf TB, Roux B (1996) Structure, energetics, and dynamics of lipid-protein interactionsa molecular-dynamics study of the
gramicidin-A channel in a DMPC bilayer.
Proteins 24: 92114
Faraldo-Gmez JD, Smith GR, Sansom MSP
(2002) Setting up and optimization of membrane protein simulations. Eur Biophys J 31:
217227
Staritzbichler R, Anselmi C, Forrest LR,
Faraldo-Gomez JD (2011) GRIFFIN: A versatile methodology for optimization of proteinlipid interfaces for membrane protein
simulations. J Chem Theory Comput 7:
11671176
Jo S, Kim T, Im W (2007) Automated builder
and database of protein/membrane complexes for molecular dynamics simulations.
PLoS ONE 2(9):e880
Cheng X, Jo S, Lee HS, Klauda JB, Im W
(2013) CHARMM-gui micelle builder for
pure/mixed micelle and protein/micelle complex
systems.
J Chem Inf Model 53(8):21712180
Wolf MG, Hoefling M, Aponte-Santamara C,
Grubmller H, Groenhof G (2010) g_membed: Efficient insertion of a membrane protein
into an equilibrated lipid bilayer with minimal
perturbation.
J
Comput
Chem
31(11):21692174
Schmidt TH, Kandt C (2012) LAMBADA
and InflateGRO2: Efficient membrane alignment and insertion of membrane proteins for
molecular dynamics simulations. J Chem Inf
Model 52(10):26572669
Lindahl E, Hess B, van der Spoel D (2001)
GROMACS 3.0: A package for molecular
simulation and trajectory analysis. J Mol
Model 7:306317
Phillips JC, Braun R, Wang W, Gumbart J,
Tajkhorshid E, Villa E et al (2005) Scalable
molecular dynamics with NAMD. J Comput
Chem 26:17811802
Lomize MA, Pogozheva ID, Joo H, Mosberg
HI, Lomize AL (2012) OPM database and
PPM web server: resources for positioning of
proteins in membranes. Nucleic Acids Res
40(Database issue):D370D376
Christen M, Van Gunsteren WF (2006)
Multigraining: An algorithm for simultaneous
fine-grained and coarse-grained simulation
of molecular systems. J Chem Phys 124:
154106.1154106.7

67. Chang R, Ayton GS, Voth GA (2005)
Multiscale coupling of mesoscopic and
atomistic-level lipid bilayer simulations. J
Chem Phys 122:244716
68. Shi Q, Izvekov S, Voth GA (2006) Mixed
atomistic and coarse-grained molecular
dynamics: simulation of a membrane-bound
ion channel. J Phys Chem B 110:
1504515048
69. Orsi M, Noro MG, Essex JW (2011) Dualresolution molecular dynamics simulation of
antimicrobials in biomembranes. J R Soc Interface 8(59):826841
70. Rzepiela AJ, Louhivuori M, Peter C, Marrink
SJ (2011) Hybrid simulations: combining
atomistic and coarse-grained force fields using
virtual sites. Phys Chem Chem Phys 13(22):
1043710448
71. Wassenaar TA, Ingolfsson HI, Priess M,
Marrink SJ, Schafer LV (2013) Mixing
MARTINI: electrostatic coupling in hybrid
atomistic-coarse-grained biomolecular simulations. J Phys Chem B 117(13):
35163530
72. Goetz R, Lipowsky R (1998) Computer simulations of bilayer membranes: self-assembly
and interfacial tension. J Chem Phys 108(17):
73977409
73. Murtola T, Falck E, Patra M, Karttunen M,
Vattulainen I (2004) Coarse-grained model
for phospholipid/cholesterol bilayer. J Chem
Phys 121:91569165
74. Shelley JC, Shelley MY, Reeder RC,
Bandyopadhyay S, Moore PB, Klein ML (2001)
Simulations of phospholipids using a coarse
grain model. J Phys Chem B 105(40):
97859792
75. Smit B, Hilbers PAJ, Esselink K, Rupert
LAM, van Os NM, Schlijper AG (1990)
Computer simulations of a water/oil interface
in the presence of micelles. Nature
348(6302):624625
76. Stevens MJ, Hoh JH, Woolf TB (2003)
Insights into the molecular mechanism of
membrane fusion from simulation: evidence
for the association of splayed tails. Phys Rev
Lett 91(18):188102
77. Whitehead L, Edge CM, Essex JW (2001)
Molecular dynamics simulation of the hydrocarbon region of a biomembrane using a
reduced representation model. J Comput
Chem 22:16221633
78. Tepper HL, Voth GA (2005) A coarsegrained model for double-helix molecules in
solution: spontaneous helix formation and
equilibrium properties. J Chem Phys
122:124906
79. Tozzini V (2005) Coarse-grained models
for proteins. Curr Opin Struct Biol 15(2):
144150
107
80. Lopez CF, Nielsen SO, Moore PB, Klein ML

(2004) Understanding natures design for a
nanosyringe. Proc Natl Acad Sci U S A 101
(13):44314434
81. Venturoli M, Smit B, Sperotto MM (2005)
Simulation studies of protein-induced bilayer
deformations, and lipid-induced protein tilting, on a mesoscopic model for lipid bilayers
with embedded proteins. Biophys J 88(3):
17781798
82. Marrink SJ, de Vries AH, Mark AE (2004)
Coarse-grained model for semiquantitative
lipid simulations. J Phys Chem 108:750760
83. Bond PJ, Sansom MSP (2006) Insertion and
assembly of membrane proteins via simulation. J Am Chem Soc 128:26972704
84. Shih AY, Arkhipov A, Freddolino PL,
Schulten K (2006) Coarse grained proteinlipid model with application to lipoprotein
particles. J Phys Chem B 110(8):36743684
85. Marrink SJ, Tieleman DP (2013) Perspective
on the Martini model. Chem Soc Rev 42(16):
68016822
86. Yesylevskyy SO, Schafer LV, Sengupta D,
Marrink SJ (2010) Polarizable water model
for the coarse-grained MARTINI force field.
PLoS Comput Biol 6(6):e1000810
87. Seo M, Rauscher S, Pomes R, Tieleman DP
(2012) Improving internal peptide dynamics
in the coarse-grained MARTINI model:
toward large-scale simulations of amyloidand Elastin-like peptides. J Chem Theory
Comput 8(5): 17741785
88. Scott KA, Bond PJ, Ivetac A, Chetwynd AP,
Khalid S, Sansom MSP (2008) Coarsegrained MD simulations of membrane
protein-bilayer
self-assembly.
Structure
16(4):621630
89. Sengupta D, Marrink SJ (2010) Lipidmediated interactions tune the association of
glycophorin A helix and its disruptive mutants
in membranes. Phys Chem Chem Phys
12(40): 1298712996
90. Periole X, Huber T, Marrink SJ, Sakmar TP
(2007) G protein-coupled receptors selfassemble in dynamics simulations of model
bilayers. J Am Chem Soc 129(33):
1012610132
91. Periole X, Knepp AM, Sakmar TP, Marrink
SJ, Huber T (2012) Structural determinants
of the supramolecular organization of G
protein-coupled receptors in bilayers. J Am
Chem Soc 134(26):1095910965
92. Deplazes E, Louhivuori M, Jayatilaka D,
Marrink SJ, Corry B (2012) Structural investigation of MscL gating using experimental
data and coarse grained MD simulations.
PLoS Comput Biol 8(9):e1002683
93. Louhivuori M, Risselada HJ, van der Giessen
E, Marrink SJ (2010) Release of content
108
94.
95.
96.
97.
98.
99.
100.

through mechano-sensitive gates in pressurized liposomes. Proc Natl Acad Sci U S A
107(46): 1985619860.
Samuli Ollila OH, Louhivuori M, Marrink SJ,
Vattulainen I (2011) Protein shape change
has a major effect on the gating energy of a
mechanosensitive channel. Biophys J 100(7):
16511659
Holdbrook DA, Leung YM, Piggot TJ,
Marius P, Williamson PT, Khalid S (2010)
Stability and membrane orientation of the
fukutin transmembrane domain: a combined
multiscale molecular dynamics and circular
dichroism study. Biochemistry 49(51):
1079610802
Domanski J, Marrink SJ, Schafer LV (2012)
Transmembrane helices can induce domain
formation in crowded model membranes.
Goose JE, Sansom MS (2013) Reduced
lateral mobility of lipids and proteins in
crowded membranes. PLoS Comput Biol
9(4):e1003033
Javanainen M, Hammaren H, Monticelli L,
Jeon JH, Miettinen MS, Martinez-Seara H
et al (2013) Anomalous and normal diffusion
of proteins and lipids in crowded lipid membranes. Faraday Discuss 161:397417, discussion 419-59
Parton DL, Tek A, Baaden M, Sansom MS
(2013) Formation of raft-like assemblies
within clusters of influenza hemagglutinin
observed by MD simulations. PLoS Comput
Biol 9(4):e1003034
Chetwynd AP, Scott KA, Mokrab Y, Sansom
MSP (2008) CGDB: A database of membrane protein/lipid interactions by coarse-
101.
102.
103.
104.
105.
106.
107.
grained molecular dynamics simulations. Mol

Memb Biol 25(8):662669
Stansfeld PJ, Sansom MSP (2011) From
coarse grained to atomistic: a serial multiscale
approach to membrane protein simulations.
Rzepiela AJ, Schafer LV, Goga N, Risselada
HJ, De Vries AH, Marrink SJ (2010)
Reconstruction of atomistic details from
coarse-grained structures. J Comput Chem
31(6):13331343
Patra M, Karttunen M, Hyvnen MT, Falck
E, Lindqvist P, Vattulainen I (2003) Molecular
dynamics simulations of lipid bilayers: major
artifacts due to truncating electrostatic interactions. Biophys J 84:36363645
Patra M, Karttunen M, Hyvnen MT, Falck
E, Vattulainen I (2004) Lipid bilayers driven
to a wrong lane in molecular dynamics simulations by subtle changes in long-range electrostatic interactions. J Phys Chem B 108:
44854494
de Vries AH, Chandraskhar I, van Gunsteren
WF, Hunenberger PH (2005) Molecular
dynamics simulations of phospholipid bilayers: influence of artificial periodicity, system
size, and simulation time. J Phys Chem B
109:1164311652
MacKerrell AD, Bashford D, Bellott M,
Dunbrack RL, Evanseck JD, Field MJ et al
(1998) All-atom empirical potential for
molecular modeling and dynamics studies of
proteins. J Phys Chem B 102:35863616
Domaski J, Stansfeld P, Sansom MP, Beckstein
O (2010) Lipidbook: a public repository for
force-field parameters used in membrane simulations. J Membr Biol 236(3):255258
Chapter 6
Membrane-Associated Proteins andPeptides
MarcF.Lensink
Abstract
This chapter discusses the practical aspects of setting up molecular dynamics simulations of membrane-
associated proteins and peptides, and the analysis thereof. Topology files for selected lipids are provided
and selected analysis tools presented. These include tools for the creation of lipid bilayers of mixed lipid
content (DOPE) and easy extraction of lipid coordinates (g_zcoor, g_xycoor), the calculation of helical
axes (g_helixaxis) and aromatic order parameters (g_arom), the determination of peptide- or protein-
interacting lipids (g_under), and the investigation of lipid-specific interactions through the calculation of
lipid-bridged residueresidue contacts (g_prolip).
Key words Molecular dynamics, Lipid bilayer, Membrane, Peptidelipid interaction, Phospholipid,
Cholesterol, GROMACS, Helix axis, DOPE, Lipid order parameter, Specific interaction
1 Introduction
The underrepresentation of membrane protein structures in the
Protein Data Bank [1] is a direct result of the inherent difficulty of
membrane protein crystallization [2], but it stands in sharp contrast with the relevance of membrane proteins to cellular functioning. Roughly 30% of all genomic sequences encode for membrane
proteins, in fact most major processes in the cell are initiated at the
membrane surface. Due to the lack of atomic resolution structural
data molecular modeling and simulation techniques are expected
to play an increasingly relevant role in the study of membrane-
related systems. Membrane protein simulations complicate matters
with respect to soluble proteins at a number of levels. The generally larger size of membrane proteins makes for slower convergence, the presence of a lipid bilayer imposes a larger system box
size, both adding up to longer simulation times. The longer simulation times produce larger trajectory files, which are more complicated and take longer to process. However, the bilayer plane
typically aligns with one of the primary system axes, offering an
easy frame of reference.
109
110
Marc F. Lensink
This chapter presents some tips and tricks for the simulation of
membrane-bound proteins and peptides. These include simulation
setup and the creation of lipid bilayers of mixed content, but also
selected analysis tools are presented, that for example allow the
easy extraction of coordinates from the trajectory, the calculation
of helical axes and aromatic order parameters, the determination of
peptide- or protein-interacting lipids, and the investigation of specific proteinlipid interactions through the calculation of lipid-
bridged residueresidue contacts. The tips and tricks presented
inthis chapter refer to the GROMACS [3] suite of programs
(see Note 1), but their principles are generally applicable to other
simulation packages. The presented simulation setup and analysis
tips are based on two simulation studies: the association of a cationic peptide to a neutral and charged lipid bilayer [4], and the
detection of proteinlipid binding in an integral membrane protein
[5]. The analysis tools I wrote are either shell-script, or using the
GROMACS C programming libraries. Most of these are package-
independent since they only require a trajectory file, which is no
more than a sequence of structures.
2 System Setup
The general setup of molecular dynamics simulations requires
three input files: a structure file, describing the atomic positions of
the molecules in the system; a topology file, in which the inter-
atomic bonded and non-bonded connections are defined; and a
parameter file, supplying the algorithm with the necessary runtime
parameters. The first two of these I will briefly discuss. Programs or
files in this section are printed in bold typeface.
2.1 The
CoordinateFile
The coordinate file contains the coordinates of all molecules in

the system, in the order in which they appear in the topology file
(or vice versa, if you want). Unless the simulation you want to
run is based upon previous work, no complete coordinate file
exists and you need to create one by combining a PDB file of
your protein with apreferentially equilibratedlipid bilayer.
2.1.1 Situation A:
Placing aPeptide onTop
oftheLipid Bilayer
In this case the coordinates of peptide and lipids do not occupy the
same space and a simple concatenation of input files suffices.
1. Prepare a PDB file with the coordinates of your peptide.
2. Rotate the structure to adopt the wanted orientation: parallel
or perpendicular to the bilayer.
3. Have its geometric center coincide with that of the lipid bilayer
and increase the z coordinates (see Note 2) by 56nm
(see Note 3).
Membrane-Associated Proteins and Peptides
111
4. Add the peptide coordinates to the (solvated) bilayer box

(see Note 4).
5. Calculate the minimum box height needed and combine this z
axis with the x and y axes of the original bilayer.
6. Remove counterions (if any) and solvate the box. In order to
avoid water molecules being placed inside the hydrophobic
bilayer core, increase the Van der Waals radius of the lipid acyl
chain atoms to 6 (see Note 5).
7. Add counterions to make the system electrostatically neutral.
2.1.2 Situation B: Placing
aProtein Inside theLipid
Bilayer
Using the procedure above will result in overlapping protein and

lipid coordinates. Simply cutting away the excess lipids will not
work since the formed vacuum is filled with water molecules, which
lead to increased equilibration times. My preferred approach is
using inflategro, a tool that expands the x and y coordinates of the
bilayer, places the protein in the center and then iteratively reduces
the lipid x and y coordinates towards their original values [6].
An alternative method is through g_membed [7]. The following
procedure uses the principles of inflategro:
1. Prepare a PDB file with the coordinates of your protein.
2. Rotate and translate the structure to adopt the desired orientation (see Note 6).
3. Place the structure in the center of an equilibrated lipid bilayer,
with all waters removed and of which the x and y coordinates
are expanded by a factor of 4.
4. Reduce the system to its reference area per lipid by shrinking
the lipid x and y coordinates iteratively by 2%, deleting all lipids that have their phosphorus atom at a distance closer than
6 to any protein C atom. Let each iterative step be followed
by 100 steps of steepest descent energy minimization while
employing tight (105kJ/nm2) position restraints on the protein non-hydrogen atoms (see Note 7).
5. Expand the system box in the z direction to avoid the protein
being too close to its own copy.
6. Solvate the box, employing increased Van der Waals radii for
the lipid acyl chain atoms, and add counterions.
2.2 The Topology File
The common approach in setting up molecular dynamics simulations is to translate a complete coordinate filecontaining the
atomic coordinates of the bilayer, protein, water and counterions
into a topology using the residue building blocks. In spite of the
maturation of molecular dynamics force fields for the simulation of
proteins, lipid parameters are not as well integrated into these as
one would like. A useful approach therefore is to separate the
112
Marc F. Lensink
phosphate
P O4
CH 2
head group
POPC:
R=
+
(CH2 )2 N(CH3 ) 3
POPE:
R=
+
(CH2 )2 NH 3
POPG:
R=
CH 2 (CHOH)2
POPA:
nothing
POPS:
R=
CH
CH 2
( CH 2 )7
CH
CH 2 CH CO 2
NH+
3
double bond
carbonyl
( CH 2 )14
CH 3
CH
( CH 2 )7
CH 3
Fig. 1 Molecular structure of selected phospholipids, here with a sn1 palmitoyl and sn2 oleoyl tail. The different
head groups determine the overall charge of the molecule: PG, PA and PS carry a negative charge, while PE
and PC are neutral. The dashed boxes indicate commonly defined groups of atoms
c reation of topology files for protein and lipids and use a container
to combine them. The full procedure then becomes:
1. Extract the coordinates of your protein and process these with
pdb2gmx to create a topology. In case you have a small peptide, consider capping the C- and/or N-terminus (see Note 8).
2. Cut everything from the topology file that is not referring to
the molecule definition and save the result into a file called
protein.itp. This file you can then include at the appropriate
position in your global container topol.top.
3. Collect the topological descriptions for the other molecules in
the system, i.e., lipids, water, and counterions, and combine
the topology files (see Note 9).
2.3 The Lipid Bilayer
The easiest approach to create a lipid bilayer is to start from one

that is publicly made available. A popular bilayer system is one that
contains 128 POPC molecules (two leaflets of 64 molecules each).
POPC contains a 15-carbon 1-palmitoyl chain and a 17-carbon
2-oleoyl chain, which has a double bond in the middle. The PC
head group contains a negatively charged phosphate and a positively charged choline moiety, ensuring electrostatic neutrality and
thus avoiding the necessity to include an additional 128 Na+ counterions. The bilayer structure and topology files are freely available
(see Note 10). Figure1 shows the molecular structure of POPC
and additional phospholipids. The head group function determines
the chemistry and overall charge of the lipid.
2.4 Modification
ofLipid Topology
andMixed-Lipid
Bilayers
113
The creation of force field parameters is a field on its own. However,

in the principle of parameter transfer, one copies parameters from
well-calibrated protein residue parts onto a newly created lipid head
group function. Charges can be obtained by fitting point charges to
reproduce the electrostatic potential following ab initio density
functional calculations. In this way one can quite easily produce
POPG, POPE, POPS, and POPA topologies [4, 5] (see Note 11).
Modification of the lipid topology files directly happens as follows:
1. Remove all atoms, bonds, pairs, angles, etc. involving the head
group from the POPC lipid topology file.
2. Increase/decrease all references to atom number to account
for the difference in number of head group atoms between
POPC and the new lipid.
3. Add the new head group function to the section listing the
atoms (see Note 12).
4. Add the parameters for the new head group function, copying
from chemically equivalent groups in residues (see Note 13).
5. Add bonded terms (bond, angle, dihedral, and 14 pair interaction) for the atoms involving the head group and the
connecting phosphate entity (see Note 14).
6. Perform a simulation of a single lipid in vacuo to check
the stability of the structure in the newly created topology
(see Note 15).
An alternative approach is to construct an appropriate rtp entry
in the residue topology database and create the lipid topology automatically by processing the single-lipid structure with pdb2gmx,
or by using public databases such as LipidBook [8] (see Note 16).
However, whatever way you create your topologies, always have in
mind that they need to be validated.
The resemblance of the various lipid types in terms of chemical
structure may be exploited in order to create mixed bilayers, e.g.,
bilayers consisting of 80% POPC and 20% POPG.Since the acyl
tail groups are the same, only the head group needs to be replaced.
A step-by-step approach is as follows:
1. Perform a least-squares RMSD fit, fitting the phosphate groups
of POPG and POPC onto each other, correctly aligning the
oxygen atoms that point towards the head group and the acyl
chains.
2. Extract the head group coordinates from the fit and the phosphate and tail coordinates from the reference.
3. Rename tail atoms to correspond to the names in the POPG
topology.
4. Copy the coordinates of remaining lipids and add the new
head group, phosphate, and acyl tail coordinates with POPG
as residue name.
114
Marc F. Lensink
Fig. 2 A solvated bilayer containing 90% DPPC and 10% cholesterol molecules. DPPC displayed as black
wireframe, cholesterol as blue sticks, and water molecules as red spheres. The configuration was created with
DOPE, the figure prepared with PyMol (The PyMOL Molecular Graphics System, version 1.6.0, Schrdinger, LLC)
5. Repeat the procedure until the desired number of POPC lipids

has been replaced.
6. Restore the original box and reorder the file (see Note 9).
In this way many lipid molecules can be incorporated as long
as the new head group and/or tail atoms occupy more or less the
same space (see Note 17). I have written a dedicated program for
this purpose, called DOPE (see Note 18), which tries to align a
replacement lipid with its target molecule, either best-fitting the
molecular axes, or by comparing atom names and chemical types.
The program will be described in a separate publication, but may
be requested by contacting the author. Briefly, the program iteratively replaces a lipid from the input file by a molecule from a
library of structures, aligning the molecules molecular structure
principal axes to the system axes. It subsequently checks overlap
with other molecules in the system and accepts the new molecule
when no clashes are found. If clashes do occur, the library molecule can be rotated about any of its axes and the new position is
re-evaluated. Figure2 shows a 9:1 DPPCCholesterol bilayer created by DOPE. The procedure started with an equilibrated 340-
lipid POPC bilayer, deleting 2 lipid tail atoms to transform POPC
into DPPC, and thereby keeping atomic positions. The system was
re-equilibrated using molecular dynamics and in a second DOPE
step 10% DPPC molecules were replaced by cholesterol, using an
a priori created library of cholesterol conformations.
115
3 Analysis
The dynamics of a protein is strongly affected by the presence of a
lipid bilayer. Backbone hydrogen bond shielding [9] and a
decreased dielectric constant in the membrane core [10] promote
the formation of secondary structure, both or . In addition, the
bilayer environment places a restraining force on the protein
dynamics due to the decreased fluidity with respect to a soluble
environment. Such external forces are characterized by slow-
motion displacements of secondary structure elements, readily
identified from RMSD plots after fitting to a common reference
frame, typically the protein transmembrane domain. An additional
frame of reference exists in the surface plane of the bilayer, ignoring eventual curvature effects. The orientation of a protein with
respect to the membrane it is binding to is especially relevant in the
case of peptidemembrane association.
3.1 Coordinate
Frame inBilayer
Simulations, g_zcoor
andg_xycoor
By choosing the appropriate reference axes for your simulation you

can significantly facilitate subsequent analyses. In general, the z axis
is made to coincide with the normal to the bilayer surface. Any property involving distance to bilayer is then calculated in the z axis
only, possibly in relation to the bilayer or head group center. The
program g_zcoor does exactly this: extract the (center-of-mass averaged) z coordinate of a combination of molecules and/or atoms.
Orthogonally to this, the x and y axes describe diffusion in the
bilayer plane or surface. The program g_xycoor extracts the x and
y coordinates of every single atom in any number of combinations
of molecules and/or atoms. Modification of these absolute coordinates into relative ones, e.g., relative to a membrane-inserted protein or membrane-bound peptide, shows the restricted diffusional
motion of annular or bound lipid molecules. For both programs,
g_zcoor and g_xycoor, Note 11 applies.
3.2 Calculation
ofHelical
Axis, g_helixaxis
Hydrophobic mismatch is defined as the difference in hydrophobic

length of a proteins transmembrane domain and the thickness of
the lipid bilayer it spans. In order to minimize the exposed hydrophobic surface, proteins or peptides may aggregate, or adopt a
tilted orientation [11]. The axis of a helix is best determined using
a rotational least-squares fitting procedure, mapping the Cs of
the helix onto itself, but one residue out of phase, i.e., residue i is
mapped onto residue i+1. A quaternion-based method identifies
the screw transform (translation along and rotation about the helix
axis) that will superimpose the two helices [12]. In the case of a
single -helix containing peptide, the angle between the helix and
the normal to the bilayer plane identifies the tilt of the helix in the
bilayer. For proteins containing multiple helices in their TM
domain, the various angles between those helices provide information as to the internal organization of those helices.
116
Marc F. Lensink
b
1.0
SL
SN = 1/2
SL = 1/2
SN = 1/2
SL = 1
0.5
0.0
SN = 1
S=
1/2
L
SN
-0.5
8
12
Time (ns)
16
20
Fig. 3 Aromatic order parameters. (a) Visualization of aromatic order parameters. Solid and dashed arrows
represent SL and SN, resp. When either arrow is aligned with the normal to the bilayer plane (long arrow), the
respective order parameter equals 1. (b) Concomitant behavior of aromatic order parameters, for a single
tryptophan residue during a 20ns molecular dynamics simulation. Both order parameters cannot simultaneously be aligned to the z axis (equal 1), but they can be orthogonal to it (equal). Notice the immediate
decrease in SL after an increase of SN
The rotational least-squares method is fast, accurate and

insensitive to noise, and thus able to deal well with imperfect helices. It has been implemented in a program that uses the gromacs
development and analysis libraries: g_helixaxis (see Note 11).
The program can read trajectory and single structure PDB files
and outputs for each helix its angle with the z-axis as well as their
inter-helical angles. Optionally, the initial point, and vector components and length of each helix are written, in a format following
PDB standards and thus easily visualized using standard molecule
viewers.
3.3 Orientation
ofAromatic Residues,
g_arom
Whereas the calculation of helical axes helps in determining the

dynamics of secondary structure elements, at a more local level aromatic order parameters are used [13]. The aromatic order parameters SN and SL are calculated relative to the normal to the bilayer
plane, through the formula S=(3cos21). SN relates to the normal to the aromatic ring, whereas SL describes the vector from C
through the ring. is the angle between the respective vector and
the bilayer normal. For S=1 these vectors are aligned, whereas
S= means orthogonality. The different orientations and order
parameter values are visualized in Fig.3a. As these vectors cannot
both simultaneously be aligned with the bilayer normal, the combinations SN=1 and SL=1 are mutually exclusive and an increase in
the one induces a decrease in the other. They can, however,
117
both equal , meaning that both vectors are orthogonal to the

bilayer normal. The mutually exclusive behavior is illustrated in
Fig.3b.
Calculation of the aromatic order parameters has been implemented in an analysis tool that can read both trajectories and PDB
files: g_arom (see Note 11). Detection of aromatic residues is automatic, but it is also possible to select the residues of interest. The vectors CC for Phe and Tyr, and CC2 for Trp, are used for the
calculation of SL, whereas the aromatic plane, defined by the atoms
C, C1, and C2 for Phe and Tyr, and C, C2, and C3 for Trp,
determines the normal to this plane, used in the calculation of SN.
3.4 Lipids
Interacting
withaProtein
or Peptide, g_under
The interaction between a peptide or protein and a lipid bilayer is

not a static quantity that can be defined as the interaction between
a certain number of residues and an arbitrary number of lipids.
Although membrane-interacting residues may be likely to sustain
this interaction once established, lateral lipid diffusion will take
place, replacing individual lipids, much like a water molecule interacting through hydrogen bonding with a protein residue may
exchange with bulk water.
Here I present a pragmatic way of defining lipids that interact
with a peptide or protein, which is purely distance-based. This definition [4] is implemented in the program g_under (see Note 11).
Essentially, lipids that come within a certain cutoff distance of the
peptide are defined as interacting with this peptide. This distance
can be calculated between any two peptide and lipid atoms, but
also be restricted to use only backbone atoms. The distance criteria
need not be the same in the x, y or z direction. In the case of a
peptide hovering above a lipid bilayer, one can picture a cylinder
with a radius 0.1 and height of 2nm, hanging below a certain
peptide atom. Any lipid atom entering this cylinder will define the
lipid as interacting with the peptide (atom). This procedure
includes lipids that have their acyl chains under, but the head group
besides the peptide (see Note 19).
The resulting time-dependent evolution of peptide-interacting
lipids can subsequently be used to calculate properties involving
these lipids only, as opposed to the entire bilayer or bilayer leaflet.
Similarly one can define a short cylinder with larger radius, or have
its main axis align with the x or y axis, to investigate properties of
lipids neighboring residues of integral membrane proteins.
3.5 Calculating
Properties
ofInteractingLipids
Most analysis programs calculate a quantity from the interaction

between the atomic coordinates of one set of atoms vs. another.
Doing so for every frame in the simulation one obtains the evolution of this quantity over time. It is usuallyand also in the case of
gromacs analysis toolsnot possible to vary one of these sets of
atoms, as one would need for example in the case of peptide-
interacting lipids or when studying a shell of water molecules
around an active site.
118
Marc F. Lensink
However, when the quantity to be calculated is cumulative,

i.e., the quantity can be calculated a posteriori from each individual
lipid molecule at the instantaneous time t, e.g., the average z coordinate of the phosphorus atoms of the protein- or peptide-
interacting lipids, or the lipid order parameters, one simply needs
to traverse the trajectory and extractfor the first examplethe z
coordinate for every single lipid. Then in a second step these can
be combined with the list of peptide-interacting lipids to get the
evolution of average z coordinate of all interacting lipids during
the course of the simulation (see Note 20). Alternatively, one could
cut the trajectory in pieces and scan these individually. This has the
advantage that for a lipid that becomes interesting during only a
fraction of the simulation not the entire trajectory needs to be
scanned, but only a (small) part of it (see Note 21).
3.6 Bilayer Structure
Lipid deuterium order parameters describe the ordering of the

lipid acyl chains with respect to the bilayer normal. They can be
measured by NMR experiments, but also calculated from the lipid
tail CC dihedral angles [14]. They are expressed as a scalar value
per lipid carbon atom that typically ranges between 0 for disordered and 0.5 for ordered lipid structure. The following example
shows the calculation of lipid order parameters for lipids that are
interacting with a peptide, following the previous section.
1. Calculate which lipids interact with the peptide for every frame
in the trajectory.
2. For each lipid tail:
a. Calculate the lipid order parameters for each individual
lipid and lipid tail for every frame in the trajectory. These
time frames should match the calculation of peptide-interacting lipids.
b. For each time frame:
Extract the residue numbers of the peptide-interacting

lipids.
Average the calculated order parameters for the given
lipid tail for these lipids at the given time frame.
c. Average these averaged order parameters over all time

frames.
Figure 4 shows the difference in order parameters between
interacting (open circles) and non-interacting (solid circles) lipids.
3.7 Lipid-Mediated
Salt Bridges
Integral membrane proteins are immersed in a lipid environment

and as such, their activity can be directly related to global membrane properties such as fluidity [15]. Anchoring of the proteins
occurs through nonspecific proteinlipid interactions [16]. Many
of these interactions are hydrophobic, e.g., lipid acyl chains that
settle on a hydrophobic surface patch created by one or several

sn1 Palmitoyl
0.30
0.30
0.25
0.25
0.20
0.20
0.15
0.15
0.10
0.10
0.05
0.05
0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Carbon atom number
9 10 11 12 13
Scd
Scd
sn2 Oleoyl
119
0.00
Carbon atom number
Fig. 4 Lipid deuterium order parameters calculated over a 50ns molecular dynamics trajectory of a 16-residue
peptide bound to a lipid bilayer. Solid circles denote order parameters calculated over all lipids, open squares
are for peptide-interacting lipids only. Peptide-interacting lipids here account for about 12% of the lipid bilayer
(15 lipids), or 2025% of the bilayer leaflet
transmembrane helices. But in addition, the lipid carbonyl or phosphate groups can act as acceptor for hydrogen bonds emanating
from the protein, or salt bridges may be formed between opposing
charges, e.g., between phosphate and arginine or lysine.
Hydrophobic interactions are the weakest of these and the strongest interactions are made by salt bridges, which are in fact a particularly strong form of a hydrogen bond. Lipids surrounding the
protein and anchoring it in the bilayer are called annular lipids.
Annular lipids show increased residence times but exchange with
bulk lipids occurs on a regular basis. For sufficiently long simulations such effects can be quantified in the lipid diffusion. However,
there is increasing evidence for the existence of non-annular lipid
binding sites, where specifically bound lipids are necessary to
achieve biological function [17, 18].
Molecular dynamics simulations may detect such strong interactions through the occurrence of lipid-mediated salt bridges, where a
single lipid bridges both a negatively and positively charged at the
same time. A striking example is a phosphatidylethanolamine (PE)
lipid binding simultaneously two neighboring charged residues
D68 and K69in lactose permease [5]. The binding to K69 is
nonspecificany phospholipid contains a phosphate groupbut
the binding of the ethanolamine moiety to D68 is PE-specific. In

addition, D68 is a highly conserved residue in the Major Facilitator
Superfamily, which groups together some 15,000 membrane transporter proteins [19]. The interaction is depicted in Fig.5.
120
Marc F. Lensink
Fig. 5 Example of a lipid-mediated salt bridge. The figure shows a POPE lipid
bound to both Asp-68 as well as Lys-69 of lactose permease (LacY). LacY is
drawn in cartoon representation, the PE lipid and residues 68 and 69in ball-and-
stick. The bond lengths displayed are between the hydrogen bond donor and the
hydrogen atom itself
Proteinlipid interactions can be investigated with the program

g_prolip (see Note 11). The program takes a gromacs trajectory and
topology file as input and then traverses this trajectory to look for
lipid-bridged residueresidue contacts. More precisely, it detects
when during the trajectory the same lipid is bound simultaneously
to two different protein residues. By calculating the cumulative presence of a given lipid-mediated salt bridge (see Note 22) and combining this with the root mean square fluctuation (MSF) of the lipid
donor and acceptor atoms (see Note 23) through the formula
Fatom = Dtcumul ( MSFmax - MSFatom + MSFmin ) ,
one gets the persistence factor F, which exists for both the donor
(Fdonor) as well as acceptor (Facceptor) interaction. The persistence
factor is an indication of the strength of interaction and is typically
correlated with residue conservation [5].
3.8 Downloadable
Files
The analysis programs
g_zcoor, to plot average z coordinate,
g_xycoor, to plot x and y coordinates,
g_helixaxis, to calculate the axis of a helix,
g_arom, to calculate aromatic order parameters,
121
g_under, to calculate which lipids interact with a protein or

peptide, and
g_prolip,
contacts,
to
calculated
lipid-bridged
residueresidue
and topology files for POPS, POPC, POPE, and POPG lipids
are made available to the scientific community (see Note 11).
Gromacs needs to be installed (see Note 24), as these programs
dynamically link to the gromacs libraries, but to be able to use
these programs the simulations need not necessarily be performed
by gromacs.
4 Notes
1. http://www.gromacs.org/
2. We assume the z axis aligns with the normal to the bilayer
plane.
3. A typical bilayer has a thickness of 44.5nm. A 16-residue
alpha-helical peptide has a length of about 2.5nm. If we want
to place the perpendicularly to the bilayer plane at a minimum
distance of about 2nm we need to overcome half the bilayer
thickness, half the peptide length, and add the extra 2nm, i.e.,
translate by at least 5.5nm.
4. You can use a solvated bilayer box since the solvation procedure will remove overlapping waters.
5. This is the file vdwradii.dat, which can be copied from the
gromacs topology directory to the working directory.
6. If your protein structure comes from the Protein Data Bank, it
likely features in the Orientation of Proteins in Membranes
database [20, 21]. The database contains membrane protein
structures with a disk of dummy atoms located at the point in
the lipid bilayer (at either side) where the hydrophilic to hydrophobic transfer energy derivative maximizes, i.e., roughly at
the height of the phosphorus atoms in a phospholipid bilayer.
The protein already contains the correct x and y orientation, so
only a translation in the z axis is needed.
7. Steps 3 and 4 can be taken care of by inflategro. After about
eight iterations, the deflation can be increased to 5% per step.
8. Capping is generally necessary to avoid artifacts from a
terminal charge caused by the artificial chain breaking. Take
especially care of capping if the simulations complement
experiments where the peptide was capped at one or both
ends. Capping is easiest performed using the residue topology database by adding a residue with the correct name
atthe terminus; hydrogens are then added automatically.
122
Marc F. Lensink
Some RMS fitting may be necessary, but the exact position

of the cap atoms is not very important because the energy
minimization will quickly relax them.
9. The force field definitions (bonded and non-bonded interactions) can be included in any order, but must appear before the
molecule type definition. The final section defines the molecules that are present in the system, in this section the order of
the molecules must reflect the order in which they are found in
the coordinate file.
10. Many such files can be found at http://moose.bio.ucalgary.ca/
11. Files can be downloaded at http://cb.iri.univ-lille1.fr/Users/
lensink/lipid/
12. Rename new lipid atoms to avoid overlap with existing ones.
This step is not necessary if the new atoms have unique names.
13. If chemically equivalent groups are not available for the force
field you are using, you will have to go through the whole
process of deriving parametersespecially partial charges
from quantum mechanical calculations, following the procedure as described in the literature for your chosen force field.
14. Also here Note 13 applies, but at this point the charges are
already known. Other parameters are less critical, e.g., angles
and dihedrals can be made to follow sp2 or sp3 hybridization
and bond lengths taken from experimentally determined values (NMR or X-ray). Moreover, in most present-day simulations bond lengths are constrained.
15. Incorrect topologies will quickly explode or collapse. Check
the final structure: if it looks okay, it probably is okay. Remove
rotational center-of-mass motion to avoid accelerated spinning. Vacuum simulations should be sufficiently long (in the
order of several nanoseconds) to allow the dissipation of energy
in the limited number of degrees of freedom.
16. http://lipidbook.bioch.ox.ac.uk/
17. Overlap with water is not a problem since the conflicting water
molecules can easily be removed, either manually or automatically via the solvation step.

18. Not to be confused with the lipid DOPE (di-oleoylphosphatidylethanolamine).
19. Specifically, first every atom that enters the cylinder is calculated
and subsequently this group is expanded into full residues. The
average distance of the resulting group of atoms to any other
group of atoms can be calculated, with the possibility of excluding itself. More concretely, one could calculate the evolution of
the distance between the average position of the phosphorus
atoms of interacting and non-interacting lipids of one bilayer
half during the course of the molecular dynamics simulation.
123
20. For a 128-lipid bilayer this still means that the trajectory has to
be traversed 128 times. When only the peptide-interacting lipids are required, a first step would be the identification of these
lipids to avoid unnecessary processing of the trajectory.
21. Scanning of a trajectory file containing all coordinates in the
system, including water, may become prohibitively slow for
extended simulation times. For many analyses not all coordinates are required and in those cases it is advised to create a
copy of the trajectory file, but containing only those coordinates needed for the analysis. This step usually results in a trajectory that is small enough to avoid the necessity of cutting it
in pieces.
22. This is defined as the combined fractional presence over the
entire simulation and can be calculated through division of the
number of frames the bridge is active by the total number of
frames in the simulation.
23. This is the root mean square fluctuation of atomic positions,
which basically gives information as to how mobile the atom is.
24. The programs compile against gromacs versions 4.5 and 4.6.
Compilation against earlier and later versions may require
minor adaptation of the code.
References
1. Berman H, Henrick K, Nakamura H, Markley
JL (2007) The worldwide Protein Data Bank
(wwPDB): ensuring a single, uniform archive
of PDB data. Nucleic Acids Res 35(Database
issue):D301D303
2. Carpenter EP, Beis K, Cameron AD, Iwata S
(2008) Overcoming the challenges of membrane protein crystallography. Curr Opin
Struct Biol 18(5):581586
3. Pronk S, Pall S, Schulz R, Larsson P, Bjelkmar P,
Apostolov R etal (2013) GROMACS 4.5: a
high-throughput and highly parallel open source
molecular simulation toolkit. Bioinformatics
29(7):845854
4. Lensink MF, Christiaens B, Vandekerckhove J,
Prochiantz A, Rosseneu M (2005) Penetratin-
membrane association: W48/R52/W56 shield
the peptide from the aqueous phase. Biophys J
88(2):939952
5. Lensink MF, Govaerts C, Ruysschaert JM
(2010) Identification of specific lipid-binding
sites in integral membrane proteins. J Biol
Chem 285(14):1051910526
6. Schmidt TH, Kandt C (2012) LAMBADA and
InflateGRO2: efficient membrane alignment
and insertion of membrane proteins for molecular dynamics simulations. J Chem Inf Model
52(10):26572669
7. Wolf MG, Hoefling M, Aponte-Santamaria C,

Grubmuller H, Groenhof G (2010) g_membed:
efficient insertion of a membrane protein into an
equilibrated lipid bilayer with minimal perturbation. J Comput Chem 31(11):21692174
8. Domanski J, Stansfeld PJ, Sansom MS,
Beckstein O (2010) Lipidbook: a public
repository for force-field parameters used in
membrane simulations. J Membr Biol
236(3):255258
9. Garcia AE, Sanbonmatsu KY (2002) Alpha-
helical stabilization by side chain shielding of
backbone hydrogen bonds. Proc Natl Acad Sci
U S A 99(5):27822787
10. Avbelj F, Luo P, Baldwin RL (2000) Energetics
of the interaction between water and the helical
peptide group and its role in determining helix
propensities. Proc Natl Acad Sci U S A 97(20):
1078610791
11. Zhang YP, Lewis RN, Hodges RS, McElhaney
RN (1992) Interaction of a peptide model of a
hydrophobic transmembrane alpha-helical segment of a membrane protein with phosphatidylcholine bilayers: differential scanning
calorimetric and FTIR spectroscopic studies.
Biochemistry 31(46):1157911588

12. Christopher JA, Swanson R, Baldwin TO
(1996) Algorithms for finding the axis of a helix:
124
Marc F. Lensink
fast rotational and parametric least-

squares
methods. Comput Chem 20(3):339345

13.
Tieleman DP, Forrest LR, Sansom MS,
Berendsen HJ (1998) Lipid properties and the
orientation of aromatic residues in OmpF,
influenza M2, and alamethicin systems: molecular dynamics simulations. Biochemistry
37(50):1755417561
14. Merz KM, Roux B (1996) Biological membranes: a molecular perspective from computation and experiment. Birkhauser, Boston, MA,
xiii, 593 p., 1 leaf of plates
15.

Vigh L, Escriba PV, Sonnleitner A,
Sonnleitner M, Piotto S, Maresca B etal
(2005) The significance of lipid composition
for membrane activity: new concepts and
ways of assessing function. Prog Lipid Res
44(5):303344
16. Nyholm TK, Ozdirekcan S, Killian JA (2007)
How protein transmembrane segments sense
the lipid environment. Biochemistry 46(6):

14571465
17. Lee AG (2005) How lipids and proteins interact in a membrane: a molecular approach. Mol
Biosyst 1(3):203212
18. Paila YD, Tiwari S, Chattopadhyay A (2009)
Are specific nonannular cholesterol binding
sites present in G-protein coupled receptors?
19. Kaback HR (2005) Structure and mechanism of
the lactose permease. C R Biol 328(6):557567

Mosberg HI (2006) OPM: orientations of proteins in membranes database. Bioinformatics
22(5):623625
21. Lomize MA, Pogozheva ID, Joo H, Mosberg
Chapter 7
Coarse-Grained Force Fields for Molecular Simulations
Abstract
Molecular dynamics (MD) simulations at the atomic scale are a powerful tool to study the structure and
dynamics of model biological systems. However, because of their high computational cost, the time and
length scales of atomistic simulations are limited. Biologically important processes, such as protein folding,
ion channel gating, signal transduction, and membrane remodeling, are difficult to investigate using atomistic simulations. Coarse-graining reduces the computational cost of calculations by reducing the number of
degrees of freedom in the model, allowing simulations of larger systems for longer times. In the first part of
this chapter we review briefly some of the coarse-grained models available for proteins, focusing on the
specific scope of each model. Then we describe in more detail the MARTINI coarse-grained force field, and
we illustrate how to set up and run a simulation of a membrane protein using the Gromacs software package.
We explain step-by-step the preparation of the protein and the membrane, the insertion of the protein in the
membrane, the equilibration of the system, the simulation itself, and the analysis of the trajectory.
Key words Coarse-graining, Molecular dynamics, Force field, MARTINI, Protein, Lipid membrane
Introduction
Atomistic molecular dynamics (MD) simulation provides structural
and dynamic information on molecular systems on a sub-nanometer
length scale, with femtosecond time resolution. It is a powerful
tool to interpret experiments, to predict structure and dynamics of
simple systems, and to get an insight into processes that are difficult
to explore experimentally due to limited length or time resolution.
Biologically relevant phenomena, like protein folding, ion channel
gating, signal transduction, and membrane remodeling often occur
on times scales of microseconds or greater [1]. These time scales
are computationally expensive for atomistic MD simulations, which
rarely extend beyond the microsecond. Sampling is often an issue:
it is not always possible to run long enough or large enough simulations. Therefore, some phenomena are out of reach for state of
the art atomistic MD.
Several strategies exist to increase the range of what can be
sampled. One strategy is to make better use of modern hardware;
125
126
for instance, the most recent versions of major MD packages

support efficient use of graphical processing units (GPU). Another
example of efficient use of hardware is Anton, a special purpose
supercomputer built by the Shaw group; Antons microchips were
designed specifically to do computations that are frequent in MD
calculations [2]. This new hardware can perform millisecond-long
MD simulations [3]. Beside hardware improvements, the use of
some algorithms can improve the sampling of energy landscapes
by overcoming free energy barriers, for example Monte-Carlo
simulations, simulated annealing, metadynamics algorithms [4],
replica exchange [5], or biased MD algorithms like umbrella
sampling [6]. Yet, these algorithms are not suitable for all problems
and may require long runs to reach convergence.
Coarse-graining is another approach to tackle sampling issues
and size limitations. It consists of reducing the complexity of the
simulated system by grouping atoms together into virtual particles.
Reducing the number of particles in the simulated system reduces
the number of interactions to calculate, and smoothens the energy
landscape; hence coarse-grained (CG) simulations are faster.
The specific acceleration of CG simulations depends on the details
of the models. The way CG models are built depends on the kind
of scientific question they aim to address. When reducing the level of
detail of the representation it is important not to oversimplify the
description of the system, i.e., to preserve the features that drive
the phenomenon of interest.
The variety of coarse-grained models is large. Solvent can be
represented by explicit particles or by a potential that mimics its
effect; the mapping of the system can vary from one particle per
protein to several particles per amino acid; the functional forms can
be the same as in all-atom simulations or entirely different, for
example they can include tabulated potentials.
In general, the range of applicability of coarse-grained models is
more limited than for atomistic models, i.e., CG models parameterized to reproduce certain phenomena and properties might fail to
reproduce other phenomena and properties. This is due to intrinsic
limitations in the transferability of the potentials between different
chemical environments, thermodynamic conditions, etc. [7].
Theory
2.1 Coarse Grain

Models for Proteins
A wide variety of CG models for proteins are currently available.

Again, the precise coarse-graining methodology depends on the
question to be answered, as the question defines the degrees of
freedom that will be removed.
The first coarse-grained models for proteins were meant to
tackle the problem of protein folding. In the pioneering work by
Warshel and Levitt [8], published in 1975, each amino acid was
127
represented by two spheres: one at the alpha carbon position,

and one at the side chain centroid. The torsion angle between two
consecutive residues was the only degree of freedom. Since then,
several models have been developed to study protein folding and
mechanics. These models can be very different from each other in
their mapping, in the functional form and in the way their parameters are derived. For instance, the UNRES [9] force field uses
anisotropic potentials to model interactions involving amino acid
side chains. This model represents the main chain by one interaction site between two alpha carbons. The alpha carbons are used
only to join the interactions sites and the side chains. The OPEP
force field [10] and the protein force field by Bereau and Deserno
[11] have similar mapping: both use a quasi-atomistic main chain
and represent the side chain by a single bead. The difference
between them lies mainly in the potential form and in target properties used in the parameterization. The PaLaCe force field [12]
also uses a quasi-atomistic representation of the protein main
chain. Yet, from the main chain, only the alpha carbon is used in
the calculation of non-bonded interactions, while the other heavy
atoms are used for bonded interactions, avoiding complicated
angle and torsion potentials and allowing an explicit treatment of
backbone hydrogen bonds. A coarser resolution is adopted for
residue-specific non-bonded interactions, with just one or two
interaction centers (except for Glycine) at the side-chain level.
Besides their different mappings, those models differ in the way
that interactions parameters are derived. The choice of the target
properties reflects the type of problem to be tackled by each model:
folding for OPEP and Desernos model, structural fluctuations
beyond the elastic regime for PaLaCe.
Other coarse-grained force fields do not tackle questions
related to protein folding. For instance, Zacharias [13, 14] developed a CG model for protein docking. This model represents proteins as rigid bodies, with each residue modeled by up to three
pseudo-atoms. This representation is used in energy minimizations
from multiple starting points to find the protein association with
the lowest energy [13]. Flexibility of the side chains is taken into
account by using several rotamers for each residue; global flexibility can be modeled with energy minimization in low frequency
normal modes. The model has been extended to docking of proteins with nucleic acids [14].
Mapping of CG models varies depending on the purpose of
the model: when the purpose is to represent motions involving
large molecular assemblies (tens of nanometers or more), then the
mapping is typically coarser. An example is provided by clathrin
self-assembly, studied by Den Otter and Matthews using a very
coarse CG model [15, 16]. Clathrins are proteins that self-assemble
to form a coat that shapes endocytic vesicles. Clathrin self-assembly
involves tens of proteins and a large lipid membrane. Such
128
simulations were made possible by modeling clathrin with only few

particles representing the proteins general shape. In this case,
small-scale structural details were sacrificed in favor of computational speed.
Multiple levels of coarse graining can be considered in a single
simulation; this approach is often referred to as multi-scale. Over
the past decade, several research groups have been developing
atomistic/CG multi-scale approaches. Such multi-scale models can
consider different molecules at different levels of description in the
same system; for example, a membrane protein can be described at
the all-atom level while the membrane itself can be coarse-grained
[17]. Another multi-scale approach considers different levels of
representation within the same molecule; a protein can be described
partly with all-atom and partly with CG representation [18, 19], or
with several levels of coarse-graining at the same time. For instance,
in their simulations of the interactions between actin and myosin,
Taylor and Katsimitsoulia [20] used simultaneously three levels of
CG representation. At the more detailed level, they represented
secondary structure elements as cylinders; then they grouped the
cylinders into domains and the domains into proteins. Lower level
elements did not interact with each other unless they collided.
Such approach allows considering interactions at the lower scale
only when it is pertinent. The level of resolution can also be
changed on-the-fly, as proposed in the Adaptive Resolution
Scheme (AdResS) by Delle Site and coworkers. The basic theoretical
principles of this methodology are illustrated in a recent review [21].
2.2 The Martini
Force Field
The Martini force field [2225] is a popular coarse-grained model

[26]. It has been developed to be fast, versatile, and to keep chemical
specificity. Carefully tuned building blocks are assembled in a way that
allows new molecules to be added to the model without redefining
the force field. This makes the force field easily extensible.
Martini was originally developed for lipid and surfactant systems
[22]. The force field was then extended to proteins [24, 25], carbohydrates [27], polymers [28, 29], carbon nanoparticles [30, 31]
and other molecules [26]. It has been used to study phenomena
like vesicle fusion [32], lipid phase transformations [33], or membrane tether formation [34]. Due to its speed, simulated systems
can contain up to several million particles [34]. Martini has been
built to be usable with Gromacs, but has been implemented in
several other simulation engines [35, 36]. Therefore, it can benefit
from the performance and the features of the most advanced simulation packages.
On average, MARTINI uses a 4:1 mapping: groups of four
heavy atoms (and the associated hydrogen atoms) are represented
by one interaction site (bead). Smaller beads are used to represent
rings, with a 2:1 mapping. This 4:1 mapping has been chosen as a
reasonable compromise between chemical specificity and
129
computational efficiency. Similar mapping can be found in other

coarse-grained models [37].
The beads are divided into four main types, depending on
their polarity: beads can be polar, intermediate, apolar or charged.
In addition, each main type is subdivided into subtypes to represent the hydrogen-bonding character (donor, acceptor, both donor
and acceptor, neither donor nor acceptor) or the degree of polarity
of the bead (five levels from low polarity to high polarity). This
gives a total of 18 different particle types (plus their small counterparts used in ring structures). Each particle is characterized by its
interactions with the other particles. These interactions are represented by a LennardJones 12-6 potential with the van der Waals
radius = 0.47 nm for regular beads and = 0.43 nm for small
beads. The strength of the interaction, represented by the well
depth ij, ranges from 2.0 to 5.6 kJ/mol; it is scaled by 75 % for
small beads. The potential is calculated with a cutoff at 1.2 nm and
a shift function starting at 0.9 nm. Electrostatics is represented by
a Coulomb potential screened with a relative dielectric constant
rel = 15. Like for the LennardJones potential, the Coulomb
potential is calculated with a cutoff of 1.2 nm and a shift function.
The shift function for the Coulomb potential starts at 0 nm.
Non-bonded interactions were parameterized based on the
properties of building block molecules, and particularly the free
energy of transfer between water and organic solvents. Free energy
of hydration and the free energy of vaporization were also used
during the parameterization of the force field. Bonded interactions
were parameterized to reproduce distributions of bond lengths,
angles, and dihedral angles from atomistic simulations.
The representation of water is consistent with the rest of the
model: four water molecules are represented by one Martini bead.
A polarizable model of water is also available [38]. In this more
complex model, each water particle consists of three beads: a central LennardJones bead and two partially charged beads, with
charges of opposite sign, bound to the central LennardJones bead
with a constraint. The two charged particles interact with all other
beads only via a Coulomb potential (no LennardJones interactions), and interact with each other only through an angle potential, which reduces the dipole to zero (at zero angle the two
particles are overlapped) in the absence of an electric field. In the
presence of an electric field (generated, for instance, by other
charged beads), the water charges are separated and give rise to a
dipole. Notice that use of the polarizable water model requires
some changes in the LennardJones interaction levels (see ref. 38
for details). This polarizable water model allows a better treatment
of electrostatics but has a cost in terms of computer efficiency (it is
a factor of 3 slower than the non-polarizable model).
MD simulations carried out with the Martini model are stable
with integration time steps from 20 to 40 fs (although a time step
130
of 2025 fs is preferable). The large time step, in addition to the

reduced number of degrees of freedom, and the short cutoffs used
for non-bonded potentials, lead to a speed-up of 23 orders of
magnitude compared to atomistic simulations. Time scales of tens
of microseconds and longer are computationally affordable.
The Martini model for proteins was introduced in 2008 by
Monticelli et al. [24] and updated in 2012 by de Jong et al. [25].
The model was built using the same philosophy as the rest of the
Martini force field, so it is compatible with the other Martini molecules. The four-to-one mapping is generally used (two-to-one for
ring moieties). One bead represents the main chain of each amino
acid, and up to four beads represent the side chain, depending on
the size of the amino acid. Particle types for each amino acid were
chosen based on experimental measures of wateroil partitioning
of side chain homologues. Hence, Martini reproduces accurately
the amino acid hydrophobic scale and the partitioning of amino
acids in a lipid bilayer, which makes the model especially suitable
for membrane protein simulations.
The large radius of the beads imposes a restraint on the minimum distance attainable by beads with opposite charge; this
restraint reduces very significantly the maximum strength of the
attractive electrostatic interaction. Placing the charges off the center of the LennardJones beads (on an extra particle that has no
LennardJones interactions) solves this issue. In version 2.2 of the
force field, this technique is used for charged residues, making
their mutual attraction more realistic [25].
As described above, the Martini force field reproduces in a
realistic way amino acid hydrophobicity, and it features a fairly
sophisticated treatment of side chain electrostatics. Nevertheless,
Martini cannot be used to predict protein folding, and maintaining
a stable protein fold requires restraining the secondary structure.
This is mostly due to the very simplified treatment of the protein
backbone. In addition, bonded interactions in Martini are tuned to
reproduce distances and angles from the Protein Data Bank (PDB),
and they are set based on the secondary structure. Once the secondary structure is chosen by the user (for the entire sequence of
the protein), it is maintained throughout the simulation, and all
bonded interactions can only fluctuate harmonically.
Possible applications of the MARTINI force field are the simulation of protein and lipid self-assembly, large-scale simulations of
membrane proteins, and investigations of protein-lipid and proteinprotein interactions. For instance, MARTINI can be used to characterize the tilt and orientation of transmembrane helical peptides
and their dimerization [39, 40]. In some cases, having fixed the
protein secondary structure, it is possible to simulate changes in
the protein tertiary structure triggered by external forces. For
example, Louhivuori et al. studied the relative motion of the transmembrane segments and gating in the MscL mechano-sensitive
channel in a pressurized liposome [41].
131
Sometimes, restraining the secondary structure elements is not

enough to maintain a correctly folded three-dimensional protein
structure. In this case, additional restraints to a given protein conformation can be applied with the ELNEDYN elastic network
[42]. This approach was shown to reproduce protein large-scale
motions. The computational cost of such an elastic network grows
with the number of springs involved. The SAHBNET elastic
network is an attempt to reduce the number of springs by using
information from the hydrogen bond network and the solvent
accessible surface of the residues [43].
Constrained secondary structures are one limitation of the
Martini force field. Other limitations have to be considered. For
instance, the shape of the potential used for non-bonded interactions tend to over-structure fluids compared to atomistic simulations. Other limitations are common to most coarse-grained force
fields, as they are due to the reduced number of degrees of freedom.
The four-to-one mapping reduces both spatial and chemical resolution. A four bead long alkane can represent a hexadecane molecule
as well as a pentadecane or a heptadecane. Because there are less
degrees of freedom, entropy is underestimated, and changes in
entropy might be represented unrealistically. As the model was
parameterized based on free energies, the balance between entropy
and enthalpy has to be considered with care, and it is often not correct. Dynamics is also a non-trivial issue. The free energy landscape
is smoothened by the coarse-graining, hence sampling is faster and
dynamics is also faster. Comparing diffusion of single molecules at
the coarse-grained and atomistic resolution, a fourfold speed-up has
been estimated on average. Yet, this speed-up factor is not the same
for all molecules and depends on simulation conditions.
Despite the limitations above, there is a broad range of applications for the MARTINI force field. Principles and applications of
the force field have been recently reviewed by Marrink and
Tieleman [26].
In the following, we will describe a simple procedure to simulate a rhodopsin dimer embedded in a lipid bilayer using the
Martini force field. None of the steps are specific to this particular
protein, so the procedure can be applied to any membrane protein
and any lipid membrane.
Materials
We will use the Gromacs software package to prepare, run, and
analyze an MD simulation. We will use Gromacs version 4.5 [44]
(or 4.6 but without GPU acceleration). See Note 1 about how to
use the Martini force field with Gromacs 4.6 and the GPU acceleration. Some additional third party scripts and programs need to
be installed too. We will display plots with xmgrace (http://
plasma-gate.weizmann.ac.il/Grace/). The martinize script can be
132
found on the tools section of the Martini Web site: http://

md.chem.rug.nl/cgmartini/index.php/tools2/proteins-andbilayers. The g_remove_water script is available at www.github.
com/hublot/g_remove_water, and the replace_atoms script at
www.github.com/jbarnoud/replace_atoms. DSSP [45] can be
downloaded from http://swift.cmbi.ru.nl/gv/dssp/. We assume
the programs to be installed and accessible in the default search
path, and the DSSP executable to be named mkdssp. Extensive
documentation on how to install Gromacs can be found on www.
gromacs.org.
We will use the Martini force field version 2.2 [25]. All force
field files are available on the Martini Web site. We will use the
particle description for the normal version of the force field
(martini_v2.2.itp). Note 2 details what to change in case one
wants to use the polarizable water model. We will also need the
description of lipids (martini_v2.0_lipids.itp), and the description
of ions (martini_v2.0_ions.itp). Finally, some structure samples
from the Martini Web site will be used: dopc_bilayer.gro and water.
gro, that can be found in the example applications of the Martini
Web site.
The instructions have been tested on GNU/Linux and MacOS
X. They should work on Microsoft Windows as well, after a few
minor adaptations.
Methods
We will describe how to perform a molecular dynamics simulation
of a box containing a rhodopsin dimer embedded in a dioleoylphosphatidyl-choline (DOPC) bilayer, with explicit water and ions.
The final box size will be about 13 nm 13 nm 12 nm in the X,
Y, and Z dimension, respectively. The membrane will lie in the XY
plane of that box. Besides defining the chemical content, starting
an MD simulation also requires a description of initial coordinates
and velocities of each particle. The initial velocities will be generated automatically to produce a Maxwell distribution. In Gromacs,
a so-called topology file (TOP) contains all the information on
the chemical content of the simulated system (types and number of
molecules) and on the force field used to calculate their mutual
interactions.
4.1 Preparation
of the Protein
We will use a protein structure resolved by X-ray crystallography

as a starting point to build a coarse-grained protein structure.
The structure of rhodopsin can be downloaded from the Protein
Data Bank (ID code 1L9H [46]). Rhodopsin was co-crystallized
with several ligands that we will ignore. Also, the first residue of
each chain is acetylated, and some residues are missing at the
C-terminus; we will ignore those too. Obtaining a clean atomistic
structure is beyond the scope of the present chapter.
133
Available on the Martini Web site, the martinize.py script does

the conversion from atomistic protein structures to their Martini
CG version. We will run the script on the PDB file downloaded
from the databank; the script will ignore the ligand, so we do not
need to edit the file:
martinize.py -f 1L9H.pdb -ff martini22 -dssp \
mkdssp-o CG-1L9H.top -x CG-1L9H.pdb
Using the 2.2 version of the Martini force field, the script outputs the CG coordinates in CG-1L9H.pdb and the CG topology
file in CG-1L9H.top. The topology file contains links to three files:
martini.itp, Protein_A.itp, and Protein_B.itp. The first link is a
placeholder for the force field file (that we will replace later with
the appropriate file name for the Martini version we want to use),
the two other files are generated by the martinize script and
describe the particles and the connectivity for each protein chain.
More precisely, ITP files list all the particles in each protein chain
(with particle type and partial charge), and describes how they are
bonded together, what are the distance constraints, the angles, and
the dihedral angles. The coordinates are written using the PDB
format. To convert the coordinate file from PDB format to GRO
format, used by Gromacs, we type:
trjconv -f CG-1L9H.pdb -s CG-1L9H.pdb -o CG-1L9H.gro
# We select the group 0 (System)
The protein structure can be visualized with the VMD software [47]. VMD does not know how to connect coarse-grained
beads, but there are several ways to display bonds in coarse-grained
structures. One way is to display the protein backbone with the
dynamic bonds representation. To do so, select the atoms named
BB in the Representation window and choose DynamicBonds
with a distance cutoff of 4 . This displays an approximation of the
protein backbone connectivity; this approximation is typically sufficient to visualize the main features of the protein structure and
the secondary structure elements. Another way to display the
bonds requires the use of a script named cg_bonds, which can be
found on the Martini Web site. The script cg_secondary_structure
allows displaying the secondary structures with the cartoons
representation of VMD. Figure 1 depicts the Martini CG representation of the protein, side by side with its atomistic structure.
4.2 Preparation
of the Membrane
Now that the protein is ready, the next step is to add a lipid bilayer.
This can be done with several methods. The main challenge is to
get an equilibrated membrane patch by spending little computational resources. The easiest method is to start from a preequilibrated patch. Such patches can be found on the Martini Web
site for some lipids; the instructions to change the lipid type are
134
Fig. 1 Chain A of 1L9H at the atomistic (left) and coarse-grained (right) resolution. The protein backbone is
represented in dark grey and the side chains in light grey. The coarse-grained structure has been drawn using
the script cg_bonds
given in the Martini tutorial on the same Web site. A membrane

can also be built by multiplying a single lipid and translating it on
a plane. The second leaflet can be built by rotating the first one
(not by mirroring it: the latter transformation would invert the
stereochemistry of the lipids). Some automated tools exist, like the
VMD [47] or CHARMM-GUI [48] membrane builders, but they
do not handle the Martini model. Yet another way to build a membrane patch is to let it self-assemble from randomly placed lipids.
This method is computationally expensive on atomistic systems,
but easily usable with coarse-grained models. We will use the
DOPC patch available on the Martini Web site that has been built
by self-assembly. In this patch, leaflets contain different numbers of
lipids, and the X and Y dimensions are different. In some cases,
these features can be practically inconvenient as they make the
analysis more difficult to interpret. Also, asymmetry of the leaflets
can affect some membrane properties, including the lateral pressure profile of the membrane.
The membrane patch contains water and a bilayer of 128
DOPC molecules. The upper leaflet counts 65 lipids while the
lower leaflet counts only 63 lipids. To fix the symmetry we will
remove two lipids from the upper leaflet. With a text editor,
remove lines 330 in the dopc_bilayer.gro file. Remove also all
water beads; we will add them back later. Water beads have W as
residue and atom name; see Note 3 on how a GRO file is structured.
135
The number of particles present in the file is written in the second

line. This number has to be exact as Gromacs tools use it. Since we
removed atoms, we need to update the number of particles and
change it from 3,292 to 1,764. Save the modified file as 126dopc.
gro. Finally, to make the patch square, we will rescale the Y axis,
so it will have the same length as the X axis:
editconf -scale 1 1.119142862 1 \
-f 126dopc.gro -o 126dopc_square.gro
(Notice that this operation also rescales the coordinates of all
particles in the system. We will get back to this below.)
Once rescaled, the box size is 7.02 nm in the X and Y dimensions. The radius of gyration of the protein is 2.7 nm; the maximum distance between two beads of the protein is about 8 nm,
therefore the patch is too small. We can multiply the patch on the
XY plane by running:
genconf -f 126dopc_square.gro \
-o 504dopc.gro -nbox 2 2 1
The patch is now 14.05 nm in the X and Y dimensions and
counts 504 lipids. The box is only 8.4 nm along the Z-axis. With
such small size, the protein will interact with its periodic images
along the Z direction, leading to possible artifacts in the simulations. Hence we need to increase the box size in the Z dimension.
For convenience, we will also have the membrane centered in
the box:
editconf -f 504dopc.gro -o 504dopc_rebox.gro \
-box 14.05010 14.05010 11 c
By modifying the membrane size (membrane rescaling step,
above), we perturbed bond lengths and angles in the lipids. This
can result in very high energies and possibly explosion of the system. A simple solution is to perform an energy minimization with
the steepest descent algorithm. To this end, we need to download
the Martini particles definition and the Martini topology for lipids
from the Martini Web site. The files are martini_v2.2.itp and martini_v2.0_lipids.itp. Then we need to write a topology file for the
membrane (topol.top) and a parameter file for the energy minimization (param_em.mdp).
---- topol.top ---#include "martini_v2.2.itp"
#include "martini_v2.0_lipids.itp"
[ system ]
Non-hydrated DOPC bilayer
[ molecules ]
136
DOPC
504
---------------------- param_em.mdp ---title
= Martini
integrator
= steep
nsteps
= 400
nstlist
= 10
rlist
= 1.4
coulombtype
= Shift
rcoulomb_switch
= 0.0
rcoulomb
= 1.2
epsilon_r
= 15
vdw_type
= Shift
rvdw_switch
= 0.9
rvdw
= 1.2
---------------------With Gromacs, energy minimizations and MD simulations are

performed in two steps. First a run input file has to be generated
by the Gromacs preprocessor program, named grompp. grompp
combines the initial structure (GRO), the topology file (TOP),
and the parameter file (MDP; this contains all simulation parameters), into one run input file, named TPR file. Then the simulation
engine, named mdrun, performs the energy minimization using
the TPR file as the only input:
grompp -f param_em.mdp -c 504dopc_rebox.gro \
-p topol.top -o 504dopc_em.tpr
mdrun -deffnm 504dopc_em -v
4.3 Protein Insertion
in the Membrane
Inserting the protein in the lipid membrane requires two steps.

First the protein has to be oriented with respect to the membrane;
then it has to be embedded in the lipid bilayer. Orientation consists
in rotating and translating the protein so that its transmembrane
part overlaps with the membrane. This step can be done manually
with visual tools like VMD. One can also refer to databases of
already oriented proteins, such as the Orientation of Proteins in
Membrane (OPM) [49, 50] database and the Protein Data Bank
of Transmembrane Protein (PDBTM) [5153]. These databases
use programs available on the PPM [50] and the TMDET [54]
Web servers, respectively. As an alternative, one can also use standalone programs, like Lambada [55], that can be integrated in an
automatized workflow. The PPM and TMDET Web servers and
Lambada do not handle Martini structures; yet, they can be applied
to the atomistic structures before coarse-graining.
137
Because Martini is parameterized based on partitioning data,

and thanks to the acceleration due to coarse-graining, proteins
typically rotate and assume reasonable orientations spontaneously
in CG simulations within a short time. Therefore, precise manual
orientation of the protein is not necessary. Still, by approximately
orienting the protein we can save some equilibration time.
We will place the protein at the center of a box of the same size
as the box containing the lipid membrane, with its principal axis
parallel to the membrane normal.
editconf -f CG-1L9H.gro -o orient.gro \
-box 14.05010 14.05010 11 -princ -rotate 90 0 0
# We select the group 1 (Protein)
At this point it is necessary to concatenate the structure file of
the oriented protein and the energy-minimized membrane:
cat orient.gro 504dopc_em.gro > concat.gro
Notice that the first line of a GRO file is the title of the system,
the second line is the number of particles and the last line contains
the dimensions of the box. These three lines need to appear only
once in the concatenated file. The system title is chosen by the
user; the box dimensions should be the same in the protein and
the membrane structure files, and they should remain the same in
the concatenated structure file; the number of particles has to be
consistent with the new structure file (there should be now 8,519
particles in the system). We assume the protein to be first in the
file, and we call the file concat.gro.
We now need to update the topology file (topol.top), to
include the topology for each protein chain. We also change the
system title and add the protein chains to the list of the molecules
in the system (in the [molecules] section). The topol.top file
now should look like:
#include "Protein_A.itp"
#include "Protein_B.itp"
[ system ]
Non-hydrated DOPC bilayer and rhodopsin dimer
[ molecules ]
Protein_A
Protein_B
DOPC
504
-------------------
138
We have the proteins embedded in the lipid bilayer, but the

system is not yet usable: some protein particles overlap with some
lipid particles, which can cause numerical instability (overlap comes
with very high forces) or artifacts (e.g., lipids sticking to the protein, when one of their beads is stuck inside a ring). The easiest way
to address the issue of overlap is to remove the overlapping molecules, and then carry out a short MD simulation to equilibrate the
lipids around the protein. This method is commonly used. Its main
drawback is that often the overlap involves only a few particles in
the lipid, while the rest of the molecule can be far from the protein.
In this case, removing the entire lipid molecule will result in a hole
around the protein, that will require some equilibration. Another
method to fix the issue of overlapping lipids is to run an MD
simulation with the protein replaced by a repulsive potential.
The Gromacs software package includes g_membed [56], a tool
that makes the protein grow progressively (by progressively
scaling the protein coordinates), pushing the lipids away. Kandt and
Tieleman [57] devised the inflategro method. Instead of growing
the protein, this method expands the membrane, to reduce the
number of overlapping lipids. Lipids that still overlap after the
expansion are removed, then the membrane is contracted in
successive steps until it reaches its initial size. At each contraction
step, the energy of the system is minimized. The inflategro method
has been updated in inflategro2 [55]. Here the number of lipids to
remove in each leaflet is calculated based on the initial area per lipid
and the protein cross-sectional area. Only the lipids close to the
protein are expended and contracted. Hence, the other lipids remain
as they were before the procedure.
To reduce overlap between protein and lipid particles, we will
use the inflategro2 software, available at http://code.google.
com/p/inflategro2/. The software involves an iterative procedure
that calls gromacs to minimize the energy at each contraction step.
To this end it requires two additional gromacs input files: a parameter file (MDP) and an index file (NDX). For the parameter file, we
copy the file param_em.mdp created earlier to a new file named
deflate.mdp. In this new file, we set the number of minimization
step (nsteps) to 50 and add the following lines at the end:
energygrps = Protein Dynamic Static
energygrp_excl = Protein Protein Static Static \
Dynamic Static Protein Static
freezegrps = Protein Static
freezedim
= Y Y Y Y Y Y
The index file defines groups of atoms. It can be generated by

make_ndx. We run the following command:
make_ndx -f concat.gro -o index.ndx
139
The program generates some groups automatically, and lists

them on the screen; then it prompts the user for further commands. As all the groups we need are generated automatically, so
we can type q to save and quit. Now, we can run inflategro2.
inflategro2 -f concat.gro -n index.ndx \
-p topol.top -m deflate.mdp
inflategro2 asks for a group to inflate. This group has to be the
one that defines the membrane, so we choose the group called
DOPC (group 13). As this group already contains the whole
membrane we do not need to include other groups to the selection.
Then, inflategro2 prompts for a group that contains the protein.
The last generated structure will be shrink.20.gro. [We may
encounter an issue for some simulation systems when using infltategro2 on multicore computers. See Note 4 on how to fix it.] inflategro2 updates automatically topol.top; if we have to run the program
again (for instance, if execution was interrupted), then we should
first check that the number of lipids in topol.top is correct.
The protein is now embedded in the membrane and the lipids
are reasonably well placed around it. A short MD simulation will
equilibrate the systemi.e., orient the protein correctly and
optimize the position of the lipids around the protein. The topology
was updated by inflategro2 to reflect the new number of lipids.
The current state of the system is illustrated in Fig. 2.
4.4 Addition
of the Solvent
The system is not yet hydrated. A solvent-free version of Martini is

planned [26], but the current Martini model requires explicit
solvent. So we will add explicit water to our system.
The program genbox, included in the Gromacs package, allows
to fill the empty space in a box with a solvent; the solvent it taken
from an equilibrated box; genbox also removes solvent molecules
overlapping with the solute. An equilibrated box of water can be
found on the Martini Web site. genbox does not know the size of
Martini beads, so it needs further information to calculate when
beads overlap. We can pass the approximate radius of a Martini water
bead as an argument to genbox. To hydrate the system, we type:
genbox -cp shrink.20.gro -cs water.gro \
-vdwd 0.23 -o hydrated.gro
Some water beads may be inserted in the bilayer. They would
go out of the membrane during the equilibration run, but they
may remain stuck for a while, which would make the equilibration
longer. We can remove the water beads from the hydrophobic part
of the membrane using the g_remove_water script, available at
www.github.com/hublot/g_remove_water.
g_remove_water.py--ipid_atom GL1 \
--ater_atom W -f hydrated.gro \
-o hydrated2.gro
140
Fig. 2 Membrane protein system viewed from the side (left) and from the top (right). The protein is represented
as spheres, with the backbone in dark grey and the side chains in light grey. The lipids are represented as licorice, with the polar head in dark grey and the tails in light grey. The black rectangle shows the border of the
simulation box; outside of the box is the periodic image
We notice that the net charge of our system is 3e, because the
protein is not neutral. To neutralize the system we will add sodium
ions. We can replace three water particles with sodium ions by
changing the atom name from W to NA+, and the residue
name from W to ION in the structure (GRO) file and in the
topology (TOP) file. Notice that, in the GRO file, the alignment
of the columns needs to be maintained.
It has been reported that water, in the Martini model, has a
melting temperature higher than 0 C, and it can freeze at room
temperature under certain conditions. To address this issue, version 2.0 of the model introduced a new water particle type with a
larger radius, which interacts with all non-water beads in exactly
the same way as the original water bead, but has slightly different
interaction with water beads: the sigma parameter of the Lennard
Jones interaction with water is increased to 0.57 nm; this way water
packing is perturbed and freezing at room temperature is avoided.
We will replace 5 % of the water particles by these antifreeze
water particles. As the box should contain about 9,990 water
particlesthis number can change between two runs of the procedure, we will replace 500 random water particles with antifreeze
141
water, by changing the atom name and the residue name from
W to WF. We will also reorder the atoms so that ions and
antifreeze beads are grouped; then we name the modified file
hydrated3.gro. The replace_atoms script automatizes such atom
manipulations. Using this script, replacing water beads by ions and
antifreeze water can be done with:
cat hydrated2.gro | replace_atoms.py -n 3 \
-o W -r ION -a NA+ | replace_atoms.py \
-n 500 -o W -r WF -a WF > hydrated3.gro
Now that we changed the content of the box, we need to
update the topology file. topol.top should include the description
of ions (also available on the Martini Web site). The 3 sodium ions,
the 500 antifreeze water particles, and the 9,487 remaining water
particles (this number can change) should be included in the list of
molecules. The topol.top file should now look like:
#include "Protein_A.itp"
#include "Protein_B.itp"
#include "martini_v2.0_ions.itp"
[ system ]
Hydrated DOPC bilayer and rhodopsin dimer
[ molecules ]
Protein_A
Protein_B
DOPC
NA+
WF
W
467
3
500
9487
------------------Recently, a program has been made available (on the MARTINI

Web site) to build membrane system in an automatic manner. That
program is called insane, and builds the membrane by repeating a
lipid model on a grid with the right area per lipid. The program can
also embed a protein in the membrane. Therefore it can replace
many of the steps described above. Yet, the system that is built by
insane is far from equilibrium, and one should carefully monitor
the equilibration steps to avoid possible freezing and other artifacts. The insane program is available at http://md.chem.rug.nl/
cgmartini/index.php/insane.
142
4.5 Minimize
the Energy
and Equilibrate
All the components of our systems are now in place, but the system
most likely still has high energy, due mostly to non-optimal packing of the lipids and to unfavorable lipid-protein contacts. In addition, because we used a large distance criterion to minimize water
overlap when we hydrated the box, the system density is probably
too low. As a first step towards equilibration, we will run an energy
minimization to reduce unfavorable contacts:
grompp -f param_em -c hydrated3.gro \
-p topol.top -o em
mdrun -deffnm em -v
Then we will run a short MD simulation to equilibrate the system
density. Like for the energy minimization, we need a parameter file
(MDP). We will run the simulation for 20 ns with a timestep of 20 fs.
This represents 1,000,000 integration steps. On a modern workstation, this should take about an hour. We will run the simulation at
310 K and 1 bar using the Berendsen weak coupling algorithm for
temperature and pressure [58]. This algorithm may not result in a
correct kinetic energy distribution, so we will use the ParrinelloBussi
thermostat [59] and the ParrinelloRahman barostat [60] for the
production simulation. The ParrinelloBussi and ParrinelloRahman
algorithms tend to produce high fluctuations when temperature and
pressure are too far from their target values, which makes equilibration longer. Write the param_eq.mdp file as follows:
---- param_eq.mdp ---integrator
dt
nsteps
nstcomm
nstxout
nstvout
nstfout
nstlog
nstenergy
nstxtcout
xtc_precision
nstlist
rlist
coulombtype
rcoulomb_switch
rcoulomb
epsilon_r
vdw_type
rvdw_switch
rvdw
tcoupl
tc-grps
tau_t
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
md
0.02
1000000
10
0
0
0
1000
100
1000
100
10
1.4
Shift
0.0
1.2
15
Shift
0.9
1.2
Berendsen
System
4.0
ref_t
Pcoupl
Pcoupltype
tau_p
compressibility
ref_p
gen_vel
gen_temp
constraints
constraint_algorithm
lincs_order
lincs_warnangle
----------------------
=
=
=
=
=
=
=
=
=
=
=
=
143
310
berendsen
semiisotropic
4.0
1e-5 1e-5
1.0 1.0
yes
310
none
Lincs
4
30
Now we can run the equilibration:

grompp -f param_eq -c em.gro \
-p topol.top -o eq
mdrun -deffnm eq -v
4.5.1 Is the System
Equilibrated?
It is important to verify that the system is equilibrated before starting a production run. To this end, visual inspection is a useful,
quick first step. The trajectory can be displayed with VMD [47].
During the first few steps of the trajectory, the protein tilts to find
a suitable orientation in the membrane. The box changes size as
water become denser and the area per lipid adjusts.
Visual inspection is not sufficient to determine if the system
reached equilibrium. At the end of the equilibration run, box
dimensions, potential energy and kinetic energy should have converged to stable values. Energies and box dimensions are stored in
eq.edr. They can be extracted using g_energy, and visualized with
xmgrace. Gromacs allows decomposing potential energy by groups
of atoms, called energy groups. It is then possible to check if the
non-bonded interaction between the protein and the lipids
converged. Energy groups need to be specified in the simulation
parameters before the run, though. See the Gromacs manual on
energy_grp on how to use energy groups. Other properties like
density profile or protein orientation can be checked too.
If the system did not reach equilibrium, the equilibration run
should be extended.
4.6 The
Production Run
After equilibration, we can run the actual simulation. The simulation parameter file for a production run is similar to the equilibration run, but some details need to be changed: the duration
of the run and (possibly, but not necessarily) the temperature
and pressure coupling algorithms. We will run the simulation for
1 s, e.g., 50.000.000 steps, so the nsteps parameter has to be
adapted. We will use the ParrinelloBussi thermostat (v-rescale)
and the ParrinelloRahman barostat. We copy param_eq.mdp to
144
param_md.mdp, and we edit the latter file, which has to contain

the following parameters:
tcoupl
tau_t
ref_t
Pcoupl
Pcoupltype
tau_p
compressibility
ref_p
=
=
=
=
=
=
=
=
v-rescale
1.0
310
parrinello-rahman
semiisotropic
12.0 12.0
3e-4 3e-4
1.0 1.0
We set gen_vel to no to start the simulation with the velocities generated during the equilibration as they are written in eq.gro.
Then we run the simulation:
grompp -f param_md.mdp -c eq.gro \
-p topol.top -o md
mdrun -deffnm md -v
4.7
Analysis
Once the simulation is done, one should again verify that system
size and energy do not have any drifts. Again, this can be investigated with g_energy.
Analyzing a Martini trajectory is not different from analyzing
any other MD trajectory. For example, one can look at the root
mean square deviation with g_rms or the root mean square fluctuations with g_rmsf (see Fig. 3).
Some Gromacs-related tools will prompt for a group to process, and you may want to run some analyses on the protein main
chain. In Martini, beads from the main chain are typically named
BB, so you can create group for them by using a BB as a command in make_ndx.
Notes
1. Gromacs 4.6 features an improved support for graphics
processing units (GPU). This support requires the use of a new
way to handle neighbor lists, the so-called Verlet cutoff
scheme. Shift functions are replaced by exact cutoffs, changing
the shape of the non-bonded potential. Reducing the cutoff to
1.1 nm instead of 1.2 nm seems to allow the use of this new
algorithm without affecting the properties of common systems
simulated with the Martini model. Overall, the change in the
cutoff scheme results in a speedup by almost 100 %, but at this
time the precise effects remain mostly untested.
2. Using the polarizable water model requires a few changes in
the protocol described here. First of all, we need to replace the
145
1.0
0.9
0.8
RMSD (nm)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
200
400
600
800
1000
Time (ns)
1.1
Chain A
1.0
Chain B
0.9
RMSF (nm)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
50
100
150
200
Residues
250
300
Fig. 3 Evolution of the root mean square deviation (RMSD) from the initial structure during the production
run (top panel), and root mean square fluctuation (RMSF) of the backbone, averaged over the entire production run (bottom panel). The RMSD is calculated on the whole protein after a least mean square fit on
the backbone
146
force field definition martini_v2.2.itp with martini_v2.2P.itp.

For the hydration of the system, use the box of polarizable
water (available on the Martini Web site). A script that converts normal Martini water into polarizable water is also available on the same Web site. In the MDP files, one should use a
dielectric constant of 2.5 instead of 15 (epsilon_r=2.5). We can
also use the particle mesh ewald (PME) algorithm for the calculation of electrostatic interactions (coulombtype=PME). In
this case, we would use a real space cutoff of 1.2 nm (rcoulomb=1.2), a Fourier grid spacing of 0.12 nm (fourierspacing=0.12), and a fourth-order interpolation (pme-order=4).
3. GRO files have a very strict format. The first line is the title of
the system, the second line is the number of atoms in the file,
and the last line defines the periodic box. All other lines
describe atoms. These atom lines are formatted by columns.
The residue number, the residue name, the atom name, and
the atom number are written, in this order, with five characters
each. They are followed by the X, Y, and Z coordinates of the
atom, written with eight characters each. Optionally are written the velocities for each dimension, also with eight characters
each. There is no column separator. The periodic box is defined
on the last line by at least three numbers representing the size
of the box in the X, Y, and Z dimension, respectively.
Optionally, six other values can follow. They are used to define
non-rhombohedric boxes. The format of this last line is free
and values are space separated.
4. Inflategro2 works by successive energy minimizations. Since
the systems contain large empty volumes, there may be issues
with the assumptions of the algorithm used to parallelize the
calculations. One may need to run the energy minimizations
on a single core. This requires modifications to the inflategro2
source code. In line 955 of inflategro2, replace:
"mdrun -s %s -v -deffnm %s -c %s"
by:
"mdrun -s %s -v -deffnm %s -c %s -nt 1"
Note that the option -nt 1 asks mdrun to use only one thread.
Acknowledgments
The authors thank Juliette Martin and Nicoletta Ceres for their
useful comments on the manuscript.
147
References
1. Dror RO, Dirks RM, Grossman JP, Xu H,
Shaw DE (2012) Biomolecular simulation: a
computational microscope for molecular
biology. Annu Rev Biophys 41:429452
2. Shaw DE, Chao JC, Eastwood MP, Gagliardo
J, Grossman JP, Ho CR et al (2008) Anton, a
special-purpose machine for molecular dynamics simulation. Commun ACM 51:91
3. Lindorff-Larsen K, Piana S, Dror RO, Shaw
DE (2011) How fast-folding proteins fold.
Science 334:517520
4. Bussi G, Laio A, Parrinello M (2006)
Equilibrium free energies from non-equilibrium
metadynamics. Phys Rev Lett 96:090601
5. Sugita Y, Okamoto Y (1999) Replica-exchange
molecular dynamics method for protein folding. Chem Phys Lett 314:141151
6. Torrie GM, Valleau JP (1977) Nonphysical
sampling distributions in Monte Carlo freeenergy estimation: umbrella sampling. J Comput
Phys 23:187199
7. Carbone P, Varzaneh HAK, Chen X, MllerPlathe F (2008) Transferability of coarsegrained force fields: the polymer case. J Chem
Phys 128:064904
8. Levitt M, Warshel A (1975) Computer simulation of protein folding. Nature 253:694698
9. Liwo A, Oldziej S, Pincus MR, Wawak RJ,
Rackovsky S, Scheraga HA (1997) A unitedresidue force field for off-lattice proteinstructure simulations. I. Functional forms and
parameters of long-range side-chain interaction
potentials from protein crystal data. J Comput
Chem 18:849873
10. Maupetit J, Tuffery P, Derreumaux P (2007) A
coarse-grained protein force field for folding
and structure prediction. Proteins 69:394408
11. Bereau T, Deserno M (2009) Generic coarsegrained model for protein folding and aggregation. J Chem Phys 130:235106
12. Pasi M, Lavery R, Ceres N (2013) PaLaCe: a
coarse-grain protein model for studying
mechanical properties. J Chem Theory Comput
9:785793
13. Zacharias M (2003) Protein-protein docking
with a reduced protein model accounting for
side-chain flexibility. Protein Sci 12:
12711282
14. Setny P, Zacharias M (2011) A coarse-grained
force field for Protein-RNA docking. Nucleic
Acids Res 39:91189129
15. Den Otter WK, Renes MR, Briels WJ (2010)
Asymmetry as the key to clathrin cage assembly. Biophys J 99:12311238
16. Matthews R, Likos CN (2013) Structures and

pathways for clathrin self-assembly in the bulk
and on membranes. Soft Matter 9:57945806.
doi:10.1039/c3sm50737h
17. Shi Q, Izvekov S, Voth GA (2006) Mixed
atomistic and coarse-grained molecular dynamics: simulation of a membrane-bound ion channel. J Phys Chem B 110:1504515048
18. Rzepiela AJ, Louhivuori M, Peter C, Marrink
SJ (2011) Hybrid simulations: combining
atomistic and coarse-grained force fields using
virtual sites. Phys Chem Chem Phys 13:
1043710448
19. Zacharias M (2013) Combining coarse-grained
nonbonded and atomistic bonded interactions
for protein modeling. Proteins 81:8192
20. Taylor WR, Katsimitsoulia Z (2010) A coarsegrained molecular model for actin-myosin simulation. J Mol Graph Model 29:266279
21. Praprotnik M, Delle Site L (2013) Multiscale
molecular modeling. Methods Mol Biol
924:567583
22. Marrink SJ, de Vries AH, Mark AE (2004)
Coarse grained model for semiquantitative lipid
simulations. J Phys Chem B 108:750760
23. Marrink SJ, Risselada HJ, Yefimov S, Tieleman
DP, De Vries AH (2007) The MARTINI force
field: coarse grained model for biomolecular
simulations. J Phys Chem B 111:78127824
24. Monticelli L, Kandasamy SK, Periole X, Larson
RG, Tieleman DP, Marrink S-J (2008) The
MARTINI coarse-grained force field: extension to proteins. J Chem Theory Comput
4:819834
25. De Jong DH, Singh G, Bennett WFD, Arnarez
C, Wassenaar TA, Schfer LV et al (2013)
Improved parameters for the martini coarsegrained protein force field. J Chem Theory
Comput 9:687697
26. Marrink SJ, Tieleman DP (2013) Perspective
on the Martini model. Chem Soc Rev 42:6801
6822. doi:10.1039/c3cs60093a
27. Lpez CA, Rzepiela AJ, de Vries AH,
Dijkhuizen L, Hnenberger PH, Marrink SJ
(2009) Martini coarse-grained force field:
extension to carbohydrates. J Chem Theory
Comput 5:31953210
28. Rossi G, Fuchs PFJ, Barnoud J, Monticelli L
(2012) A coarse-grained MARTINI model of
polyethylene glycol and of polyoxyethylene
alkyl ether surfactants. J Phys Chem B 116:
1435314362
29. Rossi G, Monticelli L, Puisto SR, Vattulainen I,
Ala-Nissila T (2011) Coarse-graining polymers
148
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.

with the MARTINI force-field: polystyrene as
a benchmark case. Soft Matter 7:698
Wong-Ekkabut J, Baoukina S, Triampo W,
Tang I-M, Tieleman DP, Monticelli L (2008)
Computer simulation study of fullerene translocation through lipid membranes. Nat
Nanotechnol 3:363368
Monticelli L (2012) On atomistic and coarsegrained models for C60 fullerene. J Chem
Theory Comput 8:13701378
Marrink SJ, Mark AE (2003) The mechanism
of vesicle fusion as revealed by molecular
dynamics simulations. J Am Chem Soc 125:
1114411145
Risselada HJ, Marrink SJ (2008) The molecular face of lipid rafts in model membranes. Proc
Natl Acad Sci U S A 105:1736717372
Baoukina S, Marrink SJ, Tieleman DP (2012)
Molecular structure of membrane tethers.
Biophys J 102:18661871
Baron R, Trzesniak D, de Vries AH, Elsener A,
Marrink SJ, van Gunsteren WF (2007)
Comparison of thermodynamic properties of
coarse-grained and atomic-level simulation
models. ChemPhysChem 8:452461
Shih AY, Arkhipov A, Freddolino PL, Schulten
K (2006) Coarse grained protein-lipid model
with application to lipoprotein particles. J Phys
Chem B 110:36743684
Shinoda W, DeVane R, Klein ML (2010)
Zwitterionic lipid assemblies: molecular
dynamics studies of monolayers, bilayers, and
vesicles using a new coarse grain force field.
J Phys Chem B 114:68366849
Yesylevskyy SO, Schfer LV, Sengupta D,
Marrink SJ (2010) Polarizable water model for
the coarse-grained MARTINI force field. PLoS
Comput Biol 6:e1000810
Monticelli L, Tieleman DP, Fuchs PFJ (2010)
Interpretation of 2H-NMR experiments on the
orientation of the transmembrane helix
WALP23 by computer simulations. Biophys J
99:14551464
Castillo N, Monticelli L, Barnoud J, Tieleman
DP (2013) Free energy of WALP23 dimer
association in DMPC, DPPC, and DOPC
bilayers. Chem Phys Lipids 169:95105
Deplazes E, Louhivuori M, Jayatilaka D,
Marrink SJ, Corry B (2012) Structural
Investigation of MscL gating using experimental data and coarse grained MD simulations.
PLoS Comput Biol 8:e1002683
Periole X, Cavalli M, Marrink S-J, Ceruso MA
(2009) Combining an elastic network with a
coarse-grained molecular force field: structure,
dynamics, and intermolecular recognition. J
Chem Theory Comput 5:25312543
43. Dony N, Crowet JM, Joris B, Brasseur R, Lins

L (2013) SAHBNET, an accessible surfacebased elastic network: an application to membrane protein. Int J Mol Sci 14:
1151011526
44. Pronk S, Pll S, Schulz R, Larsson P, Bjelkmar
P, Apostolov R et al (2013) GROMACS 4.5: a
high-throughput and highly parallel open
source
molecular
simulation
toolkit.
45. Kabsch W, Sander C (1983) Dictionary of protein secondary structure: pattern recognition
of hydrogen-bonded and geometrical features.
Biopolymers 22:25772637
46. Okada T, Fujiyoshi Y, Silow M, Navarro J,
Landau EM, Shichida Y (2002) Functional role
of internal water molecules in rhodopsin
revealed by X-ray crystallography. Proc Natl
Acad Sci U S A 99:59825987
47. Humphrey W, Dalke A, Schulten K (1996)
VMD visual molecular dynamics. J Mol
Graph 14:3338
48. Jo S, Lim JB, Klauda JB, Im W (2009)
CHARMM-GUI Membrane Builder for mixed
bilayers and its application to yeast membranes.
Biophys J 97:5058
Mosberg HI (2006) OPM: orientations of proteins in membranes database. Bioinformatics
22:623625
50. Lomize MA, Pogozheva ID, Joo H, Mosberg
40:D370D376
51. Tusndy GE, Dosztnyi Z, Simon I (2004)
Transmembrane proteins in the Protein Data
Bank:
identification
and
classification.
PDB_TM: selection and membrane localization of transmembrane proteins in the protein data bank. Nucleic Acids Res 33:
D275D278
53. Kozma D, Simon I, Tusndy GE (2013)
PDBTM: protein data bank of transmembrane
proteins after 8 years. Nucleic Acids Res
41:D524D529
TMDET: web server for detecting transmembrane regions of proteins by using their 3D
coordinates. Bioinformatics 21:12761277
55. Schmidt TH, Kandt C (2012) LAMBADA and
InflateGRO2: efficient membrane alignment
and insertion of membrane proteins for molecular dynamics simulations. J Chem Inf Model
52:26572669

56. Wolf MG, Hoefling M, Aponte-Santamara C,
Grubmller H, Groenhof G (2010) g_membed: efficient insertion of a membrane protein
into an equilibrated lipid bilayer with minimal
perturbation.
J
Comput
Chem
31:
21692174
57. Kandt C, Ash WL, Tieleman DP (2007)
Setting up and running molecular dynamics
simulations of membrane proteins. Methods
41:475488
149

WF, DiNola A, Haak JR (1984) Molecular
dynamics with coupling to an external bath.
J Chem Phys 81:3684
Canonical sampling through velocity rescaling.
60. Parrinello M, Rahman A (1981) Polymorphic
transitions in single crystals: a new molecular
dynamics method. J Appl Phys 52:7182
Chapter 8
Tackling Sampling Challenges inBiomolecular Simulations
AlessandroBarducci, JimPfaendtner, andMassimilianoBonomi
Abstract
Molecular dynamics (MD) simulations are a powerful tool to give an atomistic insight into the structure
and dynamics of proteins. However, the time scales accessible in standard simulations, which often do not
match those in which interesting biological processes occur, limit their predictive capabilities. Many
advanced sampling techniques have been proposed over the years to overcome this limitation. This chapter
focuses on metadynamics, a method based on the introduction of a time-dependent bias potential to accelerate sampling and recover equilibrium properties of a few descriptors that are able to capture the complexity of a process at a coarse-grained level. The theory of metadynamics and its combination with other
popular sampling techniques such as the replica exchange method is briefly presented. Practical applications of these techniques to the study of the Trp-Cage miniprotein folding are also illustrated. The examples contain a guide for performing these calculations with PLUMED, a plugin to perform enhanced
sampling simulations in combination with many popular MD codes.
Key words Enhanced sampling, Metadynamics, PLUMED, Replica exchange methods, Molecular
dynamics, Collective variables, Free energy
1 Introduction
Uniquely providing insights into the structure and dynamics of
complex biomolecular systems at the atomistic level, MD simulations can play a fundamental role in molecular biology.
Unfortunately, the large numbers of particles that are needed for
an accurate model of biomolecules and the complexity of their
free-energy landscape make simulations computationally expensive
and prevent exhaustive sampling by standard MD in all but the
simplest cases. Recently, the development of dedicated hardware
[1] and distributed computing protocols [2] has in part alleviated
these issues. Nevertheless, the time scales accessible to MD are still
significantly shorter than those typical of several interesting biomolecular processes as well as of many experimental techniques.
To extend the time scales of MD simulations, several advanced
sampling methods have been proposed over the years [3, 4]. Acomprehensive review of such methods is beyond the scope of this chapter.
151
152
Alessandro Barducci et al.
Here we focus on metadynamics [5] (MetaD), which is able at the

same time to enhance sampling and reconstruct the free-energy
surface (FES) as a function of a few selected degrees of freedom.
Since its introduction in 2002, MetaD has been successfully applied
to a variety of different problems in chemistry, biology, and solidstate physics [6]. Furthermore, MetaD can be seamlessly integrated
with other advanced sampling algorithms, such as replica exchange
method [7, 8] (REM), with immense synergistic benefits when
studying complex biomolecular processes in large systems.
In this chapter, we present a concise introduction to the theory
of MetaD and combined methods and a few practical applications
to the study of Trp-Cage miniprotein [9] folding. We also provide
all the scripts needed to run these examples with PLUMED [10], a
plugin that enables enhanced sampling calculations with several
popular MD codes. The chapter is organized as follows. In
Subheading2, we illustrate the basic theory of: (1) MetaD in its
well-tempered variant [11] (Subheading2.1), (2) REM, with a
particular focus on parallel tempering (PT) (Subheading2.2), (3)
possible combinations of these methods (Subheading2.3). In
Subheading3, we list the software used to carry out the simulations
described in this chapter. In Subheading4, we provide a step-by-
step guide to the following simulations of the Trp-Cage miniprotein: MetaD (Subheading4.2), PT-MetaD (Subheading4.3), and
PT-MetaD in the Well-Tempered Ensemble [12] (WTE)
(Subheading4.4). Finally, Subheading5 contains a series of notes
on detecting and solving practical issues that can rise when using
the sampling techniques presented in this chapter.
2 Theory
2.1 Metadynamics
In MetaD, an external history-dependent bias potential is constructed in the space of a few selected degrees of freedom, generally called collective variables (CVs). CVs are functions S of the
microscopic coordinates R of the system:
S ( R ) = (S1 ( R ) , , Sd ( R ) ) ,
(1)

which are able to provide a coarse-grained description of the process under study. In particular, CVs must distinguish the relevant
states of the system and include all the kinetically relevant degrees
of freedom. The MetaD bias potential (VG) can be written as a sum
of Gaussians deposited along the system trajectory in the CVs
space. In the well-tempered approach [11], VG has the following
functional form at time t:
d S R S R t
) i ( )
i (
VG (S ,t ) = dt (t ) exp
2 i2
0
i =1
))
, (2)
153
where i is the width of the Gaussian for the ith CV.The time-
dependent energy rate (t) is defined as:
V (S ,t )
(t ) = 0 exp G
,
kB T
(3)
where 0 is an initial deposition rate, kB is the Boltzmann constant,

and T is an input parameter with the dimension of a
temperature.
It has been proven [11] that, in the long time limit, VG converges to:

VG (S , t ) =
T
F (S ) + C ,
T + T

(4)
where T is the temperature of the system and C is an irrelevant

additive constant. F(S) is the free energy as a function of the CVs:
U ( R )
F (S ) = kBT log dR (S S ( R ) ) exp
,
kBT
(5)
where U(R) is the potential energy function. Equation4 is often

expressed in terms of the so-called bias-factor =(T+T)/T as:

VG (S , t ) = (1 1 ) F (S ) + C .
(6)

In the long-time limit, the CVs probability density P(S, t) can be
written as:
F (S )
P (S , t ) exp
,
kB (T + T )
(7)
which corresponds to sampling the free-energy surface (FES) at

the fictitious temperature of T+T. The extent of FES exploration
can thus be regulated by tuning T.
To understand the effect of VG, let us consider a system whose
dynamics can be captured by a one-dimensional free energy, with
local minima separated by barriers much higher than thermal fluctuations (Fig.1). In a standard MD simulation, the system would
remain trapped in one of the local minima and it would not be
possible to sample all the relevant regions in a reasonable simulation time. In a well-tempered MetaD simulation, Gaussians are
progressively deposited along the trajectory resulting in the growth
of the bias potential, which ultimately facilitates barrier crossing.
Ifthe parameters are properly chosen, the system will eventually
sample all the relevant free-energy minima and VG will converge.
At that point, the free-energy profile can be estimated using Eq.4.
One notable case of well-tempered MetaD corresponds to the
choice of the potential energy of the system as CV [12]. In this
situation, the MetaD bias leads to the sampling of a well-defined
154
Fig. 1 Well-tempered MetaD simulation in a one-dimensional model system. (a) The free-energy profile (thick
black line) is characterized by three local minima separated by energy barriers higher than kBT. The sum of the
underlying free energy and the bias potential is shown at different times of the simulation (grey lines). The bias
potential is rescaled by the bias factor, following Eq.6. (b, c) Time series of the CV S (b) and of the Gaussian
height w (c) in the first 10,000 steps of simulation. The system is prepared in S=1. As the bias potential
grows, the Gaussian height w decreases, following the well-tempered recipe of Eq.3. Around t=200, the
system escapes from the initial basin into the first minimum on the right. At this point the deposition rate suddenly increases (c), as expected in well-tempered MetaD when visiting a previously unexplored region of the
CV space. After this basin is completely filled, the system starts diffusing between the first and second minima
(800<t<1,500). When a sufficient amount of bias is accumulated, the system is pushed to visit the third basin
on the right and the deposition rate increases again. Finally, around t=4,500 the underlying free energy is
almost completely compensated by the bias potential. At this point, the system starts diffusing smoothly in the
CV space, while the Gaussian height is progressively decaying to zero
distribution called Well-Tempered Ensemble (WTE). In this

ensemble, the average energy remains close to the canonical value
but its fluctuations are enhanced in a tunable way, thus improving
sampling.
In well-tempered MetaD, the bias deposition rate decreases
with time as 1/t. The dynamics of all the microscopic variables
thus becomes progressively closer to thermodynamic equilibrium
as the simulation proceeds. This feature allows to easily recover the
equilibrium Boltzmann distribution of degrees of freedom other
than the CVs, which typically is altered by the introduction of VG.
Indeed, a simple reweighting scheme has been proposed [13] in
order to obtain on-the-fly estimates of expectation values of any
variable during a well-tempered MetaD simulation. This algorithm
significantly extends the capabilities of MetaD by allowing a quantitative contact between biased simulations and experiments [14].
2.2 Replica
ExchangeMethod
In REM [7, 8], sampling is accelerated by properly modifying the

original Hamiltonian of the system. This goal is achieved by simulating in parallel N non-interacting replicas of the system, each
155
Fig. 2 Schematic representation of the PT scheme. (a) N=5 independent copies of the system are simulated
at different temperatures. Periodically, an exchange between replicas at different temperatures (typically
neighbors) is proposed and accepted based on the Metropolis criteria defined in Eqs.8 and 9. The time needed
to span the entire temperature range, i.e., the round-trip time, is often used as measure of PT efficiency. (b)
Quasi-Gaussian potential energy distributions at different temperatures, as typically observed in simulations
of proteins in explicit solvent. The PT acceptance probability ultimately depends on the overlap of the potential
energy distributions at two temperatures. The number of replicas needed to guarantee a similar overlap at a
fixed temperature range scales as the square root of the number of degrees of freedom, thus making PT computationally prohibitive in large systems
evolving according to a different Hamiltonian. At fixed intervals,

an exchange of configurations between two replicas is attempted
while respecting detailed balance (Fig.2a). One popular case of
REM is parallel tempering (PT), in which replicas are simulated
using the same potential energy function, but different
temperatures. By accessing high temperatures, replicas are prevented from being trapped inlocal minima. In PT, exchanges are
usually attempted between adjacent temperatures with the following acceptance probability:
)}
p ( j k ) = min 1, exp jPT

,
,k
(8)
with
1
1
jPT
,k =
kBT j kBTk
(9)
U R j U ( Rk ) ,

where Rj and Rk are the configurations at temperature Tj and Tk,
respectively. Equation9 indicates that the acceptance probability is
ultimately determined by the overlap between the energy distributions of two replicas (Fig.2b).
One advantage of PT is that there is no need to select a priori
an arbitrary set of CVs, once the temperatures are chosen. However,
the efficiency of the algorithm depends on the benefits provided by
( ( )
156
sampling at high-temperature. Therefore, an efficient diffusion in

temperature space is required and configurational sampling is still
limited by entropic barriers. Finally, PT scales poorly with system
size. In fact, a sufficient overlap between the potential energy distributions of neighboring temperatures is required in order to
obtain a significant diffusion. Therefore, the number of temperatures needed to cover a given temperature range scales as the
square root of the number of degrees of freedom, making this
approach prohibitively expensive for large systems.
2.3 Combined
Approaches
MetaD and REM have been successfully applied to characterize a

variety of systems [6, 15]. However, two issues have limited the
efficiency of these two sampling algorithms in simulating large biomolecular systems. First, finding a small set of CVs that captures
the slow dynamics of complex conformational changes might be a
daunting task in MetaD.Second, the size of these systems makes
multi-replica approaches computationally demanding, especially
when using an explicit solvent model. These limitations can be alleviated by properly combining different sampling approaches.
In this spirit, one can apply the MetaD bias on some selected
CVs within a multi-replica PT scheme. In the resulting PT-MetaD
algorithm [16], N replicas performed in parallel a MetaD simulation at different temperatures, using the same set of configurational CVs. The REM acceptance probability is modified in order
to account for the presence of the MetaD bias potential:
MetaD
jPT
= jPT
,k
,k +
(( ) )
1 (j)
j
VG S R j , t VG( ) (S ( Rk ) , t )
kBT j
(( ) )
1 (k )
k
VG (S ( Rk ) , t ) VG( ) S R j , t ,
kBTk
(10)

where VG(j) and VG(k) are the bias potentials acting on the j-th and
k-th replicas, respectively.
PT-MetaD is particularly effective because it compensates for
some of the weaknesses of each method individually taken. The
negative effect of neglecting a slow degree of freedom in the choice
of the MetaD CVs is alleviated by PT, which allows the system to
cross moderately high free-energy barriers on all degrees of freedom. On the other hand, the MetaD bias potential allows crossing
higher barriers on a few selected CVs, in such a way that the sampling efficiency of PT-MetaD is greater than that of PT alone.
Nevertheless, PT-MetaD still suffers from the poor scaling of
computational resources with system size. This issue may be circumvented by including the potential energy of the system among
the set of MetaD CVs, as in the WTE approach [12]. This leads to
the so-called PT-MetaD-WTE scheme [17], in which replica diffusion in temperature space is enhanced by the increased energy fluctuations at all temperatures.
157
3 Materials
Simulations of the Trp-Cage miniprotein have been carried out
[17] using GROMACS version 4.5.3 [18] and the PLUMED plugin version 1.2.2 [10]. However, for didactical purposes the scripts
reported here have been updated to PLUMED version 2 [19].
Figures have been prepared with UCSF Chimera [20] and
Matplotlib [21]. All simulations should be run in parallel on a cluster machine. The reader should refer to GROMACS and PLUMED
user manuals for detailed instructions about how to compile and
execute the codes.
4 Methods
In this section we show a few applications of the enhanced sampling techniques introduced above to study the Trp-Cage miniprotein folding, using an atomistic description of both solute and
solvent degrees of freedom. Trp-cage is a 20-residue protein whose
structure has been determined by NMR [9] (PDB code 1L2Y),
and its folding process has been extensively studied by several
experimental [9, 2225] and computational [2631] techniques.
This section is organized as follows. In Subheading4.1, we
describe the steps needed to prepare and equilibrate the system. In
Subheading4.2, we illustrate a simple MetaD simulation that uses
2 CVs. In Subheadings4.3 and 4.4, we combine MetaD with a
multi-replica approach (PT) and with WTE, respectively. In
Subheading4.5, we present a quantitative analysis of the convergence of the simulations along with an estimate of the error in the
reconstructed FES.
4.1 System
Preparation
The initial structure of Trp-Cage is taken from Protein Data Bank

(PDB code 1L2Y). The GROMACS tools are used to create the
topology using the AMBER99-SB [32] force field and solvate the
protein in a truncated octahedral box containing 3,717 TIP3P
[33] water molecules. The initial conformation is first energy-
minimized by 5,000 steps of steepest descent and then equilibrated
at 300K and 1atm by a 1ns simulation in the isothermal-isobaric
ensemble (NPT). The final volume of the cell is 4.9nm3. To generate an initial unfolded structure, we heat the system at 600K for
1ns in the isothermal ensemble (NVT).
4.2 MetaD
Our goal is to study the folding landscape of Trp-Cage. Although this

is a small protein, its folding is a complex molecular process which
involves many degrees of freedom and a wide range of spatial and
time scales. However, in a MetaD simulation it is very inefficient to
simultaneously bias a large number of degrees of freedom due to the
158
Fig. 3 Definition of the MetaD CVs used to study the folding of the Trp-Cage
miniprotein. (a) S counts the number of hydrogen bonds (black dotted lines)
formed in and between the -helical regions (orange cartoon) of the NMR native
structure (b). Shc counts the number of contacts in the hydrophobic core, defined
here by residues Y3, W6, P12, and P18 (ball and stick)
exponential increase of the CVs space. Therefore, most applications

of MetaD have been carried out using typically 2 or 3 CVs that were
devised to capture the relevant modes of the process under study.
Here, we use 2 CVs that coarse-grain the most important driving forces in protein folding: the formation of the secondary structure and of the hydrophobic core (Fig.3). In order to define the
CVs, we take advantage of the following switching function, provided by the PLUMED plugin, that can be parameterized to quantify the formation of a contact between the two atoms i and j:
n
rij
1
r0 ,
s ij =
m
rij
1
r0
(11)

where rij is the distance between the two atoms, r0 is a characteristic contact distance, and the pair n, m defines the steepness of the
CV (see Note 1). The first CV (S) describes the number of
backbone-backbone -helical hydrogen bonds formed (Fig.3a):
N H NO
S = s ij ,
(12)

where r0=0.25 nm, n=8, m=12, and the sums are over the hydrogen and oxygen atoms that form an -helical hydrogen bond in the
native state. The second CV (Shc) describes the number of contacts
in the hydrophobic core (Fig.3b):
i =1 j =1
S hc =

i>j,
s ij ,
i , j core
(13)

159
where r0=0.50 nm, n=8, and m=12, and the sum is over representative side-chain Carbon atoms of the residues belonging to the
hydrophobic core (Y3, W6, P12, and P18).
The MetaD bias potential is constructed using an initial deposition rate equal to 2.5kJ/mol every 0.5ps. Each Gaussian has a
width equal to 0.4 for both CVs. The bias factor is set to 8 (see
Note 2). The following input file can be used to run the MetaD
simulation with PLUMED 2:
The groups are defined in terms of the atom indices in 1L2Y.

pdb. The units of measurement are: kJ/mol for energy, nm for
distances, K for temperature, and number of MD steps for time
(here the time step is set to 2fs). The MetaD simulation should be
run following the standard GROMACS instructions, by adding
the flag that activates PLUMED to the command line:
where plumed.dat is the name of the PLUMED input file defined

above.
We analyze the results of the MetaD simulation by examining
the evolution of the system in CV space along with the deposition
rate of the bias potential (Fig.4). The system starts from a configuration with an unstructured helix (S~2.6) and a collapsed hydrophobic core (Shc~8.0). In a few nanoseconds, Trp-cage is pushed
to break all the hydrogen bonds and to explore a region of configurational space with a completely broken helix (S~0).
Afterwards, the systems seems to be stuck in this region for ~100ns
although it samples a broad range in the Shc space. A significant
amount of bias is accumulated in this part of the simulation as confirmed by its deposition rate that rapidly decreases following Eq.3.
After 100ns, the system suddenly forms hydrogen bonds and
escapes this region towards more native-like conformation with
160
Fig. 4 MetaD simulations of the Trp-Cage miniprotein. Time series of the CVs (a) S and (b) Shc, along with the
Gaussian height (c). The 2 CVs seems to fail in describing all the relevant slow modes of the system, since we
do not observe a smooth exploration of the CVs space in the time scale of this simulation
high number of -helical hydrogen bonds and hydrophobic contacts. At this stage, the deposition of the bias potential suddenly
increases again. During the rest of the simulation, the trajectory
visits conformations with variable number of -helical hydrogen
bonds and hydrophobic contacts, but never returns to the region
of completely unstructured configurations that has been explored
earlier (S<0.5).
The difficulty of the bias potential in pushing Trp-Cage back
and forth from one conformational region to the other is a symptom
that our 2 CVs do not capture all the relevant modes of the system.
Differently, we would observe a smoother exploration of the CV
space and a quasi-diffusive dynamics upon convergence of the bias
potential. This outcome is not totally unexpected, since the configurational ensemble of Trp-Cage is extremely wide and characterized
by several metastable states that these CVs seem not to properly
describe. At this point one can adopt several different strategies:
1. Complementing the existing set of CVs with additional CVs;
2. Devising more effective CVs. A possible choice include CVs
that are able to describe the collective behavior of the folding
process, such as the Path Collective Variables [34] or CVs based
on dimensionality reduction [35, 36];
3. Biasing separately different CVs in a bias-exchange MetaD

approach [37];
4. Combining MetaD with other sampling algorithms.
In the following subsection, we show how to combine MetaD
in S and Shc with PT and the resulting benefit to sampling
efficiency.
4.3 PT-MetaD
To setup a PT-MetaD simulation, we need to choose the replicas

temperature distribution (see Note 3) and equilibrate each replica
prior to perform the MetaD simulation (see Notes 4 and 5).
161
Wethus perform short 500ps simulations in the NVT ensemble,

so that at the end of equilibration the C root-mean-square deviation (RMSD) from the lowest-energy NMR structure ranges from
4 to 9 across the replicas.
We use the same set of CVs (S-Shc) introduced in
Subheading4.2. One PLUMED input file per replica as the one
described above should be prepared (plumed.dat.0, plumed.
dat.1,), in which the correct temperature is specified in the line
of the METAD keyword. We have to follow the GROMACS
instructions to activate PT, in this case using 100 replicas and
exchanging every 100 steps [38]:
The first thing to check in a PT-MetaD simulation is the correctness of the REM setup, in particular the temperature distribution. This can be done by analyzing both the average acceptance
rate between neighboring temperatures (reported in the
GROMACS log file) and the overall trajectory of each replica in
temperature space (Fig.5a, see Notes 6 and 7).
As for the MetaD simulation, we examine the evolution of the
system in the CV space along with the deposition rate of the bias
potential at 300K (Fig.6). At variance with single replica MetaD, this
is not a continuous trajectory, due to the exchanges with other temperatures (see Note 6). It is clear from this analysis that the statistics
accumulated at 300K spans the entire range of CVs space throughout
the simulation (Fig.6a,b). This is confirmed by the smooth average
decrease of the bias deposition rate over the simulation time (Fig.6c).
Fig. 5 Temperature diffusion of a representative replica in PT-Metad (a) and PT-MetaD-WTE (b). Time is measured per replica. Despite having only 10 intermediate temperatures instead of 100 to cover the same temperature range (300600K), the diffusion of PT-MetaD-WTE appears to be smooth thanks to the static bias on
energy. The average round-trip times of PT-Metad and PT-MetaD-WTE are 6.3ns and 4.0ns, respectively
162
Fig. 6 PT-MetaD simulation of the Trp-Cage miniprotein. (Discontinuous) time series of (a) S and (b) Shc at
300K, along with the Gaussian height (c). Time is measured per replica. Thanks to the exchange with other
temperatures, an exhaustive sampling of the CVs space is achieved
Fig. 7 PT-MetaD simulation of the Trp-Cage miniprotein. (Continuous) time series of (a) S and (b) Shc for a
representative replica diffusing in temperature. Time is measured per replica. It is crucial to reconstruct the
continuous trajectories of the replicas to assess whether the excursions in temperature do lead to an exhaustive sampling of the CVs space
Even if the behavior is reassuring of the correctness of the simulation setup, further analysis is required to declare convergence.
Indeed, due to the REM protocol, 100 trajectories contribute to
the statistics reported in Fig.6. Since all the replicas are prepared
in different regions of the CV space, the exhaustive sampling
observed in Fig.6 could result from the exchanges between replicas, which individually still suffer from sampling problems.
Therefore we have to check the trajectories of each individual replica across temperatures to assess the diffusion in the CVs space
(Fig.7 and see Note 6). This step is crucial to verify whether diffusion in temperature space can effectively help the system crossing
barriers in degrees of freedom not included in the CVs.
4.4 PT-MetaD-WTE
To reduce the number of replica needed to cover a given temperature range, we couple PT-MetaD with WTE.Enlarging the energy
163
fluctuations by means of WTE leads to an increase overlap between

potential energy distributions, thus allowing the use of a larger
spacing between temperatures. Here we setup a simulation using
only ten replicas and choosing the WTE parameters in order to
achieve diffusion in temperature comparable to the PT-MetaD
simulation described above (see Note 7).
In typical solvated biomolecular systems, the PT-MetaD-WTE
protocol is best carried out in two steps (see Note 8). First, the bias
in the potential energy space is converged at each temperature, so
that the WTE is sampled at the end of this preliminary step. For
this simulation, we use an initial deposition rate equal to 2.0kJ/
mol every 0.25ps. Each Gaussian has a width equal to 500kJ/
mol. The bias factor is set to 24. The following input file can be
used to run the WTE simulation with PLUMED 2:
In Fig.8a we show the trajectory in energy space along a preliminary 0.5ns simulation in the NVT ensemble, followed by a
Fig. 8 Sampling WTE at two representative temperatures (300 and 324K). (a) Time series of the potential
energy during a 0.5ns preliminary NVT simulation, followed by a 1ns MetaD simulation used to converge the
bias on the potential energy. While in the NVT part of the trajectory, the potential energy distributions at the two
temperatures are well separated, in the WTE ensemble a significant overlap is obtained. (b, c) Time series of
the ratio of the average potential energy (b) and fluctuations (c) to the canonical values. At the end of the simulation, these ratios are close to the theoretical WTE values (=24)
164
1ns MetaD simulation to converge the WTE bias at two representative temperatures. At the end of the latter simulation, the potential energy averages and fluctuations are close to the theoretical
WTE values (Fig8b and see Note 9).
In a second step, we run a PT-MetaD-WTE simulation using a
history-dependent two-dimensional MetaD bias on S and Shc and
a static bias on the energy. The latter has been stored on a grid and
written to file (BIAS) at the end of the preliminary WTE run
described above (see Note 10). The following input file can be
used to run the PT-MetaD-WTE simulation with PLUMED 2:
As done for the PT-MetaD simulation, we analyze the overall

trajectory of each replica in temperature space (Fig.5b and see
Note 6). Despite having fewer replicas, diffusion in temperature is
similar to that observed in PT-MetaD, thanks to the bias on energy.
To make a more quantitative comparison, we calculate the average
round-trip time , i.e., the time needed for a replica to move from
the lowest to the highest temperature and back. This analysis shows
that the two setups result into a very similar behavior (PT
=4.0ns). Finally, the evolution in the
MetaD=6.2 ns, PTMetaDWTE
S-Shc space can be checked using the analysis described in the previous section.
4.5 Analysis
oftheFES:
Convergence andError
Estimate
The FES as a function of the MetaD CVs can be calculated by integrating the Gaussians deposited along the simulation after proper
rescaling (see Note 11). For the PT-MetaD-WTE simulations, an
additional step needs to be performed, i.e., the removal of the
effect of the static bias on energy. This can be easily done by applying a TorrieValleau correction [39] to the statistics accumulated
in the WTE ensemble.
165
Fig. 9 Trp-Cage miniprotein FES from the PT-MetaD (a) and PT-MetaD-WTE (b) simulations. (c) Convergence can
be assessed by monitoring the free-energy differences between relevant regions of the CVs space as the simulation progresses. Time is measured as the fraction of the total aggregated time (simulation time per replica
multiplied by the number of replicas), i.e., 5.0s and 2.5s for PT-MetaD and PT-MetaD-WTE, respectively
In Fig.9a,b, we report the FES obtained from the PT-MetaD

and PT-MetaD-WTE simulations. By calculating the estimate of
the FES as a function of time, we can assess more quantitatively the
convergence of our simulation. Since monitoring the evolution of
a multi-dimensional profile may represent a complex task, one typically adopts a coarse-grained representation of the FES.This can
be done by defining relevant regions in the CVs space and calculating free-energy differences between pairs of them as a function of
the simulation time (see Note 12). In our example, we define a
folded (F) and an unfolded region (U) based on the value of S
alone (U: S<3; F: S3). FFU are calculated from the PT-MetaD
and PT-MetaD-WTE simulations and plotted in Fig.9c as a function of the simulation time.
The damped fluctuations in Fig.9c provide an intuitive estimate of the precision of the free-energy reconstruction as a function of the simulation time (see Note 13). If an exact reference
profile were available, one could estimate the accuracy by measuring the deviation of the calculated FES from the reference
(see Note 14).
5 Notes
1. While in analysis the simplest definition of contact is a step
function of the interatomic distance, in MetaD we need to use
a continuous and differentiable function in order to calculate
the additional forces due to the bias potential.
2. In order to perform a well-tempered MetaD simulation, one
has to set the following parameters: the Gaussian width i (one
per CV), the initial deposition rate 0, and the well-tempered
bias factor . The Gaussian width should be comparable to the
shape of basins in the underlying FES.This can be estimated a
166
priori by performing short unbiased MD simulations and computing CV fluctuations. Typically, the Gaussian width should
not be greater than 1/3 of the fluctuations. Recently, a more
advanced approach has been proposed to automatically tune
this parameter [40]. The initial deposition rate does not affect
the long time behavior [11]. However, a small initial deposition rate would result in a longer filling time, while a too high
rate might be problematic in the transient regime if the CVs
are not properly chosen. Typically, in simulation of proteins an
initial deposition rate of at most 1kBT per ps is used. The bias
factor affects the probability distribution of the CV in the long
time limit. Therefore, the optimal bias factor should be large
enough to cross all the relevant barriers in the process under
study, and small enough to limit sampling to the relevant
regions of the CV space. As discussed in ref. 11, overestimation
of the optimal value is to be preferred to underestimation.
3. In PT-MetaD simulations, one should carefully choose the
value of the minimum and maximum temperature and how
the replicas are distributed in this interval. The lowest temperature usually corresponds to the temperature of interest
(typically 300K). The highest temperature should guarantee a
fast sampling of all the degrees of freedom other than the CVs.
Replicas should be chosen so that a sufficient overlap between
potential energy distributions of neighboring temperatures is
achieved, thus guaranteeing a good acceptance probability for
the exchanges. The proper distribution depends on the system
specific heat and its dependence on temperature. When simulating proteins in explicit solvent in the NVT ensemble, the
maximum temperature is typically set to 600700K and the
appropriate temperature distribution is given in ref. 41.
4. Equilibration of all the replicas prior to applying the MetaD
bias is of great importance in PT-MetaD.It is indeed not convenient to initiate the simulation from identical conformations
since this would result in a long transient due to the large
accumulation of bias in the initial region of the CVs space.
Equilibration can be achieved either through PT run or multiple NVT simulations.
5. Some of the algorithms routinely used to perform simulations
at constant temperature cannot reproduce the correct energy
fluctuations of the NVT ensemble. Since the exchange process
of the REM protocol is strongly dependent on the potential
energy distributions, one must implement a correct thermostat, such as NoseHoover [42], Langevin, or BussiDonadio
Parrinello [43], to avoid possible artifacts [44].
6. The trr and xtc files produced by GROMACS contain configurations sampled at constant temperature; therefore, due to
the PT exchanges between replicas, these are not continuous
167
trajectories. To reconstruct the continuous trajectory of each

replica across the temperature space, one can use the
GROMACS tools demux.pl and trjcat (see GROMACS manual for their respective use). Once these trajectories are reconstructed, one can use the PLUMED tool driver to recalculate
the continuous CVs trajectory of each replica.
7. The parameter modulates the increase of the potential energy
fluctuations in WTE and thus the overlap between potential
energy distributions of neighboring temperatures. For a given
system, one can tune in order to (1) obtain a more efficient
diffusion in temperature once the number of temperatures is
fixed, (2) reduce the number of replicas needed to span the
temperature range of interest, without affecting the exchange
efficiency. A wide range of different choices is thus possible,
depending on the computational resources available to the
user. According to ref. 17, a key quantity in PT-MetaD-WTE is
the average replica round-trip time in temperature space, which
should be minimized to achieve optimal efficiency in the FES
reconstruction. However, choosing a too high is discouraged
since it has several drawbacks: (1) slow convergence of the bias
in energy space, (2) exploration of unlikely regions of the phase
space, such as conformations in which the water is frozen, (3)
inefficiency of reweighting procedures when the simulated
ensemble is very different from the original one [45]. As an
additional check to verify the WTE part of the simulation was
performed correctly, we also strongly suggest carefully monitoring the early stages of a PT-MetaD-WTE simulation in
order to ensure the expected exchanges are obtained.
8. Typically, in solvated biomolecular systems the potential
energy and its fluctuations are dominated by the solvent
degrees of freedom. Thus, the potential energy distribution is
quasi-Gaussian and it is almost decoupled from the CVs that
are used to describe the protein conformation. If one focuses
on protein properties, it is convenient to use two different bias
potentials: a one-dimensional static bias on the solvent-
dominated potential energy, and a N-dimensional MetaD bias
acting on the protein CVs. This strategy is to be preferred to
the introduction of a (N+1)-dimensional MetaD bias acting
simultaneously on the potential energy and protein CVs.
9. Two important things should be kept in mind when performing simulations in the WTE ensemble (both PT-WTE and
PT-MetaD-WTE). First, during the course of the different
steps, one should not change those MD parameters that affect
the absolute value of the potential energy, such as the scheme
to calculate the electrostatic potential or the cutoffs for electrostatic and van der Walls interactions. Second, the energy
must be calculated at every time step to apply correct forces to
168
the dynamics. In GROMACS, this can be done by setting in

the input file nstcalcenergy=1.
10. The bias on the potential energy needed to sample the WTE
can be stored on file in a grid format for later use. The boundaries of the grid should be specified in the PLUMED input file
and must include the range of potential energy sampled at all
temperatures. In order to properly evaluate the bias potential,
the dimension of the grid side should be equal to at most half
of the Gaussian width.
11. As discussed in the Subheading2, we can estimate the FES
from the Gaussians deposited along a MetaD simulation. In
practice, this can be done using the PLUMED sum_hills tool,
which automatically integrates the Gaussians and applies the
correct scaling factor (Eq.4).
12. When assessing convergence, it is convenient to define regions
in the FES corresponding to local minima and to calculate
free-energy differences between pair of them. In order to do
that, we need to define the free energy of a basin (A) as:
F (S )
FA = kBT log dS exp
.
kBT
SA
(14)
13. The statistical error in a well-tempered MetaD simulation is

proportional to 1/t [11]. In practical applications, the precision in FES reconstruction has been estimated by the following approaches:
(a) Using the standard deviation of the FES obtained by multiple independent MetaD simulations [46];
(b) Using the standard deviation of the FES reconstructed at different times in a single simulation [47]. However, each point
should be weighted proportionally to t, following the relation between error and simulation time mentioned above.
14. Different metrics have been used to compare an estimate of
the free energy F(S) to a reference profile FREF(S) (see for example [48]). Among these:
(a) The RMSD:
dRMSD ( F (S ) , FREF (S ) ) =

2
1
dS ( F (S ) FREF (S ) ) , (15)
where is the volume of the CVs space explored. This

volume is often chosen to include only relevant regions of
the CVs space, i.e., points within a few kBT from the global
minimum.
169
(b) The KullbackLeibler divergence [49]:
F (S ) F (S ) FREF (S )
dKL ( F (S ) , FREF (S ) ) = dS exp REF
, (16)
kBT
kBT

which assigns a weight to each point based on its reference
probability density.
Since free energies are always defined modulo an irrelevant
additive constant, one should optimally align the two profiles
in order to minimize the distance between them (using
Eqs.15, 16, or other metrics). When using RMSD as metrics,
optimal alignment can be achieved by calculating the average
free energy in the volume of interest and subtract it to the
profile, so that each free energy is offset to have its average
value at zero.
Acknowledgements
AB thanks the Swiss National Science Foundation for financial
support under the Ambizione grant PZ00P2_136856. J.P.
acknowledges the support of NSF award CMMI-1032368. The
simulations of Trp-Cage miniprotein were made possible in part by
the National Science Foundation through TeraGrid resources provided by NICS.These simulations were also facilitated through the
use of computational, storage, and networking infrastructure provided by the Hyak supercomputer system, supported in part by the
University of Washington eScience Institute.
References
1. Shaw DE, Maragakis P, Lindorff-Larsen K
etal (2010) Atomic-level characterization of
the structural dynamics of proteins. Science
330:341346
2. Beberg AL, Ensign DL, Jayachandran G,
Khaliq S, Pande VS (2009) Folding@home:
lessons from eight years of volunteer distributed
computing,
IEEE
International
Symposium on, Parallel & Distributed
Processing, 2009. IPDPS 2009, 23-29 May
2009, Rome, pp.16241631
3. Chipot C, Pohorille A (2007) Free energy calculations: theory and applications in chemistry
and biology. Springer, Berlin
4. Dellago C, Bolhuis PG (2009) Transition path
sampling and other advanced simulation techniques for rare events. Adv Polym Sci
221:167233
5. Laio A, Parrinello M (2002) Escaping free-

energy minima. Proc Natl Acad Sci U S A
99:1256212566
6. Barducci A, Bonomi M, Parrinello M (2011)
Metadynamics. Wir Comput Mol Sci
1:826843
7. Sugita Y, Okamoto Y (1999) Replica-exchange
molecular dynamics method for protein folding. Chem Phys Lett 314:141151
8. Hansmann UHE (1997) Parallel tempering
algorithm for conformational studies of biological molecules. Chem Phys Lett 281:
140150
9. Neidigh JW, Fesinmeyer RM, Andersen NH
(2002) Designing a 20-residue protein. Nat
Struct Biol 9:425430
10. Bonomi M, Branduardi D, Bussi G etal (2009)
PLUMED: a portable plugin for free-energy
170
calculations with molecular dynamics. Comput

Phys Commun 180:19611972

11. Barducci A, Bussi G, Parrinello M (2008)
Well-
tempered metadynamics: a smoothly
converging and tunable free-energy method.
Phys Rev Lett 100:020603
12. Bonomi M, Parrinello M (2010) Enhanced
sampling in the well-tempered ensemble. Phys
Rev Lett 104:190601
13. Bonomi M, Barducci A, Parrinello M (2009)
Reconstructing the equilibrium Boltzmann
distribution from well-tempered metadynamics. J Comput Chem 30:16151621
14. Barducci A, Bonomi M, Parrinello M (2010)
Linking well-tempered metadynamics simulations with experiments. Biophys J 98:L44L46
15. Earl DJ, Deem MW (2005) Parallel tempering: theory, applications, and new perspectives.
Phys Chem Chem Phys 7:39103916
16. Bussi G, Gervasio FL, Laio A, Parrinello M
(2006) Free-energy landscape for beta hairpin
folding from combined parallel tempering and
metadynamics.
J
Am
Chem
Soc
128:1343513441
17. Deighan M, Bonomi M, Pfaendtner J (2012)
Efficient simulation of explicitly solvated proteins in the well-tempered ensemble. J Chem
18. Hess B, Kutzner C, van der Spoel D, Lindahl
E (2008) GROMACS 4: algorithms for
highly efficient, load-balanced, and scalable
molecular simulation. J Chem Theory
Comput 4:435447

19. Tribello GA, Bonomi M, Branduardi D,
Camilloni C, Bussi G (2014) PLUMED 2:
new feathers for an old bird, Comput Phys
Commun 185:604613
20. Pettersen EF, Goddard TD, Huang CC etal
(2004) UCSF chimera- A visualization system
for exploratory research and analysis. J Comput
Chem 25:16051612
21. Hunter JD (2007) Matplotlib: a 2D graphics
environment. Comput Sci Eng 9:9095
22. Qiu LL, Pabit SA, Roitberg AE, Hagen SJ
(2002) Smaller and faster: the 20-residue Trp-
cage protein folds in 4 mu s. J Am Chem Soc
124:1295212953

23.
Streicher WW, Makhatadze GI (2007)
Unfolding thermodynamics of Trp-cage, a 20
residue miniprotein, studied by differential
scanning calorimetry and circular dichroism
spectroscopy. Biochemistry US 46:28762880
24. Neuweiler H, Doose S, Sauer M (2005) A
microscopic view of miniprotein folding:
enhanced folding efficiency through formation
of an intermediate. Proc Natl Acad Sci U S A
102:1665016655
25. Ahmed Z, Beta IA, Mikhonin AV, Asher SA

(2005) UV-resonance Raman thermal unfolding study of Trp-cage shows that it is not a
simple two-state miniprotein. J Am Chem Soc
127:1094310950

26. Zhou RH (2003) Trp-cage: folding free
energy landscape in explicit water. Proc Natl
Acad Sci U S A 100:1328013285

27. Ota M, Ikeguchi M, Kidera A (2004)
Phylogeny of protein-folding trajectories
reveals a unique pathway to native structure.
Proc Natl Acad Sci U S A 101:1765817663
28. Juraszek J, Bolhuis PG (2006) Sampling the
multiple folding mechanisms of Trp-cage in
explicit solvent. Proc Natl Acad Sci U S A
103:1585915864
29. Paschek D, Nymeyer H, Garcia AE (2007)
Replica exchange simulation of reversible folding/unfolding of the Trp-cage miniprotein in
explicit solvent: on the structure and possible
role of internal water. J Struct Biol
157:524533
30. Marinelli F, Pietrucci F, Laio A, Piana S (2009)
A kinetic model of trp-cage folding from multiple biased molecular dynamics simulations.
PLoS Comput Biol 5:e1000452
Science 334:517520

32. Hornak V, Abel R, Okur A etal (2006)
Comparison of multiple amber force fields and
development of improved protein backbone
parameters. Proteins 65:712725
33. Jorgensen WL, Chandrasekhar J, Madura JD,
Impey RW, Klein ML (1983) Comparison of
simple potential functions for simulating liquid
water. J Chem Phys 79:926935

34. Branduardi D, Gervasio FL, Parrinello M
(2007) From A to B in free energy space.
JChem Phys 126:054103
35. Ceriotti M, Tribello GA, Parrinello M (2011)
Simplifying the representation of complex
free-energy landscapes using sketch-map. Proc
36. Spiwok V, Kralova B (2011) Metadynamics in
the conformational space nonlinearly dimensionally reduced by Isomap. J Chem Phys
135:224504

37. Piana S, Laio A (2007) A bias-exchange
approach to protein folding. J Phys Chem B
111:45534559

38. Sindhikara DJ, Emerson DJ, Roitberg AE
(2010) Exchange often and properly in replica
exchange molecular dynamics. J Chem Theory
Comput 6:28042808

39. Torrie GM, Valleau JP (1977) Non-physical
sampling distributions in monte-carlo free-energy

estimation- umbrella sampling. JComput Phys
23:187199
40. Branduardi D, Bussi G, Parrinello M (2012)
Metadynamics with adaptive gaussians. JChem
41. Prakash MK, Barducci A, Parrinello M (2011)
Replica temperatures for uniform exchange
and efficient roundtrip times in explicit solvent
parallel tempering simulations. J Chem Theory
Comput 7:20252027
42. Nose S (1984) A unified formulation of the
constant temperature molecular-dynamics
methods. J Chem Phys 81:511519

Canonical sampling through velocity rescaling.

44. Rosta E, Buchete NV, Hummer G (2009)
Thermostat artifacts in replica exchange
molecular dynamics simulations. J Chem
171

45.
Ceriotti M, Brain GAR, Riordan O,
Manolopoulos DE (2012) The inefficiency of
re-weighted sampling and the curse of system
size in high-order path integration. P Roy Soc
a-Math Phys 468:217

46. Angioletti-Uberti S, Ceriotti M, Lee PD,
Finnis MW (2010) Solid-liquid interface free
energy through metadynamics simulations.
Phys Rev B 81:125416
47. Berteotti A, Barducci A, Parrinello M (2011)
Effect of urea on the beta-hairpin conformational ensemble and protein denaturation
mechanism. J Am Chem Soc 133:
1720017206
48. Sutto L, DAbramo M, Gervasio FL (2010)
Comparing the efficiency of biased and unbiased molecular dynamics in reconstructing the
free energy landscape of met-enkephalin.
JChem Theory Comput 6:36403646
49. Kullback S, Leibler RA (1951) On Information
and Sufficiency. Ann Math Stat 22:142143
Chapter 9
Calculation ofBinding Free Energies
VytautasGapsys, ServaasMichielssens, Jan HenningPeters,
BertL.deGroot, andHadasLeonov
Abstract
Molecular dynamics simulations enable access to free energy differences governing the driving force
underlying all biological processes. In the current chapter we describe alchemical methods allowing the
calculation of relative free energy differences. We concentrate on the binding free energies that can be obtained
using non-equilibrium approaches based on the Crooks Fluctuation Theorem. Together with the theoretical
background, the chapter covers practical aspects of hybrid topology generation, simulation setup, and free
energy estimation. An important aspect of the validation of a simulation setup is illustrated by means of calculating free energy differences along a full thermodynamic cycle. We provide a number of examples, including
proteinligand and proteinprotein binding as well as ligand solvation free energy calculations.
Key words Free energy, Molecular dynamics, Alchemical transitions, Proteinligand binding, Protein
protein interaction, Non-equilibrium methods, Hybrid topology, Crooks Fluctuation Theorem
1 Introduction
Whether or not a process happens spontaneously is determined
by its free energy. This is because, without an external source of
energy, systems evolve to their lowest free energy state. Likewise,
the rate at which that state is reached depends on free energy
barriers along the pathways to that minimum. Hence, free energies
are of central importance as they determine, e.g., binding affinities (spontaneous binding or not) or protein folding (the folded
state is usually the free energy minimum). In addition, as barriers
are linked to rates via rate theory, free energy barriers determine
binding, folding, permeation, and reaction kinetics. Therefore,
free energies are among the most critical thermodynamic quantities
to accurately be derived by computational techniques, not only
because they play such a fundamental role, but also because they
can be directly and quantitatively compared to experimental data.
173
174
Vytautas Gapsys et al.
Thermodynamically, a distinction is made between the so-called

Helmholtz and Gibbs free energies. The Helmholtz free energy is
defined as:

F = U - TS
(1)
with U the internal energy, T the temperature, and S the entropy

of the system. The Gibbs free energy is defined as:

G = H - TS
(2)
with H the enthalpy of the system, defined as H=U+pV, with p

the pressure and V the volume of the system.
Importantly, note that both definitions contain a term that is
dependent on the internal (or potential) energy of the system, and
another that depends on the entropy. This has the important implication that free energies, and hence, affinities or stabilities, are
affected by both changes in interatomic interactions as well as
entropic changes. Particularly, changes in solvent entropy can play
a substantial or even dominant role as, e.g., in the hydrophobic
effect, where solvent molecules are set free upon the formation of
a hydrophobic protein or membrane core.
The difference between the Helmholtz and Gibbs free energy
lies in the pV term. The Helmholtz free energy is applicable at
constant volume, whereas the Gibbs free energy is used at constant
pressure, where the pV term quantifies the work associated with a
change in volume. Under physiological conditions and for a typical
experimental setup we usually have isobaric conditions, hence the
Gibbs free energy gradient acts as a driving force for a system.
While we will return to the concepts of the Helmholtz and Gibbs
free energies in Subheading 2.1, for the practical examples in this
chapter we will focus on the Gibbs free energy.
In addition to being the quantity that is minimized at equilibrium, the free energy is also the maximum amount of energy that
can be obtained from a spontaneous process at constant temperature, or conversely, the minimum amount of energy required to
drive an uphill process. More precisely, the free energy is the energy
obtained from (or required to drive) a process slow enough such
that it is in equilibrium with its surroundings at all times. If enforced
faster, friction results in non-equilibrium work values that on average are larger than the associated free energy change. The excess
energy is dissipated as heat. Traditional free energy methods therefore require transitions slow enough such that the systems under
investigation can be considered to be in equilibrium at all times.
Remarkably, however, with the Jarzynski equality and the Crooks
Fluctuation Theorem (CFT) given, free energies can also be
derived from non-equilibrium work distributions, as described
further below.
175
An important implication of the free energy follows from its

role in statistical mechanics, stating that the probability to be in
state x is directly related to its free energy:

p (x ) e -G (x ) / kBT
(3)
where kB is the Boltzmann constant. The term e -G (x )/ kBT is called the

Boltzmann factor. This relation is particularly useful to estimate
free energy differences between two states A and B:
p (A)
= e -(G (A )-G (B )) / kBT = e - DG / kBT
p (B)

(4)
which also presents the most straightforward way to estimate free

energies from simulation: provided that a sufficiently converged
ensemble is available, usually requiring several reversible transitions
between A and B, the free energy difference G can directly be estimated from this relation by evaluating the populations in A and B.
Alternative simulation approaches to derive free energies are
frequently tailored towards overcoming or avoiding free energy
barriers associated, e.g., with binding, unbinding, conformational
transitions, or (re)folding. Examples include, umbrella sampling
[1] and thermodynamic integration [2]. In umbrella sampling, a
biasing potential is employed to enforce transitions across high
energy states along a predefined reaction coordinate. The effect of
the biasing potential can be corrected for afterwards, yielding
the free energy profile for a one-dimensional reaction coordinate
or the free energy landscape for a multi-dimensional reaction coordinate. In thermodynamic integration and other alchemical appro
aches, the sampling of cumbersome binding and unbinding events
is prevented by instead carrying out a transition in chemical space
where the Hamiltonian describing system A is transformed into
B. Making use of the fact that the free energy is a state variable, the
relative free energy change can be estimated from the difference of
the associated transition free energy for example in the context of
a protein binding pocket and the same transformation in solution,
for the calculation of relative binding free energies.
A remarkable characteristic of the free energy is that in many
cases, its estimate converges faster than either the individual
enthalpy and entropy contributions. This is due to the fact that the
largest contributions to both arise from a solventsolvent term
that occurs in both but that exactly cancels out in the free energy
estimate [3].
This chapter mainly focuses on the use of non-equilibrium
methods for the calculation of binding free energies. For equilibrium approaches, we refer to excellent reviews in the literature
[46].
176
2 Theory
2.1 Definition
ofFreeEnergy
The free energy surface of a system completely describes its

thermodynamic and kinetic properties. An intuitive interpretation
of the free energy is provided by its relation to the probability pA of
finding the system in a phase space volume A:
pA =
e - b FA
e -bF
(5)
where FA is the free energy of a phase space volume A, F is the

Helmholtz free energy evaluated over the whole phase space, and
=1kBT, with the Boltzmann constant kB and temperature T.
In most practical cases the differences between the state populations are of interest in contrast to the absolute free energy values.
By considering two states, A and B, the free energy difference
FAB can be expressed as
DFAB = FB - FA = -
1 pB
1 Q
ln
= - ln B
b pA
b QA
(6)
Q denotes the canonical partition function: Q = Q(N,V,T), with

N number of particles and V volume of a container. The free energy
is related to the partition function via F=1lnQ(N,V,T).
From the latter expression (Eq.6) several approaches to obtaining F can be deduced. Firstly, a direct counting of events sampling phase space volumes A and B immediately enables access to
the probabilities pA and pB. For example, in case of molecular binding, this method could be employed by simulating two molecules
of interest (ligands A and B) and directly counting their respective
probabilities of being in bound/unbound forms. While the
approach is simple, its computational cost for large biologically
relevant systems is usually beyond reach for the current state-ofthe-art atomistic simulations.
Another method requires evaluation of the partition functions.
However, while for an ideal gas example, the partition function has
an analytical expression, for particles with complex interactions
a numerical integration over the coordinates and momenta is
required. A canonical partition function is defined as follows:

Q (N ,V ,T ) =
1
e - b H (p1 pN , q1 qN )dp1 dpN dq1 dqN
h N!

3N
(7)
where H(p,q) is a Hamiltonian of a system, q and p denote coordinates and momenta, respectively, h is Plancks constant. For a
multi-particle system, integration over all the degrees of freedom is
not computationally feasible. Hence, a simulation is often used to
sample the accessible phase space volume. In a simulation, high
177
energy microstates will be visited only rarely (or will not be visited
at all), hence, rendering this approach unsuitable for the estimation
of absolute free energies. For the free energy differences, however,
the inaccessible phase space regions will be discarded for both
states, A and B, thus allowing for an accurate assessment of F due
to cancellation of errors.
Up to now, we considered a canonical ensemble with the associated canonical partition function Q and Helmholtz free energy F.
In practice, however, experimental measurements are usually
performed at isothermal-isobaric conditions generating an NPT
ensemble. In such a case, a partition function is defined as
Q (N , P ,T ) =
1
e - b (H (p1 pN , q1 qN )+ PV )dVdp1 dpN dq1 dqN
h 3N N !

(8)
where P is the pressure. The Gibbs free energy is defined as

G=1lnQ(N,P,T
). For the remainder of this chapter we
assume constant pressure and therefore focus on the NPT ensemble with the associated Gibbs free energy. By considering the fact
that a Hamiltonian of a system consists of a kinetic and potential
energy components H(p,q)=K(p) + U(q), the partition function
can be separated into kinetic and configurational partition functions. As the K(p) component depends only on the particle
momenta, the kinetic partition function for states A and B does
not change as long as there is no perturbation of mass between the
states. Therefore, for the sake of simplicity, often only the configurational partition function is used. However, in this chapter, we
will use a more general approach by considering the full Hamiltonian
of a system.
In the following section, we will briefly introduce the most
popular approaches for free energy estimation from equilibrium
simulations. Further, we will concentrate on the conceptually different approaches relying on non-equilibrium transitions between
states.
2.2 Free Energy
Estimates
fromEquilibrium
Simulations:
PerturbationMethod
The free energy perturbation (FEP) method introduced by

Zwanzig [7] (Pohorille etal. [8] attribute the first derivation to
Landau [9]) can be derived from Eq.6:
DG AB = GB - G A = -
1 QB
1
ln
= - ln e - b (H B (p, q)-H A (p, q))
b QA
b
(9)

where the angular brackets denote an ensemble average. The

approach relies on equilibrium sampling at state A and subsequent
evaluation of the produced configurations with the Hamiltonian of
state B. The Hamiltonian is controlled by an external parameter
often denoted as : H l (p, q) = H (p, q, l). The accuracy of the FEP
method is strongly dependent on the phase space overlap of the
states A and B. In case the overlap is small, the configurations
178
enerated at state A will be identified as high energy microstates

g
when evaluated with the Hamiltonian HB(p,q), in turn contributing
little to the exponential average. Hence, the approach is known to
converge slowly and is only tractable for the G estimation between
the states exhibiting large overlap in the phase space. More information concerning the accuracy, convergence, and usage of FEP can be
found in [8, 1012].
Another method to estimate G from end state equilibrium
sampling was proposed by Bennet and is referred to as Bennets
Acceptance Ratio (BAR) [13]. Bennets estimate is expressed by:
DG AB =

f (H A (p, q) - H B (p, q) + C )
1
ln
b
f (H B (p, q) - H A (p, q) - C )
+C
(10)

here f represents Fermi function f(x)=1(1 + exp(x)) and

1 Q n
C = ln A B , nA and nB are the numbers of configurations
b Q B nA
generated in the states A and B, respectively. Equation (10) can be
solved numerically by finding such C, that

f (H
B
(p, q) - H B (p, q) + C ) = f (H B (p, q) - H A (p, q) - C )

A
(11)

Having determined C yields the free energy difference:

1 n
DG = - ln B + C . BAR is a minimal variance free energy estimate.
b nA
Shirts etal. [14] also derived Bennets formula using maximum
likelihood formulation, thus, demonstrating that BAR provides
the most likely free energy estimate for the observed work
distributions.
Both methods introduced so far (FEP and BAR) were defined
for free energy calculations by sampling physical end states of a
system. Simulations, however, allow accessing unphysical (alchemical) pathways as well, by coupling the Hamiltonians of the two
states H l = (1 - l)H A + lH B . Different dependent coupling
functions have been investigated [15, 16]. For the values 0 and 1,
the system is at the physical states A and B, respectively, whereas for
the values 0<<1 the system is in a mixed unphysical state. The
path between the states A and B can be divided into discrete states
(also called stratification). Performing equilibrium simulations at
the intermediate states ensures a larger phase space overlap between
the ensembles adjacent to one another along the coordinate. FEP
or BAR estimators can be employed to obtain free energy differences between the intermediate states which can later be summed
up to yield GAB between the physical end states. The weighted
histogram analysis (WHAM) [17] and multistate Bennet Acceptance
Ratio (MBAR) [18] methods were developed to estimate free
179
energy difference by considering all intermediate states simultaneously.

Ways to achieve an optimal distribution of the discrete states to
ensure maximal phase space overlap along the path have also been
investigated [19, 20].
2.3 Free Energy
Estimates
fromEquilibrium
Simulations:
Thermodynamic
Integration
A conceptually different method from the ones described above,

termed thermodynamic integration (TI) [2], obtains the free energy
difference by integrating the average force exerted on the system
along the variable during an alchemical transition:
DG AB =
1
0
H
l
dl
l
(12)

Several approaches have been developed to use thermodynamic

integration in simulations. The slow growth TI requires performing transition between the states very slowly, such that the system
remains close to equilibrium at all times. In this case, the free energy
value is equal to the work done during the transition. Another
assumption states that an instantaneous H / l value at any i is
equal to an ensemble average at that i. In practice, due to limited
sampling the system is kept at a quasi-equilibrium state, which in
turn results in inaccurate free energy estimates as some work is dissipated. Convergence issues related to the slow growth TI are well
known and have been thoroughly investigated [2123].
A discrete thermodynamic integration method (DTI) is based
on dividing the path along the coordinate into discrete steps,
similarly as for the FEP and BAR approaches discussed in the previous section. An equilibrium simulation is started at every i state
H
and an average
is evaluated. A numerical integration of
l li
the averages directly yields a G value. The accuracy of DTI
depends on the sampling time and distribution of the discrete
states. As it was the case for the multi-state FEP and BAR methods,
DTI requires phase space overlap between the neighboring ensembles along the coordinate. The numerical integration scheme was
also demonstrated to influence accuracy of the free energy estimates [20, 24, 25].
2.4 Theory ofNon-
equilibrium Free
Energy Calculations
The methods described in the current section rely on the work

measurements over non-equilibrium transitions, and hence are
sometimes termed NEW.In 1997 Jarzynski derived an identity
relating an exponential average of work during non-equilibrium
transitions to the free energy difference of the canonical ensemble
[26] (it was later shown by Cuendet [27], that the equality also
holds for an NPT ensemble):
e - bDG AB = e - bW
(13)
180
The Jarzynski equality requires the transitions to be started

from an equilibrium ensemble. The work values can be obtained
from a relation similar to that of thermodynamic integration:
W =

1
0
H
dl
l
(14)

In contrast to TI, in Eq.14 instantaneous H / l values are

integrated. The resulting work contains both a contribution from
the free energy difference and the dissipated work along the transition path. The FEP approach (Eq.9) can be viewed as a special case
of the Jarzynski equality, where a transition from A to B is performed instantaneously.
Similarly to FEP, of which the convergence strongly depends
on the phase space overlap between the end states, the accuracy of
the free energy differences calculated using Jarzynskis formula
depends on rare events, where little work is dissipated. Fast transitions, driving a system far from equilibrium, would yield large work
values, contributing little to the exponential average, hence slowing the convergence of the G estimation.
Several free energy estimators have been developed based on
the Jarzynski equality. In case the transitions are performed in the
near equilibrium regime, a Gaussian approximation for the work
distribution P(W) is valid due to the central limit theorem [28].
Cumulant expansion of a Gaussian distribution allows expressing
the free energy difference as
DG = W
b s W
2
(15)
where W n is the mean and s W is the variance of a P(W) distri

bution. The variance of a free energy estimate is given by
2
4
s W / n + b 2 s W / (2n - 2) , with n denoting number of measured
work values. The estimator ought to be used only if the assumptions for the Gaussian approximation of a work value distribution
are fulfilled, i.e. many work values are obtained and the transitions
keep the system near equilibrium [29]. Otherwise, the free energy
difference can be estimated from the Jarzynski equality directly.
Such an estimator, however, has been shown to be biased given a
limited number of observed work values [29]. The Jarzynski estimator is defined as
1
DG = - ln e - bW
b
(16)

A more general relation, enabling the combination of the work

value distributions from forward and backward transitions to obtain
the Helmholtz free energy difference was derived by Crooks [30, 31].
181
The relation is also known as the Crooks Fluctuation Theorem (CFT).

Chelli demonstrated the validity of the Crooks equation for an NPT
ensemble [32].
P f (W )
Pr (-W )
= e b (W - DG )
(17)

where Pf(W) and Pr(W) correspond to the work distributions

during forward and reverse processes, respectively. The Jarzynski
identity can be derived from the Crooks equality, as demonstrated
by Crooks [30]. CFT requires that the transitions between the
states were started from equilibrium ensembles. However, similar
to the Jarzynski identity, there is no requirement for a system to
reach an equilibrium at the final state of a transition. A number of
estimators have been developed to extract the free energy difference
from the work distributions using Crooks equality [14, 3335].
For the cases where an overlap between the work distributions is
large, the free energy difference can be estimated directly from the CFT:
ln

P f (W )
Pr (-W )
= bW - bDG
(18)

Plotting the left hand of Eq.18 against the work values yields
a line with the slope . The line intercepts the work axis at a value
equal to DG . While the approach is easy to implement, there are
several caveats regarding such a direct estimation. Firstly, a sufficient
overlap between the work histograms is achieved only for nearequilibrium transitions where little work is dissipated. Secondly,
only the work values from the overlap region will contribute to the
free energy estimate, whereas the rest of the measurements will not
be used.
To alleviate the latter problems, the work histograms can be
approximated by an analytical distribution. Following from the
CFT, the intersection point of the two distributions corresponding
to the forward and reverse transitions marks a work value equal to
G. Nanda etal. [33] proposed using a universal probability density function [36] allowing to account for the asymmetry of the
distributions. Goette and Grubmller showed that in practice a
Gaussian approximation also yields accurate free energy estimates
[35]. They derived a Crooks Gaussian Intersection (CGI) estimator which is expressed as
Wf
nf
2
DG =
s f
- Wr
2
s r
nr
1
2
( Wf
2
s f s r
1
2
s f
nf
+ Wr

2
1 - 1 ln s r
+
2
)
nr
s 2f s r2 s f
(19)
1
2
s r
182
where W f
, Wr
are work averages and s 2f , s r2 are variances

of the work distributions for the forward and reverse transitions,
respectively. The estimator predicts two intersection points, unless
the distributions overlap completely, which would only be the case
for equilibrium transitions. The intersection point in between of
the W f
nf
nf
nr
and - Wr
nr
is of relevance. The statistical error for
the estimator can be calculated by means of bootstrapping. As the

estimator depends on a Gaussian approximation, the validity of this
assumption can be assessed using a s tatistical test, e.g. Kolmogorov
Smirnov [37].
The already introduced BAR estimator (Eq.
10) can also
be used for non-equilibrium simulations. Shirts etal. provide the
following expression for BAR [14]
nf
i =1
nr
1
1 + exp(ln
nf
nr
+ b (Wi - DG ))
1
n
^
j =1
1 + exp(ln r - b (W j - DG ))
nf
(20)

where nf and nr are the numbers of transitions in forward and

reverse directions, respectively. A numerical solution of Eq.20
provides a maximum likelihood estimator of the free energy difference. The estimator is asymptotically unbiased and its variance
converges to the inverse of a Fisher information when the number
of work observations goes to infinity. An analytical expression for
the variance [14] is given by
s DG =
1
b n f +r
-1
n f +r n f +r
-
+
nf
nr
1
2 + 2 cosh(ln
nf
nr
+ b (Wi - DG ))
n f +r
(21)
where nf+r corresponds to the total number of forward and reverse

transitions and the angular brackets denote averaging over the
transitions in both directions. The maximum likelihood estimator
was later generalized by Maragakis etal. who introduced a Bayesian
free energy estimator based on CFT [34].
2.5 Concepts
ofSingle andDual
Topology
As discussed in the previous sections, alchemical free energy calculations explore unphysical pathways by combining Hamiltonians of
physical states with an external parameter . Having two (or more)
separate Hamiltonians implies the necessity to define multiple
topologies for the system at every end state. To be able to couple
Hamiltonians, a mapping between the topologies needs to be
established. Two approaches of constructing the topologies for
alchemical free energy calculations have been introduced [38, 39].
183
Fig. 1 Two topology generation approaches illustrated by an example of a methyl- to ethyl-benzamidinium

mapping. The dummy atoms are represented as transparent spheres. (a) In the single topology approach a
methyl group is morphed into an ethyl. (b) For the dual topology mapping, the whole group is annihilated upon
a transition and a new group is created
In a single topology approach every atom of state A is mapped

to an atom of state B. In case the states have different number of
atoms, dummy particles are introduced. The total number of atoms
in a system corresponds to the atom number of the larger state. The
bonded interactions of dummy atoms are usually left intact during
a transition, resulting in an ideal gas molecule state (i.e., the bonded
interactions are not switched off) in one of the end-states [40].
In a dual topology approach, atoms that are different for the
two states are defined separately, i.e. they are present in the system
simultaneously. The atoms that are different for the states A and B
do not interact with each other and are controlled by a parameter. It was found that the ideal gas molecule approach (the same as
for the single topology) ought to be used to avoid convergence
problems [40].
The G values for a transition calculated using different topology mappings will give rise to different free energy estimates, since
the actual end states will differ for the single and dual topologies.
The G values, however, do not depend on the choice of the
topology mapping as long as the same procedure is used across the
thermodynamic cycle [40, 41]. In principle, both topology mapping approaches can be combined if required. For example, the
topology for a part of a molecule may be designed to follow the
single topology approach, while the other part may be represented
by a dual topology (Fig.1).
2.6 Soft-Core
Potential
Transitions exploiting unphysical pathways across a thermodynamic cycle may require particle creation or annihilation. This is
always the case for a dual topology approach, whereas for a single
topology particle creation/annihilation is required only when the
184
number of atoms at the end states differs. In a classical molecular

mechanics force field description, non-bonded terms described by
the Coulomb and Lennard-Jones potentials contain a singularity at
inter-particle distances r=0. For the end state simulations, reaching a singularity point is prohibited by strong Pauli repulsion term
(r12) of the Lennard-Jones potential. For the unphysical transitions, however, very short inter-atomic distances may be encountered when approaching the end states. Another caveat along
the alchemical paths comes from simultaneous switching of the
electrostatic and van der Waals interactions. When the system
approaches an end state, the created/annihilated atoms have weak
van der Waals repulsion, whereas electrostatic attraction may be
strong enough to bring particles to distances close to zero. To
avoid the latter problem, decoupling electrostatic and van der
Waals interactions involved in transitions were suggested [42]. In
this approach, firstly, the Coulomb interactions are switched off. In
a subsequent separately performed transition the van der Waals
interactions are modified and finally the electrostatic interactions
are restored. The numerical instability problem, however, still
remains for the part of the procedure involving the LennardJones
interaction modification. To avoid numerical instabilities using the
classical non-bonded interaction potentials, the integration time step
needs to be decreased with the change of the parameter [43].
Another approach is to soften the non-bonded interactions
along an alchemical transition [43, 44]. The soft-core potential
for the Coulomb and van der Waals interactions can be written as
follows:
Vij (rij ) =
qi q j
4p e 0e r (a Q (1 - l) + rijp )1/ p
1
1
+ 4le ij
(a LJ (1 - l) + (rij / s ij )s )12 / s ) (a LJ (1 - l) + (rij / s ij )s )6 / s )
(22)
where p and s are integer parameters, rij is an interatomic distance,

qi and qj denote partial charges, e ij and ij represent LennardJones
parameters, e 0 and e r are the dielectric constant in vacuum and
relative dielectric constant, respectively.
Softening of the non-bonded interactions alleviates the
problems of singularity points and numerical instability. The soft-core
potential can be used for both, the Coulomb and van der Waals
interactions, hence, the switching of all the non-bonded interactions can be performed simultaneously. While the soft-
core
potential defined in Eq.22 is routinely used in the equilibrium free
energy calculations, its application for a number of non-equilibrium
simulation setups revealed potential caveats [45]. As the potential
becomes flat for the very short inter-particle distances, overlapping
particles exert no repulsive force on each other. This situation may
185
create unwanted additional minima in the potential, where particles are kept in a close proximity throughout a transition, eventually leading to a strong repulsion when reaching an end state. The
strong repulsions result in large work dissipation decreasing accuracy of the free energy estimation. A different soft-
core the
approach was proposed to solve the problems of singularity points,
numerical instabilities and additional minima [45]. In the latter
approach the softening of the non-bonded interactions is applied
at the force level by modifying the non-bonded interactions such
that a finite, but non-zero, force is reached at short inter-particle
distances.
3 Topology Generation
Topology generation is the particular aspect of an alchemical free
energy setup that makes it different from a regular molecular
dynamics simulation. The topology must describe both states of a
molecule undergoing a transition. Therefore, regardless of which
topology approach one chooses, single or dual, a mapping between
the atoms of the two states needs to be established.
At the first step, a mapping algorithm, when provided with the
structures of two molecules, should list atoms that need to be
morphed into one another or be turned into dummies. To automate such a process several approaches are available. A graph theory based connectivity analysis of the molecules can be used to find
a subset of connected atoms for morphing, e.g. a maximum common subgraph algorithm. The atoms not falling within the identified subset would be marked to become dummies in one of the
states. The drawback of a graph based approach is the fact that
while the atoms mapped for morphing may be close to each other
in a graph representation, in Cartesian coordinates the distance
between them may be large, resulting in potential convergence
issues in the simulations.
A different approach to atom mapping is based on a Euclidean
distance criterion. For the two superimposed molecules distances
between all the atom pairs need to be calculated. By defining a
threshold value (e.g., 0.5) pairs of atoms with distances below
the threshold are selected for morphing. The threshold parameter
can be adjusted depending on a specific situation. However, one
needs to be careful and avoid introducing unreasonable mappings
of spatially distant atoms. Creating fragments in a molecule connected via dummies should be avoided, since bonded interactions
of the dummies would restrain the degrees of freedom that need
not to be restricted. Similarly, breaking ring systems when morphing atoms may lead to stability and convergence issues. Therefore, it
is better to follow a dual topology approach and create/annihilate
intact rings.
186
Superpositioning of the molecules plays an important role in

the distance based topology generation. The atoms to be used for
superpositioning can either be defined in advance based on the
knowledge of the molecular structures at hand or an unsupervised
alignment and superpositioning method can be applied, as, e.g.,
implemented in Open3DALIGN algorithm [46].
In the second step, a merged topology file should be created.
The merged file needs to contain the atoms that are to be morphed
during a transition, as well as the dummies of both states. The
structure file has to be adjusted accordingly to contain all the atoms
described by the topology. In a GROMACS [47] topology file, the
atomic details, as well as bonded and non bonded parameters
for every interaction can be defined for the two =0 and =1
states separately. In that case, the generation of a merged topology
requires assigning the parameters for the two states where a change
occurs, and carefully adjusting atom numbering due to the possible introduction of dummies. The Python based package pmx
(formerly known as Pymacs [48]) contains a set of tools for the
distance based atom mapping and merged topology generation.
In case the merged topologies for a set of molecules are
intended to be used frequently, it is convenient to generate a library
containing the rules for atom mapping. For example, calculating
amino acid mutation effects is often of interest when assessing protein thermostability, proteinprotein or proteinligand binding.
The aforementioned pmx package provides a pre-calculated
mapping for all amino acid mutations, as well as the required tools
for the merged topology and mutated structure generation.
For ligands, however, establishing the mapping between the
atoms beforehand is not possible, hence, the topology generation
process needs to be carried out from the beginning for every
molecule pair of interest. In fact, it is often the case that the 3D
structures and topologies of individual ligands need to be genera
ted from a simplified molecule representation (a 2D structure, a
SMILES, or SMARTS string). As a starting point here, an empirical
3D structure generator can be employed, e.g., OpenBabel [49],
ChemAxons Marvin package tool MolConverter, CORINA [50].
A decision on the protonation state of a ligand must be made at this
step as well. If the protonation state is expected to play an important role for the binding of a particular ligand, the different protonation states can be treated as separate molecules and the topologies
for each of them could be generated. Subsequently, the calculated
free energy differences can be used to estimate the relative pKa
shifts for the titratable sites. If the molecular 3D structure was generated from scratch, it is suggested to optimize the geometry with
a quantum chemical package of choice. Several methods for the
atomic partial charge assignment are available: one such method
requires the calculation of an electrostatic potential (ESP) followed
by a procedure of fitting the ESP surface onto atoms [51, 52].
187
Alternatively, the computationally less demanding semiempirical

AM1-BCC [53] partial charges have been shown to be of comparable quality in the solvation free energy estimation [54]. Generation
of the bonded parameters for a molecule depends on the force field
of choice. Nowadays the main biomolecular force fields have been
generalized to include small organic compounds. In addition, automated atom typing and bonded parameter assignment procedures
are readily available: the Antechamber [55] m
odule allows topology
generation for the Generalized Amber Force Field (GAFF) [56];
CHARMM General Force Field (CGenFF) [57] topologies can be
created with the ParamChem [58, 59] utility; GROMOS force field
compatible topologies can be generated with an Automated
Topology Builder (ATB) [60]; ligand topologies for the OPLS
force field can be generated using MKTOP [61] software.
4 Thermodynamic Cycles
As previously mentioned, the absolute free energy of a system is
difficult to determine, but fortunately most problems can be formulated in terms of relative free energies. Free energy differences
are both more approachable and contain important information
about the system. The change in free energy due to binding
(Gbinding) can be determined experimentally (e.g., by means of
calorimetry). Although absolute binding affinities can be calculated using, for example, umbrella sampling, such calculations are
usually cumbersome, as the whole binding/unbinding process
needs to be taken into account, while its path is unknown in most
cases. Therefore, investigating the effect of a change of the system
(e.g., an amino acid mutation or a ligand modification) on binding
is usually more feasible. For this, the double free energy difference
(Gmutation,binding) is calculated.
The alchemical methods calculate the work needed to move
a system from one state to another through unphysical pathways.
As the free energy of a system is a function of its state, the free
energy difference found between the states is independent of the
path taken between them. Sufficient sampling is of critical importance in all free energy methods, and the computational difficulty
of reaching convergence dramatically increases with the magnitude
of the perturbation.
Binding free energies usually involve a large perturbationin
one state, both binding partners are free in solution, in the other
they are in a complex. The phase space overlap between these two
states is often small resulting in a slow convergence. An alchemical
transition from state A (a wild-type protein or a ligand) to the state
B (a mutated or modified molecule), while physically impossible,
requires a much smaller perturbation. As we are interested in the
difference in the binding free energies between the states A and B,
188
Fig. 2 Schematic representation of a thermodynamic cycle. To determine the

change in binding affinity upon a protein mutation or ligand modification,
the difference between the binding free energy of the ligand or protein in state A
(G1) and that of the state B (G2) need to be determined. However, the values
(G3) and (G4) are more accessible to the free energy perturbation methods.
Using the conservation of energy in this closed cycle, we can derive
DDG = DG2 - DG1 = DG 4 - DG3
we can make use of the fact that the free energy of a state does not
depend on the path taken to reach it. Hence, the free energy differences along a closed cycle of reactions (like the one depicted in
Fig.2) will always add up to zero. This feature allows calculation of
the double differences in free energy (G) of binding, thermostability, partitioning in different solvents, etc. between two states
of a system

DDG = DG 2 - DG1 = DG 4 - DG3
(23)
The resulting G is often more accurate than the G values

used to calculate it, as systematic errors (like those caused by the
presence of dummy particles and limited sampling) cancel out
when the free energy differences are subtracted. Note that this way,
relative binding affinities are obtained without studying the actual
binding/unbinding event.
The two branches of a thermodynamic cycle can be combined
into a single transition by performing both transitions in the same
box. This is important to bear in mind, when an alchemical modification involves a charge change between the states. In such a case,
the simulation box in one of the end states carries a non-zero
charge, hence, treatment of the electrostatic interactions by means
of the Ewald summation introduces artifacts. Meanwhile methods
that would not introduce such artifacts, e.g. a simple cut-off based
Coulomb interaction calculation, are generally less accurate. Hence,
the best solution is to design a specific simulation setup to keep the
system neutral at all times during a transition. One such simulation
setup is to combine both branches of a thermodynamic cycle in one
simulation box (Fig.3). In such a double-system/single-box setup,
189
Fig. 3 Double-system/single-box setup. (a) Two branches of a thermodynamic

cycle are placed in one simulation box. The different boxes in the scheme are
indicated by the broken and dotted lines. (b) An example of a simulation setup for
the ligand binding free energy calculation. During a transition the enzyme bound
ligand is being morphed from the state A to B, whereas the solvated molecule
undergoes a reverse transition from B to A state. The net charge of the simulation
box remains zero. The free energy estimate corresponds to a double free energy
difference: G=G4 G3
a state A molecule bound to an enzyme is positioned in the same

box with the solvated state B molecule. The other end state contains
the same compounds, but in inverted roles: state B molecule is in
contact with an enzyme, while molecule A is solvated. In this way,
molecules A and B are present in the simulation throughout a transition and no net charge change occurs.
It is important to position the solvated (free) and protein
bound molecules at a sufficiently large distance from each other to
prevent direct interaction, because the branches of the thermodynamic cycle, that are now in a single box, are assumed to be
190
independent. In practice, placing the structures3nm apart is sufficient to obtain accurate free energy estimates. To prevent an
interaction between the solvated molecule and the protein due to
motions during the simulation, a position restraint on a single
atom of the molecule free in solution can be imposed. Once a system is set up this way, the estimated free energy corresponds to the
G of binding.
Even if a charge changing mutation/modification is set up to
remain in a neutral simulation box during an alchemical transition, some unwanted electrostatic artifacts may persist due to the
finite size and periodicity effects [62].
5 Validation Using Closed Thermodynamic Cycles

To achieve accurate and valuable free energy estimates from simulations, it is important to aim at the best possible setup within the
approximations made. If and when available, known experimental
results can be used to validate a particular setup but even then, there
can be multiple reasons for a discrepancy between experiments and
simulations. For instance, it is often difficult to completely match
experimental conditions with simulation conditions, e.g. trace
amounts of a certain ion in the buffer could have a large effect, but
are often neglected in simulations. Also, it can never be excluded
that there is an error in the experimental results.
From the simulation side the major factors affecting the accuracy of the results are conformational sampling and the Hamiltonian/
force field. One can validate the setup independent of external
factors such as the force field or known experimental results, by utilizing closed thermodynamic cycles where all transition branches are
accessible and are explicitly computed. The result of summing up all
branches should then be zero. Any deviation from zero gives access
to the accuracy of the approach used. The simplest type of a closed
cycle is using the double-system/single-box setup (see Subheading 4),
where in one box A2B and B2A are present. A2B can be the transition of one amino acid to another in the case of protein binding, or
can be the change from one ligand to another in the case of ligand
binding.
The double system in a single box approach is used in the following example to obtain a closed cycle. A transition from leucine to
alanine (L2A) is combined with the reverse transition from alanine
to leucine (A2L) in a single system. Ten independent equilibrium
simulations of 10ns are launched and convergence is verified as a
function of the number of independent trajectories involved in
computing the free energy. Using more independent simulations
reduces the error bar. Note that using only one 10ns trajectory
can produce a result that is close to zero, but the error bar is large.
In this case, an additional 10ns of simulation can increase the
191
Fig. 4 (a) Closed thermodynamic cycle involving three states. (b) Including the
effects of the dummy atoms a closed thermodynamic cycle with three states is
transformed to one with six states
eviation from zero, but thereby indicates that 10ns of sampling is

d
not sufficient for the simulation to converge. Combining all ten
simulations results in a small error bar (with zero within).
Thus, using short simulation times may often result in seemingly good results for closed thermodynamic cycles, since in
that case all simulations stay close to the starting structure and
the conformational space is insufficiently explored. Assuming that
sufficient exploration of the conformational space is required,
using longer simulations would be preferred. However, using longer simulations also holds the risk that they will get trapped in
artificial minima caused by force field artifacts. To avoid this we
recommend using multiple short simulations as done in this example. This approach ensures better sampling and reduces the risk of
getting trapped in artificial minima. Furthermore, a more rigorous
and straightforward error estimation could be performed using
this approach, provided that a sufficient number of transitions is
achieved for each independent simulation. In such a case the free
energy can be calculated for each trajectory separately and the error
can be evaluated for the deviation among the independent G
estimates.
In the example above we used a convenient double-system/
single-box setup. However, in case of a closed cycle is constructed
by considering the branches of a thermodynamic cycle separately,
one additional caveat in terms of topology construction needs to
be taken into account. Both single and dual topology procedures
mostly involve dummy atoms, and those need to be considered
in a closed cycle. A simple thermodynamic cycle, as represented in
Fig.4a, containing three vertices, might end up in a cycle with six
vertices (Fig.4b), and edges containing only dummy transitions
that are not easily accessible [40, 41]. In a typical free energy simulation one is often interested in the difference between two G
values, where the contribution of the dummy atoms cancels out,
e.g. in protein thermostability calculations the effect of the dummy
atoms is the same in the reference (unfolded) state as in the folded
state, or for ligand binding affinities the effect is the same for the
ligand in solvent as the ligand bound to a protein. Therefore, one
possibility for the construction of a valid closed thermodynamic
192
Fig. 5 Results for a double system in a single box, having both L2A and A2L
structures in one box. The number of equilibrium simulations of 10ns used can
be read from the x-axis. Each point on the graph consists of 100 non-equilibrium
trajectories of 50ps
cycle with three vertices, might be by having G values at the

edges instead of G. As a simpler alternative, the double-system/
single-box setup as shown in the example above could be used
(Fig.5).
6 Non-equilibrium Free Energy Calculation Setup

In this section we will outline the main steps of the alchemical non-
equilibrium free energy calculation setup. The described protocol
is compatible with the G estimation based on both the Jarzynski
equality and CFT.In the first step of the procedure, equilibrium
ensembles of the system in both end states need to be generated.
Afterwards, short transition simulations are performed starting
from the structures selected from the equilibrium ensembles. The
work performed by the system is calculated by numerically integrating the H / l curves for every transition. Finally, the free
energy difference between the two end states is estimated from the
accumulated work values.
To create equilibrium ensembles for both end states of a transition, simulations of the system with the hybrid topology
(Subheading 3) can be performed by keeping the transition controlling variable constant at one of the two end states. This has
an advantage that structures from the equilibrium simulations can
directly be used as starting structures for the transitions. On the
193
other hand, if equilibrium ensembles generated without a hybrid

topology are already available, they can be re-used, e.g. when the
same residue of a protein needs to be mutated to more than one
target, or if an alternative method is used to generate the equilibrium ensemble which is not compatible with the use of hybrid
topologies. In these cases, the hybrid topology can be introduced
into structures after generating the equilibrium ensembles with a
regular (non-hybrid) topology, but it is advisable to perform
a short simulation with the hybrid topology prior to a transition, to
allow the system to relax.
For the non-equilibrium transitions, a number of structures
(from 100 as in published works [35, 48] up to 450 as in the ATP
example described later in this chapter) are picked from both equilibrium ensembles. For each of these structures, a short (usually on
the order of 50200ps) simulation is performed, during which the
topology is changed from one state to the other. For more details
on the simulation parameters see Note 1.
The generalized force H / l associated with this transition is
calculated and written out by the simulation program.
The work value for each transition can be obtained by numerical integration of the H / l curves. Then, two histograms are
created from all work values associated with transitions in a single
direction (Fig.6). The CGI, BAR, or another method (Subheading
2.4) can be employed to extract the free energy difference from the
work distributions.
The computational time necessary to calculate free energy
differences using non-equilibrium approach depends on several
factors, including first of all the size of the system and the significance of the change, but also the desired precision. It might be
preferable to get fast approximate results in a screening process,
while more computational effort would be spent to obtain a p
recise
value for a specific mutation. The quality of the simulation protocol
can be validated using closed thermodynamic cycles (Subheading 5),
but a closer look at the transition work distributions may also indicate how the results can be improved.
As the CFT is based on the assumption that the transition runs
start from an equilibrium ensemble, improving the sampling of
these ensembles usually yields the greatest improvement to the
quality of the result. This is best achieved by running several parallel simulations than a single long one as molecular dynamics simulations tend to get stuck inlocal energy minima.
The difference in work values between the forward and backward transitions is caused by the fact that the simulations are
performed in non-equilibrium conditions. The work distributions
for forward and backward transitions will be closer to each other
the slower the transitions are. Hence, increasing the length of these
simulations would improve the result if the work distributions are
not properly overlapping.
194
Fig. 6 Non-equilibrium free energy calculation setup. Two equilibrium ensembles

are generated (e.g. by MD simulations) for the system in state A and B. From
these ensembles, snapshots are selected and used to set up short non-equilibrium
transition simulations, in which the state A is turned into B and vice versa
7 Trypsin Inhibitors
Binding affinity estimation for small organic compounds is of high
importance in the search for potential drug candidates. Hence, we
will analyze in more detail a study of alchemical G calculations
for a set of trypsin inhibitors. Talhout etal. [63] performed isothermal calorimetry (ITC) measurements of the binding free energies
for a number of p-n-alkylbenzamidinium molecules. We will use this
set of ligands to illustrate the workflow of the alchemical ligand
binding free energy calculations, and the ITC measurements will
serve us as a reference to assess the quality of our estimates. More
information on the computational studies on this set of trypsin
inhibitors can be found in the publications [45, 63].
195
Fig. 7 Trypsin inhibitor analysis. (a) A set of alkylbenzamidinium molecules. Molecule ordering corresponds
to the pairs of ligands for which the free energy differences were calculated. (b) Structure 3PTB with a cocrystallized benzamidine served as a starting structure for the MD simulations. (c) G values from the ITC
measurements [63] and calculated estimates
7.1 Topology
andStarting Structure
Before starting the ligand topology generation, the molecules were

superimposed on the atoms comprising a common scaffold. In the
current example, carbons of a rigid benzene ring were well suited
for the superposition. Subsequently, the order for the alchemical
morphs was established: one of the possibilities is shown in Fig.7a.
Growing an alkyl chain on a benzamidinium several carbons at a
time ensured a sufficient phase space overlap between the end state
ligands. The ligand topologies were generated using the GAFF
methodology [56]. A single topology approach was used to establish a mapping between the molecules. Atoms that ought to be
transformed into one another were identified using a distance criterion after structural superpositioning.
The starting structure for the simulations (Fig.7b) was obtained
from a crystal structure in the Protein Data Bank (id 3PTB [64]).
Since the crystal structure contained a co-crystallized benzamidine
molecule, all the ligands of interest were superimposed onto the
experimentally determined structure. Such a construction of a starting structure rests on an assumption that the analyzed compounds
share a similar binding pose with the benzamidine moiety. In case a
co-crystallized ligand is not available or the binding pose is not well
196
defined, additional analysis is needed. For example, a molecular

docking combined with a long time scale molecular dynamics simulation can be used to establish a starting structure with a reliable
ligand pose for the alchemical free energy calculations.
7.2 Simulation Setup
The thermodynamic cycle for the binding G estimation was

constructed by considering the transitions between the two ligand
states separately: in water and in contact with trypsin. Such a simple cycle construction was possible, since no charge change was
involved in any of the transitions.
Firstly, 10ns equilibrium molecular dynamics simulations were
performed at the end states for the ligands in water and in the
bound form. While this time scale does not cover large conformational motions or significant changes in the binding pose, it provides an equilibrium sampling in a free energy minimum close to
the starting structure. For the cases, where in a simulation time the
system travels far from its initial point, longer sampling times may
be required to cover the relevant phase space volume. The non-
equilibrium transitions were spawned from the snapshots extracted
from the equilibrium trajectories. Here, 100 simulations (50ps
each) were performed going in both directions: A to B and B to A.
See also Notes 2 and 3 for more technical details on the equilibrium sampling and non-equilibrium transitions, respectively.
7.3 Estimation
oftheFree Energy
Differences
Integration over the H / l curves to obtain the work values and

subsequent application of the CGI method enabled calculation
A B
of the G values for the parts of the thermodynamic cycle: DG water
A B
(see Note 4). The final G estimates are shown in
and DG trypsin
Fig.7c. A positive value in the graph indicates that the ligand in state
B is a weaker binder in comparison with the state A molecule.
For all the ligand pairs the difference between the calculated
and experimentally measured free energy values was smaller than
1kcal/mol. In most cases the G values were estimated very
accurately, and only the transition from benzamidine to methyl-
benzamidine showed a discrepancy of1kT.Larger errors for the
estimated free energy differences can be observed for the longer
alkyl chains. This effect is a nice illustration of the increased uncertainty in the calculated results due to the larger dissipated work
values. Alchemically grown longer chains face stronger hindrance
and steric clashes, leading to an increase in the amount of work
dissipated along the path. The errors could be reduced by performing the transitions slower (e.g., 100ps for a transition).
In the current example the results were presented as the double
differences in free energy. In practice, however, an estimate of the G
values may be of interest. Conversion from the G to the single
free energy differences can be attained by setting one of the G
estimates to an experimental value (e.g., benzamidinium ITC
197
Fig. 8 G values calculated by propagating the double free energy differences

(G) along a chain of ligands (see Fig.7a) The results are less accurate than
the G values due to the error accumulation
easurement) and propagating the differences together with the

m
associated errors along the chain of compounds (Fig.8). It is clear
that the estimates appear far more inaccurate due to the accumulation of errors. This result illustrates the importance of an appropriate
simulation setup addressing a question of interest. For the case where
G estimates are to be compared to the experimental measurements,
a single reference structure could be used to create the pairs for transitions. Although the perturbations in such a scenario would be larger
than when considering a chain of ligands, the problem of error accumulation could be avoided. In addition, such a setup would allow for
internal consistency checks via closed thermodynamic cycles.
8 ProteinProtein Binding: Binding of-Chymotrypsin withIts Inhibitor

Turkey Ovomucoid Third Domain
In this example the change in binding free energy upon mutation
in the complex of the Kazal-type Ovomucoid from Turkey
(OMTKY3) and -chymotrypsin (CHT) is computed. Leucine 18
of OMTKY3 is mutated to A,I,W,F,Y,V,M,S, and T residues. The
results are compared to both experimental data [65] and other
theoretical predictions [66]. Note that this example is well suited
as a practical hands-on test for the principles learned in this chapter. The structure is available in the Protein Data Bank (id 1CHO
198
Fig. 9 The complex of Kazal-type Ovomucoid from Turkey (OMTKY3) and

-chymotrypsin (CHT). Leucine at position 18 is mutated. The structure is taken
from the pdb-database (PDB id 1CHO)
Fig. 10 Thermodynamic cycle used for proteinprotein binding
[67]), and calculations can be set up using freely available software

(pmx [48] and Gromacs [47]).
The CHT:OMTKY3 complex is shown in Fig.9, the position
(L18) that is mutated is at the interface of the two binding partners.
The thermodynamic cycle in Fig.
10 was used to compute
the binding free energy. In this cycle the reference is OMTKY3.
The shift in binding free energy upon mutation (G) is computed
from the difference between G1 and G2, since computing the
difference between G3 and G4 directly would be computationally too expensive.
Calculated binding free energy (kJ/mol)
30
199
Expected
1 kcal/mol dev.
25
2 kcal/mol dev.
20
15
L18V
L18A
L18S
10
5
L18W
L18M
L18I
L18F
5
10
L18T
L18Y
10
5
15
0
10
Experimental binding free energy (kJ/mol)
20
25
Fig. 11 Results of mutations in OMTKY3:CHT complex. The average unsigned

error is 7.9kJ/mol and correlation coefficient is 0.83
Setting up these simulations consists of four phases. First the

topology is prepared, next the equilibrium simulations are performed,
followed by non-equilibrium simulations and finally the results are
analyzed to extract the free energy. This procedure is performed
for both the complex and for free OMTKY3in solvent. The topology is initially prepared using the Gromacs [47] pdb2gmx
tool and the Amber99sb [68] force field. Equilibrium simulations
of 20ns are performed for the native protein and all mutants.
From those equilibrium simulations, 100 snapshots are selected
and hybrid structures and topologies are created using pmx [48]
(see Note 5). Those are followed by non-equilibrium transitions
of 100ps each. The free energies are extracted using the BAR
method [14] (see Note 4). BAR was used in this case, since the
CGI method requires a Gaussian distribution of the work values,
which is not always the case for all mutations in this complex. The
results of the calculations are shown in Fig.11. The correlation
coefficient with the experiment is 0.83 and the average unsigned
error is 7.9kJ/mol. Particularly, the mutants with the largest destabilization are p
redicted poorly. The same mutations were predicted
by Benedix etal. [66], where a slightly better result was obtained,
with a correlation coefficient of 0.9 and an average unsigned error of
7.9kJ/mol.
200
When comparing to experimental results, their origin has to

be taken into account. The experimental results [65] used here are
obtained by enzyme assays, where the binding affinity is not measured directly, but instead the inhibition of proteolysis is measured.
The inhibition of proteolysis might be related to the binding
affinity of OMTKY3, but other factors might play a role, e.g., a
certain mutation could trigger a conformational change in CHT
which inhibits proteolysis, without having a stronger binding affinity than another mutant.
9 ATP andMagnesium Complex inSolution

In the example given here, we show how seemingly simple free
energy calculations involving a highly charged ligand (ATP and a
bound magnesium) can result in substantial inaccuracies if not
treated correctly. Many biological functions require the presence of
an ATP molecule which is often only active with a bound magnesium ion. The binding free energy of an ATPMg complex can be
calculated via a thermodynamic cycle where the complex is annihilated once from the binding pocket of the protein and once in an
unbound (free) state which is represented by an annihilation of the
ATPMg in water. The second transformation is in fact the solvation free energy. Here, we present some of the pitfalls that might
occur merely in the calculation of the solvation free energy of
ATPMg, as it presents a challenging system by itself and may be
found in much more complex systems as well (e.g., in protein
complexes).
9.1 System Setup:
Charged Systems
The annihilation of an ATP4 molecule, even when attached to a

Mg+2 ion, incorporates a change in the overall system charge which
would influence the resulting free energy. One possibility to neutralize the system is to add 2 counter Na+ ions that would turn into
dummies during the transition as well. However, this effectively
turns the free energy of ATPMg solvation, into a combined solvation free energy of [ATPMg] + 2Na. Moreover, an issue encountered in equilibrium free energy calculations of this system is that
the interaction of the disappearing Na+ ions with the disappearing
ATPMg+2 is not sufficiently sampled in the simulated time scales.
A second and more computationally expensive possibility that
avoids the coupling of the counter ions to the transition is the
double-system/single-box setup (Subheading 4). For this example,
one ATPMg complex is annihilated and another is solvated in the
same box, while the structures are positioned4nm apart in order
to avoid any interaction between them. The Mg+2 ion itself cannot
migrate between the ATP molecules when its interactions are
turned off, because it is kept close to its respective ATP via a distance restraint. Topology generation is relatively simple, where A
201
Fig. 12 (a) Non-equilibrium free energy calculation of ATP solvation and desolvation which should serve as a
closed cycle. The left section of the plot shows the work values as a function of the transition number. Since
the transitions are started from consecutive frames it could indicate whether the equilibrium trajectory is drifting. The right region of the plot combines the work values into histograms. (b) Distribution of ATP and Mg
orientations measured according to distance from P and P. P is not shown since its distance does not vary
and fluctuates around 0.3nm
and B states are constructed such that one molecule is present at

=0 while the second is annihilated (turned to dummies), and vice
versa for =1. This type of setup allows either for directly computing the binding free energy in one simulation box (if the ligand is
annihilated while bound to a protein on one side of a box, and
solvated on the other side), or to compute a closed cycle as a consistency check. In this example we performed the latter.
9.2 Calculating
aClosed Cycle
Having constructed the system as described above, 50ns equi

librium ensembles were obtained for =0, =1 and, afterwards,
fast non-equilibrium transitions were performed 200ps each.
The results are shown in Fig.12a, indicating that the distributions
of the resulting work values do not overlap nor fit into a gaussian
function, whereas the estimated G is far from zero.
How is it possible that a transition and its opposite do not
result in a zero free energy change? Sampling and symmetry are
key issues here. The two ATPMg complexes were found not to be
entirely symmetric in the conformations covered by their equilibrium ensembles. Specifically, it is the Mg+2 ion that adopts different
orientations relative to the phosphates of the ATP.In Fig.12b one
can differentiate between two states, which are mutually exclusive
for the two ATP molecules in the box: state X, where the Mg+2 is
triply coordinated by all three phosphates that adopt a similar distance of 0.3nm to it, and state Y, where Mg+2 is doubly coordinated and found slightly further away from the P (0.45 to
0.55nm). The lower cluster in state Y, where the distance to P
202
Fig. 13 An extended closed cycle of an ATPMg complex annihilated and solvated in solution. The horizontal
arrows are performed in both directions simultaneously for a closed cycle consistency check
shortens, arises from the initial structure, but it gradually equilibrates into the top cluster of state Y within 120ns and never
re-visits the former cluster. Since these conformations do not inter
change, they can be seen as two distinct chemical species. Thus,
turning one chemical species on, and another off does not r epresent
two opposite reactions and will not converge to G=0.
Spontaneous transitions between these conformations do not
occur in MD simulations with a length of 100ns, and they are kept
separated by a steep barrier. However, the barrier is solely maintained
by the highly attractive Coulomb interactions between the phosphates
and the magnesium. Since the free energy is a state function, the
path from the fully solvated to annihilated (uninteracting dummies)
state could be chosen to pass through a reduced charge state, where
the charge on the magnesium ion is scaled to +1, while the charge
onthe phosphate groups is scaled in reverse to maintain the neutrality
of the system. This effectively creates two transitions that need to be
calculated: one from a fully appeared and fully charged state to a
reduced charge state, and then another one into dummies.
To enable the convergence of the closed cycle of annihilating
and solvating an ATPMg complex in solution, the calculation is
decomposed further into the two ATP orientations, such that transitions (annihilation or solvation) will be performed for each orientation (X and Y ) of the ATP1 as depicted by the horizontal arrows
in Fig.13. There is no need to compute the vertical transitions, but
we would like to note that both vertical transitions maintain a
G0. In the reduced charge state, conformations X and Y interchange multiple times, while they are equally populated. As for the
dummy state, the difference between X and Y is merely a difference
of orientation in space, which is imposed by a distance restraint.
The distribution of work values for each of the four horizontal
transitions shown in Fig.13 is plotted in Fig.14. These transitions
are as before, performed in both directions to close a thermodynamic cycle. Four hundred and fifty transitions were performed in
1
Distance restraints on the P, P and the Mg+2 will keep the atoms in their
respective orientation in one of the two top clusters shown in Fig.12b.
Fig. 14 Results from non-equilibrium transitions as depicted by the horizontal arrows in Fig.13. The snapshot
number represents the transitions
each direction, their length was increased to 2ns. Indeed, this time
the resulting G values are mostly within 1.5kJ/mol away from
zero2, except for the last transition (reduced charge state into dummies in state Y ). This state is much more flexible than its fully
charged version, and might need more time to converge, i.e. via
longer equilibrium simulations, however, electrostatic artifacts
could remain due to the nature of the mutation (turning off a
charge) and the finite size of the system. In principle, determining
whether the generated equilibrium ensemble is indeed at equilibrium is difficult. However, sometimes examining the series of work
values taken from consecutive initial snapshots from the equilibrium ensemble may indicate whether there is a drift and whether
more equilibration time is needed. For example, if we would only
have a quarter of the equilibrium trajectory of state X +2 to +1 and
the corresponding transitions (first 120 transitions in Fig.14), the
drift in work values might have hinted at a trajectory drift, but
further equilibration and transitions from later snapshots show that
those work values are reproduced again and again.
2
A shorter transition time of 200ps also gave a result that was fairly close to
zero, but longer times were used to reduce the error.
204
Table 1
Parameters and suggested values for the non-equilibrium free
energy calculations
Parameter
Value
init-lambda
0 or 1
delta-lambda (equilibration)
delta-lambda (transition)
1/nsteps
nstdhdl
sc-coul
yes
sc-alpha
0.3
sc-sigma
0.25
sc-power
Having arrived at a consistent zero free energy change for a

closed cycle of ATP in solution, one can be more confident in the
accuracy of ATPMg annihilation in protein complexes. For practical purposes, a similar conformation decomposition can be done
when computing the binding free energy of an ATPMg complex
in a protein. As indicated earlier, in such a case, the box will contain
ATPMg and a protein with ATPMg in its binding pocket which
will be simultaneously turned on and off. This time, rather than
receiving a zero from each transition (and an overall zero from
ATPMg annihilation and resolvation in both orientations), the
combined transition from a fully charged state to a dummy state
would yield the binding free energy of ATPMgX and ATPMgY .
It can then be combined using the following formula [69]:

DG = - b -1 ln[exp(- bDG X ) + exp(- bDGY )]
(24)
10 Notes
1. In the Gromacs simulation package the free energy code is
activated by setting the flag free-energy=yes in a molecular dynamics parameter (mdp) file. Triggering this option automatically enables the H / l output to an external file. The
initial state of a system is set by defining init-lambda to
be equal to 0 or 1 for the states A and B, respectively. Setting
the two aforementioned parameters is sufficient to perform an
equilibrium sampling simulation at one of the end states. For
the transition runs, an increment in needs to be specified by
setting the parameter delta-lambda to a non-zero value.
delta-lambda has to be estimated such that an end state is
205
reached within the defined number of integration steps. For

example, for a transition from init-lambda=1 to the end
state =0in 50ps with an integration time step of 2fs, 25,000
integration steps are required. Hence, delta-lambda has to
be set to 4e5. For the non-equilibrium free energy calculation it is important to collect all the available H values,
therefore, the frequency of output should not be reduced, i.e.
nstdhdl=1. We recommend using the soft-core potential
energy function for both the van der Waals and electrostatic
interactions. Table 1 summarizes the essential parameters
for the non-equilibrium free energy calculations as defined in
Gromacs. The suggested values should be adjusted to a particular problem at hand.
2. The equilibration simulations at the end states (=0 and =1)
are performed following the standard unbiased molecular
dynamics simulation setup. If the hybrid topology for a system
is generated, the free energy code must be switched on and the
state has to be defined (see Note 1). It is also important to
consider which ensemble, canonical or NPT, is sampled, since
on this choice depends which free energy will be calculated
Helmholtz or Gibbs. Dispersion correction for the energy and
pressure has a significant effect on the accuracy of estimated
free energies [70, 71] (simulation parameter in Gromacs
DispCorr=EnerPres).
3. When setting up non-equilibrium free energy simulations,

three parameters concerning the simulation length need to be
decided upon. Firstly, the length of an equilibrium sampling
simulation has to be defined. The time dedicated for this step
very much depends on a particular system at hand, and on the
time scale of conformational changes: an equilibration may
vary from a nanosecond to microsecond range. Secondly, the
number of spawned non-equilibrium transitions needs to be
chosen. The starting structures, with the associated velocities,
need to cover the sampled equilibrium ensemble. In practice,
100 transitions in each direction are often sufficient to reach a
converged free energy estimate. The third parameter, the time
spent for a single transition, depends on the scale of perturbation between the states A and B. Extending the time from
50ps up to several nanoseconds per simulation will improve
convergence, as for the slower transitions less work will be dissipated. A convenient method to assess the convergence is by
monitoring the extent of an overlap between the work distributions for the forward and backward transitions: for the larger
overlap free energy estimate is more accurate. While the second and third parameters are responsible for the convergence
and statistical error of the free energy estimate, it is the equilibration time which mostly contributes to the accuracy of the
G estimate.
206
4. The work required for a transition is obtained by numerically

integrating the H / l curves. It is important to bear in mind
that the integration ranges for the forward transition are defined
to be 0 and 1, not to confuse them with the actual simulation
time (which is, e.g., reported in the Gromacs H / l output).
For the reverse transition integration from 1 to 0 has to be performed and afterwards the sign of the work value needs
to be inverted (see Eq.17 in Subheading 2.4). This effectively
allows integration from 0 to 1 without the subsequent sign inversion for the reverse transition. The pmx [48] package contains the
script (analyze_crooks.py) performing the integration using
Gromacs H / l output directly for an arbitrary number of forward and reverse trajectories. Estimate of the free difference using
the CGI approach (Eq.19) or BAR (Eq.20) can also be obtained
with analyze_crooks.py in the pmx package.
5. The hybrid topology for an amino acid mutation can be created

employing the pre-generated mutation database available in
the pmx package. Firstly, the hybrid structural file containing
atoms of both states needs to be generated (script mutate.
py). In the second step, the standard topology generation is
performed using the newly created structure file (pdb2gmx in
Gromacs). As the hybrid structure (e.g., A2L amino acid
denoting alanine in state A and leucine in state B) is not present in the standard biomolecular force fields, one needs to
use the specifically modified force field files provided in the
pmx.Finally, the parameters for the B state of the system need
to be added to the topology file. This can be accomplished
with the make_bstate.py script from the pmx package.
References
1. Torrie GM, Valleau JP (1977) Nonphysical
sampling distributions in Monte Carlo free-
energy estimation: umbrella sampling. J Comput
Phys 23(2):187199
2. Kirkwood JG (1935) Statistical mechanics of
fluid mixtures. J Chem Phys 3:300313
3. Ben-Naim A (1992) Statistical thermodynamics
for chemists and biochemists. Plenum Press,
NewYork
4. Chipot C, Pohorille A (eds) (2007) Free
energy calculations. Theory and applications in
chemistry and biology. Springer series in chemical physics, vol 86
5. Christ CD, Mark AE, van Gunsteren WF (2010)
Basic ingredients of free energy calculations:
a review. J Comput Chem 31(8):15691582.
ISSN 1096-987X. doi:10.1002/jcc.21450.
http://dx.doi.org/10.1002/jcc.21450
6. Michael S, David M (2013) An introduction

to best practices in free energy calculations. In:
Biomolecular simulations: methods and protocols. Methods in molecular biology, vol 924.
Humana Press, NewYork, pp271311
7.
Zwanzig RW (1954) High-temperature
equation of state by a perturbation method.
I. Nonpolar gases. J Chem Phys 22:
14201426
8. Pohorille A, Jarzynski C, Chipot C (2010) Good
practices in free-energy calculations. J Phys
Chem B 114(32):1023510253
9. Landau LD (1938) Statistical physics. The
Clarendon Press, Oxford

10. Lu N, Kofke DA (2001) Accuracy of free-
energy perturbation calculations in molecular
simulation. I.Modeling. J Chem Phys 114:
73037311

11. Lu N, Kofke DA (2001) Accuracy of free-
energy perturbation calculations in molecular
simulation. II.Heuristics. J Chem Phys 115:
68666875
12. Shirts MR, Pande VS (2005) Comparison of
efficiency and bias of free energies computed
by exponential averaging, the Bennett acceptance ratio, and thermodynamic integration.
J Chem Phys 122(14):144107-1144107-16
13. Bennett CH (1976) Efficient estimation of free
energy differences from Monte Carlo data.
J Comput Phys 22(2):245268
14. Shirts MR, Bair E, Hooker G, Pande VS (2003)
Equilibrium free energies from nonequilibrium
measurements using maximum-
likelihood
methods. Phys Rev Lett 91(14):140601
15. Mezei M (1992) Polynomial path for the calculation of liquid state free energies from computer
simulations tested on liquid water. J Comput
Chem 13(5):651656
16. Steinbrecher T, Mobley DL, Case DA (2007)
Nonlinear scaling schemes for Lennard-Jones
interactions in free energy calculations. J Chem
Phys 127:214108
17. Kumar S, Rosenberg JM, Bouzida D, Swendsen
RH, Kollman PA (1992) The weighted histogram analysis method for free-energy calculations on biomolecules. I.The method. J Comput
Chem 13(8):10111021

18. Shirts MR, Chodera JD (2008) Statistically
optimal analysis of samples from multiple equilibrium states. J Chem Phys 129:124105
19. Shenfeld DK, Xu H, Eastwood MP, Dror RO,
Shaw DE (2009) Minimizing thermodynamic
length to select intermediate states for free-
energy calculations and replica-exchange simulations. Phys Rev E 80(4):046705

20. Buelens FP, Grubmller H (2012) Linear-
scaling soft-core scheme for alchemical free
energy calculations. J Comput Chem 33(1):
2533
21. Mazor M, Pettitt BM (1991) Convergence of
the chemical potential in aqueous simulations.
Mol Simul 6(13):14

22. Mitchell MJ, McCammon JA (1991) Free
energy difference calculations by thermodynamic integration: difficulties in obtaining a
precise value. J Comput Chem 12(2):271275
23. Straatsma TP, McCammon JA (1991) Multi
confi
guration thermodynamic integration.
J Chem Phys 95:1175

24.
Jorge M, Garrido NM, Queimada AJ,
Economou IG, Macedo EA (2010) Effect of
the integration method on the accuracy and
computational efficiency of free energy calculations using thermodynamic integration. J Chem
207

25. Bruckner S, Boresch S (2011) Efficiency of
alchemical free energy simulations. II. Impro
vements for thermodynamic integration.
J Comput Chem 32(7):13201333
26. Jarzynski C (1997) Nonequilibrium equality
for free energy differences. Phys Rev Lett
78(14):26902693

27. Cuendet MA (2006) The Jarzynski identity
derived from general hamiltonian or non-
hamiltonian dynamics reproducing NVT or
NPT ensembles. J Chem Phys 125:144109

28. Hummer G (2001) Fast-growth thermodynamic integration: error and efficiency analysis.
J Chem Phys 114:73307337
29. Gore J, Ritort F, Bustamante C (2003) Bias
and error in estimates of equilibrium free-
energy differences from nonequilibrium measurements. Proc Natl Acad Sci USA 100(22):
1256412569
30. Crooks GE (1998) Nonequilibrium measurements of free energy differences for microscopically reversible Markovian systems. J Stat Phys
90(56):14811487
31. Crooks GE (1999) Entropy production fluctuation theorem and the nonequilibrium work
relation for free energy differences. Phys Rev E
60(3):27212726

32. Chelli R, Marsili S, Barducci A, Procacci P
(2007) Recovering the Crooks equation for
dynamical systems in the isothermal-isobaric
ensemble: a strategy based on the equations of
motion. J Chem Phys 126:044502
33. Nanda H, Lu N, Woolf TB (2005) Using non-
Gaussian density functional fits to improve relative free energy calculations. J Chem Phys
122(13):134110-1134110-8
34. Maragakis P, Ritort F, Bustamante C, Karplus
M, Crooks GE (2008) Bayesian estimates of
free energies from nonequilibrium work data in
the presence of instrument noise. J Chem Phys
129:024102
35. Goette M, Grubmller H (2009) Accuracy and
convergence of free energy differences calculated from nonequilibrium switching processes.

36. Bramwell ST, Christensen K, Fortin J-Y,
Holdsworth PCW, Jensen HJ, Lise S, Lpez
JM, Nicodemi M, Pinton J-F, Sellitto M
(2000) Universal fluctuations in correlated systems. Phys Rev Lett 84(17):37443747
37. Massey FJ Jr (1951) The Kolmogorov-Smirnov
test for goodness of fit. J Am Stat Assoc
46(253):6878
38. Pearlman DA, Kollman PA (1991) The overlooked bond-stretching contribution in free
energy perturbation calculations. J Chem Phys
94:45324545
208
39. Pearlman DA (1994) A comparison of alternative approaches to free energy calculations.

J Phys Chem 98(5):14871493
40. Boresch S, Karplus M (1999) The role of bonded
terms in free energy simulations. 2. Calculation
of their influence on free energy differences of
solvation. J Phys Chem A 103(1):119136

41. Boresch S, Karplus M (1999) The role of
bonded terms in free energy simulations: 1.
Theoretical analysis. J Phys Chem A 103(1):
103118
42. Bash PA, Singh UC, Langridge R, Kollman PA
(1987) Free energy calculations by computer
simulation. Science 236(4801):564568
43. Beutler TC, Mark AE, van Schaik RC, Gerber
PR, van Gunsteren WF (1994) Avoiding singularities and numerical instabilities in free
energy calculations based on molecular simulations. Chem Phys Lett 222(6):529539. ISSN
0009-2614
44. Zacharias M, Straatsma TP, McCammon JA
(1994) Separation-shifted scaling, a new scaling method for Lennard-Jones interactions in
thermodynamic integration. J Chem Phys 100:
90259031
45. Gapsys V, Seeliger D, de Groot BL (2012)
New soft-core potential function for molecular
dynamics based alchemical free energy calcu
lations. J Chem Theory Comput 8(7):
23732382

46. Tosco P, Balle T, Shiri F (2011) Open
3DALIGN: an open-source software aimed at
unsupervised ligand alignment. J Comput
Aided Mol Des 25(8):777783
47. Hess B, Kutzner C, van der Spoel D, Lindahl E
(2008) GROMACS 4: algorithms for highly
efficient, load-balanced, and scalable molecular
simulation. J Chem Theory Comput 4(3):
435447
48. Seeliger D, De Groot BL (2010) Protein thermostability calculations using alchemical free
energy simulations. Biophys J 98(10):2309
2316. ISSN 0006-3495
49. OBoyle NM, Banck M, James CA, Morley C,
Vandermeersch T, Hutchison GR (2011) Open
Babel: an open chemical toolbox. J Cheminf
3(1):114

50. Sadowski J, Gasteiger J, Klebe G (1994)
Comparison of automatic three-dimensional
model builders using 639 X-ray structures.
J Chem Inf Comput Sci 34(4):10001008
51. Singh UC, Kollman PA (1984) An approach to
computing electrostatic charges for molecules.
52. Bayly CI, Cieplak P, Cornell W, Kollman PA
(1993) A well-behaved electrostatic potential
based method using charge restraints for deriving

atomic charges: the RESP model. J Phys Chem
97(40):1026910280
53. Jakalian A, Bush BL, Jack DB, Bayly CI (2000)
Fast, efficient generation of high-quality atomic
charges. AM1-BCC model: I. Method. J Comput
Chem 21(2):132146
54. Mobley DL, Dumont , Chodera JD, Dill KA
(2007) Comparison of charge models for fixed-
charge force fields: small-molecule hydration
free energies in explicit solvent. J Phys Chem B
111(9):22422254

55. Wang J, Wang W, Kollman PA, Case DA
(2006) Automatic atom type and bond type
perception in molecular mechanical calculations. J Mol Graph Model 25(2):247260
56. Wang J, Wolf RM, Caldwell JW, Kollman PA,
Case DA (2004) Development and testing of a
general amber force field. J Comput Chem
25(9):11571174
57. Vanommeslaeghe K, Hatcher E, Acharya C,
Kundu S, Zhong S, Shim J, Darian E, Guvench
O, Lopes P, Vorobyov I, MacKerell AD Jr
(2010) CHARMM general force field: a force
field for drug-like molecules compatible with
the CHARMM all-atom additive biological
force fields. J Comput Chem 31(4):671690
58. Vanommeslaeghe K, MacKerell AD Jr (2012)
Automation of the CHARMM General Force
Field (CGenFF) I: bond perception and atom
typing. J Chem Inf Model 52(12): 31443154
59. Vanommeslaeghe K, Raman EP, MacKerell AD
Jr (2012) Automation of the CHARMM
General Force Field (CGenFF) II: assignment
of bonded parameters and partial atomic charges.
J Chem Inf Model 52(12):31553168
60. Malde AK, Zuo L, Breeze M, Stroet M, Poger
D, Nair PC, Oostenbrink C, Mark AE (2011)
An automated force field topology builder
(ATB) and repository: version 1.0. J Chem
61. Ribeiro AAST, Horta BAC, de Alencastro RB
(2008) MKTOP: a program for automatic construction of molecular topologies. J Braz Chem
Soc 19(7):14331435
62. Rocklin GJ, Mobley DL, Dill KA, Hnenberger
PH (2013) Calculating the binding free energies of charged species based on explicit-solvent simulations employing lattice-sum
methods: an accurate correction scheme for
electrostatic finite-size effects. J Chem Phys
139(18):184103.
63. Talhout R, Villa A, Mark AE, Engberts JBFN
(2003) Understanding binding affinity: a
combined isothermal titration calorimetry/
molecular dynamics study of the binding of a

series of hydrophobically modified benzamidinium chloride inhibitors to trypsin. J Am
Chem Soc 125(35):1057010579
64. Marquart M, Walter J, Deisenhofer J, Bode W,
Huber R (1983) The geometry of the reactive site
and of the peptide groups in trypsin, trypsinogen
and its complexes with inhibitors. Acta
Crystallogr Sect B Struct Sci 39(4):480490
65. Lu W, Apostol I, Qasim MA, Warne N, Wynn R,
Zhang WL, Anderson S, Chiang YW, Ogin E,
Rothberg I, Ryan K, Laskowski M (1997)
Binding of amino acid side-chains to S1 cavities of
serine proteinases. J Mol Biol 266(2): 441461
66. Benedix A, Becker CM, de Groot BL, Caflisch
A, Bckmann RA (2009) Predicting free
energy changes using structural ensembles. Nat
Methods 6(1):34
67. Fujinaga M, Sielecki AR, Read RJ, Ardelt W,
Laskowski M, James MNG (1987) Crystal
and molecular structures of the complex of
-chymotrypsin with its inhibitor turkey ovomucoid third domain at 1.8 resolution.
J Mol Biol 195(2):397418
209

of multiple Amber force fields and development of improved protein backbone parameters. Proteins Struct Funct Bioinform 65(3):
712725
69. Mobley DL, Chodera JD, Dill KA (2006) On
the use of orientational restraints and symmetry corrections in alchemical free energy cal
culations. J Chem Phys 125(8):084902. doi:
10.1063/1.2221683.
http://link.aip.org/
link/?JCP/125/084902/1
70. Shirts MR, Pitera JW, Swope WC, Pande VS
(2003) Extremely precise free energy calculations of amino acid side chain analogs: comparison of common molecular mechanics
force fields for proteins. J Chem Phys 119(11):
57405761
71. Shirts MR, Mobley DL, Chodera JD, Pande
VS (2007) Accurate and efficient corrections
for missing dispersion interactions in mole
cular simulations. J Phys Chem B 111(45):
1305213063
Part II
Conformational Change
Chapter 10
The Use ofExperimental Structures
toModel Protein Dynamics
AtaurR.Katebi, KannanSankar, KejueJia, andRobertL.Jernigan
Abstract
The number of solved protein structures submitted in the Protein Data Bank (PDB) has increased dramatically in recent years. For some specific proteins, this number is very highfor example, there are over 550
solved structures for HIV-1 protease, one protein that is essential for the life cycle of human immunodeficiency virus (HIV) which causes acquired immunodeficiency syndrome (AIDS) in humans. The large
number of structures for the same protein and its variants include a sample of different conformational
states of the protein. A rich set of structures solved experimentally for the same protein has information
buried within the dataset that can explain the functional dynamics and structural mechanism of the protein. To extract the dynamics information and functional mechanism from the experimental structures, this
chapter focuses on two methodsPrincipal Component Analysis (PCA) and Elastic Network Models
(ENM). PCA is a widely used statistical dimensionality reduction technique to classify and visualize high-
dimensional data. On the other hand, ENMs are well-established simple biophysical method for modeling
the functionally important global motions of proteins. This chapter covers the basics of these two.
Moreover, an improved ENM version that utilizes the variations found within a given set of structures for
a protein is described. As a practical example, we have extracted the functional dynamics and mechanism
of HIV-1 protease dimeric structure by using a set of 329 PDB structures of this protein. We have
described, step by step, how to select a set of protein structures, how to extract the needed information
from the PDB files for PCA, how to extract the dynamics information using PCA, how to calculate ENM
modes, how to measure the congruency between the dynamics computed from the principal components
(PCs) and the ENM modes, and how to compute entropies using the PCs. We provide the computer
programs or references to software tools to accomplish each step and show how to use these programs and
tools. We also include computer programs to generate movies based on PCs and ENM modes and describe
how to visualize them.
Key words HIV-1 protease, Principal component analysis, Elastic network model, Protein dynamics,
Acquired immunodeficiency syndrome, Protein data bank
1 Introduction
There are large numbers of structures in the protein data bank
(PDB [1]) for many categories of enzymes. Shown in Fig.1 are the
most abundant enzyme structures ordered by enzyme commission
(EC) numbers. Some other examples for individual EC categories,
213
214
Ataur R. Katebi et al.
Fig. 1 Numbers of related protein structures available for extracting protein functional dynamicssnapshot of the PDB statistics for the largest categories of
enzymes (08/30/2013). In total, there are over 17,000 enzyme structures, and a
significant number of structures for many diverse enzyme types. The most common structure on the left of this histogram with 1,285 structures is EC 3.2.1.17 that
includes lysozymes, and at the right side is 5.2.1.8 acetylcholinesterases with 337
different structures (taken from enzyme classification data provided by PDB: http://
www.pdb.org/pdb/statistics/histogram.do?mdcat=entity&mditem=pdbx_
ec&name=Enzyme%20Classification) [1])
with the numbers of their related structures in parentheses are:

3.4.21: Serine endopeptidases (2,459), 3.4.23: Aspartic endopeptidases (1,146), 3.4.24: Metalloendopeptidases (727), 3.4.22:
Cysteine endopeptidases (720), 3.4.11: Aminopeptidases (292),
3.4.19: Omega peptidases (244), 3.4.17: Metallocarboxypeptidases
(144), 3.4.14: Dipeptidyl-peptidases (120), 3.4.25: Threonine
endopeptidases (109), 2.7.7, Nucleotidyltransferases (107),
3.4.21: Serine endopeptidases (105), 3.4.16: Serine-type carboxypeptidases (97), 2.7.7: Nucleotidyltransferases (106), 3.4.23:
Aspartic endopeptidases (77), and 3.4.19: Omega peptidases (58).
In addition, there are many structures of non-enzyme
Experimental Protein Dynamics
215
proteinsstructural proteins, immunoglobulin Fabs, viral proteins,

and many others. The PDB has many additional ways to search for
functionally related structures that are invaluable for finding structures with similar dynamics. You can search by biological process
such as gene ontology (GO), cellular component, molecular function, and transporter classification. In addition there are many
receptors with multiple reported structures. Overall, there is abundant data to investigate functional protein dynamics of many classes
of proteins directly from experimental structures.
Important conformational changes can readily be extracted
from a set of PDB structures for a protein and these are found to
relate directly to function. Experimental structures can be a rich
source of information. It is well established that functionally related
structures must have similar structures and similar dynamics
building on the broad experience of many researchers. There have
been several efforts at extracting dynamics from specific sets of
experimental structures. One approach is principal component
analysis (PCA) [24], a statistical method based on covariance
analysis. PCA can transform the original space of correlated
variables into a greatly reduced space of independent variables
(i.e., the principal components or PCs). By performing PCA, most
of a systems variance will usually be captured in a quite small subset of the PCs. PCA has been applied often to analyze trajectory
data from MD simulations to find the essential dynamics [5, 6].
Teodoro etal. applied PCA to the dataset composed of many
conformations for HIV-1 protease [7, 8]. They found that PCA
transformed the original high-dimensional representation of protein
motions into a low-dimensional one that provides the dominant
protein motions. This is a huge reduction in dimensionality from
hundreds of thousands to fewer than 50 degrees of freedom. Howe
[9] used PCA to classify the structures in NMR ensembles automatically, according to correlated structural variations, and the
results have shown that two different representations of the protein structure, the C coordinate matrix and the CC distance
matrix, gave equivalent results and permitted the identification of
structural differences between conformations. More recent efforts
include our own previous efforts in analyzing the HIV-1 protease
set [10], those of the Bahar group [11], and our efforts in developing the MAVEN program [12], as well as related efforts by the
Bahar group with their ProDy [13], and Grant with his Bio-3D
[14]. Any of these can provide a similar set of starting tools.
On the other hand, the Elastic Network Models (ENM) have
proven themselves to be highly useful in representing the global
motions for a wide variety of diverse protein structures [1519].
For modeling and simulating the dynamics of proteins, ENMs can
be applied on multiple scales [2023]. All atom ENM models give
a finer description of protein dynamics. The most common coarse-
graining involves a single-site per residue representation, in which
216
the sites are identified by the C atoms and connected by uniform

springs. The dynamics of such interconnected model can be
described by the Gaussian Network Model (GNM) [17] or the
Anisotropic Network Model (ANM) [15]. GNM has been very
successful in yielding information on the magnitudes of the fluctuations of the protein structures but provides no directional information or the 3-D nature of motion of the protein is considered in
the model. However, in reality protein fluctuations are generally
directional and anisotropic [24, 25]. ANM considers the anisotropy of the protein structure in modeling its dynamics and thus
ANM computed collective motions are more relevant to biological
function and mechanism of the protein molecule.
In this chapter, we give an example of how to use computational methods to extract protein dynamics from a large set of
experimental structures of HIV-1 protease. Behind this is the
implicit assumption that there is a significant amount of information about protein dynamics, mechanisms and allostery buried
within the structures in the PDB.We will show how to utilize PCA
to extract dynamics from the abundantly available HIV-1 protease
structures and how to compute the agreement between PCA-
based protein motion and the ANM modeled motion, and describe
how these could be used in simulations with a new structure-based
elastic network model.
2 Theory
2.1 Principal
Component Analysis
(PCA)
PCA is a multivariate technique to analyze a dataset where the

observations are described quantitatively by a set of inter-correlated
variables. The goals of PCA are to (1) extract the most important
information from the data; (2) remove noise and compress the
data set by keeping only the important information; (3) simplify
the description of the data set; and (4) analyze the structure of the
observations and the variables. This method generates a set of new
orthogonal variables called principal components (PCs). Each PC
is a linear combination of the original variables. Hence, PCA can
be considered as a mapping of the data points from the original
variable space to the PC space. PCs are rank ordered in such a way
that PC1 represents the maximum variance among all possible
choices for the first axis. Similarly, PC2 represents the second highest variance contribution, and so forth through all the modes.
Usually only a few PCs are sufficient to understand the internal
structure of the data [26].
For extracting functional dynamics from the PDB experimental
structures, PCA is performed on the structure datasets. The input
is the set of coordinates of all of the structures in the set [7, 8].
From these data, the average position of each point in the structure
217
is computed as ri and the covariances for pairs of points i and j

are computed according to

cij = (ri - ri ) rj - rj
(1)

where brackets indicate averages over the entire set of structures.
The covariance matrix C can be decomposed as
C = P DP T ,
(2)
where the eigenvectors P represent the principal components (PCs)

and the eigenvalues are the elements of the diagonal matrix .
The eigenvalues are sorted in order. Each eigenvalue is directly
proportional to the amount of the variance it captures.
2.2 Elastic Network
Model (ENM)
Anisotropic Network Model (ANM) is an elastic network model

used to compute the directions of the normal modes from a single
structure [15]. In ANM, the potential energy V is a function of the
displacement vector D of each point in the structure
V =
g
DHD T ,
2

(3)
where is the spring constant for all closely interacting points in a

structure, and H is the Hessian matrix containing the second derivatives of the energy, with respect to each of the coordinates r=x, y, z.
For a structure with n residues, the Hessian matrix H contains nn
super-elements of size 33. The Hessian matrix H can be decomposed [7, 8, 15] as

H = M LM T ,
(4)
where is a diagonal matrix comprising the eigenvalues with the

eigenvectors forming the columns of the matrix M. This decomposition generates 3n6 normal modes (the first six modes account
for the rigid body translations and rotations of the system and must
be factored out, meaning that we actually perform singular value
decomposition to extract the normal modes) reflecting the vibrational fluctuations. We like to further mention that for ANM coarse
graining, it is shown that a cutoff distance of any value from 10 to
13 is appropriate for placing the springs and such an ANM model
represents the realistic protein dynamics. In this chapter, we use a
cutoff distance of 13.
2.3 Structure-Based
New ANM
The internal distance changes in a set of structures can provide

information that can be used directly to derive new structure-based
elastic network models. We have extracted spring constants
between all residue pairs in a set of structures by simply relating
these to the inverse of the variance of internal distance changes
between pairs of residues, as the spring stiffness (normalized
218
between 0 and 1). We have applied a cutoff of 13 to limit the

range of interactions. However the difference between the conventional ANM described in the previous section and this modified
ANM is that here the values for the spring constants are obtained
directly from the structure set rather than using a uniform value or
distance dependent values, as is customary with ENM.
2.4 Comparing
Directions ofMotions
Using Overlaps
The alignment between the directions of motion, for example

between a given PC and a given normal mode, is measured by their
overlap, which was defined as the dot product of the two vector
directions by Tama and Sanejouand [27]
Oij =
Pi M j
Pi M j
(5)

where Pi is the ith PC for model P and Mj is the jth PC or normal
mode for model M.A perfect match yields an overlap value of 1.
They also defined the cumulative overlap (CO) between the first k
vectors of M and Pi as
1
k
2
CO ( k ) = Oij2
j =1
(6)
which measures how well the first k PCs for model M together can
capture the motion of a single PC for model P.
2.5 Coarse-Grained
Global Entropies
Calculated from
Principal Component
Analysis
As covariance matrix can be decomposed as in Eq.2 of

Subheading 2.1, an approximation of the entropy from the PCs
can be obtained as well [10, 28]:
N
DS = Const li ( PCi PCiT )
(7)

where PCi is the ith PC, and i is the ith eigenvalue, N is the total
number of eigenvalues.
Andricioaei etal. also reported a similar result for entropy calculation from the covariance matrices of the atomic fluctuations as
shown in equation 7 of their paper [29]. It should be noted that
this expression is different from that for normal modes of the elastic network models, which because of the averaging normally
involved the inverse of the eigenvalues.
i =1
3 Materials
There are a huge number of available HIV-1 protease structures in
the PDB (564 X-ray and three NMR structures as of 07/26/2013),
which provides a remarkably rich set of different conformational
219
Fig. 2 Description of HIV-1 protease homo-dimer and its critical structural components that facilitate the functional dynamics (a) HIV-1 protease has two symmetric subunitssubunit A (red) and subunit B (blue). (b) Each
subunit has several structural components that are important for its coordinated motions. Fulcrum (orange,
residues 921) is a comparatively less mobile region that swings up and down similar to the flap elbow. E-34
(blue)Hinge residue which is responsible for transmitting the motion from the fulcrum to the flap region. Flap
elbow (magenta, residues 3742)Hinge residue E-34 drives the motion of this region to transfer the dynamics further away from the fulcrum to the upper flap region. This loop can make top-down and bottom-up
swings. When the flap elbow swings from top to bottom, the flap domain opens up, and when it swings upward
the flap domain closes. The Flap domain (residues 4358) consists of flap tip (yellow, residues 4952) and
-hairpin flaps (dark orange, residues 4348 and 5358). Opening and closing of the flap domains enable the
protein to bind ligands and release its products after proteolysis. Cantilever (green, residues 5975) functions
as a base for the flap domain. The C-terminal -hairpin flap is held by the N-terminal end of the cantilever and
this arrangement is important to control the swinging of the flap [30, 31]
states, which can be viewed as direct structural information on the

proteins dynamics.
The approach described here computes the essential or most
important protein motions from multiple structures of the same
protein, in contrast to using just the two structures such as the
open and closed conformations, which have often been used to
define the endpoints of conformational transitions. To demonstrate
this approach, we use HIV-1 protease as an example. Its abundant
experimentally determined structures are complemented by the
relatively small size of the protein. In the next section, first, we will
give a description of the structural components that are important to
drive the motion of the HIV-1 protease structure. Then, we will
describe the dataset of HIV-1 structures that we have used to perform our computations.
3.1 HIV-1 Protease
Architecture
HIV-1 protease functions as a homo-dimer as shown in Fig.2a.

The dimer has a single active site and 99 residues per monomer.
Each monomer has three domains: a terminal domain (residues
14 and 9599 of each chain), which is important for the dimerization and stabilization; a core domain (residues 1032 and 6385
220
of each chain), for dimer stabilization and catalytic site stability;

and a flap domain that includes two solvent accessible loops
(residues 3343 of each chain) followed by two flexible flaps
(residues 4462 of each chain) important for ligand binding interactions. The conserved Asp25-Thr26-Gly27 active site triad is
located at the interface between parts of the core domains. The
active site of HIV-1 protease is formed at the homo-dimer interface. Each monomeric unit has important structural components
as identified in Fig.2b that are important for its functional dynamics. The principal advantage of this structural arrangement is that
the hinge residue E 34 causes the up-down swinging motion of the
flap elbow (residues 3742), which transmits the motion generated
in the fulcrum (residues 921) to drive the dynamics of the flap
domain (residues 4258), whose conformation switches between
open and closed states to facilitate substrate trapping in the catalytic
pocket and product release following hydrolysis [30, 31].
3.2 HIV-1 Protease
Structure Set
(X-Ray-329)
1A30
We have used 329 PDB structures of HIV-1 protease for the computations to extract protein dynamics from experimental structures.
The PDB Ids of the data set are here (see Notes 1 and 2):
1A8G
1A8K
1A94
1AXA
1B6J
1B6K
1B6L
1B6M 1B6P
1BDL
1BDQ 1BDR 1BV7
1BWA 1BWB
1C6X
1C6Y
1C6Z
1C70
1D4S
1D4Y
1DAZ 1DIF
1DMP 1DW6 1EBK
1EBW 1EBY
1EBZ
1EC0
1EC1
1EC2
1EC3
1F7A
1FEJ
1FFF
1FFI
1FG8
1FGC
1FQX
1G2K
1G35
1GNM
1GNN 1GNO 1HBV 1HIH
1HIV
1HOS 1HPO
1HPS
1HPV
1HPX
1HSG
1HSH
1HTE
1HTF 1HTG 1HVH 1HVI
1HVJ
1HVK 1HVL
1HVR 1HVS
1HWR 1HXW 1IIQ
1IZH
1IZI
1K6C
1K6P
1K6V
1KJ7
1K1U
1LZQ 1M0B
1K2B
1FF0
1K2C
1BV9
1FG6
1K6T
1AJV
1AJX
1KJ4
1KJF
1KJG
1KJH
1MET 1MEU 1MRW
1MRX 1MSM 1MSN
1MT7
1MT8
1MT9
1MTB 1MTR 1MUI 1N49
1NH0 1NPA
1NPV
1NPW 1ODW 1ODX 1PRO
1QBR
1QBS
1QBT 1QBU 1RL8
1RPI
1RQ9 1RV7
1SDT
1SDU 1SDV
1SGU
1SH9
1SP5
1T3R
1T7I
1T7J
1T7K
1TCX
1TW7 1U8G
1VIJ
1VIK
1XL2
1XL5
1YT9
1YTG
1YTH
1Z8C
1ZBG
1ZLF
1ZPK
1ZSF
1ZSR
2A1E
2A4F
2AID
2AOF
2AQU 2AVM
2AVO
2AVS
2AVV
2AZC
2B7Z
2BB9
2BBB
2BPV
2BPW 2BPX
2BPY
2BPZ
2BQV
2CEJ
2CEM 2CEN 2F3K
2F80
2F81
2F8G
2FDD
2FDE
2FGU
2FGV
2FNS
2FNT
2FXD
2FXE
2HB3
2HC0
2HS1
2HS2
2I4D
2I4U
2I4V
2I4W
2I4X
2IDW
2IEN
2IEO
2J9J
2J9K
2JE4
2NMZ 2NNK 2NNP 2O4K
2O4L
2O4P
2O4S
2P3A
2P3B
2P3C
2P3D
2PK5
2PK6
2PQZ
2PYN
2Q3K
2Q63
2Q64
2QAK
2QCI
2QD6 2QD7 2QD8 2QHC 2QHY 2QHZ 2QI0
2QI1
2QI3
2QI4
2QI5
2QI6
2QI7
2QMP 2QNN 2QNP 2QNQ 2R38

2UY0
2Z4O
1MER 1MES
1A9M 1AAQ 1AID
3A2O
3AID
2PWC 2PWR 2PYM
2R3T
2R3W
2R43
2R5P
2R5Q
2RKF
2UPJ
2UXZ
3BGB 3BGC
3BVA
3BVB
3CKT
3CYW
3CYX
3D1X
3D3T
(continued)
221
(continued)
3FX5
3GGA 3GGV 3GGX 3GI4
3GI5
3GI6
3I7E
3KF0
3KFN
3KFR
3KFS
3LZS
3LZU 3MWS 3NDU 3NDX 3NU3 3NU4 3NU5
3NU6 3NU9
3NUJ
3NUO 3O9F
3O9G
3O9H 3O9I
3OK9
3OTS
3OXC 3PWM 3PWR
3QAA 3R4B
3S43
3S53
3S56
3S85
3SO9
3T11
3U7S
3UCB 3UF3
3UHL
4DQB 4DQC
4DQE 4DQG 4DQH 4EJ8
4EJK
4EJL
4FAE
4FL8
4FLG
4HVP
4I8W
4J54
7HVP 7UPJ
8HVP
9HVP
4FM6
4I8Z
4J55
3S54
4J5J
4PHV
The following method section gives a step by step description

of how to retrieve these PDB files from the protein databank and
then how to extract the dynamics from these structures.
4 Methods
To successfully complete the procedures described in this section,
one needs the following software/programs:
Perl 5Several perl scripts are included here. Perl programming language [32] can be downloaded free at www.perl.org.
PythonA python script is used to calculate the internal distances between residue pairs for the set of 329 protein structures. A Python environment can be downloaded at http://
www.python.org/.
MatlabSeveral Matlab scripts are included here that can be executed in a Matlab programming environment [33]. Matlab
product site is http://www.mathworks.com/products/matlab/.
MAVENsThis software was developed in the Jernigan lab [12].
In our Matlab code, we have invoked several MAVEN functions:
ANM.mThis is a function from MAVEN [12] used in
experimentalDynamics.m to compute ENM normal modes
from a given PDB structure.
modeAnimator.mThis is a function from MAVEN used
in experimentalDynamics.m to visualize the ENM modes
and PCs by creating movies.
readPDB.m, writePDB.mThese two Matlab functions from
MAVEN are used to read and write PDB files, respectively.
CompareVectors.mThis function from MAVEN is used
in experimentalDynamics.m to compare the directions of
PCs and ENM modes.
plot_compareVectors.mThis function from MAVEN
plots the results obtained from the above CompareVectors.m.
mat2vec.mThis function converts a matrix to a vector.
222
MAVEN is available for download at http://maven.sourceforge.net.
MUSTANGMultiple structural alignment will be done using

this program [34]. This program can be installed only on a Linux
operating system. MUSTANG can be downloaded at http://
www.csse.monash.edu.au/~karun/Site/mustang.html.
PyMOLThis software has a free version for academic use
[35]. This can be used to visualize the structures and their
dynamics. PyMOL can be downloaded at http://pymol.org/.
The following Table1 summarizes the steps that are discussed

in this sectionstarting from processing raw PDB structures to
computing PCs and ANM modes, and comparing their dynamics.
4.1 Extracting
Cartesian Coordinates
fromRaw PDBFiles
In this section, we will describe how to prepare the dataset

X-ray-329 for PCA.The 329 PDB Ids are listed in the pdbIds.
txt file. Download these files from the protein data bank (http://
www.pdb.org/pdb/download/download.do)
(Download
options: download TypePDB File Format, Compression
Typeuncompressed) and save them in a sub folder named
data-raw under the parent folder experimentalDynamics. The
downloaded PDB files have a lot of extra information that we
will not be used.
The records of ATOM type for residue 8 and modified residue
67 of PDB file 2p3a are shown in Schema 1. The important fields
are labeled. Each residue of a protein is recorded in this way. When
a residue in the protein is modified with a non-amino acid type
molecule, HETATM keyword is used to identify that record. The
TER key word is used as an end of chain marker. The PDB file has
other detailed information and have different record identifiers.
We will retain the ATOM type records for the C atoms of each
residue or modified residue for our calculation. When more than
one alternate location is recorded, we arbitrarily retain the first
alternate location for that ATOM.
The following three subsections describe how to copy the
Cartesian coordinates from each PDB file and align these structures.
4.1.1 Preparing aData

Set forMUSTANG
fromRaw PDBFiles
Download and save the following perl scripts in the same folder
experimentalDynamics. Run these perl scripts in the same sequence
as they are listed below:
copyBackboneAtoms.plThis program copies the backbone

ATOM and HETATM from a set of PDB files.
perl copyBackboneAtoms.pl
Output files after running this program will be saved in
data-backbone subfolder of the experimentalDynamics
parent folder.
223
Table 1
Summary of the steps for extracting biomolecular dynamics
Program/file name
Function
Subheading4.1 Extracting Cartesian coordinates from raw PDB files

Subheading4.1.1 Data set preparation for MUSTANG from raw PDB files
copyBackboneAtoms.pl
Copies backbone ATOM and HETATM from a set of PDB files.
retainFirstAltLocation.pl
Retains the first alternate location for each ATOM and HETATM when
multiple locations for that ATOM/HETATM exist. It operates
on a set of PDB files.
replaceHETATM.pl
Replaces the keyword HETATM with the keyword ATOM in a set

of PDB files.
retainCA.pl
Copies the CA atoms from a set of PDB files with no TER keyword
between chains to comply with the MUSTANG input file format.
Subheading4.1.2 Multiple structural alignment using MUSTANG

Subheading4.1.3 Data set preparation for PCA from MUSTANG output
copyChainsToPDBs.pl
Copies the chains from alignAll.pdb to individual PDB files.
pdbIds.txt
This file list the PDB ids for 329 PDB structures used here.
Subheading4.2 Principal Component Analysis (PCA)

Subheading4.2.2 Comparing and visualizing PCs and ANM modes
Subheading4.3 Comparing PCs and structure-based ANM
Subheading4.4 Computing Entropy using PCs
experimentalDynamics.m
This Matlab program (1) computes principal components from aligned

structures, (2) computes ENM modes, (3) computes the overlap
between PCs and ENM modes, (4) computes entropies from PCs.
readAlignedPDBcoordinates.m
This Matlab function reads the coordinates of aligned PDB

structures and returns the coordinates of those structures.
internal.py
This program, written in Python, calculates the internal distances of

Mustang aligned structures.
calc_Entropy_PC.m
This Matlab function computes entropy from computed PCs.
The above files, the files used from MAVEN, other accessory files and dataset can be downloaded at http://ribosome.
bb.iastate.edu/4papers/2013/ataur/experimentalDynamics/
Running this program will retain the backbone atoms for

each ATOM and HETATM record. A sample output for
residue 8 and modified residue 67 is shown in Schema 2.
retainFirstAltLocation.plThis program retains the first alternate

location for each ATOM when multiple alternative locations
for that ATOM exist. It operates on a set of PDB files.
perl retainFirstAtlLocation.pl
data-backbone-singleAltLocation subfolder of the experimentalDynamics parent folder.
224
Schema 1 The records of ATOM type for residue 8 and modified residue 67 of the PDB file 2p3a
After running this program, a PDB file will have residues

with only the backbone atoms and only the first alternate
location will be retained in case of multiple alternate locations. Output for residue 8 and modified residue 67 is
shown in Schema 3.
ATOM
ATOM
ATOM
ATOM
ATOM
ATOM
72
73
74
75
76
77
N
N
CA
CA
C
C
AARG
BARG
AARG
BARG
AARG
BARG
Res No
Alt Loc Ind
Atom No
Record id
A
A
A
A
A
A
225
Cart. Coordinates
8
8
8
8
8
8
26.288
26.517
25.875
26.053
26.929
27.135
-8.483
-8.547
-8.023
-8.180
-8.501
-8.490
-5.941
-6.064
-4.614
-4.733
-3.624
-3.723
0.60
0.40
0.60
0.40
0.60
0.40
23.05
23.23
21.71
22.23
21.10
21.09
N
N
C
C
C
C
A. Records for Residue 8

HETATM
HETATM
HETATM
HETATM
HETATM
572
573
574
575
588
N
N
CA
CA
C
ACME
BCME
ACME
BCME
ACME
A
A
A
A
A
67
67
67
67
67
31.550
31.558
32.726
32.776
33.921
-12.012
-11.938
-12.421
-12.485
-12.522
8.379
8.292
7.660
7.653
8.646
0.70
0.30
0.70
0.30
0.70
29.54
29.11
31.90
30.89
32.23
N
N
C
C
C
HETATM
589
BCME A
67
33.963
-12.613
8.623
0.30
31.53
B. Records for Residue 67
ATOM
72
74
ATOM
76
ATOM
N AARG A
CA AARG A
C AARG A
Res No
Alt Loc Ind
Atom No
Record id
Schema 2 A sample output of copyBackboneAtoms.pl for residues 8 and 67
Cart. Coordinates
26.288
-8.483
-5.941
8
8
25.875
-8.023
-4.614
26.929
-8.501
-3.624
0.60
0.60
0.60
23.05
21.71
21.10

HETATM
572
ACME A
67
31.550
-12.012
8.379
0.70
29.54
HETATM
574
CA ACME A
67
32.726
-12.421
7.660
0.70
31.90
HETATM
588
67
33.921
-12.522
8.646
0.70
32.23
ACME A

Schema 3 The output of retainFirstAtlLocation.pl for residues 8 and 67
Res No
Alt Loc Ind
Atom No
Record id
226
Cart. Coordinates
ATOM
72
AARG A
26.288
-8.483
-5.941
0.60
23.05
ATOM
ATOM
74
76
CA AARG A
C AARG A
8
8
25.875
26.929
-8.023
-8.501
-4.614
-3.624
0.60
0.60
21.71
21.10
C
C

ATOM
ATOM
ATOM
572
574
588
N ACME A
CA ACME A
C ACME A
67
67
67
31.550
32.726
33.921
-12.012
-12.421
-12.522
8.379
7.660
8.646
0.70
0.70
0.70
29.54
31.90
32.23
N
C
C
Schema 4 The output of replaceHETATM.pl for residues 8 and 67
replaceHETATM.plThis program replaces the keyword

HETATM with the keyword ATOM in a set of PDB files.
perl replaceHETATM.pl
data-backbone-singleAltLocation-NoHETATM subfolder
of the experimentalDynamics parent folder.
MUSTANG [34] removes all records with the keyword
HETATM before multiple structural alignment. To prevent
the removal of needed data, replaceHETATM.pl program
replaces the keyword HETATM with ATOM so that
MUSTANG will use the HETATM coordinates as required
in the multiple structure alignment. Sample output for residue 8 and modified residue 67 is shown in Schema 4.
retainCA.pl This program copies the CA atoms from a set of

PDB files.
perl retainCA.pl
data-CA subfolder of the experimentalDynamics parent
folder.
This program retains the records of the C atoms only.
Therefore, for each residue only the record for C will be
copied. Also, the TER keyword to separate the chains will
not be retained in the output file so that MUSTANG considers the whole structure (multiple chains) as one chain.
A sample output for residue 8 and modified residue 67 is
shown in Schema 5.
74
Res No
Atom No
ATOM
Alt Loc Ind
Record id
CA AARG A
227
Cart. Coordinates
25.875
-8.023
-4.614
0.60
21.71

ATOM
574
CA ACME A
67
32.726
-12.421
7.660
0.70
31.90

Schema 5 The output of retainCA.pl for residues 8 and 67
4.1.2 Aligning PDB

Structures Using MUSTANG
There are several successful multiple structural alignment programs

such as MUSTANG [34] , TM-align [36], DaliLite [37], etc.
A Wikipedia page has a list of multiple structure alignment software/programs (http://en.wikipedia.org/wiki/Structural_alignment_software, 10/15/2013). We have used MUSTANG for
multiple structural alignments of the selected PDB structures.
MUSTANG does not consider sequence information in its alignment algorithm. Rather, it performs a structural alignment by finding maximal similar substructures. Thus it can capture the
conformational variations among the structures much better than
the alignment algorithms that rely upon sequence similarity information. Moreover, in its alignment MUSTANG uses the C backbone atoms only. The running time MUSTANG 3.2.1 for the
alignment of the selected dataset of 329 structures is approximately
5.00h on a Linux machine with the following configuration
Linux version 2.6.18-348.4.1.el5 Intel(R) Xeon(R) CPU E5630
@2.53GHz. MUSTANG needs a Linux operating system. After
installing MUSTANG under the experimentalDynamics folder on
a Linux machine, save and copy the following file description in the
same folder; and copy data-CA subfolder with the structures in the
same folder as well:
data-CA: This folder has all the backbone PDB files for multiple structural alignments.
Description: This file has the path of the source directory
where MUSTANG will find the input files for multiple structural alignment. After the path information, this file also has
the list of the PDB file names that MUSTANG will read from
the source directory. The list of the filenames in this file is in
228
the same order as the list of the PDB Ids in the pdbIds.txt file
which has the 329 PDB Ids that are listed in Subheading3.2.
Update the line in description file that records the path of the
source directory for the input files (path to the files in data-CA
subfolder) that would be aligned.
Run the following command to execute MUSTANG:
mustang-3.2.1 -f description -o alignAll -F fasta -r ON
This will create the following two files:
4.1.3 Preparing Data

forPCA fromMUSTANG
Output Files
alignAll.pdb: This file contains the aligned structures. Each

chain corresponds to a specific PDB file and the header of this
file lists the file names in the same order (see Note 3).
alignAll.afasta: This contains the alignment of the amino acid
sequences of the HIV-1 proteases based on the structural
alignment.
copyChainsToPDBs.plThis perl script will copy each chain
of alignAll.pdb file to the corresponding PDB file according to
PDB Ids listed in the pdbIds.txt file.
perl copyChainsToPDBs.pl
pdbIds.txt alignAll.pdb
This will create a subfolder alignedPDBs in the experimentalDynamics folder. This subfolder will have the 329 PDB files with
the aligned C atoms of each structure. So when the Cartesian
coordinates of each file will be placed in a matrix such that each
row corresponds to the coordinates of one PDB Id, this matrix can
be used for principal component analysis (see Note 4).
4.2 Use ofCartesian
PCs toExtract
Functional Dynamics
fromtheProtein
Structures
Matlab script experimentalDynamics.m reads the Cartesian coordinates of the structures from the MUSTANG aligned files and
perform PCA on them.
4.2.1 Significance
ofPrincipal Components
(PCs)
Figure 3 shows the distribution of the 329 PDBs Ids projected

onto the space of the first few PCs from three separate views
PC1PC2 (panel a), PC1PC3 (panel b), and PC2PC3 (panel c).
In panels a and b, open and closed structures are clearly separated
in two regions (open structures on the left side and closed structures on the right side) and the intermediate conformations (1aid,
3t11, 4ej8, etc.) spanning the middle region. The PC2PC3 view
in panel c, the structures are distributed based on conformational
differences in the flap elbow region.
We used the MAVEN function modeAnimator.m to animate
the motions of the structure along PC1, PC2, and PC3 vectors.
The following code calculates the conformations along PC1 and
can be found in experimentalDynamics.m matlab function:
Fig. 3 Distributions of the 329 PDB structures by PCA. (a) Distribution of the structures on a PC1-PC2 plot.
(b) Distribution of the structures on a PC1PC3 plot. (c) Distribution of the structures on a PC2PC3 plot. In plots
a and b, open structures are located on the left side; closed structures are located on the right side; and the
intermediate structures fall in between. Distribution of structures on PC2PC3 plot (panel c) is based on primarily
on the conformational differences along the flap elbow region. PC1, PC2, and PC3 capture 30%, 20%, and 7%
of the variances in the dataset, respectively
230
m = readPDB(ifname,1); %read the MUSTANG reference structures

c = sqrt(length(m.IND)/ sum(PC(:,1).^2));
%c controls the vector displacement amount
modeAnimator(m,PC(:,1),'',c,c/10,
ofname,'',0,'',1);
%use PC1 as the mode vector to simulate the
motion of the structure
The motion of the structure along PC1, PC2, and PC3 can be
observed by opening the corresponding file using PyMOL visualization software. It is evident that, PC1 is closely related to the
opening and closing (or expansion/contraction) of the flaps and
the ligand binding cavity as shown in Fig.4a. The two extreme
ends of PC1 motion correspond closely to the closed (+) and the
open () experimental structures (closed: PDB 1ebw, open: PDB
1rpi). The PC2 and PC3 correspond to twisting motions that are
best seen in a perpendicular direction to those of PC1. PC2 is predominantly a twisting motion of the flap domains (panel C),
whereas PC3 is predominantly a hinge motion of the core domains
moving towards and away from the flaps (panel D).
experimentalDynamics.m also has code to visualize structures
by using the ANM modes and the generated frames are saved in
PDB file format that can be visualized using PyMOL software.
4.2.2 Comparing PC
Based andANM Computed
Dynamics
Matlab program experimentalDynamics.m has the code to compute the ANM modes by using the MAVEN function ANM.m,
and it then computes the overlap and the cumulative overlaps with
the previously computed PCs by using another MAVEN function
CompareDynamics.m. Figure5, generated by MAVEN function
plot_compareDynamics.m, shows the overlaps between the first
ten PCs and the first ten ANM modes. The highest overlap is 60%
found between PC1 and ANM mode 3.
Table2 shows the cumulative overlaps between PCs and the
ANM modes. The cumulative overlap between each of the first and
the second PCs and the first 20 modes is above 80%. Interestingly,
the cumulative overlap reaches 80% between the second PC and
the first six modes. This clearly indicates that given an appropriate
experimental dataset the motions captured by the PCs conform
quite closely with the ANM motions.
4.3 New Internal

Distance Based ANM
Motions
The use of structural information in ANM improves the modeling

of the protein dynamics. Subheading2.3 describes a way to derive
spring constants from the structures. Here, we compute the inverse
of the variance of the internal distances from the aligned structures
in the MUSTANG aligned file align.pdb by using the Python program internal.py. The calculated inverse values are stored in
hiv.329.var.sc file that could be downloaded at the link in the footnote of Table1. MAVEN function ANM.m can be modified to use
231
Fig. 4 Visualization of the first three PCs of HIV-1 protease on the structures. (a)
Structures showing the closed form (left, PDB 1ebw) and open form (right, PDB
1rpi) of HIV-1 protease. The two subunits are shown in red and blue color and in
ribbon diagram. (b) Snapshots of the structures displaced along the directions of
PC1 shown in connected line segment. The direction of motions of the protein
along each PC is shown with a black arrow. It can be seen that the openingclosing motion of the flaps can be easily identified from the extrema of PC1. Two
extrema are shown for each motion in each row, together with arrows that indicate the directions for transition to the other structure. (c) PC2 images are shown
looking down from the top of those in PC1 and PC3. PC2 is a twisting of the flap
regions whereas (d) PC3 is a hinge motion between the core and flaps, with the
core and flaps moving to and fro relative to one another
these values as the spring constants to compute the normal modes.

Table3 shows the overlaps of PCs based on these internal distances
and the new ANM modes. The highest overlap is 79% that occurs
between PC1 and mode 2, which is much higher than the highest
overlap (60%) that occurred between PCs and conventional ANM
modes (Fig.5).
Table 4 shows the cumulative overlap between PCs and the
new ANM modes. We can see that cumulative overlap between
PC1 and the first three modes reaches 90% which is quite high
compared to the cumulative overlap between PC1 and the first
three modes (62% as shown in Table2). However, the cumulative
232
Fig. 5 Overlap between PCs and ANM modes. PC1 and mode 3 gives the highest overlap 60%
Table 2
Cumulative overlap between the first three PCs and sets of the ANM modes
ANM modes/PCs
PC1
PC2
PC3
3 modes
0.62
0.71
0.44
6 modes
0.64
0.80
0.54
10 modes
0.77
0.83
0.59
20 modes
0.80
0.85
0.65
CO between a PC and ANM modes is shown in bold type if it is greater than 0.80
Table 3
Overlaps between PCs and the new ANM modes
PCs/newANM modes
Mode 1
Mode 2
Mode 3
PC1
0.09
0.79
0.40
PC2
0.34
0.01
0.24
PC3
0.34
0.01
0.10
Table 4
Cumulative overlaps between PCs and the new ANM modes
New ANM modes/PCs
PC1
PC2
PC3
3 modes
0.90
0.42
0.35
6 modes
0.91
0.44
0.41
20 modes
0.95
0.89
0.84
Values in bold indicate cumulative overlaps above 80%
233
Fig. 6 Depiction of entropies of HIV-1 protease structure (PDB 1rpi) computed

from PCs. Residues are colored spectrally according to the entropy values
coloring from red for the highest entropy to blue for the lowest entropy. Some of
the residues along the flap and flap elbow regions on the subunit A (right subunit)
have higher entropies than the same residues on subunit B (left subunit).
overlap between PC2 and the first three modified ANM 42%; on
the other hand this value between PC2 and the first three conventional ANM modes is 71%, a much higher value. Therefore, in
some cases cumulative overlap between a PC and the new ANM
modes gets improved compared to the similar values between a PC
and the conventional ANM modes. But when 20 new ANM modes
are included, the values are constantly higher.
Taken together, this suggests that modified ANM can improve
the performance of the ANM models.
4.4 Computing
Entropy UsingPCs
We compute the entropy of the HIV-1 protease system using Eq.7

described in Subheading2.5. By using calc_Entropy_PC.m matlab
program, we compute the entropy from the principal components
of the 329 aligned HIV-1 protease structures. The residues of
HIV-1 protease are colored in Fig.6 according to the entropy values. It is clear from the figure that the entropies are asymmetrically
distributed in the two HIV-1 protease subunits. Subunit A (right
subunit) has higher entropies along the flap and flab elbow regions.
5 Conclusion
This chapter gives the background of two important methods
PCA and ENM.By following the steps with the set of 329 HIV-1
PDB structures, one can get a hands-on experience on how to
234
apply PCA to extract dynamics and mechanism information by

capturing the conformational variability buried in different PDB
structures of the same protein. One can also learn how to model
the functionally important global motions of the protein using the
widely accepted ANM model and compare the dynamics and
mechanism found from experimental structures by PCA and from
the ANM model. The higher overlaps between PCs and modified
ENM modes indicate that a rich dataset of protein structures can
play an important role in understanding functional dynamics and
mechanism of the protein.
Moreover, the PCs represent the variability apparent within
the sets of structures, and hence these are used as a direct measure
of the conformational entropy of the protein structure.
This approach can also be extended to other highly diverse
protein structure sets. The PDB database continues to grow rapidlyin 2008 there were ~43,000 protein structures and now in
2013 there are more than 90,000 structures [1]. In the future if
new technologies for X-ray structure determination are developed
that are much more efficient and very rapid, then there will be truly
abundant structures of related proteins, including aberrant protein
structures from patients. Among the various structures there are
many single proteins with multiple X-ray structures determined
under different conditions, as well as NMR structures. Generally
proteins are robust and not easily disturbed by different environments or mutations; and the preponderance of evidence suggests that
proteins have a limited range of conformations that are essential for
their function. Therefore, the approach described here can generally be used to extract dynamics of any protein with significant
numbers of available experimental structures.
6 Notes
1. Selecting a set of structures: There are 564 HIV-1 X-ray structures in PDB (07/26/2013). Among them, 329 PDB structures are selected so that the MUSTANG structural alignment
does not produce any gaps in the corresponding aligned
sequences. If a different set of structures is selected that produces gaps after multiple structural alignment, the residues in
a structure that fall along the gaps on the alignment need to be
removed before the PCA calculation.
2. Construction of the selected dataset: It is important to select a
dataset that represents the whole conformational landscape of
a protein structure. In panels A and B of Fig.3, the open and
closed structures are clustered on the left and the right side,
respectively, and the intermediate conformations (1aid, 3t11,
4ej8, etc) span the middle region. Though the number of
235
closed structures is much higher than the number of open and

intermediate structures, this dataset is a good selection as it has
representation from whole conformational landscape.
3. Caution in the use of MUSTANG: MUSTANG output file
align.pdb is found to break lines in some structures. Therefore,
once the align.pdb is generated from MUSTANG, it needs to
be normally scanned to detect and fix such broken lines.
4. PCA on all the backbone ATOMs: data-backbone-singleAltLocation-
NoHETATM subfolder in the experimentalDynamics folder has
the structures with all backbone atoms. These structures can, as
well, be used for MUSTANG alignment and then subsequent
PCA and other related operations.
PCA can also be done on all atoms of each structure. In that
case, first, the structures need to process to keep the same atoms
for each residue in all structures and then use MUSTANG to
align the structures. Afterwards, the Cartesian coordinates of all
structures need to be extracted and perform PCA on them. For
this, noOfAtoms variable in experimentalDynamics.m need to
be initialized accordingly.
Acknowledgments
We gratefully acknowledge the support provided by NIH Grant
R01GM072014 and NSF Grant MCB-1021785.
We used several Matlab functions from MAVEN by Zimmerman
etal. [12] as mentioned in Subheading4. MAVEN is also useful to
compute PCs, PC-plot, ANM modes, overlap between PCs and
ANM modes.
References
1. Berman HM, Westbrook J, Feng Z, Gilliland G,
Bhat TN, Weissig H, Shindyalov IN, Bourne PE
(2000) The protein data bank. Nucleic Acids
Res 28(1):235242, PMCID:PMC102472
2. Hotelling H (1993) Analysis of a complex of
statistical variables into principal components.
JEduc Psychol 24:417441
3. Manly B (1986) Multivariate statisticsa
primer. Chapman & Hall, Boca Raton
4. Pearson K (1901) On lines and planes of closest fit to systems of points in space. Philos Mag
2(6):559572
5. Amadei A, Linssen AB, Berendsen HJ (1993)
Essential dynamics of proteins. Proteins
17:412425
6. Amadei A, Linssen AB, de Groot BL, van
Aalten DM, Berendsen HJ (1996) An efficient
method for sampling the essential subspace of

proteins. J Biomol Struct Dyn 13:615625
7. Teodoro ML, Philips GN Jr, Kavraki LE
(2002) A dimensionality reduction approach
to modeling protein flexibility. J Comput Biol
10:299308
8. Teodoro ML, Philips GN Jr, Kavraki LE
(2003) Understanding protein flexibility
through dimensionality reduction. J Comput
Biol 10:617634
9. Howe PW (2001) Principal components analysis of protein structure ensembles calculated
using NMR data. J Biomol NMR 20:6170
10. Yang L, Song G, Carriquiry A, Jernigan RL
(2008) Close correspondence between the
motions from principal component analysis of
multiple HIV-1 protease structures and elastic
236
network modes. Structure 16:321330,

PMCID:PMC2350220
11. Yang LW, Eyal E, Bahar I, Kitao A (2009)
Principal component analysis of native ensembles
of biomolecular structures (PCA_NEST):
insights into functional dynamics. Bioinformatics
25:606614, PMCID:PMC2647834

12. Zimmermann MT, Kloczkowski A, Jernigan
RL (2011) MAVENs: motion analysis and
visualization of elastic networks and structural
ensembles. BMC Bioinformatics 12:264,
PMCID:PMC3213244
13. Bakan A, Meireles LM, Bahar I (2011) ProDy:
protein dynamics inferred from theory and
experiments. Bioinformatics 27:15751577,
PMCID:PMC3102222

14.
Grant BJ, Rodrigues AP, ElSawy KM,
McCammon JA, Caves LS (2006) Bio3d: an R
package for the comparative analysis of protein
structures. Bioinformatics 22:26952696
15. Atilgan AR, Durell SR, Jernigan RL, Demirel
MC, Keskin O, Bahar I (2001) Anisotropy of
fluctuation dynamics of proteins with an elastic
network model. Biophys J 80:505515,
PMCID:PMC1301252
16. Bahar I, Jernigan RL (1994) Cooperative structural transitions induced by non-homogeneous
intramolecular interactions in compact globular
proteins. Biophys J 66:467481
17. Bahar I, Atilgan AR, Erman B (1997) Direct
evaluation of thermal fluctuations in proteins
using a single-parameter harmonic potential.
Fold Des 2:173181
18. Bahar I, Erman B, Haliloglu T, Jernigan RL
(1997) Efficient characterization of collective
motions and interresidue correlations in proteins
by
low-resolution
simulations.
Biochemistry 36:1351213523

19. Bahar I, Jernigan RL (1997) Inter-residue
potentials in globular proteins and the dominance of highly specific hydrophilic interactions
at close separation. J Mol Biol 266:195214
20. Bahar I, Rader AJ (2005) Coarse-grained normal mode analysis in structural biology. Curr
Opin
Struct
Biol
15:586592,
PMCID:PMC1482533
21. Chennubhotla C, Rader AJ, Yang LW, Bahar I
(2005) Elastic network models for understanding biomolecular machinery: from
enzymes to supramolecular assemblies. Phys
Biol 2:S173S180

22. Jernigan RL, Yang L, Song G, Doruker P
(2008) Elastic network models of coarse-
grained proteins are effective for studying the
structural control exerted over their dynamics.
In: Voth G (ed) Coarse-graining of condensed

phase and biomolecular systems. Taylor and
Francis, Boca Raton, pp237254
23. Bahar I (2010) On the functional significance
of soft modes predicted by coarse-grained
models for membrane proteins. J Gen Physiol
135:563573, PMCID:PMC2888054
24. Ichiye T, Karplus M (1987) Anisotropy and
anharmonicity of atomic fluctuations in proteins: analysis of a molecular dynamics simulation. Proteins 2:236259
25. Kuriyan J, Petsko GA, Levy RM, Karplus M
(1986) Effect of anisotropy and anharmonicity
on protein crystallographic refinement. An
evaluation by molecular dynamics. J Mol Biol
190:227254
26. Abdi H, Williams LJ (2010) Principal component analysis. WIREs Comput Stat 2:433459
27. Tama F, Sanejouand YH (2001) Conformational
change of proteins arising from normal mode
calculations. Protein Eng 14:16
28. Yang L, Song G, Jernigan RL (2007) How well
can we understand large-scale protein motions
using normal modes of elastic network models?
Biophys J 93:920929, PMCID:PMC1913142
29. Andricioaei I, Karplus M (2001) On the calculation of entropy from covariance matrices of
the atomic fluctuations. J Chem Phys
115:62896292
30. Harte WE Jr, Swaminathan S, Mansuri MM,
Martin JC, Rosenberg IE, Beveridge DL
(1990) Domain communication in the dynamical structure of human immunodeficiency
virus 1 protease. Proc Natl Acad Sci U S A
87:88648868, PMCID:PMC55060
31. Hornak V, Okur A, Rizzo RC, Simmerling C
(2006) HIV-1 protease flaps spontaneously
open and reclose in molecular dynamics simulations. Proc Natl Acad Sci U S A 103:915
920, PMCID:PMC1347991
32. Larry Wall (2011) Perl 5. Version 5.12.4

33.
Matlab Version 7.11.0.584 (2010) The
MathWorks Inc., Natick, Massachusetts
34. Konagurthu AS, Whisstock JC, Stuckey PJ, Lesk
AM (2006) MUSTANG: a multiple structural
alignment algorithm. Proteins 64:559574

35.
The PyMOL Molecular Graphics System
Version 1.4. (2012) Schrdinger, LLC
36. Zhang Y, Skolnick J (2005) TM-align: a protein structure alignment algorithm based on the
TM-score. Nucleic Acids Res 33:23022309,
PMCID:PMC1084323
37. Holm L, Sander C (1996) Mapping the protein
universe. Science 273:595603
Chapter 11
Computing Ensembles of Transitions with Molecular
Dynamics Simulations
Abstract
A molecular understanding of conformational change is important for connecting structure and function.
Without the ability to sample on the meaningful large-scale conformational changes, the ability to infer
biological function and to understand the effect of mutations and changes in environment is not possible.
Our Dynamic Importance Sampling method (DIMS), part of the CHARMM simulation package, is a
method that enables sampling over ensembles of transition intermediates. This chapter outlines the context
for the method and the usage within the program.
Key words Conformational transition, Sampling intermediates, Relative free energy, Statistical
mechanics of proteins, Structurefunction
Introduction
Starting from a seminal paper by McCammon et al. [1], the field of
protein molecular dynamics has evolved rapidly. This reflects the
interest of a broad section of the biophysics community in understanding how a detailed X-ray or NMR structure can be connected
to measured function [2]. The force-fields that are integrated on
the computer to enable the sampling of motions have also improved
dramatically along with the hardware that is enabling more and
more complex systems of growing size and timescales to be
explored on a detailed level with the use of computers. Probably
the most extensive tests of force-fields and simulation times have
come from the DE Shaw group and their recent sets of microsecond to millisecond simulations of BPTI and of a WW-domain [3].
In this case the computations were able to reach conformational
substates that were not obvious from the X-ray structures, but
that are fully consistent with the available experimental information. The exciting news from their results is that the force-fields
currently in use seem to be capable of generating insights on
much longer time-scales than had ever before been attempted.
237
238
The less exciting news is that reaching these time scales, even with
supercomputer access, is challenging for larger systems and for
complex conformational change. Thus there remains a need to
develop methods to enhance sampling of rare events that are poorly
sampled in a general molecular dynamics trajectory.
We suggest that understanding large conformational change
is the next real frontier for the molecular dynamics community.
The finding that fluctuations from the X-ray or NMR starting
points agree with experiment is now pretty well agreed by most
researchers. The DE Shaw results thus further confirm and extend
the validity of the molecular dynamics method. However, this validation of the method still leaves many questions for how to best
apply it to a given biophysical question. If we want to know how
a channel gates or how a kinase switches states, then it is not obvious that the best solution is the DE Shaw one of simply running
the calculations for a very, very long time. Phrased another way, the
number of interesting biophysical questions far outnumbers the
available computer cycles. So, we need to find ways to efficiently
use the computer resources that we do have access to and to most
confidently sample on the interesting biological changes. Phrased
another way, to reach these important biologically inspired questions demands a new level of computation that will not be readily
met by even the most advanced computer systems. In general the
initial excitement about the molecular dynamics method as a panacea then led to disappointments and is now in a slowly growing
re-enthusiasm as the computer speeds have reached a stage where
more elaborate calculations with greater statistical accuracy can be
performed. This is also reflected in the recent Nobel Prize awards
for the development of the approach.
1.1 Why Transitions
Are Important
A key to understanding protein function is a detailed molecular

description of the different conformations reached by each protein
[4]. Examples of these states are protein folding transitions, protein
activation by the exposure of an otherwise buried catalytic site, the
opening/closing of channels to allow the passage of ions though
the cellular membrane, auto-inhibition, and conformational
changes that induce dimerization of monomeric units in order to
activate signaling pathways as in human epidermal growth factor
receptor(see Fig. 1). The motions involved in these changes range
from hinge, shear, or rotation of subunits to the arrangement of
amino acid side chains [7].
The density of conformational states may be dependent on
physiological conditions and on the presence/absence of ligands
that stabilize a particular conformation. Thus this ability to sample
different conformations reflects the ever changing environment
that cells and proteins are exposed to. Understanding at the atomic
level the conformational changes and, the motions involved in
them, provides insights into the nontrivial knowledge of the
Computing Transition Ensembles
239
Fig. 1 Human epidermal growth factor receptor (HER3) exhibits a large conformational change that prevents exposure of the dimerization arm (blue and pink )
by interactions between domains I and IV (blue and green ) [5, 6]
biological function associated with them. In addition, by sampling

fully on the transitions the ability to make a thermodynamic
connection to experiment is put on a stronger footing. If the transitions in addition to the stable states are understood, then all the
kinetics and the relative free energies between the states becomes a
computed outcome to be compared with experiment.
1.2 Simulation
Techniques
The basic idea of molecular dynamics is the integration of a coupled set of differential equations. By specifying the initial positions
(xyz-space) through the use of a structure determined by X-ray or
NMR methods, and by drawing an initial assignment of velocities
to be consistent with a thermodynamic temperature, the equations
of motion in classical space are well defined. The problems with the
methods currently used are largely due to limitations in the time-scale
of sampling and the relative accuracy of the potential functions.
The potential function defines the forces that are used in the
F = m a = (x) that underlies the solution of the coupled set of
240
equations. By pushing the boundaries of the largest systems that

can be explored, by understanding the limitations of the forcefields, the community moves along towards a better set of confidence bars on when the simulations can be trusted.
In order to study the rare events associated with transitions,
several methods have been proposed. Probably the earliest is the
use of minimum energy, adiabatic, pathways. The work of Fisher
[8, 9], and Elber et al. [10] may serve as examples of this research
direction [1116]. In their approach a starting point is first minimized to a low gradient RMS on the force and then minimum
energy pathways are explored to move from one conformation to
another. This has been extended with some attempts to understand the effects of temperature on the pathways [17]. Similar
ideas have also been proposed for a type of low-energy Brownian
walk to find a minimum energy path [18]. Modifications to this
basic idea have been explored by many groups. For example work
by Paci and Karplus has examined the use of experimental restraints
to try and define the intermediates [19]. Perhaps the most currently well-known example is the transition path sampling methods
developed by the Chandler group [20].
In transition path sampling, based on ideas developed by Pratt
[21] and on work of Pratt and Chandler [22], an initial path estimate
is made to connect the beginning and ending states. From this initial
path a series of possible candidate improved paths are sampled in the
Monte Carlo space generated by random moves from the initial path.
The method has been explored and expanded from its initial foundations by the Bolhuis [23] and the Dellago groups [20]. In particular
there has been considerable progress in identifying the class of
perturbations from the original candidate path that will most likely
lead to the starting and ending points rather than diverging.
Alternative pathway finding algorithms have worked with ideas
developed by the Elber group [10]. In general these approaches use
a transform of the problem to the space of finding spatial coordinates that fit the discretized space. Examples of these are the
MaxFlux formalism [24], the string methods developed by the
Vanden-Eijnden group [25, 26], the OnsagerMachlup formalism
of Eastman and Doniach [27, 28] and the elastic rubber band
approach first proposed by Jnsson et al [2933]. The class of
mathematical optimization that is used depends on solutions to diffusion equations and asks for the most optimal arrangement of conformational steps along a fixed set of time points. In this way the
problem is transformed from one of dynamics to one of sampling
on intermediate conformations that sample well on the underlying
energy surface. A problem with this approach is that entropy contributions to the free energy of the conformational change are not
well sampled. Recent work from the Elber group has developed a
related, but new class of methods based on Milestoning [34].
A problem with this method is that a set of intermediate points
needs to be defined for the transition to proceed: this relates it back
241
to transition path sampling in that a small number of key intermediate

points act as assumptions for where the transitions need to be occurring and act to systematically steer the sampling. This can be very
good if the intermediates are well chosen and systematically bad if
the intermediates are poorly selected.
Another class of methods has been based on biased selection
from the dynamics. This has been implemented in different ways, and
a survey of some of the approaches and their relative advantages
and disadvantages was assessed by the Post group [35]. For example, in biased molecular dynamics [19, 36] a one-sided potential is
added to a standard MD simulation. This bias favors moves towards
a target structure and penalizes moves away and so attempts to
gently perturb the dynamics towards a transition. This is somewhat
similar to meta-dynamics [37] in that the potential is adjusted during the simulation and that it aims at creating a flat distribution of
states between the initial and final points. A related concept is that
of weighted ensemble Brownian (WEB) dynamics [38] where a set
of equally weighted conformations is iterated forward in time to
create a set of Brownian walkers that are eventually evenly distributed over the barrier crossing space. One advantage of the WEB
dynamics is that the relative probability of a particular transition path
is determined during the trajectory production. The Zuckerman
group [39] has shown that this method can produce high quality
transitions in their test systems.
Steered and targeted methods have been employed quite frequently for sampling on large conformational change. Both of these
methods have problems with providing an unbiased estimate of
change. Steered methods will have a large systematic error, depending on their choice of pulled coordinates and the rate of pulling. In
principle, though rarely in practice, the non-equilibrium work based
method can be used to correct for a calculation of the free energy
change with steered molecular dynamics, but the distributions
required to get convergence within the steered simulations have
not been examined carefully. Plastic or targeted molecular dynamics
will generate a single transition depending on the strength of the
additional forces and so would need to be run with a systematic
variation of force constants and sampled starting velocities to generate a range of transitions [40]. As the method is generally used it
provides an estimate of conformational change and is more likely to
create a systematic bias than a genuine sample of intermediate states,
due to its use of root mean square (RMS) as a force to connect the
starting with the ending states [41]. An extreme version of both the
usefulness of the transitions and the difficulty in assessing their
truthfulness is the Database of Macromolecular Movements web
pages maintained by the Gerstein group. This has the advantage of
presenting a large number of conformational changes for many protein systems, but the disadvantage of not being able to provide any
strong metric for the confidence with which the conformational
change has been presented [42].
242
1.3 Importance
Sampling and SDE
Biased sampling with or without correction has become an important

method in the molecular dynamics community. This can be seen as
starting from the importance sampling for Monte Carlo methods
that started with the Manhattan project in the 1940s. In that method
the distribution of events that are collected is focused by a known
umbrella or bias that forces more events to be sampled within the
biasing region. Since the biasing function is known, the unbiased
distribution can be recovered by dividing out. In effect this is simply
multiplying and dividing by the same number (i.e., by one). This use
of importance sampling and its improvement in accuracy for estimating ensemble averaged events was instrumental in modeling the
diffusion of neutrons and in the whole design of the atomic bomb.
As the method was developed its application to statistical mechanical
problems seemed natural, it was widely employed for ideal gases and
other homogeneous mixes. With the development of molecular
dynamics, several biasing methods, some with correction and some
without have been developed. For example, there is a strong community that uses the steered molecular dynamics methods of the
Schulten group. These can enable a new force to be added into the
simulation that pulls on particular atoms or groups of atoms and
thus enhances sampling of events that are tied into the pulled
direction. While it is easy to see the utility of being able to steer the
simulation, it is difficult to connect the resulting trajectories into
experimental measurements. Thus, other methods have worked
strongly on how best to combine information from multiple biased
windows to come up with an unbiased estimate.
Probably the most well known of these approaches is the
weighted histogram analysis method (WHAM) developed by
Swendson and coworkers [43]. In this approach the question is
phrased as how to optimally estimate the unbiased free energy surface, given samples from multiple biased umbrellas distributed along
a one or two dimensional reaction coordinate surface. The extension
of this idea by Shirts and colleagues (MBAR) suggests that the probability arguments for optimally combining information will continue
to be an important and fruitful topic for extending the sampling of
protein motions and their free energy surfaces [44, 45].
Dynamic Importance Sampling (DIMS) uses another type of
importance sampling, with correction, the difference being in that
the corrections are applied to the drawn events as they happen, by
comparing their relative likelihood from the biasing distribution to
that inferred likelihood from the unbiased distribution [4651].
The method was developed based on ideas from stochastic differential
equations and their simulations with variance reduction methods
[52]. This enables the DIMS method to sample on large conformational change without having to define a low dimensional reaction
coordinate for the conformational change.
An intriguing suggestion is that the apo forms of a protein will
have a free energy surface that is related to the holo forms [53, 54].
That is important, because the computational methodology that is
243
pursued by DIMS requires that the same set of atoms be present at

the beginning and the end of the simulation. Thus, by focusing the
computational power on the apo states and the protein conformational changes, without the ligand, much can be learned about the
behavior of the protein. Nonetheless, the nature of the changes from
the apo to the holo state is not clearly defined and thus the framework of the population shift model should be seen as a starting point
for a more detailed analysis of the role of the ligand in conformational
change. In particular, the nature of the transition state in the absence
and the presence of the ligand is an interesting question.
Transition States and Order Parameters

Chandler has provided the most commonly cited definition for the
transition state. This can be understood as a region where a set of
random walkers will be equally likely to fall back towards the starting
point or to move along to the ending point. Finding and describing
this surface for a protein system is still in its infancy, with the concept
being reasonable, but no well-defined algorithm that all agree will
work for all systems to define and to prove that the true transition
state has been found. This means that to prove that an intermediate
is at the transition state, or even part of the pathway for a large conformational change will, in principle, require an even larger set of
calculations to prove that the committer probabilities are consistent
with the expectations for the likelihood of a transition moving in each
direction along all candidate pathways. Thus there is a need for methods that can improve on the sampling of intermediate states and that
can provide connections to the unbiased estimate of the probability
of reaching those intermediate states from the starting point.
The first application of these ideas, developed using the framework of statistical mechanics, but not yet applied in a simulation setting, was to the isomerization of butane. This simple model system is
ideal at the level of transitions, since the one-dimensional reaction
coordinate is easy to define and the sampling, even at the time of the
Chandler paper, could be exhaustive. Chandlers work showed that a
6kT difference over the barrier from the minimum could be sampled
and the transition state as the population dividing surface between
these stable states could be analyzed. This opened the computational
community up to the idea of how to define a transition state and how
to determine rates over a free energy surface.
Recent work has shown that a poorly determined order parameter may lead to systematically poor results [55]. This is important for
the work applying the effective transfer entropy formalism [6, 56],
where multiple order parameters can be defined by the application
of nonlinear time series analysis. In addition this relates to efforts
to determine the kinetics from the transitions where both sampling
and possible systematic errors will need to be considered [57].
This is illustrated in a simple conceptual surface where q is an order
244
parameter along a complex multidimensional terrain. In one situation the transition state is fully sampled along the order parameter.
In contrast, for another situation, the order parameter creates a
systematic error in sampling along the transition surface, suggesting that the barrier is found about half-way along the order parameter, rather than realizing that the bottleneck is closer to the
starting state in the projection of the dimensions. This problem
can only be considered even more difficult for a multidimensional
system with an even larger number of degrees of freedom that need
to be averaged out in order to get good sampling.
Methods for Analysis of Functional Protein Modes

Principal component analysis (PCA) is a commonly used method to
extend the molecular dynamics fluctuations sampled in a trajectory
into a cohesive story for the most important sampled motions [58].
For instance, by performing PCA on 100 ns simulations of human
growth factor (see Fig. 2) it is possible to identify collective motions
that could be easily related to the conformational changes seen in
the crystal structures (Fig. 1). Thus, PCA has become the go-to tool
for this type of analysis, since there has been little else to use for this
type of question and the formal analysis of the method has not been
pushed to the same level as the force-fields and the long-time sampling [59]. But, as the time-scales for sampling has improved the
ability to question whether the principal component analysis is valid
and under what conditions has started to become a front burner
issue. This is an important concern, because if the modes that are
inferred from the trajectories can truly be used for longer-time analysis and extension, then the modes represent an important extension, by analysis, of the original trajectories. The current research,
however, suggests that the method concentrates too much information on fluctuations from a single stable state, so that if there are
complex motions between multiple stable states, there will be
nonlinear mixing that will make the modes defined by the method
less informative on what is happening on the molecular scale.
To infer more from the trajectories there are several alternatives.
One is that a complete description of the relative free energy surface
can be determined with sufficient sampling. This is an extension of
ideas from Stillinger about the connections between adiabatic and
at temperature simulations. In the complete, and most satisfying,
form of the description, all the quasi-stable states would be clearly
defined and the relative routes (pathways) between them would be
cleanly described. In a sense this is the underlying concept behind
the Markov analysis that is being pioneered by the Vande and IBM
groups [60]. In the Markov analysis the long-time molecular
dynamics trajectory is divided into basins of attraction and transitions between those basins. Ideally there would be more than sufficient sampling so that the effective modes defined by these stable
245
Fig. 2 Most dominant mode from PCA analysis for the human epidermal growth
factor receptor (HER3). The direction of the motion is similar to the conformational change observed in crystal structures (Fig. 1)
states could be defined afterwards by understanding the relative

contributions of each stable state. That would then, in principle,
provide a mechanism for a comparison of the Markov model and
the principal component model.
Some work has been proposed to follow the low frequency
modes as a way to sample on the longer time scale events [46, 6163].
In particular work on the low frequency modes for the K-channel
has had mixed results [64], suggesting that the modes seen in the
lowest energy starting state may not immediately lead to the appropriate transition pathways [65]. This has built on the ability to compute normal modes for larger and larger systems [66], an approach
that we use to improve our own sampling of candidate motions for
dynamic importance sampling [46]. For example, the Brooks group
has shown that the method of normal mode following can be
applied to a large range of sample calculations [67]. In a somewhat
related way, elastic network models using simple harmonic springs
between heavy atom groups have been used as a coarse-grained
246
approach to enable some aspects of conformational change to be

sampled [6870]. These models have, however, been shown to be
limited in their ability to sample on the all-atom pathways seen in
more complex transitions, so they may have utility in providing a
very fast look at a candidate transition, but will not enable a confident sampling on a range of intermediate states.
There has been intriguing work from the Thorpe group on
the use of conformational restraints and accounting for the main
degrees of freedom to try and determine pathways between states
[71]. Even if the methods are approximate, the suggestion that
some pathways can be determined by a rough accounting of
domain motions and flexible links is potentially important.
An alternative to the complete enumeration is to ask if the
model can be reduced by nonlinear methods [7277]. This is the
framework we explored for the Nitrogen receptor protein C [56]
and the larger and more complex human epidermal growth factor
receptor shown in Fig. 1 [6]. The approach consists of asking if a
nonlinear reduction to a reduced description can help to predict
the most important fluctuations from the starting point. Some
precedents for this type of question are in the work of the Wriggers
group and the use of mutual information from the Gilson and
Grubmuller groups. Mathematically the precedents to this idea are
from Granger causality analysis and transfer entropy ideas [7881].
It is worth mentioning, that previous to our work, the only use of the
effective transfer entropy in the setting of biomolecular simulations
is from the van der Vaart group [82]. In their approaches they
mainly focused on the use of the method to pick out key residues
in the bound and free states of a DNA binding protein.
Implementation of Dynamic Importance Sampling (DIMS) in CHARMM

DIMS (see Notes) is implemented and available in CHARMM
since the c35a1 release. It must be enabled before compilation by
adding the DIMS flag in the install.com file. In order to support
Normal Mode biasing the VIBBLOCK flag must be added as well.
In addition, the soft-ratcheting implementation of DIMS offers
support for the parallel version of CHARMM.
The basic work flow with DIMS is simple:
1. Load structures. The target must be stored in the DIMS coordinates set using COOR COPY DIMS.
2. Setup DIMS using the DIMS command.
3. Run dynamics.
4.1 Loading
the Structures
Loading structures in CHARMM is performed via the READ

COOR command. When using DIMS, the first structure that must
be loaded is the target structure. Then it must be copied to the
DIMS set using the COOR COPY TARG command. Once the
DIMS set has been initialized the starting structure can be loaded.
247
Fig. 3 If a trial molecular dynamics step is towards the target B then the motion
is accepted. A motion away from the target is only accepted with a certain probability, and this probability decreases as the trial move is away from the target.
The algorithm is similar to a Brownian ratchet system with the general direction
of the random walk being towards the target [83]
! target configuration
OPEN READ UNIT 1 CARD NAME target.crd
READ COOR CARD UNIT 1
CLOSE UNIT 1
COOR COPY DIMS
! target must be copied to DIMS set
! starting configuration
OPEN READ UNIT 1 CARD NAME start.crd
READ COOR CARD UNIT 1
CLOSE UNIT 1
4.2
Setting Up DIMS
DIMS offers two types of biasing mechanism: soft-ratcheting and

NM following. In the first case, DIMS uses a predefined progress
variable. At every timestep (see Fig. 3), once a trial structure x has
been generated by following Langevins equation, DIMS determines
whether the motion is towards or away from the target state.
In case that the movement is towards the target B then the motion
is accepted. If the trial move x moves away from the target, then it is
only accepted with a certain probability controlled by the rejection
constant [46]. In order to enable the soft-ratcheting mechanism,
a DIMS line must be added to the CHARMM input file:
DIMS DCAR 1e-5
orient 1000
SELE mydims END
SELE orient END
COFF 1.0 halt
! set up DCAR-DIMS
! re-orient at every 1000 step

! DIMSatom selection (biased atoms)
- ! RMS-fit atom selection

! stop biasing when target is 1.0 A from target
248
This example gently moves the system toward the target without
a restriction on the total time. The rejection constant in this
example is set to =1 105, selection of the rejection constant is
system dependent, and must be explored before production runs.
If the barrier height is not high enough under some conditions this
algorithm will not converge. When the barriers to conformation
change are small this approach will converge with a better DIMS
or OnsagerMachlup (OM) score. As DIMS uses by default RMSD
as the progress parameter, it is generally required to align the two
structures in order to eliminate rotations and translations. In the
previous example the structures are aligned every 1,000 steps using
the second atom selection for the alignment.
As previously mentioned, DIMS can also use a bias based on
the Normal Modes from the initial structure, and recalculate them
as the simulation progresses. Calculation of the modes is performed
by using the block normal mode method available in CHARMM.
A typical usage of the NM-biasing method is:
DIMS DBNM DSCALe 0.1 SKIP 500 BSKIP 50 NBIAs 27
- ! dims options
SERL GENR SCAL 0.5882 TMEM 420 MEMO 20 MEMA 400 NMOD 30 - ! BNM options
COFF 2.0 HARD
- ! NM Hard Cutoff
ORIEnt 20
- ! Orient every 1000th
SELE mydims END
- ! DIMS selection
COMB 3 NBES 15
- ! Store 15 best modes and then group them

- ! by 1, 2 and 3 for each try.
NWINDow 12
- ! Avoid normal modes previously used within

- ! a window of 12 modes
MTRA @I NMUNit 10
- ! @i trajectories per file, write output to unit 10
DSUNIT
- ! Use unit 11 to store dims score
Normal modes are computationally expensive calculations as

they require calculation and diagonalization of the Hessian matrix.
Therefore, DIMS only recalculates the normal modes every s steps
(500 in the example) and uses them for a given number of steps.
Modes are scaled by a constant factor determined DSCAL. Similar
to the soft-ratcheting, when RMSD is enabled, alignment of the
structures is recommended. Additionally, in this example, selfavoidance is enabled for a window of 12 steps from the current
move. Although all modes are evaluated, only the best 15 (NBES)
are kept, and linear combinations of up to three modes are evaluated at every time-step using a previously defined progress variable
(e.g., RMSD).
249
Notes
Trajectories generated by DIMS are completely independent [46],
therefore ensembles of transitions can be generated by running
multiple instances of CHARMM. Additionally, DIMS can also
compute the OnsagerMachlup action functional for each trajectory as the simulation progresses, this calculation can be activated
by adding the flag OMSC to the DYNA command [46]. The current implementation of DIMS is not limited to just RMSD as
progress variable, it can also use interatomic distances, angles, and
dihedrals, as well as, combinations of these in order to generate
collective variables. This flexibility allows, for instance, the use of
native contacts of the target structure as the progress variable.
References
1. McCammon JA, Gellin B, Karplus M (1977)
Dynamics of folded proteins. Nature
267:585590
2. Schlick T, Collepardo-Guevara R, Halvorsen
LA, Jung S, Xiao X (2011) Biomolecular modeling and simulation: a field coming of age. Q
Rev Biophys 138
3. Shaw DE et al (2010) Atomic-level characterization of the structural dynamics of proteins.
Science 330:341346
4. Creighton TE (1993) Proteins: structures and
molecular properties. Macmillan, New York
5. Ferguson KM et al (2003) EGF activates its
receptor by removing interactions that autoinhibit ectodomain dimerization. Mol Cell
11:507517
6. Perilla JR, Leahy DJ, Woolf TB (2013) Molecular
dynamics simulations of transitions for ECD epidermal growth factor receptors show key differences between human and drosophila forms of
the receptors. Proteins 81:11131126
7. Gerstein M, Lesk AM, Chothia C (1994)
Structural mechanisms for domain movements
in proteins. Biochemistry 33:67396749
8. Fischer S (1992) Conjugate peak refinement:
an algorithm for finding reaction paths and
accurate transition states in systems with many
degrees of freedom. Chem Phys Lett 194:
252261
9. Gruia AD, Bondar A-N, Smith JC, Fischer S
(2005) Mechanism of a molecular valve in the
halorhodopsin chloride pump. Structure
13:617627
10. Elber R, Karplus M (1987) A method for
determining reaction paths in large molecules:
application to myoglobin. Chem Phys Lett
139:375380
11. Olender R, Elber R (1997) Yet another look at

the steepest descent path. J Mol Struct
398:6371
12. Czerminski R, Elber R (1990) Self-avoiding walk
between 2 fixed-points as a tool to calculate
reaction paths in large molecular-systems. Int J
Quant Chem 24:167186
13. Ulitsky A, Elber R (1990) A new technique to
calculate steepest descent paths in flexible
polyatomic systems. J Chem Phys 92:
15101511
14. Czerminski R, Elber R (1989) Reaction-path
study of conformational transitions and helix
formation in a tetrapeptide. Proc Natl Acad Sci
U S A 86:69636967
15. Czerminski R, Elber R (1990) Reaction-path
study of conformational transitions in flexible
systemsapplications to peptides. J Chem Phys
92:55805601
16. Choi C, Elber R (1991) Reaction-path study of
helix formation in tetrapeptideseffect of sidechains. J Chem Phys 94:751760
17. Elber R, Shalloway D (2000) Temperature
dependent reaction coordinates. J Chem Phys
112:55395545
18. Berkowitz M, Morgan J, Mccammon J (1983)
Generalized Langevin dynamics simulations
with arbitrary time-dependent memory kernels. J Chem Phys 78:32563261
19. Paci E, Vendruscolo M, Dobson CM, Karplus
M (2002) Determination of a transition state at
atomic resolution from protein engineering
data. J Mol Biol 324:151163
20. Dellago C, Bolhuis PG, Csajka FS, Chandler D
(1998) Transition path sampling and the calculation of rate constants. J Chem Phys 108:
19641977
250
21. Pratt LR (1986) A statistical method for identifying transition states in high dimensional
problems. J Chem Phys 85:50455048
22. Chandler D, Pratt LR (1976) Statistical
mechanics of chemical equilibria and intramolecular structures of nonrigid molecules in
condensed phases. J Chem Phys 65:
29252940
23. Bolhuis PG, Chandler D (2000) Transition
path sampling of cavitation between molecular
scale solvophobic surfaces. J Chem Phys 113:
81548160
24. Huo S, Straub JE (1997) The MaxFlux algorithm for calculating variationally optimized
reaction paths for conformational transitions in
many body systems at finite temperature. J Chem
Phys 107:50005006
25. Ren W, Eijnden EV, Maragakis P, Weinan E
(2005) Transition pathways in complex systems: application of the finite-temperature
string method to the alanine dipeptide. J Chem
Phys 123:134109
26. Maragliano L, Fischer A, Vanden-Eijnden E,
Ciccotti G (2006) String method in collective
variables: minimum free energy paths and isocommittor surfaces. J Chem Phys 125:24106
27. Eastman P, Gronbech-Jensen N, Doniach S
(2001) Simulation of protein folding by reaction path annealing. J Chem Phys 114:
38233841
28. Onsager L, Machlup S (1953) Fluctuations
and irreversible processes. Phys Rev
91:15051512
29. Jnsson H, Mills G, Jacobsen KW (1998)
Classical and quantum dynamics in condensed
phase simulations. In Berne BJ, Coker
DF. Proceedings of the International School of
Physics. LERICI, Villa Marigola. pp. 385404
30. Crehuet R, Field MJ (2003) A temperaturedependent nudged-elastic-band algorithm.
J Chem Phys 118:95639571
31. Peters B, Heyden A, Bell A, Chakraborty A
(2004) A growing string method for determining transition states: comparison to the nudged
elastic band and string methods. J Chem Phys
120:78777886
32. Trygubenko S, Wales D (2004) A doubly
nudged elastic band method for finding transition states. J Chem Phys 120:20822094
33. Mathews D, Case D (2006) Nudged elastic
band calculation of minimal energy paths for
the conformational change of a GG noncanonical pair. J Mol Biol 357:16831693
34. Kuczera K, Jas GS, Elber R (2009) Kinetics of
helix unfolding: molecular dynamics simula-
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
tions with milestoning. J Phys Chem A 113:

74617473
Huang H, Ozkirimli E, Post CB (2009)
Comparison of three perturbation molecular
dynamics methods for modeling conformational transitions. J Chem Theory Comput 5:
13041314
Marchi M, Ballone P (1999) Adiabatic bias
molecular dynamics: a method to navigate the
conformational space of complex molecular
systems. J Chem Phys 110:36973702
Laio A, Parrinello M (2002) Escaping freeenergy minima. Proc Natl Acad Sci U S A
99:1256212566
Huber G (1996) Weighted-ensemble Brownian
dynamics simulations for protein association
reactions. Biophys J 70:97110
Zhang BW, Jasnow D, Zuckerman DM (2007)
Efficient and verified simulation of a path
ensemble for conformational change in a
united-residue model of calmodulin. Proc Natl
Acad Sci U S A 104:1804318048
Maragakis P, Karplus M (2005) Large amplitude conformational change in proteins explored
with a plastic network model: Adenylate kinase.
J Mol Biol 352:807822
van der Vaart A, Karplus M (2005) Simulation
of conformational transitions by the restricted
perturbation-targeted molecular dynamics
method. J Chem Phys 122:114903
Echols N, Milburn D, Gerstein M (2003)
MolMovDB: analysis and visualization of conformational change and structural flexibility.
Nucleic Acids Res 31:478482
Kumar S, Rosenberg JM, Bouzida D, Swendsen
RH, Kollman PA (1992) The weighted histogram analysis method for free-energy calculations on biomolecules. I. The method. J Comput
Chem 13:10111021
Shirts MR, Chodera JD (2008) Statistically
optimal analysis of samples from multiple equilibrium states. J Chem Phys 129:124105
Pohorille A, Jarzynski C, Chipot C (2010)
Good practices in free-energy calculations.
J Phys Chem B 114:1023510253
Perilla JR, Beckstein O, Denning E, Woolf T
(2011) Computing ensembles of transitions
from stable states: dynamic importance sampling. J Comput Chem 32:196209
Zuckerman DM, Woolf TB (1999) Dynamic
reaction paths and rates through importancesampled stochastic dynamics. J Chem Phys
111:94759484
Jang H, Woolf TB (2006) Multiple pathways in
conformational transitions of the alanine dipep-
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
tide: an application of dynamic importance sampling. J Comput Chem 27:11361141

Zuckerman DM, Woolf TB (2002) Rapid
determination of multiple reaction pathways in
molecular systems: the soft-ratcheting algorithm. (eprint). arXiv: physics/0209098
Zuckerman DM, Woolf TB (2000) Efficient
dynamic importance sampling of rare events in
one dimension. Phys Rev E 63:016702
Woolf T (1998) Path corrected functionals of
stochastic trajectories: towards relative free
energy and reaction coordinate calculations.
Chem Phys Lett 289:433441
Wagner W (1987) Unbiased Monte Carlo evaluation of certain functional integrals. J Comput
Phys 71:2133
Swift RV, Mccammon AJ (2009) Substrate
induced population shifts and stochastic gating
in the PBCV-1 mRNA capping enzyme. J Am
Chem Soc 131
Bahar I, Chennubhotla C, Tobi D (2007)
Intrinsic dynamics of enzymes in the unbound
state and relation to allosteric regulation. Curr
Opin Struct Biol 17:633640
Dickson BM, Makarov DE, Henkelman G
(2009) Pitfalls of choosing an order parameter
for rare event calculations. J Chem Phys
131:074108
Perilla JR, Woolf TB (2012) Towards the prediction of order parameters from molecular
dynamics simulations in proteins. J Chem Phys
136(164101):164101
Xin Y, Doshi U, Hamelberg D (2010)
Examining the limits of time reweighting and
Kramers rate theory to obtain correct kinetics
from accelerated molecular dynamics. J Chem
Phys 132:224101
Garca A (1992) Large-amplitude nonlinear
motions in proteins. Phys Rev Lett 68:
26962699
Ma J (2005) Usefulness and limitations of normal mode analysis in modeling dynamics of
biomolecular complexes. Structure 13:
373380
Singhal N, Snow C, Pande V (2004) Using
path sampling to build better Markovian state
models: predicting the folding rate and mechanism of a tryptophan zipper beta hairpin.
Kitao A, Go N (1999) Investigating protein
dynamics in collective coordinate space. Curr
Opin Struct Biol 9:164169
Petrone P, Pande V (2006) Can conformational change be described by only a few normal modes? Biophys J 90:15831593
251
63. Lange O, Grubmller H (2006) Can principal

components yield a dimension reduced description of protein dynamics on long time scales? J
Phys Chem B 110:2284222852
64. Denning EJ, Woolf TB (2010) Cooperative
nature of gating transitions in K(+) channels as
seen from dynamic importance sampling calculations. Proteins 78:11051119
65. Miloshevsky GV, Jordan PC (2007) Openstate conformation of the KcsA K+ channel:
Monte Carlo normal mode following simulations. Structure 15:16541662
66. Florence TF, Xavier G, Osni M, Yves-Henri S
(2000) Building-block approach for determining low-frequency normal modes of macromolecules. Proteins 41:17
67. Zhenggo W, Brooks BR (2005) Normalmodes-based prediction of protein conformational changes guided by distance constraints.
Biophys J 88:31093117
68. Zheng W, Doniach S (2003) A comparative
study of motor-protein motions by using a
simple elastic-network model. Proc Natl Acad
Sci U S A 100:1325313258
69. Kim MK, Chirikjian GS, Jernigan RL (2002)
Elastic models of conformational transitions in
macromolecules. J Mol Graph Model 21:
151160
70. Kim MK, Jernigan RL, Chirikjian GS (2002)
Efficient generation of feasible pathways for
protein conformational transitions. Biophys J
83:16201630
71. Lei M, Zavodszky MI, Kuhn LA, Thorpe MF
(2004) Sampling protein conformations and
pathways. J Comput Chem 25:11331148
72. Schreiber T (1997) Detecting and analyzing
nonstationarity in a time series using nonlinear cross predictions. Phys Rev Lett 78:
843846
73. Schreiber T (2000) Measuring information
transfer. Phys Rev Lett 85:461464
74. Kaiser A, Schreiber T (2002) Information
transfer in continuous processes. Physica D
166:4362
75. Kantz H et al (1993) Nonlinear noise reduction:
a case study on experimental data. Phys Rev E
48:15291538
76. Kantz H, Schreiber T (2004) Nonlinear time
series analysis. Cambridge University Press,
Cambridge
77. Stamati H, Clementi C, Kavraki LE (2010)
Application of nonlinear dimensionality
reduction to characterize the conformational
landscape of small peptides. Proteins 78:
223235
252
78. Barnett L, Barrett AB, Seth AK (2009)

Granger causality and transfer entropy are
equivalent for Gaussian variables. Phys Rev
Lett 103:238701
79. Barrett AB, Barnett L, Seth AK (2010)
Multivariate Granger causality and generalized
variance. Phys Rev E 81:041907
80. Hirata Y, Aihara K (2010) Identifying hidden
common causes from bivariate time series: a
method using recurrence plots. Phys Rev E
81:016203
81. Jamsek J, Palus M, Stefanovska A (2010)

Detecting couplings between interacting oscillators with time-varying basic frequencies:
instantaneous wavelet bispectrum and information theoretic approach. Phys Rev E 81:036207
82. Kamberaj H, van der Vaart A (2009) Extracting
the causality of correlated motions from
molecular dynamics simulations. Biophys J 97:
17471755
83. Magnasco MO (1993) Forced thermal ratchets.
Phys Rev Lett 71:14771481
Chapter 12
Accelerated Molecular Dynamics andProtein
Conformational Change: ATheoretical andPractical
Guide Using aMembrane Embedded Model
Neurotransmitter Transporter
PatrickC.Gedeon, JamesR.Thomas, andJeffryD.Madura
Abstract
Molecular dynamics simulation provides a powerful and accurate method to model protein conformational
change, yet timescale limitations often prevent direct assessment of the kinetic properties of interest.
A large number of molecular dynamic steps are necessary for rare events to occur, which allow a system to
overcome energy barriers and conformationally transition from one potential energy minimum to another.
For many proteins, the energy landscape is further complicated by a multitude of potential energy wells,
each separated by high free-energy barriers and each potentially representative of a functionally important
protein conformation. To overcome these obstacles, accelerated molecular dynamics utilizes a robust bias
potential function to simulate the transition between different potential energy minima. This straightforward approach more efficiently samples conformational space in comparison to classical molecular dynamics simulation, does not require advanced knowledge of the potential energy landscape and converges to
the proper canonical distribution. Here, we review the theory behind accelerated molecular dynamics and
discuss the approach in the context of modeling protein conformational change. As a practical example,
we provide a detailed, step-by-step explanation of how to perform an accelerated molecular dynamics
simulation using a model neurotransmitter transporter embedded in a lipid cell membrane. Changes in
protein conformation of relevance to the substrate transport cycle are then examined using principle
component analysis.
Key words Biological transport, Membranes, Molecular dynamics simulation, Neurotransmitter
transport proteins, Protein conformation
1 Introduction
Classic molecular dynamics (cMD) simulations are used to study
the kinetic behavior of proteins. By implementing fundamental
laws of motion, the technique is able, on an atom-by-atom basis,
to accurately predict the dynamic behavior of proteins in various
modeled environments. The technique effectively samples conformational space in a time-dependent manner, and if conducted for
253
254
Patrick C. Gedeon et al.
a sufficient number of time-steps, produces an accurate representation

of a systems potential energy landscape. The resulting trajectories
can then be analyzed, allowing for insight to a wide variety of
dynamic processes, including protein conformational change.
This technique is invaluable given that proteins are complex
structures, interacting dynamically with their environment in
response to a variety of thermodynamic, ionic, and other factors.
While experimentally determined structures provide valuable
three-dimensional information and often a necessary starting point
for modeling a given protein, the experimental process provides a
single snapshot of a protein, frequently in a non-physiological environment. In order to obtain a detailed mechanistic understanding
of protein function, as it is necessary for rational drug development
for example, a detailed understanding of dynamic protein conformational change in an environment representative of defined physiological or pathological conditions and under parameters amenable
to subtle molecular change is necessary.
While cMD simulations are highly beneficial in this regard, simulation time is limited to the nanosecond or microsecond timescale
at best, despite implementation of state-of-the-art supercomputers
with highly advanced processing ability. Although processing ability
continues to improve, the limit in achievable timescale is further
reduced as an increased understanding of biological systems results
in the need to model more accurate and complex systems, containing multimers of large proteins in various contexts for example.
Given these limitations, cMD simulations often fall short of
effectively exploring the full energy landscape. Exceedingly large
numbers of computational steps are required for rare events to
occur allowing systems to overcome energy barriers and transition
from one potential energy minima to another. Furthermore, the
energy landscape for proteins often consists of multiple potential
energy wells separated by high free energy barriers. Still, by means
of exceedingly long all-atom cMD simulations, often conducted
on special-purpose machines using specialized force fields, investigators have achieved timescales far in excess of those previously
accessible to computational study, gaining insight into a variety of
complex dynamic protein behavior, including protein folding and
conformational change within the folded state [1, 2].
In order to avail a more widely available and expeditious technique to simulate the infrequent transition between potential
energy minima and therefore realistically capture protein dynamics, alternate computational approaches are necessary. One such
approach, termed steered molecular dynamics (SMD), involves
applying external forces to a system to explore its mechanical
responsiveness [3]. While this approach has proven effective in a
number of contexts [48], it relies on user defined forces most
accurately applied given prior knowledge of the conformational
state of interest.
Accelerated Molecular Dynamics andProtein Conformational Change
255
In contrast, accelerated molecular dynamics (aMD) provides a

straightforward and effective way to simulate infrequent events
required for protein conformational change without previous knowledge of conformational states, potential energy wells, or barriers.
The aMD method alters the amount of computational time a system
spends in a given potential energy minima by adding a bias potential,
V(r), to the true potential in such a way that potential surfaces in
vicinity of the minima are raised, while those closer to the barrier or
saddle point are not affected. This allows for a reduction in computational time spent at a potential energy minima and an increase in
the ability of a system to move over potential barriers. The effect of
the bias is removed by correcting statistics sampled on the biased
potential. The method results in preservation of the underlying
shape of the potential energy surface and converges to the correct
canonical probability distribution, allowing for an approach that
efficiently and accurately explores conformational space [9].
Today, aMD simulations are routinely performed to assess
time-dependent protein conformational change [1015] and are
fully integrated into commonly used software packages including
NAMD [16, 17] and Amber [18, 19]. Performing aMD simulations is straightforward and consists of steps similar to those in
cMD simulations, with the addition of defining aMD specific
variables.
2 Theory
Expanding on a previous hyperdynamics method explored by
Voter [20, 21], the aMD method allows for assessment of protein
conformational change by reducing the computational time a simulated protein spends in a potential energy basin, allowing the system
to transverse potential energy barriers more readily. This is accomplished by adding a bias potential to the true potential when the
systems potential energy falls below a threshold level (Fig.1). While
Voters method of implementing the bias potential requires the
Hessian matrix to be diagonalized at each time step in order for
identification of transition state regions, the aMD method is based
on a simpler bias potential proposed by Steiner etal. [22] and
implemented by Rahman and Tully [23]. In this puddles method,
the bias potential is selected so that the produced modified potentials near the minima remain constant if the true potential of the
system falls below a selected threshold level. Accordingly, diagonalization of the Hessian matrix is not required at each step, allowing for the simulation method to be applied to larger systems such
as proteins.
Specifically, a nonnegative, continuous bias boost potential
function V(r) is defined such that when the true potential of a
system, V(r), falls below a specified boost energy, E, the simulation
is carried out using the modified potential V*(r)=V(r)+V(r),
3000
V(r)
3500
4000
4500
256
2000
2500
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Fig. 1 Graphical representation of the biased potential, the threshold boost

energy, E, and the normal potential
while when V(r) is greater than or equal to E, the simulation is carried

out using the true potential such that V*(r)=V(r). The relationship between the true potential, modified potential, boost energy,
and bias boost potential is
V (r), V (r) E ,
V * (r) =
V (r) + DV (r), V (r) < E
(1)
aMD provides a robust method that, in comparison to cMD

simulation, accelerates the state to state transition of large molecules such as proteins. By using the bias potential to modify the
true potential of the system, the transition of the system from one
state to another takes place at an accelerated rate with a timescale
t* that is nonlinear, such that

Dt i* = Dt e
bDV r (t i )
(2)
Accordingly, the clock is advanced at each step based on the

strength of the bias potential, where t represents the actual time
step on the unmodified potential. If the bias potential, V(r)=0,
as is the case when V(r) is greater than or equal to the boost energy E,
the system is on true potential and t*=t. Statistics can then be
257
used to estimate the total simulation time based on the following

equations:
t * = Dt i* = Dt e
t* = t e
bDV r (t i )
bDV r (t i )
bDV r (t i )
(3)

(4)
e
is the boost factor, a measure of the simulation
where
acceleration extent, and N is the total number of simulation steps.
Importantly, the aMD method converges to the canonical distribution, allowing for accurate determination of equilibrium and
other thermodynamic properties. The phase space for the modified
potential can be reweighted at each point by multiplying individual
configurations by the bias strength at each configuration, resulting
in a corrected ensemble average equivalent to that observed with
the normal potential [24, 25]. This approach as well as other
reweighting approaches allowing for accurate free energy calculation of the resulting trajectories continue to be explored.
In the original method described by Rahman and Tully, the
boost potential V(r) is defined as EV(r). With this method,
when the true potential energy is below the threshold boost energy
E, the modified potential energy, V*(r)=E. This produces flat
regions or puddles at potential energy basins. While this method
is computationally inexpensive, complications due to the discontinuity at points where the unmodified potential meets the modified
potential result in the need for special computations which increase
the computational burden overall. More importantly, at a high
boost energy, the flat modified potential exists at a level higher
than most transition state regions. As a result the system may
undergo a random walk and is slow to converge.
Alternatively, the aMD method utilizes a snow drift approach
which fills the minima, producing a more smooth landscape. The
shape of the underlying potential energy surface is maintained even
at a high boost energy E. The method results in a smooth transition where the unmodified potential energy above the boost energy
meets the modified potential energy. In order to accomplish this,
V(r) is defined by the equation
( E - V (r))
,
a + ( E - V (r) )

2
DV (r) =

(5)
The tuning parameter determines the depth of the modified

potential energy basin, such that when =0, the energy basin is flat,
just as in the method described by Rahman and Tully, while as
increases the depth of the modified potential energy basin decreases.
It is important that the values for E and be selected carefully
in the course of an aMD simulation, as these values determine how
4000
5000
V(r)
=1
= 100
= 500
= 1000
= 5000
3000
V(r)
2000
258
Fig. 2 Graphical representation of a hypothetical potential energy function and

different bias potentials plotted at various values and at a relatively low threshold boost energy, E
aggressive a given simulation is accelerated and how accurately the

energy landscape is maintained respectively. The value of E should
be selected such that it is larger than the minimum V(r), Vmin, at
the beginning of the simulation. Consider for example a situation
in which E is selected such that it is lower than a systems potential
energy minimum. Where V(r)>E, V*(r)=V(r), the energetics of
the system will be maintained just as in a typical molecular dynamic
simulation. Alternatively, E can be selected such that it is greater
than a systems potential energy minimum, V*(r)=V(r)+V(r).
Since proteins tend to have several potential energy minima spaced
closely together, it is advisable to calculate the average true potential energy of a protein over a short period of cMD simulation, and
to use this value as the minimum potential energy from which E
can be determined.
When the value of E is set low, the modified potential energy
remains below the transition state regions and the probability of
overcoming energy barriers and enhanced energy sampling is
diminished (Fig.2). In this situation the value of is of lesser
importance to protein conformational sampling and the aMD simulation as a whole, providing that is large enough so that the
modified potential energy basins are not flat (puddles) since this
would result in the calculated force becoming discontinuous where
the modified potential energy equals the boost energy.
259
3000
2000
V(r)
4000
5000
V(r)
=1
= 100
= 500
= 1000
= 5000
Fig. 3 Graphical representation of a hypothetical potential energy function and

different bias potentials plotted at various values and at a relative high threshold boost energy, E
Alternatively, when the value of E is set to be high, the value of

becomes much more significant (Fig.3). If the value of E is set
high and the value of is set low, the modified potential becomes
equivalent in most places and the system undergoes a random
walk. In order to prevent this and to maintain the basic shape of
the underlying unmodified potential energy, when the value of E is
set high, the value of must likewise be set high.
Taken together, as indicated above, the first step in determining
the correct values for E and is to determine the minimum potential energy of a given system, which is best determined by calculating the average potential energy of the true potential over a short
period of cMD simulation, and then using this value as the minimum potential energy (Vmin). The value of E should then be chosen
so that it is greater than the minimum potential energy. The magnitude by which E is greater than Vmin will determine the aggressiveness of acceleration in the simulation. In practice, the value of E is
usually selected so that it equals Vmin plus 2.53.0 times the number of amino acids in the simulation. The value of should then be
set so that it is equal to EVmin. At this value the modified potential energy will preserve the landscape of the underlining unmodified potential energy wells and will merge with the original potential
in a smooth fashion [9].
260
While aMD simulation allows rapid sampling of multiple conformational states, analyzing the resultant trajectories can be problematic
due to the large amount of data produced. While the traditional root
mean square deviation (RMSD) method can be used to distinguish
between different conformational states, this method is not as effective as the system undergoes transitions between subtle yet significant
conformational states yielding low RMSD values. Instead, principal
component analysis (PCA) can be used as a more sensitive method to
distinguish between different conformational states.
PCA reduces the dimensionality of large data sets by calculating a covariance matrix and it eigenvectors. Vectors with the highest eigenvalues become the most significant principal components.
When principal components are plotted against each other, similar
structures cluster. Each cluster then theoretically represents a different protein conformational state.
Since PCA can be calculated using coordinates from any subset
of atoms within a given protein, atom selection can have a large
effect on observed outcomes. A common protocol used to avoid
sample noise from random fluctuations is to calculate the PCA only
for backbone carbon atoms. Alternatively, specific residues or segments of a given protein can be selected based on experimental
observations and isolated in a PCA analysis. For example, following aMD simulation of the leucine transporter, the combination of
coordinate positions for the backbone carbon atoms of the transmembrane helical domains 1b and 6a was a better discriminator of
conformations than the calculations of either the whole structure
or any other helical domains alone or in combination [15].
3 Methods
Below we present a protocol for performing aMD simulation and
assessing protein conformational change. As an example, we highlight our recent work modeling the bacterial leucine transporter
(LeuT), a homologue of the eukaryotic Na+/Cl-dependent neurotransporters responsible for terminating synaptic transmission
by driving the cellular uptake of neurotransmitters, including the
biogenic amines. These proteins are the targets of numerous pharmacological compounds and their dysfunction is associated with
disorders of the nervous system. Through the use of aMD simulation, we have gained insight to the function of this class of proteins [26, 15].
In order to provide the reader with a guide such that an equivalent technique can be extended to the study of any protein of
interest, we will discuss (1) building a simulation environment
suitable for aMD simulation and the study of protein conformational change, (2) preparatory energy minimizations, heating the
system, cMD equilibration, and aMD production runs, and (3)
analysis of protein conformational change using PCA.
261
3.1 Computer
Software
A text editor is useful for many of the manipulations described

below. Helpful functions include the ability to quickly copy and
paste several lines of text, search and replace functionality with
features such as wild-cards and line number restrictions and the
ability to quickly move to a given line number or string of text.
More complicated tasks are often facilitated by the use of a procedural programing language such that complex edits to a PDB file
are straightforward. UCSF Chimera [27] is a program for visualization, analysis, and manipulation of molecular structures and
when used in conjunction with MODELLER [28] is useful in this
context for adding missing residues during the initial preparation
of the system. The Visual Molecular Dynamics (VMD) package
[29] offers robust molecular visualization as well as a variety of
molecular building and manipulation options; it will be used here
for visualization, the creation of lipid membrane starting coordinates, and embedding the protein in the lipid membrane. AMBER
[18] offers highly scalable code designed for high-performance
simulation of large molecular systems; it will be used in the following example for placing hydrogen atoms, solvation, energy
minimization, heating the system, cMD equilibration, and aMD
production runs. The Bio3D utility package in the R statistical
computing software environment [30] will be used to perform
PCA analysis.
3.2 Obtain
theStarting Protein
Coordinates
Prior to performing an aMD simulation, it is necessary to obtain

starting coordinates for the protein of interest. These may come
from crystal structures, nuclear magnetic resonance (NMR)
structures, or various computational models. In the case of LeuT,
the crystal structure coordinates [31] can be downloaded from
the Protein Data Bank, entry 2A65 (www.rcsb.org; MMDB
accession no. 34395). When working with a new PDB file, it is a
good practice to open the PDB file with a text editor to carefully
review the header information in order to gain an overall understanding of the types of residues present, both those that are part
of the protein and otherwise, potential limitations of the model,
and any other pertinent information.
3.3 Build
Coordinates foraLipid
Membrane
For many membrane proteins, the surrounding lipid membrane

structure plays a critical role in influencing protein dynamics. In
order to realistically model and most accurately capture transporter
dynamics we will embed the LeuT in a lipid bilayer. Here we will
demonstrate this by using the VMD Membrane Plugin 1.1 to generate the starting coordinates for a lipid membrane and then use
VMDs graphic user interface (GUI) front end to align the newly
generated membrane with the existing crystal coordinates.
The VMD Membrane Plugin 1.1 uses pre-built patches of lipid
bilayers to generate a rectangular membrane of the necessary size.
The algorithm used to build the lipid bilayer patches creates two
lipid layers, with each layer consisting of a two-dimensional
262
hexagonal lattice of lipids. In order to allow for straightforward

insertion of proteins, the lipid tails are nearly fully extended.
The distance between the layers and the lattice period is set to
correspond with experimental data corresponding to actual membrane thickness and surface density respectively. In order to introduce some disorder to the patches, each lipid is placed in the lipid
plane in a random orientation and a truncated Gaussian spread is
used in the perpendicular direction. Additional disorder is introduced by a 1picosecond equilibration in vacuum, serving the additional function of eliminating steric collisions while still leaving
most of the lipid tails extended. Finally, the VMD Membrane
Plugin 1.1 properly hydrates the lipid head groups, simplifying
downstream solvation and equilibration of the protein-membrane
system. Amore detailed description of the VMD Membrane Plugin
1.1 can be found at: http://www.ks.uiuc.edu/Research/vmd/
plugins/membrane/.
In the following steps, we build the membrane coordinates
inVMD:
1. Select ExtensionsModelingMembrane Builder to launch
the VMD Membrane Plugin.
2. Under Lipid, select POPE. This defines the lipid composition.
Currently, POPE and POPC lipid structures are supported.
The lipid composition should be selected to most closely match
the lipid composition invivo as guided by experimental data.
Here we selected POPE as this is compatible with the
phosphatidylethanolamine-
predominant monoamine transporter protein milieu.
3. Under Membrane X Length and Membrane Y Length we specify the membrane dimensions in ngstrms. For this system,
we enter 110 for both dimensions.
4. The Output Prefix and Topology can be left as the default settings. Click Generate Membrane to create the specified
membrane.
3.4 Align
theMembrane
andProtein
Coordinates
Positioning the membrane properly must be done carefully and

with appropriate consideration for the underlying biology. One way
to accomplish this is to position the membrane such that polar head
groups and hydrophilic lipid tails make contact with specific residues of the transmembrane protein, as guided by experimental data.
If this data is not available, an alternate method involves calculating
the center of mass of the main center vestibule of the transmembrane protein and aligning this with the center of mass of the generated membrane. This process is explained in detail in the VMD
Membrane Proteins Tutorial by Alek Aksimentiev etal. [33]. Here,
we position the LeuT based on a combination of experimental data
263
and an analysis of the polarity of the residues at the transporter-lipid

interface as follows:
After generating the membrane and water coordinates as
detailed in Subheading3.3, without closing VMD, load the LeuT
crystal structure file (PDB 2A65).
1. Load PDB 2A65 by selecting FileNew Molecule. Under
Load Files for: ensure that New Molecule is selected. Click on
Browse and navigate to the folder where the coordinates are
saved. Selected the correct file and click Open. Under Determine
file type: PDB should be selected automatically if the file was
given the correct extension (.pdb). Select Load.
2. In the VMD Main window, the currently loaded molecules are
listed in the Molecule List Browser. Two separate molecules
should be listed in this window: one containing the generated
membrane and water coordinates and a second containing the
2A65 coordinates.
In order to better visualize our protein, membrane, and water
molecules, we will alter the graphical representation. Select
GraphicsRepresentations and adjust the representation as follows:
1. Under Selected Molecule select the molecule corresponding to
the lipid membrane. Under selected atoms type resname POPE
and not hydrogen and click on Apply.
2. Click on Create Rep. This creates a second representation for
the molecule containing the lipid coordinates. Under Selected
Atoms type resname TIP3 and click Apply. This creates two
separate representations for the coordinates generated by
Membrane Builder 1.1, one representation for the lipid molecules (without hydrogen atoms) and a second representation
for the water molecules. In the white box under Create Rep,
where the two separate representations are listed, double click
on the representation containing the water molecules. The text
should turn red and the water molecules should disappear
from the OpenGL Display. While the water molecules are
important to the simulation, positioning the membrane correctly in relation to the protein is facilitated by hiding the water
molecules.
3. Under Selected Molecule click on the dropdown menu and
select the molecule containing the LeuT coordinates. Change
the Coloring Method to ResType and Drawing Method to
Cartoon and click Apply. This allows us to distinguish between
nonpolar residues (white), polar residues (green), basic residues (blue), and acidic residues (red). This information will be
helpful in properly positioning the lipid membrane.
4. In the VMD Main window, select DisplayOrthographic.
264
Fig. 4 LeuT positioned within the POPE lipid membrane. Nonpolar residues (white) at the membrane interface
in contact with the hydrophobic lipid tails of the POPE membrane and polar residues (green) at the membrane
interface either in contact with the polar head groups our out of the transverse plane of the lipid bilayer
We are now ready to move the crystal structure coordinates

while holding the newly generated membrane coordinates fixed.
1. In the VMD Main window select MouseMoveMolecule.
All the coordinates within the 2A65 crystal structure are now
moved by clicking with the left mouse button on the transporter
in the OpenGL Display window and dragging. Holding the shift
key and moving the mouse allows the membrane and water coordinates to be rotated about the point on the screen that is clicked.
Using the left mouse button rotates about the x or y axis of the
screen, while using the middle or right button rotates the coordinates about an axis perpendicular to the screen.
2. Using the keyboard and mouse commands indicated above, we
now move the crystallized coordinates such that the transporter
is positioned correctly in relation to the membrane. The general
orientation of the transporter in the lipid membrane is described
based on experimental data obtained by Yamashita etal. [31].
More exact placement is obtained by positioning to the greatest
extent possible (1) the nonpolar residues (white) at the membrane interface in contact with the hydrophobic lipid tails of the
POPE membrane and (2) the polar residues (green) at the
membrane interface either in contact with the polar head groups
our out of the transverse plane of the lipid bilayer (Fig.4).
265
3. When satisfied with the placement of the transporter in relation

to the membrane coordinates, in the VMD Main window
select FileSave Coordinates. Under Save data from use the
dropdown menu to select the molecule containing the correctly positioned 2A65 coordinates. Under Selected atoms
select all. Ensure pdb is selected under File type. Click Save,
navigate to an appropriate directory and enter a file name to
save the lipid coordinates. Finally, click Save again. The PDB
and PSF files generated for the membrane and membrane solvation are automatically saved to the VMD working directory
when the membrane structure is generated. Since no modifications were made to the membrane it is not necessary to re-save
these files.
3.5 Prepare
theStarting Protein
Coordinates
In the case of 2A65, in addition to including the coordinates for the

amino acids that make up the transporter, the Protein Data Bank
entry includes crystallized coordinates for ions found in the binding
pocket, the substrate leucine, water molecules, and a B-octylglucoside
(BOG) residue. In our simulation, we wish to include the crystallized transporter with ions, but remove the substrate, water molecules, and B-octylglucoside residues. The substrate is removed in
this example in order to sample transporter conformations that
occur in the absence of substrate. The B-octylglucoside residues are
removed since they are not likely to influence the dynamic process
of substrate transport, the interest of this study. It is important that
decisions in regard to experimental design, such as that mentioned
previously, be made with consideration for underlying biology and
specific study objectives.
In order to accomplish this, we will use a text editor to modify
the crystal coordinates saved after alignment with the membrane in
Subheading3.4. While it is not essential, it can be helpful from an
organizational perspective to save the coordinates for each individual subunit/chain in new separate PDB files, keeping only lines
containing coordinate information. If several chains are to be
included in a single coordinate file, it is necessary to add the TER
keyword between each chain, although this is not needed between
water molecules. In the case of LeuT (PDB 2A65), we create two
new separate PDB files from the membrane aligned coordinates,
one containing the coordinate lines for the residues of the transporter and one containing the coordinate lines for the binding
pocket ions (with each line of ion coordinate information separated by the TER card).
At this stage it is also important to consider special residues
with modifiable protonation states, such as cysteine and histidine
residues. When PDB coordinate files are initially read into LEaP to
produce input files for production calculations in AMBER, regular
protonated cysteine residues should be named CYS, deprotonated
or metal atom bound cysteine residues should be named CYM,
266
and cysteine residues involved in disulphide bridges should be

named CYX.Similarly, histidine residues should be named HIE
when protonated in the epsilon position, HID when protonated in
the delta position, and HIP when protonated in both the epsilon
and delta positions. In the LeuT crystal structure, there are no
cysteine residues, however, we replace the residue name HIS with
the residue name HID.
It is helpful also to remove all hydrogen atoms included in the
newly made coordinate files. Coordinates for hydrogen atoms in
NMR structures may be inaccurate. Furthermore, NMR naming
conventions are different from PDB naming conventions. Still, if
needed, it is possible to use LEaP to correct the nonstandard
hydrogen names. In the present example, all hydrogen atoms will
be positioned using LEaP and then energy minimized prior to
cMD simulation.
Finally, note that some PDB files contain alternate coordinates
for specific residues. When these coordinates are read into AMBER
using LEaP, only the A conformation is used. If a different coordinate set is needed, this can be accomplished by modifying the PDB
file to contain only the coordinates of interest.
3.6 Build Missing
Residues
Review of the headers in the 2A65 PDB file indicates that the four
most N- and C-terminal residues and residues N133 and A134 are
missing. We will build the missing non-terminal residues (N133
and A134) into our structure, but leave out the terminal residues.
This task can be accomplished in UCSF Chimera and MODELLER
as follows:
1. In UCSF Chimera load the membrane aligned 2A65 PDB file
by selecting FileOpen and navigating to the PDB structure.
While it is often not necessary to specify the file type it is a
good practice to do so to avoid any potential unexpected
errors.
2. Select ToolsStructure EditingModel/Refine Loops. This
brings up sequence information for the protein. Missing residues are highlighted with red boxes. Selecting Model/Refine
Loops will also open UCSF Chimeras interface to MODELLER.
3. In the Model Loops/Refine Structure window select non-
terminal missing structure. This selects residues that are found
in the PDB SEQRES record, but missing from the PDB file
coordinates. This selection only selects residues constrained at
both ends by existing structures, so the N- and C-terminal
missing residues will not be built.
4. Other building parameters can also be modified. In this example, we allow 0 residues adjacent to the missing regions to
move, generate 1 model, use the Discrete Optimized Protein
Energy (DOPE) energy score [32], and run MODELLER
267
using the Web service. After selecting Ok, progress toward

completion of this step can be tracked in the lower left hand
side of the main UCSF Chimera window.
5. The new coordinates can be saved by selecting FileSave
PDB. Give the file a new name by adding text in the box to the
right of File name, select the newly created model in the box to
the right of Save models and select Save.
6. Using a text editor, open the newly saved PDB file containing
the missing residues, navigate to and highlight the newly created
residues and cut/paste the new ATOM information into the
appropriate place in the previously prepared PDB file containing
only the transporter coordinates. Note that since we have not
built the N-terminal residues, the PDB file output by UCSF
Chimera has renumbered the residues beginning in the first
position. Be careful to select and copy the correct residues.
7. Renumber the new residues and atoms appropriately in the
transporter coordinate file being prepared for input to
AMBER.While renumbering the two newly created residues is
straightforward to accomplish manually, renumbering the
atoms manually is a daunting task. To assist in doing this, the
Fortran 90 source code provided below can be used to compile
a program that reads in a PDB file containing 4056 atoms and
outputs a PDB file with identical information but appropriately
numbered atoms. Different PDB manipulation scripts can be
written in any procedural programing language of choice to
assist in performing more complex alterations to various components of a PDB file.
program renumber_atoms
implicit none
character (len = 5) :: first
character (len = 6) :: old
character (len = 64) :: last
integer :: i = 1
do i = 1, 4056
read (5,"(a5,a6,a64)"), first, old, last
write (6,"(a5,i6,a64)"), first, i, last
end do
end program
3.7 Remove
Overlapping
Membrane Structures
In Subheading3.4 PDB and PSF files were generated for a lipid

bilayer with solvated polar head groups. In Subheadings3.5 and
3.6 the LeuT transporter and binding pocket ions were correctly
positioned within the bilayer and the missing transporter residues
were built respectively. There are still, however, lipid and water
molecules that overlap with the transporter structure. In this
section VMD will be used to delete any lipid or water residues that
overlap or are too close to the transmembrane protein.
268
First, a protein structure file (PSF) is generated from the LeuT

only PDB file created in Subheading3.6. This is done in VMD
using the automatic PSF generator as follows:
1.
Open VMD and in the VMD Main window select
ExtensionsTk Console. In the VMD TkConsole that opens,
navigate to the directory where the LeuT PDB file generated
in Subheading3.6 is located by using the cd command followed by the path to the appropriate directory and pressing
enter.
2. Load the PDB file by entering the command mol new filename.pdb and pressing enter.
3. In the VMD Main window select ExtensionsModelingAu
tomatic PSF Builder.
4. Under Step 1: Input and Output Files ensure the PDB file
loaded in step 2 is selected under Molecule and that the
Output basename is leut_autopsf.
5. Under Step 2: Selections to include in the PSF/PDB make sure
that Everything is selected.
6. Under Step 3: Segments Identified select Add a new chain.
Under Chain Name enter LeuT. The First Atom and
Last Atom should be changed to the first and last LeuT
atom number in the PDB file that was loaded (1 and 4057
respectively). Under N terminal patch and C terminal
patch add NTER and CTER respectively. Finally select
Add chain.
7. Under Step 4: Patches select Apply patches and finish
PSF.This will create in the current working directory a PSF
and PDB file with the name entered under Output
basename.
Next, the previously generated membrane.psf and membrane.
pdb files are loaded, the membrane and transporter structures are
combined into a single PSF and PDB to allow for comparison,
lipid and water molecules that are within a specified distance from
the protein are deleted, and the resulting PDB and PSF files for the
system with overlapping residues deleted are output. All of these
steps are easily accomplished with the following script that can be
run from the VMD Tk Console:
# Load package
package require psfgen
# Load starting structures
resetpsf
readpsf membrane.psf
coordpdb membrane.pdb
readpsf leut_autopsf.psf
coordpdb leut_autopsf.pdb
269
# Output a combined structure

writepsf leut_mem.psf
writepdb leut_mem.pdb
# Load combined structure
mol load psf leut_mem.psf pdb leut_mem.pdb
# Delete lipid and water molecules within 1.2
A of protein
# Get segids to check
set check_sel [atomselect top "resname POPE
or water"]
set check_sel_segs [lsort -unique [$check_
sel get segid]]
foreach list_seg $check_sel_segs {
# find atoms
set current_atom [atomselect top "segid
$list_seg and within 1.2 of protein"]
# delete residues
set current_res [lsort -unique [$current_
atom get resid]]
foreach res $current_res {
delatom $list_seg $res
}
}
# Output modified structure
writepsf leut_mem_no_overlap.psf
writepdb leut_mem_no_overlap.pdb
# Load modified structure for viewing
mol load psf leut_mem_no_overlap.psf pdb
leut_mem_no_overlap.pdb
The above script is executed in VMD by inserting the commands in a text document and saving the file to VMDs current
working directory. Within the Tk Console, the script is then run by
executing the command source filename, where filename is
the name of the file the above script is saved as.
Next, assess the resulting structure. The script above can be
modified such that the cutoff distance at which residues are deleted
is changed. When satisfied with the resulting structure, save PDB
files for the membrane alone and the membrane solvation alone as
follows. In the VMD Main window select FileSave Coordinates.
Under Save data from use the dropdown menu to select the correct molecule. Under Selected atoms select resname POPE and not
hydrogen. Ensure pdb is selected under File type. Click Save, navigate to an appropriate directory and enter a file name to save the
lipid coordinates. Finally, click Save again. Repeat this process
again to save the water molecules alone by changing the Selected
atoms to waters and not hydrogen.
270
3.8 Generate
aNonstandard POPE
Lipid Unit Using
Antechamber
If we loaded the files generated thus far into AMBERs LEaP

program using a standard protein force field, the POPE lipid residue would not be recognized. In this next section we will use the
Antechamber tool set that comes with AMBER in order to generate an input file that is necessary in order to include the POPE lipid
structure in our simulation. The Antechamber tool set works in
conjunction with the general AMBER force field (GAFF), a force
field that is designed to have general atom types such that it can
provide broad applicability to a variety of different molecules. It can
also be used in conjunction with traditional AMBER force fields
that may be more appropriate for the rest of the molecules in a
simulation. As such, Antechamber and GAFF are highly useful
entities. The Antechamber tool set automatically identifies atom
and bond types, judges atomic equivalence, provides a residue
topology file, and suggests reasonable alternatives for any resulting
missing force field parameters. Given the highly automated nature
of this process, however, it is critical to carefully evaluate the resulting output to verify suitability.
Here we will use Antechamber to assign atom types and charges
for the POPE residue. To begin, open the PDF file generated previously containing the coordinate information for the POPE lipid membrane. Copy and paste all the lines that together make up a single
POPE residue (there are 125 ATOM lines that together represent a
single POPE residue) into a new text file called pope.pdb. Next, in
order to use antechamber to create the prepin file necessary to
define the new POPE unit in LEaP, run the following command:
$AMBERHOME/bin/antechamber -i pope.pdb -fi pdb
-o pope.prepin -fo prepi -c bcc -s 2
The $AMBERHOME/bin/antechamber command initiates
the antechamber utility. The -i command specifies the name of the
input file and the -fi command specifies the format of the input
file; antechamber accepts several other input file formats. The -o
command specifies the name of the output file and the -fo prepi
command tells antechamber to output the file in the PREP format
used internally by LEaP. -c bcc directs antechamber to calculate
the atomic point charges using the BCC charge model. Finally, -s
2 specifies the level of verbosity that antechamber is to provide.
Completion of this step will result in several new files in the
directory where the script is executed. The files in capital letters
are intermediate files useful for troubleshooting; should analysis of the following produce the intended results they can safely
be deleted. The divcon.out file results from quantum mechanics
calculations necessary in order to determine the point charges
for individual atoms. It is useful to assess this file to ensure that
this process has completed without error. Finally, the pope.prepin file is the file necessary to load the new POPE unit into
LEaP. This file contains a definition of the POPE unit including
271
connectivity, charges, and atom type information. Importantly,

the atom types in this file (column 3) are all listed in lower case;
this is intended to distinguish from the atom types to be evaluated using traditional AMBER force filed parameters, which are
defined in upper case.
The GAFF is designed to be extensive and therefore include
parameters for a large combination of parameters. It is possible,
however, that a new unit defined in Antechamber may contain a
combination of atom types that have not been parameterized.
Inorder to check for this the parmchk executable is called:
$AMBERHOME/bin/parmchk -i pope.prepin -f prepi
-o pope.frcmod
This command compares the parameters that are necessary for
the new POPE unit to those that are available in the GAFF.Should
any parameter be missing antechamber will either empirically calculate the parameter or find a similar, analogous parameter; the
results of which are then listed in the .frcmod file. Should it be
impossible to define parameters in this way, Antechamber will add
a place holder and a comment indicating that revision is required,
in which case other methods of parameterization should be pursued. It is important to carefully assess the rational that is used for
any parameters obtained in the .frcmod file. The pope.frcmod
file obtained here is shown below:
remark goes here
MASS
BOND
ANGLE
DIHE
IMPROPER
c3-o -c -os 10.5 180.0 2.0 General improper
torsional angle (2 general atom types)
c2-c3-c2-ha 1.1 180.0 2.0 Using default value
NONBON
3.9 Solvate
theSystem, Add
Missing Atoms,
andGenerate Input
Files forProduction
Calculations
Prior to performing production calculations such as energy minimizations or MD simulations in AMBER, it is necessary to define:
(1) a prmtop file that contains the required force field parameters and a description of the molecular topology and (2) an
inpcrd file that contains the atomic coordinates. An inpcrd file
can also contain atomic velocities and periodic box information if
defined. Here, AMBERs LEaP program will be used to generate
the files necessary for production calculations.
In order to realistically model protein behavior, it is furthermore necessary to solvate the system with water molecules and
ions. Any missing atoms including hydrogen atoms and N- and
C-terminus specific atoms which have not previously been included
can be added at this point. Conveniently, LEaP will perform all of
272
these tasks. The molecular coordinates and parameter information

developed as described above will be used as a starting point.
Here, tleap, the command line version of LEaP will be used.
The script provided below will read in the coordinate and parameter information generated thus far and define a periodic box.
Next, water molecules will be added, maintaining a 1.5 or greater
distance from any existing solute and forming a box that extends
12 in the Z axis from any existing solute. Sodium and chloride
atoms will then be added within the box surrounding solute.
The Coulombic potential on a grid will be used to determine
placement and should steric conflict occur with a water molecule,
the water molecule will be removed and the ion will be added in its
place. The number of sodium and chloride ions are selected to mimic
physiological ion concentrations and also to meet the requirement
of an overall neutral system. Finally, the resulting system is evaluated
for errors and a PDB file as well as the necessary prmtop and
inpcrd files are output for the system as a whole.
# Load force field parameters
source leaprc.gaff
source leaprc.ff99SB
# Load Antechamber files- from 3.8
loadAmberPrep pope.prepin
loadAmberParams pope.frcmod
#
Load
positioned
binding
pocket
ions- from3.5
bpions = loadpdb "bp_ions_positioned.pdb"
# Load positioned LeuT with missing residues- from 3.6
leut = loadpdb "leut_positioned_allres.pdb"
# Load lipids and lipid solvation- from 3.7
pope = loadpdb "pope.pdb"
popewater = loadpdb "pope_water.pdb"
# Combine into one unit
system
=
combine
{bpions
leut
pope
popewater}
# Set periodic box
setBox system vdw
# Solvate the system along the z axis; maintain a distance of 1.5 A from solute
solvateBox system TIP3PBOX {0 0 12} 1.5
# Add ions to produce a neutral system
addIons system Na+ 56 Cl- 60
# Examine the final unit
check system
charge system
desc system
# Output files
savePdb system leut_system.pdb
saveamberparm
system
leut_system.inpcrd
quit
273
leut_system.prmtop
With all the necessary files in the current working directory

and the above script saved to a text file called leap.in, the following command is used to execute the script:
tleap f leap.in
3.10 Perform Energy
Minimizations
Next energy minimizations are performed in order to relax the

system. This is necessary to eliminate any energetically unfavorable
interactions that may occur in the process of building the system.
For example, hydrogen atoms added as described above are positioned according to a predefined geometry and therefore the
resulting coordinates may have steric conflicts or overlap with
other residues. The minimization steps eliminate such problems,
providing an energetically favorable, more stable starting point for
molecular dynamics simulation.
Using the prmtop and inpcrd files generated in
Subheading 3.9, AMBERs sander can be used to perform the
energy minimization. The following minimize.in file is used to
direct sander to perform an energy minimization:
Minimization of solvent
&cntrl
imin
= 1,
maxcyc = 1000,
ncyc
= 400,
ntb
= 1,
ntr
= 1,
cut
= 12
/
Hold protein and lipids fixed
250.0
RES 1 636
END
END
The &cntrl command must begin with a leading black space
and specifies the type of namelist that follows. The minimization
starts at step 1 (imin=1) and continues for 1,000 steps (maxycy=1,000). The first 400 steps utilizes the steepest decent algorithm, appropriate for quickly reducing the largest strain, and the
remaining 600 steps utilizes the conjugate gradient algorithm
which is more appropriate for converging to a minima (ncyc=400).
Constant volume periodic boundary conditions will be used
(ntb=1). Position restraints will be used (ntr=1) using a force
constant of 250kcal/mol/2 to restrain the lipid and protein residues (residue numbers 1636).
The following command will execute the energy minimization:
274
$AMBERHOME/bin/sander -O -i minimize.in -o
minimize.out c leut_system.inpcrd -p leut_system.prmtop -r leut_system_min.rst
Where -O specifies that output files should be overwrite any
existing files, -I specifies the name of the mdin file, -o specifies the name of the output file, -c specifies the starting inpcrd
file, -p specifies the prmtop file, and -r specifies the name of
the final coordinates following minimization.
Beginning with the rst file from the previous minimization,
the following mdin file can be used to minimize the lipids as well:
Minimization of solvent and lipids
&cntrl
imin
= 1,
maxcyc = 1000,
ncyc
= 400,
ntb
= 1,
ntr
= 1,
cut
= 12
/
Hold protein fixed
250.0
RES 1 511
END
END
Finally, all atoms in the simulation are minimized with the following mdin file:
Minimization of all atoms
&cntrl
imin
= 1,
maxcyc = 1000,
ncyc
= 400,
ntb
= 1,
ntr
= 0,
cut
= 12
/
3.11 Heat
theSystem
andPerform anInitial
cMD Equilibration
Following energy minimization the system is ready for cMD equilibration. The system temperature is linearly increased from 0 to
310K (tempi=0.0; temp0=310.0) in order to prevent excessive
and sudden solute fluctuations. A weak restraint is placed on the
protein and lipid residues to further aid in this regard. A Langevin
temperature equilibration scheme (ntt=3) will be used to equalize
and maintain the system temperature using a collision frequency of
1.0ps1 (gamma_ln=1.0).
While equilibration will ultimately be performed using constant
pressure and temperature parameters, as the system is heating up the
pressure that is calculated can be inaccurate, leading to issues if constant pressure parameters are employed. The use of restraints with
275
constant pressure parameters further causes problems. Accordingly,

the system is initially equilibrated at constant volume (ntb=1) and
then later transitioned to constant pressure parameters. Furthermore,
since the hydrogen atom motion in solute is unlikely to affect the
overall protein dynamics, a SHAKE algorithm is used which fixes all
bonds involving hydrogen (ntc=2; ntf=2), reducing computational
burden and allowing the time step to be increased to 2fs without
introducing any instability (if hydrogen atoms were allowed to oscillate the would provide the highest frequency oscillation in the system and therefore determine the maximum time step).
The following heat.in script can be used to heat the system:
cMD heating of the system with restraints on
protein and lipid residues
&cntrl
imin = 0, irest = 0, ntx = 1,
ntb = 1,
cut = 12.0,
ntr = 1,
ntc = 2, ntf
= 2,
tempi = 0.0, temp0 = 310.0,
ntt = 3, gamma_ln = 1.0,
nstlim = 20000, dt = 0.002
ntpr = 2000, ntwx = 2000, ntwr = 2000
/
Keep protein and lipids fixed with weak
restraints
10.0
RES 1 636
END
END
In the heat.in script above the minimization is turned off
(imin=0) and initial velocities are randomly assigned from a
Boltzmann distribution (irest=0; ntx=1). A cutoff of 12 is used
(cut=12). 20,000cMD steps will be performed (nstlim=20,000)
using a time step of 2fs per step (dt=0.002), resulting in a total
simulation time of 40ps. The output file (ntpr), trajectory file
(ntwx) and restart file (ntwr) will be written to every 2,000 steps.
Finally, weak position restraint (ntr=1) using a force constant of
10kcal/mol/2 will be used to minimize lipid and protein residues
movement (residue numbers 1636). The following command will
initiate heating of the system:
$AMBERHOME/bin/sander -O -i heat.in -o heat.
out c leut_system_min3.rst -p leut_system.
prmtop -r leut_system_heated.rst
After heating the system, a short cMD equilibration using
constant temperature and pressure parameters in necessary to
276
obtain information about the system that will be required for aMD
production runs.
The following cMD.in script can be used to accomplish this:
cMD equilibration to obtain parameters necessary for aMD production runs
&cntrl
imin = 0, irest = 1, ntx = 7,
ntb = 2, pres0 = 1.0, ntp = 2, taup = 2.0,
cut = 12.0,
ntr = 0,
ntc = 2, ntf = 2,
tempi = 310.0, temp0 = 310.0,
nstlim = 500000, dt = 0.002
ntpr = 2000, ntwx = 2000, ntwr = 2000
/
In the cMD.in script above, the parameters that have
changed in comparison to the heat.in script are defined as follows. Since the simulation is being restarted, the time step is read
in from the previous run (irest=1) and the coordinates being read
in are in ASCIII restart format (ntx=7). Anisotropic pressure scaling will be used (ntp=2) to maintain a constant pressure (ntb=2)
with an average pressure of 1atm (pres0=1.0) and a relaxation
time of 2ps (taup=2.0). No position restraints are used (ntr=0).
One nanosecond of simulation time will be obtained
(nstlim=500000, dt=0.002). The following command can be
used to execute the script:
$AMBERHOME/bin/sander -O -i cMD.in -o cMD.
out c leut_system_heated.rst -p leut_system.
prmtop -r leut_system_cMD_equil.rst
3.12 Perform aMD
ProductionRuns
The cMD equilibrated system and data obtained from the cMD
equilibration can now be used for aMD production runs. aMD
production runs are conducted in a very similar fashion to the
cMD equilibration described in Subheading3.11, with the only
difference being the definition of additional aMD specific
parameters. The following aMD.in script can be used for
aMD production runs:
aMD production run
&cntrl
imin=0, irest=1, ntx=7,
ntb = 2, pres0 = 1.0, ntp = 2, taup = 2.0,
cut = 12.0,
ntr = 0,
ntc = 2, ntf = 2,
tempi = 310.0, temp0 = 310.0,
277
nstlim = 500000, dt = 0.002

ntpr = 2000, ntwx = 2000, ntwr = 2000
iamd = 3,
alphaD = 357.7,
EthreshD = 12988.9,
alphaP= 10494.2,
EthreshP = -94241.8
/
In the aMD.in script above, the simulation parameters are
the same as in the previously used cMD.in with the addition of
the following aMD specific parameters. The aMD implementation
in AMBER allows for the possibility of boosting the torsional
terms of the potential only (iamd
=
2), the whole potential
(iamd=1), or the whole potential with an additional boost to the
torsions (iamd=3). alphaD is defined as the product of 0.2,
3.5kcal/mol/residue, and the number of protein residues (511
residues are in the LeuT example here). EthreshD is defined as the
product of 3.5kcal/mol/residue and the number of protein residues, added to the average dihedral based on cMD equilibration
(11,200.4 for the system in this example). alphaP is defined as the
product of 0.2 and the total atoms (52,471 total atoms in this
example). EthreshP is defined as the sum of alphaP and the average EPtot from cMD equilibration (104,736 for the system in
this example).
The following command can be used to execute the aMD.in
script:
$AMBERHOME/bin/sander -O -i aMD.in -o aMD_01.
out c leut_system_cMD_equil.rst -p leut_system.prmtop -r leut_system_aMD_01l.rst
Several additional aMD production runs can be performed,
each time beginning with the restart coordinates output from the
previous run.
3.13 Prepare Data
forPCA Analysis
When sufficient aMD production runs have completed, principal

component analysis is used to analyze the results from the aMD simulations. The statistical program R is a platform for performing statistical calculations with large data sets. Through the use of the Bio3D
add on package, R is able to perform analysis of PDB and DCD files.
To start using the Bio3D package with R to analyze the data, the data
must be formatted into a Bio3D useable style. The first thing to note
is that Bio3D works best with protein only data, so any other molecules will need to be removed (lipids, waters, ions, etc.). The second
note is that Bio3D does not utilize parameter-topology files from
AMBER (PRMTOP) or protein structure files from NAMD (PSF);
instead, Bio3D uses information from the PDB file to determine the
protein structure. The final note on Bio3D is that it requires DCD
trajectories for processing (the default binary for NAMD), so AMBER
trajectories will need to be converted.
278
1. Use VMD and load in the structure file (PSF or PRMTOP):

FileNew Molecule then load in the PRMTOP file. VMD will
recognize that it is an AMBER topology file but not whether
or not it is one with a periodic boundary conditions. For our
LeuT simulations, and probably under most circumstances, it
will be necessary to change the file type to AMBER coordinates
with Periodic Box.
2. Load the trajectories into the created molecule: Right click the
molecule and select Load data into molecule. Select the trajectory file and load in. Depending on the size of the trajectory
and available computer resources, this may need to be broken
into smaller steps or trajectory frames will need to be skipped.
3.
Change the visualization to show the protein only:
GraphicsRepresentations. Select the molecule and change the
atom selection to the desired atoms. Replacing the selection all
with the phrase protein will usually be sufficient, however this may
change depending on the configuration of the topology file.
4. Save a PDB file of the first frame of the trajectory: Right click
the molecule from the main window and click Save Coordinates.
Change the type to PDB and change the first and last frame to 0.
Use the drop down menu to select the atom selection specified
in step 3 above. This will be the PDB Bio3D will use to determine the protein structure.
5. Create a DCD file of the trajectory frames: Follow the steps
above to Save coordinates, but change type to DCD and make
sure the first frame is 0 and the last frame is the last frame
loaded into the system.
3.14 Using Bio3D
andR toCalculate
thePCA
In an R terminal, load in the Bio3D library (ensure the package has

previously been installed). Read in the PDB file of the first frame as
the structure and then the trajectory from the DCD file.

library(bio3d)
pdb <-read.pdb("LeuT_frame1.pdb")
dcd <- read.dcd("LeuT_trj.dcd")
Select the atoms of interest for the PCA analysis. The atom.select
command pulls the indices of atoms which correspond to the atom
selection. For our analysis, we choose to use all alpha carbon positions of select residues. The elety parameter of the atom.select
command specifies the atom type (CA=alpha carbon), and the
resno parameter can be used to select residues by number.

ca.ind <- atom.select(pdb, elety="CA")
In order to eliminate translations and rotations the trajectory

structures will need to be fitted and superposed to the PBD structure. The fit.xyz command can be used to accomplish this. The
fit.xyz command takes two necessary input parametersthe structure to use as a reference and the structures to fit to the reference.
Itshould be noted that the fit.xyz command only takes coordinates
279
as input, so the $xyz object of the PDB will need to be used.

The bracket argument pulls the atom selections from above as the
coordinates to fit. The output of the fit.xyz command will be a
superposed trajectory of the selected atoms only.
trj.fit
<dcd[,ca.ind$xyz])
fit.xyz(pdb$xyz[ca.ind$xyz],
The PCA of the coordinates can be taken based on the trajectory with the pca.xyz command.
trj.pca <- pca.xyz(trj.fit)
With the Bio3D package installed, the plot command has
been overloaded to create a default PCA plot with four graphs.
Three are the z-scores of the first three principal components
plotted against each other in two dimensions. The last is a scree plot
representing how much of the variance of the data set is captured
by each principal component (Fig.5).
plot(trj.pca)
The plot points can be computationally clustered and colored
by cluster. This can be done by creating a distance matrix of the
principal components of interest (principal components 1, 2, and 3
were used in this analysis).
d <- dist(trj.pca$z[,1:3])
A dendrogram of the distance matrix can be calculated and
plotted for visualization. The plot of the dendrogram can be used
to determine the number of clusters desired from the analysis.
hc <- hclust(d)
plot(hc)
The cutree command can be used to create a color vector which
will color each point based on the number of groups desired. It
takes two arguments, the dendrogram and k which is the desired
number of clusters. Here, the LeuT data appeared to fall into seven
clusters. The PCA plots can be colored by replotting the PCA and
using the output from the cutree command as an argument to the
col parameter (Fig.6). For structure analysis, representative structures from each cluster can be isolated and analyzed in molecular
visualization software of choice like VMD.
grps <- cutree(hc,k=7)
plot(trj.pca,col=grps)
3.15 Validation
andUse ofthePCA
PCA of aMD data is useful for finding protein segments that are
most involved in structural changes. The RMSF (root mean square
fluctuations) calculation can be used to determine how much each
residue moves during the trajectory.
rf <- rmsf(trj.fit)
280
Fig. 5 Bio3D plot of principal component data
The pca.xyz command calculated a matrix with as many principal

components as number of data frames entered into it, but the question of how many principal components to use remains. One tool
for determining how many to use is the scree plot which shows how
much of the data is captured by each principal component, but for
the LeuT analysis described here, less than 50% of the variance is
captured after the three most influential principal components. The
RMSF can also be utilized in order to determine whether all of the
captured fluctuations within those first three components is a sufficient representation of the structural changes involved.
The RMSF can be used to validate the principal components
by overlaying the RMSF with the $au object from the PCA calculation. The $au object is the atomic vector which shows how much
each residue contributes to a principal component. The goal is to
281
Fig. 6 Bio3D plot of PCA data colored by cluster after calculation of cluster groups
obtain a graph where each residues RMSF value is captured by at

least one of the selected principal components. If a residue shows
on the RMSF but is not captured in one of the $au objects, then
another principal component will need to be used (Fig.7).
barplot(rf,col="purple",border="purple")
par(new=TRUE)
plot(trj.pca$au[,1],type="l",col="orange",
lwd=3)
points(trj.pca$au[,2],type="l",col="green",
lwd=3)
points(trj.pca$au[,3],type="l",col="skyblue
",lwd=3)
282
Fig. 7 Root mean square fluctuations graph superposed with the $au vectors
representing the amount of variance captured by each principal component. This
graph has all residues included, and the carboxy terminus appears to dominate
the principal component
legend.labels = c("RMSF", "PC 1", "PC 2", "PC3")

legend.cols = c("purple", "orange", "green",
"skyblue")
legend("topleft",legend.labels, fill = legend.cols)
Looking at Fig.7 from the above example, sometimes a single residue or group of residues may dominate the entire PCA and prevent
the PCA from capturing other residue fluctuations. In the LeuT example here, it appears that the last few residues of TM12 at the carboxy
terminus of the protein contribute a large portion of the PC.However,
these residues mostly just vibrate in the cytosol, and we can therefore
exclude these residues from the analysis and rerun the PCA.This can
be done by editing the indices used with the atom.select command.
Then check the $au object with the RMSF again (Fig.8).
ind.adj
<atom.select(pdb,elety="CA",re
sno=1:505)
trj.fit.adj <- fit.xyz(pdb$xyz[ind.adj$xyz],
dcd[,ind.adj$xyz])
trj.pca.adj <- pca.xyz(trj.fit.adj)
rf.adj <- rmsf(trj.fit.adj)
283
Fig. 8 RMSF graph superposed with the $au vectors of the principal components
after adjusting the PCA by removing residues 506512
barplot(rf.adj,col="purple",border="purple"
,main="Adjusted")
par(new=TRUE)
plot(trj.pca.adj$au[,1],type="l",col="orang
e",lwd=3)
points(trj.pca.adj$au[,2],type="l",col="gre
en",lwd=3)
points(trj.pca.adj$au[,3],type="l",col="sky
blue",lwd=3)
Once a suitable RMSF plot has been obtained, the PCA can be
reclustered and replotted to visualize the new PCA result (Fig.9).
d <- dist(trj.pca.adj$z[,1:3])
hc <- hclust(d)
grps <- cutree(hc,k=7)
plot(trj.pca.adj,col=grps)
The fluctuations captured by each principal component can
also be visualized as a trajectory in VMD.Use the mktrj.pca command to generate a VMD trajectory for visualizing captured
fluctuations in each principle component desired. Load the PDB
file into VMD and change the representation to Trace or Tube
since the PCA only printed out the alpha carbons.
284
Fig. 9 Adjusted PCA plot of LeuT with residues 506512 removed and each point colored by cluster
mktrj.pca(trj.pca,pc=1,mag=1, file="PC1.pdb")
As mentioned above in the background, it may be necessary to
reweigh the PCA plots to determine the validity of each point. The
following steps are necessary in order to accomplish this:
1. Extract the V[r(ti)] for each structural point from the simulation log files using text manipulation tools like grep and awk.
2. Load these values into R.
3. Calculate e
bDV r (t i )
for the vector.
4. Plot any two desired principal components against each other

bDV r (t )
and generate a contour plot using the e i vector as a
coloring parameter (or any other desired method of plotting
three dimensional data in R).
3.16 Compare
theResulting
Conformations
withAdditional
Structural Data
285
Additional structural coordinates, such as various crystallized LeuT

structures thought to be representative of portions of the substrate
transport cycle, can be projected into the PCA space using the homology functionality within Bio3D.To start, read the structures to be
projected onto the PCA into R and align to the referencePDB.
crys <- pdbaln(c(pdb,other pdb files))
The alignment of the crystal files and the alignment of the trajectory files must match. Unfortunately, since the alignment
changes the index values of the reference PDB to match the
sequence, it is necessary to match the index values of the aligned
sequence to the trajectory. There is a function (pdb2aln) to do this
planned for new releases of Bio3D; however, at the time of this
writing, this function is not yet available in the general Bio3D
package.
Gap inspection (for missing residues) may be necessary using
the gap.inspect command. In this case, the PCA may need to be
recalculated if residues previously used have gaps. One way to
avoid having to recalculate the PCA is to do the PCA of the crystals alone and project the trajectory into the crystal PCA space.
The following example will assume projecting the crystals onto
the trajectory PCA space.
1. Run a fit to the reference PDB:
crys.fit <- fit.xyz(crys$xyz[1, any index values
such as gap positions or specific residues], crys
$xyz[2:length(crys$xyz[,1]),same index values])
2. Project the crystals into the trajectory PCA space:
crys.pca <- pca.project(crys.fit, trj.pca)
plot(trj.pca)
points(crys.pca[,1],crys.pcs[,2], col= "black")
References
Science 334(6055):517520. doi:10.1126/
science.1208351
2. Shaw DE, Maragakis P, Lindorff-Larsen K,
Piana S, Dror RO, Eastwood MP, Bank JA,
Jumper JM, Salmon JK, Shan Y, Wriggers W
(2010) Atomic-level characterization of the
structural dynamics of proteins. Science
330(6002):341346.
doi:10.1126/
science.1187409
3. Genchev GZ, Kallberg M, Gursoy G, Mittal A,
Dubey L, Perisic O, Feng G, Langlois R, Lu H
(2009) Mechanical signaling on the single protein level studied using steered molecular
dynamics. Cell Biochem Biophys 55(3):141
152. doi:10.1007/s12013-009-9064-5
4. Baker JL, Biais N, Tama F (2013) Steered
molecular dynamics simulations of a type IV
pilus probe initial stages of a force-induced
conformational transition. PLoS Comput Biol
9(4):e1003032.
doi:10.1371/journal.
pcbi.1003032
5. Forti F, Boechi L, Estrin DA, Marti MA (2011)
Comparing and combining implicit ligand
sampling with multiple steered molecular
286
dynamics to study ligand migration processes

in heme proteins. J Comput Chem.
doi:10.1002/jcc.21805
6. Li W, Shen J, Liu G, Tang Y, Hoshino T
(2011) Exploring coumarin egress channels in
human cytochrome P450 2A6 by random
acceleration and steered molecular dynamics
simulations.
Proteins
79(1):271281.
doi:10.1002/prot.22880
7. Shen M, Guan J, Xu L, Yu Y, He J, Jones GW,
Song Y (2012) Steered molecular dynamics
simulations on the binding of the appendant
structure and helix-beta2in domain-swapped
human cystatin C dimer. J Biomol Struct Dyn
30(6):652661. doi:10.1080/07391102.201
2.689698
8. Xu L, Hasin N, Shen M, He J, Xue Y, Zhou X,
Perrett S, Song Y, Jones GW (2013) Using
steered molecular dynamics to predict and
assess Hsp70 substrate-binding domain
mutants that alter prion propagation. PLoS
Comput Biol 9(1):e1002896. doi:10.1371/
journal.pcbi.1002896
9. Hamelberg D, Mongan J, McCammon JA
(2004) Accelerated molecular dynamics: a
promising and efficient simulation method for
biomolecules. J Chem Phys 120(24):11919
11929. doi:10.1063/1.1755656

10.
Bucher D, Grant BJ, Markwick PR,
McCammon JA (2011) Accessing a hidden
conformation of the maltose binding protein
using accelerated molecular dynamics. PLoS
Comput Biol 7(4):e1002034. doi:10.1371/
11. Grant BJ, Gorfe AA, McCammon JA (2009)
Ras conformational switching: simulating
nucleotide-dependent conformational transitions with accelerated molecular dynamics.
PLoS
Comput
Biol
5(3):e1000325.
doi:10.1371/journal.pcbi.1000325
12. Mucksch C, Urbassek HM (2013) Enhancing
protein adsorption simulations by using accelerated molecular dynamics. PLoS One 8(6):e64883.
doi:10.1371/journal.pone.0064883
13. de Oliveira CA, Grant BJ, Zhou M, McCammon
JA (2011) Large-scale conformational changes
of Trypanosoma cruzi proline racemase predicted by accelerated molecular dynamics simulation. PLoS Comput Biol 7(10):e1002178.
14. Salmon L, Pierce L, Grimm A, Ortega Roldan
JL, Mollica L, Jensen MR, van Nuland N,
Markwick PR, McCammon JA, Blackledge M
(2012)
Multi-timescale
conformational
dynamics of the SH3 domain of CD2-
associated protein using NMR spectroscopy
and accelerated molecular dynamics. Angew
Chem 51(25):61036106. doi:10.1002/

anie.201202026
15. Thomas JR, Gedeon PC, Grant BJ, Madura JD
(2012) LeuT conformational sampling utilizing
accelerated molecular dynamics and principal
component analysis. Biophys J 103(1):L1L3.
doi:10.1016/j.bpj.2012.05.002
16. Phillips JC, Braun R, Wang W, Gumbart J,
Tajkhorshid E, Villa E, Chipot C, Skeel RD,
Kale L, Schulten K (2005) Scalable molecular
dynamics with NAMD.J Comput Chem
26(16):17811802. doi:10.1002/jcc.20289

17.
Wang Y, Harrison CB, Schulten K,
McCammon JA (2011) Implementation of
accelerated
molecular
dynamics
in
NAMD.Comput Sci Discov 4(1):pii: 015002,
doi:10.1088/1749-4699/4/1/015002

18.
Case DA, Darden TA, Cheatham TE,
Simmerling CL, Wang J, Duke RE, Luo R,
Walker RC, Zhang W, Merz KM, Roberts B,
Hayik S, Roitberg A, Seabra G, Swails J, Goetz
AW, Kolossvry I, Wong KF, Paesani F, Vanicek
J, Wolf RM, Liu J, Wu X, Brozell SR,
Steinbrecher T, Gohlke H, Cai Q, Ye X, Wang
J, Hsieh MJ, Cui G, Roe DR, Mathews DH,
Seetin MG, Salomon-Ferrer R, Sagui C, Babin
V, Luchko T, Gusarov S, Kovalenko A, Kollman
PA (2012) AMBER 12. University of
California, San Francisco
19. Pierce LC, Salomon-Ferrer R, Augusto FOC,
McCammon JA, Walker RC (2012) Routine
access to millisecond time scale events with
accelerated molecular dynamics. J Chem
Theory
Comput
8(9):29973002.
doi:10.1021/ct300284c
20. Voter AF (1997) Hyperdynamics: accelerated
molecular dynamics of infrequent events. Phys
Rev Lett 78(20):39083911
21. Voter AF (1997) A method for accelerating the
molecular dynamics simulation of infrequent
events. J Chem Phys 106(11):46654677.
doi:10.1063/1.473503
22. Steiner MM, Genilloud PA, Wilkins JW (1998)
Simple bias potential for boosting molecular
dynamics with the hyperdynamics scheme.
Phys Rev B 57(17):1023610239
23. Rahman JA, Tully JC (2002) Puddle-skimming:
an efficient sampling of multidimensional configuration space. J Chem Phys 116(20):8750
8760. doi:10.1063/1.1469605
24. Shen T, Hamelberg D (2008) A statistical analysis of the precision of reweighting-based simulations. J Chem Phys 129(3):034103.
doi:10.1063/1.2944250
25. Xin Y, Doshi U, Hamelberg D (2010) Examining
the limits of time reweighting and Kramers rate
theory to obtain correct kinetics from accelerated

molecular
dynamics.
J
Chem
Phys
132(22):224101. doi:10.1063/1.3432761
26. Gedeon PC, Indarte M, Surratt CK, Madura
JD (2010) Molecular dynamics of leucine and
dopamine transporter proteins in a model cell
membrane lipid bilayer. Proteins 78(4):797
811. doi:10.1002/prot.22601

27. Pettersen EF, Goddard TD, Huang CC,
Couch GS, Greenblatt DM, Meng EC, Ferrin
TE (2004) UCSF Chimeraa visualization
system for exploratory research and analysis.
JComput
Chem
25(13):16051612.
doi:10.1002/jcc.20084
28. Sali A, Blundell TL (1993) Comparative protein modelling by satisfaction of spatial
restraints. J Mol Biol 234(3):779815.
doi:10.1006/jmbi.1993.1626
VMD: visual molecular dynamics. J Mol Graph
14(1):3338, 27-38
287
30.

Grant BJ, Rodrigues AP, ElSawy KM,
McCammon JA, Caves LS (2006) Bio3d: an
R package for the comparative analysis of
protein
structures.
Bioinformatics
22(21):26952696. doi:10.1093/bioinformatics/btl461
31. Yamashita A, Singh SK, Kawate T, Jin Y,

Gouaux E (2005) Crystal structure of a bacterial homologue of Na+/Cl-dependent
neurotransmitter
transporters.
Nature
437(7056):215223.
doi:10.1038/
nature03978
32. Shen MY, Sali A (2006) Statistical potential
for assessment and prediction of protein
structures.
Protein
15(11):25072524.
doi:10.1110/ps.062416606
33. Aksimentiev A, Sotomayor M, Wells D (2012)
Membrane proteins tutorial. Theoretical and
Computational Biophysics Group, University
of Illinois at Urbana-Champaign, Champaign
Chapter 13
Simulations andExperiments inProtein Folding
GiovanniSettanni
Abstract
The interplay between simulations and experiments of protein folding has largely contributed to the
elucidation of many important aspects of the phenomenon. In this chapter, I briefly describe the experiments which provide information on the kinetics of the protein folding process, and help to characterize
the folding transition state. Then, I show how to probe the kinetics of protein folding using molecular
dynamics simulations, how to compare the simulations with the experiments and how to help and rationalize
the latter, ultimately offering a molecular picture of the process. After the production of suitable molecular
dynamics simulation data in the form of trajectories, the procedure involves sequentially the identification
of the stable states of the protein, the identification of the transition pathways connecting the stable states,
the identification of the transition state conformations, comparison with experimental results, and finally, the
identification of the molecular determinants or reaction coordinates of the folding process, that is, the
features that clearly help distinguishing the transition state from the stable states.
Key words Kinetics, Transition state, Committor, Phi-value, Clustering, Kinetic network
1 Introduction
Since its discovery [1], the phenomenon of reversible folding of
proteins has engaged generations of scientists from many different
fields. After decades of research the emerging picture often used
to describe the protein folding process, at least in its simplest declination, is that of a first order phase transition between the native
state of the protein, and the unfolded/denatured state, the former
being enthalpically stabilized by favorable interactions internal to
the protein chain and with the solvent, the latter being entropically
stabilized by the large number of conformations that the protein
chain can sample within this state.
This rationalization of the protein folding process, led to more
focused attempts to identify the determinants of the process.
Indeed, a first order phase transition implies the crossing of a free
energy barrier, which, then, represents a bottleneck in the transition
process. The height of the free energy barrier, which can be measured by experiments on the folding kinetics, and the way it changes
289
290
Giovanni Settanni
depending on the various components of the protein-solution system

(protein sequence, protein structure, environmental conditions,
etc.) has been used to assess which elements play a major role
and how that happens. In this framework a special role is played by
phi-value analysis.
In phi-value analysis [2, 3], the structure of a given protein is
perturbed by the introduction of conservative point mutations along
the sequence and the effect of the perturbation on the folding free
energy barrier is assessed by a combination of equilibrium and kinetic
experiments. The equilibrium experiments are used to measure
with high accuracy the stability of wild-type and mutant protein.
The kinetics experiments are used to assess the folding rate, which is
related to the height of the free energy barrier through an Arrhenius
relation kobsexpGU/kbT, where kobs is the observed folding
rate, GU is the height of the free energy barrier for folding, i.e.,
the free energy difference between the transition and unfolded state,
kb is the Boltzmann constant, T the temperature. The folding phivalue of the mutated amino acid is thus the ratio between the change
in height of the free energy barrier for folding and the change of the
protein stability upon mutation.
Ff =

WT
DGmut
- U - DG - U
DG
mut
F -U
- DG
WT
F -U
DDG-U
DDGF -U
k mut
f
= ln WT
k
f
/ DDGF -U
(1)
where GFU is the folding free energy, i.e., the free energy difference between folded and unfolded state (see Note 1). In the classical interpretation, phi-values close to one mean that the mutation
has produced the same effect on the free energy of the transition
and folded state (the free energy of the unfolded state is used as
reference), thus, the conformation of the mutated residue must be
similar in the two states. On the other hand, a phi-value close to
zero means that the mutation affected mostly the free energy of
the folded state and not the transition state, thus the conformation
of the residue in the transition state must be similar to its unfolded
state. Intermediate and non-classical (<0 or >1) phi-values
cannot be interpreted straightforwardly and each case requires
specific attention.
Systematic measurements of phi-values of many amino acids,
sometime termed alanine/glycine scans, are being made available
for many proteins. These measurements help to identify either
which amino acids take part in the formation of a folding nucleus
around which the overall folding process takes place (the amino
acids with large ), or the presence of a diffuse transition state,
where there are no amino acids with large . Early results [4] have
shown that in many cases it is possible to correlate the phi-value of
an amino acid to the fraction of native atomic contacts formed at
the transition states. This relationship offers both a way to test
simulations, and to interpret simulation data where experimental
data are not available [5].
291
Simulations allow for the interpretation of phi-value data in

terms of molecular structures of the various states of the protein.
Indeed, simulations, after being validated with experimental data,
allow, for example, for verifying which interactions are present in
the transition state vs. the native state or for investigating the presence of residual structure in the unfolded state, that is, interactions
that are present also in the unfolded state [59]. Ultimately, simulations can be used to find out which structural observables represent good reaction coordinates for the folding process, in other
words, the observables able to distinguish the transition state from
the native and from the unfolded states [10]. The simulation
approach, however, is hampered by the elusive nature of the transition state, which, being a saddle point in the free energy landscape
of the protein, with considerably higher free energy than the stable
states, is not normally densely populated and/or easy to sample.
Although nowadays there exist specialized hardware capable of
simulating classical atomistic models of relatively small proteins
with the surrounding solvent molecules up to the millisecond time
scale [11] (a fully quantum-mechanical treatment of the folding
dynamics of proteins is at present unthinkable), the time scales of
the folding process of proteins, sometimes reaching hours, are often
too large for such an approach on standard high performance hardware (i.e., linux clusters, gpu workstations). For this reason, a large
variety of methods have been devised to overcome this barrier. One
possible approach is the use of simplified coarse-grained force fields,
where interactions between atoms are substituted by a potential
function designed to have a deep minimum corresponding to the
native state of the protein. These models, referred to generally as
G models [12], have been shown to reproduce some important
characteristics of the folding process of proteins [7, 8, 1315].
These characteristics depend for a large part on the topology of the
protein, i.e., the way each amino acid interacts with the other amino
acids along the sequence. The scope of these models, however, does
not allow for addressing interesting questions about non-native
interactions, which are excluded a priori, or the specific contribution
to the interactions due to the chemical nature of each group of
atoms, which is not preserved in these models. For this reason I will
focus on classical atomistic models of proteins and peptides and in
particular on the use of these models in molecular dynamics simulations of the folding/unfolding process, aimed to reveal the structure of the transition state ensemble.
In what follows, I give an overview of the steps necessary to
(a)produce a molecular dynamics simulation suitable to study the
folding process of a protein system; (b) identify the stable states,
and transition events; (c) identify the transition states; (d) compare
with experimental data and find good reaction coordinates for
folding. These various steps will be exemplified by means of two
test cases.
292
Giovanni Settanni
2 Materials
All the simulations and analysis are performed on Linux workstations/clusters. In case study 1 simulations are performed using the
program GROMACS [16]. In case study 2 CHARMM [17] is used
to run the simulations. In either case, analysis of the trajectories is
performed using the program WORDOM [18]. Shell scripts and
AWK scripts are used for data analysis. The programs GNUPLOT
(http://www.gnuplot.info) and GRACE (http://plasma-gate.weizmann.ac.il/Grace/) are used for plotting graphs. The program
VMD [19] is used for visualization of protein structures and trajectories. The native conformation of the simulated Trpz1 peptide is
obtained from the Protein Data Bank, with pdbid 1LE0 [20].
3 Methods
3.1 Models, Force
Fields, Simulations
Techniques
The aim of our simulations is to collect as many folding/unfolding

transition events as possible, which provide the largest part of the
information we need to characterize the transition state. These
events are rare and may occur on time scales that cannot be sampled
by standard molecular dynamics simulations at constant temperature. In our two test cases we have adopted several expedients to
overcome this problem.
Case study 1: [10] High temperature is used to speed up the unfolding process (see Note 2). The Trpz1, a 12-residue peptide, is simulated at high temperature using a standard force field OPLSAA
[21]. Several unfolding simulations are collected at each temperature starting from the experimental native state with different initial velocities. Each of the simulations must be long enough to
show at least an unfolding event. In the case of Trpz1, which is a
very small peptide, this occurs within 100ns at most of the simulated temperatures (200ns at 375K). Other simulations parameters follow. The peptide is immersed in a periodic cubic simulation
box leaving at least 8 between the peptide and the box boundaries. The cutoff for van der Waal and electrostatic interactions is set
to 10 . The box is filled with 1666 SPC water molecules [22].
Simulations are run at several temperatures between 375 and
450K and pressure is set to 1atm (see Note 3). Temperature and
pressure are regulated by the Berendsen algorithm [23]. The time
step is set to 2fs and all covalent bonds are kept rigid using the
LINCS algorithm [24]. Snapshots of the coordinates of the system
are saved every 20ps into trajectory files.
Case study 2: [25] The effects of the aqueous solvent are modeled
by a mean-field approximation, avoiding the explicit representation of solvent molecules in the simulations. The same peptide
Trpz1 is simulated using the charmm19 [17] united atom force
293
field (i.e., only polar hydrogen atoms are explicitly represented)

close to the transition temperature of the peptide (330K). The
SASA model is adopted [26] as a mean-field approximation of the
solvent. The lack of friction due to the absence of explicit water
molecules helps to speed up the folding/unfolding kinetics of the
peptide without significantly affecting the thermodynamics of the
system [27]. The temperature is kept constant by the Berendsen
thermostat [23]. The SHAKE algorithm [28] is used to fix covalent bond lengths, allowing for a 2fs time step. In about 20s of
cumulated trajectories from ten different simulation runs, it is possible to collect hundreds of folding/unfolding events. The coordinates of the system are saved every 20ps into trajectory files.
In both case studies, the production runs are preceded by a
minimization, heating and equilibration phase. The simulations
are relatively straightforward and do not require specially designed
input files; we refer to Chapter 1 of this book or the CHARMM
and GROMACS online documentation for their preparation.
Ineither case, at the end of the simulations, the resulting trajectory
files (either in .xtc format or in .dcd format, for GROMACS and
CHARMM, respectively) need to be listed along with their complete path in a file called trjlist.txt.
3.2 Identifying
theStable States
oftheSystem
Transition events between the stable states of the system can be identified only after the stable states of the system have been identified.
To this extent, a set of observables relevant for the description of the
protein/peptide are selected and the time series of these observables
are measured along the trajectories. In the case of the Trpz1 peptide,
the root mean square deviation (RMSD) of the C atoms from the
native conformation and the number of native hydrogen bonds
formed along the backbone (HB) represent relevant observables.
Many ways are available to perform these measurements. I will show
how to use the program WORDOM (http://wordom.sourceforge.
net/), which is particularly fast and easy to handle. WORDOM analysis can be invoked using the following syntax:
where the wordom input file .wrd contains the following for
RMSD calculations:
or the following for HB calculations:
294
Giovanni Settanni
In the output of these calculations, the time step along the

trajectory and the observable of the corresponding conformation
are listed.
A further important analysis that needs to be carried out at this
point is a cluster analysis. In cluster analysis, similar conformations
observed along the trajectory are grouped together into clusters.
This will help to understand which conformations are more populated, and how the trajectory flows through these sets of conformations. The degree of similarity can be measured using several
metrics, e.g., RMSD, distance-RMSD, contact-map difference etc.
Here we adopt the distance-RMSD (dRMSD) of C and C atoms,
that is, for each conformation we compute the distances between
all the C and C atoms and store them in a matrix, then the
dRMSD between two conformations is the Euclidean distance
between the corresponding distance matrices. The clustering is
performed using the leader algorithm [29] (alternatively, many
other algorithms can be used). Briefly, the conformations along the
trajectory are sequentially analyzed. If the analyzed conformation
is within a cutoff distance from any of the cluster centers already
identified, then it is added to that cluster, otherwise it is used to
identify the center of a new cluster. WORDOM can be used for the
clustering, the input file in this case will be:
The cutoff for clustering must be chosen small enough so that

conformations within the same cluster belong to the same free
energy minimum. At the same time the cutoff must not be too
small otherwise we will obtain many clusters containing very few
conformations, and this may negatively affect our subsequent calculations (see Note 4). In the output of this calculation, the time
step of each trajectory snapshot is listed along with its corresponding cluster center.
Depending on the algorithm used for clustering, the volume of
each cluster in conformation space may vary (see Note 5). Assuming
that it fluctuates around a certain fixed value, we expect that when
the trajectory visits an energy-stabilized free energy minimum
(i.e., the native state) the volume of the sampled conformation
space, i.e., the sum of the volume of all the clusters visited so far, will
not increase too much, as the system fluctuates around the same
minimum conformation. On the contrary, when the trajectory will
explore an entropy-stabilized free energy minimum (i.e., the
unfolded state), the sampled conformation space volume will
increase significantly, as it will be very unlikely, given the high dimensionality of the conformation space, that very similar conformations
are sampled repeatedly in such a high-entropy state and it will be
295
Fig. 1 (a) representative time series of the observables HB, RMSD and VIR for case study 1, simulations at
450K.The horizontal lines mark the boundaries of each state as identified from the histograms of the observables. These boundaries mark the starting and ending points of transition pathways (regions shaded in
magenta). (b) Histogram of the observables from case study 1, simulations at 450K. Lines around the populated regions indicate the boundaries of the stable states of the peptide. Adapted from ref. 10 with permission
from Elsevier. (c) Histogram of the observables for case study 2. The boundaries of the native (blue) and
denatured (red) state are shown
also unlikely that we have sampled this state to saturation. The

leader clustering algorithm, then, where new clusters are defined
as the trajectory visits conformations further away from those already
visited, provides a natural way to quantify this. It allows us to define
a further observable linked to the speed of conformational sampling:
the volume increase rate (VIR), i.e., the number of new clusters
observed in the unit time, which is supposedly large in the unfolded
state and small in the native state (see Note 6).
Histograms of the observables allow for the identification of
separate populated regions, which correspond to the stable states
of the peptide (Fig.1b, c). The two dimensional histograms can be
produced by pasting the time series of the observables (Fig.1a) in
a single file and then binning the data in 2D bins (e.g., using an
AWK script). The binned data can be plotted using GNUPLOT in
pm3d mode. Alternatively, 1D histograms can be produced using
the histogram analysis tool in GRACE.In case study 1 (Fig.1b),
three populated states can be identified, the fully native N, the
denatured state D and an intermediate state I.In case study 2, two
broadly populated states have been identified, the native and the
unfolded state (see Note 7). The minor differences between the
stable states in the two case studies are due to the different simulation setups (different temperatures and force field).
It is also convenient to store in separate trajectory files all the snaphots of the original trajectory corresponding to each identified state.
This can be done using WORDOM with the following syntax:
296
Giovanni Settanni
where the listA.F file is just a list of snapshots corresponding to the

state A.The listA.F file can be generated using a simple AWK script
to parse the time series of the observables and select those snapshots that are within the boundaries of the required state. The
output file stateA.dcd is a standard .dcd trajectory file listing all the
snaphots corresponding to the state (obviously, the chronological
information in those trajectory files have no meaning).
3.3 Identifying
Folding/Unfolding
Transition Events
Transition events occur when the trajectory leaves a stable state

and reaches a different stable state. To quantitatively capture these
events, we draw a border line enclosing the most populated part of
each stable state as it occurs in the simulations (Fig.1b, c) and then
we look for events where the trajectory crosses the border of one
state in the outwards direction and the border of another state in
the inward direction (Fig.1a). Is it possible that the chosen observables do not describe the system optimally. In that case fast
recrossing events may be observed, i.e., the system may leave the
region of observables space corresponding to one state, reach the
region belonging to another state and quickly go back in the initial
state. These events are not true transitions, they are due to the
inability of the observables to completely separate the two states.
The use of several independent observables and a conservative
choice of the borders of the states so that only the most populated
regions are enclosed, minimizes the probability to observe
recrossings.
In case study one, where we identified 3 stable states, in principle
we could observe six different transition events, i.e., from N to I,
and back, from I to D and back and from N to D and back. In reality, we observe only four kinds of transitions N to I and back and I
to D and back, although we have very few representatives of the
backward (refolding) transitions. Analysis of the average transition
times shows that the N to I transition is the rate limiting step, for
this reason, in what follows, we will focus on the determination of
the transition state for the N to I transition.
3.4 Identifying
theTransition State
Ensemble
For the identification of the transition state ensemble (TSE) in a

two-state transition it is useful to introduce the concept of committor pc of a conformation, i.e., the probability that the simulations starting from the conformation reach one state before the
other [30]. The TSE is defined as the set of conformations that
are equally committed to end in one state or the other, so their
committor (sometime called folding probability) pc equals approximately 0.5. In other words, simulations starting from conformations of the TSE have a 50% probability to reach one of the two
states before the other. Because of this definition, the committor
represents an optimal reaction coordinate, in the sense that it
allows to distinguish each phase of the transition process, that is,
the reactant state, the product state and the transition state in
297
between. The committor of conformations sampled along our

simulations can be computed by starting many additional simulations from the given conformation and counting the fraction that
reaches one of the states before the other. These calculations are
computationally very expensive as we need at least 20 trial runs for
each conformation to obtain a relatively precise estimate of pc (see
Note 8) and each run needs to be long enough to allow the system to commit to either state. Strategies need to be employed to
reduce as much as possible those calculations.
Strategy 1: The committor is either precisely 0 or 1 for conformations in the unfolded or folded state, respectively. It is different
from those values only for conformations along transition pathways and it has generally a monotonic behavior along these pathways [31]. In case study 1 we can limit the committor calculations
to few conformations selected along the transition pathways and
use a binary search approach to converge to those conformations
with committor ~0.5 (i.e., the TSE). In practice, this approach
requires a Linux shell script and can be solved stepwise:
(a) Select the conformation for committor calculation along transition pathways using a binary search strategy, targeting conformations with committor ~0.5 (i.e., select a conformation at
the midpoint of the trajectory between two conformations one
with pc>0.5 and the other with pc<0.5).
(b) Extract the selected conformation from the trajectory and use
it to start 20 (or more) short (2ns for Trpz1) simulations.
(c) Measure the observables (RMSD, HB, VIR) along the committor simulations and use them do determine which stable
state was reached first.
(d) Evaluate the committor as the fraction of the simulations that
reached one of the stable states first (see Note 9). Go back to (a).
Strategy 2: Instead of computing the committor for single
conformations, we can assume that similar conformations (i.e.,
those with a small pairwise dRMSD) have also similar committors.
Then, we can measure the committor of the clusters which will
approximate those of their member conformations [32]. We do
that without starting any new committor calculation, rather we use
the simulation data that we have already collected [33]. The latter
must contain information about a sufficient number of transition
events in order to get a precise estimate of the cluster committors.
We adopt this strategy in case study 2. In practice, this approach
requires the following steps:
(a) The data about cluster membership and observables (RMSD,
HB) along the trajectory are placed in a single file.
(b) A script is made (e.g., using awk) that analyzes the file from
the previous step and assesses which stable state is reached
first along the trajectory after it leaves each conformation.
298
Giovanni Settanni
The script also counts the fraction of the conformations in

each cluster that reach one state before the other. This is the
cluster committor.
(c) The cluster committor is finally printed along the trajectory,
i.e., each snapshot is assigned with its cluster committor.
At the end of these calculations, in both case study 1 and 2, we
obtain the committor for at least several relevant conformations
along the transition pathways. In particular, with both strategies
we end up with a set of conformations with a committor ~0.5
(Fig.2).
3.5 Validating
andCharacterizing
theTransition State
Ensemble
In case study 1, we have measured the committor for few conformations along the transition pathways then we need to find out, by
analyzing those conformations, if there are observables that may
correlate with the committor. We identified several possible observables and tested them by measuring the degree of correlation with
the committor (Fig.3). In practice, this can be achieved by:
(a) Using wordom or VMD to measure possible observables along
the trajectory.
(b) Extracting (e.g., using awk) the value of the observables for
the conformations with known committor.
(c) Plotting the committor vs the observable.
In Fig.3 only a fraction of all the tested observables is shown.
Visual inspection of the structures with pc~0.5 and comparison
with those from the two stable states must be used to guide the
selection of the relevant observables. As it is often the case, the
observable that best represents the transition is a complex
combination of factors that describes the reciprocal position of
several groups of atoms. In the case of Trpz1 we found that the
reciprocal position of the four tryptophan side chains is crucial in
determining the stage of the transition. When the distance between
the side chains of Trp9 and Trp4 becomes smaller than the distance
between Trp4 and Trp2, then the committor increases. In other
words, when the Trp side chains align with each other as in the
native state, although the packing may not yet be native, then the
committor increases and those conformations are more likely to
fold rather than unfold.
In case study 2, we have measured an approximate committor
c for all the sampled conformations, then we can test if this
approximation represent a good reaction variable. To this extent
we adopt the procedure suggested by Hummer and coworkers
[34] where we compute the conditional probability p(TP|c)
that conformations with a given c belong to a transition pathway.
In practice, we need to:
(a) Combine the c data with the RMSD and HB data in one file,
so that for every sequential snapshot of the trajectory we have
at the same time c, RMSD and HB.
299
Fig. 2 Conformations of the Trpz1 peptide with pc>0.7 (a), pc~0.5 (b) and pc<0.2 (c) from case study 1, simulations at 450K.Adapted from ref. 10 with permission from Elsevier. Superimposed conformations of the native
(d), the TS (e) and the denatured state (f) in case study 2. Adapted with permission from ref. 25. Copyright (2011)
American Chemical Society. In all cases the Trp side chains have been rendered as licorice sticks, while the
other side chains have been hidden for clarity. In df the backbone has been colored according to the prevalent
secondary structure (red extended beta-sheet, blue beta-bridge, yellow turn, gray random coil)
(b) Write a script (e.g., using awk), that bins the conformations
according to c and counts the fraction of conformations in
each c bin that belong to a transition pathway (i.e., p(TP|c)).
(c) Plot the resulting p(TP|c).
300
Giovanni Settanni
Fig. 3 Committor pc (folding probability) plotted as a function of several observables for case study 1. The difference in the distance between the C atom paris of Trp2-Trp4 and Trp9-Trp4 (Dzip) has a good correlation with
pc especially in the transition region, thus it may represent a good reaction coordinate
In the ideal diffusive case where pc is the true committor and

optimal reaction coordinate, p(TP|pc) is a bell shaped curve reaching a maximum of 0.5 at the transition state [34]. The results that
we obtain with our approximate c are shown in Fig.4. The bell
shaped curve distribution of c from our simulations reach values
in between 0.48 and 0.49, as expected from a very good reaction
coordinate. These results demonstrate that the approximate
committor c is almost an optimal reaction coordinate, as it can
separate particularly well the product and reactant and the intermediate species, including the elusive TSE.
3.6 Comparison
withExperiment
In the previous two paragraphs we have shown how it is possible to

check for internal consistency the protein folding model emerging
from the simulations. We still need a way to compare with available
experimental data and/or make predictions about possible experiments regarding the folding kinetics we are trying to characterize.
As mentioned earlier, phi-value analysis is often the experimental
technique employed for studying the folding kinetics of proteins.
Thus, we need a way to translate the structural information about
301
Fig. 4 (a) Distribution of p'c for case study 2. (b) The conditional probability to be
on a transition pathway given the value of p'c (p(TP|p'c)). Adapted with permission
from ref. 25. Copyright 2011 American Chemical Society
the TSE from our simulations, to something like phi-values. Early

studies on protein engineering showed that stability changes upon
mutation correlate with the number of atomic contacts removed/
added with the mutation [4, 35]. This led to assume that phi-
values, that are a ratio of free energy changes in TSE and in native
state upon mutation, would correlate with the fraction of native
contacts formed by the residue in the TSE [5]. Then, a structural
phi-value S(R) for residue R can be defined from the
simulationsas:
SF ( R ) =

iNC ( R )
rTSE ( i )
iNC ( R ) r N ( i )
(2)

302
Giovanni Settanni
where the sums are extended to all the native atomic contacts
NC(R) of residue R. N(i) and TSE(i) are the fraction of conformations where the contacts i is formed in the native (N) and TS
ensemble, respectively. The assumption about the correspondence
of the phi-value and the structural phi-value S has been later
verified by measuring in silico the folding kinetics of a model peptide and its mutants using the techniques described above [33].
In practice, to measure S along the trajectory, we need to
count the atomic contacts along the trajectory. As per commonly
adopted definition, an atomic contact is present when the distance
between a pair of heavy atoms is lower than 6.0. For the contact
calculations, we initially include all the possible pairs of heavy
atoms, with the exclusion of those involving the same residue,
nearest neighbor residues and backbone atoms. To do that we create an AWK script that reads the pdb file of the peptide and generates a WORDOM input file listing all the mentioned pairs of atoms.
The resulting WORDOM input file has the following structure:
WORDOM, with the previous input, generates a large output

file with as many columns as pairs of atoms and as many lines as
snapshots in the trajectory. To limit the extent of the output, we do
not run the analysis on the whole trajectory but only on the subsets
of interest. In particular we run it on the state trajectory file of the
native state and of the transition state, separately. First, we identify
the native atomic contacts as those that are present in at least 2/3
of the snapshots of the native state (N>2/3). Then we measure
the TSE and N only for those contacts and we apply Eq.2 to obtain
the Ss. The latter steps can be achieved with AWK scripts.
The Ss of our simulated systems provide a prediction of the
experimental phi values that can be verified, if those are available.
In the case of the Trpz1 peptide this is not the case, as the peptide
is very small and the folding transition is very broad. However,
kinetics measurements are available for similar peptides like Trpz2,
Trzp3, Trpz4 and the GB1 hairpin [36, 37]. In these experiments
several characteristics of the peptides have been compared and
their effect on the kinetics evaluated. The main outcome of the
experiments is that mutations on the turn region mostly affect the
folding rate, while mutations of the hydrophobic core (i.e., the four
Tryptophan side chains in Trpz1) affect mostly the unfolding rate.
303
Fig. 5 Structural phi-values S(R) for the TSE identified in case study 2. Adapted
with permission from ref. 25. Copyright 2011 American Chemical Society
Simulations data can be used to rationalize those experiments.

Indeed, the structural phi-values obtained from the simulations of
case study 2 (Fig.5) are large in the turn region and lower for
regions increasingly distant from the turn. Similarly, in case study
1 the TSE conformations (Fig.2a) show a hydrophobic core with
a non-native packing of side-chains (but native-like reciprocal distances) and a natively in-register turn region thanks to the closure
of the loop between Thr3 and Thr10. Thus, the turn region is
formed already in the TSE (but not in the denatured/unfolded
state), so that any change in the turn region would mostly affect
the free energy difference between the TSE, where the turn is present and the unfolded state, where the turn is not present, and as a
consequence, the folding rate. A similar argument explains the
effects of perturbations to the hydrophobic core region: the packing of the hydrophobic residues is not-native like in the TSE so
that any change of these residues will mostly affect the free energy
difference between the native state, where they are natively
packed, and the TSE, where they are not, ultimately affecting the
unfolding rate.
Concluding, I have presented detailed procedures to simulate,
using MD, the kinetics of folding of proteins/peptides and obtain
data directly comparable with the relevant experiments, i.e., phi-
value analysis and/or folding kinetics measurements. I have shown
how to practically apply the procedures to real data and pointed
out possible difficulties dependent on the system under investigation. The information provided here should serve the reader to
extend the methodology to an increasingly larger variety of systems
where both computational and experimental investigations are
possible. All the mentioned scripts and input files are available for
download upon request to the author (settanni@uni-mainz.de).
304
Giovanni Settanni
4 Notes
1. Experimentally, the unfolding phi-value is often preferred to
the folding phi-value as it is affected by smaller uncertainties
DDG- F
k mut
Funf =
= ln unf
/ DDGU - F . The two phis are linked
WT
DDGU - F
kunf
by the following relationship: f=1 unf.
2. In this kind of simulations, mostly unfolding transitions are
observed, although by citing the reversibility of the Newtonian
dynamics, they have been assumed to provide information on
the folding pathway, as well [5]. In general, however, such a
large change in environmental conditions may induce changes
in the free energy landscape of the protein and on the folding/
unfolding pathways. Extensive test of the method with the
experiments has shown that this does not occur often, possibly
because of the robustness of the folding/unfolding pathways
of proteins that are the end result of a very long evolution
process.
3. Notwithstanding the conditions used for the simulations, the
limited size of the system and the simulation protocol involving
a slow heating phase prevent the observation of the liquid-vapor
phase transition of water.
4. For the choice of the right cutoff , when the native state of the
peptide is known, our suggested strategy consists in measuring
the dRMSD with respect to the native state along the trajectory
and producing an histogram of dRMSD values. This histogram
typically shows either a small peak at low dRMSD or, at least, a
shoulder. That peak is due to the other conformations of the
native state observed along the trajectory and its location in
terms of dRMSD is a good starting point for the clustering cutoff
because the native state is possibly the narrowest free energy
minimum to be observed in the simulations.
5. In the present case, clusters are portions of n-spheres in the
space of distance matrices, where n is the number of independent
elements of the distance matrix n=N(N1)/2 with N the
number of atoms used for the distance matrix.
6. The number of visited clusters changes in a discrete way. To
measure the VIR, we smooth the number of visited clusters by
convoluting it with a narrow Gaussian kernel (width 100ps)
and then taking the derivative.
7. In reality, for the case study 2 it is possible to identify four
different states using a Markov-state-model approach, however from the slowest relaxation in the system it is possible to
establish that the largest free energy barrier separates the denatured states from the variably folded states. Within those broad
305
states, it is possible to identify further and smaller free energy

barriers, which help distinguish the states of the peptide. This
complex analysis of the kinetics of protein folding is out of the
scope of this book chapter and the interested reader should
refer to the original publication [25] for further details.
8. The committor pc of a conformation follow a scaled binomial
distribution, thus the standard error on the mean value is equal
to pc (1 - pc ) / N t where Nt is the number of trials and it is
largest at committor ~0.5, that is, at the TSE.
9. A more refined approach consist of defining a separatrix in
between the boundaries of the two states and measuring the fraction of commitment runs on the folding side of the separatrix
as a function of the elapsed time along the run. This value follows
an exponential relaxation which converges to the pc (see ref. 10
for details).
References
1. Anfinsen CB, Haber E, Sela M, White FH Jr
(1961) The kinetics of formation of native
ribonuclease during oxidation of the reduced
polypeptide chain. Proc Natl Acad Sci U S A
47:13091314
2. Matouschek A, Kellis JT Jr, Serrano L, Fersht
AR (1989) Mapping the transition state and
pathway of protein folding by protein engineering. Nature 340(6229):122126
3. Fersht AR, Matouschek A, Serrano L (1992)
The folding of an enzyme. I.Theory of protein
engineering analysis of stability and pathway of
protein folding. J Mol Biol 224(3):771782
4. Jackson SE, Moracci M, elMasry N, Johnson
CM, Fersht AR (1993) Effect of cavity-creating
mutations in the hydrophobic core of chymotrypsin inhibitor 2. Biochemistry 32(42):
1125911269
5. Li A, Daggett V (1996) Identification and characterization of the unfolding transition state of
chymotrypsin inhibitor 2 by molecular dynamics simulations. J Mol Biol 257(2):412429
6. Gsponer J, Caflisch A (2002) Molecular
dynamics simulations of protein folding from
the transition state. Proc Natl Acad Sci U S A
99(10):67196724
7. Vendruscolo M, Paci E, Dobson CM, Karplus
M (2001) Three key residues form a critical
contact network in a protein folding transition
state. Nature 409(6820):641645
8. Clementi C, Nymeyer H, Onuchic JN (2000)
Topological and energetic factors: what determines the structural details of the transition
state ensemble and en-route intermediates
for protein folding? An investigation for small

globular proteins. J Mol Biol 298(5):
937953
9. Settanni G, Gsponer J, Caflisch A (2004)
Formation of the folding nucleus of an SH3
domain investigated by loosely coupled molecular dynamics simulations. Biophys J 86(3):
16911701
10. Settanni G, Fersht AR (2008) High temperature unfolding simulations of the TRPZ1 peptide. Biophys J 94(11):44444453
11. Shaw DE, Deneroff MM, Dror RO, Kuskin JS,
Larson RH, Salmon JK etal (2008) Anton, a
special-purpose machine for molecular dynamics simulation. Commun ACM 51(7):9197
12. Abe H, Go N (1981) Noninteracting local-
structure model of folding and unfolding transition in globular proteins. II.Application to
two-dimensional lattice proteins. Biopolymers
20(5):10131031
13. Klimov DK, Thirumalai D (2000) Mechanisms
and kinetics of beta-hairpin formation. Proc
Natl Acad Sci U S A 97(6):25442549

14. Settanni G, Cattaneo A, Maritan A (2001)
Role of native-state topology in the stabilization of intracellular antibodies. Biophys J
81(5):29352945
15. Settanni G, Hoang TX, Micheletti C, Maritan
A (2002) Folding pathways of prion and doppel. Biophys J 83(6):35333541
16. Lindahl E, Hess B, van der Spoel D (2001)
GROMACS 3.0: a package for molecular simulation and trajectory analysis. J Mol Model
7(8):306317
306
Giovanni Settanni
17. Brooks BR, Bruccoleri RE, Olafson BD, States

DJ, Swaminathan S, Karplus M (1983)
Charmma program for macromolecular
energy, minimization, and dynamics calculations. J Comput Chem 4(2):187217
1 8. Seeber M, Cecchini M, Rao F, Settanni G,
Caflisch A (2007) Wordom: a program for
efficient analysis of molecular dynamics
simulations.
Bioinformatics
23(19):
26252627
VMD: visual molecular dynamics. J Mol Graph
14(1):3338, 2738

20. Cochran AG, Skelton NJ, Starovasnik MA
(2001) Tryptophan zippers: stable, monomeric
beta-hairpins. Proc Natl Acad Sci U S A
98(10):55785583

21. Jorgensen WL, Maxwell DS, TiradoRives J
(1996) Development and testing of the OPLS
all-atom force field on conformational energetics and properties of organic liquids. J Am
Chem Soc 118(45):1122511236
WF, Hermans, J (1981) Intermolecular foces.
In: Pullman B (ed). Reidel, Dordrecht, The
Netherlands
23. Berendsen HJC, Postma JPM, Vangunsteren
WF, Dinola A, Haak JR (1984) Molecular-
dynamics with coupling to an external bath.
J Chem Phys 81(8):36843690
24. Hess B, Bekker H, Berendsen HJC, Fraaije
JGEM (1997) LINCS: a linear constraint
solver for molecular simulations. J Comput
Chem 18(12):14631472
25. Radford IH, Fersht AR, Settanni G (2011)
Combination of Markov state models and
kinetic networks for the analysis of molecular
dynamics simulations of peptide folding. J Phys
Chem B 115(22):74597471

26. Ferrara P, Apostolakis J, Caflisch A (2002)
Evaluation of a fast implicit solvent model for
molecular dynamics simulations. Proteins 46(1):
2433
27. Cavalli A, Ferrara P, Caflisch A (2002) Weak

temperature dependence of the free energy
surface and folding pathways of structured
peptides. Proteins 47(3):305314

28. Ryckaert JP, Ciccotti G, Berendsen HJC
(1977) Numerical-integration of Cartesian
equations of motion of a system with constraintsmolecular-dynamics of N-alkanes.
JComput Phys 23(3):327341

29. Hartigan JA (1975) Clustering algorithms.
Wiley series in probability and mathematical
statistics. Wiley, NewYork. xiii, 351 p

30. Du R, Pande VS, Grosberg AY, Tanaka T,
Shakhnovich ES (1998) On the transition
coordinate for protein folding. J Chem Phys
108(1):334350

31. Frenkel D, Smit B (2002) Understanding
molecular simulation : from algorithms to
applications, 2nd ed. Computational science
series. Academic, San Diego. xxii, 638 p

32. Rao F, Settanni G, Guarnera E, Caflisch A
(2005) Estimation of protein folding probability from equilibrium simulations. J Chem Phys
122(18):184901
33. Settanni G, Rao F, Caflisch A (2005) Phi-value
analysis by molecular dynamics simulations of
reversible folding. Proc Natl Acad Sci U S A
102(3):628633
34. Best RB, Hummer G (2005) Reaction coordinates and rates from transition paths. Proc Natl
Acad Sci U S A 102(19):67326737
35. Serrano L, Matouschek A, Fersht AR (1992)
The folding of an enzyme. III.Structure of the
transition state for unfolding of barnase analysed by a protein engineering procedure. J Mol
Biol 224(3):805818

36. Du D, Zhu Y, Huang CY, Gai F (2004)
Understanding the key factors that control the
rate of beta-hairpin folding. Proc Natl Acad Sci
U S A 101(45):1591515920
37. Du D, Tucker MJ, Gai F (2006) Understanding
the mechanism of beta-hairpin folding via phi-
value analysis. Biochemistry 45(8):26682678
Part III
Protein Structure Determination
Chapter 14
Comparative Modeling of Proteins
Abstract
Much of the biochemistry that underlies health, medicine, and numerous biotechnology applications is
regulated by proteins, whereby the ability of proteins to effect such processes is dictated by the threedimensional structural assembly of the proteins. Thus, a detailed understanding of biochemistry requires
not only knowledge of the constituent sequence of proteins, but also a detailed understanding of how that
sequence folds spatially. Three-dimensional analysis of protein structures is thus proving to be a critical
mode of biological and medical discovery in the early twenty-first century, providing fundamental insight
into function that produces useful biochemistry and dysfunction that leads to disease. The large number
of distinct proteins precludes rigorous laboratory characterization of the complete structural proteome,
but fortunately efficient in silico structure prediction is possible for many proteins that have not been
experimentally characterized. One technique that continues to provide accurate and efficient protein structure predictions, called comparative modeling, has become a critical tool in many biological disciplines.
The discussion herein is an updated version of a previous 2008 treatise focusing on the general philosophy
of comparative modeling methods and on specific strategies for successfully achieving reliable and accurate
models. The chapter discusses basic aspects of template selection, sequence alignment, spatial alignment,
loop and gap modeling, side chain modeling, structural refinement and validation, and provides an important new discussion on automated computational tools for protein structure prediction.
Key words Proteins, Comparative modeling, Homology, Threading, Sequence alignment, Structure
alignment, Loop modeling, Structure refinement, Structure validation
Abbreviations
AA
BSE
CASP
CATH
C
DNA
H-bond
MD
NMR
NR
PDB
Amino acid (plural: AAs)

Bovine spongiform encephalopathy
Critical assessment of techniques for protein structure prediction
Class architecture, topology, and homologous superfamily
Alpha carbon on the amino acid backbone
Deoxyribonucleic acid
Hydrogen-bond
Molecular dynamics
Nuclear magnetic resonance
Number of residues
Protein data bank
309
310
PrPC
PrPSc
ps
PSI
RMSD
3D
Prion protein cellular

Prion protein scrapie
Picosecond(s)
Protein structure initiative
Root-mean-squared deviation
Three-dimensional
Introduction
One of the wonders of life is its tremendous diversity, not only in
terms of organisms of vastly different sizes and characteristics but
even within any one single organism. The lowly bivalve, for example, is composed of different tissues that range across an incredible
breadth of color, transparency, flexibility, hardness, adhesiveness,
and electrical conductivity. Such variation is owed primarily to proteins: a class of materials composed of a modest set of ~20 distinct
amino acids (AA) building blocks that evolution has chemically
permuted into the most diverse collection of unique, naturally
occurring substances of any molecular class in existence. By varying the length and sequence of constituent AA chains, proteins can
assemble to form the fundamental matrices of materials that are
harder than stone, softer than soap, as translucent as glass, as
opaque as soot, soluble in water or grease, excellent conductors or
insulators, and among the most efficient known fluorophores,
capacitors, diodes, and catalysts known to man. One of the most
important keys to rationalizing, exploiting, and refining such attributes, and seeking corrective measures when they go awry, is a
detailed understanding of the three-dimensional (3D) assembled
structure(s) of proteins, for it is in this form that proteins adopt
their unique functional properties and exert their intended influence on their surrounding environment.
Beginning with the first atomic-level resolution of a protein
structure (whale myoglobin by Kendrew et al. [1]), 3D protein
models have provided a wealth of insight into biomolecular properties and processes, inspiring a growing thirst for structural detail.
However, whereas biomolecular sequencing has become highly
amenable to efficient, high throughput characterization, the experimental resolution of the 3D protein structure remains timeconsuming and frequently poses significant challenges for
traditional characterization techniques such as X-ray crystallography, with eventual success contingent on luck, persistence, ingenuity, or through the application of innovative techniques [2].
Seminal early work by Anfinsen that will be discussed later, implied,
however, that the structure of any protein was uniquely dictated by
its constituent sequence [3], which suggests that the combination
of a comprehensive catalog of protein sequences and an accurate
understanding of the relationship between sequence and structure
311
could provide the basis for correspondingly comprehensive understanding of the structural proteome. In this spirit, the Protein
Structure Initiative (PSI) (http://www.nigms.nih.gov/Research/
FeaturedPrograms/PSI/) was initiated in 2000 with the expressed
goal of making the three-dimensional, atomic-level structures of
most proteins easily obtainable from knowledge of their corresponding
DNA sequences. As for the human genome project, such objectives are of a scope (around 100,000 proteins, not counting posttranslational modifications and conformational variants) that
requires that conventional resolution methods such as X-ray crystallography and NMR by supplemented by efficient and analytically rigorous computational modeling techniques that perceive
and exploit relationships between primary AA sequence and the
biologically observed 3D structural manifestation of the protein.
This chapter thus offers a brief discussion of in silico protein structure prediction, focusing mainly on comparative modeling.
Whereas other papers (e.g., Baker and ali [4], Mart-Renom et al.
[5]) provide comprehensive reviews of the underlying method
development and research achievements in the field, this chapter
discusses the motivation for using comparative modeling and outlines the practical considerations to be made in assembling such a
model. In the years since the original version of this chapter was
composed [6], computational developments have taken place that
have substantially changed the practice of computational structural
biology: many of the protocols described within the original text
have been effectively implemented as systematically automated
protocols that are often capable of producing plausible (and often
excellent) results with minimal human intervention. This revised
version recognizes this valuable service as an important contribution to the efficient acquisition of knowledge and has added a section which profiles some of the best current (ca. 2013) resources
for automated comparative modeling. However, prudent application of automated protocols still requires the confidence of understanding the underlying manipulations and computations, thus the
meat of this chapter remains a detailed discussion of how protein
structure predictions are computationally effected.
Conceptual Basis for Comparative Modeling

A key inspiration for projects such as the PSI was realization that
the human genome sequencing project was not the panacea for
biological understanding that some had hoped for, but rather
established a basis that helped to illuminate factors other than the
genetic coding portions of our DNA that determine tissue function and differentiation, inter- and intra-species diversity, medical
causality, and other key issues of organism-scale biology. A more
apt currency for appreciating the diversity inherent in life may lie in
312
proteomics wherein one can identify specific units responsible for a

given biological function or dysfunction, quantify their relative
abundance as a function of tissue behavior, and thus help to unravel
the underlying molecular mechanisms. A key source of insight for
the latter is protein structure. Combined with simple rules of physics (e.g., thermodynamics and electrostatics) and chemistry (bond
formation/breaking), 3D molecular structure information can be
readily extrapolated toward understanding (or confidently predicting) molecular physical attributes, inter-molecular association,
enzyme function, and structural response. Structural information
thus serves as a basis for rationalizing relationships between constituent molecules and specific functional or dysfunctional processes
evident in an organism.
Pharmaceutical research frequently follows the paradigm of
discovering a biomolecular target for a given disease, characterizing the target, and finding ways of beneficially modulating its
behavior. Since most targets are proteins, a tool that often plays a
key role in the first step is comparison of protein expression signatures for diseased and healthy tissue samples. Proteins whose
presence or function is noticeably amplified or suppressed in dysfunctional tissue are flagged as potential diagnostic biomarkers and
they and their close partners in biochemical pathways often prove
to be a fertile source of possible therapeutic targets. Therapeutics
are then sought to either restore normal healthy balance in target
function or to compensate for imbalances. Knowledge of protein
3D structure can have some applicability in rationalizing protein
protein interaction partners in the search for prospective targets,
but is especially invaluable for the latter step of therapeutics discovery, providing key insight into potential receptors that might be
pharmacologically targeted, as well as providing a basis for optimizing drug candidates according to their propensity for binding
to specific target receptor. Insight into nontarget protein structures is also very useful for intuiting potential drug side effects, as
proteins with receptors similar to the target are at pronounced risk
of inadvertent modulation by chemicals designed for the latter.
Broad understanding of protein structures should further aid in
expression-based target identification by helping to pinpoint proteinprotein interactions that may obscure whether a deviant
expression profile in a protein is a primary cause of a given dysfunction or a secondary symptom. Specifically, knowing 3D structures
can inform us of which proteins should colocalize in a given cellular environment, and which are likely to engage in direct (i.e.,
actual physical association) or indirect (e.g., exchange of metabolite or transmitter) interactions.
Protein 3D structures may be obtained in a number of ways.
Most high accuracy structures in the protein databank (PDB;
http://www.rcsb.org/pdb) were resolved through crystallographic
techniques, with the rest mainly arising from NMR analysis.
313
While some proteins have been resolved via both techniques,

crystallography is best suited to high-resolution determination of
structures for soluble, predominantly globular, proteins most amenable to crystallization, whereas NMR affords more a dynamic
interpretation of the structure (i.e., multiple conformers are usually resolved from the same data) and is viable for many proteins
that resist crystallization, albeit generally at a lower level of resolution. Theoretically, the entire proteome should be accessible to
either NMR or crystallography, although the lipophilicity of transmembrane proteins poses experimental complications while some
of the most massive proteins still entail major challenges of data
deconvolution. While the trajectory of decades of structural
biology offers hope that technical advances will eventually conquer
the remaining obstacles, it is more uncertain whether experimental
resolution of the entire known proteome will ever be achieved, as
both crystallographic and NMR remain technically very demanding, time-consuming, and expensive. The underlying assumption
of the PSI is that computational modeling may yet accomplish
what is spectroscopically impractical.
The main schemes for computational protein structure prediction include: (1) self assembly simulations, (2) associative models based on sequence pattern recognition, and (3) comparative
modeling. All of these methods were inspired in part by seminal
findings of Christian Anfinsen et al. that an unraveled ribonuclease
AA chain could, in plain solution, coalesce within a reasonable
time frame (minutes to hours) to form a protein functionally
indistinguishable from native in vivo ribonuclease [3]. This
implied that: (a) a proteins can be uniquely identified by AA
sequence, (b) this sequence uniquely encodes the in vivo protein
function, and (c) the AA sequence is capable of consistent selfassembly into functional form, based exclusively on intramolecular interactions among sequence AAs plus interatomic interactions
with surrounding solvent. Observation (c) thus directly suggested
that assembly simulations could be a reliable mode for protein
structure prediction, whereas observations (a) and (b) were key to
the eventual formulation of pattern recognition and comparative
modeling techniques.
Practical formulation of associative and comparative modeling
methods required a substantial basis in empirical understanding of
protein structure. One key development was the accumulation of a
reasonable volume of protein sequence data beginning in the early
1950s, leading to gradual elucidation of strong correlative patterns
between protein sequence similarity on one hand and analogous
function on the other. This permitted the classification of proteins
into families and superfamilies [7], with an underlying assumption
being that members of the same protein superfamily are all evolutionarily related (a.k.a., homologous), and members of the same
family are closely related.
314
A second important precursor to protein structure prediction

was the acquisition of crystal structures from the late 1950s
onwards. These structures generally validated the earlier family and
superfamily classifications in that proteins with similar sequences
and function were usually found to have commensurately similar
3D structure. Analysis of structural data also revealed that all proteins tended to assemble as a sum of a limited number of unique
substructure forms. At a fine grained level, it was noted that all
proteins tended to adopt similar H-bond stabilized features such as
alpha helices, beta sheets, and a number of distinct turn and hairpin structures, collectively referred to as secondary structure.
At coarser levels, it also became apparent that the full manifold of
protein structures could be classified into a relatively small number
of unique folds: characteristic self-stabilizing collections of secondary structure elements. Estimates for the precise number of distinct
folds that one should expect to find in the proteome have varied
significantly over the past 20 years [8] but may be converging to
approximately 2,000. In a practical sense, while there are almost
certainly unique protein classes that have not yet been structurally
resolved, the number of distinct, experimentally characterized and
validated folds has recently stagnated: as of July 2013, the lists of
distinct folds registered by CATH [9] and SCOP [10] from the
body of PDB structures had not been added to since 2009 (1,282)
and 2008 (1,393) respectively, where the parenthetical numbers
reflect the number of distinct folds registered in each of these two
classification schemes. In the heyday of the PSI (ending in around
2007), approximately 20 % of newly solved protein structures corresponded to new folds, but that ratio has obviously dwindled,
perhaps indicating that the remaining portion of fold space might
be dominated by challenging (e.g., poor solubility, or substantially
disordered) proteins that are inherently resistant to conventional
crystallographic and NMR techniques.
In light of the current state of the practice, it thus appears that
comparative modeling might be destined to applicability within a
subset of the proteome, but fortunately within this regime the
underlying basis for the technique is solid. Not surprisingly, proteins within the same family (as classified by sequence) almost
invariably correspond to similar fold classifications. Such correspondence between sequence similarity trends versus structural
similarity completes the basis for both associative and comparative
modeling techniques. Specifically, associative models are based on
probabilistic tendencies of certain AA combinations to adopt a
given secondary structure and tendencies of certain secondary
structure elements to adopt a specific fold, whereas the more general tendency of proteins with similar sequences to adopt similar
3D structures is the foundation for comparative modeling.
While each of the different structure prediction method is
potentially applicable to modeling a given target protein, each has
limitations to heed. Self-assembly simulations, commonly known
315
as protein folding models, are especially computationally demanding

due to huge numbers of unique conformers that must be considered for even modest-sized proteins. With such large conformer
manifolds, one may also have to empirically choose between multiple unique but comparably favorable structures. A more severe
problem exists for the fraction of protein population whose thermodynamically optimal conformer is not the one observed in vivo,
the most famous example of which being the prion PrPC whose
native form appears to be higher in energy than a rogue form,
PrPSc. Presence of the latter rogue form of the protein in living
tissue is believed to catalyze the exothermic refolding of healthy
PrPC to the biologically inutile PrPSc, with the result being prion
diseases such as scrapie, Creutzfeldt-Jakob Disease or BSE [11].
Given the greater thermodynamic stability of the latter, a good
folding algorithm based on the underlying free energy effects
assumed by Anfinsen to drive the assembly process [3] should converge to the rogue PrPSc form rather than the biologically healthy
PrPC. In truth, energy is often not the sole driving force behind
assembly, in that life forms have developed sophisticated tools
called chaperonins that perform protein assembly quality control,
correcting misfolds and sometimes preferentially effecting assembly of higher energy structural forms. Relevant interactions
between chaperonins and their client polypeptides are very complex and the associated mechanisms of mediated folding are still in
the process of being elucidated [12, 13]. Folding simulations that
explicitly consider chaperonin contributions have shown promise
in de novo structure prediction [1214]; however, it is not yet
clear whether they will sustain general success over the diverse collection of proteins that require some form of folding mediation.
Accurate associative modeling is also fairly challenging due to
prospects for compounding errors. Even for the most sophisticated
prediction algorithms, the first step of secondary structure prediction from raw sequence remains limited to about 80 % success on
a per-residue basis [15], thus leading for a medium-sized protein
(hundreds of AAs) to dozens of local errors at the outset. Since
local errors in backbone torsion can easily extrapolate to large
errors in the global structure (e.g., mislocation of entire lobes),
many proteins are poorly predicted. Furthermore, the stated 80 %
success rate applies primarily to globular proteins, and indications
are that similar predictions for transmembrane proteins are worse
on average [16] largely because there is a much smaller pool of
secondary structure features resolved membrane-spanning proteins available for training predictive models.
Assessment of results from the Critical Assessment of
Techniques for Protein Structure Prediction (CASP) competitions
reveal that the average predictive performance of comparative models remains consistently superior to nontemplate methods [17].
For example, in the 6th CASP competition (2004), on the full set
of systems for which both comparative and folding models were
316
generated, with fully automated comparative models achieving a

mean RMSD (root mean-squared deviation) relative of 2.17 relative to experimental crystal structures and further improvement to
2.09 RMSD given human intervention in template selection,
alignment modification, etc. By contrast, folding methods (with
and without intervention) attained mean RMSD values of 2.41 and
2.51 respectively. Comparative models are also relatively computationally expedient; however, their dependence on quality 3D templates with structural and functional similarity to the target precludes
many targets from consideration. Fortunately, the current body of
structurally resolved representative fold structures is believed to
encompass the majority (perhaps two thirds) of all distinct protein
classes, thus in the majority of scientific studies it should be possible
to identify a prospective template or modeling cases.
Methods
In practice, comparative modeling is best viewed not as one technique
but rather as a strategy for assembling information from various component methods (including assembly and associative techniques)
toward a unified 3D structure prediction. In general, these component steps can be approximately summarized as follows:
1. identify template proteins with structural similarity to the target as gauged (optimally) from sequence-based homology, or
from physicochemical similarity.
2. align the target sequence with all relevant template sequences
according to the same arguments of homology or physicochemical similarity employed in step 1.
3. spatially align all of the template structures into a single framework, and use the sequence alignment to project the target
protein backbone onto this framework.
4. estimate structures for target protein fragments that are illrepresented by the template manifold, or else omit them from
the predicted structure.
5. align target side chains with analogous side chains from the
template structures, or intelligently guess their disposition
according to known spatial and torsional preferences.
6. refine unphysical contacts and strains via conformational
searches, and,
7. evaluate the final relaxed model for physical tenability.
Each step above entails various methodological and strategic considerations, some of which provide opportunities for
iterative feedback to prior steps, as is shown graphically in Fig. 1.
These considerations will be elaborated upon in the remainder
of this section.
317
Fig. 1 Flow diagram for comparative modeling of proteins showing standard process (solid arrows) and
feedback/refinement mechanisms (dashed arrows)
3.1 Template
Identification
The two important objectives in identifying templates are quality

and quantity. Specifically, one wishes to find templates with a high
quality match to your target of interest so that they offer a strong
likelihood of corresponding to similar structures, which will thus
provide an apt framework for assembling your target structure.
Secondly, unless one is able to uncover a very high quality template
that covers the full span of interesting folds in the target with a
high degree of homology, it is very useful to identify multiple possible templates. The strength in numbers is derived from several
factors: the different templates often provide a useful source of
validation in that they should display comparable structure in
regions of functional overlap, and templates that cover distinct
nonoverlapping regions of the target can collectively serve as a
complementary basis with which to stitch together a more complete model of the target that might be possible from a single template alone. There is even some benefit in the case where multiple
templates cover largely the same regions and features of the target:
if different plausible templates exhibit some compositional diversity but have similar 3D folding patterns, the quality of the resulting sequence alignment and comparative model has been generally
found to benefit since multiple sequence alignments are typically
more accurate than pair-wise alignments, and employing simultaneous input from multiple structures may tend to average out local
structural aberrations.
318
When identifying prospective templates via sequence homology,

a reasonable template-target sequence identity is important for
minimizing the risk of errors in sequence alignment and for maximizing the likelihood that the entire template fold is to the target.
The widely used minimum sequence identity criterion for homologous templates is 30 % sequence identity over the extent of the
mutual target-template alignment: this criterion is expected to provide reasonable confidence that the target and template do indeed
share a common fold. The precise origin of the 30 % identity criterion is difficult to trace, in that the seminal paper typically cited for
template identification, by Chothia and Lesk, offered a more conservative criterion of 50 % identity, although their objective was to
find highly similar structures with less than 1.0 RMSD in the
position of backbone atoms [18]. Conversely, others have claimed
that identity levels as low as 2022 % are frequently still viable
[19], provided one has clear empirical evidence that suggests a
clear functional analogy between the target and template proteins.
Nonetheless, the empirical 30 % criterion is broadly accepted, has
stood the test of careful analyses [20] and, when complemented by
a careful, well-scrutinized sequence alignment, generally affords
confidence for a plausible structure prediction.
While many targets exist for which no templates with at least
30 % sequence identity are available, it has been estimated via fold
statistics that close to two thirds of all possible targets of interest
should have a template of reasonable structural similarity already
present in the PDB. This wealth of templates may be explained
from the fact that structural conservation can often be largely
retained even in cases of very distant ancestral commonality.
Furthermore, shared structure is also possible courtesy of evolutionary convergence (i.e., unrelated proteins may gradually adapt
similar structures due to innate functional benefits that are uniquely
available to such conformations) or in spite of evolutionary divergence (i.e., a protein retains a desirable conformation attained by a
distant ancestor, in spite of an extensive mutational history that
reduces the sequence identity to well below 30 %). Thus if no template meeting the 30 % sequence identity is identified through
homology searches, a reasonable model may still be achieved via a
technique known as threading that evaluates target sequences
against a library of unique fold representatives (a collection of
structurally unique proteins such as those collected in the CATH
database [9]) according to residue by residue similarity in terms of
shape, spatial volume, hydrophilicity/hydrophobicity, helix- or
sheet-forming propensity, etc. Functionally special residues such as
prolines (whose closed-ring structure induces a kink in AA chains),
cysteines (well known for their fold-stabilizing disulfide bonds),
and ionic residues (aspartate, glutamate, lysine, and arginine all
tend to occur predominantly near solvent accessible surfaces of proteins, except in cases where they form salt bridges) are frequently
319
Fig. 2 Observed correspondence between PROSPECT Z-scores computed for

pairs of proteins and the likelihood of their sharing a common fold or having a
familial relationship
given additional weight in any assessment. As sequence identity is

generally less useful for assessing threading templates, a consensus
Z-score is instead used to gauge statistical significance for the quality of one template candidate relative to others being considered.
Z-score scales vary across the manifold of different threading programs, but typically provide an indication of how likely it is that a
given target/template pair are actually members of a common protein family or super family, whether they merely share a common
fold, or whether they have no obvious relationship. An example of
such ranges, as reported for the widely used PROSPECT-II program [21], is provided in Fig. 2.
The Z-score provides a reasonable scheme for identifying the
plausible templates, but has a margin of error that must be heeded.
A study contrasting performance of PROSPECT-II with other
threading programs found that the top-scoring PROSPECT-II
match was a valid fold-conserving template 84.1 % of the time
when the template manifold contained species within the same
family as the target, 52.6 % of the time when superfamilial (but no
familial) templates were available, and 27.7 % if the best templates
merely shared a common fold [21]. These numbers improved to
88.2, 64.8 and 50.3 % when examining the top five matches for
possible valid templates. These are reasonably successful ratios relative to competing threading models at the time, but do highlight
the possibility that invalid templates may achieve high scores, and
that one should scrutinize multiple top-scoring candidates, looking for templates known to prefer similar regions of cells as the
target species, known to perform functions at least vaguely similar
to the target, etc.
A more recent and more rigorous assessment was performed
by Brylinski and Lingam [22] explored the capacity of threading
programs to correctly identify multiple legitimate templates by
choosing not just the top-scoring candidate but also successively
ranked options, while recognizing the increasing risk of selecting
false positives as one scans down the list. Surveying ten popular
threading techniques via searches for distant analogs (i.e., any templates with sequence identity of greater than 40 % relative to the
query were considered to be trivial matches and thus were omitted
320
from statistics) of a diverse collection of test molecules, the HHPred

method [23] was found to give the best performance of any single
model, correctly identifying 50 % of real structural analogs (true
positives) before incorrectly selecting as many as 5 % of those
prospective templates that were not actually structurally related
(i.e., false positives). The eThread method, which selected templates based on a consensus of the ten distinct models, was able to
improve on this performance by achieving on average a true positive rate of 60 % before suffering a 5 % false positive selection. In
cases where a reasonable sampling of templates has been identified,
a high degree of modeling performance can be attained by
comparing the structures of the different templates and discarding
obvious outliers (i.e., poor alignment with the bulk of the selected
templates), but scenarios where few high scoring templates are
identified run the risk of producing poor structure predictions.
Note that in cases where templates have been identified by
homology, but the sequence identity is somewhat below the safe
range (i.e., less than 40 %, and especially if less than 30 %), threading can provide an excellent point of validation. If threading analysis on the target template does not rank the target at a level that
seems to assure a conserved fold, the putatively homologous template should be scrutinized with care and skepticism.
Finally, although one may rely substantially on the numerical
guidance of sequence identity or threading Z-scores for identification and specification of useful templates, there are aspects of the
process that can benefit from informed human intervention. For
example, if you seek to model a protein in a specific functional
conformation (e.g., many proteins exhibit sizable structural variations between activated and inhibited states) it is important to bias
your template manifold toward that specific conformation.
Whenever possible, one should avoid using multiple templates that
span conformationally different states as this may produce a final
model reflective of an unphysical hybrid of those states. Finally, for
metalloproteins, where the presence or absence of the metal ions
can substantially impact the resulting protein conformation, one
should identify the desired metal ion state of the target and choose
templates with analogous ion residency.
3.2 Sequence
Alignment
Most of the various programs commonly used for template identification generally also yield a tentative sequence alignment relative
to the target. In homologous cases with greater than 50 % targettemplate sequence conservation over the mutually aligned portion
of the structure, it is generally assumed that the alignment prediction algorithm will produce a qualitatively reliable alignment with
only modest local misalignments (no positional errors more than
several residues). Over a data set of broadly varying protein similarity, the PROSPECT-II assessment of threading reliability was that
the program could achieve about a 60 % average accuracy in
321
prediction the alignment position of any given residue, and typically

located each residue within 4 AA positions of the correct spot
around 80 % of the time [21]. In fairly strong threading models
(e.g., PROSPECT Z-scores >10) one can likely assume better
performance, perhaps on par with good homology alignments.
However in all cases, and especially those with poorer sequence
identity or Z-scores, careful manual validation is a good policy.
This can be achieved by comparing the alignment relative to the
known 3D (template) structure to identify any of the following
cases:
sizeable gaps (i.e., greater than 2 or 3 AAs) present in template

core regions.
gaps greater than 1 AA in known template sheets or helices.
target prolines located within known template helices.
positional displacement of more than 1 or 2 AAs for template

cysteine residues known to engage in disulfide bonds.
displacement of more than 23 AAs for template ionic residues

known to form salt bridges.
Alignments containing instances such as those above can yield

unphysical comparative models for which fold conservation may be
less favored or even no longer feasible. Manual adjustments may
thus be made to alleviate such errors, at the expense of opening or
extending gaps in exterior loops, as long as relatively modest penalties are incurred in the overall alignment score. Any templates that
have a significant number of alignment problems that cannot be
alleviated through minor manual adjustment should be viewed
with skepticism. Furthermore, the number of irreconcilable alignment problems observed for a given target/template pair is a good
source of evaluating and prioritizing that template's suitability relative to other candidates.
The presence of template gaps in regions that should correspond solvent exposed loops in the target species is rarely a key
criterion in evaluating a template or a target-template alignment.
With the exception of large deletions that might lead to excessive
loop shortening (which might produce untenable strain in the
model), the impact of minor modifications to such loops has minimal influence on the basic fold of a protein, thus from an evolutionary perspective they tend to have the largest frequency of
noncritical point mutations, insertions, and deletions among
homologs. If the gap in such a loop is relative small (less than 5
AAs), and is not believed to play an important role in properties of
interest (i.e., not being part of a known receptor or proteinprotein
interaction site) it might be justifiably omitted, although it is often
more satisfying to piece it together according to empirical loop
libraries (e.g., [24, 25]) or as an arbitrary structure such as a beta
turn, with the expectation that its relatively large mobility will be
322
respond well to subsequent refinement steps. For gaps that are

significantly longer than 10 AAs, however, the structural integrity
of your predicted model is best served by finding a legitimate
template for the unrepresented loop itself. This can entail addition
of a lower ranked template that may be less suitable for global
alignment, but that does afford reasonable alignment to the gap
regions. If such a template is believed to have regions other than
the gap in question that are inappropriate for modeling the rest of
the target (i.e., very poor alignment, or known to have an inappropriate conformation), one may discard all of the template except
that specifically corresponding to the gap region.
3.3 Spatial
Alignment
of the Target
Once templates have been selected and a sequence alignment is in

place, construction of a preliminary structural model of the target
is straightforward and can usually be accomplished reliably with
black-box processes implemented in most comparative modeling
programs. The assembly methods typically construct a backbone
model first, and then incorporate side chains into the resulting
framework. Some methods may assemble the core region of a protein (solvent-inaccessible portions of the structure plus conserved
secondary structure elements) first, then treat exposed loops. Most
backbone modeling methods begin by superimposing all templates
onto a common framework and computing a consensus backbone
defined by mean positions of corresponding Cs. One of three
different schemes is then typically used: (1) rigid body assembly of
target fragments corresponding to nongapped portions of the target/template alignment as is implemented in the COMPOSER
program [26], (2) segment matching whereby target protein fragments are projected onto the consensus backbone with torsional
angles being established through reference to fragment polypeptide libraries (e.g., SEGMOD [27]), or (3) construction of a compromise model that minimizes steric and torsional restraints of the
target backbone as it is projected onto the template framework
(e.g., MODELLER [28]). All of the above methods have proponents, and it is unclear whether the ultimate model accuracy varies
much as a function of method choice, in that subsequent refinement is usually required regardless of technique. The restraintsbased formalism is probably the most widely used by comparative
modelers at this point.
One major complication to the assembly process may arise
when the target protein contains a terminal domain that is represented by a different template than is used for the main core of the
protein. In such cases, it may be unclear from the relevant templates how the terminal domain should pack onto the core framework. One method that we have applied for such a scenario [29] is
to assemble the main core and the terminal domain as separate
models, and to estimate the preferred coreterminus packing via
proteinprotein docking analysis as performed by GRAMM [30],
323
ZDock [31] and other programs. Protein-protein docking methods

typically offer a suitability score for each predicted complex, thus
helping to guide the selection of complex to serve as the packing
model. Another practical and critical consideration is whether a
predicted complex places the two units in a position where it is
possible to chemically rejoin the broken AA chain without unphysical strain on the backbone. Given a very low probability of predicting a complex that perfectly places the final AA of one unit in
covalent binding distance to the initial AA of the next unit, our
strategy has been to omit a portion of the target sequence within
the core-terminus boundary region during initial model construction for these two domains, with the intention of re-integrating
this portion as a flexibly modeled gap (see next section). Specifically,
our recommended protocol is to omit a number of residues (say
NR) that are not predicted to be part of a defined secondary structure element (i.e., unstructured coil), and then constrain the choice
of docked complexes to those that place the final C of the first
domain within a distance of less than NR 3.0 from the first C
of the next domain (i.e., somewhat less than the maximum
unstrained CC distance of approximately NR 3.8 ). This
method has not yet been exhaustively validated, but is based on
reasonable logic and provides a potentially workable solution to an
otherwise very challenging problem.
3.4 Loop and Gap
Modeling
While the previous backbone assembly step should yield structure

predictions for many (hopefully most) of the loops in your target, it
is common for some to be represented by poor sequence conservation or to appear in gapped regions of the alignment. Gaps covering
5 AA positions or less pose minimal concerns as they can typically
be patched in with reasonable accuracy by referring to polypeptide
structure libraries (e.g., [24, 25, 27]), but gaps of greater length are
difficult to reliably model via methods other than template comparison. In cases where no template for the gap can be found from
either homology searches or threading, it is still possible to attempt
a prediction based either on the same peptide libraries available for
short loops or an algorithmic scheme akin to protein folding strategies (e.g., for molecular dynamics see Chapter 1 and for Monte
Carlo conformational searches see Chapter 2). Neither of these
strategies is guaranteed to produce a model with close correspondence to its real optimal structure, especially for gap lengths larger
than 10 AAs. In general, however, many loops without a suitable
template remain unresolved precisely because they inherently have
high conformational mobility. In such cases, nonoptimal conformers may well still correspond to in vivo accessible structures.
3.5 Side Chain

Modeling
Conserved disulfide bonds and salt bridges are typically incorporated into the target model directly during the backbone assembly
process. Beyond this, side chain positions for highly conserved
324
residues may also be inferred directly from the template, although

their conformations are known to vary significantly from one
protein to another when sequence identity is less than 50 %, and
are considered to be poorly conserved for identities less than 30 %
[32]. Fortunately a number of effective side-chain packing algorithms have been developed and validated (see Huang et al. [33]
for an excellent review), and are often implemented as black-box
features in comparative modeling programs. For the fastidious, the
graphical program PyMol [34] provides an excellent qualitative
side chain sampling and evaluation tool embedded within the program's mutagenesis toolkit. This tool permits users to click their
way through a manifold of distinct, plausible side chain conformers, which each structure graphically rendered according to favorable and unfavorable contacts with the surrounding protein.
Deepview [35] offers a somewhat similar mutation module that,
although a bit less user friendly, provides the added benefit of
molecular mechanics side chain optimization.
3.6
Refinement
The nature of comparative modeling, whereby protein structures

are predicted by analogy to different but related proteins, invariably leads to some degree of structural error. If the template
(or template manifold) contains no sequence gaps relative to the
target and only minor variations in sequence, the resulting error
may be less than the resolution accuracy of the template(s) thus
further correction to the predicted target structure would be
unnecessary. However if the target-template alignment contains
gaps or has sequence identity less than 70 %, some structural refinement is advisable. In cases where some templates have been selected
to cover regions of the sequence poorly represented by primary
templates, boundary effects arise where one template may significantly perturb the projected target backbone derived from other
templates and vice versa. Library models used for loop patching
lack specific environment information necessary to adapt the loop
to the target of interest and thus yield imperfect backbones. It is
finally important to note that most backbone errors are amplified
in the predicted side chain conformation. Fortunately, most models that have been assembled with care and without unrealistic
assumptions will possess only modest errors that may be corrected
in a physically natural manner. In vivo, a protein that has been perturbed from its native conformation due to some minor disturbance (e.g., intermolecular interaction and change in ionic state )
stands a reasonable chance of reverting to normal structure when a
stable normal environment is restored. By analogy, one may consider the deviations arising from a careful but imperfect protein
assembly to be comparable to an environmental perturbation, thus
any simulation that subjects the protein model to conditions akin
to a normal in vivo environment should encourage reversion to a
reasonable structure. Molecular dynamics (MD) methods are
exceptionally well suited to this task: protein simulation is one of
325
the most common MD applications (see Chapter 1), and excellent

force fields (Chapter 4) have thus been developed for simulating
the interactions of proteins (and their constituent cofactors, ions,
etc.) with solvent environments comparable to in vivo conditions.
Various reviews on the MD strategies, techniques, and resources
applicable to protein structure prediction and refinement are available (e.g., Refs. [36, 37]). The main drawback to MD simulations
is computational expense, however, thus when seeking to mitigate
unphysical structural aspects arising from model assembly one may
find that constant-temperature simulations may not accomplish all
necessary refinements in a viable time frame. An alternative is to
perform simulated annealing calculations wherein the protein is
gradually warmed up to a fairly hot temperature (1,000 K is a practical upper limit for short pulse heating; for temperatures hotter
than this or for heating durations longer than 10 ps, one may
observe potentially irreversible unphysical structural changes) and
then slowly cooled back to ambient temperature. After multiple
thermal cycles of this sort (typically 510, with each cycle lasting
from 10100 ps), most moderate errors in the original structure
are corrected, generally leaving a fairly plausible conformer. Some
comparative modeling programs such as MODELLER [28] contain an embedded MD code and perform annealing as an option of
the structure prediction process.
Special caution is due when modeling transmembrane proteins. Many MD methods and parameters have been tailored for
the specific case of soluble proteins immersed in a polar (generally
aqueous) media, and may yield unphysical conformations for
inherently hydrophobic membrane-binding portions of a protein.
Special methods for simulating transmembrane proteins are discussed in Chapters 811 of this volume and e.g., in a review by Im
and Brooks [38].
3.7
Validation
Any protein structure determination, from a rough comparative

model to a high-resolution synchrotron-based crystal structure,
will differ somewhat from the real in vivo conformation. A number
of validation tools have thus arisen to evaluate structural models
and detect aspects that appear to differ conformationally from
standard bond distance, angle, torsion, or contact ranges derived
from extensive assessment of known structures. Conveniently,
many of the best validation tools are available online, e.g.:
SAVES (http://nihserver.mbi.ucla.edu/SAVES/): uses multiple

distinct servers running PROCHECK, WHAT_CHECK,
ERRAT, VERIFY3D, and PROVE to examine amino acid stereochemistry, appropriateness of nonbonded contacts, secondary
structure populations, and packing.
Harmony (http://caps.ncbs.res.in/harmony/): evaluates

structures according to backbone conformation, solvent accessibility, and hydrogen bonding.
326
MolProbity (http://molprobity.biochem.duke.edu): validation

relative to Ramachandran distributions, steric clashes and
hydrogen bonding [39].
PROSESS (http://www.prosess.ca/): validates structures

according to various measures including covalent geometry,
nonbonded packing, clashes, His & Asn side chain flips, and
global and local atomic-level energies [40].
In comparing the above tools, one finds some redundant

checks, but each of these utilities has unique features (albeit SAVES
is a unique aggregation of other separate services), thus giving
consideration to the feedback from all of these services is a valuable
step toward producing a high quality model. Among the various
warnings that are likely to be issued for a comparative model, some
minor problems may be alleviated by more refinement, however
care should be exercised in that over-refinement is often itself a
source of error. In the most serious cases, excessive use of simulated annealing can produce effects akin to protein denaturation,
while a more moderate scenario could involve removal of conformational attributes that might have corresponded in the template
to intermolecular interactions (e.g., association with another protein) that might have an analogous influence on the function of the
target system. More serious errors may be indicative of poor (hopefully correctable) choices in the original sequence alignment or
loop assembly in the region highlighted. Other issues may arise not
from the modeling process but rather from using templates that
themselves contain errors. To reduce instances of the latter, one
may perform similar validation analysis on candidate templates
prior to use and thus screen out inferior structures.
Judgments on whether a validation warning warrants countermeasures hinge on how severe the error appears to be, whether it
lies in a particular interesting region of the molecule, and on
whether it might impact analysis planned for the model. For small
molecule docking studies, accurate structures are required in the
receptor region, and side chain conformations can be critical, thus
one would likely try to alleviate any appreciable error within reasonable nonbonding radius (typically about 8 ) of the putative
binding site. Protein-protein docking studies and MD analyses are
generally much less sensitive.
Automated Structure Predictions

The years since the first incarnation of this chapter appeared have
been witness to a key development in the field of comparative modeling: automated modeling. The practice of computational protein
structure prediction has become greatly popularized, in part thanks
to a proliferation of user-friendly and generally effective protein
structure prediction services. This is a natural outgrowth of the
327
prestigious CASP competitions that have placed a premium on the

development of intelligent web-based services that can automatically implement the complete comparative protocol based only on
user specification of a specific target protein sequence. There are
few computational science tasks as complex as protein structure
prediction that have been so effectively implemented as black box
protocols.
The primary focus of this chapter (and indeed much of this
book) has been practical instruction on how biological researchers
can effectively implement complex computational analyses toward
better understanding of numerous facets of protein modeling.
To some extent the availability of black box solutions diminishes
the need for detailed how-to instruction, especially if one can generate quality results based on a process as trivial as go to site X, paste
in your protein sequence, hit compute and wait for the results to
appear. What value, then, is retained from detailed instructions
such as those presented herein?
The primary value of a rigorous and fundamental methodological understanding is the ability to objectively perceive whether a
protocol is sensible and provides a good probability of generating
accurate structures. In most cases, black box services such as the
ones that will be briefly reviewed in this section do permit diligent
users to examine log files that clearly specify the specific manipulations undertaken en route to the final answer. Users should not use
black box services that do not conveniently enable such queries,
and should make the effort to evaluate the chosen templates, alignments, and refinement strategies at least once for any given target
protein. Since black boxes can be as fallible as humans, careful
quality control remains important: if validation data is provided by
the automated service, the user should scrutinize it for every run
that produces a structure that the user plans to use for subsequent
analysis. If detailed validation is not provided automatically, users
should subject the generated structures to various independent
structure validation services discussed in the previous section.
While there are too many servers that perform automated comparative protein modeling to treat each comprehensively in this
chapter, it is helpful to at least list some of the most noteworthy:
SwissModel (http://swissmodel.expasy.org/): Perhaps the oldest
structure prediction server, SwissModel emerged in 1995 as a
value-added service in conjunction with the SwissProt collection of
curated protein structure templates [41]. It remains widely used,
but is limited to identification of templates with unequivocal
homology to the query sequence, which constrains the scope of
the service.
I-TASSER (http://zhanglab.ccmb.med.umich.edu/I-TASSER/):
Currently the state of the practice protein structure prediction
service, receiving top server accolades in each of the CASP7, 8, 9
328
and 10 competitions, I-TASSER constructs models based on

multiple threading alignments [42]. The server is extensively used
and relies on an actively updated template library.
RaptorX (http://raptorx.uchicago.edu/about/): An excellent
server for cases where the best templates have very low sequence
identity relative to the query protein [43], RaptorX is the latest
version of a line of servers from the Jinbo Xu group. RaptorX is
complemented by various analytical tools including protein binding site prediction, residue contact map analysis, etc.
HHPred (http://toolkit.tuebingen.mpg.de/hhpred): Strictly
based on homology modeling (i.e., technically not employing
threading), HHPred achieves high quality structure predictions
[44] for most of the targets well addressed by threading (it consistently ranks well in CASP competitions) while achieving a very
high degree of computational efficiency.
ROBETTA (http://robetta.bakerlab.org/): A unique combination of homology modeling and ab initio self assembly routines,
ROBETTA is able to stitch together plausible models of proteins
that may be missing sizable portions when modeled by comparative modeling alone [45]. The self assembly portion of this service
is performed by ROSETTA which, although much more computationally demanding than purely comparative modeling, employs
steering algorithms that render it much more efficient than conventional molecular dynamics methods.
The question of whether it is better to build comparative models in an automated manner or to rely on human rationalization is
an issue worthy of concluding this chapter. Fairly thoughtful studies have concluded that although fully automated models perform
nearly as well on average as those that with operator intervention,
expert involvement is nonetheless beneficial [46, 47], especially in
guiding the somewhat instinctive processes of prioritizing machinerecommended templates (in particular, determining which templates correspond to functional states analogous to that which one
wishes to model), checking automated alignments, and recognizing functional motifs whose conservation is critical to generating a
model that adheres to a function of interest. Conversely, automated threading procedures are often capable of identifying compositionally dissimilar but functionally valid templates that might
be perceived even by domain experts, and are probably better
equipped to perform the corresponding (often nonintuitive)
sequence alignment. Furthermore, a well-validated automated
model will probably outperform nonexperts on average. The basic
conclusion, perhaps, is that human and machine are converging on
a synergistic partnership, with the relative roles of each varying
according to the complexities of each case, but remaining complementary. Regardless of the underlying protocol, a well-constructed
329
and validated structure model can open many doors for subsequent
analysis. In addition to valuable insight derived from simple visual
inspection, the model can form a reliable basis for many other
modeling analyses, as are discussed extensively in other chapters of
this book.
References
1. Kendrew JC, Bodo G, Dintzis HM, Parrish RG,
Wyckoff H, Phillips DC (1958) A threedimensional model of the myoglobin molecule
obtained by x-ray analysis. Nature 181:662666
2. Konermann L, Pan Y (2012) Exploring membrane protein structural features by oxidative
labeling and mass spectrometry. Expert Rev
Proteomics 9:497504
3. Anfinsen CB, Redfield RR, Choate WI, Page J,
Carroll WR (1965) Studies on the gross structure, cross-linkages, and terminal sequences in
ribonuclease. J Biol Chem 207:201210
4. Baker D, ali A (2001) Protein structure prediction and structural genomics. Science 294:
9396
5. Mart-Renom MA, Stuart AC, Fiser A, Snchez
R, Melo R, Sali A (2000) Comparative protein
structure modeling of genes and genomes.
Annu Rev Biophys Biomol Struct 29:291325
6. Lushington GH (2008) Comparative modeling
of proteins. Methods Mol Biol 443:199212
7. Dayhoff MO (1972) Atlas of protein sequence
and structure. National Biomedical Research
Foundation,
Georgetown
University,
Washington DC
8. Schaeffer RD, Daggett V (2011) Protein folds
and protein folding. Protein Eng Des Sel
24:1119
9. Pearl F, Todd A, Sillitoe I, Dibley M, Redfern
O, Lewis T, Bennett C, Marsden R, Grant A,
Lee D, Akpor A, Maibaum M, Harrison A,
Dallman T, Reeves G, Diboun I, Addou S, Lise
S, Johnston C, Sillero A, Thornton J, Orengo
C (2005) The CATH domain structure database and related resources Gene3D and DHS
provide comprehensive domain family information for genome analysis. Nucleic Acids Res
33:D247D251
10. Murzin AG, Brenner SE, Hubbard T, Chothia C
(1995) SCOP: a structural classification of proteins database for the investigation of sequences
and structures. J Mol Biol 247:536540
11. Prusiner SB (1991) Molecular biology of prion
diseases. Science 252:15151522
12. Takagi F, Koga N, Takada S (2003) How protein thermodynamics and folding mechanisms
are altered by the chaperonin cage: molecular
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
simulations. Proc Natl Acad Sci U S A 100:

1136711372
Kmiecik S, Kolinski A (2011) Simulation of
chaperonin effect on protein folding: a shift from
nucleationcondensation to framework mechanism. J Am Chem Soc 133:1028310289
Baumketner A, Jewett A, Shea JE (2003)
Effects of confinement in chaperonin assisted
protein folding: rate enhancement by decreasing the roughness of the folding energy landscape. J Mol Biol 332:70113
Dor O, Zhou Y (2006) Achieving 80% tenfold
cross-validated accuracy for secondary structure prediction by large-scale training. Proteins
66:838845
Rost B (2001) Protein secondary structure
prediction continues to rise. J Struct Biol 134:
204218
Kryshtafovych A, Fidelis K (2009) Protein
structure prediction and model quality assessment. Drug Discov Today 14:386393
Chothia C, Lesk AM (1986) The relation
between the divergence of sequence and structure in proteins. EMBO J 5:823826
Chung SY, Subbiah S (1996) How similar
must a template protein be for homology modeling by side-chain packing methods? Pac Symp
Biocomput 126141
Forrest LR, Tang CL, Honig B (2006) On the
accuracy of homology modeling and sequence
alignment methods applied to membrane proteins. Biophys J 91:508517
Dongsup K, Xu D, Guo JT, Elrott K, Xi Y
(2003) PROSPECT II: protein structure prediction program for genome-scale applications.
Protein Eng 16:641650
Brylinski M, Lingam D (2012) eThread: a highly
optimized machine learning-based approach to
meta-threading and the modeling of protein tertiary structures. PLoS One 7:e50200
Soding J (2005) Protein homology detection
by HMM-HMM comparison. Bioinformatics
21:951960
Fernandez-Fuentes N, Oliva B, Fiser A (2006)
A supersecondary structure library and search
algorithm for modeling loops in protein structures. Nucleic Acids Res 34:20852097
330
25. Kolodny R, Koehl P, Guibas L, Levitt M

(2002) Small libraries of protein fragments
model native protein structures accurately.
J Mol Biol 323:297307
26. Sutcliffe MJ, Haneef I, Carney D, Blundell TL
(1987) Knowledge-based modelling of homologous proteins. Part I. Three dimensional
frameworks derived from the simultaneous
superposition of multiple structures. Protein
Eng 1:377384
27. Levitt M (1992) Accurate modeling of protein
conformation by automatic segment matching.
J Mol Biol 226:507533
28. ali A, Blundell TL (1993) Comparative protein modeling by satisfaction of spatial
restraints. J Mol Biol 234:779815
29. Lushington GH, Zaidi A, Michaelis ML (2005)
Theoretically predicted structures of plasma
membrane Ca2+-ATPase and their susceptibilities to oxidation. J Mol Graph Model 24:
175185
30. Tovchigrechko A, Vakser IA (2005)
Development and testing of an automated
approach to protein docking. Proteins 60:
296301
31. Wiehe K, Pierce B, Mintseris J, Tong W,
Anderson R, Chen R, Weng Z (2005) ZDOCK
and RDOCK performance in CAPRI rounds 3,
4, and 5. Proteins 60:207221
32. Chung SY, Subbiah S (1996) A structural
explanation for the twilight zone of protein
sequence homology. Structure 4:11231127
33. Huang ES, Koehl P, Leavitt M, Pappu RV,
Ponder JW (1998) Accuracy of side-chain prediction upon the near-native protein backbones
developed by ab initio folding methods.
Proteins 33:204217
34. PyMol v. 1.6 (2013) http://www.pymol.org/
35. Johansson MU, Zoete V, Michielin O, Guex N
(2012) Defining and searching for structural
motifs using DeepView/Swiss-PdbViewer.
BMC Bioinformatics 13:173
36. Caflisch A, Paci E (2005) Molecular dynamics
simulations to study protein folding and
unfolding. In: Buchner J, Kiefhaber T (eds)
Protein folding handbook, vol 2. Weinheim,
Wiley-VCH Verlag GmbH, pp 11431169
37. Karplus M, Kuriyan J (2005) Molecular

dynamics and protein function. Proc Natl Acad
Sci USA 102:66796685
38. Im W, Brooks CL III (2005) Interfacial folding and membrane insertion of designed
peptides studied by molecular dynamics simulations. Proc Natl Acad Sci USA 102:
67716776
39. Chen VB, Arendall WB, Headd JJ, Keedy DA,
Immormino RM, Kapral GJ, Murray LW,
Richardson JS, Richardson DC (2010)
MolProbity: all-atom structure validation for
macromolecular
crystallography.
Acta
Crystallogr D Biol Crystallogr D66:1221
40. Berjanskii M, Liang Y, Zhou J, Tang P,
Stothard P, Zhou Y, Cruz J, Macdonell C, Lin
G, Lu P, Wishart DS (2010) PROSESS: a protein structure evaluation suite and server.
Nucleic Acids Res 38:W633W640 Webserver
edition
41. Peitsch MC (1995) Protein modeling by
E-mail. Biotechnology 13:658660
42. Roy A, Kucukural A, Zhang Y (2010)
I-TASSER: a unified platform for automated
protein structure and function prediction. Nat
Protoc 5:725738
43. Kllberg M, Wang H, Wang S, Peng J, Wang
Z, Lu H, Xu J (2012) Template-based protein
structure modeling using the RaptorX web
server. Nat Protoc 7:15111522
44. Hildebrand A, Remmert M, Biegert A, Soding J
(2009) Fast and accurate automatic structure
prediction with HHpred". Proteins 77:12832
45. Song Y, DiMaio F, Yu-Ruei Wang R, Kim D,
Miles C, Brunette TJ, Thompson J, Baker D
(2013) High resolution comparative modeling
with RosettaCM 21(10):173542
46. Das R, Qian B, Raman S, Vernon R, Thompson
J, Bradley P, Khare S, Tyka MD, Bhat D,
Chivian D, Kim DE, Sheffler WH, Malmstrom
L, Wollacott AM, Wang C, Andre I, Baker D
(2007) Structure prediction for CASP7 targets
using extensive all-atom refinement with
Rosetta@home. Proteins 69:118128
47. Zhang Y (2009) I-TASSER: fully automated
protein structure prediction in CASP8.
Proteins 77:100113
Chapter 15
De Novo Membrane Protein Structure Prediction
Timothy Nugent
Abstract
Recent advances in identifying residueresidue contacts from large multiple sequence alignments have
enabled impressive gains to be made in the field of protein structure prediction. In this chapter, we discuss
these advances and provide a step-by-step guide to applying the latest tools to the de novo modelling of
alpha-helical transmembrane proteins. As a practical example, we demonstrate the process of building an
accurate 3D model of a G protein-coupled receptor, correctly orientated in the membrane, using only its
primary protein sequence.
Key words Transmembrane protein, De novo modelling, Contact prediction, Structural bioinformatics
Introduction
Membrane proteins are encoded by approximately 30 % of the
genes of a typical genome and perform crucial roles in a diverse
range of essential biological processes including transport of ions
and small molecules, intercellular communication and signal transduction. They are also important drug targets, with estimates suggesting that about 60 % of current drug targets are membrane
proteins [1]. Recently, there has been encouraging progress in
structure determination led by structural genomics initiatives that
explicitly target integral membrane proteins, resulting in increasing
coverage of important protein families [2, 3]. Despite this, coverage of membrane protein fold space remains sparse as only about
1 % of structures in the Protein Data Bank (PDB) describe membrane proteins, of which about 300 are unique. However, the technical difficulties associated with purification and structure
determination by X-ray crystallography and NMR spectroscopy are
likely to prohibit a rapid increase in these numbers. Computational
structure prediction therefore provides a vital alternative approach
with which to further our understanding of both the structure and
function of this important class of proteins.
331
332
Timothy Nugent
Historically, tools to analyse membrane proteins have focused

on topology prediction methods which typically use machine learning algorithms trained on the limited structural data that is available [410]. Predicting membrane protein structure by homology
modelling can be highly effective when a suitable template is available, especially when membrane protein-specific tools are used
[11, 12]. Given that fold preservation in transmembrane regions
requires less sequence conservation than for globular proteins, it
should be possible to obtain relatively accurate 3D models even at
low (<20 %) sequence identity [13], although the small number of
solved membrane protein structures will limit the number of families that such methods are applicable to. De novo modelling
approaches, which attempt to build 3D models from sequence
information alone, have typically relied upon knowledge-based
potentials derived from the statistical analysis of known structures.
A number of membrane protein de novo modelling tools exist
including FILM [14] and Rosetta Membrane [1517].
FILM (Folding In Lipid Membranes) is based on the globular
protein structure prediction method FRAGFOLD [18, 19] which
performs a conformational search guided by simulated annealing
using highly resolved super secondary structural fragments to
assemble the tertiary fold. FILM adds a knowledge-based membrane potential to the standard FRAGFOLD energy terms (pairwise, solvation, steric, and hydrogen bonding), derived from the
statistical analysis of a data set of 640 transmembrane helices with
experimentally defined topologies. By assessing the relative frequencies of each amino acid at fixed distances from the membrane
centre, the membrane term is calculated by transforming these values using the inverse Boltzmann equation. Applying the method
to small membrane proteins of known 3D structure showed that it
was capable of predicting both the helix topology and the conformations of these proteins at a reasonable accuracy level. However,
reproducing the compactness of large transmembrane helix bundles was challenging, since transmembrane helix bundles are usually not optimally compact, although neighboring helices are
closely packed. Subsequent modification of the method allowed
the prediction of larger bundles by incorporating lipid exposure
prediction into the potential function, allowing models of seven
transmembrane helix bacteriorhodopsin and rhodposin to be generated with 67 root mean square deviation (rmsd) to the experimentally determined structure [20].
RosettaMembrane, a modification of the Rosetta method [21,
22], which also assembles folds using fragments of know structures
using simulated annealing, added terms to the energy function that
described intraprotein and proteinsolvent interactions in the
anisotropic membrane environment. The method describes interactions between protein residues in atomic detail while applying
continuum solvent models to the water, hydrophobic core, and
333
lipid head group regions of the membrane. RosettaMembrane was

able to successfully predict the structures of 12 small transmembrane protein domains (up to 150 residues) to within 4 rmsd of
the native structures, with near-atomic resolution predictions
(<2.5 rmsd) achieved for three domains, suggesting that the
model captures the essential physical properties that govern the
solvation and stability of membrane proteins. More recently, the
method was extended to incorporate distance constraints into the
predictions by constraining helixhelix interactions, derived from
both experimental data or predicted from sequence [2325]. This
allowed larger (90300 residues) structures with more complicated topologies to be successfully modelled to within 4 rmsd in
the best four cases, with results indicating that only a single constraint was sometimes sufficient to enrich the population of nearnative models.
While the use of knowledge-based potentials derived from statistical analyses of known membrane protein structures has been the
standard approach for de novo structure prediction, the field has
changed dramatically over the last 5 years as new methods capable of
accurately inferring residueresidue contacts from large multiple
sequence alignments (MSAs) have emerged, allowing the direct
computation of structures from sequence. The progress of these
methods has been largely driven by the rapid growth in the size of
sequence databases; while 5 years ago a database may have contained
a few hundred sequences for a typical protein family, the number can
easily be ten times as many today [26]. It is this wealth of sequence
data that allows detection of correlated mutations between sites in
MSAs. The key idea behind correlated mutations is that residues
proximal in 3D space are likely to impose constraints on each other,
which should lead to a correlation in their substitution patterns in a
MSA. Should either residue mutate, the stability of the contact
might be disrupted, therefore impacting on the stability of the structure. Should one or both residues mutate to a more physicochemically complementary pairing, the contact is more likely to be retained.
Residue pairs that form contacts in the native structure are therefore
seen to coevolve in tandem, and it is this property that modern contact prediction methods seek to exploit.
Many different strategies have been applied to predicting contacts from sequence data, typically based on the direct recognition
of substitution patterns, and often using machine learning-based
approaches. However, the major obstacle has been in dealing with
indirect coupling effects: should a direct physical coupling exist
between sites AB and BC, an apparent coupling may emerge
between AC even though no direct interaction exists. It is this indirect coupling that has the effect of blurring the signal and has been
the bottleneck in prediction accuracy. A recent implementation by
Lapedes et al. [27] dealt with this chaining problem by applying a
maximum entropy approach but at a high computational cost.
334
Timothy Nugent
Another method, Direct Coupling Analysis (DCA), reduced the

problem to one of maximum entropy inference of the strength of
the interaction parameters between columns in the MSA, and
applied a heuristic message passing approach to determine the
solution of the contact weights [28]. This allowed the approach of
Lapedes et al. to be put to practical use, with prediction accuracy
achieving sufficient quality to be useful in structure prediction
[29]. Another method, PSICOV, is based on the properties of the
inverse covariance (or precision) matrix, which will only contain
nonzero elements where there is direct coupling influence.
Calculating the inverse covariance matrix, while not always possible, is usually costly, therefore PSICOV constrains the solution
based on sparsity using the graphical lasso method [30, 31]. Where
sufficient numbers of aligned sequences were available, PSICOV
was able to predict contacts with an accuracy approaching 80 %
even for long-range contactsthose separated by >23 residues in
the sequencesufficient to identify to the native fold for medium
sized (<200 residue) globular proteins [32]. More recently, the
plmDCA method [33], which uses a pseudolikelihood method
applied to the Potts models, was shown to significantly outperform
existing DCA-based approaches, while a consensus approach
PconsC that combines PSICOV and plmDCA predictions using a
random forest classifier was shown to provide a relative improvement of 20 % in comparison to the best single method [34].
The prediction accuracy of these recent methods has led to the
development of a number of de novo structure prediction methods
capable of generating accurate models for even large domains,
guided primarily by predicted contacts. EVfold [35], which uses
DCA in combination with the CNS molecular dynamics software
suite to generate 3D models, was able to determine accurate structures for fifteen targets ranging in size from 50 to 260 residues to
within 2.74.8 rmsd of their native structures over at least twothirds of the protein. Evfold_membrane [36], which incorporates
predicted transmembrane topology into the model, was able to
generate model with correct folds (TM-score >0.5, see Note 1)
amongst the top ten scoring candidate models in 22 out of 25 targets. The latest version of FILM, FILM3, replaced the statistical
potential with a single scoring function based on predicted contacts and their estimated probabilities [37]. Benchmark results
using contacts predicted by PSICOV indicated that models with
TM-scores >0.5 could be generated for 25 out of 28 membrane
protein targets with complex topologies and an average length
over 300 residues. In the best case, it was possible to build a model
of cytochrome c oxidase polypeptide I with a TM-score >0.75 over
all 514 residues. While these results are extremely encouraging,
data suggests that even with perfect distance constraints, folding
methods are unable to generate models less than 2 rmsd of the
native structure, suggesting that protein refinement will play an
important role in generating higher accuracy models.
335
Methods
This chapter will describe the process of generating a 3D model of
an alpha-helical membrane protein starting with only its primary
sequence and in the absence of structural homologues. A number
of steps are required that will be discussed in detail: (1) predicting
residueresidue contacts, (2) predicting secondary structure and
transmembrane topology, (3) generating candidate 3D structures,
(4) recombining candidate structures to generate a final model, (5)
refinement, (6) orientating of the refined model in the membrane,
and (7) model quality assessment. To reproduce these steps you will
need access to a UNIX/Linux workstation with a number of software packages and databases installed. While we will focus on tools
developed in-house at UCL, many of the methods can easily be
substituted or combined with other programs. As a target sequence,
we will use the 329-residue bovine rhodopsin protein (PDB code
1GZM), a prototypical G protein-coupled receptor (GPCR).
2.1 Predicting
ResidueResidue
Contacts
The key to building de novo 3D models using FILM3 is the use of

predicted contacts, which are generated here with PSICOV (see
Note 2). The input file for PSICOV is a large and diverse MSA. In
the original PSICOV paper, the example with the lowest number
of sequences had 511 homologous sequences, so an alignment size
upwards of this is critical for accurate predictions. Alignments
smaller than this, or with low sequence diversity, may result in
PSICOV failing to converge on a solution or aborting, although
in both cases this behaviour can be overridden. A number of
programs can be used to generate MSAs; here we will use HHblits
(see Note 3) which uses profile hidden Markov models to generate
alignments [38]. To build a MSA from the target sequence (in
FASTA file format) using a Uniprot [39] database, use the following
command:
./hhblits -i 1gzmA.fasta -d uniprot20_2013_03
-oa3m 1gzmA.a3m -mact 0 -n 3 -diff inf -cov 60
The options generate a MSA in a3m format (-oa3m), setting
the maximum accuracy (MAC) algorithm realignment threshold to
0 therefore generating quasi-global alignments, running three iterations (-n), filtering both query and databases to generate the most
diverse alignment (-diff), and setting the minimum coverage of the
target sequence to 60 % (-cov). Full details of HHblits command
line parameters can be found by passing the help all flag. The
following command can then be used to convert the a3m formatted
MSA into the PSICOV input format, which simply consists of a
single sequence per row, with the target sequencewhich must
contain no gapsas the first line. Duplicate rows are also removed:
egrep -v "^>" 1gzmA.a3m | sed 's/[a-z]//g' |
sort -u > 1gzmA.aln
336
Timothy Nugent
Another method that can be used to build a MSA is jackhmmer,

part of the HMMER 3.0 package [40] (see Note 4). To build a
MSA using jackhmmer against a UNIREF100 database, use the
following command:
./jackhmmer -N 3 -E 1e-6--ncE 1e-6--otextw -A
1gzmA.sto 1gzmA.fasta uniref100.fasta
Here the command line flags set the number of iterations to 3
(-N), and the E-value threshold to 1e-6. The resulting Stockholm
formatted file (-A) can be converted to a3m format with the reformat.pl Perl script, which is included in the HH-suite package:
./reformat.pl 1gzmA.sto 1gzmA.a3m
This can then be converted to the PSICOV format using the
previous egrep command. If you use an alternative method to generate the MSA which produces a FASTA formatted alignment file,
this can be converted to the PSICOV format using the fasta2aln
tool included with PSICOV:
./fasta2aln 1gzmA.msa.fasta > 1gzmA.aln
Having successfully generated the MSA, PSICOV can be run
with the following command:
./psicov -p -d 0.03 1gzmA.aln > 1gzmA.con
Here, the command line options output positive predictive values (PPV) rather than raw output scores (-p) and set the precision
matrix sparsity to 0.03 (-d)justified by the observation that on
average only ~3 % of all residue pairs are observed to be in direct
contact. The resulting output file is in 5 column CASP RR format
(http://predictioncenter.org/casp10/index.cgi?page=format):
i j d1 d2 p
Where i and j are residue numbers, d1 and d2 are real numbers
where values of d1 = 0 and d2 = 8 indicate a predicted contact, and
p (probability) is the PSICOV PPV. When passed to FILM3, only
contacts with a probability >0.5 will be used to build the model, so
this list of contacts can easily be replaced or augmented with contacts determined by experimental methods or other predictors
such as plmDCA or PconsC, ensuring that the probability is above
this threshold. Contacts that are more confidently predicted will
result in a lower pseudo-energy when satisfied in the 3D model.
2.2 Predicting
Secondary Structure
and Transmembrane
Topology
Fragment selection by FILM3 is based upon predicted secondary

structure using PSIPRED and transmembrane helix predictions
using MEMSAT-SVM. Predictions are combined using a decision
tree scheme ensuring that transmembrane helix positions are
enforced, additionally inserting a small amount of coil in the centre of predicted transmembrane loops if it did not already exist.
337
Where transmembrane loop regions are predicted, helix, coil or

helix/coil is predicted, depending on the PSIPRED confidence.
To generate this ensemble secondary structure string, the
PSIPRED and MEMSAT-SVM predictions are combined with the
Combine-MEMSAT-SVM-PSIPRED.pl Perl script, which is
included in the FILM3 package. To generate the PSIPRED prediction (see Note 5), run it passing in the target FASTA file:
./runpsipred 1gzmA.fasta
This will generate two output files containing the predicted
secondary structure and neural network probabilities used to make
the prediction1gzmA.ss2 and 1gzmA.horiz.
Run the MEMSAT-SVM (see Note 6) Perl script as follows:
./run_memsat-svm.pl 1gzmA.fasta
A number of output files are produced including 1gzmA.
memsat_svm, which contains the predicted transmembrane topology, and 1gzmA_SVM_ALL.out, which contains the raw support
vector machine (SVM) scores. To generate the ensemble secondary structure, place these four files in the same directory as the
Combine-MEMSAT-SVM-PSIPRED.pl Perl script and run it passing in just the base name of the target:
./Combine-MEMSAT-SVM-PSIPRED.pl 1gzmA
This will generate two files called 1gzmA.ess and 1gzmA.zcoord. The third line of 1gzmA.ess contains the ensemble secondary
structure string, where C = coil, H = helix, E = sheet, h = helix or
coil, and ? = unknown. This string can easily be modified by hand if
you want to enforce a different secondary structure using alternative methods, or the transmembrane helix boundaries in the
Topology line at the end of the 1gzmA.memsat_svm file can be
modified and the Combine-MEMSAT-SVM-PSIPRED.pl re-run.
The file 1gzmA.zcoord contains predicted Z-axis (perpendicular to
the Cartesian plane formed by the membrane surface) coordinates
for each residue, which are linearly extrapolated from the transmembrane topology prediction. These are used to provide minimum distance constraints to FILM3, ensuring that the meandering
nature of the target proteins transmembrane topology is enforced.
These coordinates can easily be replaced by those generated by
more elaborate schemes such as ZPRED [41].
2.3 Generating
Candidate 3D
Structures
Using FILM3
The FILM3 input file, 1gzmA.nfpar, has the following format:

----------1gzmA.nfpar---------ALNFILE 1gzmA_FILM3.aln
INITEMP 0.6
MAXSTEPS 20000000
POOLSIZE 9
338
Timothy Nugent
TRATIO 0.6
MAXFRAGS 5
MAXFRAGS2 25
CONFILE 1gzmA.con
ZFILE 1gzmA.zcoord
------------------------------Three files are referenced: the first is an alignment file, although
the format is slightly different to that used by PSICOV. Instead,
the first three lines from the 1gzmA.ess file are added to the beginning of the PSICOV alignment file. Additionally, the second line is
set to the size of the PSICOV alignment file. The file can therefore
be generated as follows:
head -3 1gzmA.ess > 1gzmA_FILM3.aln
cat 1gzmA.aln >> 1gzmA_FILM3.aln
Then simply change the second line in 1gzmA_FILM3.aln to
the alignment size, which can be determined with:
cat 1gzmA.aln | wc l
The file should then appear as follows, assuming the original
PSICOV alignment file contained 1,969 sequences:
----------1gzmA_FILM3.aln---------1gzmA
1969
CCCCCCCChhhhCCCCCCCCCCCCCCCCCCCCChHHHHHHHHHHHH
HHHHHHHHHHHHHHHHHhhCCCCChhHH
MNGTEGPNFYVPFSNKTGVVRSPFEAPQYYLAEPWQFSMLAAYMFL
LIMLGFPINFLTLYVTVQHKKLRTPLNY
etc.
----------------------------------The other two files contained in 1gzmA.nfpar are the contacts
predicted by PSICOV1gzmA.conand the predicted Z-axis
coordinate file1gzmA.zcoord. In the FILM3 paper, the use of
Z-coordinate distance constraints generated lower energy models
in eight cases out of 28; the use of these coordinates can therefore
be suppressed by removing or commenting out the corresponding
line. The remaining options in the 1gzmA.nfpar control aspects of
the Replica Exchange Monte Carlo function, which in general can
be left at their default values. The MAXSTEPS parameter is the
total number of fragment swaps to make, POOLSIZE is the number of replica conformations to use, and INITEMP and TRATIO
indicate the starting temperature and temperate ratio between
each replica. MAXFRAG and MAXFRAG2 are the minimum and
339
maximum number of fragments to use for each residue. With the

input file complete, FILM3 can be run as follows:
./film3 1gzmA.nfpar > 1gzmA.pdb
The FILM3 binary (see Note 7) expects to find four files in the
working directoryatomdist.dat, rotalib.dat, tdb.dat, and tor.
lstthese contain atomic distance, side-chain and fragment library
information, and are include in the FILM3 data directory. Running
FILM3 once will generate a single model. For the following recombination step, an ensemble of say 100 models or more are required,
and usually the more models that are generated, the more accurate
the final model will be. We would recommend generating 200
models per target100 models with Z-coordinate distance constraints, and 100 models without. Below we assume that the models are named 1gzmA_i.pdb, with i running from 1 to the number
of models. While FILM3 can be run on a single computer, generating large numbers of models is typically achieved using a cluster
of machines.
2.4 Recombining
Candidate Structures
to Generate a Final
Model
Models generated by FILM3 have their energy written in the

HEADER record of the PDB file. Rather than simply selecting the
final model with the lowest energy, a combinatorial refinement step
can be used to generate a single final model from the ensemble of
models. Here, the lowest energy model is identified and the next
100 lowest energy models, selected from the pooled set of
Z-coordinate constrained and unconstrained models, are fitted to
it by rigid body superposition. The recombination randomly selects
fragments from the ensemble and transfers them on to the lowest
energy structure to see if a lower energy model can be produced.
This greedy search procedure is repeated until no further improvement in energy is observed. To generate the ensemble, use the
super_models.csh script included in the FILM3 package. This uses
the ProFit tool for superpositioning (see Note 8), but many alternatives exists. Once installed, identify the lowest energy model by
parsing the HEADER records:
cat *.pdb | grep HEADER | sort n | tail -1 |
cut -c 21-30 | xargs -i grep -e {} *.pdb
Assuming this command identifies 1gzmA_3.pdb as the lowest
energy target, and the 100 lowest energy candidate structures are
in a directory named models/, we now modify the super_models.csh script accordingly:
----------super_models.csh---------set refpdb = models/1gzmA_3.pdb
# Superpose
directory
all
models
found
in
models
340
Timothy Nugent
foreach pdb (models/1gzmA_*.pdb)

echo -n "ref " $refpdb > profit.cmds
echo "" >>profit.cmds
echo -n "mob " $pdb >>profit.cmds
echo "" >>profit.cmds
echo "atoms CA" >>profit.cmds
echo "fit" >>profit.cmds
echo "write temp.pdb" >>profit.cmds
echo "quit" >>profit.cmds
# Assume that profit command is in command path
profit < profit.cmds >/dev/null
cat temp.pdb
echo "END"
end
-----------------------------------Then run the scripts as follows to generate the ensemble:
csh super_models.csh > 1gzmA_ensemble.pdb
Now run contactrecomb, included in the FILM3 package,
passing it the PSICOV contact file and the ensemble secondary
structure file that were generated previously. This command will
generate a single recombined model, 1gzmA_recomb.pdb:
./contactrecomb 1gzmA_ensemble.pdb
1gzmA.ess 1gzmA_recomb.pdb
2.5 Model
Refinement
1gzmA.con
After recombination, the final model can be refined using

MODELLER (see Note 9) to produce reasonable loop and side
chain conformations [42]. The recombined model is simply used
as a template for MODELLER, but with additional secondary
structure restraints applied to regions predicted to be alpha-helical
by MEMSAT-SVM. The following two files need to be generated,
firstly an alignment file that can be read by MODELLER:
----------1gzmA_refine.ali--------->P1;1gzmA_recomb
structureX:1gzmA_recomb:::::PDB::0.00:0.00
MNGTEGPNFYVPFSNKTGVVRSPFEAPQYYLAEPWQFSMLAAYM
FLLIMLGFPINFLTLYVTVQHKKLRTPLNYILLNLAVADLFMVFG
GFTTTLYTSLHGYFVFGPTGCNLEGFFATLGGEIALWSLVVLAIER
YVVVCKPMSNFRFGENHAIMGVAFTWVMALACAAPPLVGWSRY
IPEGMQCSCGIDYYTPHEETNNESFVIYMFVVHFIIPLIVIFFCY
GQLVFTVKEAAAQQQESATTQKAEKEVTRMVIIMVIAFLICWLPY
341
AGVAFYIFTHQGSDFGPIFMTIPAFFAKTSAVYNPVIYIMMNKQ
FRNCMVTTLCCGKNDDE*
>P1;SEQ
sequence:SEQ:::::SEQ::0.00:0.00
MNGTEGPNFYVPFSNKTGVVRSPFEAPQYYLAEPWQFSMLAA
YMFLLIMLGFPINFLTLYVTVQHKKLRTPLNYILLNLAVADL
FMVFGGFTTTLYTSLHGYFVFGPTGCNLEGFFATLGGEIALW
SLVVLAIERYVVVCKPMSNFRFGENHAIMGVAFTWVMALACA
APPLVGWSRYIPEGMQCSCGIDYYTPHEETNNESFVIYMFVV
HFIIPLIVIFFCYGQLVFTVKEAAAQQQESATTQKAEKEV
TRMVIIMVIAFLICWLPYAGVAFYIFTHQGSDFGPIFMTIPAF
FAKTSAVYNPVIYIMMNKQFRNCMVTTLCCGKNDDE*
-----------------------------------In this file, the name of the structure in the first and second lines
must match the name of the final recombined model, in this case
1gzmA_recomb. The MODELLER Python script is as follows:
----------1gzmA_refine.py---------from modeller import *
from modeller.automodel import *
log.verbose()
env = environ()
class MyModel(automodel):
def special_restraints(self, aln):
rsr = self.restraints
at = self.atoms
rsr.add(secondary_structure.alpha(self.
residue_range(39,63)))
a = MyModel(env, alnfile
= '1gzmA_refine.ali',
knowns = '1gzmA_recomb', sequence = '1gzmA_
recomb')
342
Timothy Nugent
a.starting_model= 1
a.ending_model
= 1
a.md_level = refine.slow
a.make()
----------------------------------Here, the residue ranges to which alpha-helical secondary
structure restraints are applied, according to MEMSAT-SVM
transmembrane helix boundary predictions, can be added using
the rsr.add command. The alignment file and model name must be
referenced accordingly on the following line. The actual refinement step is initiated by the refine.slow command, which uses
molecular dynamics with simulated annealing [43]. Finally, run the
script using Python:
python 1gzmA_refine.py
This will generate the final recombined and refined model
(Fig. 1).
Fig. 1 The final recombined and refined model of rhodopsin
2.6 Orientation
of the Model
in the Membrane
343
Identifying the correct orientation of the model within the lipid

bilayer allows us to study the complex relationship between
sequence, structure, and the lipid environment. Using
MEMEMBED (see Note 10), which couples a knowledge-based
membrane potential, calculated by the statistical analysis of
transmembrane protein structures, with a combination of genetic
and direct search algorithms, we can quickly and accurately orientate membrane proteins within the lipid bilayer [44]. To orientate
the model, run the following command:
./memembed -n out -s 3 q 1 1gzmA_refined.pdb
This should generate a file called 1gzmA_refined_EMBED.
pdb (Fig. 2). Here, the optional flags tell the program that the
N-terminus of the model has an extracellular location (-n out), as
determined by the MEMSAT-SVM prediction, the s 3 flag tells
the program to search for the optimal orientation using a genetic
algorithm, repeating the search five times, and then returning the
lowest energy orientation, while the -q 1 flag will optimise the
hydrophobic thickness of the bilayer after orientation, returning
a value in angstroms, and positioning the membrane leaflets
appropriately rather than at the default values of 15 and 15 .
Fig. 2 The model orientated in the membrane. The blue plane indicates the
membrane inner leaflet; the red plane is the membrane outer leaflet
344
Timothy Nugent
By passing in the models predicted transmembrane helix boundaries,

the tilt angle relative to the membrane for each helix and the model
as whole can also be calculated for the orientated model:
./memembed -z -r "A" -t 39,63,73,96,109,133,154
,173,203,224,252,274,287,309
1gzmA_refined_EMBED.pdb
The z flag tells the program to calculate tilt angles using
the comma separated helix boundary list provided by t. The
r flag indicates which chain the topology refers toif the model
has no chain identifier, pass an empty character in quotes.
Finally, if PyMOL [45] is installed, the included Python script
can easily be used to generate an image of the orientated structure:
./pymol_membrane_image.py
pdb 1gzmA_refined_EMBED.png
1gzmA_refined_EMBED.
In addition to MEMEMBED, a number alternative methods

exists that can position structures accurately in the membrane.
These include the TMDET, Ez-3D, and PPM servers [4648].
2.7 Model Quality
Assessment
Making an assessment of the quality of a model is clearly important

step since it provides users with a measure of confidence in the
prediction. Unfortunately, for de novo models, this is challenging
due to the absence of an experimental structure to compare against.
There are however a number of ways in which an assessment of
model quality can be made. Firstly, the estimated precision of contacts predicted by PSICOV can be used. In development of FILM3,
it was possible to build models with TM-scores >0.5, therefore
indicating approximately the correct fold [49], for 26 out of 28
targets where there were at least 20 predicted contacts with a PPV
>0.5. The two targets that met this criteria but had a TM-score
<0.5 were both part of a larger complex and appear to be heavily
stabilised by additional chains in their native states. The second
method is to calculate the mean pairwise TM-score between all
models in the ensemble of candidate structures. Where predicted
contacts are sufficient to determine the correct fold, there should
be a high degree of similarity in the ensemble. We found that the
mean pairwise TM-score showed a strong correlation with the
TM-score of the final model, therefore allowing the final model
TM-score to be predicted using linear regression. You can use the
film3mqap program to calculate these values using the ensemble
that was generated previously:
./film3mqap 1gzmA_ensemble.pdb
This will generate a mean pairwise TM-score for the ensemble
and a predicted TM-score for the final modela value >0.5 indicating the model probably has the correct fold. The program also
computes a pseudo-temperature-factor from the ensemble of models
345
Fig. 3 The final model coloured by pseudo-temperature-factor. Warmer colours

indicate high variation in the ensemble. These regions, particularly the N-terminus
at the top of the image, correspond to areas of the contact map where there are
few predicted contacts
and writes these to an output file (Fig. 3). When displayed in a

graphics package such as PyMOL with the residues are coloured by
temperature, regions in the ensemble that are clearly defined and
thus more likely to be correct can easily be identified:
./film3mqap 1gzmA_ensemble.pdb 1gzmA_refined.pdb
1gzmA_refined_bfactors.pdb
It is also possible to use the predicted transmembrane topology to assess models. Modern topology prediction methods can
achieve accuracies approaching 90 % on certain data sets; therefore,
ensuring that the model displays the correct topology is a relatively
straightforward way of assessing its quality. By loading the orientated structure into PyMOL, the predicted transmembrane helices
can be selected and coloured as follows:
select tmh, resi 39-63 + resi 73-96 + resi 109133 + resi 54-173 + resi 203-224 + resi 252-274
+ resi 287-309; color orange, tmh
346
Timothy Nugent
Fig. 4 Transmembrane helices are coloured in orange. It is clear that all seven
helices lie within the membrane plane
It should be fairly clear if any predicted transmembrane helices

lie outside the plane of the membrane (Fig. 4). While it is unreasonable to expect perfect bundles of helices lying perpendicular to
the membrane, excessive tilt angles should be inspected closely.
At this stage, it may be worth regenerating the ensemble using different subsets of candidate structures. For example, the Z-coordinate
distance constraints are only useful in a minority of cases. If the
topology of the final model built from an ensemble which included
Z-coordinate constrained candidates was implausible, it may be
worth generating the ensemble using only the unconstrained candidates. Similarly, the recombination step produced better models
in only 18 out of 28 targets. It is probably wise to inspect the lowest energy candidate model, and compare this with the recombined
structure.
2.8 Managing
Expectation
The example used here, bovine rhodopsin, represents a good target, consisting of a single transmembrane domain with enough
aligned sequences to produce a reliable model. However, other
347
targets may prove more challenging for a number of reasons.

Firstly, if the target membrane domain is part of a multi-domain
protein, it will need to be parsed from the rest of the sequence
before generating alignments, predicting contacts and modelling
using FILM3. A number of tools including DomPred [50] can be
used for this task. Care also needs to be taken when modelling
chains that are part of complexes. In the cases of homo-multimeric
chains, it can be expected that both inter and intrachain contacts
will coevolve, and contact predictors should identify both types.
Currently, we are unable to distinguish between the two, which is
likely to be problematic for methods such as FILM3 that will
attempt to satisfy as many predicted contacts as possible, with the
expectation that they are all intrachain contacts. In general, chains
that are stabilised via interactions with other monomers in a complex tend to produce poor models and should be avoided where
possible. Other targets that are likely to be challenging are those
which undergo significant conformational change upon activation,
for example transporters that adopt distinct alternate conformations, therefore requiring different sets of contacts to stabilise each
state. Again, the FILM3 objective function will attempt to satisfy
such multiple sets of contacts simultaneously, therefore producing
a model that may represent the average of two conformations. On
the other hand, targets which undergo relatively little conformational change upon activation are likely to produce good models.
Conclusions
This chapter should provide a useful introduction to de novo 3D
modelling of alpha-helical membrane protein using predicted contacts. Presented here are a number of powerful tools that, in combination, are capable of generating accurate models of large
transmembrane protein domains. Such models should be particularly useful for directing experimental studies on families where
structural data is unavailable. It is clear that the use of contacts
predicted by methods such as PSICOV provide extremely powerful constraints for de novo modelling, and it is likely that this strategy will become applicable to even more protein families as
sequence databases continue to grow. We estimate that the PFAM
database [51] contains more than 500 single architecture transmembrane domains with >400 aligned sequencesenough to
accurately predict contacts using methods such as PSICOV, plmDCA, or PconsCbut no experimentally determined 3D structure. Applying FILM3 to these families has the potential to
significantly expand our knowledge of transmembrane fold space,
and it is likely that many of these families will be of significant biomedical and pharmacological interest.
348
Timothy Nugent
Notes
1. The TM-score is intended to be a more accurate measure of
structural alignment compared to rmsd or GDT. Scores are in
the range (0, 1], with 1 indicating a perfect match between
two structures, scores below 0.20 typically correspond to randomly chosen unrelated proteins, while scores >0.5 are roughly
the same fold [49].
2. PSICOV can be downloaded from http://bioinfadmin.cs.ucl.
ac.uk/downloads/PSICOV/. Follow the included compilation instructions to build the PSICOV binary.
3. HHblits binaries and source code and accompanying databases
can be downloaded from: http://toolkit.genzentrum.lmu.de/
hhblits/.
4. HMMER binaries and source code can be downloaded from
http://hmmer.janelia.org/.
5. PSIPRED can be downloaded from http://bioinfadmin.
cs.ucl.ac.uk/downloads/psipred/. The NCBI toolkit (ftp://
ftp.ncbi.nih.gov) and PSI-BLAST (ftp://ftp.ncbi.nih.gov/
blast) are also required. Configure the PSIPRED script by
adding the NCBI binary directory and database paths.
Follow the included compilation instructions to build the
PSIPRED binary.
6. Download MEMSAT-SVM from http://bioinfadmin.cs.ucl.
ac.uk/downloads/memsat-svm/, configuring it in exactly the
same way as PSIPRED.
7. FILM3 can be downloaded from http:vbioinfadmin.cs.ucl.
ac.uk/downloads/FILM3/. Compile the three programs as
per the instructions.
8. ProFit can be downloaded from http:vwww.bioinf.org.uk/
software/profit/.
9. Download MODELLER from http:vsalilab.org/modeller/.
You will need to register to receive the license key required to
run it.
10. MEMEMBED can be downloaded from: http://bioinf.cs.ucl.
ac.uk/downloads/memembed/. Follow the included compilation instructions to build the MEMEMBED binary.
References
1. Hopkins AL, Groom CR (2002) The druggable genome. Nat Rev Drug Discov 1:727730
2. Kloppmann E, Punta M, Rost B (2012)
Structural genomics plucks high-hanging
membrane proteins. Curr Opin Struct Biol
22:326332
3. Pieper U, Schlessinger A, Kloppmann E et al

(2013) Coordinating the impact of structural
genomics on the human -helical transmembrane
proteome. Nat Struct Mol Biol 20:135138
4. Kll L, Krogh A, Sonnhammer ELL (2005) An
HMM posterior decoder for sequence feature
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
prediction that includes homology information.

Bioinformatics 21(Suppl 1):i251i257
Bernsel A, Viklund H, Falk J et al (2008)
Prediction of membrane-protein topology
from first principles. Proc Natl Acad Sci U S A
105:71777181
Viklund H, Bernsel A, Skwark M et al (2008)
SPOCTOPUS: a combined predictor of signal
peptides and membrane protein topology.
Viklund H, Elofsson A (2008) OCTOPUS:
improving topology prediction by two-track
ANN-based preference scores and an extended
topological grammar. Bioinformatics 24:
16621668
Bernsel A, Viklund H, Hennerdal A et al
(2009) TOPCONS: consensus prediction of
membrane protein topology. Nucleic Acids Res
37:W465W468
Nugent T, Jones DT (2009) Transmembrane
protein topology prediction using support
vector machines. BMC Bioinformatics
10:159
Nugent T, Jones DT (2012) Detecting porelining regions in transmembrane protein
sequences. BMC Bioinformatics 13:169
Kelm S, Shi J, Deane CM (2010) MEDELLER:
homology-based coordinate generation for
membrane proteins. Bioinformatics 26:
28332840
Hill JR, Deane CM (2013) MP-T: improving
membrane protein alignment for structure prediction. Bioinformatics 29:5461
Olivella M, Gonzalez A, Pardo L et al (2013)
Relation between sequence and structure in
15891592
Pellegrini-Calace M, Carotti A, Jones DT (2003)
Folding in lipid membranes (FILM): a novel
method for the prediction of small membrane
protein 3D structures. Proteins 50:537545
Yarov-Yarovoy V, Schonbrun J, Baker D (2006)
Multipass membrane protein structure prediction using Rosetta. Proteins 62:10101025
Barth P, Wallner B, Baker D (2009) Prediction
of membrane protein structures with complex
topologies using limited constraints. Proc Natl
Acad Sci U S A 106:14091414
Barth P, Schonbrun J, Baker D (2007) Toward
high-resolution prediction and design of transmembrane helical protein structures. Proc Natl
Acad Sci U S A 104:1568215687
Jones DT (1997) Successful ab initio prediction of the tertiary structure of NK-lysin
using multiple sequences and recognized
supersecondary structural motifs. Proteins
Suppl 1:185191
349
19. Jones DT, McGuffin LJ (2003) Assembling

novel protein folds from super-secondary structural fragments. Proteins 53(Suppl 6):480485
20. Hurwitz N, Pellegrini-Calace M, Jones DT
(2006) Towards genome-scale structure prediction for transmembrane proteins. Phil Trans
Roy Soc Lond B Biol Sci 361:465475
21. Simons KT, Bonneau R, Ruczinski I et al
(1999) Ab initio protein structure prediction
of CASP III targets using ROSETTA. Proteins
Suppl 3:171176
22. Rohl CA, Strauss CEM, Misura KMS et al
(2004) Protein structure prediction using
Rosetta. Methods Enzymol 383:6693
23. Fuchs A, Kirschner A, Frishman D (2009)
Prediction of helix-helix contacts and interacting helices in polytopic membrane proteins
using neural networks. Proteins 74:857871
24. Lo A, Chiu Y-Y, Rdland EA et al (2009)
Predicting helix-helix interactions from residue
contacts in membrane proteins. Bioinformatics
25:9961003
25. Nugent T, Jones DT (2010) Predicting transmembrane helix packing arrangements using
residue contacts and a force-directed algorithm. PLoS Comput Biol 6:e1000714
26. Sadowski MI, Taylor WR (2013) Prediction of
protein contacts from correlated sequence substitutions. Sci Prog 96:3342
27. Lapedes A, Giraud B, Liu L et al (1999)
Correlated mutations in models of protein
sequences: phylogenetic and structural effects.
Stat Mol Biol Genet 33:236256
28. Weigt M, White RA, Szurmant H et al (2008)
Identification of direct residue contacts in proteinprotein interaction by message passing.
Proc Natl Acad Sci U S A 106(1):6772
29. Taylor WR, Sadowski MI (2011) Structural
constraints on the covariance matrix derived
from multiple aligned protein sequences. PLoS
One 6:e28265
30. Friedman J, Hastie T, Tibshirani R (2008)
Sparse inverse covariance estimation with the
graphical lasso. Biostatistics 9:432441
31. Jones DT, Buchan DWA, Cozzetto D et al
(2012) PSICOV: precise structural contact
prediction using sparse inverse covariance estimation on large multiple sequence alignments.
32. Taylor WR, Jones DT, Sadowski MI (2012)
Protein topology from predicted residue contacts. Protein Sci 21:299305
33. Ekeberg M, Lvkvist C, Lan Y et al (2013)
Improved contact prediction in proteins:
using pseudolikelihoods to infer Potts models.
Phys Rev E Stat Nonlin Soft Matter Phys
87:012707
350
Timothy Nugent
34. Skwark MJ, Abdel-Rehim A, Elofsson A (2013)

PconsC: combination of direct information
methods and alignments improves contact prediction. Bioinformatics 29:18151816
35. Marks DS, Colwell LJ, Sheridan R et al (2011)
Protein 3D structure computed from evolutionary sequence variation. PLoS One
6:e28766
36. Hopf TA, Colwell LJ, Sheridan R et al (2012)
Three-dimensional structures of membrane
proteins from genomic sequencing. Cell 149:
16071621
37. Nugent T, Jones DT (2012) Accurate de novo
structure prediction of large transmembrane
protein domains using fragment-assembly and
correlated mutation analysis. Proc Natl Acad
Sci U S A 109:E1540E1547
38. Remmert M, Biegert A, Hauser A et al (2011)
HHblits: lightning-fast iterative protein
sequence searching by HMM-HMM alignment. Nat Methods 9:173175
39. Magrane M, Consortium U (2011) UniProt
knowledgebase: a hub of integrated protein
data. J Biol Databases Curat, Database, p 2011
40. Finn RD, Clements J, Eddy SR (2011)
HMMER web server: interactive sequence
similarity searching. Nucleic Acids Res
39:W29W37
41. Granseth E, Viklund H, Elofsson A (2006)
ZPRED: predicting the distance to the
membrane center for residues in -helical
e191e196
42. Mart-Renom MA, Stuart AC, Fiser A et al
(2000) Comparative protein structure model-
43.
44.
45.
46.
47.
48.
49.
50.
51.
ing of genes and genomes. Annu Rev Biophys

Biomol Struct 29:291325
Sali A, Blundell TL (1993) Comparative protein modelling by satisfaction of spatial
Nugent T, Jones DT (2013) Membrane protein orientation and refinement using a
knowledge-based statistical potential. BMC
Schrdinger L (2010) The PyMOL molecular
graphics system, version 1.3r1
Tusndy GE, Dosztnyi Z, Simon I (2005)
TMDET: web server for detecting transmembrane regions of proteins by using their 3D
coordinates. Bioinformatics 21:12761277
Senes A, Chadi DC, Law PB et al (2007) E(z), a
depth-dependent potential for assessing the
energies of insertion of amino acid side-chains
into membranes: derivation and applications to
determining the orientation of transmembrane
and interfacial helices. J Mol Biol 366:436448
Lomize MA, Pogozheva ID, Joo H et al (2012)
OPM database and PPM web server: resources
for positioning of proteins in membranes.
Nucleic Acids Res 40:D370D376
Xu J, Zhang Y (2010) How significant is a protein structure similarity with TM-score = 0.5?
Bryson K, Cozzetto D, Jones DT (2007)
Computer-assisted protein domain boundary
prediction using the DomPred server. Curr
Protein Pept Sci 8:181188
Punta M, Coggill PC, Eberhardt RY et al
(2012) The Pfam protein families database.
Nucleic Acids Res 40:D290D301
Chapter 16
NMR-Based Modeling and Refinement
of Protein 3D Structures
Wim F. Vranken, Geerten W. Vuister,
and Alexandre M.J.J. Bonvin
Abstract
NMR is a well-established method to characterize the structure and dynamics of biomolecules in solution.
High-quality structures can now be produced thanks to both experimental advances and computational
developments that incorporate new NMR parameters and improved protocols and force fields in the structure calculation and refinement process. In this chapter, we give a short overview of the various types of
NMR data that can provide structural information, and then focus on the structure calculation methodology itself. We discuss and illustrate with tutorial examples classical structure calculation, refinement, and
structure validation approaches.
Key words NMR, Structure calculation, Structure refinement, Structure validation
Introduction
The first step of a structure determination by NMR spectroscopy
consists of the acquisition of NMR data, typically using heteronuclear multidimensional experiments, that allow the assignment
of all atoms/spins of a molecule (1H, 15N, 13C) to their chemical
shift values (Fig. 1). Once this chemical shift assignment step is
completed, 13C- and 15N-edited 3D NOESY spectra are generally
used to obtain inter-atomic distances from nuclear Overhauser
effects (NOE). These NOESY spectra provide the most detailed
structural information that can be obtained from NMR and are
still the most common core data used to define the 3D structure of
the protein [1, 2]. In addition to distance information, other
parameters, such as J-couplings [3], residual dipolar couplings
(RDCs) [4], paramagnetic relaxation enhancements (PRE) and
pseudo-contact shifts [5] can be measured, providing additional
information to define the protein structure. Recent developments
have enabled the calculation of the structures of relatively small
351
352
Wim F. Vranken et al.
Fig. 1 Schematic overview of the structure calculation process in NMR. This chapter deals with the boxes in
black, with indication of the software covered
proteins (less than ~12 kDa) from chemical shift values alone [68],
a procedure that will be briefly outlined in this paper.
The experimental NMR parameters are converted to in silico
restraints and 3D structures are generated from restrained molecular dynamics simulations following usually some form of molecular
dynamics simulated annealing scheme (MD/SA) [9]. Multiple
structures are calculated in this way, starting from the same experimental data but different random starting conditions. Provided
that enough data of sufficient quality are available, the structures
will converge onto the same overall fold. These structures are nowadays often further refined in explicit solvent (water), which has
been shown to significantly improve their quality [10, 11]. Finally,
the structures that best satisfy as many experimental restraints as
possible, together with proper general chemical properties of proteins (such as bond lengths and angles), are then selected to form
an ensemble of structures that represents the definitive solution of
the structure calculation process.
In this chapter we will discuss the classical NMR structure
calculation, refinement, and validation methods, with some reference to new chemical shift-based approaches. These will be illustrated with tutorial examples making use of the programs CYANA
NMR 3D Structure Modelling of Proteins
353
[12] and CNS [13, 14] using the RECOORD [11] approach for
water refinement [10], and the CS-ROSETTA protocol [7] based
on chemical shift data. These are followed by a description of
structure validation with the program CING [15].
Theory
This section gives a very brief overview of the NMR data relevant
in the determination of 3D protein structures.
2.1 NMR Structural

Information Sources
Several NMR parameters providing structural information can be

measured for use in structure calculations and refinement; these
are briefly described here.
2.1.1 Chemical Shifts
The first step in analyzing NMR experiments is to assign the

observed resonances, each of which has a particular chemical shift
value, to atoms in the molecule being studied. Although chemical
shifts are very sensitive probes of the chemical environment of a
spin, their dependency on the 3D structure is complex and their
usage in 3D structure calculation is still evolving. A common use is
to determine per-amino acid residue secondary structure preferences using the deviations of actual chemical shifts values from
those of random coil peptides. This approach can be used to restrict
the local conformation of a residue to a given region of the
Ramachandran plot, either through torsion angle restraints [16] or
by special database potential functions [17]. The last years have
also seen the development of methods that can calculate protein
structures from chemical shifts only [68]. In those, the chemical
shifts are typically used to select peptide fragments that are then
assembled to produce a 3D model. One of these, CS-ROSETTA
[7], will be briefly illustrated in this chapter.
Many software programs are now available that perform the
reverse operation, i.e. predict chemical shifts from in silico 3D
structures (ShiftX [18], ShiftX2 [19], SHIFTS [20], SHIFTCALC
[21], PROSHIFT [22] and SPARTA+ [23]). They can be used for
example for structure validation.
2.1.2
Classical protein structure determination by NMR relies on obtaining a dense network of distance restraints derived from nuclear
Overhauser effects (NOEs) between nearby hydrogen atoms in a
protein [1, 2]. Together, these restraints provide the essential
information for defining the tertiary structure of a protein.
The NOE originates from cross-relaxation between dipolar
coupled spins as a result of through-space spinspin interactions
that result in the transfer of magnetization from one spin to
another. The NOE approximately scales with the distance r between
the two spins as 1/r6. Because of this 1/r6 dependency, NOEs are
NOEs
354
only detected between protons less than 56 away in space and

are thus strongly biased towards the shorter distances. In addition,
they are sensitive to dynamics averaging and suffer from so-called
spin diffusion effects. Nevertheless, the large number of potential
NOE-derived restraints renders them very valuable when treated
appropriately.
2.1.3 J-Couplings
Scalar or J-couplings are mediated through chemical bonds

connecting two spins. Particularly informative are the vicinal 3J scalar coupling constants between atoms separated by three covalent
bonds, which are correlated to the enclosed torsion angle, , by
the empirical Karplus equation [24]. In particular, 3J(HNH)
and 3J(HH) give information about the backbone -angle and
the side-chain 1 angle in an amino acid, respectively. For amino
acid with diastereotopic protons (H2/H3), the use of 3J(H
H) coupling does require their stereospecific assignments.
In contrast to the NOEs, scalar coupling constants only provide
information on the local conformation of a polypeptide chain.
J-couplings are commonly converted into dihedral angle restraints
[1] or directly used as J-coupling restraints [25, 26] in NMR structure calculations. Nowadays, the less accurate dihedral angle
restraints predicted from chemical shift values are in more common use.
2.1.4 Hydrogen Bonds
Some hydrogens, such as those present in the backbone amide

groups, can chemically exchange with hydrogens from the water
(H2O) solvent. Experiments where the H2O solvent is replaced by
the protonNMR-inactive D2O can determine which of these
hydrogens only exchange slowly; such hydrogens are assumed to
be protected from the solvent and/or involved in a hydrogen bond
[27]. Identification of the acceptor atom requires either additional
experimentation or implicit assumptions. The latter is based on
NOEs and/or the regularity of secondary structures and the
inferred hydrogen bond restraints should be used with caution.
Hydrogen bonds can also be detected directly from cross-hydrogen
bond scalar coupling measured from constant time HNCO spectra
[28, 29], thus providing accurate restraints for structure calculations [30]. Hydrogen bond restraints are usually introduced into
the structure calculation protocol as distance restraints, typically by
confining the donor-hydrogen/acceptor distance to a given range.
2.1.5 Residual Dipolar

Couplings
Residual dipolar couplings (RDCs) are now a well-established

source of structural information [31, 32]. They can be measured in
solution by weakly aligning the molecule using a variety of methods [33]. RDCs provide orientational information of the internuclear vector of the two atoms for which the RDC is measured
relative to three globally defined axes in the molecule, i.e. those of
the alignment tensor. Note that if a RDC is measured between two
355
atoms that are not at a fixed distance from each other, there is also
a distance dependence and hence usually only RDCs measured for
inter-nuclear vectors with a fixed distance are used in the structure
calculations. Residual dipolar couplings can be added as orientational restraints to the target function of the structure calculation
algorithm [34].
2.1.6 Diffusion
Anisotropy
For non-isotropical tumbling molecules, NMR relaxation data

contain orientational information comparable to RDCs as result of
the diffusion anisotropy [35]. NMR relaxation is characterized by
relaxation times T1 and T2, and the T1/T2 ratio can be used to
define diffusion anisotropy restraints in NMR structure calculations [36]. Again the orientation information comes from the
angles of inter-nuclear vectors in an external frame, which, in the
case of diffusion anisotropy data, corresponds to the orientational
diffusion tensor frame. In practice, 15N T1 and T2 relaxation data
are most often used.
2.1.7 Paramagnetic
Relaxation Effects
If a paramagnetic metal ion is present in a protein, or if it is introduced via for example a chelating agent chemically bound to the
protein, the NMR signals of the nuclei in a shell around it will be
affected [37] by several effects including contact and pseudocontact shifts, relaxation rate enhancements, and cross-correlation
effects. Analogously to RDC and diffusion anisotropy, these,
depending on their type, can provide both distance and orientation
information which can be converted into restraints to be used in
various structure calculation softwares [38, 39].
2.2 Structure
Calculation Software
The experimental information sources discussed above can be converted into restraints that can be used in the structure calculation
process. Several computer programs have provisions for using the
experimental NMR restraints; the most commonly used ones are
CNS [13, 14], Xplor-NIH [40] and CYANA [12, 41], although
many others are available, e.g. SCULPTOR [42], the SANDER
module of AMBER [43], GROMACS [44] and YASARA [45].
Structure calculations in essence transform the experimental
data (as restraints) into in silico atomic coordinate information. The
calculations are usually based on some molecular dynamic simulated annealing protocol performed in torsion angle and/or
Cartesian space, followed by a final refinement phase in explicit solvent (water). A general feature of all these protocols is the usage of
a target function: lower values of this function for a calculated
structure indicates better agreement with the experimental data and
with known molecular information. This molecular information is
defined by a force field that contains physical energy terms for interactions such as van der Waals interactions and electrostatics, as well
as terms describing the molecular geometry such as bond lengths,
bond angles, etc. During the initial stages of a structure calculation
356
often the description of some of these terms is simplified to increase

the computational speed and/or simplify the energy landscape to
be search. For example, long-range nonbonded interactions are
reduced to only repulsions between atoms and electrostatic interactions are at first neglected. A full nonbonded representation, including van der Waals (Lennard-Jones) and electrostatic (Coulomb)
interactions, is then typically reintroduced for final refinement in
explicit solvent [10].
2.3 Structure
Selection
and Structural Quality
Typically a large pool of structures is generated during the structure calculation process, from which a final ensemble of best
structures is then selected. This choice of an ensemble of structures, rather than one single one, reflects the uncertainty in the
experimental NMR data: often structures that agree with the
experimental data equally well but differ locally (such as in loop
regions) can be obtained. The most widely used structure selection
procedure is based on the agreement with the experimental data
(rather arbitrarily defined as a small number of restraint violations)
and a low (overall) energy of the structures. Typically ensembles
containing the 20 lowest energy models are selected, although this
number is arbitrary. Ideally, the selected ensemble should represent
the available conformational space accessible to the structure while
simultaneously satisfying the experimental restraints. From this
ensemble, a representative structure is usually defined; no real consensus exists, however, on how it should be selected. The wwPDB
NMR validation taskforce recommends to select the structure that
differs the least from all other structures within the ensemble, i.e.
the mediod (the structure with the lowest atom coordinate RMSD
from all other structures)[46].
The final ensemble is subsequently subjected to structure validation procedures in order to verify its quality. It is useful to distinguish the well-defined from the ill-defined regions of the ensemble
during this process: these can be defined with, for example,
CYRANGE [47], FindCore [48] or circular variance methods
[49]. In practice, the quality indicators that are most commonly
used to assess especially the well-defined regions of an NMR
ensemble are [50]:
the goodness of fit to the experimental data, by analyzing

restraint violations;
the precision of the ensemble, measured by positional root

mean square deviation (RMSD);
several physical and stereochemical quality indicators that

assess the local and overall quality of protein structures, many
of them based on knowledge from high-resolution X-ray
structures.
Table 1 lists the most commonly used validation programs;

some of which will be described later.
357
Table 1
Internet resources of NMR-related programs and databases mentioned in this chapter
Software
Internet address
Purpose
CNS
http://cns.csb.yale.edu/v1.3
Multilevel hierachical approach for the most

commonly used algorithms in
macromolecular structure determination
(NMR, crystallography)
RECOORD
http://www.ebi.ac.uk/pdbeapps/nmr/recoord/
Database of recalculated NMR structures

with the CNS scripts used in the tutorial
example
PDBe
http://www.pdbe.org/
An information portal to biological

macromolecules structures
BMRB
http://www.bmrb.wisc.edu
Biological magnetic resonance data bank
CCPN
http://www.ccpn.ac.uk
A collaborative computing project for NMR
CING
http://nmr.cmbi.ru.nl/icing/
A server to validate the quality of NMR

structures
PSVS
http://psvs-1_3.nesg.org/
A server to validate the quality of NMR

structures
TALOS+
http://spin.niddk.nih.gov/bax/
nmrserver/talos/
Protein backbone angle restraints from

searching a database for chemical shift
and sequence homology
CYANA
http://www.cyana.org
Structure calculation program (paid license

required)
Below accessible via WeNMR framework

CS-ROSETTA
http://haddock.science.uu.nl/
enmr/services/CS-ROSETTA3
Structure calculation program using only

chemical shift information
TALOS+
http://haddock.science.uu.nl/
enmr/services/TALOS
Protein backbone angle restraints from

searching a database for chemical shift
and sequence homology
CYANA
http://www.enmr.eu/webportal/
cyana.html
Structure calculation program (paid license

required)
FormatConverter
http://haddock.chem.uu.nl/
enmr/format-converter.html
Conversion between file formats (this web

version limited compared to local install)
Methods
In this tutorial section we describe the procedures to generate various types of NMR restraints and how they can then be used in
structure calculations using CYANA or CNS with the RECOORD
scripts. The CS-ROSETTA chemical shift-only approach is also
discussed, followed by a description of structure validation using
the CING webserver. A classical, NOE distance-based structure
358
determination is recommended when possible since it typically

leads to more accurate structures. The protocol to follow is however more complex than that of chemical shift-based methods like
CS-ROSETTA. The latter is very simple to use and has been shown
to generate accurate models for small systems [7, 51]. The
commands to be executed are in bold Courier font with gray shading. Information about the various programs and web pages used
in this tutorial can be found in Table 1. As an example project for
these tutorials we provide a target from the CASD-NMR experiment [52, 53], OR36 (pdb 2LCI). This protein is 128 amino acids
long. Besides backbone chemical shift data, NOE data are also provided that can be used in the other tutorials.
3.1 NMR Data
and File Formats
A common problem in NMR is the abundance of file formats to

store the NMR data (such as spectral data, chemical shifts, and
peaks lists) and resulting restraint files for structure calculation.
This makes data exchange between different programs a tedious
process. The CCPN Data Model for macromolecular NMR [54] is
intended to cover all data needed for macromolecular NMR spectroscopy from the initial experimental data to the final validation.
The ccpNmr FormatConverter application allows consistent
conversion between a large variety of data formats via import to
and export from CCPN, and we recommend its local installation
(see Note 1) and use for this purpose. The general steps to follow
for importing data files into the FormatConverter from scratch are:
1. Start the FormatConverter on the command line:
formatConverter
2. Under the Project menu option, click New and enter a name
for your project; you will not be able to use the FormatConverter
until you have done this.
3. Under the Import menu option, go to Single files, select the
type of data you want to import, then the name of the software
the data comes from. A window will pop up, click on the Select
file button, navigate to the file, and click Select. Click
IMPORT to start the importing process.
4. Depending on the type of data, you might get popups that ask
you to validate information or provide additional input. Press
on the ? button in these popups for additional information.
5. After successful import, repeat the process from step 3. Due to
the way the CCPN framework stores information, you will
need to import both the sequence of your molecule and experimental information to create a complete CCPN project. At
this point you will get a prompt to initiate the linkResonances
process. Click Yes when asked, and Yes again to perfom this
359
process automatically (if possible). For more information on

this process, see the links in Note 1.
6. Under the Project menu option, click Save or Save as to store
your project. The data will be saved in a directory with subdirectories containing a set of XML files; all these have to stay
together to remain valid for the CCPN framework.
If you created a CCPN project or downloaded a publicly available one (see Note 1) you can export data using the following
general steps:
1. Under the Project menu option, click Open, navigate to the
main directory containing your CCPN project in the pop-up
window, and click Open.
2. Under the Export menu option, select the format/software to
which you want to export the data. In the pop-up window,
select the type of data you want to export to file, then select
the specific data item(s) from the CCPN project. Click on the
Select export file button, identify the file name you want to
export the data to, and click Select. Click Export file to
start the export process.
3. Depending on the type of format and data, you might get popups that ask you to validate information or provide additional
input. Press on the ? button in these popups for additional
information.
3.2 TALOS+:
Chemical ShiftDerived Dihedral Angle
Restraints
J-coupling and secondary chemical shifts can be used to define

restraints on the torsion angles of the chemical bonds, typically ,
, and 1 angles can be generated and included. They can be calculated applying the Karplus equation [24] to the measured
J-couplings, or estimated from chemical shifts in programs such as
TALOS+ [55], DANGLE [56] or CSI [57]. The chemical shiftbased methods are now the most commonly used and have good
accuracy; DANGLE can be run from the CCPN software, but we
will describe here the more commonly used TALOS+ approach.
TALOS+ predicts and backbone torsion angles from a
combination of available chemical shifts (H, C, C, CO, N) for a
protein sequence; it is an empirical method supported by a database
of existing information on chemical shifts and backbone torsion
angles. Before executing TALOS+, it is essential to ensure that your
chemical shifts are correctly referenced (see Note 2), and that you
have the correct input file (see Note 3). We here describe the use of
TALOS + from the web-based server; other options are available
(local installation or the WeNMR server, see Subheading 3.3).
1. Connect to the TALOS+ Web server:
http://spin.niddk.nih.gov/bax/nmrserver/talos.
2. Under the Chemical Shift Input tab, click Browse and select
your chemical shift input file (see Note 3).
360
3. Under the Prediction Options tab, you can deselect the Apply
Offset Correction tick mark if you are sure your chemical
shifts are correctly referenced (see Note 2).
4. Under the Submission Details tab, enter your email address
(twice) and click Submit to start TALOS+. The screen will
change, and you will after a few minutes receive an email from
the NMR Server Agent with details on where to retrieve the
results.
5. Create a directory where you want to store the results, open
the email, and save all six files to this directory (detailed information on these files is available from the TALOS+ server).
6. In order to convert the TALOS+ predictions to accurate dihedral angle restraints, they have to be manually inspected.
Connect to the jRAMA+ online viewer: http://spin.niddk.
nih.gov/bax/software/TALOS+/JRAMA+.
The first time you access this page you have to explicitly state
that you trust this server; click the tick mark box followed by Run.
7. Click on the TALOS+ image in the top left corner; a pop-up
will appear. In this popup, click on the File menu option (top
left) and select Open Prediction Files. In the file window that
appears, navigate to the directory you created to store the
TALOS+ results, and select the pred.tab file. A set of other
windows will now appear; the RCI-S2 and Secondary Structure
Plot gives an overview of the chemical shift predicted backbone
dynamics (from RCI [58]) and secondary structure for your
protein.
8. Go to the window that contains your protein sequence with a
color-coded box for each amino acid. The color coding indicates the following:
(a) Green: the prediction is Good for this residue and can
be used to create a dihedral angle restraint.
(b) Yellow: the prediction is Ambiguous and should be
manually inspected before use.
(c) Red: the prediction is Bad and should not be used.
(d) Blue: the residue is highly Dynamic and a prediction is
not possible.
If you click on a box, the other windows will update to show
detailed information about the prediction for this residue, and
you can examine the / distributions of the detected matches
and override the TALOS+ decisions on which residues should
be included in the prediction and which ones are outliers. You
can change the prediction status of this residue selecting a box
at the bottom of the main windowonly residues with
Good status will be written out the dihedral restraint file as
discussed in the next point.
361
9. When you are satisfied that only well-predicted residues have

Good status, you can export the dihedral restraints from
the Tools menu option. Write out both Xplor and Cyana
angle restraint files.
3.3 CS-ROSETTA:
Chemical Shift-Based
Structure Calculation
The CS-ROSETTA3 protocol [51] makes use of the original

CS-ROSETTA protocol [7], using a new fragment selection
method (Vernont et al. personal communication) and Rosetta 3.3,
and is powered by the WeNMR Grid [59]; see Note 4 for more
information on the WeNMR project and how you can access its
resources. This tutorial guides you through the steps for setting up
a basic CS-ROSETTA run, and how to interpret the results. First
create a working directory, download the supplementary data associated with this chapter from the Springer extra Website at http://
extras.springer.com, and unpack it with the following commands:
mkdir OR36
cd OR36
tar xvfz OR36_blind.tar.gz
Before submitting your chemical shift list to CS-ROSETTA, it

is recommended to check the flexibility of your protein. If there are
flexible ends, we suggest you truncate them from your chemical
shift list. CS-ROSETTA predicts structure, so performs best for
parts of a protein that have structure. To identify flexible residues,
run the TALOS+ server as described before or via the WeNMR
infrastructure:
1. Connect to the WeNMR TALOS+ Web server:
http://haddock.science.uu.nl/enmr/services/TALOS.
2. Give a name to your run.
3. Select as file type to submit: BMRB3.1.
4. Select as chemical shift file to submit the originalData/
bmrb31.str file from the OR36 directory you created
previously.
5. Click on Submit.
6. You will be presented with a link to the result page. The result
page for this system (Fig. 2) shows the predicted phi and psi
angles for the residues, and the uncertainty in the prediction.
If the respective bar is colored green, then the prediction is
reliable. If the bar is red, the prediction is ambiguous. If the
bar is absent, no prediction is made, or the respective residue is
dynamic. Several consecutive dynamic residues suggest the
presence of flexible parts. TALOS+ does not predict the torsion angles for the first and last residue in the sequence. Our
target in this tutorial case seems thus to be ordered (i.e. does
362
Fig. 2 Example of TALOS output from the WeNMR server based on the chemical shifts for entry 2lci. (The server
is accessible via the NMR ->Chemical Shifts menu of www.wenmr.eu)
363
not seem to contain flexible ends). The last Histidines in the

sequence are not predicted because no chemical shifts are available for those.
7. Before proceeding save the talos.tab file provided on the
results page. We will use it for the CS-ROSETTA server
submission. Edit this file and remove from the sequence the
last six histidines since no chemical shifts are available for them.
Also delete the chemical shift entries for residue 131 (one of
the last histidines). It does not make sense to include them for
CS-ROSETTA calculations.
Now set up and run the CS-ROSETTA calculations:
1. Connect to the WeNMR CS-Rosetta3 web portal:
h t t p : // h a d d o c k . s c i e n c e . u u . n l / e n m r / s e r v i c e s /
CS-ROSETTA3.
2. Figure 3 shows the web form you have to fill in. There are
seven fields you can fill in:
(a) The run name: has to be unique, has to be alpha numerical, and maximally 20 characters long.
(b) The data format of the chemical shift list: you can supply
your chemical shift list in three data formats: TALOS,
BMRB2.1, and BMRB3.1.
(c) The chemical shift list: the location of the chemical shift
file on your local computer.
(d) The number of models to generate: 10,000 is the maximum for a default account.
(e) Option to automatically exclude flexible tails: we suggest
you to do this manually as explained above.
(f) Rescoring options: After generating your models, they
have to be scored. ROSETTA3.3 does this with an all atom
energy score (raw score). There are two additional rescoring algorithms: chemical shift (CS) rescoring, and DP
rescoring (based on unassigned NOE peak lists) [60]. For
DP rescoring different types of NOE data can be supplied
(outside the scope of this tutorial).
3. To set up your tutorial run, choose the following options for
the above:
(a) Run name: OR36.
(b) Type of file to submit: TALOS.
(c) Chemical shift list: use the talos.tab file you just
saved and edited.
(d) Leave the number of models to their default value.
364
Fig. 3 CS-ROSETTA web form on the WeNMR server. (The server is accessible via the NMR -> Structure
Calculation menu of www.wenmr.eu)
(e) Leave the remove flexible parts box unticked.

(f) Chemical shift rescoring is selected by default; leave this.
(g) Enter your username and password and submit the
calculations.
365
After a successful submission you get a link to your personal

result page (which is also mailed to you), on which you can track
your run, and find the results when the run will have completed
(which can take a few days depending on the system size and load
on the server). The result page present the top five models based
on various scoring scheme and also indicates the reliability of the
predictions based on convergence and energetics criteria.
You can directly validate the CS-ROSETTA calculation results
using iCING (see Subheading 3.6).
3.4 CYANA Structure
Calculation
from NOESY Peak Lists
A cross-peak in a NOESY spectrum indicates spatial proximity

between two atoms; under idealized asumptions the intensity of
the NOE peak (I) is proportional to the interatomic distance (r) as
I ~ 1/r6. Each NOESY peak can therefore be converted into an
interatomic distance after determination of the proportionality
constant (a process called distance calibration). Because the
intensitydistance relationship is only approximate in practice, due
to multiple complicating effects such as dynamics and the presence
of many interacting spins, and as result of limitations in most structure calculation protocols, a distance range is usually assumed
when converting the NOE peak intensity into a distance restraint.
This range tends to be between 1.8 and 6.0 angstroms, with various ranges depending on the intensity of the NOE peaks.
An NOE peak can only be used as a distance restraint if it can
be assigned to atoms in the molecule based on their chemical shift
values: this step is of crucial importance, as the structure calculation process may not be able to deal with too many wrong or
highly ambiguous (when many atoms with similar chemical shift
values can be assigned to an NOE) assignments. The manual
assignment of NOEs is an intensive and time-consuming job: many
protocols are available to do this automatically starting from spectra [61] or peak lists [12, 62, 63]. We will here describe the procedure starting from peak lists as implemented in CYANA 2.1: note
that CYANA requires a license, but because of its speed and ease of
implementation it is in our view the most convenient software to
use at this stage.
1. Create a directory for the structure calculation with CYANA.
The following files have to be present in this directory:
(a) A sequence file with the protein amino acid sequence in
XEASY/CYANA format (see Note 5). We will refer to this
file as protein.seq, and assume it has 100 residues.
(b) A chemical shift file in XEASY/CYANA format (see Note
5). We will refer to this file as shifts.prot.
(c) A dihedral angle file (e.g. from TALOS+ in Subheading 3.2).
We will refer to this file as dihedrals.aco.
366
(d) One or more peak lists from NOESY spectra in XEASY/

CYANA format (see Note 5). We will refer to these files as
peaks1.xpk, peaks2.xpk, .
(e) A CYANA initialization file (init.cya) containing the lines:
rmsdrange:=1..100
cyanalib
read seq protein.seq
Note that you can adjust the RMSD range to only include
well-defined regions of the protein structure (if this information is known).
(f) A CYANA file with information for the structure calculation run (AUTO.cya) containing the lines:
peaks
prot
constraints
tolerance
structures
steps
randomseed
:=
:=
:=
:=
:=
:=
:=
peaks1.xpk,peaks2.xpk
shifts.prot
dihedrals.aco
0.030,0.040,0.25
100,20
10000
34983434
#
#
#
#
#
#
#
NOESY peak lists

Chemical shifts
Restraints
Tolerances for assignment
Initial/final structures
Calculation steps
Calculation seed
noeassign peaks=$peaks prot=$prot autoaco
The tolerances determine which atoms (based on their individual chemical shift values) are assigned to the NOESY peaks;
wider tolerances will result in CYANA detecting more possible
assignments for the NOESY peaks. If the tolerances are too narrow, less assignment possibilities will be found, but correct
assignments might be missed. The above protocol assumes that
the NOESY peaks are not yet assigned; additional protocols are
available from the CYANA Web site and other resources.
2. Start CYANA with the command:
cyana
The program should automatically pick up all the relevant files

and will assign the peak lists and calculate structures based on
its internal procedures.
3. If CYANA successfully finished, you will see a set of files with
final in the name that contain the following information:
(a) final.pdb: The atom coordinates of all models in the
final structure ensemble.
367
(b) final.upl: An overview file containing information on

the structure calculation. Particularly relevant here is the
list of violated constraints: if they are violated in most
structures (full sequence of + or * signs) you should go
back to the original peak lists with the program you were
using for spectrum analysis and check whether the NOE
peak and/or assignments make sense.
(c) final.aco, final.upl: Constraint files used in final calculation
round.
(d) protein-final.prot: An XEASY file containing the final
chemical shift assignments.
You should also check for *** WARNING messages in
the AUTO.out file to make sure there were no problems during the calculations. Finally, the files ending in cycle7.peaks
and cycle7-ref.peaks contain the original peak lists with information adapted based on final information from the structure
calculation.
4. You can directly validate the CYANA calculation results using
iCING (see Subheading 3.6). As CYANA uses a limited force
field, we also recommend to water-refine the final structures
with the protocol from Subheading 3.5 (see Notes 5 and 7 on
how to generate the CNS input via CCPN).
3.5 NMR Structure
Refinement with CNS/
RECOORD
For the structure calculation part we are going to describe the use
of the program CNS [13, 14] with a simulated annealing protocol
derived from ARIA [64] followed by refinement in explicit solvent
[10]. All the scripts mentioned in this section can be downloaded
from the RECOORD [11] Webpage (see Table 1).
1. Download: Create a folder where you will run the calculations, download there the tar file containing the RECOORD
scripts and decompress it:
mkdir struct-calc
cd struct-calc/
wget http://www.ebi.ac.uk/pdbeapps/nmr/data/recoord/RECOORDscripts-cns1.3.tgz
tar xzfv RECOORDscripts-cns1.3.tgz
In case the wget command does not work, use a Web browser
to download the scripts manually from the RECOORD webpage (see Table 1). This tutorial uses CNS version 1.3.
2. Initialise: Before starting the calculations, you need to set up
your current path for the scripts to work. In order to do this,
you need to edit the file changeScriptsDir.sh located
in RECOORDscripts-cns1.3 and change the directory path for
newDir in line 8 with your current path and execute it:
368
./changeScriptsDir.sh
3. Get data: Make sure you have a working version of CNS set
up; the last step is then to create a working directory and
assigning a project name for the protein you are working on.
This project name will be used to generate the file names at the
different stages of the protocol. We will use as example the
OR36 data for the PDB 2LCI structure with the corresponding NMR restraints available for this entry from the
BioMagResBank (BMRB) [65]. First download and rename
the PDB structure file:
mkdir 2lci
cd 2lci
wget http://www.ebi.ac.uk/pdbe-srv/view/files/2lci.ent.gz
gunzip pdb2lci.ent.gz
mv pdb2lci.ent 2lci.pdb
Then get the NMR restraints from this link:
http://www.bmrb.wisc.edu/servlets/MRGridServlet?pdb_id=2l
ci&min_items=0&block_text_type=3-converted-DOCR
In the result table, click on the number in the distance row
under the XPLOR/CNS column, then click on the link in
the mrblock_id column. Copy and paste these restraints in a
text file called unambig.tbl in the 2lci/ directory (see Note 6).
Alternatively you can export CNS restraints from CCPN projects (see Note 7) or use the data for OR36 directly from the
restraints/cns/ directory of the example project.
4. Generation of molecular topology files: We can generate the
molecular topology either from the primary sequence or from
a PDB coordinate file, depending on availability (see Note 8).
We will use here the downloaded PDB file:
../RECOORDscripts-cns1.3/generate.sh 2lci.pdb
A topology file called 2lci_cns.mtf will be generated (note

that this name is based on the name of the input file and will
be different when you try other examples). You should check
the ERRORS_generate file created inside the 2lci/ folder:
in this particular case, you can see that the script reported
many nomenclature errors which can be ignored at this stage.
369
A new pdb file called 2lci_cns.pdb is also generated with

the proper CNS nomenclature. You can display this structure
in your favorite molecule viewer to verify it is correct.
5. Generation of extended starting structure: The next step is
the generation of an extended starting conformation which
will be used as input in the simulated annealing protocol:
../RECOORDscripts-cns1.3/generate_extended.sh 2lci_cns.mtf
The extended structure is in the file 2lci_cns_extended.
pdb. You should check the ERRORS_generate_extended
file for errors, and again check the generated file in your favourite molecule viewer.
6. Simulated annealing stage: For the structure calculation
itself, we can use three different types of restraints (if available): unambig.tbl (NOE distance restraints), hbonds.
tbl (hydrogen bond restraints) and dihedrals.tbl (dihedral angle restraints). The script annealing.sh will generate a CNS parameter file (run.cns) with all details and
specifications for the structure calculation protocol, and will
start the calculation. This script should be run from a higher
level than the previous two:
cd ..
RECOORDscripts-cns1.3/annealing.sh 2lci
Individual job files will be generated and executed for each
model you want to calculate. By default two models will be
generated in the created str/ folder, with name similar to
2lci_cns_[1-2].pdb. The CNS input and output files can
be found in the directory cnsRef/, together with possible
error files (see Note 9). The header of every PDB file generated
contains information about violations and energy values.
7. Water refinement stage: Once the simulated annealing phase
is finished and all resulting structures have been written into
the str/ directory, we can proceed to water refinement:
RECOORDscripts-cns1.3/re_h2o.sh 2lci
In the str/ directory, a new directory called wt/ will be created, the best energy structures will be copied there and subsequently refined (see Note 10).
370
3.6 Structure
Validation and Quality
Assessment
The 3D protein structures are the end result of a highly complex

procedure involving interpretation of experimental data and transformation of this data to in silico atomic coordinates. It is therefore
essential that these structures are well-validated and analyzed to
ensure their correctness before they can be used as models that
represent the conformation of the protein in solution.
This validation and analysis happens on two levels: comparison
of the structures to original experimental data such as restraints
and chemical shifts, and assessment of the physical quality of the in
silico molecule description in terms of known physicochemical
properties, such as atom bonds, angles, dihedrals, packing,
hydrogen-bond, and electrostatics. Many programs are available to
do parts of these validations (see Vuister et al. [49] for a recent
overview); we will here focus on CING [15], a recently developed
package that assembles the results of different validation programs
and presents them in a joint, interactive, Web-2.0-based validation
report. As an equivalent alternative, PSVS can be used [66].
The CING package is free to download and install; however,
its large number of dependencies on external software makes this a
nontrivial exercise. A virtual linux image with all public-domain
software installed can be requested from the authors, requiring
only the installation of the external licensed packages. Here, we
will use the recommended iCing web server to validate the
structures.
The iCing server can accept structure coordinates in the form
of PDB files. However, it is highly advisable to also supply the
server with the experimental data. The prefered input format of the
iCing server is a self-contained CCPN project with the structure,
restraints, chemical shifts, peaks, etc, compressed as a .tgz file. Such
a file can be generated and submitted directly from the ccpNmr
Analysis program or can be generated using the FormatConverter
(see Notes 1, 5, and 7).
1. Access the iCING server: To upload the CCPN project data,
first create an archive file of your project:
tar cvfz myCcpnProjectArchive.tgz myCcpnProjectDirectory/
Point your browser to http://nmr.le.ac.uk and select the iCing

option. In the upload panel (Fig. 4a) use the Choose File
button, select your ccpn.tgz file, which will then be uploaded.
For Program we select CCPN (the default). Note that
iCing also natively accepts PDB formatted coordinate files,
multiple files in Cyana format as a .tgz file (see Note 11) or a
CING project as a .tgz file. Selecting Forward allows one
to proceed to the next panel with Criteria. Leave these
371
Fig. 4 iCing server and results. (a) iCing server upload page. Selection area is indicated by the red box. (b)
Summary page for entry 2lci. (c) Residue-specific page for Leu14 of entry 2lci. The residue has a red ROG
score as result of poor sidechain 1-2 dihedral angle conformation (Janin plot not visible), Many related elements (previous and next residues, restraints, chemical shifts, etc.) can directly be accessed through the links
on this page. (d) DihedralByResidue plots. A quick overview is obtained by scrolling down. Individual residue
pages can be accessed through the links
set to their default values. Selecting Forward again, yields

the Options panel that allows for selection of residues, or
models in the ensemble. You can use this to set a range of residues to consider if you want to override the CING routines for
determining this automatically. Generally, this is not required.
Pressing Forward once more yield the Run-panel.
2. Running CING: Press Submit to start the CING validation analysis. The Output panel will appear, which can be manually updated to display the progress in the analysis. A typical
CING validation analysis will take ~1015 min. This is caused
by the many programs that are run and the extensive graphical
output that will be generated. When the program is finished, a
link is presented to the on-line webpages with the results. The
whole CING project file that includes all the data, the results
from the individual programs and the all the webpages can also
be downloaded. The results are stored anonymously on the
iCing server and are automatically removed after 2 days.
3. Using the CING results: CING generates easily accessible,
comprehensive, interactive HTML/Javascript-based validation
372
reports and directs the NMR spectroscopist to areas of the

structure that have problems. Hyperlinks connect all elements
at the different levels; e.g. the full molecule, chains, residues,
atoms, as well as restraints, peaks, and chemical shifts. Lists are
interactive entities that can be reordered and filtered by query.
CING uses at all data levels a simple RedOrangeGreen
(ROG) scoring, which depends upon the combined analysis of
all results and allows CING to direct the user to important
issues:
(a) Red: Potentially serious issues with the structures.
(b) Orange: Potential issues with the structures.
(c) Green: No issues detected.
In addition to this classification mechanism, CING displays the
validation results in direct relation to the experimental data.
4. Structure level results: The link to the on-line webpages with
the results first presents an overview of the various sections of
the CING report. A link to the summary is best examined first
(Fig. 4b). Here, an overview is given of the rmsd values within
the ensemble, the overall CING [15], WHATIF [67] and
PROCHECK_NMR [68] scores, as well as the results of the
restraints analysis. CING also analyses, based upon circular
variance criteria of the backbone dihedral angles and , which
residues should be considered well-defined (see Note 12); only
these regions are used for the superposition of the structures
and the related rmsd values, as are the PROCHECK_NMR
scores and one of the two overal CING ROG scores; the other
pertains to the full molecule. An ensemble of structures that
has low green (< ~20 %) and high red (> ~50 %) overall ROG
score should be labeled as highly troublesome.
5. Residue level results: Within the CING philosophy, a large
emphasis is placed upon the validation of the individual residues. NMR assignment strategies are almost exclusively
residue-based, NMR related parameters are residue dependent
and the local nature of the NMR-derived restraints also correlates well with a residue-based approach. Structural properties
can also be conveniently summarized at the residue level, and
hence all residue-specifc elements are specified on the residue
pages. Residue-specific Ramachandran plots (Fig. 4c) are generated for each residue with all the individual conformers of
the ensemble marked and , restraints, if present, indicated.
Inspection of the sidechain 12 dihedral angle distributions
of the conformers in the ensemble is possible by means of the
Janin plot, as also automatically generated by CING for each
residue. The D1D2 plots display the conformation of the residue relative to the preceding and following residue in the
chain. A residue page also contains all the experimental restraint
373
data involving the featured residue, as well as the critiques defining its ROG scores. Thus, a residue page presents a comprehensive account of all relevant information
pertaining to a specific residue. Hyperlinks connect the page to
all other relevant pages; e.g. the previous and next residue in
the chain, but also all residues connected through restraints or
all its atoms and their chemical shifts. The analysis of the conformation of residues and identification of potential problems
can also conveniently be done using the Dihedral plots
per residue page (Fig. 4d), which displays the relevant
Ramachandran, Janin and D1D2 plots of all residues sequentially, in one scrollable interface from which the relevant residue can also be selected.
6. Other information: It is useful to display the residue-specific
ROG scores mapped onto a structure of the molecule. For
this, macros for the structure visualization programs JMOL
[69], YASARA [45], PyMol (The PyMOL Molecular Graphics
System, Schrdinger, LLC.) and MOLMOL [70] are provided. They can be accessed by following the Programs
flat- > Macros link from the home page. Closely clustered
red/orange residues in the structure are highly suspect and
warrant further investigation.
The CING results are based, in part, upon the results of the
programs PROCHECK_NMR [68] and WHATIF [67]. Direct
links to the output of these programs are also provided.
Notes
1. To use the FormatConverter, install the latest stable version of
the CCPN software, which can be downloaded from http://
www.ccpn.ac.uk/downloads/stable. A less versatile but easier
to use alternative is the FormatConverter web version at
http://haddock.chem.uu.nl/enmr/format-converter.html.
Further information on the FormatConverter is available from
http://www.ccpn.ac.uk/software/fcfolder, as well as a detailed
tutorial at http://www.ccpn.ac.uk/software/tutorials/intro.
Examples of full CCPN projects that can be used for structure
calculation are available from the RECOORD project at
http://www.ebi.ac.uk/pdbe-apps/nmr/recoord/ by clicking
the CCPN projects link near the bottom of the page.
2. Chemical shifts are calculated from absolute frequencies relative
to a reference frequency; this way the positions of NMR resonances can be expressed independently of the magnetic field
strength. If this reference frequency is not correctly set all chemical shift values calculated with it will be equally offset from their
true value. There are a host of programs available to check if
374
chemical shifts are correctly referenced, and to suggest a correction

to obtain their true value (re-referencing of the chemical
shifts): AVS [71], LACS [72, 73], SPARTA+ [23], PANAV
[74], CheckShift [75], ShiftX2 [19], and VASCO [76].
3. Load a CCPN project with the FormatConverter or import a
sequence and chemical shift list. Go to the Export menu and
select TALOS. Open the project export menu, click on Select
export file and define the file you want to write to. The chemical shift data to export can be selected next to Select shift list to
export: (in the example OR36 project this will be 1:bmrb21.
str, indicating a shift list with CCPN serial number 1 and
name bmrb21.str). Click Export project file, then OK in the
popup window (you can here change the sequence numbering
should you wish to do so), then OK again if the export finished
successfully. The exported TALOS file should contain sequence
as well as chemical shift information and can be uploaded to the
TALOS+ server. The procedure for CS-ROSETTA is the same,
although in this case it is useful to first run TALOS+ and then,
in the input file to be uploaded to the CS-ROSETTA server, first
manually remove N- and C-terminal regions from the sequence
that have no chemical shifts, as well as residues that are labeled
as Dynamic by TALOS.
4. WeNMR [59] provides an extensive infrastructure that allows
users to run up-to-date and computationally demanding
software online. In order to use it, you will have to obtain a
personal X509 certificate and register with the WeNMR project and virtual organization. Instructions for this can be found
at the following site: http://www.wenmr.eu/wenmr/access/
registration. Relevant software for this protocol available on
WeNMR is indicated in Table 1.
5. (a) Export of sequence, peak list and chemical shift data for
CYANA. First load a CCPN project with these data with the
FormatConverter (e.g. the example OR36 project). (i)
Sequence export: Go to the Export menu and select Cyana.
Open the sequence export menu and select the molecule for
which you want to export the sequence next to Select chains to
export: (in the example OR36 project this will be
Molecularsystem:A, indicating a molecule with chain
code A in the molecular system Molecularsystem).
Click on Select export file and define the file you want to
write to, then select the CYANA version next to Cyana version:
Click Export sequence file, then OK in the popup window
(you can here change the sequence numbering should you
wish to do so), then OK again if the export finished successfully. (ii) Chemical shift list export: Go to the Export menu and
select XEasy. Open the shift export menu and select the
chemical shift list you want to export next to Select shift list to
375
export: (in the example OR36 project this is 1:bmrb21.

str). Click on Select export file and define the file you want
to write to, then select the button next to Use CYANA2.1
atom names: if you are using the CYANA 2.1 version or higher.
Click Export shifts file, then OK in the popup window, and
OK again if the export finished successfully. (iii) Peak list
export: Go to the Export menu and select XEasy. Open the
peaks export menu and select the peak list you want to export
next to Select peak list to export: (in the example OR36 project
there will be three peak lists, ending in nnoe_raw,
alinoe_raw in aronoe_raw; the text in front of this
indicates the CCPN experiment and processing information).
write to, then select the button next to Write as CYANA format: Click Export peaks file, then OK in the popup window.
In the next window you can define the order of the peak list
dimensions in the output file, press OK when this is fine.
Finally, press OK again if the export finished successfully. Note
that for reasons particular to the XEASY format assignments in
your peak list can only be written out if you first exported a
chemical shift list.
(b) Import of coordinates, distance and dihedral restraint
lists from CYANA. Create a new CCPN project or load an
existing one in the FormatConverter. (i) Coordinate import:
Go to the Import menu, Single files, Coordinates and select
PseudoPdb (note that selecting Pdb will only import official
files from the PDB). Click on Select files, select the file (for
OR36, try restraints/coordinates.pdb) and click
Select. Click IMPORT. Then two situations are possible. (1)
If your sequence was already loaded into the CCPN previously, click OK in the popup window after selecting the correct CCPN molecular system, OK again after successful
import. (2) If you have no sequence in your project, you will
first have to create one matching the molecule from the coordinates. In the first popup window you can change the
sequence information and/or molecule name. Click OK when
done, in the next popup give a name to your molecular system
and press OK, then click No in the next popup unless you
have a homomultimer. You might then get additional popups
asking you to disambiguate atom names from the coordinate
data. Finally click OK again after successful import (ii) Distance
restraints: Go to the Import menu, Single files, Distance constraints and select Cyana. Click on Select file, select the file
(for OR36, try restraints/cyana/distances.upl)
and click Select. You can also select the matching lower distance limits distances.lol file at Select file (optional).
Select the correct CYANA version (use 2.1 for the most recent
versions), and click IMPORT. In the next popup you will be
376
able to select the name of an existing CCPN structure generation,

or create a new one. In the next popup give your distance
restraint list a name; another popup will appear asking you to
read in a lower distance limit file in case you havent specified it.
Click OK again after successful import. Finally, click No for the
popup asking you to run linkResonanceswe will do this after
the next step. (iii) Dihedral restraints: Go to the Import menu,
Single files, Dihedral constraints and select Cyana. Click on
Select file, select the file (for OR36, try restraints/
cyana/dihedrals.aco) and click Select. Select the correct
CYANA version (use 2.1 for the most recent versions), and click
IMPORT. In the next popup you will be able to select the name
of an existing CCPN structure generation, or create a new one.
In the next popup give your dihedral restraint list a name; click
OK again after successful import. When you are done importing
all connected restraint files, click Yes to the linkResonances question, then Yes again (see Subheading 3.1, step 5).
6. If dihedrals or any other types of restraints are available, they
can be obtained in a similar way. The names assigned will be
dihedrals.tbl and hbonds.tbl.
7. Export of distance and dihedral restraint lists as well as coordinates (PDB-like files) for CNS/XPLOR. First load a CCPN
project with these data with the FormatConverter (e.g. the
example OR36_withRestraints project). (i) Distance restraints
export: Go to the Export menu and select Cns. Open the distanceConstraints export menu and select the distance constraint list for which you want to export the sequence next to
Select distance constraint list to export: (in the example OR36
project this will be 1:1:distance constraint list,
indicating a distance constraint list with serial number 1 in
the structure generation with serial number 1). Click on
Select export file and define the file you want to write to.
Click Export distanceConstraints file, then OK in the popup
window (you can here change the chain codes and sequence
numbering should you wish to do so), then OK again if the
export finished successfully. (ii) Dihedral restraints export:
This is the same procedure as for distance restraints except use
the dihedralConstraints menu, and in the example OR36 project select 1:2:dihedral constraint list). (iii)
Coordinates export: Open the coordinates export menu and
select the structure(s) you want to export next to Select structures to export: in the example OR36 project there will be one
structure, MolecularSystem:1:model_1, indicating a
set of coordinates for model_1 relating to molecular system
MolecularSystem in structure ensemble with serial 1.
write to, then click Export coordinates file.
377
8. This only works if a PDB coordinates file is available. Otherwise

use generate_seq.inp and generate_template.inp from CNS to
create such a PDB.
9. Once everything is set up in a proper way, you can edit the
script and change the protocol parameters where necessary.
You can for example change the number of models to generate. It is set to 2 by default, but more common numbers would
be 100 or 200. For more complex systems, you can switch to
a longer annealing protocol, by doubling the number of steps
to be carried out. Depending if you are going to use a cluster
or your own computer, you should change the submit command. Remember also to change the sleep time between submitting jobs, especially if you are not using a cluster and you
do not want to have 100 jobs running in your computer at the
same time! In such a case choose a sleep time that matches the
time needed for one structure calculation.
10. You should also edit this script and change the number of
structures to refine since by default it is set to only 1. Increase
this number to 25 models. They will be assigned names like
2lci_cns_w_[1-25].pdb. The CNS input and output will
be directed to the directory cnsWtRef/.
11. Preparing Cyana input to iCing. The Cyana myMolecule.
cyana.tgz of myMolecule should contain myMolecule.
seq (the cyana sequence file), myMolecule.prot (the corresponding cyana prot file), myMolecule.pdb (the Cyana
generated structure), myMolecule.aco (optional dihedral
restraints), myMolecule.lol (optional lower bounds),
myMolecul.upl (optional upper bounds).
12. The precision of a structure can be estimated by measuring the
conformational variance over an ensemble of models. Usually,
this variance has been expressed as the positional root-meansquare deviation (rmsd) of the individual models from the mean
structure. This parameter is useful for estimating the precision
of the calculation, but does not report on the accuracy. The
later can only be calculated if a standard reference is available.
References
1. Wuthrich K (1986) Nmr of proteins and
nucleic acids. Wiley, New York, NY
2. Neuhaus D, Williamson MP (2000) The
nuclear overhauser effect in structural and conformational analysis. Wiley, New York, NY
3. Altona C (1996) Vicinal coupling constants
and conformation of biomolecules. In Harris
DMG, a K R (eds) Encyclopedia of nuclear
magnetic
resonance.
Wiley,
London.
pp 49094922
4. Bax A, Kontaxis G, Tjandra N (2001) Dipolar

couplings in macromolecular structure determination. Methods Enzymol 339:127174
5. Bertini I, Luchinat C, Parigi G (2012) Towards
mechanistic systems biology. Wiley-VCH Verlag
GmbH, Weinheim, Germany. pp 154171
6. Cavalli A, Salvatella X, Dobson CM,
Vendruscolo M (2007) Protein structure
determination from NMR chemical shifts. Proc
378
7. Shen Y et al (2008) Consistent blind protein

structure generation from NMR chemical shift
data. Proc Natl Acad Sci U S A 105:46854690
8. Wishart DS et al (2008) CS23D: a web server
for rapid protein structure generation using
NMR chemical shifts and sequence data.
Nucleic Acids Res 36:W496W502
9. Gntert P (1998) Structure calculation of biological macromolecules from NMR data. Q
Rev Biophys 31:145237
10. Linge JP, Williams MA, Spronk CAEM,
Bonvin AMJJ, Nilges M (2003) Refinement
of protein structures in explicit solvent.
Proteins 50:496506
11. Nederveen AJ et al (2005) RECOORD: a
recalculated coordinate database of 500+ proteins from the PDB using restraints from the
BioMagResBank. Proteins 59:662672
12. Gntert P (2004) Automated NMR structure
calculation with CYANA. Methods Mol Biol
278:353378
13. Brnger AT et al (1998) Crystallography &
NMR system: a new software suite for macromolecular structure determination. Acta
Crystallogr D Biol Crystallogr 54:905921
14. Brunger AT (2007) Version 1.2 of the
Crystallography and NMR system. Nat Protoc
2:27282733
15. Doreleijers JF et al (2012) CING: an integrated residue-based structure validation program suite. J Biomol NMR 54:267283
16. Luginbhl P, Szyperski T, Wthrich K (1995)
Statistical basis for the use of 13c chemical
shifts in protein structure determination.
J Magn Res 109:92
17. Kuszewski J, Qin J, Gronenborn AM, Clore
GM (1995) The impact of direct refinement
against 13C alpha and 13C beta chemical shifts
on protein structure determination by NMR.
J Magn Reson B 106:9296
18. Neal S, Nip AM, Zhang H, Wishart DS (2003)
Rapid and accurate calculation of protein 1H,
13C and 15N chemical shifts. J Biomol NMR
26:215240
19. Han B, Liu Y, Ginzinger SW, Wishart DS
(2011) SHIFTX2: significantly improved protein chemical shift prediction. J Biomol NMR
50:4357
20. Xu XP, Case DA (2001) Automated prediction
of 15N, 13Calpha, 13Cbeta and 13C chemical shifts in proteins using a density functional
database. J Biomol NMR 21:321333
21. Williamson MP, Kikuchi J, Asakura T (1995)
Application of 1H NMR chemical shifts to
measure the quality of protein structures.
J Mol Biol 247:541546
22. Meiler J (2003) PROSHIFT: protein chemical

shift prediction using artificial neural networks.
J Biomol NMR 26:2537
23. Shen Y, Bax A (2010) SPARTA+: a modest
improvement in empirical NMR chemical shift
prediction by means of an artificial neural network. J Biomol NMR 48:1322
24. Karplus M (1963) Vicinal proton coupling in
nuclear magnetic resonance. J Am Chem Soc
85:28702871
25. Kim Y, Prestegard JH (1990) Refinement of
the NMR structures for acyl carrier protein
with scalar coupling data. Proteins 8:377385
26. Torda AE, Brunne RM, Huber T, Kessler H,
Van Gunsteren WF (1993) Structure refinement using time-averaged J-coupling constant
restraints. J Biomol NMR 3:5566
27. Wagner G, Wthrich K (1982) Amide protein
exchange and surface conformation of the
basic pancreatic trypsin inhibitor in solution.
Studies with two-dimensional nuclear magnetic resonance. J Mol Biol 160:343361
28. Pervushin K et al (1998) NMR scalar couplings
across Watson-Crick base pair hydrogen bonds
in DNA observed by transverse relaxationoptimized spectroscopy. Proc Natl Acad Sci U
S A 95:1414714151
29. Cordier F, Rogowski M, Grzesiek S, Bax A
(1999) Observation of through-hydrogenbond 2hJHC' in a perdeuterated protein.
J Magnet Res 140:510512
30. Bonvin AMJJ, Houben K, Guenneugues M,
Kaptein R, Boelens R (2001) Rapid protein
fold determination using secondary chemical
shifts and cross-hydrogen bond 15N-13C
scalar couplings (3hbJNC). J Biomol NMR
21:221233
31. Bax A (2003) Weak alignment offers new
NMR opportunities to study protein structure
and dynamics. Protein Sci 12:116
32. Bax A, Grishaev A (2005) Weak alignment
NMR: a hawk-eyed view of biomolecular structure. Curr Opin Struct Biol 15:563570
33. Prestegard JH, Bougault CM, Kishore AI
(2004) Residual dipolar couplings in structure
determination of biomolecules. Chem Rev
104:35193540
34. Tjandra N, Omichinski JG, Gronenborn AM,
Clore GM, Bax A (1997) Use of dipolar
1H-15N and 1H-13C couplings in the structure determination of magnetically oriented
macromolecules in solution. Nat Struct Biol
4:732738
35. Fushman D, Varadan R, Assfalg M (2004)
Determining domain orientation in macromolecules by using spin-relaxation and residual
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
dipolar coupling measurements. Prog Nucl

Magn Reson Spectrosc 44:189214
Tjandra N, Garrett DS, Gronenborn AM, Bax
A, Clore GM (1997) Defining long range
order in NMR structure determination from
the dependence of heteronuclear relaxation
times on rotational diffusion anisotropy. Nat
Struct Biol 4:443449
Bertini I, Luchinat C, Parigi G, Pierattelli R
(2005) NMR spectroscopy of paramagnetic
metalloproteins. ChemBioChem 6:15361549
Banci L et al (2004) Paramagnetism-based
restraints for Xplor-NIH. J Biomol NMR
28:249261
Bertini I, Luchinat C, Parigi G (2002)
Paramagnetic constraints: an aid for quick
solution structure determination of paramagnetic metalloproteins. Concepts Magn Reson
14:259286
Schwieters CD, Kuszewski JJ, Tjandra N,
Clore GM (2003) The Xplor-NIH NMR
molecular structure determination package.
J Magnet Res 160:6573
Gntert P, Mumenthaler C, Wthrich K
(1997) Torsion angle dynamics for NMR
structure calculation with the new program
DYANA. J Mol Biol 273:283298
Hus JC, Marion D, Blackledge M (2000) De
novo determination of protein structure by
NMR using orientational and long-range order
Case DA et al (2005) The Amber biomolecular simulation programs. J Comput Chem
26:16681688
van der Spoel D et al (2005) GROMACS:
fast, flexible, and free. J Comput Chem 26:
17011718
Krieger E, Koraimann G, Vriend G (2002)
Increasing the precision of comparative models
with YASARA NOVAa self-parameterizing
force field. Proteins 47:393402
Montelione GT et al (2013) Recommendations
of the wwPDB NMR validation task force.
Structure 21(9):15631570
Kirchner DK, Gntert P (2011) Objective
identification of residue ranges for the superposition of protein structures. BMC Bioinformatics
12:170
Mao B, Guan R, Montelione GT (2011)
Improved technologies now routinely provide
protein NMR structures useful for molecular
replacement. Structure 19:757766
Vuister GW, Fogh RH, Hendrickx PMS,
Doreleijers JF, Gutmanas A (2013) An overview of tools for the validation of protein NMR
structures. J Biomol NMR 58(4):259285
379
50. Spronk C, Nabuurs SB, Krieger E (2004)

Validation of protein structures derived by
NMR spectroscopy. Prog Nucl Magn Reson
Spectrosc 45:315337
51. van der Schot G et al (2013) Improving 3D
structure prediction from chemical shift data.
52. Rosato A et al (2009) CASD-NMR: critical
assessment of automated structure determination by NMR. Nat Methods 6:625626
53. Rosato A et al (2012) Blind testing of routine,
fully automated determination of protein structures from NMR data. Structure 20:227236
54. Vranken WF et al (2005) The CCPN data
model for NMR spectroscopy: development of
a software pipeline. Proteins 59:687696
55. Shen Y, Delaglio F, Cornilescu G, Bax A (2009)
TALOS+: a hybrid method for predicting protein backbone torsion angles from NMR chemical shifts. J Biomol NMR 44:213223
56. Cheung M-S, Maguire ML, Stevens TJ,
Broadhurst RW (2010) DANGLE: A Bayesian
inferential method for predicting protein backbone dihedral angles and secondary structure.
J Magn Reson 202:223233
57. Wishart DS, Sykes BD (1994) Chemical shifts
as a tool for structure determination. Methods
Enzymol 239:363392
58. Berjanskii MV, Wishart DS (2005) A simple
method to predict protein flexibility using secondary chemical shifts. J Am Chem Soc 127:
1497014971
59. Wassenaar TA, et al (2012) WeNMR: structural
biology on the grid. J Grid Comp 10:743767
60. Huang YJ, Powers R, Montelione GT (2005)
Protein NMR recall, precision, and F-measure
scores (RPF scores): structure quality assessment measures based on information retrieval
statistics. J Am Chem Soc 127:16651674
61. Guerry P, Herrmann T (2012) Comprehensive
automation for NMR structure determination
of proteins. Methods Mol Biol 831:429451
62. Linge JP, Habeck M, Rieping W, Nilges M
(2003) ARIA: automated NOE assignment
and NMR structure calculation. Bioinformatics
19:315316
63. Raman S et al (2010) Accurate automated protein
NMR structure determination using unassigned
NOESY data. J Am Chem Soc 132:202207
64. Linge JP, O'Donoghue SI, Nilges M (2001)
Automated assignment of ambiguous nuclear
overhauser effects with ARIA. Methods
Enzymol 339:7190
65. Seavey BR, Farr EA, Westler WM, Markley JL
(1991) A relational database for sequence-specific
protein NMR data. J Biomol NMR 1:217236
380
66. Bhattacharya A, Tejero R, Montelione GT

(2007) Evaluating protein structures determined by structural genomics consortia.
Proteins 66:778795
67. Vriend G (1990) WHAT IF: a molecular modeling and drug design program. J Mol Graph
8:5256
68. Laskowski RA, Rullmann J, MacArthur MW
(1996) AQUA and PROCHECK-NMR: programs for checking the quality of protein structures
solved by NMR. J Biomol NMR 8:477486
69. Herrez A (2006) Biomolecules in the computer: Jmol to the rescue. Biochem Mol Biol
Educ 34:255261
70. Koradi R, Billeter M, Wthrich K (1996)
MOLMOL: a program for display and analysis
of macromolecular structures. J Mol Graph
14(515):2932
71. Moseley HNB, Sahota G, Montelione GT (2004)
Assignment validation software suite for the evaluation and presentation of protein resonance
assignment data. J Biomol NMR 28:341355
72. Wang L, Eghbalnia HR, Bahrami A, Markley

JL (2005) Linear analysis of carbon-13 chemical shift differences and its application to the
detection and correction of errors in referencing and spin system identifications. J Biomol
NMR 32:1322
73. Wang L, Markley JL (2009) Empirical correlation between protein backbone 15N and 13C
secondary chemical shifts and its application
to nitrogen chemical shift re-referencing.
74. Wang B, Wang Y, Wishart DS (2010) A probabilistic approach for validating protein NMR
chemical shift assignments. J Biomol NMR
47:8599
75. Ginzinger SW, Gerick F, Coles M, Heun V
(2007) CheckShift: automatic correction of
inconsistent chemical shift referencing.
76. Rieping W, Vranken WF (2010) Validation of
archived chemical shifts through atomic coordinates. Proteins 78:24822489
Part IV
ProteinLigand Interactions
Chapter 17
Methods for Predicting ProteinLigand Binding Sites
Abstract
Ligand binding is required for many proteins to function properly. A large number of bioinformatics tools
have been developed to predict ligand binding sites as a first step in understanding a proteins function or
to facilitate docking computations in virtual screening based drug design. The prediction usually requires
only the three-dimensional structure (experimentally determined or computationally modeled) of the
target protein to be searched for ligand binding site(s), and Web servers have been built, allowing the free
and simple use of prediction tools. In this chapter, we review the underlying concepts of the methods used
by various tools, and discuss their different features and the related issues of ligand binding site prediction.
Some cautionary notes about the use of these tools are also provided.
Key words Structural bioinformatics, Proteinligand interaction, Protein surface grid, Molecular probe,
Surface pocket and cavity, Ligand binding site prediction, Bioinformatics software and servers
Introduction
Interaction with a ligand molecule is essential for many proteins to
carry out their biological function. This interaction is generally
specific, not only in terms of the molecules involved in the interaction, but also in the location (i.e., the site of ligand binding) in
which the interaction takes place. In order to gain knowledge
about the interaction and, by extension, the proteins function
and how to influence its activity by, for example, designing small
molecule drugs, considerable efforts have been made to develop
methods that can predict ligand binding sites (LBSs) of proteins
computationally, and a very large number of bioinformatics tools
are now available for LBS prediction (reviewed in [14]). In general,
because of the location specificity of LBSs, most of these methods
have exploited one or more of four types of properties (evolutionary, geometric, energetic, and statistical) in order to distinguish
the binding site from other parts of the protein surface. In this
review, we will survey the many LBS prediction methods and
classify them on the basis of the site-distinguishing properties they
383
384
use into one of the following categories: (1) template-based

those that utilize homologous and/or similar structures with
known binding sites, (2) geometry-basedthose that perform
some kind of geometric computation to identify binding site pockets, (3) energy-basedthose that compute the interaction energy
usually using imaginary ligands as binding probes, (4) propensitybasedthose that compute the propensity of a certain property
that shows a statistically significant preference for known LBSs,
rather than non-LBSs, and (5) combination-based and others
those that make the prediction based on the results of other methods and those that cannot be easily classified into one of the above
categories. We will focus our review on the basic concepts and features of each of these categories; access information for, and notes
about, the different methods surveyed are given in Table 1.
Methods
2.1 Template-Based
Methods
Proteins sharing sequence homology are known to adopt similar

three-dimensional (3D) structures and usually perform similar
biological functions [5]. This is the basic idea behind all templatebased methods for LBS prediction. The first step in these methods is
to identify one or more ligand-bound complex structures (to serve
as a template to find potential LBSs) that share sufficient sequence
similarity with the target protein (the protein in which LBSs are to
be predicted). By superimposing the target protein and identified
templates, which include information on the location of known
LBSs, a consensus site for ligand binding can be revealed and its
characteristics as a putative LBS for the target protein evaluated by
comparison to those of known LBSs [613]. As a result, the similarity between the template(s) and target protein and the accuracy
of the sequence and structure alignments used in the procedure
can affect the prediction accuracy of the methods in this category.
In general, template-based methods tend to yield better accuracy
in LBS prediction than methods in other categories (see Note 1).
Thus, for example, in CASP (Critical Assessment of Protein
Structure Prediction) competition experiments in the category of
LBS prediction [14, 15], most participating groups employed
some form of template-based approach [810, 16]. However, one
caveat about the use of template-based methods is the requisite for
a suitable template structure(s) containing known LBSs.
If the structure of the target protein has not yet been experimentally determined, template-based predictions of LBSs are still
possible using homology-derived model(s) of the target protein as
a substitute. It has been shown [17, 18] that, for about 90 % of
proteins studied, at least one structure analog exists in the structure database PDB [19]. Thus, the large majority of all proteins
can be structurally modeled using standard homology modeling
Reference
[8, 86]
[9, 8789]
[10]
[11, 90]
[21, 28]
[27]
Firestar
I-TASSER
Lees method
IntFOLD
ProBis
Ims method
[30]
[31]
PocketPicker
VICE
II. Geometry-based methods

LIGSITEcsc
[29, 32]
[7, 79]
FINDSITE
I. Template-based methods
3DLigandSite
[6, 16]
Method1
Table 1
List of LBS prediction methods
http://projects.biotec.
tu-dresden.de/pocket/
http://www.reading.ac.uk/
bioinf/IntFOLD/
http://probis.cmm.ki.si
http://firedb.bioinfo.cnio.es/
Php/FireStar.php
http://zhanglab.ccmb.med.
umich.edu/I-TASSER/
http://www.sbg.bio.ic.ac.
uk/~3dligandsite/
http://www.cssb.biology.
gatech.edu/findsite
Web server2
Yes
Yes
Yes
Yes
Structure
viewer3
(continued)
Option for re-ranking by conservation score and changing the parameters

Two widely used established datasets: 210 bound structures and 48 pairs
of bound/unbound structures
Option for input of single or multiple chains
Calculates buriedness values for each empty grid and clusters high
buriedness grids
Similar concept to PocketPicker [30], but with new indices, such as
cavityness and shaping factor
Evaluated on a large benchmark set containing>1,000 nonredundant

structures and structure models
Option for subsequent docking computation
Among the most highly scored methods in CASP9 [15]
Single chain input
Needs registration for job submission.
Includes structure prediction and functional annotation
Multiple prediction functions: structure model, domain, disorder, and
binding sites, plus quality assessment of models
Uses graph theory to find local structure similarity by comparison with
about 24000 PDB structures
Concept similar to ProBis [21]
Uses spatial proximity to identify the 100 highest scoring templates
Note

385
[35]
[36]
[3739]
[40]
[41, 42]
[93]
[94]
[95]
SCREEN
POCASA
CASTp
MSPocket
fpocket
PocketDepth
DEPTH
DoGSiteScorer
III. Energy-based methods

SiteHound
[46, 53]
Reference
Method1
Table 1
(continued)
http://scbx.mssm.edu/
sitehound/sitehound-web/
Input.html
http://dogsite.zbh.unihamburg.de/
http://bioserv.rpbs.univparis-diderot.fr/cgi-bin/
fpocket
http://proline.physics.iisc.
ernet.in/pocketdepth/
http://mspc.bii.a-star.edu.
sg/tankp/run_depth.html
http://altair.sci.hokudai.ac.
jp/g6/service/pocasa/
http://sts.bioengr.uic.edu/
castp/
http://bhapp.c2b2.columbia.
edu/screen2/cgi-bin/
screen2.cgi
Web server2
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Structure
viewer3
Option to use one of 4 types of probes (methyl carbon, aromatic carbon,

phosphate oxygen, or hydroxyl oxygen) to compute the interaction
energy
Similar concept to LIGSITEcsc [29, 32]

Uses a depth factor to identify surface grids of a pocket
The main purpose of this server is to compute the depth of protein
residues and the solvent accessible surface area (SASA), but it can also be
used to predict the ligand binding cavity
Basic assumption: ligand binding residues are often both deep in a cavity
and solvent-exposed
Option to adjust operating parameters
Uses a variety of geometric and physicochemical properties (volume,
depth, surface, ellipsoid main axes, site lining atoms and residues, and
functional groups)
Uses two spheres of different radii to identify empty spaces

Uses random forests, a machine learning method, to improve prediction
accuracy
Batch job is allowed
Similar concept to SCREEN [35]
Excludes noise points, e.g., points deep inside the protein structure
Maps residue information from Swiss-port [91] and SNP data from
OMIM [92] to predicted LBS residues
Source code available
Multiple functions: pocket prediction, molecular dynamics simulation of
the pocket, and viewing of conserved pockets in homologous proteins
Note
386
[57, 58]
http://lise.ibms.sinica.edu.tw
http://opus.bch.ed.ac.uk/
stp/
http://ftsite.bu.edu
[62]
[65]
[67]
MetaPocket
2.0
MEDock
Thorntons
method
http://projects.biotec.
tu-dresden.de/
metapocket/
V. Combination-based methods and others

ConCavity
[60]
LISE
see Note 7
see Note 8
3
see Note 9
Yes
[50]
IV. Propensity-based methods

Hirayamas
[54]
method
STP
[56]
[49]
Yes
Yes
Moritas
method
FTSite
http://www.modelling.leeds.
ac.uk/qsitefinder/
[48]
Structure
viewer3
Q-SiteFinder
Web server2
Reference
Method1
Uses docking to search for LBSs

Uses a neural network to predict active site residues in enzymes
Combines scores from 3 different methods with residue conservation

data
Reports the consensus of the results of 8 different methods
Re-ranks the pockets predicted by SURFNET [34] by amino acid

composition propensities for LBSs
Re-ranks the cavities identified by SURFNET [34] according to the
number of highly scored atoms
Does not provide the predicted pockets, but assigns a score to each atom
Scored by binding site enrichment factors of protein triangles
Tested on two widely used small datasets and a large nonredundant set
[57, 58]
Reports success rates for different types of proteins [58].
Option to use16 different small molecular probes separately on grids
Similar to Q-SiteFinder, with modified probe distribution and clustering
Compiles a widely used dataset containing 35 pairs of bound/unbound

complexes
Uses the methyl group (CH3) as probe
Note

387
388
routines to facilitate LBS prediction (see Chapter 16 of this volume).

Most template-based LBS prediction servers (e.g., see Table 1)
can accept the target protein sequence as input, but the accuracy of
the prediction will depend on the quality of the derived homology
model (see Note 2).
Since protein structures are more highly conserved than amino
acid sequences [20], structure comparison has the potential to
discover suitable templates for LBS prediction in cases in which
sequence comparison may fail. As structure determines function,
some methods, such as ProBis [21], focus on the structural similarity
between target and templates rather than sequence homology.
However, even if template(s) can be identified by structure comparison, accurate prediction of LBSs still cannot be guaranteed
because proteins with a globally similar 3D structure may have different functions and thus distinct LBSs [22, 23]. On the other
hand, binding sites for similar ligands or similar enzymatic mechanisms have been found to be conserved in proteins sharing no
overall structural similarity [2426], and this has been harnessed
with success by methods utilizing structure comparisons to identify
similarity only at the level of local structures, i.e., the area surrounding the binding site [21, 27, 28].
2.2 Geometry-Based
Methods
The main task of geometry-based methods is to identify, by computing some types of geometric measures, pocket(s) on the protein
surface that can accommodate small ligand molecules. Statistical
studies made on proteinligand complex structures archived in PDB
indicate that small molecule ligands tend to bind at deflated regions
of the protein surface, in particular, its largest (and/or deepest)
cavities [29]. Consequently, most geometry-based methods have
focused on identifying the largest pockets in proteins. However, how
to determine and identify cavities on the protein surface is a more
complicated problem than it might appear at first sight, and, over the
years, many diverse and creative approaches have been explored.
The first step in many LBS prediction methods in this category
is to find empty space on the protein surface, and one popular
approach is to spray grid points on the target protein and find
empty grids (those not occupied by protein atoms) [11, 2932].
For example, LIGSITEcsc [29] scans grids in seven directions (x, y, z,
and four cubic diagonals) to identify surface-empty-surface
connections, then clusters the empty grids from these to identify
empty spaces for potential ligand binding. Another approach is to
place empty spheres on the protein surface [33, 34]. For example,
to find large empty spaces, SURFNET [34] places empty spheres
between every pair of protein atoms that have no intervening
protein atoms. One variation of the empty sphere approach involves
rolling two spheres with different radii on the target protein to
generate an inner and outer "surface" and reveal empty spaces, i.e.,
empty pockets, between these two surfaces [35, 36]. Yet another
389
approach is to compute Delaunay triangulation to find voids on

the protein surface [3741]. For example, fpocket [41, 42] uses
the alpha sphere [37] to identify empty spaces on a protein structure
(an alpha sphere is a sphere that contacts four atoms and contains
no internal atoms).
In the next step, the empty grids, probes, spheres, or voids are
clustered to identify the largest pocket (cavity), which is often
assigned as the best (top-ranked) predicted LBS. In the development of these methods, many parameters have been devised to
help in the assignment of the top-ranked LBSs, including the use
of a buriedness index [30, 31], path length and shaping
index [31], and depth factor [36]. Other properties of protein
structures, including the B-factor [43] and packing density [44],
have been found to be highly relevant to LBSs and deserve further
investigation. However, despite considerable success, the assumption that the largest pocket is the true LBS is not always correct.
According to one study, of a rather small set of 210 proteinligand
complex structures surveyed, only about 70 % of ligands were
found to bind to the largest cavity in the protein, while 87 % of the
ligands bound to one of the three largest cavities [29]. To avoid
relying on this inherently incorrect assumption, many of the
geometry-based methods resort to re-ranking the identified cavities
by incorporating other geometrical, physicochemical, and/or evolutionary (e.g., sequence conservation) properties [29, 35, 41, 45],
although re-ranking does not usually dramatically improve the prediction results based on the cavities identified.
2.3 Energy-Based
Methods
The aim of energy-based methods is to find patches of the protein

surface that are energetically favorable for ligand binding [4652].
This is usually done by devising a probe molecule and computing
the interaction energy between the probe and surrounding protein
atoms. Many of the energy-based methods are also grid-based, as
they place probes on empty grids of the protein surface in order to
perform the energy computation [4650]. For example, SiteHound
uses Molecular Interaction Fields (MIFs) to calculate the interaction energy between a target protein and a probe, grid points with
a high interaction energy are then clustered to identify potential
LBSs [46]. Generally speaking, energy-based methods are less
diverse and much less numerous than geometry-based methods.
The main differences between the various energy-based methods
are in the design of the probes and how they are distributed on the
protein surface. In general, there is a trade-off between prediction
quality and computational load, which increases with increased
probe complexity. The server SiteHound-web [46, 53] allows users
to choose one of four types of probes to compute the interaction
energy; the FTSite server uses up to 16 different small molecular
probes to identify consensus clusters of grids; the server
Q-SiteFinder [48] uses only the methyl group (CH3) as the
390
probe, while Moritas method [49], which is similar to Q-SiteFinder,

improves the prediction by introducing a new probe distribution
and clustering scheme.
2.4 PropensityBased Methods
Another class of methods computes neither geometry nor interaction energy, but the statistics of certain properties for their propensities to be at, or associated with, known LBSs. One typical
propensity-based method is Hirayamas method [54], in which the
predicted binding pockets generated by a program named Alpha
Site Finder [55] are re-ranked by the amino acid composition,
which shows small, but statistically significant, differences between
LBSs and non-LBSs. Propensity-based methods often re-rank the
pockets predicted by other methods, mainly geometry-based
methods. For example, the surface triplet propensities (STP) algorithm [56] assigns a propensity score to each atom located in the
binding pockets predicted by SURFNET [34], then re-ranks the
SURFNET pockets by simply counting the number of high-scoring
atoms. In contrast, a new propensity-based method developed in
our laboratory [57, 58] does not rely on pockets pre-identified by
other methods. This new method was named LISE for Ligand
Interacting and binding Site Enriched protein triangles. In LISE,
the protein triangles are a triplet of three protein surface atoms
simultaneously interacting with a ligand molecule and the three
protein atoms are concomitantly enriched at LBSs and are assigned
an enrichment factor deduced from a statistical analysis of a set of
proteinligand complex structures [59].
2.5 CombinationBased Methods

and Others
Because geometry and energy are two distinct attributes of LBSs

and because different methods may complement and/or compensate each other, it is not surprising that a combination-based
approach can prove successful in LBS prediction (see Note 3). For
example, ConCavity [60] uses a weighted equation to combine the
raw scores from three different methods (LIGSITEcs [29],
SURFNET [34], and PocketFinder [61]) with surface residue
conservation data to generate a final, composite score for prediction, while MetaPocket 2.0 [62, 63] combines the results of eight
methods (LIGSITEcs [29], PASS [33], Q-SiteFinder [48],
SURFNET [34], Fpocket [41], GHECOM [64], ConCavity [60],
and POCASA [36]) and uses their consensus for the prediction.
The publications describing ConCavity and Metapocket 2.0
reported that a better prediction performance was achieved with
the combined methods used in these two approaches than with
each of the individual methods alone. Finally, there are other methods for predicting LBSs, such as MEDOCK [65] and MolSite
[66], which use a genetic algorithm and docking computation,
and Thornton's Method [67], which uses neural network learning,
to predict LBSs.
2.6
Related Issues
391
As mentioned above, many creative and useful LBS prediction

methods have been developed, especially during the past decade.
The success rates of prediction for recently reported methods
now routinely exceed 80 % for the top-ranked site and 90 % for
the top-three sites, and each reported new method tends to claim
a superior performance compared to previously reported methods; however, the comparison has often been made on rather
small datasets and, in addition, a head-to-head comparison is difficult to achieve because different methods may employ different
evaluation criteria. Other compounding factors, discussed below,
are also involved.
Most geometry-based methods are evaluated by a distance
criterion, which checks the distances between a specific single
point, usually the geometric center of the predicted LBS pocket,
and all the ligand atoms, and dictates that at least one of these distances must be within 4 for the prediction to be considered a
success [29]. Because the aim of geometry-based methods is to
identify the largest pockets of the target protein, pinpointing the
exact binding location within a pocket is often not a primary concern. In contrast, since energy-based methods often identify a small
patch within a binding pocket as the most energetically favorable
site for ligand binding, most energy-based methods are instead
assessed using a precision criterion, which checks the volume
percentage of the predicted pocket occupied by the ligand [48,
49]. Although this seems to be more stringent than the distance
criterion, it should be noted that a predicted pocket, even one
with a 100 % precision, may cover just a small part of the binding
site, and this may not be sufficient to adequately represent the
whole binding site, especially in the case of large pockets that can
bind diverse ligand molecules. Alternatively, especially in templatebased methods, one can evaluate whether each of the predicted
ligand-binding amino acids is in contact with (e.g., within a certain
distance of) the ligand, thus allowing the predictions sensitivity
and specificity to be calculated from the percentage of amino acid
residues that are predicted to be LBS-associated [16, 60]. However,
cases abound in which different ligands bind to the same protein in
the same pocket, but at different locations, or even in a different
pocket, and, thus, a residue considered to be a binding site residue
for one ligand can be a non-binding site residue for another,
making the computed sensitivity and specificity values somewhat
dependent on the particular ligand and also on the particular complex structure used [57, 58].
The issue of how to derive a dataset of proteinligand complexes with which to evaluate LBS predictions is also complicated.
For example, although a certain threshold of amino acid sequence
identity (say 30 %) can be used to remove homologous proteins to
build a so-called nonredundant set, many structures may, as
392
mentioned above, share a similar 3D structure and thus a similar

LBS despite having a sequence identify below the threshold, making the dataset not really redundant. An example of this includes
the finding that, in a dataset used widely for evaluating LBS prediction methods [29], there are six kinase structures that, along with
their LBS pockets, can be well superimposed even though their
sequence identities are below the threshold used to remove homologous structures [57]. On the other hand, as also mentioned
above, even homologous proteins can have distinct LBSs, so the
inclusion of homologous structures does not necessarily make a
dataset redundant with regard to LBS prediction. Compounding
the issue are the different criteria, such as different chain length
and X-ray resolution, used to select protein structures. Thus, for
example, we may expect geometry-based methods to perform better than energy-based methods on a dataset of large proteins, since
pocket size increases linearly with protein size [48, 68] and
geometry-based methods are designed to find large pockets.
Furthermore, the experimentally determined protein structures
deposited in PDB are heavily biased, as the protein in about 70 %
of ligand-bound protein complex structures is a cytosolic enzyme
[58], which, compared to other types of protein, is more likely to
have a better resolution owing to, for example, interest in using
them as targets for drug development. As a result, methods trained
on PDB structures, while performing well on cytosolic enzymes,
may perform badly on other types of proteins, such as membrane
proteins, which are the targets of more than 50 % of the drugs currently in use [6971], but have a rather small percentage of representation in PDB (see Note 4). This is also one of the main reasons
that, when applied to proteome-wide data, many methods usually
produce a much lower accuracy than those reported using small
datasets [58].
Most LBS predictions have been carried out using structures
of single chains. However, in order to function, many proteins
form a complex of multiple homo- or hetero-chains, and their LBS
may be located at the interface between protein chains. The biological unit of a protein should therefore be considered for LBS
prediction whenever possible (see Note 5). Biological unit structures can be downloaded from PDBsum [72]. Finally, as more and
more new methods are developed, the prediction algorithms
involved become more and more sophisticated and often include
new indices (shaping index, buriedness, etc.), modified clustering
procedures, and novel properties for re-ranking. As a result, more
and more parameters have to be optimized. While most LBS prediction servers provide a default set of parameters, users should be
aware that some methods have been tested only on a small dataset
or even require the use of different parameter values for different
types or datasets of proteins (see Note 6).
393
Conclusions and Prospects

Notwithstanding the existence of promiscuous proteins (those that
can bind many different ligand molecules) and promiscuous ligands
(those that can bind to different sites on the same or different
proteins) [7376], proteinligand interactions are generally specific, a fact that underlies the success of the various prediction
methods reviewed above which use certain properties to distinguish between LBSs (specific interactions) and other parts of the
protein surface (nonspecific interactions). Nevertheless, protein
LBSs come in different sizes and shapes and may bind a variety of
ligand molecules, making it difficult to evaluate LBS prediction
methods, many of which have been developed and tested based
only on a small, or biased, set of complex structures. In view of
this, it is worth making additional efforts to build LBS databases
for specific protein families, such as kinases [77, 78], whereby
different types of LBSs can be systematically compared and studied
to reveal the common principles of proteinligand recognition,
and this will help in developing a new generation of methods with
improved accuracy in predicting LBSs across the spectrum of proteins. In the meantime, for the purpose of facilitating docking
computations, it would be useful to have a success-measuring criterion, perhaps less sophisticated than the distance, precision,
or residue-based sensitivity/specificity mentioned above, to
evaluate different types of LBS prediction methods based on their
ability to correctly identify only the ligand binding pocket and not
necessarily the precise location or residues involved.
Notes
1. The first step in LBS prediction should be to search for a
homologous structure(s). If ligand-bound homologous
structure(s) exist in PDB, a template-based prediction is usually quite reliable, particularly if the prediction is supported by
other types of methods.
2. With the exception of FINDSITE [7, 79] and perhaps a few
others, most methods do not report their prediction accuracy
on homology-derived models, although, according to
FINDSITE, homology models can be tolerated for templatebased LBS predictions.
3. It is generally recommended to use multiple different methods
to find consensus and/or to identify potential LBSs for
evaluation.
4. Several LBS prediction methods have been developed for specific protein families [8084]; for target proteins belonging
394
to these specific protein families, these methods can be more

reliable than those reviewed in this chapter. General-purpose
methods may have a reduced accuracy in predicting the
LBSs of certain specific protein families, such as membrane
receptors [58].
5. Some methods and their Web server and/or stand-alone program may only accept single protein chains and may miss
potential LBSs at chainchain interfaces.
6. It may be easier for users to experiment using stand-alone programs, if available, to optimize the parameters for their specific
application.
7. A difficult case for LBS prediction is when a significant conformational change occurs upon ligand binding. This is a problem
that has not been sufficiently addressed in the literature for
LBS prediction.
8. Submitting a batch of jobs to run on an online server is usually
not recommended (or not allowed), as it may take up too
much computing time. For this type of application, as in largescale predictions, it is more feasible to run the batch job on a
stand-alone version installed on the users local computer.
9. Many structure viewers, such as Jmol [85], are freely available
and can be installed to view the predictions for servers that do
not offer an online viewing option.
References
1. Leis S, Schneider S, Zacharias M (2010) In
silico prediction of binding sites on proteins.
Curr Med Chem 17(15):15501562
2. Laurie AT, Jackson RM (2006) Methods for the
prediction of proteinligand binding sites for
structure-based drug design and virtual ligand
screening. Curr Protein Pept Sci 7(5):395406
3. Perot S, Sperandio O, Miteva MA, Camproux
AC, Villoutreix BO (2010) Druggable pockets
and binding site centric chemical space: a paradigm shift in drug discovery. Drug Discov
Today 15(1516):656667
4. Henrich S, Salo-Ahen OM, Huang B,
Rippmann FF, Cruciani G, Wade RC (2010)
Computational approaches to identifying and
characterizing protein binding sites for ligand
design. J Mol Recognit 23(2):209219
5. Lee D, Redfern O, Orengo C (2007) Predicting
protein function from sequence and structure.
Nat Rev Mol Cell Biol 8(12):9951005
6. Wass MN, Kelley LA, Sternberg MJ (2010)
3DLigandSite: predicting ligand-binding sites
using similar structures. Nucleic Acids Res
38(Web Server issue):W469W473
7. Brylinski M, Skolnick J (2008) A threadingbased method (FINDSITE) for ligand-binding

site prediction and functional annotation. Proc
Natl Acad Sci U S A 105(1):129134
8. Lopez G, Maietta P, Rodriguez JM, Valencia
A, Tress ML (2011) Firestar advances in the
prediction of functionally important residues.
Nucleic
Acids
Res
39(Web
Server
issue):W235W241
9. Roy A, Kucukural A, Zhang Y (2010)
I-TASSER: a unified platform for automated
protein structure and function prediction. Nat
Protoc 5(4):725738
10. Oh M, Joo K, Lee J (2009) Protein-binding
site prediction based on three-dimensional
protein modeling. Proteins 77(Suppl 9):
152156
11. Roche DB, Tetchner SJ, McGuffin LJ (2011)
FunFOLD: an improved automated method
for the prediction of ligand binding residues
using 3D models of proteins. BMC
12. Brylinski M, Feinstein WP (2013) eFindSite:
improved prediction of ligand binding sites in
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
protein models using meta-threading, machine

learning and auxiliary ligands. J Comput Aided
Mol Des 27(6):551567
Roy A, Zhang Y (2012) Recognizing protein
ligand binding sites by global structural alignment and local geometry refinement. Structure
20(6):987997
Lopez G, Ezkurdia I, Tress ML (2009)
Assessment of ligand binding residue predictions in CASP8. Proteins 77(Suppl 9):
138146
Schmidt T, Haas J, Gallo Cassarino T, Schwede
T (2011) Assessment of ligand-binding residue
predictions in CASP9. Proteins 79(Suppl
10):126136
Wass MN, Sternberg MJ (2009) Prediction of
ligand binding sites using homologous structures and conservation at CASP8. Proteins
77(Suppl 9):147151
Zhang Y (2008) Progress and challenges in
protein structure prediction. Curr Opin Struct
Biol 18(3):342348
Liu J, Rost B (2001) Comparing function and
structure between entire proteomes. Protein
Sci 10(10):19701979
Rose PW, Bi C, Bluhm WF, Christie CH,
Dimitropoulos D, Dutta S et al (2013) The
RCSB Protein Data Bank: new resources for
research and education. Nucleic Acids Res
Illergard K, Ardell DH, Elofsson A (2009)
Structure is three to ten times more conserved
than sequence a study of structural response
in protein cores. Proteins 77(3):499508
Konc J, Janezic D (2010) ProBiS algorithm for
detection of structurally similar protein binding sites by local structural alignment.
Bioinformatics 26(9):11601168
Keskin O, Tsai CJ, Wolfson H, Nussinov R
(2004) A new, structurally nonredundant,
diverse data set of proteinprotein interfaces
and its implications. Protein Sci 13(4):
10431055
Keskin O, Nussinov R (2005) Favorable scaffolds: proteins with different sequence, structure and function may associate in similar ways.
Protein Eng Des Sel 18(1):1124
Gherardini PF, Wass MN, Helmer-Citterich M,
Sternberg MJ (2007) Convergent evolution of
enzyme active sites is not a rare phenomenon. J
Mol Biol 372(3):817845
Totrov M (2011) Ligand binding site superposition and comparison based on atomic
property fields: identification of distant homologues, convergent evolution and PDB-wide
clustering of binding sites. BMC Bioinformatics
12(Suppl 1):S35
395
26. Xie L, Bourne PE (2008) Detecting evolutionary

relationships across existing fold space, using
sequence order-independent profile-profile
alignments. Proc Natl Acad Sci U S A
105(14):54415446
27. Lee HS, Im W (2013) Ligand binding site
detection by local structure alignment and its
performance complementarity. J Chem Inf
Model 53(9):24622470
28. Konc J, Janezic D (2010) ProBiS: a web server
for detection of structurally similar protein
binding sites. Nucleic Acids Res 38(Web Server
issue):W436W440
29. Huang B, Schroeder M (2006) LIGSITEcsc:
predicting ligand binding sites using the
Connolly surface and degree of conservation.
BMC Struct Biol 6:19
30. Weisel M, Proschak E, Schneider G (2007)
PocketPicker: analysis of ligand binding-sites
with shape descriptors. Chem Cent J 1:7
31. Tripathi A, Kellogg GE (2010) A novel and
efficient tool for locating and characterizing
protein cavities and binding sites. Proteins
78(4):825842
32. Hendlich M, Rippmann F, Barnickel G (1997)
LIGSITE: automatic and efficient detection of
potential small molecule-binding sites in proteins. J Mol Graph Model 15(6):359363, 389
33. Brady GP Jr, Stouten PF (2000) Fast prediction and visualization of protein binding pockets with PASS. J Comput Aided Mol Des
14(4):383401
34. Laskowski RA (1995) SURFNET: a program
for visualizing molecular surfaces, cavities, and
intermolecular interactions. J Mol Graph
13(5):323330, 307308
35. Nayal M, Honig B (2006) On the nature of
cavities on protein surfaces: application to the
identification of drug-binding sites. Proteins
63(4):892906
36. Yu J, Zhou Y, Tanaka I, Yao M (2010) Roll: a
new algorithm for the detection of protein
pockets and cavities with a rolling probe sphere.
37. Liang J, Edelsbrunner H, Woodward C (1998)
Anatomy of protein pockets and cavities: measurement of binding site geometry and implications for ligand design. Protein Sci 7(9):
18841897
38. Binkowski TA, Naghibzadeh S, Liang J (2003)
CASTp: computed atlas of surface topography
of proteins. Nucleic Acids Res 31(13):
33523355
39. Dundas J, Ouyang Z, Tseng J, Binkowski
A, Turpaz Y, Liang J (2006) CASTp: computed atlas of surface topography of proteins
with structural and topographical mapping of
396
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.

functionally annotated residues. Nucleic Acids
Res 34(Web Server issue):W116W118
Zhu H, Pisabarro MT (2011) MSPocket: an
orientation-independent algorithm for the
detection of ligand binding pockets.
Le Guilloux V, Schmidtke P, Tuffery P (2009)
Fpocket: an open source platform for ligand
pocket detection. BMC Bioinformatics 10:168
Schmidtke P, Le Guilloux V, Maupetit J,
Tuffery P (2010) fpocket: online tools for protein ensemble pocket detection and tracking.
Nucleic Acids Res 38(Web Server issue):
W582W589
Yang LW, Bahar I (2005) Coupling between
catalytic site and collective dynamics: a requirement for mechanochemical activity of enzymes.
Structure 13(6):893904
Shih CH, Chang CM, Lin YS, Lo WC, Hwang
JK (2012) Evolutionary information hidden in
a single protein structure. Proteins 80(6):
16471657
Glaser F, Morris RJ, Najmanovich RJ,
Laskowski RA, Thornton JM (2006) A method
for localizing ligand binding pockets in protein
structures. Proteins 62(2):479488
Ghersi D, Sanchez R (2009) EasyMIFS and
SiteHound: a toolkit for the identification of
ligand-binding sites in protein structures.
Silberstein M, Dennis S, Brown L, Kortvelyesi
T, Clodfelter K, Vajda S (2003) Identification
of substrate binding sites in enzymes by computational solvent mapping. J Mol Biol
332(5):10951113
Laurie AT, Jackson RM (2005) Q-SiteFinder:
an energy-based method for the prediction of
proteinligand binding sites. Bioinformatics
21(9):19081916
Morita M, Nakamura S, Shimizu K (2008)
Highly accurate method for ligand-binding site
prediction in unbound state (apo) protein
structures. Proteins 73(2):468479
Ngan CH, Hall DR, Zerbe B, Grove LE,
Kozakov D, Vajda S (2012) FTSite: high accuracy detection of ligand binding sites on
unbound protein structures. Bioinformatics
28(2):286287
An J, Totrov M, Abagyan R (2004)
Comprehensive identification of druggable
protein ligand binding sites. Genome Inform
15(2):3141
Cheng G, Qian B, Samudrala R, Baker D (2005)
Improvement in protein functional site prediction by distinguishing structural and functional
constraints on protein family evolution using
computational design. Nucleic Acids Res
33(18):58615867
53. Hernandez M, Ghersi D, Sanchez R (2009)

SITEHOUND-web: a server for ligand binding
site identification in protein structures. Nucleic
Acids Res 37(Web Server issue):W413W416
54. Soga S, Shirai H, Kobori M, Hirayama N
(2007) Use of amino acid composition to predict ligand-binding sites. J Chem Inf Model
47(2):400406
55. Edelsbrunner H, Facello M, Fu R, Liang J
(1995) Measuring proteins and voids in proteins. In proceedings of the twenty-eighth
Hawaii international conference on system sciences, Vol. 5: Biotechnology Computing,
IEEE Computer Society Press, Los Alamitos,
CA. pp 256264
56. Mehio W, Kemp GJ, Taylor P, Walkinshaw MD
(2010) Identification of protein binding surfaces using surface triplet propensities.
57. Xie ZR, Hwang MJ (2012) Ligand-binding
site prediction using ligand-interacting and
binding site-enriched protein triangles.
58. Xie ZR, Liu CK, Hsiao FC, Yao A, Hwang MJ
(2013) LISE: a server using ligand-interacting
and site-enriched protein triangles for prediction of ligand-binding sites. Nucleic Acids Res
41(Web Server issue):W292W296
59. Xie ZR, Hwang MJ (2010) An interactionmotif-based scoring function for protein
ligand docking. BMC Bioinformatics 11:298
60. Capra JA, Laskowski RA, Thornton JM, Singh
M, Funkhouser TA (2009) Predicting protein
ligand binding sites by combining evolutionary
sequence conservation and 3D structure. PLoS
Comput Biol 5(12):e1000585
61. An J, Totrov M, Abagyan R (2005) Pocketome
via comprehensive identification and classification of ligand binding envelopes. Mol Cell
Proteomics 4(6):752761
62. Zhang Z, Li Y, Lin B, Schroeder M, Huang B
(2011) Identification of cavities on protein surface using multiple computational approaches
for drug binding site prediction. Bioinformatics
27(15):20832088
63. Huang B (2009) MetaPocket: a meta approach
to improve protein ligand binding site prediction. OMICS 13(4):325330
64. Kawabata T (2010) Detection of multiscale
pockets on protein surfaces using mathematical
morphology. Proteins 78(5):11951211
65. Chang DT, Oyang YJ, Lin JH (2005)
MEDock: a web server for efficient prediction
of ligand binding sites based on a novel optimization algorithm. Nucleic Acids Res 33(Web
Server issue):W233W238
66. Fukunishi Y, Nakamura H (2011) Prediction
of ligand-binding sites of proteins by molecular
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
docking calculation for a random ligand library.

Protein Sci 20(1):95106
Gutteridge A, Bartlett GJ, Thornton JM
(2003) Using a neural network and spatial
clustering to predict the location of active
sites in enzymes. J Mol Biol 330(4):
719734
Laskowski RA, Luscombe NM, Swindells MB,
Thornton JM (1996) Protein clefts in molecular recognition and function. Protein Sci
5(12):24382452
Liang X, Zhao J, Hajivandi M, Wu R, Tao J,
Amshey JW et al (2006) Quantification of
membrane and membrane-bound proteins in
normal and malignant breast cancer cells isolated from the same patient with primary
breast
carcinoma.
J
Proteome
Res
5(10):26322641
Cole ST (2002) Comparative mycobacterial
genomics as a tool for drug target and antigen
discovery. Eur Respir J Suppl 36:78s86s
Almen MS, Nordstrom KJ, Fredriksson R,
Schioth HB (2009) Mapping the human membrane proteome: a majority of the human membrane proteins can be classified according to
function and evolutionary origin. BMC Biol 7:50
Laskowski RA (2001) PDBsum: summaries
and analyses of PDB structures. Nucleic Acids
Res 29(1):221222
Singla N, Goldgur Y, Xu K, Paavilainen S,
Nikolov DB, Himanen JP (2010) Crystal
structure of the ligand-binding domain of the
promiscuous EphA4 receptor reveals two distinct conformations. Biochem Biophys Res
Commun 399(4):555559
Ekins S, Kortagere S, Iyer M, Reschly EJ, Lill
MA, Redinbo MR et al (2009) Challenges predicting ligandreceptor interactions of promiscuous proteins: the nuclear receptor PXR. PLoS
Comput Biol 5(12):e1000594
Burris TP, Montrose C, Houck KA, Osborne
HE, Bocchinfuso WP, Yaden BC et al (2005)
The hypolipidemic natural product guggulsterone is a promiscuous steroid receptor
ligand. Mol Pharmacol 67(3):948954
von Eichborn J, Murgueitio MS, Dunkel M,
Koerner S, Bourne PE, Preissner R (2011)
PROMISCUOUS: a database for networkbased drug-repositioning. Nucleic Acids Res
Chiu YY, Lin CT, Huang JW, Hsu KC, Tseng
JH, You SR et al (2013) KIDFamMap: a database of kinase-inhibitor-disease family maps for
kinase inhibitor selectivity and binding mechanisms. Nucleic Acids Res 41(Database
issue):D430D440
van Linden OP, Kooistra AJ, Leurs R, de
Esch IJ, de Graaf C (2013) KLIFS: a
79.
80.
81.
82.
83.
84.
85.
86.
87.
88.
89.
90.
91.
397
knowledge-based structural database to navigate

kinaseligand interaction space. J Med Chem
57(2):249277
Skolnick J, Brylinski M (2009) FINDSITE: a
combined evolution/structure-based approach
to protein function prediction. Brief Bioinform
10(4):378391
Gandhi NS, Mancera RL (2012) Prediction of
heparin binding sites in bone morphogenetic
proteins (BMPs). Biochim Biophys Acta
1824(12):13741381
Krick R, Busse RA, Scacioc A, Stephan M,
Janshoff A, Thumm M et al (2012) Structural
and functional characterization of the two
phosphoinositide binding sites of PROPPINs,
a beta-propeller protein family. Proc Natl Acad
Sci U S A 109(30):E2042E2049
Yu DJ, Hu J, Huang Y, Shen HB, Qi Y, Tang
ZM et al (2013) TargetATPsite: a templatefree method for ATP-binding sites prediction
with residue evolution image sparse representation and classifier ensemble. J Comput Chem
34(11):974985
Khare H, Ratnaparkhi V, Chavan S, Jayraman
V (2012) Prediction of protein-mannose binding sites using random forest. Bioinformation
8(24):12021205
Gandhi NS, Freeman C, Parish CR, Mancera
RL (2012) Computational analyses of the catalytic and heparin-binding sites and their interactions with glycosaminoglycans in glycoside
hydrolase family 79 endo-beta-d-glucuronidase
(heparanase). Glycobiology 22(1):3555
Jmol: an open-source Java viewer for chemical
structure s in 3D. http://www.jmol.lorg
Lopez G, Valencia A, Tress ML (2007)
Firestar prediction of functionally important
residues using structural templates and alignment reliability. Nucleic Acids Res 35(Web
Zhang Y (2009) I-TASSER: fully automated
protein structure prediction in CASP8.
Proteins 77(Suppl 9):100113
Zhang Y (2007) Template-based modeling and
free modeling by I-TASSER in CASP7.
Proteins 69(Suppl 8):108117
Zhang Y (2008) I-TASSER server for protein
3D structure prediction. BMC Bioinformatics
9:40
Roche DB, Buenavista MT, Tetchner SJ,
McGuffin LJ (2011) The IntFOLD server: an
integrated web resource for protein fold recognition, 3D model quality assessment, intrinsic
disorder prediction, domain prediction and
ligand binding site prediction. Nucleic Acids
Res 39(Web Server issue):W171W176
Boeckmann B, Bairoch A, Apweiler R, Blatter
MC, Estreicher A, Gasteiger E et al (2003)
398
The SWISS-PROT protein knowledgebase and

its supplement TrEMBL in 2003. Nucleic
Acids Res 31(1):365370
92. McKusick VA (1998) On the naming of clinical disorders, with particular reference to
eponyms. Medicine (Baltimore) 77(1):12
93. Kalidas Y, Chandra N (2008) PocketDepth: a
new depth based algorithm for identification of
ligand binding sites in proteins. J Struct Biol
161(1):3142
94. Tan KP, Varadarajan R, Madhusudhan MS

(2011) DEPTH: a web server to compute
depth and predict small-molecule binding cavities in proteins. Nucleic Acids Res 39(Web
95. Volkamer A, Kuhn D, Rippmann F, Rarey M
(2012) DoGSiteScorer: a web server for automatic binding site prediction, analysis and druggability assessment. Bioinformatics 28(15):
20742075
Chapter 18
Information-Driven Structural Modelling
ofProteinProtein Interactions
JooP.G.L.M.Rodrigues, EzgiKaraca, andAlexandreM.J.J.Bonvin
Abstract
Proteinprotein docking aims at predicting the three-dimensional structure of a protein complex starting
from the free forms of the individual partners. As assessed in the CAPRI community-wide experiment, the
most successful docking algorithms combine pure laws of physics with information derived from various
experimental or bioinformatics sources. Of these so-called information-driven approaches, HADDOCK
stands out as one of the most successful representatives. In this chapter, we briefly summarize which
experimental information can be used to drive the docking prediction in HADDOCK, and then focus on
the docking protocol itself. We discuss and illustrate with a tutorial example a classical proteinprotein
docking prediction, as well as more recent developments for modelling multi-body systems and large
conformational changes.
Key words Biomolecular interactions, Information-driven docking, Conformational changes,
Multi-body docking, HADDOCK, Molecular modelling
1 Introduction
Docking is defined as the modelling of the three dimensional (3D)
structure of a molecular complex from its known unbound
constituents. It was developed to aid in the structural elucidation
of transient or weak interactions, which can be challenging to characterize experimentally due to, for example, difficulties in crystallization or because the molecular weight rules out a thorough
classical NMR analysis. The advent of explicit treatment of molecular flexibility, together with better and more efficient algorithms
for both sampling and scoring, has earned docking a solid reputation amongst experimentalists. In turn, this attention brought new
challenges such as the prediction of large molecular assemblies,
proteinnucleic acid complexes, high-throughput predictions of
entire metabolic pathways, or understanding the molecular origins
of binding affinity and specificity [1, 2].
399
400
Joo P.G.L.M. Rodrigues et al.
Docking predictions rely usually on a combination of shape

complementary and some energy functions to define the conformational landscape of the interacting molecules and identify near-
native models, respectively. Since most of these functions are
borrowed from molecular dynamics and structure prediction
software, they suffer from the same limitations due to their approximate character [3]. To gain the upper hand, available (experimental)
information was incorporated into the algorithms to greatly
enhance their accuracy. In fact, this approachinformation-driven
dockinghas become the most reliable and successful in the docking community, as shown in the latest assessment of the community-
wide docking assessment experiment (CAPRI) [4, 5].
Traditionally, information was incorporated in docking predictions as a post-sampling filter. This filter approach is simple and
straightforward: it first blindly generates a large pool of models and
then removes or penalizes models that do not agree with the data.
Its disadvantage lies in the need for a large number of solutions that
should cover, ideally, the entire conformational search space. Often,
as in the case of large or extremely flexible systems, the computational cost associated with exhaustive sampling of the search space is
prohibitively high. Therefore, an alternative is to incorporate the
data directly during the sampling stagethe so-called informationdriven docking. This effectively biases the algorithm to visit only
regions of the interaction space that respect the information contained
in the data and thus enriches the pool of generated models with
correct models, provided of course the information is correct
(which can be a pitfall of data-driven approaches) [6].
HADDOCK is an information-driven docking software [7]
that was originally adapted from the NMR automated structure
determination approach ARIA [8]. It uses CNS (Crystallography
and NMR System) [9, 10] as computation engine and, as such, has
access to a variety of energy functions that allow inclusion of
various NMR-derived parameters such as chemical shift perturbations, NOE distances, residual dipolar couplings, and pseudocontact shifts to drive the docking prediction (please see the
Chapter 16 by Vranken etal. for a description of the various NMR
data). Throughout the years, HADDOCK has been extended to
support other types of information commonly generated in the
wet-lab such as mutagenesis, chemical cross-linking, and SAXS,
as well as dry-lab interface predictions from bioinformatics
methods [1113]. How this information is incorporated into
HADDOCK depends on its characteristics. For instance, information such as NMR chemical shift perturbation (CSP), alanine scanning mutagenesis, chemical cross-linking, or EPR distances are
translated into distance restraints. These restraints can have different levels of accuracy and ambiguity (one-to-many or one-to-one
relationships, i.e., some highly ambiguous, e.g., from CSP,
some very specific, e.g., EPR distances). These are generally
Information-Driven Protein Docking
401
implemented as an additional term in the energy function used to

describe the conformational landscape. Lower resolution sources
of information such as SAXS, Cryo-EM, or collision cross-section
data (CCS) from ion-mobility mass spectrometry are currently
used only for scoring in HADDOCK; this is done by measuring
the discrepancy between the structural properties back-calculated
from the generated models and the experimentally measured
values. Information about the radius of gyration of the molecule
can also be extracted from SAXS data; this one-dimensional value
can in principle also be restrained during the sampling.
In this chapter, we will focus on information-driven docking
with HADDOCK and describe how different types of data can be
used in the modelling. We will illustrate this in a protocol example
making use of HADDOCK and NMR chemical shift perturbation
data to model the phosphoryl transfer complex between the signal
transducing proteins HPr and IIa(glucose) of E. coli (PDB entry
1GGR) [14].
2 Theory
This section briefly discusses various useful information sources,
how these can be used to drive docking predictions, and describes
the HADDOCK strategy to produce structural models of biomolecular complexes. Since HADDOCK uses CNS for its structure
calculations, details on the implementation of particular restraint
type is best found in the CNS Web site (http://cns-online.org/
v1.3) or related publications [9, 10].
2.1 Sources
ofInformation
forData-Driven
Docking
2.1.1 Common NMR
Structural Information
Sources
2.1.2 NMR Chemical
Shift Perturbations
The majority of data derived from NMR, such as NOE inter-proton

distances, hydrogen bonds, residual dipolar couplings, relaxation
diffusion anisotropy, pseudo-contact shifts and paramagnetic relaxation enhancements, can be applied directly to docking. Please refer
to the Chapter 16 for a detailed explanation of these restraints.
Some NMR experiments and other techniques particularly
useful for the structural elucidation of interactions are shortly
described below.
Chemical Shift Perturbations are a simple strategy to identify
regions of a protein that interact with the partner molecule. Atomic
nuclei register changes in their chemical environment as small perturbations in their well-defined chemical shifts. By labelling one of
the components of the complex and titrating the other partner
unlabelled, it is possible to follow which chemical shifts are displaced when compared to the spectra of the isolated partner, since
the proximity of the partner will alter the chemical environment of
those residues in the interface. The reverse experiment (first partner unlabelled and second labelled) can be done to map the
402
interface on the other partner. The downside of this technique is

that it cannot distinguish between perturbations caused by the
proximity of the partner molecule and those caused by allosteric
effects, conformational changes upon binding, or solvent reshuffling at the interface.
2.1.3 NMR
Cross-Saturation
In cross-saturation experiments, one of the interacting partners is

2
H/15N uniformly labelled. The only observable protons are those
that can exchange back with protons of water (e.g., amide protons). Upon irradiation with a radiofrequency pulse, the unlabeled
protein protons become instantly saturated due to spin diffusion
effects. Afterwards, cross-relaxation phenomena transfer this saturation to neighboring interfacial protons on the labelled protein,
reducing their peak intensity in a 15N HSQC spectrum. Reversing
the labelling on the partners allows this experiment to pinpoint
accurate information on the binding interface. Since this experiment relies on direct through-space interactions, it is more reliable
than chemical shift perturbation experiments, particularly for interactions where large conformational changes occur upon binding.
2.1.4 Hydrogen/
Deuterium Exchange
H/D exchange provides information on the solvent accessible residues of a protein. In a deuterated medium, amide protons exposed
to the solvent exchange rapidly while those buried by the protein
structure do not. Upon interaction, the interface of the proteins
also becomes inaccessible to solvent exchange. Following this event
by either NMR with 15N HSQC spectra or by mass spectrometry
reveals the solvent-accessible surface of the bound complex and,
indirectly, the interfacial residues.
2.1.5 NMR
PseudocontactShifts
Pseudocontact shift (PCS) experiments require a paramagnetic ion

attached to the protein, much like paramagnetic relaxation
enhancement experiments, and are usually measured in 15N HSQC
or 13C HSQC spectra [15]. When comparing a reference (diamagnetic) spectrum with a paramagnetic spectrum recorded in the
presence of, for example, a paramagnetic lanthanide ion, PCSs can
be measured as the differences in the chemical shifts between both
spectra. Intramolecular PCSs can be used to optimize the tensor parameters of the protein to which the lanthanide is attached,
while intermolecular PCSs can be used to obtain the anisotropic
tensor parameters with respect to the second protein. Since
both tensors are theoretically equal, they can be used to derive
relative orientations between the two interacting proteins.
Furthermore, given the distance-dependence of the PCS effect,
this experiment also provides (long-range) distance information
between the proteins.
2.1.6 Residual Dipolar

Couplings
Residual dipolar couplings (RDCs) are a chemical phenomenon

manifested as an increase or decrease in the magnitudes of multiplet splittings that can be seen in undecoupled NMR spectra [16].
403
These couplings can be measured in solution by inducing a weak

alignment of the molecule, which can be done using a variety of
methods [17]. RDCs provide information on the orientation of
the internuclear vector of the two atoms for which the RDC is
measured (e.g., NC, NH) with respect to the three global axes
of the alignment tensor. This information can be used in the
context of docking to orient the binding partners with respect to
each other, thus reducing the number of degrees of freedom to be
sampled during the docking calculation [18, 19].
2.1.7 Long Distance
Information
NOE-derived distances are typically short, below 56. Larger

distances can be reported by other experimental sources, such as
chemical cross-linking detected by mass spectrometry (the distance
depends on the linker length and flexibility), EPR (2080),
FRET (~50). These can be used to define upper limits in the
docking predictions. These techniques been used, for instance, in
the characterization of very large molecular assemblies that are not
amenable to NMR [2022].
2.1.8 Low-Resolution
Shape Information
(Cryo-EM, SAXS,CCS)
Low-resolution information obtained through Cryo-EM, SAXS,

and CCS experiments provide overall shape information that can
be used either to limit the conformational space search in sampling
stages, or to filter incorrect models after the sampling. Cryo-EM
information is often simpler and faster to obtain compared to
NMR experiments, in particular for very large multi-body assemblies like chaperonins or the nuclear pore complex, since there are
less requirements for sample preparation and there is no peak
assignment step to be carried out, although it still requires plenty
of manual intervention and can become challenging when conformational heterogeneity is present. The resulting electron density
maps allow, depending on the resolution, the identification and
positioning of the interacting partners in the density map. SAXS
provides a scattering curve and allows the calculation of the radius
of gyration of the assembly. It can also be used to generate molecular envelopes that can be used in a similar manner as Cryo-EM
maps. Collision Cross Section (CCS) data from ion mobility mass
spectrometry provides a rotationally averaged two-dimensional
projection of the molecules. Currently, HADDOCK implements
SAXS (and CCS data) as a filter [23]. These can be best used after
the rigid-body energy minimization step, and only for complexes
whose components show asymmetry in their molecular shapes.
2.1.9 Bioinformatics
Predictions
In the absence of experimental data, there is a wealth of information

stored in sequence and structure databases that can be manipulated to provide predictions on the interface of the complex [11].
Of the many algorithms and Web servers available to predict protein interfaces, one is routinely used in tandem with HADDOCK:
CPORT [24]. This server makes use of the structure of the free
proteins to search for possible interfacial residues by sequence and
404
structure homology. Other approaches based on correlated mutation

from evolutionary records have been proposed that predict
unambiguous contacts between interacting partners [25, 26].
2.2 ProteinProtein
HADDOCKing
Docking and flexible refinement in HADDOCK are performed in

three successive stages:
2.2.1 Docking Protocol
it0: Rigid-Body Energy Minimization (RBEM).

it1: Semi-Flexible Simulated Annealing (SA) in Torsion Angle
Space (TAD/SA).
water: Restrained Molecular Dynamics in Explicit Solvent.
These are preceded by a structure/topology generation stage
that rebuilds missing atoms if necessary and a post-processing stage
in which various energy terms, restraint violations, and intermolecular contact are analyzed.
Rigid-Body Energy
Minimization (RBEM,
it0)
In the initial docking stage, the interacting partners are first separated
in space and each is randomly rotated around its center of mass to
remove any orientational bias. They are then subjected to a rigidbody energy minimization protocol, where first only the orientation
of the partners is optimized, and then both rotations and translations
are allowed, effectively resulting in the docking of the molecules.
Given the fast calculation of each docking model at this stage, it is
typically worth generating a large number of models to cover the
interaction space. By default, 1,000 are written to disk, although
10,000 are sampledeach model is the result of five internal docking
trials with, for each, the 180-rotated solution around the normal to
the interface being sampled as well. These models are then ranked
according to the HADDOCK score (see below), and a fraction of
these is selected for further flexible refinementtypically 200.
Semi-flexible
Simulated Annealing
inTorsion Angle Space
(TAD/SA, it1)
The second stage of the HADDOCK protocol fine-tunes each complex by flexible refinement of its interface. This second stage starts
with a rigid-body SA step to optimize the orientation of the components. Then, the side chains of the interfaceautomatically defined
for each docking model as all residues within 10 of a partner
moleculeare allowed to move in a second SA stage. Athird and
final SA stage optimizes both backbone and side-chains of the interface residues to allow for some conformational rearrangements.
Finally, a short energy minimization in Cartesian space relaxes the
models. A new ranking of the models is produced at this step, but,
usually, all models are allowed to undergo the third and final
refinement step in explicit solvent.
Restrained Molecular
Dynamics inExplicit
Solvent (Water)
By default, both the RBEM and TAD/SA stages do not include

any explicit description of the solvent (see Note 1). The docking is
performed in vacuum with a dielectric constant () of 10 for the
electrostatic Coulomb energy term (should be set to 78in case of
405
proteinDNA docking). To improve the network of hydrogen

bonds and electrostatic interactions at the interface, as well as
increase the realism of the prediction, the third and final step of the
HADDOCK protocol is a short restrained molecular dynamics
simulation in Cartesian space in a shell of explicit solvent, either
TIP3P (water, 8 shell) or DMSO (lipid-mimic, 12.5 shell).
2.2.2 HADDOCKScore
All structure calculations in HADDOCK are bound to a set of

energetics terms, which together form the HADDOCK score.
This score is a weighted sum of terms whose weights depend on
the stage of the HADDOCK protocol:
(it0) E=0.01 EvdW+0.1 Eelec+1.0 Edesolv0.01 BSA+0.01 EAIR
(it1) E=1.0 EvdW+1.0 Eelec+1.0 Edesolv0.01 BSA+0.1 EAIR
(water) E=1.0 EvdW+0.2 Eelec+1.0 Edesolv+0.01 EAIR
where the van der Waals and electrostatic energy terms, EvdW and
Eelec, are represented by LennardJones and Coulomb potentials,
Edesolv is an empirical desolvation term developed by FernandezRecio et. al. [27], BSA is the buried surface area of the model in
ngstrom, and EAIR is the energy reflecting the accordance of the
model to the input restraints (the distance-based ones). Other
terms might be included, such as Esym for symmetry restraints or
ERDCs for residual dipolar coupling, depending on the application.
2.2.3 Clustering ofFinal

Solutions
HADDOCK models are not analyzed on a per model basis. Instead,

it is assumed that groups of structurally similar models with overall
low HADDOCK score are the best representatives of the near-
native conformation. To form these groups, HADDOCK offers
two clustering algorithms. The default choice in HADDOCK2.2 is
a contact-based algorithm that groups models based on the similarity of their contact networks at the interfacefraction of common contacts (FCC) [28]. This is particularly efficient for large
molecules, multi-body assemblies, and symmetrical complexes.
The other choice is a standard RMSD-based clustering algorithm
[29] that uses the backbone interface-ligand RMSD as a distance
measure. The backbone interface-ligand RMSD is calculated by
first fitting the models on the interface of the first component of
the complex (which should be the largest). Then, the RMSD is
computed on the interface of the remaining components. The
defaults cutoffs for each algorithm are 0.75 for FCC clustering and
7.5 for backbone interface-ligand RMSD.
2.3 Restraints
Implemented
inHADDOCK
Several experimental techniques provide information on residues

that are potentially involved in the interaction, but fail to report on
the specificity of the residue pair interactions (i.e., they produce
surface patches but not the specific pairwise residue contacts).
HADDOCK implements this information as Ambiguous Interaction
Restraints (AIRs). This concept, similar to ambiguous NOEs [30],
is designed to create an attraction between the interfaces during the
2.3.1 Ambiguous
Interaction
Restraints(AIRs)
406
docking without favoring a particular orientation. To support this

implementation, HADDOCK divides interacting residues in two
classesactive and passivethat differ in their contribution to the
binding event. Active residues are usually those involved directly in
the interaction, for which there is strong experimental or predictive
information, while passive residues comprise those that are surface
accessible neighbors of active residues. Active residues will be forced
to be at the interface, otherwise generating a restraint violation,
while passive residues may or may not be at the interface (if not, no
violation is generated). AIRs are generated between each active
residue on one partner and all active and passive residues on the
other partner(s). The total number of AIRs is equal to the sum of
active residues. False negatives (missing information) are dealt with
usually by automatic definition of passive residues, while false positives can be minimized by randomly removing a fraction (50%, by
default) of the restraints for each docking trial. Internally,
HADDOCK (or rather CNS) uses lists of active and passive residues
for each interacting partner to define an effective distance for each
active residue. This distance is defined between an active residue i
of protein A and all active and passive residues of protein B as,
eff
diAB
N A atom N resB N B atom 1

= 6
m k =1 n dm n
kB =1
iA kB
iA =1
where NA atom indicates all atoms of residue i on protein A, NresB

indicates each active and passive residue on protein B, NB atom indicates all atoms of the residue k on protein B, and d is the Euclidean
distance between atoms miA and nkB. The effective distances are
restrained to a target value by a flat bottom harmonic potential
with upper and lower bounds. This potential switches to a linear
potential beyond the upper bound to avoid large forces that could
cause the calculations to fail. By default, HADDOCK sets the
effective upper bound to 2 and the lower bound to 0 (the van
der Waals energy term prevents potential atom overlap), meaning
that if an effective distance deff greater than 2 separates pair of
restrained residues, these feel an attractive force. This seemingly
small distance is a by-product of the mathematical formula
(it grows smaller with the number of distances entering the sum),
and in reality translates to real distances of 35 between atoms of
the active/passive residues in the model.
2.3.2 Unambiguous
Distance Restraints
When the information is sufficiently accurate to allow unambiguous pairing of atoms or residues between partner molecules, it is
possible to incorporate it in HADDOCK as unambiguous distance
restraints. These use the same functional form of the distance
restraining potential as the AIRs (flat bottom harmonic potential
with upper and lower bounds transitioning to a linear potential
407
after the upper bound to avoid large forces). In practice, ambiguous

and unambiguous boils down to the syntax in which the restraints
are written: either with multiple atom selections linking one residue or group of atoms to many in the partner molecule, or just to
one specific pair of atoms. The main difference is that unambiguous restraints are never randomly discarded while ambiguous
restraints are by default. Ambiguous and unambiguous distance
restraints can therefore be combined and written in any of the two
types of distance restraint files. Next to these two types, HADDOCK
also allows the definition of a third class of restraints as hydrogen
bond restraints. These are treated in a similar manner as unambiguous restraints. Each restraint file can be activated or deactivated at
various stages of the protocol.
2.3.3 Symmetry
Restraints
If there is a priori information that hints or determines that the

complex assumes a symmetrical arrangement, it is possible to
incorporate this in the docking prediction and thus restrict quite
substantially the conformational search space. HADDOCK supports several types of cyclic and dihedral symmetries, implemented
through combinations of symmetrical distance restraints [31, 32]:
C2, C3, C4, D2, and C5. Next to the symmetry restraints,
non-crystallographic symmetry restraints (NCS restraints) can also
be defined: these ensure that two molecules are similar without
imposing any symmetry operation between them, equivalent to
restraining the RMSD between the two molecules to be zero.
2.3.4 Center ofMass

Restraints
To ensure compactness of solutions, for example when using symmetry restraints without experimental information, HADDOCK
allows the definition of distance restraints between the geometric
centers of mass of the molecules (based only on CA atoms). These
so-called center of mass restraints can also be used when there is no
information about the binding interface (ab initio docking), albeit
with a lower chance of success since the only factor in play are the
physical terms of the scoring function.
2.3.5 Other Orientational

Restraints
Next to the distance-based restraints defined above, HADDOCK

supports a variety of orientational restraints from NMR, including
residual dipolar couplings [19] and relaxation anisotropy data that
are useful in defining the relative orientation of the molecules, and
pseudo-contact shifts that provide both distance and orientational
information [33]. For details refer to the Chapter 16.
3 Materials
The following Subheading4 describes an example protocol on
the usage of HADDOCK to model the interaction between two
proteins by using experimental information. In order to follow the
408
example, the reader must have a number of software programs

installed on his/her computer, together with the data provided in
the Extra Materials (extras.springer.com). The protocol should be
run on a GNU/Linux system (or under Mac OSX).
3.1 HADDOCK v2.2
HADDOCK can be obtained free of charge for academic users at

http://nmr.chem.uu.nl/haddock. In principle, little is required to
install and configure HADDOCK (detailed instructions are provided with the software). Also, a mailing-list is available to all current and potential users to provide advice and troubleshooting for
any problem(s) encountered during installation and usage (http://
groups.yahoo.com/group/haddock-discuss). Furthermore, the
WeNMR project provides several tutorials, tips, and a help center
related to HADDOCK and might also be of interest to all users
(http://www.wenmr.eu/wenmr/support/documentation/nmr-
services/haddock). Finally, a Web server version of HADDOCK is
also available, free of charge for academic users that offers a user-
friendly interface to all the powerful features of this docking software (http://www.haddocking.org).
3.2 Crystallography
andNMR Suite (CNS)
v1.3
CNS can be obtained free of charge for academic users at http://

cns-online.org/v1.3. The usage of particular experimental information to drive the docking, such as pseudo-contact shifts (PCS)
requires the installation of an additional module and recompilation of
CNS.Otherwise, the binaries provided at the Web address are sufficient to run HADDOCK.HADDOCK v2.2 is designed for CNS
v1.3, so ensure that the versions of the software are appropriate.
3.3 ProFit
The ProFit software calculates the root mean square deviation of

atomic coordinates between molecules. It is designed to be the
ultimate protein least squares fitting program and it supports a
powerful zone selection syntax. It can be obtained free of charge
for academic users at the authors Web site (http://www.bioinf.
org.uk/software/profit/).
3.4 NACCESS
NACCESS is a software program that uses the Lee & Richards

method [34] to calculate the solvent accessible area of a molecule
from its three-dimensional PDB coordinates. It is available free of
charge to academic and nonprofit organizations (http://bioinf.
manchester.ac.uk/naccess/).
3.5 PyMOL
PyMOL is an open-source molecular visualization system offering

a large number of features and extensible through Python scripts.
It is available in several formats depending on the affiliation and
needs of the user at www.pymol.org.
3.6 Grace (xmgrace)
409
Grace is a simple 2D plotting software program that can be

obtained free of charge here: http://plasma-gate.weizmann.ac.il/
Grace.
4 Methods
4.1 Modelling
ofComplexes with
HADDOCK
We describe here the use of the HADDOCK2.2 package for the

modelling of a proteinprotein complex using chemical shift perturbation data. We will use data from the haddock2.2/examples/e2a-hpr directory. You should first copy this directory to
the directory you are working in (see Note 2):
4.1.1 Preparation ofPDB

Files andInput Data
Make sure that your input models are compliant with the PDB
format, particularly, the presence of an END statement as the last
line of the file. Furthermore, the segment identifier (characters
7376in each ATOM statement) and chain identifier fields (character 22in each ATOM statement) should be empty strings (i.e., filled
with spaces). If you use a crystal structure, make sure that there are
no double occupancies or residue insertions. If you are using an
ensemble of models, split the file in individual files that contain
only one structure (see Note 3).
As input data, you should combine chemical shift perturbation
data (or other data indicating residues at the interface) and solvent
accessibility data calculated with NACCESS: use only those
residues that have both a high enough chemical shift perturbation
(see Note 4) and a high enough relative accessibility. In the example, the residue solvent accessibilities calculated with NACCESS
are already provided in the files e2a_1F3G.rsa and hpr/hpr_
rsa_ave.lis (the latter containing the average for the 10 starting models for hpr). From these files you can select the residues
with high enough (e.g., >~40%) accessibility (see Note 5). You
could calculate the accessibility values yourself using the following
command:
4.1.2 Definition ofActive

andPassive Residues
Passive residues are defined as the solvent accessible surface neighbors of active residues. To define and visualize them you can use a
molecular visualization program, for example PyMOL,
410
Start by coloring the active residues, for example in red. Then,

filter out the residues with a low solvent accessibility, using either
the output of NACCESS, recommended, or an embedded tool of
the visualization program (e.g., get_area command in PyMOL).
Next, select all surface neighbors within a certain cutoff radius (e.g.,
5), and that are solvent accessible, to define the passive residues
and color them for example in green. In the e2a-hpr example,
several PyMOL scripts are provided with the respective residues
already colored according to this scheme: e2a_pymol_active.
pml, e2a_pymol_active_passive.pml and similar for hpr.
You can load these scripts in PyMOL using the following
commands:
You will use the active and passive residues for both molecules
to generate Ambiguous Interaction Restraints (AIRs); for this go
to the HADDOCK GenTBL service (http://haddock.chem.
uu.nl/services/GenTBL/) and follow the instructions. You should
save the resulting file as ambig.tbl in the working directory;
note that, in the e2a-hpr example directory, an example file
named e2a-hpr_air.tbl is already present and can be used for
comparison (see Note 6).
4.1.3 Setup ofaNew
Run: new.html
To set up a new run, go to the project setup page on http://www.

nmr.chem.uu.nl/haddock, click on start a new project and follow the instructions. Depending on the experimental data you
have available, you can input various data files such as ambiguous
restraints, unambiguous restraints, RDCs etc. PCS restraints are
not yet supported in the Web site, but an example case is provided
with the HADDOCK software. After saving the new.html file to
disk, type haddock2.2 in the same directory. This will generate a
run directory containing all necessary information to run haddock.
An example of a new.html file can be found in the e2a-hpr
directory as new.html-refe. (see Note 7) and is displayed
below. Such a file can in principle also be created by manual
editing.
4.1.4 Run.cns
411
The next step is to define all parameters to perform the docking

run. For this, enter the newly created directory:
You will find a file called run.cns containing all the parameters to run the docking, which deserves special attention. You need
to edit this file and define a few parameters such as the location of
the CNS executable and the queue command to use. Other options
such as the semi-flexible segments at the interface, or fully flexible
segments (see Note 8), the number of models to generate at each
stage, the clustering algorithm and cutoff, and the force constants
for the several energy terms are also defined there. You can edit
your run.cns file manually or via Project Setup on http://
www.nmr.chem.uu.nl/haddock. More information is available via
the run.cns option in the manual section on http://www.nmr.
chem.uu.nl/haddock.
412
4.1.5 Docking Run
To actually start the docking run with HADDOCK type in the

directory containing the run.cns file (see Note 9).
As more extensively explained in Subheading2 before and
the docking section in the HADDOCK manual, the entire protocol consists of three stages. An initial topology and structure
generation step validates and builds the structure files to be used in
the docking. The initial models as provided by the user are written
to data/sequence/.
Topology and model generation. The resulting topologies

(*.psf) and coordinates (*.pdb) files are written to the
begin/directory (see Notes 10 and 11). There is one output
file per chaingenerate_X.out, where X is the segment
identifier given in run.cnsthat must be checked for errors
if there is a problem at this stage (see Note 12).
Randomization and rigid body energy minimization. The
docked models are written to structures/it0/. When all
models have been generated, HADDOCK will write the PDB
files with names sorted according to the HADDOCK score
(weights defined in the run.cns) to file.cns, file.list,
and file.nam in the same directory. The number of trials
(ntrials, by default 5) and the sampling of 180 rotated
solutions (rotate180_0, by default true) can be modified in
run.cns.
Semi-flexible simulated annealing. The best models after rigid

body docking (defined at structures_1 in run.cns
and by default 200) will be subjected to a semi-flexible simulated annealing (SA) in torsion angle space. The temperatures
and number of steps for the various stages are also defined in
run.cns. The resulting refined models are written into
structures/it1. The numbering of the file names reflects
their rank from the previous step (e.g., complex_1.pdb is
the refined best ranked structure in it0 according to the
HADDOCK score). At the end of the calculation, HADDOCK
generates the file.cns, file.list, and file.nam files as
in the previous stage (see Note 13). At the end of this stage,
the models are analyzed and the results are placed in the
structures/it1/analysis directory (see the analysis
sections below).
Flexible explicit solvent refinement. The choice of the solvent in
which to refine the models is defined in run.cns (solvent)
and can be either water or dmso. The resulting models are
written in the structures/it1/water directory. The
numbering in the files here matches that of the previous stage
(e.g., complex_1w.pdb is the water refined complex_1.
413
pdb of it1). At the end of the explicit solvent refinement,

HADDOCK generates the file.cns, file.list, and file.
nam files. Finally, the models are analyzed and the results are
placed in the structures/it1/water/analysis directory (see the analysis sections).
4.1.6 Automatic Analysis
A number of analysis scripts are automatically run after the semi-

flexible and explicit solvent refinement stages and the results placed
in structures/it1/analysis and structures/it1/
water/analysis, respectively. Here we discuss a few of the
most relevant output files.
e2a-hpr_fcc.disp: contains the pairwise FCC matrix; this
file is used as input for FCC clustering. If the clustering algorithm is RMSD, then the filename is e2a-hpr_rmsd.disp.
The FCC measure, unlike RMSD, is asymmetric (FCC(AB)
FCC(BA)) so it produces a full matrix.
cluster.out: contains the clusters generated from the
abovementioned matrix. The clusters are numbered according
to their size (number of models in the cluster) and not according to their HADDOCK score. This is related to the algorithm
used to cluster the models.
noe.disp: contains the number of distance restraints violations per structure and averaged over the ensemble over all
distance restraint classes and for each class (unambiguous,
ambiguous, hbonds) separately. Similar files are generated
when you have RDC (sani.disp), relaxation anisotropy
(dani.disp), or PCS (pcs.disp) restraints.
energies.disp: contains the various energy terms per
model and averaged over the ensemble.
ana_*.lis: there is a set of files called ana*.lis where * can
be dihed_viol, dist_viol_all, hbond_viol, hbonds, nbcontacts,
noe_viol_all, noe_viol_ambig, noe_viol_unambig. The viol
refers to violations, and those files contain listings of violations
including the number of times a restraint is violated as well as
the average distance and violation per restraint. In addition,
ana_hbonds.lis gives a listing of hydrogen bonds, and
ana_nbcontacts.lis a listing of non-bonded contacts.
ene-residue.disp: contains intermolecular energies for all
interface residues.
nbcontacts.disp: contains non-bonded contacts.
4.1.7 Manual Analysis
An important part of the analysis needs to be performed manually.

A number of analysis scripts and programs are provided in the
tools directory. These allow you to collect various statistics on the
generated models and more importantly to perform re-clustering of
solutions and their analysis on a per-cluster basis.
414
Fig. 1 Plot of HADDOCK scores versus interface RMSD from the lowest energy model for the three stages of
the docking protocol (blue, green, and red, for it0, it1, and water refinement, respectively). One can clearly see
a funnel at low RMSD values becoming more apparent after flexible refinement
Collecting statistics of the models with ana_structure.csh: This

script should be run once the file.list file has been created.
It extracts various energy terms, violation statistics, and the
buried surface area from each PDB file and calculates the
RMSD of each structure compared to the lowest energy one (if
the location of ProFit is defined (see installation and software
links on http://www.nmr.chem.uu.nl/haddock)). The output
are several files named structures*.stat that contain
the same information sorted in different ways. Usually, the
most important file is structures_haddock-sorted.
stat. From this file, you can generate a plot of the HADDOCK
score as a function of the RMSD to the lowest energy model
and investigate if the run produces an energy funnel, meaning that the low energy models should have small RMSD values and the high energy models should have large RMSD
values (see Fig.1). A script called make_ene-rmsd_graph.
csh is provided in $HADDOCKTOOLS and it produces a Grace
compatible plot file. Specify two columns to extract data from
and a filename:
This will generate a file called ene_rmsd.xmgr, which

you can display using xmgrace:
415
Clustering of solutions: The clustering is run automatically in it1/

analysis and it1/water/analysis, based on the criteria
defined in the run.cns file. However, try using different cutoffs
for the clustering since it is difficult to know a priori the best
RMSD/FCC cutoff. This value depends on the system under
study and the number of experimental restraints used to drive the
docking (see Note 14). For FCC clustering, the script to use it
cluster_fcc.py, while for RMSD clustering, use the C program cluster_struc (this should have been compiled during
the installation of HADDOCK). The scripts read the appropriate
e2a-hpr_*.disp file containing the pairwise matrix and generate clusters. The usage is (in the analysis directory):
or
Here cutoff indicates the FCC/RMSD cutoff to determine if two models belong in the same cluster. For FCC clustering, there are several options that can be modulated (type
python cluster_fcc.py h for the list and their explanation). In the RMSD clustering script, min_size is the minimum number of models in a cluster (typically a number like 4)
and -f is an optional full-linkage clustering algorithm.
In either case, the output looks like the following:
The numbers after the arrow correspond to the rank in

file.nam. The sorted models are also in the analysis
directory. For example, 2 corresponds to the second model in
the analysis directory, which is the second structure listed in
file.list in it1 or it1/water.
Analysis of the clusters with ana_clusters.csh: This script takes

the output of the clustering script, by default analysis/
cluster.out, to perform an analysis of the various clusters,
calculating average energies, RMSDs, and buried surface area
per cluster. The following runs the analysis on all clusters:
The -best # is an optional argument to generate additional files with cluster averages calculated only on the best (#)
ranked models of a cluster according to their HADDOCK
score. This recommended option allows removing the dependency of the cluster averages on the size of the respective clusters (see Note 15). The ana_clusters.csh script analyses
the clusters in a similar way as the ana_structures.csh
script, but in addition generates average values over the models
416
belonging to one cluster. Itcreates a number of files for each

cluster containing the cluster number clustX in the name. It
also creates files containing various averages over the clusters,
cluster_xxx.txt; these contain the average and standard
deviation of various terms such as intermolecular energy
(xxx=ene) etc. Also, files combining all the above information and sorted based on various criteria are provided: clusters.stat that contains the various cluster averages and
clusters_xxx-sorted.stat where xxx is the energy
term according to which the values are sorted (e.g., xxx=ene
for intermolecular energy, etc.). Again, the most relevant output file is clusters_haddock-sorted.stat, or rather
clusters_haddock-sorted.stat_bestX.
Rerunning the HADDOCK analysis on a cluster basis: Having

performed the cluster analysis, you can now rerun the
HADDOCK analysis for the best models of each cluster to
obtain violation and energetics details and statistics. To run
this analysis, we need the cluster-specific file.nam_clust#,
file.list_clust# and file.cns_clust# files. A script in
the tools/directory called make_links.csh will move the
original file.nam, file.list and file.cns files to file.
nam_all, file.list_all, file.cns_all and the same
with the analysis directory. It will then create links to the
appropriate files (file.nam_clust#, ) and to a new
analysis_clust#/directory.
For example, to rerun the analysis for the best 10 models of

the first cluster type in the water directory:
The cd command brings you back into the main run directory
from where you start again HADDOCK.Only the analysis of the
best 10 models of the first cluster in the water will be run. Once
this is finished, go to the respective analysis directory and inspect
the various files. The RMSD from the average models should now
be low (check rmsave.disp).
Having run the HADDOCK analysis on a cluster basis for each
cluster, you should now have new directories in the water directory, called analysis_clustX_best10. Each of these analysis
directories contains now cluster specific statistics. You can also
visualize the clusters, using for example PyMOL.We provide a Perl
script in the tools/directory, joinpdb, which allows concatenation of the various PDB files into one single ensemble file:
417
Fig. 2 Superimposition of the top model of the best scoring cluster onto the native
structure (PDB ID 1GGR). The molecules were superimposed on backbone atoms
of E2A, which is shown in white surface representation with the phosphorylated
histidine colored according to the atom types (blue, red, and orange, for nitrogen,
oxygen, and phosphorous, respectively). The HPR molecules are shown in cartoon
representation (the model in blue and the native in peach) and the histidine residue involved in the phosphate transfer in ball-and-stick. The model is in excellent
agreement with the native structure (interface RMSD=0.97). The proximity of
the two histidines across the interface, which was not defined as a restraint in
HADDOCK, is consistent with the biological function of this phosphotransferase
complex
In general, the top ranked models of the cluster with the lowest
HADDOCK score are considered the representatives of the biological system. However, scoring in docking remains a difficult
problem and we do recommend, if possible, using additional independent information to validate the results (e.g., mutagenesis
data). The selected model should explain as much as possible all
what is known about the system (see Fig.2).
4.2 Other Docking
Scenarios
Although the previous example illustrates a canonical dimer docking,

HADDOCK also supports more advanced protocols. Users can
model large macromolecular complexes, address substrate-induced
conformational changes, and deal with extremely flexible peptides.
These protocols are briefly explained below.
4.2.1 Multi-body
Docking
HADDOCK allows users to include up to six molecules and dock

them simultaneously [31]. Multi-body HADDOCKing, as this
protocol is called, follows the same rules of the original pairwise
docking protocol requiring only that each molecule is restrained to
418
at least another one of the system. The restraints between the

different molecules are defined with the same syntax described in
the case of dimer docking, and can be generated via the Web server
indicated before: http://haddock.chem.uu.nl/services/GenTBL/.
Here, we recommend the user to also turn on center of mass
restraints, in order to ensure the compactness of the resulting models.
If there is any cyclic and/or dihedral symmetry present, the user can
activate the built-in symmetry restraints and impose them between
and/or within each molecule (see Subheading2.3.3).
4.2.2 Flexible Multi-
domain Docking
Modelling binding-induced large conformational changes is a

major challenge of the docking community, since it requires sampling a vast and intricate conformational space. Unfortunately,
addressing such a number of degrees of freedom is often out of
reach for most of the current sampling methods, including the one
of HADDOCK.To tackle this challenge, we developed a special
application of the multi-body docking protocol [35] that divides
to conquer: the flexible partner is cut at hinge regions, and thus
dissected into rigid domains that allow HADDOCK to sample a
wider range of motions during rigid-body energy minimization.
The identification of the hinge regions can be carried out using
normal mode analysis, such as that provided by the Web server
HingeProt (see Note 16). To ensure molecular integrity and biological realism, we define connectivity restraints (in the form of
distance restraints) between the separated domains. These are first
defined with a maximum distance of 10, to allow sampling of
large range of motions. They are then shortened to a peptide bond
distance (1.3) at the water refinement step. Imposing different
connectivity restraints is possible by submitting both unambig.
tbl and hbonds.tbl restraint files (a copy except for the connectivity distance), and changing the stages when these are active
in run.cns (options unambig_firstit (0), hbond_firstit
(2) and unambig_lastit (1), hbind_lastit (2)). It is
also necessary to define the artificial termini as uncharged and the
first three residues starting from the cut hinge as fully flexible.
4.2.3 ProteinPeptide
Docking
Albeit the other end of the size spectrum, small systems such as
peptides are also challenging regarding sampling. Their extreme
flexibility and the many conformations they can adopt upon binding makes them challenging to model and require usually long
molecular dynamics simulations or other advanced sampling
methods, none of which is possible or feasible, time-wise, for use
in HADDOCK.
To cover the conformational landscape of peptides, we developed a shortcut approach. In this custom-tailored protocol,
the peptide is provided as an ensemble of three most common
conformations: -helical, -strand, and polyproline II (see Note 17).
Additionally, the number of MD steps in the flexible refinement
419
stage needs to be increased fourfold to improve sampling efficiency

(from 500/500/1000/1000 to 2000/2000/4000/4000).
Finally, the peptide is defined entirely as fully flexible and the clustering algorithms are adapted for small molecules (see Note 14).
5 Notes
1. HADDOCK has a special featuresolvated dockingthat
allows water molecules to be introduced at the interface of the
complex for entire duration of the docking protocol. This feature should only be used when the experimental information is
accurate enough to drive the docking and the interface is
expected to be wet. In short, solvated docking starts by surrounding each molecule by a shell (approximately 4 wide) of
water molecules, optimized via a short MD simulation, prior to
the RBEM stage. After the minimization, all water molecules
that are not at the interface are removed. At the interface, only
a fraction of the molecules is kept (by default 25%), with the
removal being carried out via a biased Monte Carlo sampling
method whose criteria is based on a statistical potential of amino
acidwater contact propensities. Finally, energetically unfavorable water molecules (those with a positive intermolecular
energy) are removed, which might lead to a complete desolvation of the interface, and another round of RBEM is performed
to optimize the final complex. The remaining of the HADDOCK
protocol remains unchanged, with the difference that interfacial
water molecules might be included in the further refinement.
We refer the reader to the following references for an in-depth
explanation of solvated docking in HADDOCK: [3639].
2. HADDOCK must be correctly installed for the $HADDOCK
environment variable to be defined. Check the installation
instructions provided with the software.
3. If your input PDBs contains missing segments, this might lead
to domains drifting away during the refinement stage. To avoid
this, simply define a few unambiguous distance restraints
between CA atoms from the various sub-domains, setting
the actual measured distance as a target distance and the
bounds to 0.0. The same can be done to ensure that an ion
coordination geometry is properly maintained. Missing residues at the interface or in hinge regions must be handled with
extreme care not to compromise the biological integrity of the
models. Missing atoms, on the other hand, are not problematic since HADDOCK rebuilds them based on the topology
files of the force field, as long as the residue name is defined in
them. Also, termini charges are very important for the docking
protocol, as they can lead to artificial interactions. By default,
420
termini are charged but they can be neutralized by using an

appropriate linkage file (protein-allhdg5-4-*.link
files in the toppar/directory). For example, to have both termini of molecule A uncharged, simply add in run.cns (option
prot_link_A) the appropriate linkage file (protein-allhdg5-4-noter.link). Another important point concerns
ions; if proper care is not taken, they can be problematic during the torsion angle dynamics stage. HADDOCK has an inbuilt mechanism that defines artificial bonds to chelate the
ion to the protein but it relies on proper ion naming. Check
these names in the covalions.cns script and add yours if
necessary. Also, make sure that their name in the PDB file
matches the ion names defined in the ion.top file in the
toppar/directory. To avoid that a N- or C-terminal patch be
applied to them, they should also be defined in the
topallhdg5.4.pep file (look for the "first IONS" and
"last IONS" statements).
4. We have developed an automated method to discriminate the

significant CSPSAMPLEX [40]. It compares two sets of
chemical shifts from two different samples (e.g., bound/
unbound), and using the three-dimensional structure of the
molecules, returns the confidence for each residue to be in a
perturbed or unperturbed state.
5. The accessibility cutoff is not a hard limit; check the accessibilities and possibly include residues with lower accessibilities but
with functionally important groups.
6. The syntax of the restraints is what determines their (un)

ambiguous character, not the filename where they are stored:
ambig.tbl, unambig.tbl, or hbonds.tbl. This allows
for example to mix unambiguous and ambiguous restraints in
the same file. The difference lies in the random removal option
(noecv=true), which is applied only to ambig.tbl. In
principle, ambig.tbl; unambig.tbl and hbonds.tbl
could be used concurrently, for example, to provide extra
NOEs or other data (e.g., FMD connectivity restraints) for
which one wants to use different force constants or for which
there is exceptional certainty.
7. An important setting in new.html is the value of N_COMP.
This should be set equal to the number of components of the
complex (2in case of a dimer, 3 for a trimer, etc.). Note that it
can also be set to 1, in which case HADDOCK can be used for
refinement instead of docking.
8. HADDOCK allows the definition of fully flexible regions:
these are treated as fully flexible throughout all stages, except
the initial rigid-body docking. This should be useful for cases
where part of a structure are disordered or unstructured or
when docking small flexible molecules onto a protein.
421
9. This command causes HADDOCK to run in the background

and all output to be redirected to the haddock.out file. If at
some stage HADDOCK stops producing new models and the
run is not yet finished, unzip and search for error messages in
the output files:
xxx.out.gz is a particular output file, in which you should
look for ERR. Also, kill the current HADDOCK process:
Here id is the process id that is returned by the ps ef

command.
You can restart a HADDOCK run, but before doing that,
make sure to delete any *.out file from the run directory (see
Note 12). HADDOCK will proceed from where it was in the
calculations.
10. The OPLS force field used by HADDOCK is a mixed united/
all-atom force field. This means that protons do not have van
der Waals parameters of their own. Instead these are accounted
for in the heavy atom parameters to which they are attached. In
HADDOCK, by default, nonpolar hydrogen atoms are deleted
in order to speed up the calculation. This does not affect the
resulting models significantly since the missing hydrogen
atoms are actually accounted for in the united atoms parameters. You can change this behavior by setting delenph=true
in run.cns. This should be done in case classical NOE distance restraints are used, or in situations where the hydrogen
atoms are extremely relevant (e.g., small molecule docking).
11. Topology generation is often the most problematic stage in
HADDOCK.While most amino acids and their most common
modifications are supported in HADDOCK, small ligands and
exotic modifications to amino acids or nucleic acid bases that
are not described in the force field will give an error and halt
the docking protocol. A list of the supported modified amino
acids is given here: http://haddock.science.uu.nl/services/
HADDOCK/library.html. For those molecules not in the list,
the user is left to generate their own parameterization scheme,
using for example PRODRG [41], ACPYPE [42], or ATB
[43], and provide the necessary files (topology and parameters
in CNS format) in the toppar/directory: ligand.top and
ligand.par. Molecule parameterization is not simple
though, and must be approached and carried out with extreme
care and expertise (for a best-practices guide for parameterization, although under a different force-field, check the following reference: [44]).
422
12. If a particular stage of HADDOCK fails, in case of a model not

being generated, the run being stopped accidentally, etc.,
reissuing the command haddock2.2 might not be enough.
HADDOCK makes use of the output files to control the flow
of the docking run. When a particular step is initiated,
HADDOCK write a.job file and a.out file, and when it is
completed, the .out file is compressed to a.out.gz file. To
safely restart a HADDOCK run, remove all.out files prior to
the issuing of the haddock2.2 command.
13. A typical error at this stage is that a couple of models in it1

are not successfully generated. Often, this can be solved by
changing the random seed in run.cns (iniseed, by default
917) and restart HADDOCK (see Note 12). Otherwise, try to
decrease the timestep (e.g., 0.001 instead of 0.002) and/or
the temperature of the first two SA stages (e.g., 1,000 or 500K
instead of 2,000). HADDOCK, by default, tries this automatically in case a model fails. If none of this works, simply copy
the missing models from the it0 directory so that the run can
proceed. This can be done using the copy-missing.csh
script provided in the tools directory with as arguments the
file root name and the number of the missing model.
14. If only a small fraction of the models do fall into clusters, try
decreasing the cutoff in case of FCC clustering, or increasing it
in case of RMSD clustering. If all models fall in one single
cluster, and the restraints are not that restrictive, try the reverse.
This is particularly relevant for proteinsmall molecule docking, for which a tailored FCC clustering algorithm using small
molecule atomsprotein residue contacts is available. For the
RMSD clustering, due to the nature of interface-ligand RMSD,
the resulting RMSD values are larger than would be obtained
by fitting on all chains of the complex, which explains the large
cutoff value that is used by default (7.5).
15. It is better to use matching number of models (e.g., 4) to compare the cluster statistics in order to remove cluster size effects.
In our experience, the size of a cluster does not always correlate with its quality/score and as such cannot be used as an
indicator of the quality of the cluster.
16. The choice of the hinge region(s) where to cut the molecules should be made with the structural integrity of the molecule in mind. As such, hinges are favored if located at the end
of -helices and -strands, or in loops. The experimental temperature factors should also be taken into account when deciding between possible hinges. The molecule should then be cut
at the first peptide bond following the predicted hinge region.
17. The peptide conformations can be generated using PyMOL
and setting the / dihedral angles according to the desired
secondary structure: 57 and 47 for -helix, 139 and
135 for -strand, and 78 and 149 for polyproline II.
423
References
1. Melquiond AS, Karaca E, Kastritis PL etal
(2012) Next challenges in proteinprotein
docking: from proteome to interactome and
beyond. Comput Mol Sci 2:642651
2. Kastritis PL, Bonvin AM (2013) Molecular origins of binding affinity: seeking the
Archimedean point. Curr Opin Struct Biol
23(6):868877
3. Schlick T, Collepardo-Guevara R, Halvorsen
LA etal (2011) Biomolecular modeling and
simulation: a field coming of age. Quart Rev
Biophys 44:191228
4. Janin J (2013) The targets of CAPRI rounds
2027. Proteins 81(12):20752081
5. Lensink MF, Janin J (2013) Docking, scoring
and affinity prediction in CAPRI.Proteins
81(12):20822095
6. de Vries SJ, Melquiond ASJ, Kastritis PL etal
(2010) Strengths and weaknesses of data-
driven docking in critical assessment of prediction of interactions. Proteins 78:32423249
7. Dominguez C, Boelens R, Bonvin AMJJ
(2003) HADDOCK: a proteinprotein docking approach based on biochemical or biophysical information. J Am Chem Soc 125:
17311737
8. Linge JP, Habeck M, Rieping W etal (2003)
ARIA: automated NOE assignment and
NMR structure calculation. Bioinformatics
19:315316
9. Brnger AT, Adams PD, Clore GM etal
(1998) Crystallography & NMR system: a new
software suite for macromolecular structure
determination. Acta Crystallogr D Biol
Crystallogr 54:905921
10. Brunger AT (2007) Version 1.2 of the crystallography and NMR system. Nat Protocol
2:27282733
11. de Vries SJ, Bonvin AMJJ (2008) How proteins get in touch: interface prediction in the
study of biomolecular complexes. Curr Protein
Pept Sci 9:394406
12. Karaca E, Bonvin AMJJ (2013) Advances in
integrative modeling of biomolecular complexes. Methods 59:372381
13. Schmitz C, Melquiond AS, de Vries SJ etal
(2012)
Proteinprotein
docking
with
HADDOCK, NMR of biomolecules: towards
mechanistic systems biology, 1st edn. Wiley-
VCH Verlag GmbH & Co. KGaA, Weinheim,
pp521535
14. Wang G, Louis JM, Sondej M etal (2000)
Solution structure of the phosphoryl transfer
complex between the signal transducing proteins
HPr and IIA(glucose) of the Escherichia coli

phosphoenolpyruvate:sugar phosphotransferase
system. EMBO J 19:56355649

15. Bertini I, Calderone V, Cerofolini L etal
(2012) The catalytic domain of MMP-1 studied through tagged lanthanides. FEBS Lett
586:557567

16. Bax A (2003) Weak alignment offers new
NMR opportunities to study protein structure
and dynamics. Protein Sci 12:116

17. Prestegard JH, Bougault CM, Kishore AI
(2004) Residual dipolar couplings in structure
determination of biomolecules. Chem Rev
104:35193540
18. Tjandra N (1997) Direct measurement of distances and angles in biomolecules by NMR in a
dilute liquid crystalline medium. Science
278:11111114

19. van Dijk ADJ, Fushman D, Bonvin AMJJ
(2005) Various strategies of using residual
dipolar couplings in NMR-driven protein
docking: application to Lys48-linked di-
ubiquitin and validation against 15N-relaxation
data. Proteins 60:367381

20. Kalisman N, Adams CM, Levitt M (2012)
Subunit order of eukaryotic TRiC/CCT chaperonin by cross-linking, mass spectrometry,
and combinatorial homology modeling. Proc
21. Choi UB, Strop P, Vrljic M etal (2010) Single-
molecule FRET-derived model of the synaptotagmin 1-SNARE fusion complex. Nature
17:318324
22. Brunger AT, Strop P, Vrljic M etal (2011)
Three-dimensional molecular modeling with
single molecule FRET.J Struct Biol 173:
497505
23. Karaca E, Bonvin AMJJ (2013) On the usefulness of ion-mobility mass spectrometry and
SAXS data in scoring docking decoys. Acta
Cryst D69:683694, 112
24. de Vries SJ, Bonvin AMJJ (2011) CPORT: a
consensus interface predictor and its performance in prediction-driven docking with
HADDOCK.PloS One 6:e17695
25. Weigt M, White RA, Szurmant H etal (2009)
Identification of direct residue contacts in proteinprotein interaction by message passing.
Proc Natl Acad Sci U S A 106:6772
26. Marks DS, Hopf TA, Sander C (2012) Protein
structure prediction from sequence variation.
Nat Biotechnol 30:10721080
27. Fernndez-Recio J, Totrov M, Abagyan R (2004)
Identification of proteinprotein interaction
424
sites from docking energy landscapes. J Mol

Biol 335:843865
28. Rodrigues JPGLM, Trellet M, Schmitz C etal
(2012) Clustering biomolecular complexes by
residue contacts similarity. Proteins 80:
18101817
29. Daura X, Gademann K, Jaun B etal (1999)
Peptide folding: when simulation meets experiment. Angew Chem Int Ed 38:236240
30. Nilges M, ODonoghue SI (1998) Ambiguous
NOEs and automated NOE assignment. Progr
Nucl Magn Reson Spectros 32:107139
31. Karaca E, Melquiond ASJ, de Vries SJ etal
(2010) Building macromolecular assemblies by
information-driven docking: introducing the
HADDOCK multibody docking server. Mol
Cell Proteomics 9:17841794
32. Nilges M (1993) A calculation strategy for the
structure determination of symmetric dimers
by 1H NMR.Proteins 17:297309

33. Schmitz C, Bonvin AMJJ (2011) Protein
protein HADDocking using exclusively pseudocontact shifts. J Biomol NMR 50:263266
34. Lee B, Richards FM (1971) The interpretation
of protein structures: estimation of static accessibility. J Mol Biol 55:379400
35. Karaca E, Bonvin AMJJ (2011) A multidomain
flexible docking approach to deal with large
conformational changes in the modeling of
biomolecular complexes. Structure 19:
555565
36. van Dijk ADJ, Bonvin AMJJ (2006) Solvated
docking: introducing water into the modelling
of biomolecular complexes. Bioinformatics

22:23402347

37. Kastritis PL, van Dijk ADJ, Bonvin AMJJ
(2012) Explicit treatment of water molecules
in data-driven proteinprotein docking: the
solvated HADDOCKing approach. Methods
Mol Biol 819:355374
38. Kastritis PL, Visscher KM, van Dijk ADJ etal
(2013) Solvated proteinprotein docking using
Kyte-Doolittle-based water preferences. Proteins
81:510518
39. van Dijk M, Visscher KM, Kastritis PL etal
(2013) Solvated protein-DNA docking using
HADDOCK.J Biomol NMR 56:5163
40. Krzeminski M, Loth K, Boelens R etal (2010)
SAMPLEX: automatic mapping of perturbed
and unperturbed regions of proteins and complexes. BMC Bioinformatics 11:51

41. Schttelkopf AW, van Aalten DMF (2004)
PRODRG: a tool for high-throughput crystallography of proteinligand complexes. Acta
Crystallogr D Biol Crystallogr 60:13551363

42. Sousa da Silva AW, Vranken WF (2012)
ACPYPE AnteChamber PYthon Parser interfacE.BMC Res Notes 5:367
43. Malde AK, Zuo L, Breeze M etal (2011) An
automated force field topology builder (ATB)
and repository: version 1.0. J Chem Theor
Comput 7:40264037

44. Lemkul JA, Allen WJ, Bevan DR (2010)
Practical considerations for building GROMOScompatible small-molecule topologies. J Chem
Inf Model 50:22212235
Chapter 19
Identifying Putative Drug Targets and Potential
Drug Leads: Starting Points for Virtual
Screening and Docking
David S. Wishart
Abstract
The availability of 3D models of both drug leads (small molecule ligands) and drug targets (proteins) is
essential to molecular docking and computational drug discovery. This chapter describes a simple approach
that can be used to identify both drug leads and drug targets using two popular Web-accessible databases:
(1) DrugBank and (2) The Human Metabolome Database. First, it is illustrated how putative drug targets
and drug leads for exogenous diseases (i.e., infectious diseases) can be readily identified and their 3D structures selected using only the genomic sequences from pathogenic bacteria or viruses as input. The second
part illustrates how putative drug targets and drug leads for endogenous diseases (i.e., noninfectious diseases or chronic conditions) can be identified using similar databases and similar sequence input. This
chapter is intended to illustrate how bioinformatics and cheminformatics can work synergistically to help
provide the necessary inputs for computer-aided drug design.
Key words Drug, Disease, Drug target, Metabolite, Bioinformatics, Sequence comparison, Chemical
similarity, Exogenous disease, Endogenous disease
Introduction
As most readers have already seen in previous chapters, protein
modeling is a mature field that allows many interesting biological
questions to be addressed using only a computer. Insights gained
through computational modeling have helped us to better understand proteins and their many important structurefunction relationships. While macromolecular modeling has helped enormously
to advance basic biology, one of the central justifications for the
enormous resources that have gone into this field over the past
30 years is the hope that molecular modeling could, one day,
accelerate both drug discovery and drug design [13].
Computational drug discovery is a subfield of macromolecular
modeling that involves the docking or virtual screening of one or
more small-molecule compounds against a chosen protein target.
425
426
David S. Wishart
We distinguish between receptor-based virtual screening and

ligand-based virtual screening, of which the latter is a powerful
technique to identify new ligands based on a set of existing
known ligands. This is, however, outside the scope of this receptor-protein focused book. The small-molecule ligands are called
drug leads and the protein of interest is called the drug target.
Both computer-aided docking and receptor-based virtual screening employ a variety of algorithms that allow the small molecule(s)
to be rapidly rotated and translated around the protein surface
or active site and scored on the basis of their steric fit and/or
predicted free energy [47]. In more advanced packages the
ligand (and even the protein) is allowed to exhibit some conformational flexibility. When an optimal orientation is found or a
particularly high scoring molecule is identified, a drug lead or a
drug mechanism is said to have been discovered. The results
of these computational experiments are used in an iterative fashion by synthetic organic chemists to help design or select
improved lead compounds.
What distinguishes virtual screening from docking is the number of molecules used (screening uses 1,000s, docking uses one),
the objective of the search process (screening identifies drug leads,
docking identifies active sites or mechanisms), and the robustness
or complexity of the docking energy function (docking uses a complex force field, screening does not). There are now many excellent
docking and/or virtual screening software packages such as Dock
[8], AutoDock [9], Gold [10], Glide [11], and AutoDock Vina
[12]. Almost all are freely available. These will be discussed in more
detail in the next chapter.
However it is important to remember that before either virtual
screening or macromolecular docking can begin, a protein target
needs to be identified (and modeled) and a set of potential drug
leads needs to be assembled. This chapter describes how both drug
targets and drug leads can be identified through several easily
accessible Web resources. Specifically we show how putative drug
targets for pathogenic viruses or bacteria can be identified directly
from their genomic sequences and how the 3D structures of putative drug leads and drug targets can be subsequently extracted
from a comprehensive drug and drug-target database called
DrugBank [13]. This chapter also illustrates how human drug targets and potential drug leads for prostate cancer can be similarly
identified and extracted for docking/screening programs using the
Human Metabolome Database (HMDB) [14] and DrugBank.
The intent of this chapter is to give readers the necessary input files
and know-how to proceed to the next steps (docking and screening) in computational drug discovery.
Finding Drug Targets and Drug Leads
427
Theory
When medicinal chemists or pharmaceutical scientists think about
drugs and drug targets they generally classify them into two separate groups: (1) those that are associated with endogenous
human diseases and (2) those that are associated infectious or
exogenous diseases. Endogenous diseases are typically chronic
human disorders or conditions that arise due to germ-line mutations (genetic diseases), somatic mutations (cancer), age (atherosclerosis, immune disorders), or some other internal factors. On
the other hand, exogenous diseases are typically temporary diseases
or conditions that arise from external, nonhuman agents such as
viruses, bacteria, fungi, protozoans, or poisonous animals (snakes,
insects). The vast majority of drug targets (96 %) and drugs (89 %)
are associated with endogenous diseases, while only a tiny minority
of drugs targets (4 %) and drugs (11 %) are actually associated with
exogenous or infectious diseases [13, 15].
2.1 Identifying Drug

Targets and Drug
Leads for Exogenous
Diseases
The identification of putative drug targets and drug leads for exogenous diseases can take one of two paths, both of which depend
substantially on bioinformatics and sequence database comparisons. One can either attempt to identify a completely novel drug
target/drug lead or one can attempt to identify a drug target/
drug lead that is similar (or even identical) to an existing class of
drug targets or drug leads. In both cases, one needs either the
complete protein or DNA sequence of the pathogen of interest.
Fortunately, with the advent of next-generation DNA sequencing,
the entire DNA sequence for hundreds of infectious agents of
interest is already known or can be determined in as little as a day.
If one chooses to identify a completely novel drug target or
drug lead the task is then to identify those genes or proteins in the
genome that are: (1) essential to viability; (2) disease causing; or
(3) presented on the surface of the organism. Surface-bound proteins may be identified by sequence analysis by looking for transmembrane segments using such tools as TMHMM [16] or
PSORTb [17]. Essential genes, especially for bacteria, may be
identified by comparing sequences to existing databases of essential
genes such as in the Database of Essential Genes [18]. Likewise
disease-causing genes can be identified by comparing sequences
between non-pathogenic forms of the microbe with pathogenic
forms (say E. coli O157 vs. E. coli MG1655) or through the identification of pathogenicity islands using tools such as IslandViewer
[19]. Alternately essential genes or disease causing genes may be
experimentally identified through knock-out mutations or deletions. Generally all viral genes in a viral genome are essential while
only 200-300 bacterial genes in a given bacterial genome are essential. Furthermore, among most pathogens, only a small fraction of
428
David S. Wishart
proteins or genes (<20) are typically disease-causing. Once these

druggable genes or protein targets are identified, one must select
for those that are sufficiently different (<35 % identity) from any
human homologues. This prevents any cross-reactivity between
the hosts proteins and the pathogens target proteins. It also prevents any potentially adverse drug interactions. After these nonhomologous protein targets are found one can either search/screen
for an inhibitory molecule or develop a vaccine (using parts of the
surface proteins). When working with completely novel drug targets, it is often difficult to know which lead compounds might
work, so widespread chemical library screening is often used.
If one wishes to find matches to an existing class of drug targets or drug leads the task involves identifying those genes or proteins in the genome of the organism that are similar to known drug
targets. The underlying assumption is that if a novel virus or a
newly identified pathogenic bacterium shares some significant
sequence similarity to a protein that is a known drug target from
another organism, then the same (or similar) drugs may be used to
combat or kill this pathogen. Alternately, these previously known
drugs may serve as potential drug leads for further synthetic modification so as to develop more effective therapies for the organism
of interest. What is needed for this process to work is a database of
known drug target sequences, each of which is linked to a set of
associated drugs. Ideally this database should also include the 3D
structures (known or predicted) of the drug targets and the drugs
themselves. Fortunately such a database exists. It is called DrugBank
[13]. DrugBank is a comprehensive drug database or drug encyclopedia that combines detailed drug (i.e., chemical) data with
comprehensive drug target (i.e., protein) information. Since it first
appeared in 2006 [20], the database has been expanded and
updated multiple times [13, 21]. The latest version of DrugBank
contains more than 7,300 drug entries including >1,550 FDA
approved small molecule and biotech drugs as well as >5,000
experimental drugs. Each compound entry contains detailed structure files in SDF, MOL, and PDB formats. Additionally, nearly
15,000 protein or drug target sequences are linked to these drug
entries, many of which have 3D structures or 3D homology models associated with them.
DrugBank supports standard BLAST sequence queries, including appropriately formatted multiple sequence inputs (i.e., complete proteomes). The output from these queries includes the
name(s) and hyperlinks to the associated drugs and the 3D structures of the drug targets. Once the drugs are identified, it is possible to use DrugBank again to search for similar drugs (based on
structure similarity). The structures of all of these chemical hits
may be downloaded, either as PDB, MOL, or SDF files. SDF files
can be converted to PDB files using the freely available tools
MolConverter (ChemAxon), CACTVS [22] or the Cactus
429
Converter (http://cactus.nci.nih.gov/services/translate/). Thus

by using DrugBank it is possible to rapidly obtain 3D structures of
putative drug targets and the 3D structures of 100s or even 1,000s
of drug leads. These data sets would obviously serve as the basis for
docking or virtual screening studies.
Targets and Drug
Leads for Endogenous
Diseases
Identifying drug targets for endogenous diseases is often far more

challenging than identifying drug targets for infectious or exogenous diseases. This is because most endogenous human diseases
have a complex etiology. With the exception of about 500 [23]
relatively rare, monogenic (single gene) disorders, the vast majority of endogenous diseases are multifactorial or partially polygenic
(multi-gene) in origin. Nevertheless, with the advent of such techniques as microarray analysis, GWAS (genome wide association
studies) or high throughput proteomics, it is now possible to identify large numbers of disease-associated genes relatively rapidly
[24]. To date more than 6,500 human disease genes (mutations,
polymorphisms, copy number variants) for both monogenic and
complex, polygenic diseases have been identified. This information
is being catalogued in many online databases such as OMIM [23],
the Human Metabolome Database [14], VnD [25], and GeneCards
[26]. It is also possible to find diseasegene associations directly
through PubMed or other Web servers such as MedGene [27] and
PolySearch [28]. A comprehensive list of druggable genes is maintained at the druggene interaction database (DGIdb) [29].
Once a list of genes or proteins associated with a given disease
is available (along with their sequences) then it is a matter of doing
a series of similar kinds of sequence searches against DrugBank as
described for Subheading 2.1. However it is also possible to find
additional or even novel drug leads by looking through the Human
Metabolome Database (HMDB). The HMDB, like DrugBank, is
a multipurpose bioinformatics-cheminformatics database containing detailed information about metabolites, their associated
enzymes or transporters and their disease-related properties. The
utility of the HMDB in drug discovery lies in the fact that most
drugs are actually analogs of existing metabolites, cofactors, or signaling molecules. Therefore if one identifies a protein or proteins
in a disease-specific pathway that require a certain metabolite or
cofactor, then these proteins may prove to be good drug targets
and their cofactors or metabolites could prove to be good drug
leads. Indeed many inborn errors of metabolism (phenylketonuria,
alkaptonuria, and galactosemia) are treated through the addition
or removal of metabolites in the diet.
Both DrugBank and the Human Metabolome Database
(HMDB) support single and multiple protein sequence queries and
both produce results that include the name(s) and hyperlinks to the
associated drugs or metabolites and the 3D structures of the corresponding proteins. Once the small molecule leads are identified,
430
David S. Wishart
it is possible to use DrugBank or HMDB again to search for

structurally similar drugs or metabolites. The structures of all these
chemical hits may be downloaded, either as PDB, MOL, or SDF
files. The SDF files can then be converted to PDB files using the
freely available tools MolConverter (ChemAxon) or the Cactus
Web server. Thus by judiciously using DrugBank and/or HMDB it
is possible to rapidly and easily obtain 3D structures of putative
drug targets and the 3D structures of numerous drug leads for
endogenous diseases.
2.3 Sequence
Matching
and Chemical
Compound Matching
This particular chapter focuses on using two different matching

protocols, one for sequence matching and another for chemical
structure matching. Sequence matching, or sequence alignment is
a central feature to much of bioinformatics while chemical structure matching is a central feature to much of cheminformatics.
Sequence alignment is often based on a technique called
dynamic programming. Strictly speaking dynamic programming is
an efficient mathematical technique that can be used to find optimal
paths or routes to multiple destinations or in locating paths that
could be combined to score some maximum value. The application
of dynamic programming to sequence alignment was first demonstrated more than 40 years ago by Needleman and Wunsch [30]. As
these two researchers demonstrated, dynamic programming allows
two or more sequences to be efficiently and automatically lined up,
permitting gaps to be inserted, extended, or deleted to make an
optimal pairwise alignment. In dynamic programming, the two
sequences being compared (say sequence A and sequence B) are
put on either axis of a table. Sequence A might be on the X-axis,
while sequence B might be on the Y-axis. Each letter in the query
sequence is compared to each letter the reference sequence and a
number (based on a scoring matrix and a special recursive function)
is placed in every box or cell that intersects each pair of letters. Once
the table of numbers is filled out, a second stage (called the traceback stage) is undertaken wherein the table is scanned in a diagonal
fashion from the lower right to upper left to look for the highest
scores. The path that is chosen is actually a series of maximum
numbers. When all the scores in this optimal path are added
together, it gives a quantitative measure of the pairwise sequence
similarity while at the same time defining which letters in sequence
A should be matched with the letters in sequence B.
Dynamic programming is a relatively slow, memory intensive
process. However, it can be sped up considerably. For instance, if
look-up tables are used, if advanced statistics are employed, if more
than one letter at a time (a tuple) is scored and if the traceback
search is limited to sections close to the diagonal line then you
have the essence of the BLAST algorithm [31]. This is the very fast
algorithm used to perform most alignments against large sequence
databases. It is also the algorithm used in the sequence searches for
DrugBank and HMDB.
431
Chemical compound matching is fundamentally different than

sequence matching. Instead of trying to match strings of characters, one tries to match substructures, coordinates or geometric
shapes. This is somewhat similar to the idea of structure superposition, which is done with protein structure comparison. However,
because the structures of chemical compounds are far more diverse
than what is seen for proteins, the structure matching utilities in
chemistry have to be slightly more sophisticated. In particular,
chemists must use the concept of subgraph isomorphisms [32] and
adjacency matrices to identify chemical similarity. For substructure
searching the 2D chemical structures of both the query and database compounds must be re-cast as tables that indicate the bond
connectivity between each pair of atoms. These tables, which have
1s for connected atoms and 0s for unconnected atoms are called
adjacency matrices. The name (adjacency matrix) comes from the
fact that they indicate which atoms are adjacent (connected) to
each other. Once prepared, the adjacency matrix from the query
structure is compared to every adjacency matrix in the database.
If substantial sections of the query matrix match to an adjacency
matrix (or portion thereof) in the database, then it is likely that the
two structures are similar. Different scoring schemes and adjustable threshold cutoffs may be used to distinguish strong matches
from weak matches or to identify compounds with particularly
important substructures.
As will be seen in the examples to follow, both sequence searching and chemical structure similarity searching can play an important role in drug target and lead compound identification.
Methods
For this section we will describe two protocols. One will describe the
identification of drug targets and drug leads for a novel retrovirus
that exhibits strong similarity to the AIDS virus (HIV) (see Notes 18).
The other will describe the identification of drug leads (from a
preexisting list of putative drug targets) for prostate cancer.

Targets and Drug
Leads for a Novel Virus
In this example we will use sequence data derived from a recently

sequenced, but unnamed virus that exhibits strong sequence similarity to the human immunodeficiency virus (HIV). This particular
virus has a total of 15 identifiable open reading frames or polyprotein
fragments, which have been fully translated. We will use this sequence
information, in combination with DrugBank to identify several drug
targets, several drug leads and the necessary coordinate files to conduct rational drug design efforts via docking and virtual screening.
1. Start your local Web browser and go to the DrugBank Web
site at http://www.drugbank.ca. The DrugBank homepage
should be visible as should the grey menu bar located near the
432
David S. Wishart
top of the page with the eight clickable titles Home, Browse,
Search, Downloads, About, Help, Tools, Contact Us.
2. Click on the Search link. A submenu should appear that
displays several search options including ChemQuery, Text
Query, Interax Interaction Search, Sequence Search, and
Data Extractor. Select the Sequence Search option. A window with the title Sequence Search should appear (Fig. 1). As
seen in the figure the window contains a standard online
BLAST search form with a text box window, with eight different BLASTP parameter settings. There are also options for the
Drug type and Database to be searched, with a variety of
options. In almost all cases users can leave everything (except
the Drug type and Database selection) in their default position. A unique feature of the Sequence Search program is its
capacity to handle multiple FASTA-formatted sequences. This
allows users to BLAST multiple sequencesor even entire
proteomes.
3. For this example we will be looking for potential drug targets
to a newly isolated retrovirus. To obtain the set of sequences to
paste into the Sequence Search text box, launch a new browser
window and go to: http://www.wishartlab.com/molecularmodelingproteins/virus. Click on the Virus hyperlink. A list
of 15 viral protein sequences should be visible. Select all 15
sequences by clicking a dragging through the window with
your mouse. Copy the sequences (using the Copy option on
your browser or using Ctrl + C).
4. Now click on the Sequence Search browser window to activate
it and paste the sequences into the Sequence Search text box by
clicking your mouse in the text box and using the Paste option
on your browser (or Ctrl + V). You have now pasted 15 different
protein sequences from the newly sequenced retrovirus. Use the
scroll bars on the right side of the text box to see if all 15
sequences are there (numbered Peptide 1 to Peptide 15).
5. Now select the DrugBank sequence database to search. For
this example go down to the bottom of the Sequence Search
window and select Drug Type Approved and Database
Target. This means you will search through all known protein targets of FDA approved drugs. Once this is done, press
the Search button. Within a few seconds the BLAST search for
all 15 input sequences should be completed. The program will
return a text-based BLAST summary for each of the 15 proteins that were submitted. The top portion of the Sequence
Search output consists of a summary of the submitted
sequences. Below that is the BLAST result for the first sequence
(Peptide 1) listing the E-value, the bit score, the query length,
the name of the closest match, and the alignment with the
query sequence at the top and the DrugBank database match
Fig. 1 Screenshot of the DrugBank BLAST search page
433
434
David S. Wishart
Fig. 2 Screeen shot of the output from the DrugBank BLAST search using the 16 viral protein sequences
belonging to a novel retrovirus
below. Matched residues will be displayed in the middle as red

letters. Below the alignment is a series of hyperlinks to a number of drug names (see Fig. 2). Clicking on any of these drug
name hyperlinks will reveal that Peptide 1 may be acted upon
by protease inhibitors.
6. For Peptide 1, click on the hyperlink to Indinavir (users may
select any one of the many hyperlinks in this list). This should
take you to the DrugCard for Indinavir. This page describes
the drug and its mode of action in detail and it suggests that
Indinavir may be able to target this viral protein target as well.
7. Scroll down further through the Sequence Search output
page and look for other sequences in this retrovirus that exhibit
hits to known DrugBank compounds and for drugs that would
be likely to work on these protein targets. In total you should
find that there are at least 24 FDA approved drugs for at least
four different target proteins in this novel retrovirus.
8. Your task now is to create a library of 3D structures for each of
these potential anti-viral drugs. To do so it is necessary to click
on each of the drug names and scroll down the DrugCard page
that is displayed (Fig. 3). Near the top of each page is a picture
of the drug. Below each drug image is a set of hyperlinks indicating Download: MOL, SDF, SMILES, InChI, PDB. Click on
the PDB link and download the PDB text file of the drug lead
(Indinavir in this case). Each of these PDB files was generated
435
Fig. 3 A view of the tabular output found in the DrugCard for Indinavir
using the MolConverter 3D structure generator. You should

repeat this step for all drug structures found in steps 57. You
may also obtain additional drug leads and drug structures by
going to the top of each DrugCard page and clicking on the
button located on the top right corner called Show Drugs
with Similar Structures for All drugs. This will generate a
table of chemically similar drugs (including approved, withdrawn and experimental) that may exhibit potential activity
against these viral proteins. Download the PDB structures for
these compounds as well. You should now have a large collection of PDB files (i.e., 3D structures) of possible drug leads for
each of the unique proteins associate with the virus.
9. To perform docking or virtual screening experiments it will be
necessary to generate 3D structures of each of the protein targets identified through steps 5, 6. For many of the proteins
identified in this exercise it is possible to generate a 3D homology model using Modeller [33], which is a downloadable program or Proteus2 [34] or Swiss-Model (see Chapter 16 for
more details), which are Web servers. Further information
about protein structure modeling is available in Chapters 14
and 15 of this volume. Once you have done this for all the
proteins that can be modeled (not all will have 3D homologues) you should have a large collection of PDB files (i.e.,
3D structures) of the key drug targets for this virus.
10. Use these two sets of structures (one for the small molecule
drug leads, the other for the drug targets) to initiate a virtual
screening run or attempt to dock selected compounds into
their corresponding protein targets.
436
David S. Wishart

Targets and Drug
Leads
for Prostate Cancer
Prostate cancer is the second most common type of cancer in men in

North America. It is responsible for more male deaths than any other
cancer except lung cancer. It is a disease that generally strikes men
over the age of 50, however many factors beyond age, including
genetics and diet, have been implicated in its development. In this
example we will show how a large list of candidate target proteins can
be easily obtained and then quickly reduced. From this list we will
show how potential drug candidates or (anti)metabolites may be
identified using DrugBank and the HMDB. We will also demonstrate how the necessary coordinate files can be obtained to conduct
rational drug design efforts via docking and virtual screening.
1. The first step is to identify a set of disease genes or disease proteins that are upregulated or altered in prostate cancer. This is
best done via a literature review. Users could consider using
the Prostate Gene Database or PGDB (http://www.urogene.
org/pgdb/), a general PubMed literature search, a search
through PolySearch [28], data found in the GeneCards database [26] or the NCBI Gene Expression Omnibus [35]. Once
an initial set of protein names or gene names has been compiled, one should try to select those proteins that appear to be:
(a) enzymes; (b) soluble proteins; (c) able to bind or act upon
relatively unique small molecules. The reason for these selection criteria is that if one wants to develop a small molecule
drug, the drug target should exhibit some propensity to bind a
small molecule. Furthermore, if one wants to perform docking
or virtual screening studies, the protein structure needs to be
known or at least modeled. Since 99 % of all proteins in the
PDB are of soluble proteins or soluble fragments, the need for
soluble protein targets is obviously important, although
advanced methods for membrane protein structure modeling
are covered in this volume. For the purposes of this example, a
reasonably good list of candidate protein targets that fit these
three criteria is given below:
(a) Alpha-methyl-acyl-CoA racemase.
(b) Glyceraldehyde 3-phosphate dehydrogenase.
(c) Pyruvate kinase dehydrogenase.
(d) Pyruvate kinase.
(e) Glycine-N-methyl transferase.
(f) Pipecolic acid oxidase.
(g) Sarcosine dehydrogenase.
(h) Hydroxymethylglutaryl-CoA synthase.
(i) Acetyl-CoA-acetyltransferase.
(j) 3-Oxo-5-alpha-steroid 4-dehydrogenase 1.
2. To retrieve the protein sequences (which are necessary for the
next steps in the analysis) you may start your local Web browser
437
Fig. 4 Screenshot of the UniProt databasee search page
and go to the UniProt Web site at http://www.uniprot.org/.

In the Query box at the top of the page (Fig. 4) type in the
name of each protein candidate and press the return key. A list
of hits from multiple organisms will appear in a tabular format.
Ensure that you select the proteins from Homo sapiens only.
Click on the corresponding protein name or Uniprot Accession
number to open its UniProt protein page. Scroll down the
protein page until the Sequence field is reached. A hyperlink
with the world FASTA should be located just above the
sequence. By clicking on this hyperlink it is possible to retrieve
a FASTA formatted protein sequence file. This process should
be repeated for each protein in the above list. However, to help
save time, a FASTA sequence file for all ten proteins is also
available for download at http://www.wishartlab.com/
molecularmodelingproteins/cancer. To obtain these sequences,
click on the Cancer hyperlink. Select all ten sequences by clicking a dragging through the window with your mouse. Copy
the sequences (using the Copy option on your browser or
using Ctrl + C) into your computer memory buffer.
3. The next step is aimed at finding metabolites or drugs that may
bind, antagonize, inhibit or deactivate these proteins. To find
these molecules, launch a new window within your current
browser and go to the HMDB Web site at http://www.hmdb.
ca. The HMDB homepage should be visible with a simple
438
David S. Wishart
Fig. 5 Screenshot of the HMDB BLAST search page
menu bar located near the top of the page with the seven clickable titles Home, Browse, Search, About, Downloads,
Metabolomics Toolbox, and Contact Us.
4. Click on the Search link. A submenu should appear that displays
nine different search options including Chem Query, Text
Query, Sequence Search, Advanced Search, etc. Select the
Sequence Search option. A window with the title Sequence
Search should appear (Fig. 5). As seen in the figure the window
contains a standard online BLAST search form with a text box
window, with eight different BLASTP parameter settings. A
unique feature of the Sequence Search program is its capacity
439
Fig. 6 Screenshot of the output from a BLAST search against the HMDB using the ten protein sequences identified as potential prostate cancer drug targets
to handle multiple FASTA-formatted sequences. This allows

users to BLAST multiple sequencesor even entire proteomes.
5. Now click on the Sequence Search browser window to activate it and paste the sequences into the Sequence Search text
box by clicking your mouse in the text box and using the Paste
option on your browser (or Ctrl + V). You have now pasted ten
different protein sequences that are potential drug/metabolite
targets. Use the scroll bars on the right side of the text box to
see if all ten sequences are there.
6. Once you have confirmed that all ten sequences have been
pasted in, press the Search button. Within a few seconds the
BLAST search for all ten input sequences should be completed.
The program will return a text-based BLAST summary for each
of the ten proteins that were submitted. The top portion of the
Sequence Search output consists of a summary of the submitted sequences. Below that is the BLAST result for the first
sequence (Alpha-methylacyl-CoA racemase) listing the E-value,
the bit score, the query length, the name of the closest match
and the alignment with the query sequence at the top and the
HMDB database match below. Matched residues will be displayed in the middle as red letters. Below the alignment is a
series of hyperlinks to a number of compound names (see Fig. 6).
7. Scroll down the list until you see the word Sarcosine as one
of the metabolites. Click on the word Sarcosine. This should
take you to the MetaboCard for Sarcosine. This page describes
the metabolite, its structure, its metabolic importance, its metabolic pathways, and the enzymes that act on it.
440
David S. Wishart
Fig. 7 Screenshot of the MetaboCard for Sarcosine. The hyperlinks for the MOL, SDF, and PDB structure files
(below the structure) are also visible
8. Scroll down further through the Sequence Search output

page and look for other sequences that exhibit hits to known
human metabolites and for metabolites that would be likely to
work on these protein targets. Ideal lead metabolites should
be larger, polyatomic molecules (not metals) that are nonessential (not ATP). Many of these compounds are substrates or
products for enzyme reactions. By overloading an enzyme with
a product, it is possible to inhibit its reaction rate. Alternately,
by identifying a chemical analog of an enzyme substrate it is
possible to completely arrest the activity of the enzyme.
9. Your task now is to create a library of 3D structures for each of
these potential anti-cancer drugs. To do so it is necessary to
click on each of the metabolite names and scroll down the
MetaboCard page that is displayed (Fig. 7). Near the top of
each page is a picture of the compound. Below each metabolite
441
image is a set of hyperlinks indicating Download: MOL, SDF,

SMILES, InChI, PDB. Click on the PDB link and download
the PDB text file of each metabolite of interest. You may also
obtain additional drug/metabolite leads and drug structures
by going to the top of each MetaboCard page and clicking on
the button located on the top right corner called Show Similar
Structures. This will generate a table of chemically similar
compounds that may exhibit potential activity against these
proteins. Download the PDB structures for these compounds
as well. You should now have a collection of 1520 PDB files
(i.e., 3D structures) of possible drug leads for each of the prostate cancer associated proteins.
10. Users may also want to employ DrugBank (as described in
Subheading 3.1) to identify additional drug leads using the
ChemQuery tool in this database. Indeed, these efforts would
prove to be quite fruitful for this particular example as
DrugBank contains a number of well known enzyme antagonists. To generate models for the protein targets, we suggest
that users follow steps 9 and 10, as described in Subheading 3.1.
This will allow them to complete the necessary steps required
to set up docking and virtual screening efforts.
Notes
1. The examples given in Subheadings 3.1 and 3.2 are realistic
but somewhat simplified compared to what might be necessary
for real life drug discovery. In particular, the identification of
drug targets always requires some critical assessment of the
utility and viability of the drug target or drug lead. This typically requires a good deal of library research and additional
experimentation. For instance, one must determine whether
the drug target(s) should be inhibited (therefore requiring an
antagonist) or activated (therefore requiring an agonist). As a
general rule, the development of antagonists is generally much
easier than agonists.
2. It is usually a good idea to determine whether the putative
drug target has been previously identified and whether experimental lead compounds have already be explored. Even if a
drug target appears viable, one should take particular care to
determine if the protein is essential, unique, or conditionally
expressed for the associated disease or condition. Nonessential,
nonunique, or continuously expressed proteins are generally
not good drug targets. Likewise proteins with generally weak
affinities (i.e., most carbohydrate binding proteins) or poor
turnover rates often turn out to be poor drug targets.
3. The selection of drug leads also requires some careful
consideration. While DrugBank, HMDB, and PubChem can
442
David S. Wishart
offer many useful suggestions, they are not the only sources for
drug leads. Surveys through the literature or careful searches
through specialized drug-screening databases can often yield
very useful ideas. Once a collection of drug leads has been
identified, it is usually prudent to assess the suitability of the
compound as a drug. Drug compounds must not be too soluble, too lipophilic, too unstable, or too toxic. These requirements are closely related to their physicochemical properties,
which are also related to their Absorption, Distribution,
Metabolism, Excretion, and Toxicity (ADMET).
4. ADMET prediction is becoming increasingly common in
early-stage drug discovery, drug screening, and drug design.
Indeed, many computational chemists would argue that
ADMET prediction is something that should always be done
in the early phases of drug-lead selection. Fortunately there are
now a number of software packages, online servers, and standardized rules (Lipinskis rule of five) to determining the likely
success or drug-likeness that a compound might have.
5. Among existing tools, AdmetSAR [36] and PreADMET
(http://preadmet.bmdrc.org/) probably represent two of the
most comprehensive and complete ADMET servers currently
available.
6. AdmetSAR is both a server and a database with more than
210,000 literature-derived ADMET data values for nearly
100,000 compounds corresponding to 45 kinds of ADMETassociated properties obtained for different proteins, cell types,
and organisms. Through database matches, machine learning
classifiers and rule-based regression models derived from its
large database and various molecular descriptors, the
AdmetSAR server also allows users to predict up to 27 ADMET
properties for query compounds. Some of these properties
include probabilities for bloodbrain barrier penetration,
Caco-2 permeability, intestinal absorption, P-gp inhibition/
substrate status, CYP isotype inhibitor or substrate status, renal
cation transporter substrate status, carcinogenicity, and Ames,
fish, or honeybee toxicity. The server accepts SMILES string
data as input and rapidly returns a hyperlinked list of values,
probabilities, or qualitative classification statements (noninhibitor, toxic, nontoxic, etc.). Each entry is also hyperlinked
to a brief description of the ADMET feature.
7. The PreADMET server (http://preadmet.bmdrc.org/) supports a variety of applications including molecular descriptor
calculations (2,000+ values), drug likeness calculations, Caco-2
cell permeability, MDCK cell permeability, human intestinal
absorption (HIA), skin permeability, bloodbrain barrier
permeability, plasma protein binding, Ames toxicity, and rodent
443
carcinogenicity. The server is nicely designed and provides

detailed references and descriptions about the server output
and how it should be interpreted.
8. Perhaps the most important point to remember for each of the
methods outlined here is that one is generating computerbased predictions. There is no guarantee that any of these predictions (drug targets or drug leads) will turn out to yield a
viable therapeutic or even an interesting lead compound. As
with any prediction in life science, one must always be prepared to thoroughly test the predictions using in vitro assays
and animal models. In many cases the computer predictions
will turn out to be wrong. In rare cases, the initial predictions
may prove to be quite promising. Nevertheless, the results
from any well-constructed wet-bench experiments can and
should be used to help guide subsequent steps involved in the
computational design, docking, and selection of drug leads.
References
1. Geldenhuys WJ, Gaasch KE, Watson M, Allen
DD, Van der Schyf CJ (2006) Optimizing the
use of open-source software applications in drug
discovery. Drug Discov Today 11:127132
2. Kirkpatrick DL, Watson S, Ulhaq S (1999)
Structure-based drug design: combinatorial
chemistry and molecular modeling. Comb
Chem High Throughput Screen 2:211221
3. Wlodawer A, Vondrasek J (1998) Inhibitors of
HIV-1 protease: a major success of structureassisted drug design. Ann Rev Biophys Biomol
Struct 27:249284
4. Mohan V, Gibbs AC, Cummings MD, Jaeger
EP, DesJarlais RL (2005) Docking: successes
and challenges. Curr Pharm Des 11:323333
5. Jain AN (2004) Virtual screening in lead discovery and optimization. Curr Opin Drug
Discov Devel 7:396403
6. Sousa SF, Ribeiro AJ, Coimbra JT, Neves RP,
Martins SA, Moorthy NS, Fernandes PA, Ramos
MJ (2013) Protein-ligand docking in the new
millennium a retrospective of 10 years in the
field. Curr Med Chem 20:22962314
7. Taylor RD, Jewsbury PJ, Essex JW (2002) A
review of protein-small molecule docking methods. J Comput Aided Mol Des 16:151166
8. Shoichet BK, Kuntz ID (1993) Matching
chemistry and shape in molecular docking.
Protein Eng 6:723732
9. Goodsell DS, Morris GM, Olson AJ (1996)
Automated docking of flexible ligands: applications of AutoDock. J Mol Recognit 9:15
10. Jones G, Willett P, Glen RC, Leach AR, Taylor

R (1997) Development and validation of a
genetic algorithm for flexible docking. J Mol
Biol 267:727748
11. Friesner RA, Banks JL, Murphy RB, Halgren
TA, Klicic JJ, Mainz DT, Repasky MP, Knoll
EH, Shelley M, Perry JK, Shaw DE, Francis P,
Shenkin PS (2004) Glide: a new approach for
rapid, accurate docking and scoring. 1. Method
and assessment of docking accuracy. J Med
Chem 47:17391749
12. Trott O, Olson AJ (2010) AutoDock Vina:
improving the speed and accuracy of docking
with a new scoring function, efficient optimization, and multithreading. J Comput Chem
31:455461
13. Knox C, Law V, Jewison T, Liu P, Ly S, Frolkis
A, Pon A, Banco K, Mak C, Neveu V,
Djoumbou Y, Eisner R, Guo AC, Wishart DS
(2011) DrugBank 3.0: a comprehensive
resource for omics research on drugs. Nucleic
Acids Res 39(Database issue):D1035D1041
14. Wishart DS, Jewison T, Guo AC, Wilson M,
Knox C, Liu Y, Djoumbou Y, Mandal R, Aziat
F, Dong E, Bouatra S, Sinelnikov I, Arndt D,
Xia J, Liu P, Yallou F, Bjorndahl T, PerezPineiro R, Eisner R, Allen F, Neveu V, Greiner
R, Scalbert A (2013) HMDB 3.0 the human
metabolome database in 2013. Nucleic Acids
Res 41(Database issue):D801D807
15. Sweetman S (2004) Martindale: the complete
drug reference, 34th edn. Pharmaceutical
Press, New York, NY
444
David S. Wishart
16. Krogh A, Larsson B, von Heijne G,

Sonnhammer EL (2001) Predicting transmembrane protein topology with a hidden Markov
model: application to complete genomes.
J Mol Biol 305:567580
17. Yu NY, Wagner JR, Laird MR, Melli G, Rey S,
Lo R, Dao P, Sahinalp SC, Ester M, Foster LJ,
Brinkman FS (2010) PSORTb 3.0: improved
protein subcellular localization prediction with
refined localization subcategories and predictive
capabilities
for
all
prokaryotes.
18. Zhang R, Lin Y (2009) DEG 5.0, a database
of essential genes in both prokaryotes and
eukaryotes. Nucleic Acids Res 37(Database
issue):D455D458
19. Langille
MG,
Brinkman
FS
(2009)
IslandViewer: an integrated interface for computational identification and visualization of
genomic islands. Bioinformatics 25:664665
20. Wishart DS, Knox C, Guo AC, Shrivastava S,
Hassanali M, Stothard P, Chang Z, Woolsey J
(2006) DrugBank: a comprehensive resource for
in silico drug discovery and exploration. Nucleic
21. Wishart DS, Knox C, Guo AC, Cheng D,
Shrivastava S, Tzur D, Gautam B, Hassanali M
(2008) DrugBank: a knowledgebase for drugs,
drug actions and drug targets. Nucleic Acids
Res 36(Database issue):D901D906
22. Ihlenfeldt WD, Voigt JH, Bienfait B, Oellien F,
Nicklaus MC (2002) Enhanced CACTVS
browser of the Open NCI Database. J Chem
Inf Comput Sci 42:4657
23. Hamosh A, Scott AF, Amberger JS, Bocchini
CA, McKusick VA (2005) Online Mendelian
Inheritance in Man (OMIM), a knowledgebase
of human genes and genetic disorders. Nucleic
24. Wagner MJ (2013) Rare-variant genome-wide
association studies: a new frontier in genetic
analysis of complex traits. Pharmacogenomics
14:413424
25. Yang JO, Oh S, Ko G, Park SJ, Kim WY, Lee B,
Lee S (2011) VnD: a structure-centric database of disease-related SNPs and drugs. Nucleic
26. Stelzer G, Dalah I, Stein TI, Satanower Y,
Rosen N, Nativ N, Oz-Levi D, Olender T,
Belinky F, Bahir I, Krug H, Perco P, Mayer B,
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
Kolker E, Safran M, Lancet D (2011) In-silico

human genomics with GeneCards. Hum
Genomics 5:709717
Hu Y, Hines LM, Weng H, Zuo D, Rivera M,
Richardson A, LaBaer J (2003) Analysis of
genomic and proteomic data using advanced
literature mining. J Proteome Res 2:405412
Cheng D, Knox C, Young N, Stothard P,
Damaraju S, Wishart DS (2008) PolySearch: a
web-based text mining system for extracting
relationships between human diseases, genes,
mutations, drugs and metabolites. Nucleic
Acids Res 36(Web Server issue):W399W405
Griffith M, Griffith OL, Coffman AC, Weible
JV, McMichael JF, Spies NC et al (2013)
DGIdb: mining the druggable genome. Nat
Methods doi: 10.1038/nmeth.2689. [Epub
ahead of print]
Needleman SB, Wunsch CD (1970) A general
method applicable to the search for similarities
in the amino acid sequence of two proteins.
J Mol Biol 48:443453
Altschul SF, Madden TL, Schaffer AA, Zhang
J, Zhang Z, Miller W, Lipman DJ (1997)
Gapped BLAST and PSI-BLAST: a new generation of protein database search programs.
Nucleic Acids Res 25:33893402
Ullman JR (1976) An algorithm for sub-graph
isomorphism. J ACM 23:3142
Fiser A, Sali A (2003) Modeller: generation and
refinement of homology-based protein structure models. Methods Enzymol 374:461491
Montgomerie S, Cruz JA, Shrivastava S, Arndt
D, Berjanskii M, Wishart DS (2008)
PROTEUS2: a web server for comprehensive
protein structure prediction and structurebased annotation. Nucleic Acids Res 36(Web
Barrett T, Wilhite SE, Ledoux P, Evangelista C,
Kim IF, Tomashevsky M, Marshall KA, Phillippy
KH, Sherman PM, Holko M, Yefanov A, Lee H,
Zhang N, Robertson CL, Serova N, Davis S,
Soboleva A (2013) NCBI GEO: archive for
functional genomics data sets-update. Nucleic
Cheng F, Li W, Zhou Y, Shen J, Wu Z, Liu G,
Lee PW, Tang Y (2012) admetSAR: a comprehensive source and free tool for assessment of
chemical ADMET properties. J Chem Inf
Model 52:30993105
Chapter 20
Molecular Docking toFlexible Targets
JesperSrensen, zlemDemir, RobertV.Swift, VictoriaA.Feher,
andRommieE.Amaro
Abstract
It is widely accepted that protein receptors exist as an ensemble of conformations in solution. How best to
incorporate receptor flexibility into virtual screening protocols used for drug discovery remains a significant challenge. Here, stepwise methodologies are described to generate and select relevant protein conformations for virtual screening in the context of the relaxed complex scheme (RCS), to design small molecule
libraries for docking, and to perform statistical analyses on the virtual screening results. Methods include
equidistant spacing, RMSD-based clustering, and QR factorization protocols for ensemble generation and
ROC analysis for ensemble selection.
Key words Relaxed complex scheme, Ligand filtering, Protein flexibility, QR factorization, RMSD-based
clustering, ROC analysis
1 Introduction
It is widely accepted that proteins do not exist in solution as a single
rigid structure but rather as an ensemble of conformations [13].
The atomic fluctuations that give rise to this ensemble range from
small rotations of an individual amino acid methyl group to much
larger fluctuations concerted between groups of residues and the
protein backbone, loops, or domains. The necessity to consider
alternate conformations, including subtle structural changes in a
binding pocket, is highlighted by the difficulties reported for accurate ranking in cross-docking exercises [47]. Put another way, no
single structure can represent the binding modes for all the competent inhibitors of a drug target. As computational chemists are often
seeking new inhibitors that bind to receptor pockets, these fluctuations need to be accounted for in our computational methods [811].
Modeling these protein ensembles thus provide an opportunity and
has demonstrated success in discovering novel and/or selective
inhibitors that bind to subpockets or alternate conformations not
obvious in the snapshot of a given crystal structure [1215].
445
446
Jesper Srensen et al.
A variety of approaches have been used to incorporate protein

flexibility into virtual screening (VS) methodologies and have been
reviewed extensively elsewhere [1619]. Briefly, one of the first
methods (and now offered by nearly all docking software programs) is the ability to soften the van der Waals potential to allow
for some receptorligand overlap, the so-called soft docking [20].
While this method does not increase the computational resources
needed for virtual screening it has the less desirable characteristic
of creating softness globally to the binding pocket despite the
observation that proteins typically have regions with a range of
relative flexibility. Alternatively, docking programs have been
developed to include a side chain rotamer library search to introduce residue flexibility [21], a combined induced-fit docking minimization protocol [22, 23], docking followed by rescoring across
alternate conformations [24], stochastic Monte Carlo search of
side chain and backbone flexibility with a bound ligand [2527] or
cross-over mutations (genetic algorithm) of multiple side-chain
and backbone conformations during docking [28]. Ultimately, the
use of many representative explicit receptor conformations has the
benefit of providing additional experimental or physical information despite the increased time to dock to multiple versions of the
receptor structure.
Generation of the protein ensemble can be achieved through
the use of experimental data, namely, multiple co-crystal structures,
an NMR structure ensemble or through computational methods
that explore a proteins conformational energy landscape such as
elastic network normal mode analysis [29, 30], Monte Carlo simulations [25], and molecular dynamic (MD) simulations [9, 31].
In this chapter we focus on the Relaxed Complex Scheme
(RCS), which is a strategy that utilizes a physics-based molecular
dynamics methodology to generate an ensemble of receptor structures and then exploits that predictive, simulation-based structural
knowledge in the discovery and design of small molecule compounds [9, 32, 33]. This method has proven successful for the
target, Trypanosoma brucei RNA-editing ligase 1 (TbREL1 [34]),
investigated here, and other relevant drug targets [12, 14, 15, 35
40]. The advantage of this physics-based approach is that it allows
for larger scale and concerted conformational changes to be captured, and it offers the potential to understand the ligand induced
structural changes.
TbREL1 is part of the RNA editing complex for the parasite
Trypanosoma brucei, a causative agent in African Sleeping Sickness.
The role of TbREL1 is the catalytic ligation of two RNA molecules
driven by the hydrolysis of ATP [34]. TbREL1 is critical for the
survival of the parasite, which makes it a promising target.
447
2 Materials
Software used:
Classical and accelerated MD simulations were performed with
NAMD 2.9 [41, 42], using the Amber99SB force field [43]. The
docking programs used are Schrdinger Glide v. 6.0, 2013 with the
SP scoring function (Schrdinger, LLC, NewYork, NewYork, www.
schrodinger.com) [44, 45] and Autodock Vina version 1.1.2
(http://vina.scripps.edu) [46]. Statistics calculations were performed in MATLAB version R2011b 7.13.0.564 (MathWorks,
Natick, MA, www.mathworks.com/matlab). Figures of the protein
conformations were produced using VMD 1.9.1 [47].
Ensemble docking starting materials are
1. A crystal structure, an NMR ensemble, or a homology model
used for MD simulation.
2. A set of ligand files formatted properly for the docking program used.
3. The docking program.
In this case, the single 1.20 resolution crystal structure for
TbREL1 (PDB ID: 1XDN) [48] was used.
Multiple ligand files were compiled for docking; 121 and 40
known binders for TbREL1 from the Drug Discovery Unit (DDU)
Diversity compounds and Kinase set [49], respectively, the ATP
ligand extracted from the TbREL1 co-crystal structure [48] and
additional known binders found in previous virtual screening
efforts for TbREL1 [39, 40]. We have used these known actives to
generate a set of nonbinders (decoys) using DUD-E [50].
3 Methods
Here, we outline the steps in selecting a representative receptor
ensemble for docking from a dataset of molecular dynamics trajectories using the RCS, the preparation of the ligand files for docking, and the statistical analysis of the virtual screening results.
Virtual screening efforts incorporating receptor flexibility have
previously been reported for TbREL1 [39, 40, 51]. Discovery of
inhibitors towards RNA editing enzymes in trypanosomatid pathogens has also been reviewed recently [52].
3.1 Generating
aConformational
Ensemble
A set of receptor structure coordinates is a prerequisite to generating an ensemble and can be derived from X-ray crystallography or
NMR spectroscopy. If no receptor structure is available for the
target, a homology model based on a structure of a related protein
can be used, preferably with a high sequence identity [5356].
Inthe event that several crystal structures of the biomolecular
target are available, these should be incorporated, as they will often
448
represent different conformations of the protein. To generate a

diverse conformational ensemble we used two sets of initial structure coordinates, a set where ATP and the magnesium ion are
retained, and a set where they were deleted. Typically, using a crystal structure determined without a ligand is not advised unless
there are no liganded examples [57]. We have used only the simulations containing ATP in the VS, as the simulations without ATP
bound show partial occlusion of the binding pocket.
We performed both conventional and accelerated [58, 59]
MD simulations using NAMD v2.9 [41] with the AMBER force
fields [43] described in our recent study of the TbREL1RNA
complex [60]. Minor modifications to the force field employed a
new magnesium ion model [61] to more accurately capture the
dynamics around the ion (see Note 1) and a more recent water
model, TIP4P-ew [62]. Detailed accounts of the required preparation prior to MD simulations have previously been described elsewhere for this system [63, 64]. Moreover, protein preparation in
general has recently been reviewed [65]. Inclusion or exclusion of
various crystallographic waters, ligands, and binding site ions or
co-factors are among the details to be considered.
Simulation lengths will vary based on the biomolecular target
and the level of conformational change is a factor to take into
account. The same is true for whether to perform conventional or
accelerated MD (see Note 2) since the latter can be used to enhance
conformational sampling. Current typical simulation lengths vary
from tens to thousands of nanoseconds, where snapshots are
extracted at regularly spaced intervals (see Note 3).
3.2 Filtering
theConformational
Ensemble into
aMeaningful Subset
The MD simulations, depending on the simulation length, will

result in tens to hundreds of thousands of snapshots of the biomolecule sampled in different conformations; in our example we have
generated a total of 60,000 snapshots. It is not feasible to dock a
(large) library of ligands into each and every one of these conformations, and recent studies where this has been attempted have
shown that it is not necessarily an advantage to include a higher
number of conformations of the biomolecular target [31, 66]. The
goal then is to extract meaningful structures that will enrich the
predictive power of the virtual screen.
This reduction can be achieved using a number of different
algorithms, including QR factorization [39, 67], atomic or residue
based root-mean-square-deviation (RMSD) clustering [68, 69],
and active-site shape-based methods [7074]. We outline here the
steps for creating ensembles by: (1) equidistant frame samples, (2)
RMSD-based clustering, and (3) QR factorization.
The simplest approach is to extract a predetermined number of
structures with regular time intervals (equidistant spacing) between
them from the MD simulations (see Note 4). In our example, the
structures have been extracted using ptraj [75] in AmberTools 13
[7678] with a spacing of every 10ns. We executed ptraj with amber
449
parameter file (prmtop) and a script that reads a number of commands for ptraj. We specified for ptraj to read in (using the trajin
command) frames 110,000 (the number of frames in the simulated
trajectory), but only to read every 1,000th frame and output these,
resulting in ptraj outputting ten frames (using the trajout command), with an equidistant spacing of 10ns. The advantage of ptraj
is that we can specify it to align the protein structures to a reference
structure, i.e., the crystal structure, using the rms command. In the
input file below for ptraj, the file system.inpcrd contains the crystal
structure conformation, which is loaded in and used as a reference
structure for aligning. The output pdb files are used for docking.
(Script 1)
RMSD-based clustering is fairly common [68, 69], yet has a number of variables to be determined by the user, which should be
chosen based on the problem at hand. There are too many variations to outline here, instead we refer the reader to an excellent
paper reviewing the possibilities [69] and provide a simple example
of a commonly used method. When developing an ensemble for
virtual screening, in which the goal is to capture the most diverse
conformations of the active site, the RMSD calculation should be
performed with a small set of atoms or residues that line the active
site or are within a certain distance from the ligand (ifone is bound
in the protein). However, if one is interested in larger scale conformational changes as one may encounter with large loops near an
active or allosteric ligand site, then the protein backbone atoms
are most likely a better selection. Here we have chosen the former approach (see Note 5 for the residue selection). The RMSDbased clustering was performed in ptraj [75], although several
other programs are available for this task. As a first step, thetrajectory snapshots were aligned to the crystal structure conformation based on the same residue selection, but only taking the
backbone CA atoms into account. The remaining variables refer
to the different variations of RMSD-based clustering, which have
been described in great detail by Shao etal. [69]. In ptraj we
have used the cluster command, employing the average-linkage
450
algorithm, requesting ten clusters based on pairwise RMSD.

The average-linkage clustering refers to merging of structures into
clusters; initially each structure is assigned to its own cluster, the
distance between each cluster is then calculated and the clusters
with the shortest distance are merged iteratively until the target
number of clusters is reached; when several protein conformations
are part of a cluster it is then the average distance of each conformation in that cluster to other clusters that are used as the distance metric. We had ptraj print out the average and the most
representative structures in pdb format for each cluster. For VS we
have used therepresentative structure from each cluster. Note that
we are only reading in every 5th frame from the simulation; RMSDbased clustering is time consuming and the time scales with N2,
with N representing the number of frames used.
(Script 2)
For the classical MD simulations, we have excluded clusters with
populations lower than 5% of the trajectory, although in some
cases lowly populated states may also be viable for discovery. This
results in seven clusters. For the accelerated MD simulations, we
have chosen to keep all ten clusters, because we expect a much better conformational sampling, while not necessarily visiting the
same conformation as often as in the case of conventional MD.
An alternative ensemble clustering methodology can also be
performed using QR factorization which enables one to efficiently
reduce the number of MD snapshots to a minimal set without compromising the loss of diversity in the geometric characteristics of
the binding pocket [39]. QR factorization is a mathematical technique that performs repeated Householder transformations with
column pivoting to reorder the ensemble of structures such that
they are arranged with increasing linear dependence. The steps
required for preprocessing the MD trajectory files for QR factorization are provided in a tutorial at the NBCR Web site listed below.
451
http://nbcr.ucsd.edu/wiki/index.php/
SI2011_track3_CADD_QR_factorization_tutorial
The processed files can then be submitted to the publicly available server on the same Web site listed below.
http://nbcr-222.ucsd.edu/opal2/
CreateSubmissionForm.do?serviceURL=http://
localhost:8080/opal2/services%2Ftrajqr_1.0
Structure extraction techniques based on the shape and chemical properties of the active site are also emerging [7074, 79, 80],
but not used here (see Note 6). Furthermore, structural water
molecules in the active site should be considered (see Note 7).
The final protein conformations used for VS were extracted using
ptraj as detailed above in scripts 1 and 2 and output in the pdb format.
The pdb files were then converted to the pdbqt format for Autodock
Vina (see Note 8). The active site center was defined by X, Y, and Z
coordinates using a fixed square box to enclose the active site (see
Note 9). For Glide docking a receptor grid file was generated using
the XGlide script provided by Schrdinger (see Note 10). The PDB
files were used as input. In total 25 protein conformations were used
for docking: eight different setups of the crystal structure, varying
inclusion of structural water molecules, and 17 structures from ATPbound simulations seven of which were extracted by RMS clustering
of conventional MD trajectories and the remaining ten extracted from
accelerated MD trajectories.We first explored which crystal structure
setup was best able to discriminate binders from nonbinders and then
added this one setup to the 17 clusters from MD.Thus, 18 protein
conformations were included for ensemble statistics, resulting in the
evaluation of 262,143 different ensembles.
3.3 Ligand Library
Construction
andPreparation
forDocking
Assembly of a high-quality compound collection for virtual screening

should be developed with the target of interest in mind. In the
TbREL1 case, the DDU Kinase set and part of their Diversity set
were utilized [49]. This collection of compounds was developed specifically for screening against a diverse set of novel parasitic enzyme
targets and kinase homologues to human kinases [81], however, the
protocol used incorporated many best practice criteria that are general to any target-focused library development [8286]. A general
workflow, loosely based on the DDU methodology is shown in
Fig.1. This workflow starts by culling compounds from commercial
suppliers, followed by extensive filtering, clustering and visualization
steps. Filtering first removes redundant compounds and subsequent
steps can be added to remove unwanted compound functionalities,
such as those with reactive groups, compounds with low functional
complexity, compounds without lead-like chemical properties and
known aggregators. These protocols can be accomplished using
OpenEyes ToolKit as described in [81] or with alternate software
(see Note11). However, a method for the filtering steps is provided
by the OpenEyes Filter tool (Santa Fe, NM. www.eyesopen.com).
452
Fig. 1 A general workflow to generate a compound library for virtual screening.

*For compound sources see Note 11
This program provides a default filter file named lead containing

many commonly accepted best practices for pre-filtering compounds
used in a virtual screen or it can be modified by the user to use a custom set of rules specific to a dataset or target.
>OpenEye/bin/filter in inputfilename.sdf filter
lead prefix clean fail failed out outputfilename
Clustering of the compounds remaining after filtering can be
performed to: (1) further assist in visualizing groups of your selections, (2) understand the relative diversity of the scaffold types and
functionalization, and (3) easily select a set by retaining only cluster centroids or another selection criterion. In the DDU Dundee
Diversity set, the JarvisPatrick algorithm within the Daylight cluster package (Daylight, Aliso Viejo) was used to give clusters containing at least nine neighbors, while removing singletons and
other compounds within a cluster having a Tanimoto coefficient
>0.9 to the centroid [81].
Decoy molecules were generated with DUD-E [50]. DUD-E
decoys were pulled from the ZINC45 database [87] based on matching physicochemical properties to the known binders, such as molecular weight, estimated wateroctanol partition coefficient (miLogP),
rotatable bonds, number of hydrogen bond acceptors and donors,
and net charge. Moreover, the decoys were selected to be topographically dissimilar. DUD-E aims to return 50 unique decoys per input
known binder. The input for DUD-E are SMILES strings of the molecules, which we generated using Schrodingers Maestro software.
2D depictions of the compounds in either .sdf or .mae format
were subsequently converted into the 3D coordinates docking format
with Schrodingers LigPrep module and the following GUI settings:
453
Tautomers with predicted probability <25% were removed

manually.
The Schrodinger LigPreps output .mae file format was used
directly as input for Schrodingers Glide docking protocol, else
exported as .mol2 format and converted to .pdbqt format for
Autodock Vina (see Note 12).
3.4 Docking
Protocols
Docking protocols typically do not differ significantly whether

including receptor ensembles or a single receptor example, but they
do usually extend the time-to-completion of the VS in real-time
since the number of receptor structures to dock into increases.
Modifications that enhance the speed or parallelization of the protocol are therefore broadly useful. Time-to-completion may be an
important consideration when selecting a docking program for the
VS.However, it is also known that not all docking programs perform
well on all types of protein receptors [4, 88]. Therefore, selection of
the docking program may depend more on the results of customary
validation, namely, re-docking of the co-crystallized ligand to the
original crystal structure. As many co-crystallized compounds as possible should be re-docked with an RMSD value below 2, and a
favorable interaction energy during the validation process.
In previous screens we used Autodock4 for docking, which
successfully predicted compounds with inhibitory activity toward
TbREL1 [39, 40]. However, for this study in which a larger number of protein receptor conformations and a larger library of ligands
were employed, we have used more efficient docking programs,
namely, Autodock Vina and Glide from Schrdinger (see Note 13).
The re-docking was performed with Glide and the RSMD of the
predicted pose of ATP compared to the crystal structure was
0.30, and the binding energy was very favorable. Furthermore,
poses of the inhibitors found in our previous studies with Autodock4
were reproduced by Glide (data not shown).
3.5 Generating
Docking Statistics
When one has a set of known binders and known nonbinders for a
target, statistical methods can be used to assist in the selection of
454
ensemble conformations to include for the best virtual screening

performance. Although there are various VS performance metrics
available in the literature (see Note 14), the area under the curve
(AUC) of the Receiver Operating Characteristic (ROC) plot is one
of the most popular performance evaluation metrics, and is utilized
in this study. AUC values are also used to compare docking protocols as well as different ensembles of receptor conformations that
exhibit the best performance for a system of interest.
VS essentially reduces to a binary classification problem that
tries to discriminate compounds as either binders or nonbinders.
The predicted binding affinities or binding free energies generated
for each compound in a VS experiment are used to rank the compounds, and the compounds with higher ranks are more likely to
be experimentally assayed. In a VS experiment in which the binders
and nonbinders are known, the probability distribution function
(PDF) of predicted binding affinity of binders and nonbinders can
be constructed separately (Fig.2a). This allows us to choose a
threshold docking score that we would use to pick compounds for
testing. Integrating the PDF curve of binders from negative infinity to the chosen threshold value gives us the true positive rate
(TPR), (see Note 15). Integrating the PDF curve for nonbinders
from negative infinity to the chosen threshold value gives us the
false positive rate (FPR) (see Note 16). Continuously evaluating
TPR and FPR at each threshold value and plotting TPR versus
FPR constructs a ROC plot (Fig.2b). A docking protocol that
yields a higher TPR value at the same FPR value compared to
another protocol has a better chance of discovering larger number
of hits using the same threshold value (see Note 17).
The area under the curve (AUC) of the ROC plot is a significant measure of performance, and it represents the probability that
a randomly selected active will have a higher rank than a randomly
selected inactive [89, 90]. The TPR and FPR can take on values
Fig. 2 (a) Probability distribution function of binding free energies of binders and nonbinders in a virtual screening
experiment represented with the solid and dashed lines, respectively. (b) ROC plot corresponding to the virtual
screening in panel a. ROC plot for a random selection is also depicted with a dashed line for comparison
455
ranging from 0 to 1. Thus, the maximum AUC value possible is 1

(see Note 18) corresponding to perfect discriminatory power of a
VS protocol. The higher the AUC value, the better the VS performance. When reporting the AUC value for a VS experiment, a
good practice is to report its 95% confidence intervals as well. If
virtual screens are performed repeatedly on different databases, the
mean AUC value will be found in the range of the confidence
intervals 95% of the time. In practice, 95% confidence intervals
can be calculated using the formula below:

CI 95% = ( 2 SEcalculated erfinv ( 0.95) )
in which erfinv is the inverse error function and the standard error
calculated using the formula [91, 92]:

2
SEcalculated = (s NB
( AUC ) / N NB ) + (s B2 ( AUC ) / N B )
in which NNB and NB are number of nonbinders and binders,

respectively.
In the case of a random selection of compounds, the PDF
curve for binders and nonbinders will be identical resulting in
identical TPR and FPR values for any threshold value. In such a
case, the plot of TPR versus FPR, or the ROC plot, will be a
straight line bisecting the unit square (the dashed line in Fig.2b),
and the AUC will be equal to 0.5. A VS protocol should yield an
AUC value better than 0.5 to demonstrate discriminatory power
for binders and nonbinders. However, more effort is needed in
order to determine whether a VS protocol performs better than
random selection in a statistically significant way.
It is illustrative to consider an example. In the case when an
AUC value of 0.7 and a standard error of 0.1 is obtained for a particular VS protocol, a PDF curve can be constructed as in Fig.3a
(solid-line). There is still a slight probability that this particular
protocol performs randomly on average, but had a high performance for this instance. How large is this chance? To answer this
question, first, we make use of the null-hypothesis, our protocol
performs randomly and construct the null distribution, or the distribution of AUC values that corresponds to a protocol that performs randomly on average. Assuming the number of actives and
inactives is large enough to justify use of the central limit theorem,
the null distribution will be a Gaussian, centered on a mean AUC
of 0.5 as depicted in the dashed-line curve in Fig.3a according to
the central limit theorem (see Note 19). Now, the probability of
obtaining an AUC of 0.7, assuming the null hypothesis is true, is
just the area under the null distribution for AUC values greater
than or equal to 0.7. And this probability is called a p-value. The
smaller the p-value, the less likely it is that the null-hypothesis is true.
Generally, if the p-value is less than or equal to 5% or 1% [92], the
null hypothesis is rejected and the VS protocol is deemed to
perform statistically better than random. Once the null hypothesis
456
Fig. 3 (a) Probability distribution function of AUC for a real virtual screening protocol with a mean value of 0.7 and
a standard deviation of 0.1 (depicted with a solid-line curve), and for a random selection with a mean value of 0.5
and a standard deviation of 0.1 (depicted with a dashed-line curve). The shaded area corresponds to the p-value
to evaluate whether the real virtual screening protocol performs better than random. (b) Probability distribution
function of AUC for a set of two real virtual screening experiments with a mean value of 0.2 and a standard
deviation of 0.1 (depicted with a solid-line curve), and for a set of two identical virtual screening experiments with
a mean value of 0.0 and a standard deviation of 0.1 (depicted with a dashed-line curve). The shaded area
corresponds to the p-value to evaluate whether the two real virtual screening experiments perform identically
is rejected, it can be replaced with the alternative hypothesis, which

assumes that on average, the protocol does in fact perform better
than random. Generally, the alternative hypothesis is centered on a
mean identical to the observed performance and can now be used
to estimate confidence intervals.
In order to determine whether one of the two VS protocols
performs better in a statistically significant way, the null hypothesis
of the two VS methods perform identically must be evaluated
[91]. The PDF curve of AUC for this null hypothesis (the dashedline curve in Fig.3b) will be centered at 0.0, and will have the same
standard deviation as the alternative curve. Assuming the difference
in the mean AUC values for the two VS protocols is 0.2, the probability distribution of the alternative curve (the solid-line curve in
Fig.3b) will be centered at 0.2 with a standard deviation calculated
using the standard deviations of the two protocols (see Note 20).
In this case, the p-value will be equal to the area under the curve of
the null distribution for |AUC|
>
=0.2 (the shaded areas in
Fig.3b). If the p-value is too small and rejected, then statistically
one protocol is deemed to perform better than the other protocol.
In practice, one needs to evaluate all possible combinations of
receptor conformations to distinguish which ensemble of receptor conformations among the N conformations predicts the true
binders best. The following Matlab scripts can be utilized to
monitor performance of all possible ensembles of conformations.
The scripts require an input matrix total in which the first column has ligand identifiers, the second column has either 0 or 1
457
for nonbinders and binders, respectively, followed by columns

containing the docking scores for each receptor conformation
(also see Note21).
458
459
These analyses have been performed to determine the best

ensemble for TbREL1 dataset described above. The results show
only mild enrichment with the maximum AUC value reaching
0.58653. The result shows that three protein conformations are
460
Fig. 4 (a) The crystal structure of TbREL1 with ATP bound (1XDN.pdb), also highlighting the magnesium ion and
three water molecules bound deep in the protein that interact with ATP.Black markers highlight important
interactions, the E60-R111 salt-bridge, Mg2+-triphosphate tail of ATP, E86 and V88 backbone hydrogen bonds
to ATP, Y58-D10 hydrogen bond, D210-R288 salt-bridge, R288-Water-N7 hydrogen bond, and stacking of F209
and the adenosine moeity. K87 is highlighted, as it is the catalytic residue that gets adenylated when attacking
P in ATP. (b) a setup of the crystal structure with one specific water molecule at the deep end of the pocket
has shown to improve the VS enrichment, (c) a representative structure from conventional MD, and (d) a representative structure from accelerated MD
performing the best, when it comes to discriminating binders from

presumed nonbinders; the crystal structure, and a structural representative from both conventional and accelerated MD.These protein conformations are shown in Fig.4, with comparison to the
crystal structure conformation. Testing with different setups of the
crystal structure showed that inclusion of one specific water molecule shown in Fig.4b, improves the enrichment over other conformations. Further specifics about these receptor conformations and
their ability to discriminate the binders from the nonbinders will
461
appear in a separate publication. The mild enrichment demonstrated by this example could be a result of many factors, the most
likely is the challenging example we posed to these docking protocols. Here we have a set of known binders with low affinity
(10M<IC50<100 M) and have asked these programs to distinguish them from a set of DUD-E ligands of similar physiochemical
and topological properties, a task that continues to be an important area of research in computer-aided drug design.
4 Notes
1. The magnesium parameters can be downloaded from the
Bryce group AMBER parameter database (http://www.pharmacy.manchester.ac.uk/bryce/amber) where we have contributed the parameter files with permission from the parameter
developers [61].
2. We used dual-boost accelerated MD in NAMD [41, 42],
which applies a boost to the entire potential energy, and also a
boost to the dihedral potential. For each boosting term two
parameters are set: the energy threshold (E) and a tuning
parameter (), which determines the depth of the potential
energy well. To determine the boost energy a short (15ns)
classical MD simulation is performed, and the average of the
POTENTIAL and DIHED terms in the NAMD output are
calculated. The boost parameters are then determined according to the following formula. The factor 4 in the dihedral
terms is not a fixed factor; values of 3.56 have been reported
in the literature [58, 59, 94]. Since we have used the TIP4P-ew
water model [62], we have counted the extra particle on the
water model as an extra atom.
E ( dihed ) = DIHEDNAMD + 4 * # residues
a ( dihed ) = 1 + 4 * # residues
5
E ( total ) = POTENTIAL NAMD + 0.16 * # atoms
a ( total ) = 0.16 + # atoms
3. Here we have extracted conformations every 10ps. This is
specified in NAMD using the DCDfreq variable.
4. Equidistant means that there is an even spacing in time
between the snapshots extracted from the simulations. The
ideal is to select a number that will allow diverse conformations of the binding site; however, such a number is highly
system specific. Another aspect to take into account is not to
set this number too low, as this will lead to too many protein
conformations, which will require more computational
462
resources for the VS, and adding too many protein conformations is not recommended [31, 66].
5. The RMS clustering included the following residues: Tyr58,
Glu60, Glu86, Lys87, Asn92, Arg111, Asp159, Phe209,
Asp210, Glu283, Val286, Arg288, Arg292, Lys307, and
Arg309. These have all previously been highlighted in structural analyses as belonging to the active site [48, 64].
6. Osguthorpe etal. have created an ensemble based on shape
diversity of the active site that had an enriching effect on their
VS [70, 71]. The MDpocket utility found in Fpocketalgorithm
will calculate the volume and specific chemical properties of a
specified pocket from each snapshot of an MD simulation [72,
73]. Subsequent clustering can then be performed on these
data to extract a representative set of structures describing variability in the active site. Alternatively, the FTMap algorithm
floods the protein surface with a set of small organic molecules
and calculates an interaction energy, thereby predicting druggable hotspots in the protein [79, 80]. This algorithm has
recently been extended for the analysis of MD trajectories
[95]. This method can also be useful in identifying, visualizing, and characterizing new subpockets of the target site.
7. There are three water molecules in the cavity wat1, wat2, and
wat3, following the nomenclature of our previously published
work [64]. Thus, we have made the following combinations:
no waters, wat1, wat2, wat3, wat1
+
wat2, wat1
+
wat3,
wat2+wat3, wat1+wat2+wat3. In total, there are seven different receptor configurations. In recent years programs for the
analysis of structurally resolved water molecules have been
developed [16]. Schrdinger has developed the WaterMap
framework to explore and exploit water molecules bound inside
the ligand binding site in drug discovery [96, 97]. Molegro
Virtual Docker [98] has developed a docking algorithm with
attached water molecules that are then retained or displaced in
the docked pose based on energy contributions [99].
8. This conversion was done with the utility prepare_receptor4.py
in Autodock, which takes a pdb file as input and outputs a
pdbqt file.
9. The center was specified as x=41.1100, y=34.9382, and
z=35.8160, based on the ATP binding site. The box size was
defined as a square with the box length set to 25. As all the
structures used were previously aligned to the crystal structure
the square box should encapsulate the active site in all the
protein conformations.
10. The script is available on the Schrdinger Web site script center (http://www.schrodinger.com/scriptcenter/) and is
already preinstalled in Maestro. The script can be used to easily
463
calculate a receptor grid for each protein conformation in an

automated fashion. The script and its application are described
in the Schrdinger knowledge base article ID 560 (http://
www.schrodinger.com/kb/560). As mentioned the PDB files
were all aligned to the crystal structure before docking, so a
grid center and box dimensions could be specified on the command line. This is very helpful when you have a large number
of protein receptor conformations and want to avoid going
through the graphical user interface wizard for each and every
conformation. The same grid center that was specified for
Autodock Vina, was specified for XGlide. The inner and outer
box dimensions were set to 10 and 20, respectively.
11. Depending upon your goals for purchasing compounds, either
the publicly available ZINC database (www.zincdocking.org)
or other large compendiums can be used to explore an extensive chemical space or if you prefer to only screen compounds
that are currently available or within a specific pricing range,
contacting individual commercial companies may be preferred
to reduce the size needing to be screened. These databases are
available for download as sdf or SMILES format. Among filtering software alternatives to OpenEyes Filter are: Schrodingers
Canvas (Schrdinger, LLC, NewYork, NY, www.schrodinger.
com), CCG MOE (Chemical Computing Group Inc., 1010
Sherbooke St. West, Suite #910, Montreal, QC, Canada, H3A
2R7, 2012; www.chemcomp.com), Molsoft (Molsoft, LLC,
San Diego, CA; www.molsoft.com), or Accelrys Pipeline Pilot
(Accelrys, San Diego, CA; http://accelrys.com).
12. The ligands can be converted using the prepare_ligand4.py
utility in Autodock. This script takes either .pdb or .mol2 files
as input. The current charges were predicted using LigPrep
from Schrdinger, but Autodock Vina uses an internal charge
scheme, so the input charges were ignored.
13. We have used the SP scoring function.
14. There are many methods available for performance analysis
like enrichment factor (EF), Boltzmann-Enhanced
Discrimination of ROC (BEDROC), and robust initial
enhancement (RIE) methods [45, 100, 101].
15. True positives are the compounds that are predicted to be
binders, and are actually binders.
16. False positives are the compounds that are predicted to be
binders, but are actually not binders.
17. To compare VS enrichments of different protocols, it is recommended to compare TPR values at FPR values of 0.5%,
1%, 2%, or 5% [90].
18. The area of a unit square is 1.
464
19. The central limit theorem (CLT) states that if enough independent measurements of a property are performed on the
same system, the average property will be distributed like a
Gaussian. The center of the CLT curve will be the true mean,
and its width will change with the variance value.
20. If methods A and B are combined, the standard deviation for
this new method is s A + B = (s A2 + s B2 ) .
21. In this example, the best docking score that each ligand gets
among the specified receptor conformations is picked.
Alternatively, one could pick the average of the docking scores
of each ligand for the specified receptor conformations. Or one
could choose to compute a weighted-average of docking scores
using the population percentages of each cluster if the receptor
conformations are extracted by RMSD-based clustering.
Acknowledgements
This work was funded in part by through the NIH Directors New
Innovator Award Program DP2-OD007237 and the National
Science Foundations XSEDE Supercomputer resources grant
LRAC CHE060073N to R.E.A.Support from the National
Biomedical Computation Resource (P41 GM103426), the Center
for Theoretical Biophysics, and UCSD Drug Discovery Institute is
gratefully acknowledged. J.S. thanks the Alfred Benzon Foundation
for a generous postdoctoral fellowship.
References
1. Frauenfelder H, Sligar SG, Wolynes PG
(1991) The energy landscapes and motions of
proteins. Science 254(5038):15981603
2. Boehr DD, Nussinov R, Wright PE (2009)
The role of dynamic conformational ensembles
in biomolecular recognition. Nat Chem Biol
5(11):789796. doi:10.1038/nchembio.232
3. Forman-Kay JD (1999) The dynamics in
the thermodynamics of binding. Nat Struct
Biol 6(12):10861087. doi:10.1038/70008
4. Cross JB, Thompson DC, Rai BK, Baber JC,
Fan KY, Hu Y, Humblet C (2009)
Comparison of several molecular docking
programs: pose prediction and virtual screening accuracy. J Chem Inf Model 49(6):1455
1474. doi:10.1021/ci900056c
5. Cheng T, Li X, Li Y, Liu Z, Wang R (2009)
Comparative assessment of scoring functions
on a diverse test set. J Chem Inf Model
49(4):10791093. doi:10.1021/ci9000053
6. Armen RS, Chen J, Brooks CL 3rd (2009)
An evaluation of explicit receptor flexibility in
molecular docking using molecular dynamics

and torsion angle molecular dynamics.
JChem Theory Comput 5(10):29092923.
doi:10.1021/ct900262t
7. Sutherland JJ, Nandigam RK, Erickson JA,
Vieth M (2007) Lessons in molecular recognition. 2. Assessing and improving cross-
docking accuracy. J Chem Inf Model
47(6):22932302. doi:10.1021/ci700253h
8. Verkhivker GM, Bouzida D, Gehlhaar DK,
Rejto PA, Freer ST, Rose PW (2002) Complexity
and simplicity of ligand-macromolecule interactions: the energy landscape perspective. Curr
Opin Struct Biol 12(2):197203
9.
Lin J-H, Perryman AL, Schames JR,
McCammon JA (2002) Computational drug
design accommodating receptor flexibility: the
relaxed complex scheme. J Am Chem Soc
124(20):56325633. doi:10.1021/ja0260162
10. Teague SJ (2003) Implications of protein
flexibility for drug discovery. Nat Rev Drug
Discov 2(7):527541. doi:10.1038/nrd1129

11. Cozzini P, Kellogg GE, Spyrakis F, Abraham
DJ, Costantino G, Emerson A, Fanelli F,
Gohlke H, Kuhn LA, Morris GM, Orozco M,
Pertinhez TA, Rizzi M, Sotriffer CA (2008)
Target flexibility: an emerging consideration in
drug discovery and design. J Med Chem
51(20):62376255. doi:10.1021/jm800562d
12. Schames JR, Henchman RH, Siegel JS,
Sotriffer CA, Ni H, McCammon JA (2004)
Discovery of a novel binding trench in HIV
integrase. J Med Chem 47(8):18791881.
doi:10.1021/jm0341913
13. Gorfe AA, Caflisch A (2005) Functional plasticity in the substrate binding site of beta-
secretase. Structure 13(10):14871498.
doi:10.1016/j.str.2005.06.015
14. Cheng LS, Amaro RE, Xu D, Li WW, Arzberger
PW, McCammon JA (2008) Ensemble-based
virtual screening reveals potential novel antiviral compounds for avian influenza neuraminidase. J Med Chem 51(13):38783894.
doi:10.1021/jm8001197
15. Baron R, McCammon JA (2007) Dynamics,
hydration, and motional averaging of a loop-
gated artificial protein cavity: the W191G
mutant of cytochrome c peroxidase in water
as revealed by molecular dynamics simulations. Biochemistry 46(37):1062910642.
doi:10.1021/bi700866x
16. Yuriev E, Agostino M, Ramsland PA (2011)
Challenges and advances in computational
docking: 2009in review. J Mol Recognit
24(2):149164. doi:10.1002/jmr.1077
17. Lavecchia A, Di Giovanni C (2013) Virtual
screening strategies in drug discovery: a critical
review. Curr Med Chem 20(23):28392860
18. B-Rao C, Subramanian J, Sharma SD (2009)
Managing protein flexibility in docking and its
applications. Drug Discov Today 14(78):394
400. doi:10.1016/j.drudis.2009.01.003
19. Sinko W, Lindert S, McCammon JA (2013)
Accounting for receptor flexibility and
enhanced sampling methods in computer-
aided drug design. Chem Biol Drug Des
81(1):4149. doi:10.1111/cbdd.12051
20. Jiang F, Kim SH (1991) Soft docking:
matching of molecular surface cubes. J Mol
Biol 219(1):79102
21. Cerqueira NM, Bras NF, Fernandes PA,
Ramos MJ (2009) MADAMM: a multistaged
docking with an automated molecular modeling protocol. Proteins 74(1):192206.
doi:10.1002/prot.22146
22. Sherman W, Day T, Jacobson MP, Friesner
RA, Farid R (2006) Novel procedure for
modeling ligand/receptor induced fit effects.
J Med Chem 49(2):534553. doi:10.1021/
jm050540c
465
23. Sherman W, Beard HS, Farid R (2006) Use of

an induced fit receptor structure in virtual
screening. Chem Biol Drug Des 67(1):8384.
doi:10.1111/j.1747-0285.2005.00327.x
24. Jain AN (2009) Effects of protein conformation in docking: improved pose prediction
through protein pocket adaptation. J Comput
Aided Mol Des 23(6):355374. doi:10.1007/
s10822-009-9266-3
25. Davis IW, Baker D (2009) RosettaLigand
docking with full ligand and receptor flexibility.
J
Mol
Biol
385(2):381392.
doi:10.1016/j.jmb.2008.11.010
26. Lemmon G, Meiler J (2012) Rosetta ligand
docking with flexible XML protocols.
Methods
Mol
Biol
819:143155.
doi:10.1007/978-1-61779-465-0_10
27. Abagyan R, Totrov M, Kuznetsov D (1994)
ICM a new method for protein modeling
and design: applications to docking and structure prediction from the distorted native conformation. J Comput Chem 15(5):488506.
doi:10.1002/jcc.540150503
28. Corbeil CR, Moitessier N (2009) Docking
ligands into flexible and solvated macromolecules. 3. Impact of input ligand conformation, protein flexibility, and water molecules
on the accuracy of docking programs. J Chem
Inf Model 49(4):9971009. doi:10.1021/
ci8004176
29. Cavasotto CN, Kovacs JA, Abagyan RA
(2005) Representing receptor flexibility in
ligand docking through relevant normal
modes. J Am Chem Soc 127(26):96329640.
doi:10.1021/ja042260c
30. Cavasotto CN (2012) Normal mode-based
approaches in receptor ensemble docking.
Methods
Mol
Biol
819:157168.
doi:10.1007/978-1-61779-465-0_11
31. Nichols SE, Baron R, Ivetac A, McCammon
JA (2011) Predictive power of molecular
dynamics receptor structures in virtual screening. J Chem Inf Model 51(6):14391446.
doi:10.1021/ci200117n
32. Amaro RE, Baron R, McCammon JA (2008)
An improved relaxed complex scheme for
receptor flexibility in computer-aided drug
design. J Comput Aided Mol Des 22(9):693
705. doi:10.1007/s10822-007-9159-2
33.
Lin J-H, Perryman AL, Schames JR,
McCammon JA (2003) The relaxed complexmethod: accommodating receptor flexibility for drug design with an improved
scoring scheme. Biopolymers 68(1):4762.
doi:10.1002/bip.10218
34. Schnaufer A, Ernst NL, Palazzo SS, O'Rear J,
Salavati R, Stuart K (2003) Separate insertion
and
deletion
subcomplexes
of
the
466
Trypanosoma brucei RNA editing complex.

Mol Cell 12(2):307319
35. Landon MR, Amaro RE, Baron R, Ngan CH,
Ozonoff D, Andrew McCammon J, Vajda S
(2008) Novel druggable hot spots in avian
influenza neuraminidase H5N1 revealed by
computational solvent mapping of a reduced
and representative receptor ensemble. Chem
Biol Drug Des 71(2):106116. doi:10.1111/
j.1747-0285.2007.00614.x
36. Babakhani A, Talley TT, Taylor P, McCammon
JA (2009) A virtual screening study of the acetylcholine binding protein using a relaxed-complex
approach. Comput Biol Chem 33(2):160170.
doi:10.1016/j.compbiolchem.2008.12.002
37. Durrant JD, de Oliveira CAF, McCammon
JA (2010) Including receptor flexibility and
induced fit effects into the design of MMP-2
inhibitors. J Mol Recognit 23(2):173182.
doi:10.1002/jmr.989
38. Demir , Baronio R, Salehi F, Wassman CD,
Hall L, Hatfield GW, Chamberlin R, Kaiser P,
Lathrop RH, Amaro RE (2011) Ensemble-
based computational approach discriminates
functional activity of p53 cancer and rescue
mutants. PLoS Comput Biol 7(10):e1002238.
39. Amaro RE, Schnaufer A, Interthal H, Hol W,
Stuart KD, McCammon JA (2008) Discovery
of drug-like inhibitors of an essential RNA-
editing ligase in Trypanosoma brucei. Proc
Natl Acad Sci U S A 105(45):1727817283.
doi:10.1073/pnas.0805820105
40. Durrant JD, Hall L, Swift RV, Landon M,
Schnaufer A, Amaro RE (2010) Novel
naphthalene-
based inhibitors of Trypanosoma
brucei RNA editing ligase 1. PLoS Negl Trop Dis
4(8):e803. doi:10.1371/journal.pntd.0000803
41. Phillips JC, Braun R, Wang W, Gumbart J,
Tajkhorshid E, Villa E, Chipot C, Skeel RD,
Kal L, Schulten K (2005) Scalable molecular
dynamics with NAMD.J Comput Chem 26:
17811802

42.
Wang Y, Harrison CB, Schulten K,
McCammon JA (2011) Implementation of
accelerated molecular dynamics in NAMD.
Comput Sci Discov 4(1):015002
of multiple Amber force fields and development of improved protein backbone parameters. Proteins 65(3):712725. doi:10.1002/
prot.21123
44. Halgren TA, Murphy RB, Friesner RA, Beard
HS, Frye LL, Pollard WT, Banks JL (2004)
Glide: a new approach for rapid, accurate
docking and scoring. 2. Enrichment factors
indatabase screening. J Med Chem 47(7):
17501759. doi:10.1021/jm030644s
45. Friesner RA, Banks JL, Murphy RB, Halgren

TA, Klicic JJ, Mainz DT, Repasky MP, Knoll
EH, Shelley M, Perry JK, Shaw DE, Francis P,
Shenkin PS (2004) Glide: a new approach
forrapid, accurate docking and scoring. 1.
Method and assessment of docking accuracy. J
Med Chem 47(7):17391749. doi:10.1021/
jm0306430
46. Trott O, Olson AJ (2010) AutoDock Vina:
improving the speed and accuracy of docking
with a new scoring function, efficient optimization, and multithreading. J Comput Chem
31(2):455461. doi:10.1002/jcc.21334
VMD: visual molecular dynamics. J Mol
Graph 14(1):3338, 2738
48. Deng J, Schnaufer A, Salavati R, Stuart KD,
Hol WG (2004) High resolution crystal
structure of a key editosome enzyme from
Trypanosoma brucei: RNA editing ligase 1. J
Mol Biol 343(3):601613. doi:10.1016/j.
jmb.2004.08.041
49. Drug Discovery Unit UoD DDU Library
Collections (2013) http://www.drugdiscovery.dundee.ac.uk/libraries.html
50. Mysinger MM, Carchia M, Irwin JJ, Shoichet
BK (2012) Directory of useful decoys,
enhanced (DUD-E): better ligands and decoys
for better benchmarking. J Med Chem
55(14):65826594. doi:10.1021/jm300687e
51. Moshiri H, Acoca S, Kala S, Najafabadi HS,
Hogues H, Purisima E, Salavati R (2011)
Naphthalene-based RNA editing inhibitor
blocks RNA editing activities and editosome
assembly in Trypanosoma brucei. J Biol
Chem 286(16):1417814189. doi:10.1074/
jbc.M110.199646
52. Salavati R, Moshiri H, Kala S, Shateri
Najafabadi H (2012) Inhibitors of RNA editing as potential chemotherapeutics against
trypanosomatid pathogens. Int J Parasitol
Drugs Drug Resist 2:3646. doi:10.1016/j.
ijpddr.2011.10.003
53. Srensen J, Palmer DS, Qvist KB, Schitt B
(2011) Initial stage of cheese production: a
molecular modeling study of bovine and
camel chymosin complexed with peptides
from the chymosin-sensitive region of kappa-
casein. J Agric Food Chem 59(10):5636
5647. doi:10.1021/jf104898w
54. Feher VA, Lawson JD (2009) Approaches to
kinase homology modeling: successes and
considerations for the structural kinome. In:
Rongshi L, Stafford JA (eds) Kinase inhibitor
drugs. Wiley, Hoboken, NJ, pp 433460.
doi: 10.1002/9780470524961.ch17
55. Marti-Renom MA, Stuart AC, Fiser A, Sanchez
R, Melo F, Sali A (2000) Comparative protein
structure modeling of genes and genomes.

Annu Rev Biophys Biomol Struct 29:291325.
doi:10.1146/annurev.biophys.29.1.291
56. Cavasotto CN, Phatak SS (2009) Homology
modeling in drug discovery: current
trendsand applications. Drug Discov
Today14(1314):676683. doi:10.1016/j.
drudis.2009.04.006
57. Damm-Ganamet KL, Smith RD, Dunbar JB
Jr, Stuckey JA, Carlson HA (2013) CSAR
benchmark exercise 20112012: evaluation
of results from docking and relative ranking
of blinded congeneric series. J Chem Inf
Model. doi:10.1021/ci400025f
58. Hamelberg D, Mongan J, McCammon JA
(2004) Accelerated molecular dynamics: a
promising and efficient simulation method
forbiomolecules. J Chem Phys 120(24):
1191911929. doi:10.1063/1.1755656
59. Pierce LCT, Salomon-Ferrer R, Augusto F,
de Oliveira C, McCammon JA, Walker RC
(2012) Routine access to millisecond time
scale events with accelerated molecular
dynamics. J Chem Theory Comput
8(9):29973002. doi:10.1021/ct300284c
60. Swift RV, Durrant J, Amaro RE, McCammon
JA (2009) Toward understanding the conformational dynamics of RNA ligation.
Biochemistry 48(4):709719. doi:10.1021/
bi8018114
61.
Allnr O, Nilsson L, Villa A (2012)
Magnesium ionwater coordination and
exchange in biomolecular simulations. J
Chem Theory Comput 8(4):14931502.
doi:10.1021/ct3000734
62. Horn HW, Swope WC, Pitera JW, Madura
JD, Dick TJ, Hura GL, Head-Gordon T
(2004) Development of an improved four-site
water model for biomolecular simulations:
TIP4P-Ew. J Chem Phys 120(20):9665
9678. doi:10.1063/1.1683075
63. Demir O, Amaro RE (2013) Designing novel
inhibitors of Trypanosoma brucei. In:
Kortagere S (ed) Methods in molecular biology: in silico models for drug discovery, vol
993. Humana Press, Totowa, NJ, pp231
243, doi: 10.1007/978-1-62703-342-8_15
64. Amaro RE, Swift RV, McCammon JA (2007)
Functional and structural insights revealed by
molecular dynamics simulations of an essential RNA editing ligase in Trypanosoma brucei. PLoS Negl Trop Dis 1(2):e68.
doi:10.1371/journal.pntd.0000068
65. Shang Y, Simmerling C (2012) Molecular
dynamics applied in drug discovery: the case of
HIV-1 protease. Methods Mol Biol819:527
549. doi:10.1007/978-1-61779-465-0_31
66. Nichols S, Baron R, McCammon JA (2012)
On the use of molecular dynamics receptor
conformations for virtual screening. In: Baron
467
R (ed) Computational drug discovery and

design, vol 819, Methods in molecular biology. Springer, NewYork, pp93103
67. ODonoghue P, Luthey-Schulten Z (2003)
On the evolution of structure in aminoacyl-
tRNA synthetases. Microbiol Mol Biol Rev
67(4):550573.
doi:10.1128/MMBR.67.
4.550-573.2003
68. Baron R, McCammon JA (2008) (Thermo)
dynamic role of receptor flexibility, entropy,
and motional correlation in proteinligand
binding. Chem Phys Chem 9(7):983988.
doi:10.1002/cphc.200700857
69. Shao J, Tanner SW, Thompson N, Cheatham
TE (2007) Clustering molecular dynamics
trajectories: 1. Characterizing the performance of different clustering algorithms. J
Chem Theory Comput 3(6):23122334.
doi:10.1021/ct700119m
70. Osguthorpe DJ, Sherman W, Hagler AT
(2012) Generation of receptor structural
ensembles for virtual screening using binding
site shape analysis and clustering. Chem Biol
Drug Des 80(2):182193. doi:10.1111/
j.1747-0285.2012.01396.x
71. Osguthorpe DJ, Sherman W, Hagler AT
(2012) Exploring protein flexibility: incorporating structural ensembles from crystal structures and simulation into virtual screening
protocols. J Phys Chem B 116(23):6952
6959. doi:10.1021/jp3003992
72. Schmidtke P, Barril X (2010) Understanding
and predicting druggability. A high-
throughput method for detection of drug
binding sites. J Med Chem 53(15):5858
5867. doi:10.1021/jm100574m
73. Le Guilloux V, Schmidtke P, Tuffery P (2009)
Fpocket: an open source platform for ligand
pocket detection. BMC Bioinformatics
10:168. doi:10.1186/1471-2105-10-168
74. Durrant JD, de Oliveira CAF, McCammon
JA (2011) POVME: an algorithm for measuring binding-pocket volumes. J Mol Graph
Model 29(5):773776. doi:10.1016/j.jmgm.
2010.10.007
75. Roe DR, Cheatham TE (2013) PTRAJ and
CPPTRAJ: software for processing and analysis of molecular dynamics trajectory data.
JChem Theory Comput 9(7):30843095.
doi:10.1021/ct400341p
76. Salomon-Ferrer R, Case DA, Walker RC
(2013) An overview of the Amber biomolecular simulation package. Wiley Interdiscip
Rev Comput Mol Sci 3(2):198210.
doi:10.1002/wcms.1121
77. Case DA, Darden TA, Cheatham III TE,
Simmerling CL, Wang J, Duke RE, Luo R,
Walker RC, Zhang W, Merz KM, Roberts B,
Hayik S, Roitberg A, Seabra G, Swails J, Goetz
468
AW, Kolossvry I, Wong KF, Paesani F, Vanicek

J, Wolf RM, Liu J, Wu X, Brozell SR,
Steinbrecher T, Gohlke H, Cai Q, Ye X, Wang
J, Hsieh M-J, Cui G, Roe DR, Mathews DH,
Seetin MG, Salomon-Ferrer R, Sagui C, Babin
V, Luchko T, Gusarov S, Kovalenko A, Kollman
PA (2012) AMBER. 12 edn. University of
California, San Francisco, CA, USA
78. Case DA, Cheatham TE, Darden T, Gohlke H,
Luo R, Merz KM, Onufriev A, Simmerling C,
Wang B, Woods RJ (2005) The Amber biomolecular simulation programs. J Comput Chem
26(16):16681688. doi:10.1002/jcc.20290
79. Kozakov D, Hall DR, Chuang G-Y, Cencic R,
Brenke R, Grove LE, Beglov D, Pelletier J,
Whitty A, Vajda S (2011) Structural conservation of druggable hot spots in proteinprotein
interfaces. Proc Natl Acad Sci 108(33):13528
13533. doi:10.1073/pnas.1101835108
80. Brenke R, Kozakov D, Chuang G-Y, Beglov
D, Hall D, Landon MR, Mattos C, Vajda S
(2009) Fragment-based identification of
druggable hot spots of proteins using
Fourier domain correlation techniques.
Bioinformatics 25(5):621627. doi:10.1093/
bioinformatics/btp036
81. Brenk R, Schipani A, James D, Krasowski A,
Gilbert IH, Frearson J, Wyatt PG (2008)
Lessons learnt from assembling screening
libraries for drug discovery for neglected diseases. Chem Med Chem 3(3):435444.
doi:10.1002/cmdc.200700139
82. Rishton GM (2003) Nonleadlikeness and
leadlikeness in biochemical screening. Drug
Discov Today 8(2):8696
83. Seidler J, McGovern SL, Doman TN, Shoichet
BK (2003) Identification and prediction of
promiscuous aggregating inhibitors among
known drugs. J Med Chem 46(21):4477
4486. doi:10.1021/jm030191r
84. McGovern SL, Caselli E, Grigorieff N,
Shoichet BK (2002) A common mechanism
underlying promiscuous inhibitors from virtual and high-throughput screening. J Med
Chem 45(8):17121722
85. Baell JB, Holloway GA (2010) New substructure filters for removal of pan assay interference compounds (PAINS) from screening
libraries and for their exclusion in bioassays. J
Med Chem 53(7):27192740. doi:10.1021/
jm901137j
86. Walters WP, Stahl MT, Murcko MA (1998)
Virtual screening an overview. Drug Discov
Today 3(4):160178
87. Irwin JJ, Sterling T, Mysinger MM, Bolstad ES,
Coleman RG (2012) ZINC: a free tool to discover chemistry for biology. J Chem Inf Model
52(7):17571768. doi:10.1021/ci3001277
88. Warren GL, Andrews CW, Capelli AM, Clarke

B, LaLonde J, Lambert MH, Lindvall M,
Nevins N, Semus SF, Senger S, Tedesco G,
Wall ID, Woolven JM, Peishoff CE, Head MS
(2006) A critical assessment of docking programs and scoring functions. J Med Chem
49(20):59125931. doi:10.1021/jm050362n
89. Nicholls A (2011) What do we know?: simple
statistical techniques that help. In: Bajorath J
(ed) Chemoinformatics and computational
chemical biology, vol 672, Methods in molecular biology. Humana Press, Totowa, NJ, pp531
581. doi:10.1007/978-1-60761-839-3_22
90. Jain AN (2008) Bias, reporting, and sharing:
computational evaluations of docking methods. J Comput Aided Mol Des 22(34):201
212. doi:10.1007/s10822-007-9151-x
91. Nichols SE, Swift RV, Amaro RE (2012)
Rational prediction with molecular dynamics
for hit identification. Curr Top Med Chem
12(18):20022012. doi:10.2174/15680261
2804910313
92. du Prel JB, Hommel G, Rohrig B, Blettner M
(2009) Confidence interval or p-value?: part 4
of a series on evaluation of scientific publications. Dtsch Arztebl Int 106(19):335339. doi:
D NLM: PMC2689604 OTO NOTNLM
93. Craig IR, Essex JW, Spiegel K (2010)
Ensemble docking into multiple crystallographically derived protein structures: an
evaluation based on the statistical analysis of
enrichments. J Chem Inf Model 50(4):511
524. doi:10.1021/ci900407c

94.
Bucher D, Grant BJ, Markwick PR,
McCammon JA (2011) Accessing a hidden
conformation of the maltose binding protein
using accelerated molecular dynamics. PLoS
Comput Biol 7(4):e1002034. doi:10.1371/
95. Votapka L, Amaro RE (2013) Multistructural
hot spot characterization with FTProd.
Bioinformatics 29(3):393394. doi:10.1093/
bioinformatics/bts689
96. Abel R, Young T, Farid R, Berne BJ, Friesner
RA (2008) Role of the active-site solvent in
the thermodynamics of factor Xa ligand binding. J Am Chem Soc 130(9):28172831.
doi:10.1021/ja0771033
97. Beuming T, Farid R, Sherman W (2009)
High-energy water sites determine peptide
binding affinity and specificity of PDZ
domains. Protein Sci 18(8):16091619.
doi:10.1002/pro.177

98.
Thomsen R, Christensen MH (2006)
MolDock: a new technique for high-accuracy
molecular docking. J Med Chem 49(11):3315
3321. doi:10.1021/jm051197e

99. Lie MA, Thomsen R, Pedersen CNS, Schitt
B, Christensen MH (2011) Molecular docking with ligand attached water molecules. J
Chem
Inf
Model
51(4):909917.
doi:10.1021/ci100510m
100. Truchon J-F, Bayly CI (2007) Evaluating virtual screening methods: good and bad metrics for the early recognition problem. J
469
Chem
Inf
Model
47(2):488508.
doi:10.1021/ci600426e
101. Sheridan RP, Singh SB, Fluder EM, Kearsley
SK (2001) Protocols for bridging the
peptide to nonpeptide gap in topological
similarity searches. J Chem Inf Comput
Sci
41(5):13951406.
doi:10.1021/
ci0100144
INDEX
A
Accelerated molecular dynamics ...............................253285
Alchemical transitions ...................................... 179, 184, 187
-chymotrypsin ........................................................197200
AMBER ................................... 9, 62, 65, 261, 265267, 270,
271, 273, 277, 278, 355, 448, 461
Ambiguous interaction restraints (AIRs) ............405406, 410
AMOEBA........................................................ 51, 52, 56, 61
Antechamber .................................................... 187, 270272
Area per lipid ................................ 85, 88, 93, 95, 98, 99, 102,
111, 138, 141, 143
Aromatic order parameters ....................... 110, 116, 117, 120
Associative modeling ........................................................315
ATP-magnesium complex ........................................200201
Automated structure prediction ................................326329
Conformational ensemble .................................. 61, 447451

Conformational transition ........................................175, 219
Contact prediction ............................................................333
Crooks equality.................................................................181
Crooks fluctuation theorem (CFT) ......................... 174, 181,
182, 192, 193
CS-ROSETTA .........................353, 357, 358, 361365, 374
CVs. See Collective variables (CVs)
CYANA............................ 352353, 355, 357, 361, 365367,
370, 374377
Bacterial leucine transporter .............................................260

Bias potential ............................152154, 156, 159161, 165,
167, 168, 255, 256, 258, 259
Binding site ............... 119, 326, 328, 383394, 448, 461, 462
Biochemical reactions .........................................................27
Bond constraint ....................................................................8
Bovine pancreatic trypsin inhibitor (BPTI)...................911,
13, 1619, 237
Data-driven docking ................................................401404

De novo modelling ........................................... 332, 344, 347
Deuterium order parameter ................................ 85, 118, 119
Dihedral angle restraints........................... 354, 359361, 369
DIMS. See Dynamic importance sampling method (DIMS)
Direct coupling analysis (DCA) ....................................... 334
Discrete optimized protein energy
(DOPE)................................................ 114, 122, 266
Drude polarizable force field ........................................5051
Drugbank ......................................... 426, 428434, 436, 441
Drug target ..................................91, 331, 425443, 445, 446
Dual topology................................... 182183, 185, 191, 205
Dynamic importance sampling method
(DIMS) ........................................ 242, 243, 245249
CAPRI .............................................................................400
CFT. See Crooks fluctuation theorem (CFT)
CHARMM ................................ 9, 4850, 52, 53, 55, 58, 62,
65, 74, 94, 96, 103, 246249, 292, 293
Chemical similarity ..................................................316, 431
Cholesterol ...............................................................103, 114
CING ................................................353, 357, 365, 370373
Clustering ......................... 294, 295, 304, 387, 390, 392, 405,
411, 413, 415, 419, 422, 448452, 462, 464
CNS ..........................353, 355, 357, 367369, 376, 377, 400,
401, 406, 408, 411413, 415, 416, 418, 420422
Coarse-grained (CG) method ....................................98102
Coarse-graining (CG) ..........................27, 98101, 126128,
131, 133, 136, 137, 215
Collective variables (CVs) ................ 152162, 164168, 249
Committor ......................................... 43, 296298, 300, 305
Comparative modeling .............................................309329
Computing rate constants ............................................3234
Elastic network model (ENM) .......................... 99, 215218,

221, 223, 233, 234, 245
Electron density profile ................................................8789
Electronic polarization .................................................53, 65
Endogenous disease .......................................... 427, 429430
Energy-based methods ............................. 384, 386, 389392
Energy minimization............................ 5, 6, 9, 12, 14, 19, 22,
75, 80, 81, 99, 111, 122, 127, 135, 136, 142, 146, 260,
261, 271, 273274, 403, 404, 412, 418
Enhanced sampling ..................................................152, 157
ENM. See Elastic network model (ENM)
Exogenous disease ....................................................427429
F
FILM. See Folding in lipid membrane (FILM)
Folding ................................. 4, 16, 18, 24, 2729, 43, 47, 49,
125127, 130, 152, 157, 158, 160, 173, 175, 238, 254,
289305, 315317, 323, 334
DOI 10.1007/978-1-4939-1465-4, Springer Science+Business Media New York 2015
471
MOLECULAR MODELING OF PROTEINS

472 Index
Folding in lipid membrane (FILM) ......................... 332, 334
Force field ...................................5, 6, 911, 15, 18, 20, 21, 23,
24, 4765, 7376, 85, 89, 94, 96, 104, 111, 113, 122,
125146, 157, 184, 187, 190, 191, 199, 206, 237, 240,
244, 254, 270272, 291293, 295, 325, 355, 367, 419,
421, 426, 447, 448
Free energy ................................. 4, 28, 49, 93, 126, 151, 173,
239, 284, 289, 315, 426, 454
Ligand filtering.........................................................448, 451

Lipid bilayer ...................................7381, 83, 85, 86, 88, 93,
9597, 102, 103, 109115, 117, 119, 121, 123, 130,
131, 133, 136, 138, 261, 264, 267, 343
Lipid order parameter.......................................................118
Lipid topology .............................................. 7475, 113114
g_arom.............................................................. 116117, 120

Geometry-based methods ........................ 384, 385, 388392
g_helixaxis ........................................................ 115116, 120
g_membed .......................................................... 96, 111, 138
g_prolip ....................................................................120, 121
g_remove_water........................................................132, 139
GRId-based Force Field INput (GRIFFIN)......................96
GROMACS ............................... 9, 11, 1315, 18, 2024, 29,
3436, 42, 74, 75, 82, 86, 88, 96, 103, 110, 116, 117,
120, 121, 123, 128, 131133, 135, 136, 138, 139, 143,
144, 157, 159, 161, 166168, 186, 198, 199,
204206, 292, 293, 355
GROMOS ..................................................... 48, 62, 94, 187
g_under.....................................................................117, 121
g_xycoor ...................................................................115, 120
g_zcoor .....................................................................115, 120
MARTINI .................................................. 73, 98, 128137,

139141, 144, 146
Martinize script ........................................................131132
MAVEN........................................... 215, 221223, 228, 230
Membrane protein....................24, 7389, 91104, 109123,
128, 130, 131, 140, 261, 262, 331348, 392, 436
Metadynamics ...................................... 19, 43, 126, 152154
Mixed-lipid bilayers..................................................113114
Molecular dynamics simulations.................. 324, 28, 4765,
91104, 110, 111, 116, 119, 122, 132, 185, 193, 196,
205, 237249, 273, 291, 292, 352, 386, 405, 418
Movie ................................................................... 17, 21, 221
Multi-body docking .................................................417418
H
HADDOCK ............. 357, 361, 363, 373, 400, 401, 403422
HIV-1 protease.................. 215, 216, 218221, 228, 231, 233
HMDB. See Human metabolome database (HMDB)
Homology .................................. 94, 101, 130, 260, 285, 313,
316318, 320, 321, 323, 327, 328, 332, 335, 357, 384,
386, 388, 391393, 404, 428, 435, 447, 451
Human metabolome database (HMDB) ................. 426, 429,
430, 436439, 441
Hybrid topology ............................................... 192, 193, 206
Hydrogen bonds ............................. 12, 1617, 19, 24, 27, 52,
54, 55, 94, 98, 99, 115, 117, 119, 120, 127, 129, 131,
158160, 293, 325, 326, 332, 354, 369, 370, 401, 405,
407, 413, 452, 460
I
Importance sampling ................................................242243
Inflategro ...........................................111, 121, 138, 139, 146
Ion channels ....................................................... 3, 4, 92, 125
J
Jarzynski equality.............................................. 174, 180, 192
K
Kinetic network ........................................................290, 302
Kinetics............................... 5, 8, 28, 142, 143, 173, 176, 177,
239, 243, 253, 289, 290, 293, 300, 302, 303, 305
N
NAMD...............................................9, 50, 65, 96, 255, 277,
447, 448, 461
NMR spectroscopy. See Nuclear magnetic resonance (NMR)
spectroscopy
Non-equilibrium methods ................................................175
Normal modes .......................... 127, 217, 218, 221, 231, 245,
246, 248, 418, 446
Nuclear magnetic resonance (NMR)
spectroscopy ...........................331, 351, 358, 372, 447
O
OnsagerMachlup (OM) ......................................... 240, 248
Open3DALIGN ..............................................................186
OPLS ........................................................... 21, 62, 187, 421
Ovomucoid ...............................................................197200
P
1-Palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine
(POPE) ....................................72, 74, 75, 79, 8587,
113, 120, 121, 262264, 269271
Parameter optimization .............................. 48, 52, 58, 6061
Parametrization ......................................................50, 5261
Particle-Mesh-Ewald summation (PME) ................. 7, 8, 18,
19, 22
PCA. See Principal component analysis (PCA)
Peptidelipid interaction ..................................................117
Periodic boundary conditions ............................... 7, 273, 278
Perturbation method ........................................ 177179, 188
Phi-value .................................................. 290, 291, 300304

473
Index
PLUMED ........................................152, 157159, 161, 163,
164, 167, 168
Polarizable force field ...................................................5065
POPE. See 1-Palmitoyl-2-oleoyl-sn-glycero-3phosphoethanolamine (POPE)
Position restrained equilibration ...................................1213
Position restraints ........................... 10, 13, 23, 24, 35, 43, 81,
111, 190, 273, 275, 276
Pressure coupling ............................................ 13, 23, 24, 143
Principal component analysis (PCA).....................215218,
222, 223, 228, 229, 233235, 244, 245, 260, 261,
277285
Propensity-based methods................................ 384, 387, 390
Protein dynamics ................................47, 115, 138, 213235,
254, 261, 275
Protein folding.......................................... 289305, 315, 323
Proteinligand binding ..................................... 186, 383394
Proteinligand interaction ................................................393
Proteinprotein interactions .............................. 47, 130, 312,
321, 399422
PSICOV .................................................. 334336, 338, 340,
344, 347, 348
Q
QM/MM ......................................................... 20, 28, 29, 50
QR factorization............................................... 448, 450, 451
Quantum/classical mechanics.......................................2744
R
Radius of gyration ................................ 1617, 135, 401, 403
Rate calculations .................................................................29
Reaction-field electrostatics ..........................................19, 23
Reaction rates .......................................................2744, 440
Reactive path .........................................2931, 34, 36, 37, 39
Receiver operating characteristic (ROC)
analysis..................................................................463
RECOORD ..................................... 353, 357, 367369, 373
Relative free energy .................................. 175, 187, 239, 244
Relaxed complex scheme (RCS) ............................... 446, 447
Replica exchange method ................................. 152, 154156
Residueresidue contacts ................................. 110, 120, 121,
333, 335336
Rhodopsin dimer ...................................... 131, 132, 137, 141
Rhombic dodecahedron.......................................... 18, 19, 24
RMSD. See Root-mean-square displacement (RMSD)
RMSF. See Root-mean-square-fluctuation (RMSF)
ROC analysis. See Receiver operating characteristic (ROC)
analysis
Root-mean-square displacement (RMSD) .................... 14, 15,
24, 63, 93, 113, 115, 145, 161, 168169, 248, 249, 260,
293295, 297, 298, 316, 318, 332334, 348, 356, 366,
372, 377, 405, 407, 413417, 422, 448450, 453, 464
Root-mean-square-fluctuation (RMSF) ..................... 15, 16,
120, 123, 144, 145, 279283
S
Sampling intermediates .................................... 240, 241, 246
Secondary structure ............................1516, 19, 63, 64, 115,
116, 128, 130, 131, 133, 158, 299, 314, 315, 322, 323,
325, 335337, 340342, 353, 354, 360, 422
Self-diffusion constant .................................................50, 85
Simulation .................................... 3, 27, 47, 73, 91, 109, 125,
151, 175, 215, 237, 253, 289, 313, 352, 386, 405, 446
Steepest descent.......................................6, 12, 111, 135, 157
Structural bioinformatics ....................................................93
Structure alignment .......................................... 226, 227, 384
Structure calculation .........................352358, 361367, 369,
373, 377, 401, 405
Structure refinement.................................................351377
Structure validation ......................................... 327, 353, 356,
357, 370373
System setup ............................................. 110114, 200201
T
TALOS+ ............................................357, 359361, 365, 374
Temperature coupling .........................................................23
Temperature factors ...........................9, 15, 16, 344, 345, 422
Template identification.............................................317320
Thermodynamic cycles .............................. 94, 183, 187193,
196198, 200, 202
Thermodynamic integration ............................. 175, 179, 180
Thermostat ...................................................... 8, 13, 23, 142,
143, 166, 293
Threading ................................................. 318321, 323, 328
TIP3P water ....................................................... 11, 157, 405
Topology ..................................9, 11, 21, 38, 7476, 79, 81, 82,
85, 89, 110113, 120122, 132, 133, 135137,
139141, 157, 182183, 185187, 191193, 195196,
199, 200, 205, 206, 262, 270, 271, 277, 278, 291,
332334, 337, 344346, 368, 404, 412, 419, 421
Topology generation .........................183, 185187, 195, 200,
206, 404, 421
TPS. See Transition path sampling (TPS)
Trajectory analysis ........................................................1416
Transition path ...................................2832, 34, 3644, 180,
241, 245, 295, 297299, 301
Transition path sampling (TPS) ................... 2744, 240, 241
Transition state .............................36, 43, 243244, 255, 257,
258, 290292, 296300, 302
Transmembrane protein ...........................262, 267, 313, 315,
325, 333, 343, 347
Transmembrane topology ................................. 334337, 345
Trp-cage miniprotein ...................................... 152, 157160,
162, 165
Trypsin inhibitors .....................................................194197
U
Unfolding ....................................47, 291293, 296, 302304

474 Index
V
Validation ..............52, 56, 62, 74, 190192, 238, 279284, 317,

320, 321, 325327, 352, 353, 356358, 370372, 453
Virtual screening ..............................425443, 446, 447, 449,
451, 452, 454, 456
Virtual sites ........................................................................18
Volume per lipid .................................................................85
Well-tempered ensemble (WTE) ........................... 152, 154,

156, 157, 162164, 167, 168
WORDOM ............................................................292295,
298, 302
WTE. See Well-tempered ensemble (WTE)

Molecular Modeling of Proteins PDF

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Molecular Modeling of Proteins PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Molecular Modeling of Proteins PDF

Uploaded by

Copyright:

Available Formats

Methods in

Molecular Biology 1215

Andreas Kukol Editor

For further volumes:

Molecular Modeling of Proteins

Additional material to this book can be downloaded from http://extras.springer.com

1 Molecular Dynamics Simulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10 The Use of Experimental Structures to Model Protein Dynamics. . . . . . . . . . .

PROTEIN STRUCTURE DETERMINATION

14 Comparative Modeling of Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17 Methods for Predicting ProteinLigand Binding Sites . . . . . . . . . . . . . . . . . . .

MARC F. LENSINK Interdisciplinary Research Institute, CNRS USR3078, University Lille1,

and/or function of proteins (rather than studying the dynamics of

Molecular Dynamics Simulations

Accessible to atomic-detail simulation today (2013)

the equations. For most purposes these approximations work great,

The updated coordinates are then used to evaluate the potential

Molecular Dynamics Simulations

Initial input data:

Repeat for millions of steps

Compute potential V(r) and

Update coordinates &

Collect statistics and write

Fig. 3 Simplified flowchart of a typical molecular dynamics simulation. The basic

Even the smallest chemical sample we can imagine is far too

Molecular Dynamics Simulations

The Bovine Pancreatic Trypsin Inhibitor (BPTI) is a small 58-residue

In addition to the coordinates/velocities that change each step,

pdb2gmx f 6PTI.pdb water tip3p

Molecular Dynamics Simulations

overlapping water molecules removed. The BPTI system will

grompp f ions.mdp p topol.top c solvated.gro \

The added hydrogens and broken hydrogen bond network in water

To avoid unnecessary distortion of the protein when the molecular

Molecular Dynamics Simulations

dynamics we also control the temperature with the Bussi thermostat

conflicts with reality and available computer resources. We will

One of the most important fundamental properties to analyze is

Molecular Dynamics Simulations

roughly the resolution of the X-ray structure. The difference is

Vibrations around the equilibrium are not random, but depend on

Another measure of stability is the protein secondary structure. This

removed. The DSSP secondary structure definition is pretty tight,

Molecular Dynamics Simulations

The result will be written to the file gyrate.xvg, which includes

A normal movie uses roughly 30 frames/second, so a 10-s movie

4 Speeding Things uptoSolve aReal Problem

Molecular Dynamics Simulations

Fig. 11 Two-dimensional example of how a hexagonal box leads to lower volume

reaction-field electrostatics instead. This difference is easyjust

have not advanced our knowledge either of simulation methods or

Molecular Dynamics Simulations

GTX780 or GTX TITAN.Beware that this development is even

as SPC/E that improve bulk properties, but the standard

Molecular Dynamics Simulations

enough it is worth investigating reaction-field electrostatics,

13. The Bussi thermostat is a great advance for simulations. It is