Estimating Functional Connectivity and Topology in Large-Scale Neuronal Assemblies

Springer Theses
Recognizing Outstanding Ph.D. Research
Vito Paolo Pastore
Estimating
Functional
Connectivity
and Topology
in Large-Scale
Neuronal Assemblies
Statistical and Computational Methods
Springer Theses
Recognizing Outstanding Ph.D. Research

Aims and Scope
The series “Springer Theses” brings together a selection of the very best Ph.D.
theses from around the world and across the physical sciences. Nominated and
endorsed by two recognized specialists, each published volume has been selected
for its scientific excellence and the high impact of its contents for the pertinent field
of research. For greater accessibility to non-specialists, the published versions
include an extended introduction, as well as a foreword by the student’s supervisor
explaining the special relevance of the work for the field. As a whole, the series will
provide a valuable resource both for newcomers to the research fields described,
and for other scientists seeking detailed background information on special
questions. Finally, it provides an accredited documentation of the valuable
contributions made by today’s younger generation of scientists.
Theses are accepted into the series by invited nomination only

and must fulfill all of the following criteria
• They must be written in good English.
• The topic should fall within the confines of Chemistry, Physics, Earth Sciences,
Engineering and related interdisciplinary fields such as Materials, Nanoscience,
Chemical Engineering, Complex Systems and Biophysics.
• The work reported in the thesis must represent a significant scientific advance.
• If the thesis includes previously published material, permission to reproduce this
must be gained from the respective copyright holder.
• They must have been examined and passed during the 12 months prior to
nomination.
• Each thesis should include a foreword by the supervisor outlining the signifi-
cance of its content.
• The theses should have a clearly defined structure including an introduction
accessible to scientists not expert in that particular field.
Indexed by zbMATH.
More information about this series at http://www.springer.com/series/8790

Vito Paolo Pastore
Estimating Functional
Connectivity and Topology
in Large-Scale Neuronal
Assemblies
Statistical and Computational Methods
Doctoral Thesis accepted by
The University of Genova, Italy
123
Author Supervisor
Dr. Vito Paolo Pastore Prof. Sergio Martinoia
Department of Informatics, Bioengineering, Department of Informatics
Robotics and System Engineering (DIBRIS) Bioengineering, Robotics
University of Genova and System Engineering (DIBRIS)
Genova, Italy University of Genova
Genova, Italy
Institute of Biophysics
National Research Council (CNR)
Genova, Italy
ISSN 2190-5053 ISSN 2190-5061 (electronic)

Springer Theses
ISBN 978-3-030-59041-3 ISBN 978-3-030-59042-0 (eBook)
https://doi.org/10.1007/978-3-030-59042-0
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed 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.
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.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Parvum addas parvo,
magnus acervum erit
Supervisor’s Foreword
It is a great pleasure to introduce the thesis work of Dr. Vito Paolo Pastore accepted
for publication within Springer Theses and awarded with a prize for an outstanding
original work. Vito Paolo Pastore performed his Master thesis under my supervision
and then joined my research group on Neuro-engineering (Department of
Informatics, Bioengineering, Robotics and Systems Engineering) for his Ph.D.
program on Bioengineering and Robotics at the University of Genova.
He started in November 2014 and completed his studies with a successful thesis
defense on February 13, 2018. The thesis reports an innovative study dealing with
high-density large-scale microelectrode arrays (HDMEAs) and with the specific
design and implementation of new algorithms for the characterization of the elec-
trophysiological behavior of neuronal cultures coupled to such HDMEAs. Part
of the presented results has been awarded by the National Bioengineering Group
with the “Award for young scientists’’ in 2018, and most of them have been
published in Plos Computational Biology on August 2018. Other results have been
preliminary published on Neuroinformatics on January 2018.
In vitro neuronal networks for studying neuronal physiological and pathological
functions are well-established experimental models in neuroscience since decades.
They are widely used for molecular, cellular and networks studies, mainly for their
possibility of manipulation, control and long-term culturing. They have recently
experienced a new attention with the advent of human models thanks to the
introduction of induced pluripotent stem cells (iPSCs). In the last decade, there was
an increasing attention in the development of cultured networks from human iPSCs
by recognizing the advantages, from the patho-physiological point of view of such
preparations, to model neuronal diseases for translational research. Specific interest
in functional connectivity studies related to the introduction of the concept of
connectome has experiences as well a new and renowned interest. Notwithstanding
the increasing attention of human in vitro neuronal models coupled to high-density
MEAs, most of the electrophysiological studies are still lacking of new and ad hoc
algorithms for reliable functional-effective connectivity estimation.
vii
viii Supervisor’s Foreword
In his thesis work, Vito Paolo Pastore greatly contributed to the development of
new experimental algorithms and their innovative implementation specifically tai-
lored for in vitro models coupled to HDMEAs. These new computational tools were
systematically validated first with computational model systems and then with
experimental model systems. His work included new analysis tools (design and
implementation) as well as optimization of computational models and experimental
validations. Vito Paolo Pastore demonstrated the possibility of estimating both
excitatory and inhibitory connection in large scale in vitro neuronal networks
coupled to HDMEAs.
This validated analysis method and approach opens new possibilities in the
network electrophysiology field, allowing estimating functional-effective connec-
tivity maps in vitro neuronal systems supporting characterization for the many neural
diseases in which the balance excitation–inhibition is compromised. Moreover, it
constitutes a step forward in the field of neuro-engineering with specific relevance
for the support to the analysis of human-based in vitro experimental model systems
under the perspective of personalized and precision medicine.
Genoa, Italy Prof. Sergio Martinoia

June 2020
Parts of this thesis have been published in the following documents:
Journal Publications
Pastore, V. P., Massobrio, P., Godjoski, A. & Martinoia, S. Excitatory-inhibitory

links and network topology in large scale neuronal assemblies from multi-electrode
recordings. PLoS Comput Biol. 2018;14(8): e1006381. Published 2018 Aug 27.
https://doi.org/10.1371/journal.pcbi.100638
Pastore, V. P., Godjoski, A., Martinoia, S. & Massobrio, P. SPICODYN: A Toolbox
for the analysis of neuronal network dynamics and connectivity from multi-site spike
signal recordings. Neuroinformatics, https://doi.org/10.1007/s12021-017-9343-z
(2017).
Pastore, V. P., Poli, D., Godjoski, A., Martinoia, S. & Massobrio, P. ToolConnect:
a functional connectivity toolbox for in vitro networks. Frontiers in Neuroinformatics
10, https://doi.org/10.3389/fninf.2016.00013 (2016).
Poli, D., Pastore, V. P., Martinoia, S. & Massobrio, P. From functional to structural
connectivity using partial correlation in neuronal assemblies. Journal of Neural
Engineering 13, 026023 (2016).
Poli, D., Pastore, V. P. & Massobrio, P. Functional connectivity in in vitro neuronal
assemblies. Frontiers in Neural Circuits 9, https://doi.org/10.3389/fncir.2015.00057
(2015).
ix
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 1
1.1 Bridging Structure and Function of Neural Circuits . . . . . . . .... 1
1.2 Different Types of Connectivity to Describe Neuronal
Assemblies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 2
1.3 The Relevance of the In Vitro Model . . . . . . . . . . . . . . . . . .... 3
1.4 Micro Electrode Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 3
1.5 How to Infer Functional Connectivity in In Vitro Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 4
1.6 Functional Connectivity Analysis Requires Ad-Hoc Software
and Algorithms Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.7 Goal of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.8 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Correlation-Based Connectivity Methods . . . . . . . . . . . . . . . . . . . 11
2.1.1 Cross-Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2 Cross-Correlation Histogram . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Filtered and Normalized Cross-Correlation Histogram
(FNCCH) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.4 Partial Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Information-Theory Based Connectivity Methods . . . . . . . . . . . . . 17
2.2.1 Transfer Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Delayed Transfer Entropy . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.3 Delayed High Order Transfer Entropy . . . . . . . . . . . . . . . 19
2.2.4 Joint Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Thresholding Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Hard Threshold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2 Shuffling Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
xi
xii Contents
2.4 Metrics to Evaluate the Connectivity Methods’ Performances . . . . 24

2.4.1 Receiver Operating Characteristic (ROC) Curve . . . . . . . . 24
2.4.2 Matthews Correlation Coefficient (MCC) Curve . . . . . . . . 25
2.5 Graph Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.6 BIOCAM and Active Pixel Sensor (APS) Array . . . . . . . . . . . . . . 29
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1 Transfer Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.1 Validation of the TE by Means of in Silico Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.2 DTE Application to Neural Networks Coupled to the
MEA-4k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Joint Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.1 Shuffling Thresholding Procedure Validation on in
Silico Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.2 Hard Threshold Validation on in Silico Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.3 JE Application to Cortex Network Coupled to the
HD-MEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 FNCCH Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.1 Validation of the FNCCH by Means of in Silico Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.2 Functional Connectivity and Emergent Network
Topologies in In Vitro Large-Scale Neural Networks . . . . 45
3.4 FNCCH Results: Discussion and Observations . . . . . . . . . . . . . . . 50
3.4.1 Identification of Inhibition . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.2 The Emergence of a Scale Free and Small-World
Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4.3 Comparison with a Transfer Entropy Based Algorithm . . . 53
3.4.4 Partial Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.5 Validation of the PC by Means of in Silico Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.6 PC Application to Experimental Data . . . . . . . . . . . . . . . 56
3.4.7 Connectivity During Development . . . . . . . . . . . . . . . . . 56
3.5 Algorithm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.1 FNCCH: Algorithm Optimization . . . . . . . . . . . . . . . . . . 60
3.5.2 Transfer Entropy Algorithm Optimization . . . . . . . . . . . . 61
3.5.3 Joint Entropy Algorithm Optimization . . . . . . . . . . . . . . . 62
3.6 Software Development for Functional Connectivity Analysis . . . . . 63
3.6.1 Main Graphical Interface . . . . . . . . . . . . . . . . . . . . . . . . 63
3.6.2 Example of TOOLCONNECT’s Application to Experimental
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.6.3 Multi-threading Implementation . . . . . . . . . . . . . . . . . . . 68
Contents xiii
3.6.4 Input Data Processing and Conversion Procedures . . . . . . 69

3.6.5 Spiking and Bursting Features Analysis . . . . . . . . . . . . . . 71
3.6.6 Spike Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.6.7 Burst Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.6.8 Spiking and Bursting Statistics . . . . . . . . . . . . . . . . . . . . 73
3.6.9 Comparison with Other Software . . . . . . . . . . . . . . . . . . 75
3.6.10 Use of SPICODYN to Process Long-Lasting Recordings
on MEA-4k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6.11 Dynamic Characterization . . . . . . . . . . . . . . . . . . . . . . . . 77
3.6.12 Connectivity Characterization . . . . . . . . . . . . . . . . . . . . . 77
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Acronyms
APS Active Pixel Sensor

CC Cluster Coefficient
CCH Cross-Correlation Histogram
DHOTE Delayed High-Order Transfer Entropy
DTE Delayed Transfer Entropy
FN False Negative
FNCCH Filtered Normalized Cross-Correlation Histogram
FP False Positive
FPR False Positive Rate
id in-degree
JE Joint Entropy
MCC Matthews Correlation Coefficient Curve
MEA Micro-Electrodes Array
NCCH Normalized Cross-Correlation Histogram
od out-degree
PC Partial Correlation
PL Path Length
PSTH Post-Stimulus Time Histogram
RCC Rich-Club Coefficient
ROC Receiver Operating Characteristic Curve
SWI Small World Index
SWM Synaptic Weight Matrix
TCM Thresholded Connectivity Matrix
td total degree
TE Transfer Entropy
TN True Negative
TP True Positive
TPR True Positive Rate
xv
Chapter 1
Introduction
1.1 Bridging Structure and Function of Neural Circuits
One of the most fundamental features of a neural circuit is its connectivity since
the single neuron activity is not due only to its intrinsic properties but especially to
the direct or indirect influence of other neurons [1]. As recently reviewed by Yuste
[2], thanks to technology advancements, the era of the “neuron doctrine” has faded:
neuronal assemblies can be considered the physiological units of the brain which
generate and sustain the functional properties as well as the dynamical states of
the entire system. Nervous systems are complex networks par excellence, capable of
elaborating and integrating information from multiple external and internal sources in
real time. It has been proposed that such information is managed by means of neural
networks complying with two competing demands, which might also be consid-
ered as fundamental organizational principles: functional segregation and functional
integration, enabling both the rapid extraction of information and the generation of
coherent brain states [3]. As confirmed by recent studies reporting structural anal-
yses of brain networks carried out on datasets describing the cerebral cortex of
mammalian animal models (e.g. rat, cat, monkey), cortical areas were found to be
neither completely connected with each other nor randomly linked; instead, their
interconnections show a specific and intricate organization [4]. It is fundamental to
elaborate and adopt research strategies aimed at a comprehensive structural descrip-
tion of such interconnections as well as the networks’ elements forming the human
connectome [5]. The connectome will significantly increase our understanding of
how functional brain states emerge from their underlying structural substrate, and
will provide new mechanistic insights into how brain function is affected if this
structural substrate is disrupted.
© The Editor(s) (if applicable) and The Author(s), under exclusive license 1
to Springer Nature Switzerland AG 2021
V. P. Pastore, Estimating Functional Connectivity and Topology in Large-Scale
Neuronal Assemblies, Springer Theses, https://doi.org/10.1007/978-3-030-59042-0_1
2 1 Introduction
1.2 Different Types of Connectivity to Describe Neuronal

Assemblies
The types of connectivity used to describe the interactions of neuronal networks are:
structural, functional and effective.
Structural connectivity (Fig. 1.1a): the structural or anatomical connectivity
indicates the physical interactions (i.e., a chemical or electrical synapses) that link
network’s neurons at a given time [6]. Therefore, we can determine which neural
units can directly interact with each other. The structural connectivity ranges over
multiple spatial scales, since we can detect morphological connections both in local
microcircuits and in long-range interactions that link different sub-networks. In a
short time scale (about., less than one minute), such morphological connections
mediated by dendritic spines are static, while in a longer time scale, they are dynamic,
since physiological mechanisms of learning, plasticity and development can shape
the morphological circuits [7].
Functional connectivity (Fig. 1.1b): functional connection indicates the correla-
tion between time series of spikes coming from different neurons. It measures statis-
tical interdependence without considering any causal effects; it is time-dependent and
“model-free”. Therefore, two neurons are functionally linked, if we can predict the
activity of one of the two neurons on the basis of the activity of the other neuron. Func-
tional properties of single neurons are strongly driven by their anatomical connec-
tions, dendritic arborizations and synaptic distributions [6]. Moreover, functional
Fig. 1.1 Classification of the neural network connections. a Structural connectivity involving the
physical connections between neurons. b Functional connectivity, involving functional connections
and activity synchrony and correlation between neurons, without any causal model. c Effective
connectivity. It is derived from functional connectivity introducing a causal model. Reproduced
from [8]
1.2 Different Types of Connectivity to Describe Neuronal Assemblies 3
interactions can contribute to the shaping of the underlying anatomical substrate

through activity-dependent synaptic modifications.
Effective connectivity (Fig. 1.1c): effective connectivity indicates the presence of
a connection when a neuron on the network directly affects another neuron through a
causal relationship between the activities of the two neurons. In other words, “effec-
tive” means any observable interactions between neurons that alters their firing
activity; so it is not “model-free” like functional connectivity, but can require the
specification of a causal model including structural parameters.
1.3 The Relevance of the In Vitro Model
The final goal of connectivity analysis is the reconstruction of the human connec-
tome, thus, the application of statistical measures to the in vivo model in both phys-
iological and pathological states. Since the system under study (i.e. brain areas, cell
assemblies) is highly complex and not controllable, to achieve the purpose described
above, it is useful to adopt a reductionist approach. A possible strategy to reduce such
a complexity makes use of in vitro experimental models coupled to Micro-Electrode
Arrays (MEAs). Nowadays, dissociated neuronal cultures coupled to MEAs are
fairly used to better understand the complexity of brain networks. One of the main
advantage to use dissociated neuronal assemblies is the possibility to manipulate and
control their connectivity: in other words, it is feasible to drive the connectivity of a
network and to study how such a topological configuration can shape the emergent
dynamics. Examples of engineered networks started in 1975 with the pioneering
work of Letourneau [9] who investigated the role of different adhesion substrates for
promoting the initiation, elongation and branching of the axons. The possibility to
use the in vitro model, that is a valuable but at the same time reduced and simplified
experimental model can be considered a great breakthrough in understanding the
functional properties of neuronal networks.
1.4 Micro Electrode Arrays
MEAs are a powerful tool for simultaneously monitoring and acquiring the electro-
physiological activity of neural preparations at many sites (Fig. 1.2a). The electrodes
embedded in such devices can record electrophysiological activity in a non-invasive
way (i.e., extracellularly) and therefore, under proper maintenance conditions, can
allow long-term recordings (i.e., from hours up to months) [10]. Commercial avail-
able MEAs usually provide 60–120 electrodes with 100–500 μm inter-electrode
spacing (Fig. 1.2b). However, recent advances in multichannel recording techniques
have made possible to observe the activities of thousands of neurons simultane-
ously, and made routine the acquisition of massive amounts of empirical data or
4 1 Introduction
Fig. 1.2 MEA and extracellular signals: a The activity of a cortical neural network (28 DIVs)
presents a mix of bursting and spiking activity (top). Applying a spike detection algorithm, time
series are converted into a serial point process (bottom). b–c Examples of Micro-Electrode Arrays
(MEAs) made up of b 60, c 4096 electrodes. Reproduced from [8]
high-density configurations with thousands of microelectrodes (4000–10,000) with

a spatial resolution of some tens of micrometers (Fig. 1.2c) [11, 12].
The characteristics of these devices allow different studies on neuronal networks
like electrical [13] and chemical manipulation [14], and/or physical segregation in
sub-populations (e.g. [15]).
More recently the scientific community is beginning to use MEAs for charac-
terizing the underlying functional connectivity, and its interplay with the expressed
dynamics [16], especially by exploiting the high-density systems which allow a more
accurate reconstruction of the network topology [17].
1.5 How to Infer Functional Connectivity in In Vitro

Neural Networks
In vitro networks represent an excellent benchmark for the validation of functional-

effective connectivity methods [8, 18]. Indeed, reconstructing the detailed connec-
tivity of a neuronal network from spike data (functional connectivity) is not trivial,
and it is still an open issue, due to the complexities introduced by neuron dynamics and
the high number of network connections [19, 20]. Researchers have often studied the
status of structural connectivity and its relationship to functional-effective connec-
tivity, and tried to relate connectivity to dynamics [21]. In order to gain understanding
of the exchange of information within and between hypothetical neuronal assemblies
1.5 How to Infer Functional Connectivity in In Vitro Neural Networks 5
and to extract topological characteristics, it is necessary to monitor, in parallel, the

activities of many neurons and to analyze large amount of data from these elec-
trophysiological measurements. For this purpose, the recent developed high-density
MEAs technologies are certainly very attractive since they allow monitoring the on-
going electrophysiological spatio-temporal patterns of complex networks [12, 22],
from the majority of neurons in the network.
To estimate the functional-effective connectivity of in vitro networks, there are two
different strategies: the first one relies on the direct analysis of the acquired sequence
of voltage values (Fig. 1.2a top) from each recording electrode (i.e., the time series).
The other approach deals with point processes (e.g., spike trains). Practically, a spike
train is a sequence of samples equal to 1 if a spike is detected and 0 otherwise
(Fig. 1.1a bottom) [23–26].
Statistical analysis of spike train data was pioneered by Perkel [27] and followed
by more than four decades of methodology development in this area [28]. Analyt-
ically, cross-correlation based methods remain the main statistics for evaluating
interactions among the elements in a neuronal network, and produce a weighted
assessment of the connections strength. Weak and non-significant connections may
tend to obscure the relevant network topology constituted by strong and significant
links, and therefore are often discarded by applying an absolute, or a proportional
weighted threshold [29]. Correlation based techniques include: independent compo-
nents analysis and various measures of synchrony, smoothed ratio of spiking activity
[30], cross-correlation and cross-correlogram [31, 32], correlation coefficient [33],
partial-correlation [34]. Cross-Correlation (CC) measures the frequency at which
one particular neuron or electrode fires (“target”) as a function of time, relative to the
firing of an event in another network (“reference”). Partial Correlation (PC) is able to
distinguish between direct and indirect connections by removing the portion of the
relationship between two spike trains that can be attributed to linear relationships with
recorded spike trains from other neurons [34, 35]. Other popular techniques to infer
functional-effective connectivity are based on Information Theory (IT) methods [36,
37], Granger causality and dynamical causal modeling [38]. Commonly used infor-
mation theoretic measures able to estimate causal relationships are transfer entropy
[39] and joint entropy [18]. Transfer Entropy (TE) is an information-theory based
method that estimates the part of the activity of a neuron which depend from the past
activity of another neuron. Joint Entropy (JE) is an entropic measure of the cross
Inter-Spike Intervals (cISI).
In my PhD work, I focused on CC and Cross-Correlation Histogram (CCH) based
methods, PC, TE and JE. With few exceptions [40, 41], all the recently introduced and
revisited methods deal with excitation, ignoring inhibition or admitting the failure
in reliably identifying inhibitory links [39]. To overcome such a limitation, I intro-
duced a new CCH-based algorithm able to efficiently and accurately infer effective
excitatory-inhibitory links (cf. Filtered Normalized Cross-Correlation Histogram
(FNCCH) section).
In fact, the cross-correlation is able by definition to detect inhibition, but, from
some experimental works related to the analysis of cortical connectivity from in vivo
multi-unit recordings, it was shown that the sensitivity for excitation is much higher
6 1 Introduction
than the sensitivity for inhibition [42] (due to the low firing rates of neurons). A
general lower sensitivity of cross-correlation for inhibition vs. excitation has also
been proved theoretically [43] thus making the task of inhibitory connection’s iden-
tification particularly difficult. However, by using my ad-hoc developed FNCCH
algorithm, I could derive effective connectivity maps (both for excitation and inhi-
bition) reliably extracting topological characteristics from multiple spike trains in
large-scale networks (i.e., thousands of neurons) monitored by large-scale MEAs
(i.e., with thousands of micro-transducers).
Relative to the information theory-based methods, I implemented customized
versions of TE and JE. In particular, I studied a new version of the TE algorithm,
in which I extended the method to deal with multiple time delays (temporal exten-
sion) and with multiple binary patterns (high order extension), obtaining improved
reliability-precision and increasing computational performances. I also worked on
the JE algorithm; in fact, in its original definition the JE provides only a value that
estimates the probability of two neurons to be functionally connected, but in my work
I added the possibility to extract the temporal delay that characterizes a detected func-
tional connection. It is worth to consider that every connectivity method provides
a full n x n connectivity map, if n is the number of the analyzed electrodes. Thus,
a thresholding procedure is required to throw away those values that are close to
or in the noise, and not real connections. This requires setting a threshold for the
connectivity matrix. Exploring the available works in the literature about the anal-
ysis of functional connectivity of in vitro neural networks, it is possible to see several
thresholding procedures, with different levels of complexity. The simplest of such
procedures is to use a hard threshold [34] defined as (μ + n · σ), where μ and σ
are the mean and the standard deviation computed among all the CM’s elements,
respectively, and n is an integer. Another possibility is to use shuffling techniques,
that allow to destroy the information stored in the spike timing, obtaining indepen-
dent spike trains (i.e., surrogate data). In my Ph.D., I focused on both the typologies
of thresholding procedures, introducing a customized shuffling approach for the JE
algorithm. Such an approach, required high computational efficiency and was made
possible only by the optimization strategies that I adopted for the implementation of
the connectivity methods.
1.6 Functional Connectivity Analysis Requires Ad-Hoc

Software and Algorithms Implementation
The analysis of the interactions in large-scale neuronal networks to infer functional

connectivity is a demanding computational process. Recent advances in multichannel
extracellular recording techniques have made possible the simultaneous recording of
the electrophysiological activity of thousands of neurons, with in vitro arrays that can
contain up to thousands of microelectrodes [7, 11, 22, 44, 45]. Such relevant advance-
ments in the technology have made routine the acquisition of massive amounts of
1.6 Functional Connectivity Analysis Requires … 7
data. Moreover, quite often the experimental protocols researchers perform require
long recordings (e.g., tens of minutes, hours) which generate a huge amount of raw
data. Furthermore, to validate an experimental finding, several trials of the same
protocol are necessary, or in other cases, several experimental conditions have to
be tested. For these reasons, the three main features that software tools have to
satisfy are: (i) reasonable times of computation (efficiency); (ii) simple procedures
to process several experiments (automation); (iii) possibility to analyze data coming
from different acquisition systems (flexibility). Indeed, the two first conditions should
not compromise on the accuracy and the efficacy of the developed algorithms. For
the above reasons, new computational strategies are requested for optimizing the
management and analysis of the acquired data.
I developed a new software package, named SpiCoDyn, which implements,
starting from raw data, several algorithms to characterize the spiking and bursting
properties of large-scale neuronal networks. Such a tool is independent from the
acquisition system and is conceived for the analysis of multi-site neuronal spike
signals in HDF5 format (HDFGroup 2013). Particularly, it is specifically tailored to
process recordings acquired by the commercial High-Density Multi-Electrode Arrays
(HD-MEAs) (e.g., 4096 microelectrodes from 3 Brain or 4225 microelectrodes from
Multi Channel Systems, MCS, Reutlingen, Germany).
1.7 Goal of the Thesis
Goal of my Ph.D. work has been to develop a group of statistical measures to infer
functional connectivity in in vitro neural networks. I worked on correlation-based
methods (cross-correlation and partial correlation), and information-theory based
methods [Transfer Entropy (TE) and Joint Entropy (JE)]. More in detail, by means
of the developed methods, I tried to study the interplay between dynamics and
connectivity, using high density resolution systems, with thousands of microelec-
trodes. To fulfill such an aim, I re-adapted the computational logic operations of the
aforementioned connectivity methods to reduce the time requested for the statistical
computations. Moreover, I worked on a new method based on the cross-correlogram,
able to detect both inhibitory and excitatory links. The FNCCH shows a very high
precision in detecting both inhibitory and excitatory functional links when applied
to a computational model of neuronal networks. Concerning the information-theory
based algorithms, I worked on the temporal and pattern extension of the already
existing TE algorithm, by developing the Delayed TE (DTE) and the Delayed High
Order TE (DHOTE) algorithms. Finally, starting from the mathematical definition
in [18] of the JE, I developed a new customized version of the algorithm capable to
detect the delay associated to a functional link, together with a customized shuffling
based thresholding approach. Finally, I embedded all of these connectivity methods
into a user-friendly open source software named SpiCoDyn[46] which inherits the
backbone structure from ToolConnect[47]. SpiCoDyn allows the user to perform
a complete analysis on data acquired from any acquisition system, including: raw
8 1 Introduction
data viewing, spike and burst detection and analysis, functional connectivity analysis,
graph theory and topological analysis.
1.8 Structure of the Thesis
In the first section, Materials and methods, I will describe at first the statistical
measures that I used to infer functional connectivity in in vitro neural networks
coupled to the MEAs. Then, I will introduce the problem of thresholding the connec-
tivity matrix, together with the description of possible solutions that I studied during
my work. Finally, I will describe the metrics that I used to evaluate the connectivity
methods’ performances and all the graph theory’s mathematical instruments funda-
mental to treat the connectivity matrix as a graph, inferring and characterizing the
neural networks’ topological features. The section Results is divided into two sub-
sections. In the first one, I will describe the results obtained by applying each of the
implemented connectivity methods on computational models to evaluate the compu-
tational accuracy, as well as an application to experimental neural networks coupled
to both low and high density MEAs. In the second one, I will describe the logics of
optimization that I adopted in order to reduce the requested time for computation and
the software package SpiCoDyn, explaining the informatics tools and implementa-
tion choices that allowed an improvement of the computational efficiency, making
possible to perform a complete analysis of neural networks coupled to high density
acquisition systems. Finally, a last section summarizes all the obtained results and
suggests a possible perspective for further extension of my work.
References
1. Makarov VA, Panetsos F, de Feo O (2005) A method for determining neural connectivity and
inferring the underlying network dynamics using extracellular spike recordings. J Neurosci
Methods 144:265–279. https://doi.org/10.1016/j.jneumeth.2004.11.013
2. Yuste R (2015) From the neuron doctrine to neural networks. Nat Rev Neurosci 16:487–497
3. Sporns O, Tononi G, Edelman GM (2000) Connectivity and complexity: the relationship
between neuroanatomy and brain dynamics. Neural Netw 13:909–922
4. Sporns O (2011) The human connectome: a complex network. Ann N Y Acad Sci 1224:109–
125. https://doi.org/10.1111/j.1749-6632.2010.05888.x
5. Sporns O, Tononi G, Kotter R (2005) The human connectome: a structural description of the
human brain. PLoS Comput Biol 1:e42. https://doi.org/10.1371/journal.pcbi.0010042
6. Sporns O, Tononi G (2002) Classes of network connetivity and dynamics. Complexity 7:28–38
7. Buchs P-A, Muller D (1996) Induction of long-term potentiation is associated with major
ultrastructural changes of activated synapses. Proc Natl Acad Sci USA 93:8040–8045
8. Poli D, Pastore VP, Massobrio P (2015) Functional connectivity in in vitro neuronal assemblies.
Front Neural Cir 9:57. https://doi.org/10.3389/fncir.2015.00057
9. Letourneau PC (1975) Possible roles of cell to substratum adhesion in neuronal morphogenesis.
Dev Biol 44:77–91
References 9
10. Potter SM, DeMarse TB (2001) A new approach to neural cell culture for long-term studies. J
Neurosci Methods 110:17–24
11. Berdondini L et al (2009) Active pixel sensor array for high spatio-temporal resolution elec-
trophysiological recordings from single cell to large scale neuronal networks. Lab Chip
9:2644–2651
12. Frey U, Egert U, Heer F, Hafizovic S, Hierlemann A (2009) Microelectronic system for high-
resolution mapping of extracellular electric fields applied to brain slices. Biosens Bioelectron
24:2191–2198. https://doi.org/10.1016/j.bios.2008.11.028
13. Wagenaar DA, Madhavan R, Pine J, Potter SM (2005) Controlling bursting in cortical cultures
with closed-loop multi-electrode stimulation. J Neurosci 25:680–688. https://doi.org/10.1523/
JNEUROSCI.4209-04.2005
14. Pancrazio JJ et al (2003) A portable microelectrode array recording system incorporating
cultured neuronal networks for neurotoxin detection. Biosens Bioelectron 18:1339–1347
15. Levy O, Ziv NE, Marom S (2012) Enhancement of neural representation capacity by modular
architecture in networks of cortical neurons. Eur J Neurosci 35:1753–1760. https://doi.org/10.
1111/j.1460-9568.2012.08094.x
16. Massobrio P, de Arcangelis L, Pasquale V, Jensen HJ, Plenz D (2015) Criticality as a signature
of healthy neural systems. Front Syst Neurosci 9:22. https://doi.org/10.3389/fnsys.2015.00022
17. Maccione A et al (2012) Multiscale functional connectivity estimation on low-density neuronal
cultures recorded by high-density CMOS micro electrode arrays. J Neurosci Methods 207:161–
171. https://doi.org/10.1016/j.jneumeth.2012.04.002
18. Garofalo M, Nieus T, Massobrio P, Martinoia S (2009) Evaluation of the performance of
information theory-based methods and cross-correlation to estimate the functional connectivity
in cortical networks. PLoS ONE 4:e6482. https://doi.org/10.1371/journal.pone.0006482
19. Van Bussel F, Kriener B, Timme M (2011) Inferring synaptic connectivity from spatio-temporal
spike patterns. Front Comput Neurosci 5:3. https://doi.org/10.3389/fncom.2011.00003
20. Doiron B, Litwin-Kumar A, Rosenbaum R, Ocker GK, Josic K (2016) The mechanics of state-
dependent neural correlations. Nat Neurosci 19:383–393. https://doi.org/10.1038/nn.4242
21. Sporns O (2013) Structure and function of complex brain networks. Dialogues Clin Neurosci
15:247–262
22. Eversmann B et al (2003) A 128 x 128 CMOS biosensor array for extracellular recording of
neural activity. IEEE J Solid-State Circ 38:2306–2317
23. Maccione A et al (2009) A novel algorithm for precise identification of spikes in extracellularly
recorded neuronal signals. J Neurosci Methods 177:241–249
24. Ide AN, Andruska A, Boehler M, Wheeler BC, Brewer GJ (2010) Chronic network stimulation
enhances evoked action potentials. J Neural Eng 7:016008. https://doi.org/10.1088/1741-2560/
7/1/016008
25. Egert U et al (2002) MEA-Tools: an open source toolbox for the analysis of multi-electrode
data with MATLAB. J Neurosci Methods 117:33–42
26. Borghi T, Gusmeroli R, Spinelli AS, Baranauskas G (2007) A simple method for efficient spike
detection in multiunit recordings. J Neurosci Methods 163:176–180
27. Gerstein GL, Perkel DH (1969) Simultaneously recorded trains of action potentials: analysis
and functional interpretation. Science 164:828–830
28. Brown EN, Kass RE, Mitra PP (2004) Multiple neural spike train data analysis: state-of-the-art
and future challenges. Nat Neurosci 7:456–461
29. Friston KJ (2009) Modalities, modes and models in functional neuroimaging. Science 326:399–
403. https://doi.org/10.1126/science.1174521
30. Ventura V, Cai C, Kass RE (2005) Statistical assessment of time-varying dependency between
two neurons. J Neurophysiol 94:2940
31. Brosch M, Schreiner CE (1999) Correlations between neural discharges are related to receptive
field properties in cat primary auditory cortex. Eur J Neurosci 11:3517–3530. https://doi.org/
10.1046/j.1460-9568.1999.00770.x
32. Salinas E, Sejnowski TJ (2001) Correlated neuronal activity and the flow of neural information.
Nat Rev Neurosci 2:539–550
10 1 Introduction
33. Bedenbaugh P, Gerstein GL (1997) Multiunit normalized cross correlation differs from the
average single-unit normalized correlation. Neural Comput 9:1265–1275
34. Poli D, Pastore VP, Martinoia S, Massobrio P (2016) From functional to structural connectivity
using partial correlation in neuronal assemblies. J Neural Eng 13:026023
35. Eichler M, Dahlhaus R, Sandkuhler J (2003) Partial correlation analysis for the identification
of synaptic connections. Biol Cybern 89:289–302
36. Quiroga QR, Panzeri S (2009) Extracting information from neuronal populations: information
theory and decoding approaches. Nat Rev Neurosci 10:173–185
37. Nigam S et al (2016) Rich-club organization in effective connectivity among cortical neurons.
J Neurosci 36:670–684. https://doi.org/10.1523/JNEUROSCI.2177-15.2016
38. Friston KJ (2011) Functional and effective connectivity: a review. Brain Connectivity 1:13–36.
https://doi.org/10.1089/brain.2011.0008
39. Ito S et al (2011) Extending Transfer entropy improves identification of effective connectivity
in a spiking cortical network model. PLoS ONE 6:e27431. https://doi.org/10.1371/journal.
pone.0027431
40. Freeman GM Jr, Krock RM, Aton SJ, Thaben P, Herzog ED (2013) GABA networks destabilize
genetic oscillations in the circadian pacemaker. Neuron 78:799–806. https://doi.org/10.1016/
j.neuron.2013.04.003
41. Zaytsev YV, Morrison A, Deger M (2015) Reconstruction of recurrent synaptic connectivity of
thousands of neurons from simulated spiking activity. J Comput Neurosci 39:77–103. https://
doi.org/10.1007/s10827-015-0565-5
42. Abeles M (1991) Corticonics: neural circuits of the cerebral cortex. Cambridge University
Press, Cambridge
43. Aertsen A et al (1991) Neural interactions in the frontal cortex of a behaving monkey: signs of
dependence on stimulus context and behavioral state. J Hirnforsch 32:735–743
44. Muller J et al (2015) High-resolution CMOS MEA platform to study neurons at subcellular,
cellular, and network levels. Lab Chip 15:2767–2780. https://doi.org/10.1039/c5lc00133a
45. Viswam V, Dragas J, Muller J, Hierlemann A In: IEEE international solid-state circuits
conference. IEEE, pp 394–396
46. Pastore VP, Godjoski A, Martinoia S, Massobrio P (2017) SPICODYN: a toolbox for the
analysis of neuronal network dynamics and connectivity from multi-site spike signal recordings.
Neuroinformatics 1:15–30. https://doi.org/10.1007/s12021-017-9343-z
47. Pastore VP, Poli D, Godjoski A, Martinoia S, Massobrio P (2016) ToolConnect: a functional
connectivity toolbox for in vitro networks. Front Neuroinf https://doi.org/10.3389/fninf.2016.
00013
Chapter 2
Materials and Methods
2.1 Correlation-Based Connectivity Methods
2.1.1 Cross-Correlation
Cross-correlation [1] measures the frequency at which one particular neuron or elec-
trode fires (“target”) as a function of time, relative to the firing of an event in another
one (“reference”). Mathematically, the correlation function is a statistic representing
the average value of the product of two random processes (the spike trains). Given
a reference electrode x and a target electrode y, the correlation function reduces to a
simple probability Cxy (τ) of observing a spike in one train y at time (t + τ ), given
that there was a spike in a second train x at time t; τ is called the time shift or the
time lag. For my Ph.D. studies, I used the standard definition for the cross-correlation
computation, following a known normalization approach on the CC values [2]. We
can define the cross-correlation as follows:
1 Nx
C x y (τ ) = x(ts )y(ts − τ ) (2.1)
N x N y s=1
where t s indicates the timing of a spike in the x train, N x is the total number
of spikes in the x train and N y is the total number of spikes in the y train. Cross-
correlation is limited to the interval [0, 1] and is symmetric Cxy (τ) = Cyx (−τ).
The cross-correlogram is then defined as the correlation function computed over a
chosen correlation window (W, τ = [−W/2, W/2]). Figure 2.1 describes a possible
strategy to compute the Cross-Correlation Histogram (CCH). The Normalized CCH
(NCCH) is obtained by applying the normalization factor represented by the squared
product of the target and the reference number of spikes to the CCH. Different
shapes of cross-correlograms can be obtained from pairs of analyzed spike trains.
When the two spike trains are independent, the cross-correlogram is flat; if there
12 2 Materials and Methods
Fig. 2.1 Computation of the cross-correlation histogram. The cross-correlation histogram (CCH)
is computed according to a discrete convolution. Thus, the correlation window is centered in each of
the reference train spikes (red squares). The number of target spikes inside each correlation window’s
bin (green squares) are counted, and the cumulative number of spikes, after normalization, represent
the cross-correlation histogram
is any covariation in the spike trains, one or more peaks appear (Fig. 2.2c). The
occurrence of significant departures from a flat background in the cross-correlogram
(i.e., a peak or a trough) is an indication of a functional connection. In particular, a
peak corresponds to an excitatory connection and a trough is relative to an inhibition.
The different amplitude of the peaks can be related to the existence of different levels
of correlation between neural spike trains. A peak can appear also for other kinds of
interaction (e.g., covariations over in response latency, and in neuronal excitability).
Generally, a correlogram can reflect a so-called direct excitatory connection between
the two neurons when a one-sided peak is evident (Fig. 2.2c). This peak is displaced
from the origin of time by latency corresponding to the synaptic delay.
2.1.2 Cross-Correlation Histogram
The use of spike trains data offers the possibility to optimize the cross-correlation
algorithm efficiency. Thus, to overcome the lack of efficiency of many of the proposed
cross-correlation computation strategies, here I present an alternative approach based
on the “direct” spike time stamps inspection that avoids un-necessary calculations
on the binarized spike trains. The only important information is stored in the bins
containing a spike (i.e., spike time stamp), that are significantly less than null bins. If
we consider that, typically, the average mean firing rate in neural networks oscillates
between 0.2 spikes/s and 20 spikes/s [3], acquiring with a sampling frequency of
10 kHz means we will have only 2% of bin with spikes. To account only for the spike
trains time stamps, I developed a new logical version of the CCH.
2.1 Correlation-Based Connectivity Methods 13
Fig. 2.2 Example of FNCCH detection for excitatory and inhibitory links in a network model.
a NCCH computed between two spike trains correspondent to two neurons linked by an inhibitory
link in the model. The NCCH might detect a false excitatory peak (blue circle). b FNCCH of the
two neurons of panel c. The filtering procedure allows to recognize the trough and to detect the
negative peak correspondent to the inhibitory link (blue circle). a NCCH computed between two
spike trains correspondent to two neurons linked by an excitatory link in the model (identified by a
red circle). b FNCCH of the two neurons of panel a. The “entity peak” allows a better recognition
of the excitatory link. Adapted from [4]
2.1.3 Filtered and Normalized Cross-Correlation Histogram

(FNCCH)
Let us consider a reference neuron x and a target neuron y, and let us suppose that we
computed the NCCH between x and y. After the NCCH computation, the maximum
value (i.e., the peak) is commonly used as a value reflecting the strength of the
estimated functional link. If x and y share an excitatory link this procedure works
well (Fig. 2.2a). When x inhibits y, instead, the inhibitory trough will be discarded
in favor of the NCCH peak (Fig. 2.2c), with a misleading excitatory link detection.
Equation (2.2) gives the mathematical definition of the FNCCH computation, that
overcomes this problem.

v= W2

1
F N CC Hx y_ peak
= C x y (τ ) τ = arg maxC x y (t) − C x y (v) (2.2)
t W
v=− W

2
I refer to the filtered peak value as entity of the peak. In this way, it is possible to
distinguish between peaks and troughs by taking into account the sign. A negative
peak is referred to an inhibitory link (Fig. 2.2b), while conversely a positive peak
is referred to an excitatory link (Fig. 2.2d). I also implemented and applied a post
computation filtering procedure to improve the detectability of inhibitory links on
noisy spike trains.
2.1.3.1 Post Computation FNCCH Filtering
The temporal occurrence of the extracted peak of the NCCH represents the time delay
of the identified connection between two neurons. When we deal with experimental
recordings, the NCCH could be very jagged, with a shape characterized by oscilla-
tions in the central region of the correlation window and a typical decrease of the
NCCH values in the borders. The FNCCH computation procedure permits to extract
the peaks evaluating their sign, but it is sensitive to a decrease in synchrony due to
uncorrelated activity. Such a decrease appears at the boundaries of the correlation
window (Fig. 2.3, black curves), and can be exchanged for a decrease in synchrony
related to an inhibitory connection. Thus, we can expect that in specific cases char-
acterized by a very jagged correlogram (e.g., due to a low firing rate), this procedure
Fig. 2.3 FNCCH post filtering procedure. In this illustrative case, correspondent to weak correla-
tion, the filtering procedure infers a negative value in the boundary region of the correlation window
(black line) leading to a false positive inhibitory link. To avoid this, heuristic post filtering procedure
is formed by a peak search re-applied in a smaller region of the correlation window (green line)
discarding part of the tail. The resulting peak, in this example, is excitatory and with a shorter delay.
Reproduced from [4]
will introduce some mistakes in the inhibitory connections detection. I defined this
artifact as tail effect on the FNCCH. For this reason, I implemented a post FNCCH
filtering operation that allows to account for the presence of these artifacts removing
them from the set of identified inhibitory connections. More specifically, the filtering
procedure (Fig. 2.3) consists in few steps applied to every negative FNCCH values
falling in one of the two boundary regions of the correlation window (defined as the
15% of half the amplitude of the correlation window, Fig. 2.3 black curves). If such
detected inhibitory link is an artifact due to the tail effect, the FNCCH values in the
boundary region will be all negative and likely with a decreasing trend. In this case,
the corresponding putative peak is discarded and a new search starts for another peak
excluding the boundary regions. We then search for a peak in the central region of
the correlation window that I called “peak re-computation window” [Fig. 2.3, green
curve, by applying Eq. (2.1)]. Indeed, the “new” peak can be related either to an
inhibitory effective connection or to an excitatory one. On the other hand, if we find
an increase in the FNCCH reaching a positive value before the end of the correlation
window, we can state that the negative peak is correctly identified (i.e., an actual
minimum).
2.1.4 Partial Correlation
Cross-correlation based methods are not able to distinguish between direct and indi-
rect connections. In order to overcome this limitation it was introduced the notion
of partial coherence [5] in which the effects of the activity of all other spike trains
(assumed to be additive) could be removed. Eichler [6], instead, presented a partial-
ization method in the time domain, based on a scaled version of the partial covariance
density known as Scaled Partial Covariance Density (SPCD). The SPCD combines
the advantages of the cross-correlation histograms and the partialization analysis
in the frequency domain: (i) it interprets, in the same way of the cross-correlation
histograms, peaks and troughs as excitatory and inhibitory connections respectively;
(ii) it allows to discriminate direct and indirect connections and common inputs (see
Fig. 2.4).
In general, let us consider a neuronal population V and two specific neurons x,
y ∈ V. Let Rx y (τ ) be their Standard Correlation, and Rx x (τ ) and R yy (τ ) the Auto-
Correlation of x and y respectively [7]. The Fourier Transform of Rx y (τ ), i.e. the
cross-spectral density Sx y (ω), will define the Spectral Coherence C x y (ω) as follows:
Sx y (ω)
C x y (ω) = (2.3)
Sx x (ω)S yy (ω)
where Sx x (ω) and S yy (ω) are the Fourier Transform of Rx x (τ ) and R yy (τ ),

respectively.
Fig. 2.4 Schematic

representation of the
partialization process. CM
built using cross-correlation
(b) and partial correlation
(c) for a simple network
made up of 3 neurons,
a correspondent sketch of
connections. PC (differently
from CC) does not detect the
indirect link between neuron
1 and neuron 3 (yellow
circle)
The partialization process [5] removes from Sx y (ω) the effect Z related to all other
(possibly multivariate) spike trains of the population V in the following way:
Z = V − [x, y] (2.4)
Sx y|Z (ω) = Sx y (ω) − Sx Z (ω)S Z−1Z (ω)S Z y (ω) (2.5)
where S Z Z (ω) is the auto-correlation of Z in the frequency domain and the

Inverse Fourier transform of Sxy|Z (ω), Rx y|Z (t), is the partial covariance density.
Basically, the partialization process consists in iterating the cross-spectral power
density computation by removing a node per time, subtracting the new computed
value from the cross-spectral power density correspondent to the couple (x, y). If
there is the influence of a third node, that is, if the investigated connection is likely
to be indirect, such an operation will significantly modify the cross-spectrum value.
C x y|Z (ω) corresponds to the partial coherence function. It can be defined by the
inversion of the spectral matrix S(ω) of the whole set of nodes [6, 8]. If G(ω) =
S(ω)−1 , then it can be mathematically proved, that the partial auto spectrum densities
correspond to:
1
Sx x|V \{x} (ω) = (2.6)
G x x (ω)
1
S yy|V \{y} (ω) = (2.7)
G yy (ω)
Moreover, it can be proved that the partial coherence function can be expressed
in function of G(ω) as:
G x y (ω)
C x y|Z (ω) = − (2.8)
G x x (ω)G yy (ω)
We can use the partial coherence function to compute the partial power spectral
density Sxy|Z (ω):
Cxy|Z (ω)
Sxy|Z (ω) = 2 (2.9)
1 − Cxy|Z (ω)
And substituting Eqs. (2.6), (2.7) and (2.8) into Eq. (2.9), it is possible to obtain:
G x y (ω) C x y|Z (ω)

Sx y|Z (ω) = − 2 Sx x|V \{x} (ω)S yy|V \{y} (ω) (2.10)
G x x (ω)G yy (ω) 1 − C x y|Z (ω)
The partial covariance density Rx y|Z (t) corresponds to the Fourier inverse trans-
form of the partial power spectral density. The partial correlation function (or SPCD)
that I used in my Ph.D.’s work is a scaled version defined of Rx y|Z (t) defined as:
Rx y|Z (t)
sx y|Z (t) = √ (2.11)
rx r y
where r x and r y are the maximum peak values of the autocorrelation function.
Finally, PC function, as well as CC, permits to recognize the directionality of the
connections by observing the peak latency from zero.
2.2 Information-Theory Based Connectivity Methods
2.2.1 Transfer Entropy
Transfer Entropy (TE) is an asymmetric directed information theoretic measure

which allows to extract causal relationships from time series [9], estimating the
part of the activity of one single neuron which does not depend only on own past,
but also by neural past activity of another cell. In other words, TE measures the flow
of information between the activity of two cells. Let us consider a reference spike
train y, and a target spike train x. The mathematical definition of transfer entropy is
given by Eq. (2.12):
p(xt+1 |xt , yt+1 )
T E y≥x = p(xt+1 , xt , yt+1 ) log (2.12)
t
p(xt+1 , xt )
Fig. 2.5 Schematic representation of the Transfer entropy computation process. a Recorded in vitro
neural networks or in silico model from which binned spike trains (b) are extracted. Considering
the reference neuron y and the target neuron x (b, bottom), Eq. (2.12) is applied by extracting all
the possible patterns (c) for the binary trains, obtaining the TE’s values (d)
where x t+1 is the bin representing the present activity of the reference train, while
x t and yt are the past activity of the two trains. See Fig. 2.5 for a schematic description
of the computation process.
2.2.2 Delayed Transfer Entropy
When two neurons are activity-dependent and connected in a causal dependent way,
the consequent information flow usually appears in the correspondent spike trains
with a specific delay, different for each couple of neurons. If TE’s bin is not large
enough, it is highly probable to miss the detection of such effective connection. On
the other side, if the bin is too large, more spikes collapse into a single bin causing
a likely decreasing in the estimated amount of activity dependence. A possibility to
overcome these limitations is to temporally extend TE [10, 11]. The Delayed Transfer
Entropy (DTE) is defined according to Eq. (2.13):
p(xt+1 ||xt , yt+1−d )
DT E y≥x (d) = p(xt+1 , xt , yt+1−d ) log (2.13)
t p(xt+1 , xt )
In particular, we can start to consider the past bins for the target train at a specific
temporal distance [parameter d in Eq. (2.12)] before the present bin in the reference
train. Considering d varying from 1 to a fixed number, we can build a temporal
function TE(d). The estimated connection among the analyzed spike trains can be
extracted using the peak of the TE(d) function and/or the Coincidence Index (CI) [12],
i.e. the ratio of the integral of the function in a specified area around the maximum
peak to the integral of the total area.
2.2 Information-Theory Based Connectivity Methods 19
Fig. 2.6 Schematic representation of the DTE and DHOTE computation process. a The past bins
for the reference train are shifted of a quantity d, that is the parameter of the function DTE(d) (b).
If multiple bins are considered (blue squares in panel (c), the DHOTE is obtained)
2.2.3 Delayed High Order Transfer Entropy
It is possible to extend the TE to deal with multiple delays. In Eq. (2.12), the part
of past activity to take into account in the TE’s computation (number of bins) for
the reference train is indicated by k while l represents the number of bin to consider
for the target train. Thus, we can define the Delayed High Order Transfer Entropy
(DHOTE) according to Eq. (2.14):

p xt+1 |xt(k) , yt+1−d
(l)
D H O T E y≥x (d) = p xt+1 , xt(k) , yt+1−d

(l)
log (2.14)
t
p xt+1 , xt(k)
In this way, the couple (k, l) defines the TE’s order. For (k, l) different from the
couple (1, 1) we define the High Order Transfer Entropy (HOTE). The more is the
increase of the TE’s order (taking into account more bins in the two trains) the more
precise is the estimation of the information flow with a subsequent increasing of the
time required for the computation. See Fig. 2.6 for a schematic description of the
computation process.
2.2.4 Joint Entropy
The JE algorithm is based on the computation of the cross-Inter Spike Intervals

(cISI). It is based on the idea that if two spike trains are correlated, their activity will
show a specific degree of synchrony. Let us introduce a reference spike train x and
a target spike train y. We can define the cISI as:
cI S I = tx − t y (2.15)
Fig. 2.7 Schematic representation of the JE computation process. The cISIs between reference
and target trains are computed. Taking into account the occurrences of each cISI it is possible to
build the cISI histogram. The application of Eq. (2.16) allows to compute the JE between the two
investigated electrodes
where t x is the time stamp correspondent to a reference spike train, while ty is the first
time stamp temporally following the reference one in the target train. By iterating the
cISI computation for every reference spike, we can build the cISI histogram. Thus,
if cISIk corresponds to a cISI of size k bins, we can introduce the JE measure as:

J E(x, y) = − p(cI S I k ) ∗ log( p(cI S I k )) (2.16)
The higher the probability of two neurons being correlated in their activity, the
lower the JE value. In the ideal case of autocorrelation, JE will be equal to 0. The more
jagged is the cISI histogram, the less is the correlation of two spike trains, causing
the growing of the JE value, according to Eq. (2.16). In its original definition [13],
JE provides only a number for each couple of neurons, with no temporal information
of the delay of a functional connection.
To overcome such a limitation, I introduced and tested a method for extracting the
delay. Such procedure is simply based on extracting the most frequent cISI’s size in
bins (see Fig. 2.7). This value corresponds to the most likely delay of the investigated
functional connection.
2.3 Thresholding Approaches
The statistical measures described in the previous paragraphs enable building a

Connectivity Matrix (CM). The CM is a n ×n matrix (where n is the number of
analyzed nodes (i.e., electrodes or neurons)) whose generic element (i, j) is the esti-
mation of the strength of connection between electrodes i and j. Any connectivity
method provides a value for each couple of analyzed electrodes, even if such elec-
trodes do not share a real functional connection. As a consequence, the CM is a
2.3 Thresholding Approaches 21
Fig. 2.8 Hard Thresholding dependence on the parameter n. Connectivity graph of a cortex neural
network coupled to the MEA-4 k acquisition system, correspondent to: 1 (a), 2 (b), 3 (c)
full matrix of n [14] elements, and a thresholding procedure is required to throw

away those values that are not significant since close to the noise level, and then not
representing real connections.
2.3.1 Hard Threshold
The simplest thresholding procedure is to apply a hard threshold [15] to the CM. One
of the simplest choices is to use a threshold equal to μ + n * σ, where n is an integer,
while μ and σ are the mean value and the standard deviation of the connectivity matrix
elements, respectively [16]. Such a procedure is based on the assumption that the
strongest a link, the most likely it corresponds to a true functional link. The main limit
of the hard threshold approach is the choice of the parameter n, that is completely
arbitrary. A little variation on the threshold (i.e., the value of n) can drastically alter
the results of a functional connectivity analysis, providing different graphs with
different topological parameters (see Fig. 2.8). Even if the basic assumption on the
links strength can be correct, the problem of defining how strong is enough, could
still affect the obtained results with a bias dependent on the choice of the parameter
n in the thresholding procedure.
2.3.2 Shuffling Approach
Another possible solution to threshold the connectivity matrix is the shuffling thresh-
olding approach [17]. This method is based on the generation of spike trains surrogate
data. A surrogate data is generated from a spike train by preserving all the statistical
features except for the one that we want to test for the significance (i.e., the timing
of the spikes). In fact, the correlation between two spike trains is embedded in the
temporal positions and synchrony of the spikes. If we destroy such information, by
randomly changing the positions of the target spikes, we can build a null hypoth-
esis, that is correspondent to a non-likely connection. I will describe the shuffling
approach that I implemented and tailored for the JE algorithm (see Fig. 2.9).
In detail, after generating surrogate data (see Fig. 2.10), it is possible to test
the significance of a found functional connection between spike trains x and y, by
applying the JE method to the pair (x, shuffled y) for every surrogate target train.
Such JE value is correspondent to a non-connection, thus, we can test if the JE value
correspondent to the pair (x, y) is lower than the shuffled one, indicating a likely
significant functional connection. If we generate hundreds or thousands of shuffled
trains, we can repeat such test a large number of times. Finally, we can define a
Fig. 2.9 Schematic

representation of the
customized thresholding
approach based on surrogate Shuffling test with 10 surrogate
data generation train per target
Distribuon Dist of the connecons

that passes the thresholding
procedure
JE(x,y) < JE (x, yshuffled)
Delta = mean(Dist) + n*std(Dist), n

integer
Shuffling test with the desired

number of surrogate train per
target. The thresholding procedure:
(JE(x,y) < JE (x, yshuffled) and
abs(JE(x,y) - JE (x, yshuffled)>delta
2.3 Thresholding Approaches 23
Fig. 2.10 Schematic representation of the shuffling process. If we consider two strongly correlated
spike trains x and y as in panel (a), they will provide high values for every applied connectivity
method. As an example, panel b reports the computed DTE’s values. If we destroy the correlation,
just randomly changing the position of the spike in the neuron y (red squares in all the panels), the
detected values decrease (panel (b), bottom). The idea behind the shuffling process is to destroy
correlation randomly changing the position of the spikes, as reported in panel (c) (green squares
represent the shuffled spikes). It is possible to shuffle the reference, the target or both the spike trains,
comparing the obtained values with the ones corresponding to the original trains. As an example,
panel d shows how the NCCH correspondent to original trains (blue line) and shuffled target train
(red line). We can notice that the shuffled peak is lower than the original one, suggesting that the
investigated functional connection is significant. Otherwise, if the shuffled peak is higher or equal
to the original one, we can assess that the functional connection is not significant, discarding it
minimum p-value for accepting as significant a functional connection, and apply the
thresholding procedure to the entire dataset.
In my work, I introduced a parameter that I called delta, that is a minimum differ-
ence between the value representing the null hypothesis and the original statistical
one, in order to increase the precision reducing the number of false positives. In
detail, I implemented a bootstrap shuffling analysis to extract the parameter delta.
Thus, I start a pre-processing analysis performing the shuffling test with 10 shuffled
spike trains for each electrode. I compute the distribution of the differences between
the original JE value and the shuffled one, among all the statistically significant esti-
mated connections, that is, considering those differences for which the original JE
value is lower than the shuffled one. At this stage, I define the delta value as an integer
multiple of the mean value of such distribution (see Fig. 2.9).
As we can see in the results, such a parameter hardly influences the performances.
In the literature, several methods to generate surrogate data from spike trains can be
found. In my work, I followed the spike time dithering approach [17]. The spike time
dithering procedure consists in randomly shuffling each spike inside a window of
specified width. Thus, each spike at time t is randomly substituted by a spike at a

time [t-w, t + w], where w is the shuffling window’s width.
2.4 Metrics to Evaluate the Connectivity Methods’

Performances
After the thresholding procedure, any connectivity method works as a binary clas-
sifier. In fact, the method classifies whether two electrodesi and j share a functional
link or not. It is evident that all the metrics used to evaluate a binary classifier can
be applied to quantify a connectivity method’s computational accuracy. For this
purpose, the testing of a connectivity method requires the use of a neural network
computational model, and the consequent comparison between prediction and obser-
vation. Using such computational models, we have the prediction represented by the
computed thresholded connectivity matrix and the observation represented by the
model’s Synaptic Weight Matrix (SWM). The higher the correspondence between
prediction and observation, the higher the computational accuracy of a method.
2.4.1 Receiver Operating Characteristic (ROC) Curve
The ROC curve [18] is a common metrics used to evaluate the performances of a
binary classifier by comparing prediction and observation. In my PhD study, the
prediction is represented by the computed Thresholded Connectivity Matrix (TCM),
while the observation corresponds to the SWM of the neural network model (i.e., the
ground truth).
We can define the True Positive Rate (TPR) and the False Positive Rate (FPR) as
follows:
TP
T PR = (2.17)
T P + FN
FP
FPR = (2.18)
FP + T N
The ROC curve (Fig. 2.11a) is then obtained by plotting TPR versus FPR. The
Area Under Curve (AUC) is a main parameter extracted to have a single number
describing the performances of a binary classifier: a random guess will correspond
to 0.5 (straight line in Fig. 2.11a), while a perfect classifier will have a value of 1.
Another important metric that can be extracted from a ROC analysis is the accuracy,
defined as:
TP +TN
ACC = (2.19)
T P + T N + FP + FN
2.4 Metrics to Evaluate the Connectivity Methods’ Performances 25
Fig. 2.11 Metrics to evaluate a binary classifier. a Example of ROC curve. The random guess
corresponds to the ROC curve being a line coincident with the bisector. b Example of MCC curve
2.4.2 Matthews Correlation Coefficient (MCC) Curve
The MCC curve [18] is a common metrics, alternative to the ROC analysis, used
to evaluate the performances of a binary classifier by comparing prediction and
observation. Using the quantities defined in the previous paragraph, changing the
threshold used to compute the TCM, we can define the MCC as:
T P ∗ T N − FP ∗ FN
MCC = √ (2.20)
(T P + F P)(T P + F N )(T N + F P)(T N + F N )
The MCC assumes values in the interval [−1,1] and the MCC curve is obtained
by plotting the MCC value versus the false positive rate (Fig. 2.11b).
2.5 Graph Theory
Graphs are made up of nodes which represent the neurons and edges which model
the connections (morphological or functional) among the neurons. If we consider
the directionality of the connection (i.e., from a pre- to a post-synaptic neuron), the
graph is named directed, otherwise it is called undirected. The structure of the graph
is described by the adjacency matrix (often named connectivity matrix (CM)), a
square symmetric matrix of size equal to the number of nodes N with binary entries.
If the element aij = 1, a connection between the node j to i is present, otherwise aij
= 0 means the absence of connections. To allow a mathematical analysis, the graph,
and consequently the network topology, can be characterized by a large variety of
parameters [19]. In the field of neuronal networks, the simplest metrics which allow
to have a simple but clear indication of the kind of underling connectivity are the
Node Degree, the Cluster Coefficient and the Average Path Length [20] which will
be briefly described below.
Node Degree: the in-degree (id) and the out-degree (od) of a single node
are defined as the number of incoming (afferent) and outcoming (efferent) edges
respectively, and the total degree (td) is their sum.
td = id + od (2.21)
High in-degree values indicate neural units influenced by a larger number of nodes,
while high out-degree values show a large number of dynamic sources. Depending
on the node degree distribution, we can identify three stereotyped graphs: scale-free,
regular, and random (Fig. 2.12).
Scale-free networks (Fig. 2.12b, panel a) [21] are characterized by high-connected
units called hubs. Hubs are nodes with a degree at least one standard deviation above
Fig. 2.12 Basic graph measures and network structures. a Node degree is the number of connections
of a given node; this panel shows a simple network divided in four different modules: Module 1, in
which we can see a high-connected unit called hub, and Module 2, that presents a low connectivity
case. Modules 3 and 4 show two units with high and low values of Cluster Coefficient respectively,
and an example of shortest path length; the nodes X and Y are connected by the shortest possible
path (three links), and two different units that we call intermediaries. b Classification of the network
structure [scale-free (panel a), regular (panel b), small-world (panel c) and random (panel d)] and
corresponding degree distributions. Reproduced from [15]
2.5 Graph Theory 27
the network mean. Thanks to this peculiarity, hubs play a significant role on the
neural dynamics [22]. In the scale-free network, the probability that a generic node
i has k connections is given by a power law relationship:
p(k) ∝ k −ϒ (2.22)
where U is the characteristic exponent which ranges experimentally from 1.3 (slice
recordings [23]) to 2 (fMRI recordings [24]).
Regular networks (Fig. 2.12b, panel b) are ordered and characterized by high
segregation values. The integration levels of the network grow by increasing the
number of graph units. In this case, the probability that i has k connections is given
by:
p(k) = c (2.23)
where c is a constant.
Random networks (Fig. 2.12b, panel d) show each node with a different connec-
tivity degree and the probability that a single unit has k connections is modeled by a
Poisson distribution:
e−δ δ k
p(k) ∝ (2.24)
k!
where δ is the average connectivity degree of the network. The random graph has
few local connections and therefore it shows low segregation values. The integration
levels of the network, instead, follow the logarithm of the number of nodes.
A last case is the small-world network (Fig. 2.12b, panel c): it shares the same
characteristics of regular and random networks, constituting a sort of composite
model. By increasing the probability p of rewiring, the order of a regular lattice is
disrupted, and when p = 1 a random graph is generated. Increasing the probability of
rewiring, both the integration and the segregation levels decrease. In a small-world
network, the distance between two nodes grows according to the logarithm of the
number of nodes of the graph [25].
Cluster Coefficient (CC)
Let x be a generic node and vx the total number of nodes adjacent to x (and including
x). Let u be the total number of edges that actually exist between x and its neighbors.
The maximum number of edges that can exist among all units within the neighbor-
hood of x is given by vx (vx − 1)/2. The Cluster Coefficient (Cx ) for the node x, is
defined as follows
2∗u
Cx = (2.25)
vx (vx − 1)
The Average Cluster Coefficient, obtained by averaging the cluster coefficient of

all the network’s nodes, is a global metric often used to quantify the segregation at
network level (Fig. 2.12a).
Average Shortest Path Length (PL)
Let x and y be two generic nodes of a network V. Let d (x, y) be the shortest distance
between the nodes x and y. We define the Average Path Length (L) as follows:
2
PL = d(x, y) (2.26)
n(n − 1) x= y
This topological parameter is commonly used to evaluate the network’s level of

integration (Fig. 2.12a).
Small World Index (SWI)
To detect the emergence of small-world network [26], it is possible to combine the
metrics previously introduced, defining the Small-World Index (SWI):
Cnet
Cr nd
SW I = L net
(2.27)
L r nd
where C net and L net are the cluster coefficient and the path length of the investigated
network; while C rnd and L rnd correspond to the cluster coefficient and the path length
of random networks equivalent to the original network (i.e., with the same number
of nodes and links). A SWI higher than 1 suggests the emergence of a small-world
topology.
Rich Club Coefficient
The Rich-Club Coefficient (RCC) [27] Φ at a specific k level is computed by
evaluating the cluster coefficient between the nodes with a degree higher than k:
2E >k
(k) = (2.28)
N>k (N>k − 1)
where E >k and N>k represent the number of edges and nodes with a degree higher
than k, respectively. The RCC curve is obtained by evaluating the RCC at the varying
of k from 1 to the maximum degree. The RCC is normalized by the corresponding
average value for a set of surrogates random neural networks equivalent to the inves-
tigated one (i.e., networks with the same number of nodes and edges). A privileged
sub-network (i.e., a rich club) emerges, if the computed normalized coefficient value
is higher than one.
2.6 BIOCAM and Active Pixel Sensor (APS) Array 29
2.6 BIOCAM and Active Pixel Sensor (APS) Array
During my Ph.D., I studied neural cultures coupled to both low resolution and high
resolution acquisition system. Briefly, the low resolution acquisition system is repre-
sented by the MEA60 and MEA2100 by Multi-Channel System (www.multichannel
systems.com; MCS, Reutlingen, Germany) 60 electrodes device, while the 3Brain
APS (www.3brain.com; Landquart, Switzerland) was used as a high resolution acqu-
sition system. In this section, I will briefly describe the BIOCAM X and the APS
MEA-4 k chip. The 3Brain APS chip is a platform based on monolithic Comple-
mentary Metal Oxide Semiconductor technology enabling acquisitions from 4096
microelectrodes at a full frame rate of 9043 kHz and at a spatial resolution up to
21 μm (electrode separation) [28]. Based on the Active Pixel Sensor concept (APS),
the platform development, electrical characterization and preliminary validation was
performed on cardiac tissue. To introduce the platform, Fig. 2.13 presents a schematic
description.
Briefly, the two essential elements are: (i) the metallic microelectrode array imple-
mented similarly to a light imager, with in-pixel integrated microelectrodes and low-
noise amplifiers; and (ii) a real-time acquisition and processing board. Based on
image/video concepts implemented in hardware, the signal and data processing is
furthermore adapted for handling a very large data flow (typically 0.5 Gbit/s). The
platform enables acquisitions with a spatial resolution comparable to mammalian
neuronal cell bodies (i.e., microelectrode size of 21 μm, electrodes separation
from 21 to 81 μm), and a temporal resolution down to 8 ms/pixel on 64 selected
pixelmicroelectrodes. The 3Brain system allows accessing in real-time the spatial–
temporal correlation of activities at both local and network levels acquired from 4096
microelectrodes.
The BIOCAM (Fig. 2.14) is the hardware tool that allows to sample the 4096
recording electrodes simultaneously at 18 kHz, providing orders of magnitude more
data points compared to conventional passive MEAs), thus leading to the recording,
Fig. 2.13 Overview of the

high resolution
electrophysiological
platform. The system is
composed of three hardware
levels, i.e. the CMOS–MEA
chip, the interface board and
a work station equipped with
a frame grabber for capturing
and storing the video stream.
Adapted from www.3brain.
com
Fig. 2.14 3Brain BIOCAM. Reproduced from www.3brain.com
over a large bandwidth, of electrophysiological signals ranging from slow large-

population field potentials to fast single-cell spiking activity. In addition, it incor-
porates further optional functionalities, which come as separate modules in most
MEA-systems, such as a temperature control system and an electrical programmable
current-driven stimulator.
References
2. Eytan D, Minerbi A, Ziv N, Marom S (2004) Dopamine-induced dispersion of correlations
between action potentials in networks of cortical neurons. J Neurophysiol 92:1817–1824
3. Wagenaar DA, Pine J, Potter SM (2006) An extremely rich repertoire of bursting patterns
during the development of cortical cultures. BMC Neurosci 7:11
4. Pastore VP, Massobrio P, Godjoski A, Martinoia S (2018) Excitatory-inhibitory links and
network topology in large scale neuronal assemblies from multi-electrode recordings. PLoS
Comput Biol 14(8):e1006381. https://doi.org/10.1371/journal.pcbi.100638
5. Brillinger DR, Bryant HL, Segundo JP (1976) Identification of synaptic interactions. Biol
Cybern 22:213–228. https://doi.org/10.1007/bf00365087
7. Papoulis A, Pillai SU (2002) Probability, random variables and stochastic processes, 4th edn.
Mc Graw Hill, New York
8. Dahlhaus R, Eichler M, Sandkuhler J (1997) Identification of synaptic connections in neural
ensembles by graphical models. J Neurosci Methods 77:93–107
9. Schreiber T (2000) Measuring information transfer. Phys Rev Lett 85:461–464
10. Bennett K, Robertson J (2011) MATLAB—A ubiquitous tool for the practical engineer. In:
Ionescu CM (ed)
11. Overbey LA, Todd MD (2009) Dynamic system change detection using a modification of the
transfer entropy. J Sound Vib 322:438–453. https://doi.org/10.1016/j.jsv.2008.11.025
References 31
12. Berdondini L et al (2006) A microelectrode array (MEA) integrated with clustering structures
for investigating in vitro neurodynamics in confined interconnected sub-populations of neurons.
Sens Actuat B Chem 114:530–541
14. Taketani M, Baudry M (2006) Advances in network electrophysiology: using multi-electrode
arrays. Springer, New York
Front Neural Circ 9:57. https://doi.org/10.3389/fncir.2015.00057
16. Nigam S et al (2016) Rich-club organization in effective connectivity among cortical neurons.
J Neurosci 36:670–684. https://doi.org/10.1523/JNEUROSCI.2177-15.2016
17. Grun S, Rotter S (2010) Series in computational neuroscience. Springer, p 444
18. Fawcett T (2006) An introduction to ROC analysis. Pattern Recogn Lett 27:861–874
19. Rubinov M, Sporns O (2010) Complex network measures of brain connectivity: uses
and interpretations. NeuroImage 52:1059–1069. https://doi.org/10.1016/j.neuroimage.2009.
10.003
20. Sporns O, Tononi G, Edelman GM (2000) Connectivity and complexity: the relationship
between neuroanatomy and brain dynamics. Neural Netw 13:909–922
21. Barabasi A-L, Bonabeau E (2003) Scale-free networks. Sci Am 50–59
22. Sporns O, Honey CJ, Kotter R (2007) Identification and classification of hubs in brain networks.
PLoS ONE 2:e1049. https://doi.org/10.1371/journal.pone.0001049
23. Bonifazi P et al (2009) GABAergic hub neurons orchestrate synchrony in developing
hippocampal networks. Science 326:1419–1424
24. Eguiluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV (2005) Scale-free brain
functional networks. Phys Rev Lett 94:018102
25. Watts DJ, Strogatz SH (1998) Collective dynamics of ‘small-world’ networks. Nature 393:440–
442
26. Downes JH et al (2012) Emergence of a small-world functional network in cultured neurons.
PLoS Comput Biol 8:e1002522. https://doi.org/10.1371/journal.pcbi.1002522
27. Schroeter MS, Charlesworth P, Kitzbichler MG, Paulsen O, Bullmore ET (2015) Emergence
of rich-club topology and coordinated dynamics in development of hippocampal functional
networks in vitro. J Neurosci 35:5459–5470. https://doi.org/10.1523/JNEUROSCI.4259-14.
2015
9:2644–2651
Chapter 3
Results
In the first section of the Results, I will consider each of the connectivity methods
that I developed and implemented during my Ph.D. In detail, I will at first show
the evaluation procedure’s results. Such a procedure exploits the use of an in silico
neural network model made up of 1000 randomly connected neurons, characterized
by an average ratio between inhibitory and excitatory connections of 1/4. Simulations
display the typical signature characterized by a mix of spiking and bursting activity,
as displayed by the raster plot and the Instantaneous Firing Rate (IFR) traces of
the excitatory (red) and inhibitory (blue) neuronal populations of Fig. 3.6a. From a
topological point of view, both the excitatory and inhibitory structural sub-networks
follow a random connectivity, as the degree distributions histograms of Fig. 3.6b
display. More details about the computational model can be found in the appendix
of this work. The evaluation procedure assesses the computational accuracy of the
connectivity methods considering them as binary classifiers, building ROC and MCC
curves. After evaluating the capability to detect the in silico functional connections
(e.g., looking at the ratio between true and false positive), I checked the capability
of the connectivity methods to reconstruct the topology of the model (looking at
topological parameters like the degree distribution, see section Graph theory) and
the delay distribution of the in silico functional connections. At first I will report
the results relative to the information-theory based methods: TE (and its extensions
DTE, DHOTE), JE, and then I will describe the correlation-based ones: FNCCH and
PC.
34 3 Results
3.1 Transfer Entropy
3.1.1 Validation of the TE by Means of in Silico Neural

Networks
I tested the developed DTE and DHOTE algorithms on 5 realizations of the in silico
neuronal network model described in the appendix. I randomly extracted 500 neurons
from such simulated neuronal populations, and I run the TE, the DTE and DHOTE
algorithms over one hour of simulation (sampled at 1 kHz).
First of all, I tested and evaluated our classic TE algorithm. To determine the
optimal bin value that maximizes the accuracy, I considered the TPR value corre-
sponding to a FPR = 0.01. Figure 3.1a shows the obtained results. The accuracy
displays a sharp increasing trend, reaching a maximum for a bin of 9 ms. Then,
in order to compare classic TE, DTE and DHOTE, I evaluated the accuracy of the
methods by means of the Receiver Operating Characteristic (ROC) curves [1]. Within
this framework, the True Positive Rate (TPR) and the False Positive Rate (FPR) are
defined on the basis of the true positive (TP), true negative (TN), false positive (FP)
and false negative (FN) values. First of all, I explored the variation of the accuracy
with respect to the order k and l for DTE. I considered a temporal window of 30 ms for
computing the DTE, corresponding to a classic TE (with bin equal to 1 ms) computed
for 30 different delays, as explained in section Delayed Transfer Entropy. Following
the approach used in [2], I took into account the TPR value corresponding to a very
low value of the FPR (0.01) performing the evaluation by varying the parameters
from order (1, 1) to order (3, 3). Figure 3.1b, c show in terms of false color maps the
results obtained considering the peak value (Fig. 3.1b) and maximum CI (Fig. 3.1c)
of DHOTE. The best accuracy performances in both cases are achieved by using k
= 1 and l = 3 (red asterisk). As expected, the computation time (Fig. 3.1d) increases
with the total order (k, l) although it remains reasonably low for all the explored
orders. For 60 min of simulation of 500 neurons, the maximum computation time is
27 min (54 s for a single TE computation) for order (3, 3). The computation time
required for the order (1, 1) is equal to 9 min (18 s for a single TE computation).
Finally, I compared the TE (using the optimum bin of 9 ms) with the DTE and the
DHOTE with k = 1 and l = 3 by means of the ROC curves. Figure 3.1e shows the
obtained results, while in Fig. 3.1f we can see the corresponding Area Under Curve
(AUC) values. There is a relevant improvement in the accuracy consequent to the
temporal extension of the TE method, while the extension of the TE’s order to k =
1 and l = 3 produces only little improvement with respect to TE (Fig. 3.1f). In the
light of the previous analysis and considerations, I used the DHOTE showing the
best performances (k = 1 and l = 3), to explore the capability of such a method to
reconstruct the temporal delay and the degree distribution of the in silico neuronal
network. Therefore, a hard threshold procedure to extract the strongest effective
links from the TE’s connectivity matrix was used. In particular, following the results
achieved in [3], I used a threshold equal to μ + σ, where μ and σ are the mean and the
standard deviation values of the TE’s connectivity matrix, respectively. Figure 3.1g,
3.1 Transfer Entropy 35
Fig. 3.1 TE’s testing on in silico neural networks. a TPR (corresponding to FPR = 0.01) for classic
TE computed by increasing the bin size from 1 to 30 ms. b, c, TPR corresponding to FPR = 0.01 for
increasing TE’s orders for reference train (k) and for target train (l) is considered. The red symbol
“*” indicates the (k, l) pair that guarantees the best accuracy. (B) Values computed using TE’s peak.
c Values computed using TE’s CI. d Computational times for increasing TE’s orders. e ROC curves
corresponding to TE (black), DTE (peak and CI, red and green, respectively) and DHOTE (peak
and CI, blue and gray, respectively); inset of panel (e) shows a zoom of the same curve. f AUC
derived from the ROC curves of panel e. g Delay distribution obtained using the DHOTE (k = 1,
l = 3) CI. The reconstructed delay distribution is coherent to the one used to generate the neural
model. h Degree distribution obtained using the DHOTE (k = 1, l = 3) CI. The reconstructed degree
distribution fits a Gaussian curve that is the distibution used to generate the underlying connectivity
of the in silico neuronal model (inset). Reproduced from [4]
36 3 Results
h display how DHOTE reconstructs both delay and degree distribution of the simu-
lated network. Figure 3.1g shows that the reconstructed delay distribution from the
functional-effective links is qualitatively similar to the uniform distribution of the
delays of the neuronal network model. Figure 3.1h shows that the estimated degree
distribution can be fitted with a Gaussian distribution (R2 = 0.89), in accordance
with the original distribution of the structural connectivity of the in silico neuronal
network (Fig. 3.1h inset).
3.1.2 DTE Application to Neural Networks Coupled

to the MEA-4k
In this section, I report a demonstrative example of the use of DTE to process and
analyze 30 min of spontaneous activity of a cortical network recorded by means of a
4096 HD-MEA (3Brain) at the sampling frequency of 9043 Hz. I computed the TE
on the extracted spike trains in order to characterize the topological features of the
network connectivity. I used a first order (k = 1, l = 1) DTE with a temporal window
of 20 ms and with a bin size equal to the sampling period (0.12 ms). Thus, to cover the
20 ms temporal window I computed 170 TE corresponding to 170 different delays.
Figure 3.2e shows the connectivity graph corresponding to the DTE analysis of an
illustrative cortical network. To provide some quantitative results, I performed some
topological analysis (Fig. 3.2f) by evaluating the CC, PL, SWI, and percentage of
hubs. The SWI is equal to 4.8 suggesting that the considered network exhibits small
world features; the degree distribution is plotted in a log–log plane in Fig. 3.2g.
3.2 Joint Entropy
3.2.1 Shuffling Thresholding Procedure Validation

on in Silico Networks
The thresholding approach based on surrogate data described in the section Shuffling
thresholding procedure, requires the choice of three main parameters. The first one
is the number of surrogates to generate for the statistical significance test (i.e., the
number of times the electrodes’ time stamps will be randomly displaced and shuf-
fled), while the second one is the parameter that I called delta, used to compute the
minimum difference between shuffled JE and original JE value to take into account
for rejecting the null hypothesis, considering a functional connection significant (see
section Shuffling approach). In detail, the minimum difference will be defined as
delta multiplied by the average value of a difference distribution built with a boot-
strap shuffling approach (see, Shuffling thresholding procedure). The last parameter
is the p-value, that represents the number of times over the total allowing the null
3.2 Joint Entropy 37
Fig. 3.2 Dynamic, connectivity and topological analysis of a cortical neural network coupled to
the 4096 electrodes 3Brain system. a ISI histogram (bin = 5 ms). b IBI histogram (bin = 50 ms).
c Burst duration histogram (bin = 100 ms). d Summary of the spiking and bursting statistics,
e Connectivity graph obtained by means of DTE thresholded with μ + 4σ to show only the strongest
links. f Summary of the topological parameters. g Degree distribution. Reproduced from [4]
hypothesis to be rejected before considering as not significant an investigated connec-

tion. For instance, a p-value of 0.01 indicates that if the original JE value is higher
than the shuffled one, more than 1 time over 100, I will discard the correspondent link
considering it as not significant. To validate the method, and tuning plausible values
for such parameters, I applied the JE algorithm to 10 realizations of the in silico
neural networks, described in detail in the appendix. I tested the shuffling approach
using 100 and 1000 surrogates for each neuron, sweeping the parameter delta from
0 to 3 with unitary step. I evaluated the shuffling performances in terms of precision,
evaluating the ratio between TP and FP. In detail, the precision is defined as TP/(TP
+ FP). The higher the precision, the higher the performances of a binary classifier.
In our case, the binary classifier is based on the JE algorithm, with different perfor-
mances in function of the parameter delta, the p-value and the number of surrogates
used. At first, I tested the classical shuffling approach, in which both the reference
and the target train are managed with random spike displacements to destroy an
38 3 Results
eventual correlation between spike trains building a null hypothesis correspondent

to non-connections. Using 100 surrogates, I obtained a precision value of 0.15 ±
0.04. Then, I tested our shuffling approach, in which only the target train is shuffled,
to be more restrictive obtaining a lower number of functional links with a higher
precision. Figure 3.3a–d show the average results obtained for 100 and 1000 surro-
gates. The 3-D graph correlates precision with the parameter delta and the p-value
used for the significance statistical test. As we can see, a delta value of 3 with a
p-value of 0.01 corresponds to a maximum value of precision for both the number of
surrogates. The generation of the surrogate data, and above all, the statistical tests,
need the execution of the JE algorithm for each couple reference electrode-shuffled
target electrode, thus require a computational time that is function of the number
of surrogates. According to this, considering that the developed algorithm has to be
used to analyze neural networks coupled to the HD-MEA acquisition system, with
4096 electrodes, the best choice seems to be the use of 100 surrogates with a p-value
of 0.01 and a delta value of 3. I use such parameters to analyze two cortex neural
networks coupled to the HD-MEA system (JE Application to cortex network coupled
to the HD-MEA).
3.2.2 Hard Threshold Validation on in Silico Neural

Networks
The shuffling approach in thresholding the connectivity matrix is aimed to provide

an efficient method, with statistical basis, to select the meaningful functional connec-
tivity links. However, the simplest approach remains the use of a hard threshold proce-
dure as described in Shuffling thresholding procedure section. The main problem with
such a thresholding approach, is the choice of the parameter n. As Fig. 3.3e shows, the
precision critically depends on the choice of n. In particular, the precision increases
at the increasing of n, with a consequent decreasing of the number of true positive
links. The hard threshold approach is based on the assumption that the strongest the
links (i.e., the lowest the obtained value for the JE), the most likely a found connec-
tion can be an actual one. To support such a hypothesis, I considered the JE values
correspondent to the functional links that survived the shuffling thresholding proce-
dure. As Fig. 3.3f shows, the JE values that survived to the shuffling thresholding
procedure, correspond to the lowest ones in the CM, and would have been selected
by the hard thresholding procedure, being a subset of the links surviving such a
procedure (red line n = 3, blue line n = 4). However, it is worth noticing that the
shuffling approach is independent from the user in the thresholding procedure. The
AUC value is equal to 0.6848 ± 0.033, that is higher than 0.5 (value correspondent
to random guess), but not descriptive of good classification performances. In fact,
when applying a threshold to select the most significant functional connections, we
work in an area of FPR rate that are very low. Thus, the TPR rate should be high in
this region to be representative of good classification performances. The CC method
Fig. 3.3 Shuffling thresholding procedure validation on in silico neural networks. a, b Average
values of precision for the shuffling approach at the varying of the parameters p_value and Delta
in a 3d graph for a number of surrogates = 1000 (a) and 100 (b). c, d Same information in a 2d
graph for a number of surrogates = 1000 (c) and 100 (d). e Average values of the precision and the
total number of detected links for the hard thresholding approach at the varying of n. f JE values
corresponding to the links surviving the shuffling thresholding procedure. These links are a subset
of the hard thresholding procedure survival links (with n = 3, red line and n = 4, blue line)
(blue curve) shows a behavior similar to the DTE one, correspondent to poor perfor-
mances in detecting and identifying the model functional connections. The MCC
curves (Fig. 3.4g) confirm the information provided by the ROC curves. In fact, the
JE (red curve) presents its maximum value (0.601 ± 0.005) in correspondence to a
FPR value of 0.01, indicating good classification performances. Both the DTE (black
curve) and the CCH (blue curve), instead, show negative values in correspondence
of FPR value of 0.01, representing a complete disagreement between prediction and
observation. The reason of such impressive differences between JE, CC and DTE has
to be searched in the inhibitory links detection and identification. In fact, it is worth
to notice that JE is not able, by definition, nor to detect or to identify inhibitory links.
Thus, the curve showed in Fig. 3.4c, actually, is relative only to the excitatory links.
The DTE, as well as the CC, are instead able to detect, by definition, inhibitory links,
but they are not able to identify it, that is, they are not able to distinguish excitatory
links from inhibitory links. It is worth to notice, that inhibitory neurons have a firing
rate higher than the excitatory ones (5–6 spikes/s versus 2 spikes/s in the model).
According to this, both the DTE and the CC detect strong functional links corre-
spondent to inhibitory links, due to artifacts and not to actual connections. To prove
such statement, that is to show that DTE and CC are hardly influenced and wasted
by the inhibitory neurons and links, I considered for the ROC curve analysis only
the excitatory neurons (the first 800). Figure 3.4d, h show the correspondent ROC
curve and MCC curves. As we can see, the DTE and the CC hardly improve their
40 3 Results
Fig. 3.4 JE validation on in silico neural networks. a Raster Plot and b mean Instantaneous Firing
Rate (IFR) representative of the simulated spiking activity. c ROC curves and MCC curves (d) rela-
tive to DTE (black curve), JE (red curve) and CCH (blue curve) considering both the excitatory and
the inhibitory neurons. Correspondent AUCs in the inset. When only the excitatory sub-networks
are considered, DTE and CCH severally improve their performances, as shown by the ROC curves
(g) and MCC curves (h). Using the shuffling approach, the degree (f) and the delay distribution (f,
inset) is reconstructed. Such distributions reflect the ones used to generate the in-silico networks
model (e)
performances, while the JE’s curve is almost the same. In the excitatory subnetwork,
the CC algorithm shows the best performances reaching a TPR value of 0.72 ± 0.01
in correspondence to the false positive rate of 0.01. The MCC curve confirms what
emerges from the ROC analysis, with the CC reaching a peak of 0.77 ± 0.02. The
DTE shows a similar behavior with a TPR value of 0.50 ± 0.01 in correspondence of
a FPR value of 0.01, and a MCC peak equal to 0.57 ± 0.01. The JE algorithm, instead,
has a TPR value of 0.580 ± 0.005 in correspondence of a FPR value of 0.01, and a
MCC peak equal to 0.616 ± 0.004. Later, I applied the shuffling approach with 100
surrogates and a delta value of 3 (see Shuffling validation on in silico networks) to
the JE connectivity matrices. I used the thresholded connectivity matrices to extract
the mean degree distribution (Fig. 3.4f). The computed average distribution fits a
Gaussian Curve (R2 = 0.92, black curve), that is the distribution used to generate
the model (Fig. 3.4e). Finally, I computed the delay distribution for the excitatory
links (Fig. 3.4f, inset). The obtained distribution reflects the uniform distribution in
[0, 20] ms used to generate the in silico model.
3.2.3 JE Application to Cortex Network Coupled

to the HD-MEA
Finally, I applied the developed JE algorithm to two cortical neural networks coupled
to the HD-MEA acquisition system, after they reached a stable stage (i.e., after
21 days in vitro). In particular, I analyzed two networks characterized by two different
seeding densities: a low density network (400 cell/mm2 ) and a high density one
(1000 cell/mm2 ). It is possible to see an example of the recorded spontaneous elec-
trophysiological activity in form of raster plot (Fig. 3.5a) and instantaneous firing
rate (Fig. 3.5b). I used our developed JE algorithm with the implemented customized
shuffling approach to infer the functional connectivity of the two neural networks.
Figure 3.5c represents an example of the thresholded connectivity graph (using the
shuffling thresholding approach with 100 surrogates, delta = 3 and a p-value of
0.01), where nodes and edges represent electrodes and functional links, respectively.
I performed a topological analysis (Fig. 3.5e) extracting the cluster coefficient and
the average path length [5]. The low density neural network showed a higher cluster
coefficient than the high density one (0.192 vs. 0.042). Such a difference is likely the
reflection of the cross-talking effect with high seeding densities in the in vitro model.
I extracted the Small World Index (SWI) by comparing the cluster coefficient and
the average path length of our networks with the average values extracted from 100
random networks equivalent to the investigated one (i.e., with the same number of
nodes and links, as done in [6]). Both the networks showed the emergence of a small
world topology (Fig. 3.5e). Such a topology is coherent with the degree distribution,
represented in Fig. 3.5d. Again, the difference in the absolute value of the SWI (10.57
vs. 2.15) is likely referred to the different cellular seeding densities.
42 3 Results
Fig. 3.5 JE application to cortex networks coupled to the HD-MEA acquisition system. a Raster
Plot and b mean Instantaneous Firing Rate (IFR) representative of the recorded spiking activity.
c Connectivity graph in false colors and correspondent degree distribution (d) obtained by means
of the JE algorithm and the customized shuffling approach. e Table reporting the main topological
parameters extracted for the two networks with different seeding densities (400 cell/mm2 and 1000
mm2 )
3.2.3.1 JE Results Discussion
The JE algorithm is characterized by both high computational accuracy and effi-

ciency. The computational efficiency is a result of the implementation I did, that
adapted the algorithm to be directly applied on the spike time stamps. The computa-
tional accuracy is linked to the selectivity of the JE algorithm for the excitatory links.
In fact, JE is not able, by definition, to neither detect nor to identify the inhibitory
links. The inhibitory neurons are usually characterized by high (and likely tonic)
firing rate, and as I showed in section JE Validation on in siliconeural networks,
they can be responsible for the appearance of false positive artifacts when using
other connectivity methods like DTE and CC. Moreover, the incapability to detect
inhibitory links could be very useful when combining the JE algorithm to another
one able to detect and identify inhibitory links. In this way, the JE could be used as a
sort of inhibition filter to improve the classification performances. Thresholding the
connectivity matrix is a main problem in the functional connectivity analysis, and can
dramatically alter the results influencing even the kind of topology that is detected [7].
The hard thresholding approach is hardly dependent on the choice of the threshold
value as I showed in section Hard threshold validation on in siliconeural networks. To

overcome this problem, I developed a customized version of the shuffling approach.
The basic thresholding approach corresponded to low values of precision, due to the
presence of a high number of false positives; thus, I introduced another threshold
that is dependent on the shuffling values distribution, and so it is dependent on the
network firing rate and dynamic parameters. Such thresholding approach brought to
very high level of precision when tested on in silico networks.
3.3 FNCCH Results
3.3.1 Validation of the FNCCH by Means of in Silico Neural

Networks
I applied the FNCCH to 10 realizations of the in silico neural networks. Figure 3.6c
shows the ROC curves obtained by comparing the Synaptic Weight Matrix (SWM) of
the model (i.e., the ground truth) with the computed Functional Connectivity Matrix
(FCM). Figure 3.6d shows the MCC curve (cf., Methods). The ROC curve relative
to the detection of inhibitory connections (blue curve in Fig. 3.6c) is very close to
the perfect classifier, with an Area Under Curve (AUC) of 0.98 ± 0.01(blue bar in
the inset of Fig. 3.6c). The MCC curve relative to the inhibitory links (blue curve in
Fig. 3.6d) has a maximum value of 0.87 ± 0.04. Then, I compared the sensitivity
of the FNCCH for the detection of excitatory links (red curves in Fig. 3.6c, d) to
the standard CCH’s one (for excitation, black curves in Fig. 3.6c, d), to underline
the improved detection capabilities with the filtering procedure. I observed not only
a significant (p < 0.001) AUC increase (0.92 ± 0.01 vs. 0.72 ± 0.02, Fig. 3.6c
inset), but also significant improvements in both ROC and MCC curves shape for
low values of false positive rates (FPR). In particular, we can notice (Fig. 3.6d), that
the FNCCH excitatory curve has a maximum value of about 0.75 with respect to
the correspondent NCCH’s value (for the same false positive rate) that is negative
(suggesting a disagreement between prediction and observation). The above results
justify the use of a hard threshold procedure (cf., Methods) to select the strongest and
significant functional connections. The Thresholded Connectivity Matrix (TCM) is
thus directly computed from the FCM using a threshold equal to μ + 1 · σ, (mean plus
one standard deviation of the connections strength) for the inhibitory links, and μ + 2
· σ for the excitatory ones, obtaining estimated links with a very high level of accuracy
(cf. Methods): R2 = 0.99 for the inhibitory links and R2 = 0.94 for the excitatory ones.
To investigate whether the reconstructed functional connectivity network resembles
the one of the model, I calculated the excitatory and the inhibitory (Fig. 3.6e) links’
degree distribution from the TCM. The computed degree distributions fit a Gaussian
distribution (Fig. 3.6e, R2 = 0.99 for the inhibitory links and R2 = 0.98 for the
excitatory ones), in accordance to the original distributions used to generate the
structural (random) connectivity of the model (Fig. 3.6b). It can be noticed, that
44 3 Results
3.3 FNCCH Results 45
Fig. 3.6 Effective connectivity estimation from 10 in silico neural networks. a Raster Plot and
mean Instantaneous Firing Rate (IFR) representative of the simulated spiking activity. b Structural
degree distribution of the model (red curve for excitatory links and blue curve for the inhibitory
ones). c ROC curves relative to inhibitory (blue curve) and excitatory (red curve) links computed by
applying the FNCCH; the black curve, related to only excitatory links extracted with the standard
NCCH, has been superimposed for comparison. Corresponding AUCs are represented in the inset.
d MCC curves related to inhibitory and excitatory links computed by applying the FNCCH; the black
curve, related to only excitatory links extracted with the standard NCCH, has been superimposed for
comparison. e functional degree distribution reconstructed by means of FNCCH. The reconstructed
degree distributions fit the Gaussian distribution used to generate the structural connectivity of the
in silico model. f Box plot of the excitatory and inhibitory delay distribution obtained by means of
the FNCCH. Adapted from [9]
the mean values of the functional Gaussian distributions are slightly higher than the
structural ones due to the presence of false positives. Finally, I computed the delay
distribution for both the excitatory and the inhibitory links of the TCM (Fig. 3.6f);
both the extracted distributions reflected the ones used to generate the model. In fact,
the excitatory delay distribution was uniform in the interval [0, 20] ms, while all the
inhibitory links introduce a constant delay set at 5 ms (cf., Methods).
3.3.2 Functional Connectivity and Emergent Network

Topologies in In Vitro Large-Scale Neural Networks
The FNCCH was applied to neuronal networks coupled to two different devices:
MEA-60 and MEA-4k. Figure 3.7 shows the two utilized micro transducers (Fig. 3.7a,
d) and the relative images of representative networks coupled to the two devices
(Fig. 3.7b, e). Such networks are the morphological substrate originating the complex
electrophysiological activity characterized by an extensive bursting dynamics (i.e.,
highly synchronized network bursts) and random spiking activity. Figure 3.7c, f
show two examples of spontaneous activity recorded by a MEA-60 (Fig. 3.7c) and a
MEA-4k (Fig. 3.7f). We can observe silent periods, (where very few activity appears),
desynchronized spiking activity and huge bumps of activity (of different duration)
called network bursts which cause a fast increasing of the firing rate as the mean
Instantaneous Firing Rate (IFR) plots show (Fig. 3.7c, f). I analyzed three cortical and
three striatal networks coupled to the MEA-60 (FNCCH parameters: time window W
= 25 ms and time bin of 0.1 ms) and three cortical networks coupled to the MEA-4k
(FNCCH parameters: time windows W = 24 ms and time bin of 0.12 ms) after they
reached a stable stage (i.e., after 21 Days In Vitro, 21 DIV).
Figure 3.8a, g show a directed graph relative to a cortical and a striatal network
coupled to a MEA60 device (Fig. 3.8b, c, h and i show the contribution of exci-
tation and inhibition respectively). All the graphs were obtained by applying the
hard threshold approach and a spatio-temporal filtering to prune co-activations (cf.
46 3 Results
Fig. 3.7 Micro-Electrode Arrays (MEAs) used in the presented experiments. a MEA-60 device,
b Cortical network coupled to a MEA-60. c Example of 100 s recording of spontaneous electro-
physiological activity and mean Instantaneous Firing Rate (IFR) plot. d MEA-4k device, e cortical
network coupled to the MEA-4k. f Example of 100 s recording of spontaneous electrophysiolog-
ical activity and mean IFR plot. Both the recordings come from cortical assemblies at DIV 25.
Reproduced from [9]
Methods). For the striatal culture, the qualitative prevalence of inhibitory connec-
tions is clearly visible. To characterize the detected links for the cortical cultures, I
computed the box plots of the functional connections’ delay (Fig. 3.8d) and connec-
tion length (Fig. 3.8e) of excitatory (red) and inhibitory (blue) connections. Interest-
ingly, I found that the inhibitory links are slower and with longer connections than the
excitatory ones, as previously reported in brain slices [8] for structural and functional
connectivity. Similar representative graphs are shown in Fig. 3.9, where a directed
graph relative to a cortical network coupled to a MEA-4k is presented (Fig. 3.9a).
Link strengths are represented by two distinct color codes (arbitrary unit) to facili-
tate the observation of excitation (hot-red color code) and inhibition (cold-blue color
code). The two detected sub-networks are also shown in Fig. 3.9b c. Moreover, box
plots showing the connectivity peak delays and lengths are presented in Fig. 3.9d, e.
The obtained results are in agreement with what obtained with MEA-60 devices.
I also computed the inhibitory links percentage with respect to the total number of
detected links for the three different experimental conditions (n = 3 experiments for
each condition). As expected, I found that striatal cultures have a higher percentage of
inhibition and inhibitory links (about 60%) [10, 11] than cortical ones (about 25%).
It is worth noticing that for the cortical cultures the excitatory/inhibitory ratio is
Fig. 3.8 Functional Connectivity analysis on different neural network populations coupled to MEA-
60 device. a Functional connectivity graph obtained by applying the FNCCH to a cortical network
at DIV 25. Excitatory and inhibitory links are separately thresholded and shown for convenience in
panel b (excitation, red color map) and c (inhibition, blue color map). Color scales are indicative of
the relative connection’s strength based on the peak of FNCCH. Yellow circles in panel b and cyan
circles in panel c represent the identified rich club nodes. d Box plot of the delays of the detected
functional links. e Box plot of the connection lengths of the detected links. f Mean percentage of
the inhibitory links revealed by the FNCCH at the varying of the recording time length. g Example
of functional connectivity graph relative to a striatal network at DIV 21 coupled to a MEA-60
device. Panels h, i show the excitatory (red color map) and inhibitory (blue color map) networks,
respectively. Reproduced from [9]
48 3 Results
Fig. 3.9 Functional Connectivity analysis on cortex neural networks populations coupled to the
MEA-4k. a Functional connectivity graph obtained by applying the FNCCH to a cortical network
at DIV 21 coupled to a MEA-4k device. Excitatory and inhibitory links are separately thresholded
and shown for convenience in panel b (excitation, red color map) and c (inhibition, blue color map).
Color scales are indicative of the relative connection’s strength based on the peak of FNCCH. Cyan
circles in panel (b), and pink circles in panel (c), represent the identified rich club sub-networks.
d Box plot of statistical distribution of the delays of the detected functional links. e Box plot of the
statistical distribution of the connection lengths of the detected links. f Percentage of the inhibitory
links revealed by the FNCCH at the varying of the recording time length. Adapted from [9]
detected quite independently from the number of recording sites (Fig. 3.8f) although
it tends to stabilize with a shorter recording time for the MEA-4k. Interestingly,
the found ratio, in cortical networks, between inhibitory and excitatory links (about
1/4) is about the same as the ratio of inhibitory and excitatory neurons estimated by
immunostaining [12].
In order to derive the topological features [13] of the analyzed cortical networks,
I computed the Clustering Coefficient, CC (Fig. 3.10a) and the average shortest
Path Length, PL (Fig. 3.10b). Then, I extracted the Small-World Index (SWI) by
comparing the CC and the PL of the analyzed networks to the mean values of CC
and PL of 100 realizations of a random network with the same degree-distribution,
as recently proposed [2]. I found that when the cortical networks are coupled to
MEA-4k devices we could see the emergence of a clear small-world (SW) topology
(Fig. 3.10c), while for cortical networks coupled to MEA-60 s we could not infer
Fig. 3.10 Topological features of the detected functional networks. a Mean Cluster Coefficient
(CC). b Average shortest Path Length (PL). c Small-World Index (SWI). Red and blue colors
indicate excitatory and inhibitory population, respectively. Degree Distributions of d excitatory,
e inhibitory, and total links (inset). f Schematic representation of the procedure used to decrease the
electrodes density to analyze the SWI dependence on the electrodes’ resolution. g SWI evaluation
as a function of the electrodes density from 60 to 4096 microelectrodes. Reproduced from [9]
any SW topology. With the measurements performed with MEA-4k’s I can state
that, both inhibitory and excitatory links with their small world index, SWI \gg 1
(9.2 ± 3.5 for the inhibitory links and 5.2 ± 2 for the excitatory ones) contribute
to ‘segregation’ with the emergence of small world networks. Moreover, inhibitory
links, with their somehow longer connections, might contribute also to network ‘inte-
gration’ (i.e., communication among the SWs). To further characterize the topology
of these neuronal assemblies, I also investigated the possible emergence of scale-free
topologies [14] by evaluating the presence of hubs [15] and power laws for the exci-
tatory (Fig. 3.10d), inhibitory (Fig. 3.10e) and global (Fig. 3.10e, inset) link degree
distributions. In agreement with what was obtained in previously published model
systems [16] and in other studies [8], I obtained that such distributions fit a power law
with a R2 higher than 0.92, in all the three cases. Finally, I searched for the presence
of privileged sub-networks, constituted by the most connected nodes (i.e., rich club)
of the investigated networks [17]. I identified the Rich Club Coefficient (RCC). For
the analyzed cortical cultures, I found privileged sub-networks as indicated by the
computed RCC with a max value of 2.7 ± 0.5. Figure 3.9b, c show the rich club
networks identified for one neural network coupled to the MEA-4k, represented by
means of blue circles (for excitatory subnetwork) and pink circles (for inhibitory
subnetwork). Figure 3.8b, c are the analogous for a cortical neural network coupled
the MEA-60 (yellow for the excitatory nodes and light blue for the inhibitory ones).
50 3 Results
When analyzing similar cortical networks but coupled to the MEA-60 devices,
as pointed out above, no SW topology is identified (Fig. 3.10c); they seem to be
characterized by a sub-random topology with a SWI of 0.4 ± 0.1 for the excitatory
and 0.2 ± 0.2 for the inhibitory links. These cortical networks are of the same type
as the ones coupled to the MEA-4k (i.e., similar density of neurons, same age, same
culture medium) and the apparent estimated random topology should be attributed to
the low number of recording sites (i.e., 60 channels) that are not enough to reliably
infer topological features. For determining how number and density of electrodes
are crucial, I computed the SWI by considering a reduced number of electrodes for
the functional connectivity analysis from the MEA-4k recording, as described in
Fig. 3.10f. In particular, I started from the full resolution of the MEA-4k (i.e., 4096
electrodes), and I progressively decreased the electrode density, down to 60 electrodes
(inter-electrode distance of 189 μm, electrode density of 19 electrode/mm2 ) to obtain
a situation comparable with the MEA-60 devices, as previously reported [18]. The
obtained results are shown in Fig. 3.10g: the SWI decreases down to a random
topology becoming variable and unstable when the number of considered electrodes
is less than 100. This last result is referred to the excitatory links and the same analysis
was not applied to the inhibitory connections. In fact, such inhibitory links are much
less than the excitatory ones, thus leading to an inhibitory topology reconstruction
that is strongly influenced by the decimation scheme applied to reduce the number
of electrodes.
3.4 FNCCH Results: Discussion and Observations
The computation of the correlation of firing activity in the framework of multiple

neural spike trains has been around since the 1960s. For over thirty years, cross-
correlation, its generalizations [19], and its homolog in the frequency domain [20]
have been the main tools to characterize interactions between neurons organizing into
functional groups, or “neuronal assemblies”. A common established technique was
to build a cross-correlogram (CCH), describing the firing probability of a neuron as
a function of time that elapsed after a spike occurred in the other one. Nevertheless,
in the literature, there has been no standard definition of the CC and the strength of a
connection can be estimated by different means. To make the correlation coefficient
independent of modulations in the firing rate, and to provide a basis for the evaluation
of the significance in correlation measurements, and in turn, to allow the interpre-
tation within an experiment and comparison between experiments, it is essential to
have procedures for correction, normalization and thresholding of the coincidence
counts obtained from cross-correlation calculations. Commonly used normalization
procedures are related to Normalized Cross-Correlation Histogram (NCCH) [21, 22],
event synchronization [23], Normalized Cross-Correlation (NCC—Pearson Coeffi-
cient) [24], Coincidence Index of the CCH [2]. Once that a Functional Connectivity
Matrix (FCM) is obtained, a thresholding procedure is necessary to discard those
values that are not related to putative real connections. All these approaches present
3.4 FNCCH Results: Discussion and Observations 51
advantages and disadvantages but none of them have been applied to reliably iden-
tify inhibitory connections on large-scale network from spiking activity. I introduced
a filtered and normalized CC based algorithm (i.e., FNCCH) from which thresh-
olded functional connectivity matrices for excitation and inhibition can be robustly
obtained.
From the analysis of the data, I identified both small-world and scale free topolo-
gies in cortical networks for the excitatory and inhibitory sub-populations. More
specifically, I extracted inhibitory subnetworks in cortical and striatal neuronal
cultures demonstrating the capability of the method and offering new under-
standing of neuronal interactions. Finally, the proposed algorithm, while confirming
already presented preliminary results in the literature, demonstrates a new way (i.e.,
through large-scale MEAs and CCH based analysis) to investigate network topology
and opens up new perspective for the analysis of multisite electrophysiological
recordings.
3.4.1 Identification of Inhibition
Generally, by inspecting the CCH we can notice an increase or a decrease of the

fluctuations [25]. In some studies it was noticed that the primary effect of inhibi-
tion on the cross-correlogram is a trough near the origin, and for this interaction
to be visible there must be present a background of postsynaptic spiking against
which the inhibitory effect may be exercised (high-tonic firing rate regime) [26,
27]. From some experimental works related to the analysis of connectivity from
cortical multi-site and multi-unit recordings [28], reporting the normalized strength
of the identified connections it appears that for excitation a good sensitivity is fairly
commonly obtained, while the situation is considerably worse for inhibition. This is
due to a low sensitivity of CCH for inhibition, especially under conditions of low
firing rates [29], as previously demonstrated [26]. The noted difference in sensitivity
may amount to an order of magnitude, and it was demonstrated that for inhibition
the magnitude of the departure relative to the flat background is equal to the strength
of the connection, whereas for excitation it involves an additional gain factor [26].
As a whole, the lack of efficiency in the detection of inhibition simply reflects the
disproportionate sensitivity of the analysis tool [30]. In our work, I introduced a
cross-correlogram filtering approach (FNCCH) that I developed to overcome the
inhibition detectability issue. As Fig. 3.6 shows, the FNCCH is able to detect with
very high accuracy the inhibitory links when applied to in silico neural networks
with similar dynamics with respect to actual ones. The filtering procedure improves
also the detectability of excitatory links resulting in a reshaping of the ROC Curve
(Fig. 3.6c) with an increase of both precision (MCC Curve, Fig. 3.6d) and AUC
with respect to the standard cross-correlation. However, being a cross-correlation
based method, the presented FNCCH has indeed some limitations in the inhibitory
links detection that I tried to investigate with the in silico models. The major factor
affecting the detectability of inhibition is the variability of the cross-correlogram.
52 3 Results
In order to reduce this variability, it is possible to increase the number of coin-

cidences per bin by increasing the bin-width (that is, down-sampling with loss of
information in the acquired electrophysiological data), or by increasing the number
of events involved (which can be obtained with high firing rate and/or by increasing
the recording time) [31]. Another influencing factor depends on the balance of exci-
tatory and inhibitory neuronal inputs (i.e., balanced model) and it is referred to the
relative strength between inhibitory and excitatory inputs. In fact, when the neuron
is not balanced, excitation is, on average, stronger than inhibition. Conversely, when
the neuron is balanced, both excitation and inhibition are strong and thus detection of
inhibitory links improves [25, 30, 32]. Starting from the in silico model, I was able to
investigate the impact of rates variability on excitation/inhibition detectability, and
try to define a reasonable threshold (criterion for detectability [25, 29]). In partic-
ular, I varied the firing rate of the inhibitory neurons from 20 to 2 spikes/s while
maintaining a firing rate of 2–3 spikes/s for the excitatory neurons. I found that the
detectability of functional inhibitory links is preserved with our method down to a
firing of about 6 spikes/s and then decreases significantly (data not shown). I also
investigated the inhibition identification with respect to the recording time. Starting
from 1 h of simulation, I reduced (10 min steps) the recording time and I found
that there is a decrease in the inhibition detectability below 30 min of recording (cf.
Fig. 3.9f). The obtained results enabled us to apply the FNCCH to in vitro large-scale
neural networks, and allowed us to infer topology and functional organization. The
described procedure can be also directly applied to Multi Unit Activity (MUA) from
in vivo multi-site measurement recordings.
3.4.2 The Emergence of a Scale Free and Small-World

Topology
The cortical networks probed with MEA-4k’s showed an unambiguous small-world

topology. The inhibitory functional links had a SWI equal to 9.2 ± 3.5, higher than
the value extracted from the excitatory links (5.1 ± 1.9). Conversely, the cortical
networks coupled to the MEA-60 showed a random organization topology (0.21 ±
0.212 for the inhibitory links and 0.38 ± 0.1 for the excitatory ones). The apparent
random organizations are due to the low number of recording site of the acquisition
system; in fact, it is worth to remember that the SWI is computed by comparing cluster
coefficient (CC) and average shortest path length (PL) of the analyzed networks to
the corresponding values for surrogate random equivalent networks (same number of
nodes and links). From the obtained results, contrarily with what has been recently
presented [17], I demonstrated that the emergence of small-worldness, and above
all, deviations from a random topology on a reduced set of electrodes (<100 in the
case of the MEA-60) cannot be reliably derived or observed in a neuronal population
probed by a reduced number of recording sites. Definitively, besides the crucial
importance of well-defined statistical tools used for the analysis, it is fundamental
to probe network activity by using large-scale microtransducer arrays (i.e., with at

least 200 electrodes). As a whole, the issue related to the low number of recording
sites should be carefully taken into account when extracting dynamical features as
well as organizational principles of complex networks.
Finally, it should be underlined that I limited and focused the work on the CC
based methods. I mentioned in the Introduction the widespread and continuously
increase use of Information Theory based techniques. Beside the relative novelties
of such methods and the good performances (for a review see [33] and references
therein), they showed high computational costs and, to our knowledge, the inability
to reliably estimate inhibitory connections [2]. Although theoretically, information
theory based methods, like Transfer Entropy (TE) and Mutual Information (MI) are,
in principle, able to detect inhibitory links, I am not aware of studies consistently
reporting a successful identification of inhibitory connections. The problem might
be more in the lack of possibility to distinguish between excitatory and inhibitory
links, rather than the detection of inhibition, as I will show in the next paragraph.
3.4.3 Comparison with a Transfer Entropy Based Algorithm
In this section, we compare FNCCH with Delayed Transfer Entropy (DTE) [4].
Figure 3.11a shows the ROC curve and the correspondent AUC (Fig. 3.11b) when
DTE is applied to an in silico network made up of 1000 neurons (cf., Methods).
The ROC curve relative to the total number of links (black curve) displays a shape
similar to the one correspondent to the NCCH (black line of Fig. 3.11c, main text).
However, if we analyze separately excitatory and inhibitory links, the results change
dramatically. If we restrict the detection only to the excitatory links, the DTE ROC
curve greatly improves (Fig. 3.11a), together with the AUC value (Fig. 3.11b). The
ROC curve related to the inhibitory links (Fig. 3.11a) shows a good TPR value higher
than 0.4 in correspondence of a FPR of 0.01. Finally, the ROC curve relative to the
Fig. 3.11 DTE effective connectivity estimation relative to an in silico neuronal network. Effective
links are estimated starting from the simulated multi-site electrophysiological activity. a ROC
curves relative to the total links (black), to the excitatory versus excitatory neurons’ links (red), to
the inhibitory versus excitatory neurons’ links (blue) and to the to the inhibitory versus inhibitory
(green). b Correspondent AUCs. c DTE weighted connectivity matrix. Reproduced from [9]
54 3 Results
links between inhibitory neurons (green curve) shows that only false positives are
detected, since in the model there are no links among inhibitory neurons. Moreover,
these false detected connections correspond to higher values with respect to all the
other detected functional links (as confirmed by the connectivity matrix, represented
with false color in Fig. 3.11c) with the consequence of worsening the total links
detection. Thus, it is indeed possible to “detect” inhibitory links with a TE based
method, but is not possible to “identify” (i.e., to distinguish) excitatory and inhibitory
links.
3.4.4 Partial Correlation
I applied PC to the in silico neural networks described in detail in the appendix.

I compared PC to TE and basic CC. In particular, I extracted subsets made up of
60, 120, and 240 neurons. Furthermore, I systematically tested how the aforemen-
tioned methods are able to account for different connectivity degrees. To this end,
we increased the average connectivity degree of each node of the network model by
considering low- (~20%), medium- (~50%) and high-connected (~100%) values.
3.4.5 Validation of the PC by Means of in Silico Neural

Networks
The performances of each connectivity method were evaluated by means of ROC

and ACC curves, averaged over 10 realizations (lasting 10 min) for each network
model. The ROC curves of our partialization procedure, observed considering low
(L) connected assemblies and low number of nodes (60 MNeu), were very close to
the ideal case (i.e. TPR values equal to 1); by increasing the complexity of the neural
networks from medium (M) to high (H) connectivity degree, PC maintained higher
performances than one-delay TE and CC (Fig. 3.12a, first row). PC showed very good
performances also in larger (120 and 240) assemblies at low and medium connectivity
degrees (Fig. 3.12a, second and third row), corresponding to reasonable physiological
conditions [12]. The extreme situation of high connected networks is considered as
limit case to test the performances of the connectivity methods. For this condition,
we observed a clear degradation of the partialization approach (Fig. 3.12a, third
column). This result might be explained with the so-called “married-nodes” effect
in which two generic nodes could be interpreted as direct (i.e., mono-synaptically)
connected if they are connected to a third common neuron [34]. These results were
confirmed by the AUC analysis that showed PC statistically different (and better)
with respect to CC and TE (Wilcoxon rank sum test; Fig. 3.12b) not only in small
(60 nodes: p < 0.001 at L, M and H connectivity degree) but also in larger networks
Fig. 3.12 Partial Correlation performances. a ROC curves. Mean shape was extracted from 21
simulations; PC algorithm shows significant performances at low and medium connectivity degrees,
not only in small (60) but also in larger (120–240) assemblies. Only the high-connected networks,
increasing the number of “married nodes”, worsen the partialization process. b Areas under the
curves (AUC). Values of the areas under ROC curves: PC shows very good performances, statis-
tically different respect to CC and one-delay TE (Kruskal–Wallis test. p < 0.05, 0.01, and 0.001
identified by one, two and three stars, respectively). Reproduced from [3] with permission from
2016 IOP Publishing Ltd; permission conveyed through Copyright Clearance Center, Inc ©
(120 nodes: p < 0.001 at L and M connectivity degree; 240 nodes: p < 0.01 and p <
0.05 at L and M connectivity degree, respectively).
56 3 Results
3.4.6 PC Application to Experimental Data
I also applied the developed partial correlation algorithm on neural networks coupled
to the MCS60 low resolution acquisition system.
Modularity Analysis
I first considered data coming from segregated yet structurally and function-
ally connected neuronal populations, and compared with homogeneous networks
(control). The rationale was to further support, with experimental data, whether the
partialization approach and the implementations of CC and one-delay TE proposed
in this work, were capable of clearly individuating the two segregated populations by
means of the modularity index Q [5, 35]. Such an index was extracted by applying
first the three methods to n = 8 simulations coming from an interconnected network
model, and then to the spontaneous activity of n = 8 cortical assemblies grown
in vitro by using a dual-compartment experimental set-up. I quantified whether and
to what extent the methods identified the segregation effects. Figure 3.13a, b shows
two examples of functional connectivity graphs, obtained by means of PC applied to
the two experimental set-ups (i.e., the standard MEA device (Fig. 3.13a) and the dual
compartment system (Fig. 3.13b: the black lines indicate the functional connections
that cross the physical barrier between the compartments). I analyzed data during the
third week in vitro since it is considered a stable and mature stage of development
of dissociated cortical networks [36].
The compartmentalization effect was identified by all connectivity methods both
in silico (Fig. 3.13c) and in vitro (Fig. 3.13d) models; however the modularity
detected by PC was much higher and more significantly different than CC and one-
delay TE, both in simulated (Fig. 3.13c; PC modularity ~0.59 ± 0.01, p < 0.001;
CC and one-delay TE modularity <0.5 ± 0.02, p < 0.01 and p < 0.05 respectively)
and experimental conditions (Fig. 3.13d; PC modularity ~0.63 ± 0.04, p < 0.001;
CC and one-delay TE modularity <0.5 ± 0.03, p < 0.01 and p < 0.05 respectively).
This is a clear and additional proof about the actual performances of the connectivity
methods used in these experimental conditions.
3.4.7 Connectivity During Development
As application of our partialization method, I characterized the functional-effective

connections and I derived the topological structures in developing neuronal cultures
(i.e., from the second to the fourth week in vitro). I considered two datasets: the first
one composed by n = 7 spontaneous homogeneous cortical assemblies (2 batches),
the second consisting of n = 10 chronically stimulated cultures (3 batches) by low-
frequency electrical pulses (cf. Sect. 2.1). I estimated functional connectivity by
means of the PC method and I evaluated basic graph measures. I normalized them
by the average of the same graph indices extracted by 100 randomized surrogates
Fig. 3.13 a Example of functional connectivity map obtained by applying PC algorithm to a

homogeneous cortical network. b Connectivity map obtained by applying PC algorithm to a dual
compartment network. c Modularity Index obtained by applying PC, CC and one-delay TE methods
to the homogeneous (grey) and modular cortical network model (dark grey): PC method is a very
good choice to detect the compartmentalization effect introduced by the structure of the experimental
set-up (p < 0.001). d Modularity Index obtained by applying PC, CC and one-delay TE methods to
the experimental data (homogeneous—grey-, and interconnected network—dark grey-). The results
obtained in the computational model were confirmed also during experimental analysis. Reproduced
from [3] with permission from © 2016 IOP Publishing Ltd; permission conveyed through Copyright
Clearance Center, Inc.
graphs (graph characterized by the same number of links and involved nodes with
respect to the experimental data), following the approach devised in [6].
When monitored during development, in vitro cortical cultures started to display
relevant spontaneous electrophysiological activity after the first week in vitro. For
this reason, I considered the evolution of the cortical networks from the second to
the fourth week.
The normalized network statistics (Fig. 3.14a) did not show a significant trend
during development (Wilcoxon rank sum test; p > 0.05). The increasing entropy
values that I found from the second to the fourth weeks in vitro, instead, provided
an important contribution to identify a significant emergence of a random structure
(Fig. 3.14c, red line; Wilcoxon rank sum test; p < 0.001), supporting the hypothesis
that cortical assemblies had a tendency to integrate rather than to segregate.
In addition to spontaneous-evolving networks, I studied how the chronic delivery
of electrical stimulation could affect the development of functional links. I considered
the development of n = 10 cortical networks during the pre-stimulus phases, from the
58 3 Results
Fig. 3.14 a Cluster Coefficient (red line) and Path Length (black line) analysis of 7 spontaneous
homogeneous networks during development. b Cluster Coefficient (red line) and Path Length (black
line) analysis of 10 stimulated homogeneous networks during development. c Entropy of sponta-
neous (red) and stimulated (black) homogeneous networks. d, e Link distributions during develop-
ment. f Histogram of number of connections extracted from spontaneous (dark grey) and stimulated
(grey) homogeneous networks. Reproduced from [3] with permission from © 2016 IOP Publishing
Ltd; permission conveyed through Copyright Clearance Center, Inc
second to the fourth week by studying how the topological structures changed from
the pre-stimulus to post-stimulus phase during the second week. By evaluating the
cluster coefficient, I observed that young cultures (II week in vitro) exhibited a more
segregated structure. In the following weeks (IIIrd and IVth) the cultures exhib-
ited the emergence of an integrated structure. This is clearly visible in Fig. 3.14b
with a decrease of the clusterization effect while keeping constant the normalized
values of Path Length. I observed a significant integration trend during development
(Wilcoxon rank sum test: II week pre or post stimulus/III-week pre-stimulus p
0.01; II week pre or post stimulus/IV week pre-stimulus p 0.001). Moreover, the
cluster coefficients that I found were significantly different from those extracted by
analyzing random network models (Wilcoxon rank sum test: II week pre or post
stimulus/random case p 0.001; III-week pre-stimulus/random case p 0.001; IV
week pre-stimulus/random case p < 0.05). These results were confirmed also by the
increasing Entropy levels, evaluated for each network node having non zeros cluster
coefficient values (Fig. 3.14c, black line). Our analysis shows how the stimulation
promoted a significant change in the network topology favouring the emergence of
a random functional structure as indicated by the decrease of the normalized Cluster
Coefficient values while keeping the Path Length measures on the maximum integra-
tion levels (levels equal to 1) during development (Fig. 3.14b). Moreover, I observed
that the stimulation did not have any effect on the connections length, since their
distributions did not change from the spontaneous to the stimulated homogeneous
networks (Fig. 3.14d, e). I did not find significant differences at level of number of
links (Fig. 3.14f). Finally, I also characterized the functional topological structures of
the same spontaneously developing and electrically stimulated networks by means
of CC and one-delay TE methods. The obtained results did not show significant
trends during development [3]. Altogether these results support the idea that PC can
be conveniently applied to estimate structural connections from functional-effective
links and that is capable to discriminate subtle changes in the network topology.
3.5 Algorithm Optimization
The functional connectivity analysis is a quite demanding computational process,

and all of the previous described connectivity methods for analyzing multiple spike
trains rely on quantities that need to be computed through intensive calculations
[4]. In order to reduce the time requested for the computation and to improve the
Random Access Memory (RAM) usage, I decided to combine informatics tool for
the implementation of the statistic methods, with a computational logic approach
based on a redefinition of the connectivity algorithms to be tailored for spike train
data. In this section, I will describe the latter approach, while in the next one, I will
introduce and define the informatics instruments as well as the informatics tools that
I developed for the complete analysis of dynamics and functional connectivity of
spike trains data. The computational logic approach modifications allowed a very
drastic reduction of the time requested for the computation and the RAM usage. In
fact, I implemented a C# optimized version of the transfer entropy algorithm using
informatics optmization strategies to reduce the time needed for the computation. I
used an in silico neural networks with 500 Izhikevich-based neurons simulated for
1 h to test the developed algorithm. That implementation required 42 h to provide the
full 500 × 500 connectivity matrix with a 4 cores i7 2.5 GhZ, 16 GB of RAM and
solid state disk. Such a computational time was not optimal in the optic of analysing
neural networks coupled to MEA-4k with 4096 electrodes. Spike trains are very
sparse data and, as a consequence, the possibility to consider only the non-zeros
value in the algorithm computation, can drastically reduce the requested time for
computation. The chosen strategy is based on the efficient use of time stamps (i.e.,
the index of a sample correspondent to a spike) as input and working data. In fact, in
the in silico networks I used for the testing, the firing rate of neurons are around 2–3
spikes/s, meaning that there are totally 7200–10,800 spikes with respect to 3600,000
total number of samples (sampling frequency set at 10 kHz), with a reduction of
the data dimensionality of more than 3 orders of magnitude. In the next paragraphs,
I will describe the logic optimization and implementation of FNCCH, TE and JE
algorithms.
60 3 Results
3.5.1 FNCCH: Algorithm Optimization
The optimization procedure for the FNCCH is based on time stamps relative to the
occurrence of a spike on a specific electrode. The block diagram and pseudocode
depicted in Fig. 3.15 shows the computational operations executed by the optimized
CC algorithm (FNCCH). We start from the first bin containing a spike in the target
train. The binning procedure is directly performed on the time stamps. For each pair
of neurons, starting from the first spike of the target train, we slide the time stamps
of the reference electrode to find the first spike whose correlation window contains
the target investigated spike. Then, we continue to move over the target train to
build the entire cross-correlogram (for that reference spike). When the correlation
window for the reference spike is completed (i.e., when we have counted the number
of spikes for all the bin of the target spike train), we move to the next spike of the
reference train, and re-iterate the procedure starting from the first target spike into
the correlation window, centered in the current reference spike. Thus, we normalize
the CC and repeat all the aforementioned operations for the other electrodes. An
important optimization strategy exploits the symmetry of the CC function allowing
considering only half of the electrodes for the computation. Moreover, we choose, as
target train, the one presenting the smallest number of spikes to reduce the number of
operations for each pair. Once the NCCH is obtained, we apply the filtering operation
described by Eq. (2.2) to compute the FNCCH values. Finally, we take the maximum
absolute value as estimation of the correlation between the two electrodes. If it is
Fig. 3.15 Schematic representation and description of the algorithm to obtain the FNCCH.
Reproduced from [9]
3.5 Algorithm Optimization 61
negative, the found connection is likely to be an inhibitory link, otherwise the found
connection is likely to be an excitatory link.
3.5.2 Transfer Entropy Algorithm Optimization
Figure 3.16 schematically represents and describes the main operations of the imple-
mented TE algorithm. First of all, the customized algorithm performs a binning
procedure (with a bin width set by the user via the GUI) directly on the time stamps.
Then, we start the computation of the information transfer among each couple of
electrodes. We compute the TE order, that is equal to x_order (i.e., number of bins in
the reference train) + y_order ( i.e., number of bins in the target train) + 1 (past bin
in the reference train). At this stage, we build a pattern vector with TE order elements
initialized with x_order bins of the reference train, the one past bin of the reference
train, and the y_order past bins of the target train. The number of all the possible
patterns is 2TEorder , and we use a decimal code (0, 1, 2,…, TE order) for coding each
binary pattern. Thus, starting from the first element of the pattern vector, we select
the minimum bin (i.e., minimum spike time stamp value) and initialize the code to
the simplest pattern (1 in that bin, and 0 for all the others). Then, we substitute this
element of the pattern vector with the next spike (from the correspondent train) and
Fig. 3.16 Schematic representation and description of the algorithm to obtain the TE. Reproduced
from [4]
62 3 Results
select again the minimum bin indifferently from the reference or the target train. If
this bin belongs to a common pattern with the precedent bins the code is updated
and increased, otherwise, a vector of length TE order, representing the frequency
of occurrence of each code is updated and the code is re-initialized to the simplest
pattern for the updated minimum bin. We repeat all these operations until the number
of spikes of the reference or the target train is ended. The null code (related to the
pattern formed by only empty bins) is obtained as the total number of recorded bins
minus the sum of all other patterns frequency of occurrences. At this stage, we use the
frequency of occurrences to extract each probability of occurrences and we compute
the TE value among the analyzed electrode by applying Eq. (2.12). For delayed
transfer entropy, I used a multithreading implementation to have a different thread
for every different delay. At the end of the computation, I save the maximum value
of TE and the correspondent delay.
3.5.3 Joint Entropy Algorithm Optimization
Figure 3.17 schematically describes and represents in form of pseudocode the main
operations executed by the customized JE algorithm’s implementation. Considering
the first spike of the reference train, we slide the target train to find the first spike
Fig. 3.17 Schematic representation and description of the algorithm to obtain the JE
3.5 Algorithm Optimization 63
next to the reference one. We compute the cISI value and update the cISI histogram
(i.e., the counting of cISI of specified sizes). We iterate such a procedure until all
the reference’s spikes have been analyzed. Finally, we compute the cISI probability
normalizing the number of cISI of a specific width for the total number of cISIs. The
application of (15) allows the computation of the JE between the analyzed electrodes.
3.6 Software Development for Functional Connectivity

Analysis
Recent advances in multichannel extracellular recording techniques have made

possible the simultaneous recording of the electrophysiological activity of thousands
of neurons, with in vitro arrays that can contain up to thousands of microelectrodes
[37–41]. Such relevant advancements in the technology have made routine the acqui-
sition of massive amounts of data. Moreover, quite often the experimental protocols
researchers perform require long recordings (e.g., tens of minutes, hours) which
generate a huge amount of raw data. For the above reasons, new computational strate-
gies and informatics toolboxes are requested for optimizing the management and
analysis of the acquired data. When I started my PhD, to the best of my knowledge,
there was no available dedicated software that put together a set of different func-
tional connectivity analysis methods, with a friendly Graphical User Interface (GUI)
and appositely tailored to be applied on spike train data. For this reason, I decided
to develop a user-friendly toolbox [42] in order to provide the researchers commu-
nity with a powerful tool to perform functional connectivity analysis on in-vitro
neuronal networks coupled to standard and high-density MEAs, while guaranteeing
computational efficiency and high accuracy.
3.6.1 Main Graphical Interface
I designed and developed TOOLCONNECT taking care of the user-friendliness.

According to this, our software package has a Graphical User Interface (GUI), which
permits also to inexperienced users to perform functional connectivity analysis, to
graphically represent the results, without knowing the details of the algorithms and
the organization of the software’s code. Figure 3.18 shows in form of block diagrams
the organization of TOOLCONNECT, while Fig. 3.19 shows an example of the graphical
tools provided to the users by TOOLCONNECT, and inherited by SPICODYN.
64 3 Results
Fig. 3.18 Schematic overview of TOOLCONNECT. The Functional blocks show the computational
and graphical tools embedded in the software package. The flow chart starts with the pre-processing
section; it includes the data acquisition procedure and all the other operations necessary to create
the spike trains, which are the software’s input data. Reproduced from [42]
3.6.2 Example of ToolConnect’s Application

to Experimental Data
3.6.2.1 Spontaneous versus Stimulated Networks
I applied the cross-correlation algorithm (in the frequency domain) to cultured

hippocampal neural networks (see Supplementary Material for cell culture) during
both spontaneous and stimulus-evoked activity. Electrical stimulation was performed
by applying bi-phasic voltage stimuli delivered at the frequency of 0.2 Hz, with a
peak-to-peak amplitude of 2.0 V, from a single electrode. Analysis was performed
on recordings lasting 10 min (sampling frequency of 10 kHz). Figure 3.20 shows
the connectivity graphs and the connectivity matrices obtained for the spontaneous
conditions (Fig. 3.20a, d) and stimulation phases relative to a stimulation phase from
channel 45 (Fig. 3.20b, e) and 21 (Fig. 3.20c, f). All these graphs and matrices were
obtained by thresholding the CM using a threshold equal to μ + 2·σ, where μ and
σ are the mean value and the standard deviation of the CM’s elements, respectively
(cf. hard threshold). After the thresholding procedure, during spontaneous activity, I
detected more links (about 90%) than during stimulated activity (Fig. 3.21c). These
functional links involve a larger number of electrodes (about 53%) in the spontaneous
condition with respect to the stimulated one (Fig. 3.21d). Moreover, the correlation
values obtained for the spontaneous activity are lower but more homogeneous than
the values corresponding to the stimulated one (maximum values’ difference of
3.6 Software Development for Functional Connectivity Analysis 65
Fig. 3.19 Example of TOOLCONNECT’s graphical output. a Sketch of connections among three
neurons in a simulated neuronal network. b, c Cross-correlograms computed by using the frequency
and the time approach respectively. d Partial correlograms computed among the three electrodes. e,
f Correspondent TE and the JE matrices. The peak appearing for the cross- and partial correlograms
12vs13 and 12vs21 corresponds to a connection found between the aforementioned electrodes. In
the same way, the TE presents the highest values for the couples (12, 13) and (12, 21) and the
joint entropy shows the smallest values for the same couples; meaning that these methods found
a connection between the aforementioned electrodes. Regarding the electrodes 13 and 21, instead,
the absence of peak in the cross-and partial correlograms, the low values for TE and the high values
for JE suggest the absence of a connection, as the sketch of the panel a shows. Reproduced from
[42]
0.16, standard deviation 0.008 versus 0.045 for spontaneous and stimulated activity
respectively, Fig. 3.20d, e, f). I also evaluated the in- and the out-degree distribu-
tion. We can observe that during spontaneous activity, the in- and the out-degree
for the different electrode are almost equally distributed among the active electrodes
(Fig. 3.20g). During stimulus-evoked activity, the stimulated electrode showed an
increased number of outgoing connections, reaching a difference of 14 links with
the other electrodes (Fig. 3.20h, i). In the case corresponding to the stimulation of
electrode 21, we had only one electrode with more than 3 outgoing connections.
When stimulating the electrode 45, I found only one electrode with more than 4
outgoing-ingoing connections. In the spontaneous condition, instead, there were 35
and 28 electrodes with an out-degree of connections greater than 3 and 4 links, respec-
tively. Finally, I used two metrics form graph theory (namely, cluster coefficient and
path length) to evaluate the topological characteristics of the neural networks, in the
two different experimental conditions. Figure 3.21a, b show the results relative to the
cluster coefficient and path length, respectively. I could observe that the path length is
not affected by the electrical stimulation (i.e., same values for the spontaneous and the
66 3 Results
Fig. 3.20 TOOLCONNECT’s application to mature hippocampal assemblies. TOOLCONNECT’s appli-

cation to mature hippocampal assemblies. Cross-correlation algorithm was applied to hippocampal
networks in spontaneous and stimulus-evoked conditions. a, d, g connectivity graph, connectivity
matrix and degree distribution for spontaneous activity. b, e, h connectivity graph, connectivity
matrix and degree distribution for stimulated activity (site of stimulation, electrode 45). c, f,
i Connectivity graph, connectivity matrix and degree distribution for stimulated activity (site of
stimulation, electrode 21). Reproduced from [42]
stimulus-evoked activity). On the contrary, I found a higher cluster coefficient for the
spontaneous conditions than for the stimulated ones. This is probably due to the effect
of the stimulation on the global network’s dynamics: the stimulation empowers the
outgoing connections from the stimulated electrode, thus, these connections become
stronger, and correspond to higher cross-correlation values. We can observe that, the
connectivity graph for the stimulation of the electrode 21 and 45 are quite different.
In particular, the outgoing connections of the electrode 45, are weaker than the ones
relative to the electrode 21 (Fig. 3.20b, c). Moreover, the degree distribution relative
to the electrode 45 shows a higher variability in the number of in-coming and out-
going connections than the one relative to the electrode 21. Figure 3.21e, f shows the
population Post Stimulus Time Histogram (PSTH) relative to the two different sites
of stimulation (i.e., the PSTH averaged over all the electrodes); we can observe that
the number of evoked spikes by the stimulation of channel 45 (Fig. 3.20e) is lower
than the number correspondent to the stimulation of the electrode 21 (Fig. 3.21f).
Fig. 3.21 Graph Theory’s analysis of hippocampal neural assemblies. a Cluster coefficient. b Path
length. c, d Number of links and number of neurons found after the thresholding procedure, respec-
tively. e PSTH relative to the stimulation of electrode 45. f PSTH relative to the stimulation of
electrode 21. Reproduced from [42]
It is possible to ascribe that the electrode 45 is less involved in the dynamics of the
analyzed network with respect to the electrode 21. Reproduced from [42].
3.6.2.2 SPICODYN
I implemented an automated and efficient open-source software for the analysis

of multi-site neuronal spike signals. The software package, named SPICODYN, has
been developed as a standalone windows GUI application, using C# programming
language with Microsoft Visual Studio based on.NET framework 4.5 develop-
ment environment. Accepted input data formats are HDF5, level 5 MAT and text
files, containing recorded or generated time series spike signals data. SPICODYN
processes such electrophysiological signals focusing on: spiking and bursting
dynamics and functional-effective connectivity analysis. In particular, for infer-
ring network connectivity, a new implementation of the transfer entropy (TE)
method is presented dealing with multiple time delays (temporal extension) and with
multiple binary patterns (high order extension). SPICODYN is specifically tailored
to process data coming from different MEAs setups, guarantying, in those specific
cases, automated processing. The optimized implementation of the Delayed Transfer
Entropy (DTE) and the High-Order Transfer Entropy (HOTE) algorithms, allows
performing accurate and rapid analysis on multiple spike trains from thousands of
electrodes. The new developed software SPICODYN inherits the overall architecture
from TOOLCONNECT[42], no more supported, and includes new relevant features such
as spike detection, first-order statistics for spikes and bursts, fully revisited TE based
algorithms, and topological analysis. Specifically, it includes a brand new temporal
68 3 Results
Fig. 3.22 Screenshot of the main interface of SPICODYN. Reproduced from [4]
extension of the TE dealing with multiple time delays and orders higher than one [2,
43]. Figure 3.22 displays a screenshot of the main raw data processing user interface
providing all the functionalities that will be described in the next sections.
3.6.3 Multi-threading Implementation
SPICODYN is a Multiple Document Interface (MDI) windows form application. A

father form is a frame in which all the embedded computational and graphical tools’
interfaces are opened and displayed. I implemented each of these interfaces inde-
pendently from the others, as a windows form. A friendly GUI and a complete set of
feedback information are fundamental in the definition of the toolbox. According to
this, each windows form has a multi-thread implementation: one thread of the appli-
cation is used to update the graphical interface and another one executes the effec-
tive code of the selected connectivity method. In this way, it is possible to update a
progress bar indicating the state of the application and other controls reporting useful
information (e.g. the start time of the algorithm and the percentage of progress), and
to make this interface accessible any time in order to resize, minimize or move the
relative window. Some of the threads were directly implemented at code level, while
other were implemented by use a.Net Framework 4.5’s powerful embedded control to
handle the multi-threading application: the backgroundworker. This object provides
the user with the methods DoWork, RunWorkerCompleted and ProgressChanged.
The first one contains all the code that must be executed in the thread (in our case
the connectivity method’s code); the second one indicates the operations to run after
the conclusion of the thread. The last one indicates the operations to do when a
change in the progress occurs (e.g., the update of a progress bar). When performing
functional connectivity analysis, the need of a large amount of RAM and the high
computational time of execution are the two main issues to manage. The aforemen-
tioned multi-threading implementation permits the various connectivity methods to
be performed simultaneously on the different threads of the different available CPUs,
significantly reducing the computational requested time.
3.6.4 Input Data Processing and Conversion Procedures
Researchers need a reliable format for exchanging large datasets for specific data
analysis. A well-recognized possibility of a common file format is based on the HDF5
(Hierarchical Data Format) structure [44]. HDF5 format (structured binary format)
allows fast loading and low processing times when dealing with huge amounts of
data [45].
For this reason, I adopted this standardized raw input data format, making our tool
compatible to deal with data coming from different acquisition systems. In fact, most
of the acquisitions system’s related software provides tools and packages to convert
the specific format or to save the acquired electrode’s raw data into a standardized
level-2 HDF5 format (e.g., “DataManager” from MCS Systems—“mcd” format,
“BrainWaveX” form 3Brain—“brw” format). However, SPICODYN offer the possi-
bility to read also the level 5 MAT-files format (v5-v6-v7, The Mathworks, Natick,
US) and convert them into textual or binary files, which are the supported formats for
spike trains files to be used as input for the spike analysis and connectivity methods.
Such a conversion procedure is done by using dedicated open-source.NET libraries
and adopting an optimized implementation strategy to extract the structured data
elements values. More specifically, I implemented a procedure to iteratively move
through the file by finding a tag and then reading ahead in memory the specific number
of bytes until the next tag. The conversion procedure offers also the possibility to join
spike trains related to distinct phases of the same experiment, provided in separate
level 5 MAT-files. Figure 3.23 graphically illustrates the MAT-files format providing
details about the structured elements. Finally, SPICODYN guarantees the compati-
bility with spike trains format data generated by SpyCode [46]: a previously devel-
oped software from our research group. The main raw data pre-processing features
offered by SPICODYN includes: (i) electrode’s raw signal visualization with a time-
varying plot of the selected electrodes in a specific interval; (ii) selectable algorithms
of spike detection (Cf., Sect. 2.3.1); (iii) raster plot of the identified spiking activity
while the electrode’s raw signal is being displayed. Figure 3.24 represents a block
diagram of SPICODYN’s functionalities and process flow, schematically describing
the operations performed by the user to select the input data files, to choose the
analysis method to perform, and to set the input parameters. It also depicts the opera-
tions performed by SPICODYN’s modules, related to the input raw data management,
conversion, spike detection and spike trains analysis (dynamics and spiking/bursting
70 3 Results
Fig. 3.23 Matlab Level 5 MAT file-format. Reproduced from [4].

Fig. 3.24 Functionalities block diagram and process flow chart describing the flux of operations
performable using SPICODYN. The blue line identifies a possible path to perform a complete
analysis starting from the raw data in HDF5 format, and performing spike detection, spiking and
bursting analysis as well as connectivity analysis. Reproduced from [4]
statistics, functional-effective connectivity analysis, connectivity matrix and graphs

visualization, topological features analysis and graph theory metrics).
3.6.5 Spiking and Bursting Features Analysis
The creation of huge amounts of data from simultaneous recording of many (thou-
sands of) neurons with HD-MEA acquisition setups implies the development of
efficient methods and specific tailored algorithms to analyze such neuronal spike
trains. Spike detection methods, applied to recorded electrodes raw signals, rely on
quantities that require intensive calculations. SPICODYN’s interface provides the user
with commonly used algorithms to perform spike detection, and some subsequent
statistical analysis on the extracted spike trains data.
3.6.6 Spike Detection
The current version of SPICODYN implements two spike-detection algorithms: “Spike

Detection Differential Threshold” (SDDT) and “Precision Timing Spike Detection”
(PTSD).
I developed a dedicated form that provides the necessary fields to specify all
the required input parameters. The spike detection algorithms involve a common
step related to the threshold computation. I implemented two different strategies to
compute distinct threshold values for all the channels:
72 3 Results
(i) “automatic” sliding-window threshold computation procedure: the user sets the
number of positions of the sliding window (k: typical values between 10 and
20), the percentage of coverage of the total signal length (or the complementary
sliding window length—w), and the multiplication factor (n) for the computed
standard deviation (SD) of the analyzed signal’s portion. The SD value is eval-
uated for all the windows positions (k), and the minimum value represents the
basal noise threshold value. The threshold on every channel can be written as:
thr esholdchannel_i = n ∗ min k {S D(channel_i_samples(wk ))} (3.1)
(ii) “manual” user-set interval threshold computation procedure: the user sets an
interval for the threshold computation on the basis of the signal’s amplitude
trend by visual inspection. The threshold value for every channel corresponds
to the SD value of that precise portion of the signal, multiplied by the SD factor
n. Taking into account the setup-dependent signal’s amplitude and variability,
the automatic procedure for the threshold computation, when applied to 3Brain
data, uses the mean SD value computed in all the sliding window positions
instead of the minimum SD value. In both methods, I implemented a preliminary
control (of the noise levels) procedure, to exclude from the spike detection those
channels that exhibits higher levels of noise with respect to a maximum admitted
value. In the case of the 3Brain data, this procedure is based on a maximum
threshold value that is set by the user. In the case of the MCS data instead, the
control is performed only with the “manual” threshold computation option, and
the maximum threshold value is computed on a “reference” electrode (set by
the user).
I implemented a version of the SDDT procedure following the algorithm as

described in [47]. The PTSD procedure is implemented on the basis of the specifics
described in [48]. Both procedures have been adapted and customized to deal with
multiple electrodes raw data in the HDF5 file format. I took specific care to guarantee
an optimized memory data management and to reduce the computational costs. To
achieve this, in the worst case of data coming from the 4096 electrodes from the
3Brain setup (sampling frequency of 9043 Hz), I implemented a memory buffer that
stores only a short portion of 10 s (i.e., 10 × 9043 samples per buffer) of the data file
in an iterative manner. I also applied a multi-threading approach to parallelize spike
detection for different electrodes on different threads.
3.6.7 Burst Detection
A burst consists of a sequence of spikes separated by short time intervals within

the sequence, and by a relatively long interval between two spikes belonging to
two different sequences [49]. However, the definitions of bursts and burst detec-
tion methods differ among studies. Commonly accepted analysis tools employ burst
detection algorithms based on predefined criteria [50]. The specific parameters of
the methods adopted to identify bursts from MEA recorded spike trains are usually
tailored according to the experimental conditions, and generally differs in the inter
spike interval (ISI) thresholds definition and usage and in the minimum number of
consecutive spikes for burst identification. Some studies use the average ISIs of the
measurements [51], average firing rates or logarithmic histogram of ISIs [52] to
calculate an ISI threshold for detecting bursts. In the current version of SPICODYN,
I implemented the detection of bursts as sequences of at least N consecutive spikes
spaced less than a chosen ISI threshold value, according to [53]. The two threshold
parameters are provided directly by the user through the interactive burst detection
interface. The first one is the maximum ISI for spikes within a burst. The second one is
the minimum number of consecutive spikes belonging to a burst (N). Our algorithm
applies an efficient strategy to scan the spike train time stamps values and calcu-
late the ISIs while at same time evaluating the threshold conditions. Thus avoiding
the need to scan the entire sparse binary spike train improving the computational
performances and data storage resources usage.
3.6.8 Spiking and Bursting Statistics
Different metrics, such as overall spiking activity, firing rates, bursting frequency and
duration, can be used to have a first overview of the dynamic behavior of a neuronal
network. Indeed, such simple metrics are extensively applied when MEAs recorded
data are used for neuro-toxicology and neuro-pharmacological studies [54, 55]. In
SPICODYN, I implemented several spiking and bursting statistics to characterize the
dynamical features exhibited by the underlying neuronal networks.
(i) Spiking activity (Raster Plot) and spiking statistics: SPICODYN provides the
computation of the single electrode’s firing rates (FR) and the overall network’s
mean firing rate (MFR), as well as the evaluation of the number of “active”
electrodes (i.e., with FR > FR_threshold_value). After the spiking statistics
computation, the user can choose the requested time intervals to plot the raster
plot trough the dedicated interface.
(ii) Inter Spike Interval (ISI): is defined as the time interval between consecutive
spikes and it is a commonly used metric in the analysis of neuronal recordings.
SPICODYN provides the network’s mean and relative standard deviation of ISI,
and the ISI coefficient of variation (CV). It also provides the ISI histogram (ISIH)
computation and plots for the whole set of electrodes. The latter is constructed
evaluating the time (samples) intervals between every couple of consecutive
spikes while binning at a user specified bin width:
74 3 Results
(ti − ti−1 )

I S I H (bin k ) = k−1≤ < k , k = 1, . . . , m; i = (2, N ) (3.2)
i
w
where w is the bin width, k is the k-th bin index, m is the maximum bin number
(user defined limit for the histogram plot), ti is the time (or timesamp) of occur-
rence of i-th spike, (ti − ti−1 ) is the ISI between spikes i and i−1, N is the total
number of spikes on the entire set of electrodes.
(iii) Inter Burst Interval (IBI): is the time interval between two consecutive bursts
(spike sequences). It is computed for the whole set of electrodes burst data spike
evaluated during the burst detection procedure. The IBI Histogram (IBIH) is
computed evaluating the time (samples) intervals between every couple of
consecutive bursts while time binning at user specified bin width:
(tbi − tb(i−1) )

I B I H (bin k ) = k−1≤ < k , k = 1, . . . , m; i = (2, B)
i
w
(3.3)
where w is the bin width, k is the k-th bin index, m is the maximum bin number (user
defined limit for the histogram plot), tbi is the time of occurrence of burst bi (first
spike of the sequence), (tbi − tb(i−1) ) is the IBI between bursts i and (i−1), B is the
total number of bursts on the entire set of electrodes. Indeed, network’s mean and
standard deviation of the IBI is also provided.
(iv) Burst Duration (BD): is the sum of the ISIs of the sequence of consecutive spikes
within the burst itself, i.e., the time interval between the first and the last spike
in the spikes sequence. I applied an efficient strategy by storing only the first
and last spike time stamps (and the number of spikes within the burst) allowing
the computation of the BD as the difference between the two time stamps. The
BD histogram (BDH) is constructed by summing the whole network’s total
number of bursts durations at a user specified time bin:
(tbi,n − tbi0 )

B D H (bin k ) = k−1≤ < k , k = 1, . . . , m; i = (2, B)
i
w
(3.4)
where w is the bin width, k is the k’th bin index, m is the maximum bin number (user
set limit for the histogram plot), tbi,0 is the time of occurrence of the first spike in
burst bi (first spike of the sequence), tbi,n is the time of occurrence of the last spike
in burst bi, B is the total number of bursts on the entire set of electrodes.
(v) Bursting statistics: SPICODYN provides the computation of single electrode’s

bursting rates (BR) and the overall network’s mean bursting rate (MBR). It
also provides the calculation of the percentage of random spikes (that can be
alternatively expressed as the complementary of the percentage of bursts related
spikes [56] for a single electrode and the entire network’s mean and standard
deviation values). Random spikes are the spikes that are not involved in the
sequences within the bursts. Finally, the percentage of bursting time (with related
network’s mean value and standard deviation) is provided. It is defined as the
fraction of the recording length (for each electrode) in which bursts occur:

% coverage = L/ B Di , i = 1, . . . , n (3.5)
i
where L is the total recording’s length, n is the total number of electrode’s bursts,
and B Di is the i-th burst’s duration.
3.6.9 Comparison with Other Software
I performed a detailed comparison between SPICODYN and SpyCode based on the

times needed to perform the most resources-requiring dynamics analysis operations,
and the computation of a subset of the main spiking/bursting parameters. Table 3.1
summarizes the results in terms of the time needed to complete each of the main
steps of the reported set of operations, on a 10-min recording of a mature cortical
neuronal network coupled to a 60-electrodes MEA. The software run under Windows
10 on 4 cores i7, 2.5 GHz, 16 GB of RAM and solid state disk. It is important to
underline that while obtaining appreciably lower computational times, the results
in terms of the produced output values of the procedures are identical compared to
those obtained with SpyCode (some not significant differences emerged only due
Table 3.1 Comparison of the SPICODYN and SpyCode computation times evaluated for the
processing of a neural network coupled to a 60 microelectrodes MEA (from MCS). All the
computational tests have been performed on a 4 cores i7 2.5 GhZ, 16 GB of RAM and solid
state disk. Reproduced from [4]
Algorithm SPICODYN[s] SpyCode [s]
Automatic threshold (30 windows of 200 ms) 1 10
Manual threshold (50 s interval) 2 10
Spike detection (SDDT) 2 15
Spike detection (PTSD) 12 25
Spike analysis (MFR + Raster Plot + ISI) 1+5+2 5 + 30 + 5
Burst analysis (MBR + MBD + IBI) 1+2+2 5+5+5
76 3 Results
Fig. 3.25 Schematic representation and relative computational times of the algorithm used to
analyze 30 min of spontaneous activity recorded by high density MEA with 4096 microelectrodes.
All the computational tests have been performed on a 4 cores i7 2.5 GhZ, 16 GB of RAM and solid
state disk. Reproduced from [4]
to some customized strategies that I applied in the spiking and bursting statistics
computation).
Spiking and connectivity analysis performed on HD-MEAs show fast computa-
tional time, considered the large amounts of calculations to be performed (Fig. 3.25).
When considering spike detection (the most demanding computational task),
SPICODYN requires about the same time that BrainWaveX software (3Brain) needs
to run the same algorithm.
3.6.10 Use of SpiCo Dyn to Process Long-Lasting

Recordings on MEA-4k
In this section, I will report an example of SPICODYN’s application to cortex neural

networks coupled to the MEA-4k. As Fig. 3.25 shows, the analysis is composed
of a part related to the dynamics and one related to the evaluation of the topolog-
ical properties of the network. The total time required to analyze both the spiking
and bursting dynamics, as well as the functional-effective topological properties is
812 min. It is worth noticing that the dynamical characterization requires less than
31 min: 30 min for the spike detection and 26 s to produce all the output (figures and
text files). Most of the computational time is consumed by the functional-effective
connectivity algorithm. However, it is worth to underline that it was necessary the
computation of 170 TEs to cover a temporal window of 20 ms, while a single TE
required only 4.5 min to be evaluated. Finally, the topological features extraction
on the active electrodes (i.e. 1980 over 4096 electrodes) required only 95 s to be
processed.
3.6.11 Dynamic Characterization
SPICODYN is independent from the acquisition system used to acquire the electro-
physiological data; according to this, it uses the HDF5 standardized input data format.
The 3Brain acquisition system allows to directly export the raw data in HDF5 format
online during the recordings. Once the HDF5 files has been stored, the user can
select the folder containing the raw data. In this way, SPICODYN permits to analyze
in parallel all the experiments and phases contained in the input data folder. After
reading the HDF5 files, the PTSD algorithm is used to spike detect the raw data. I
report the results of a complete analysis for a neural network coupled to the MEA-4k.
The mean MFR computed among all the active electrodes (1916 over 4096) was 0.55
± 1.63 spikes/s. Figure 3.2a represents the ISIH, -where it can be noticed that most of
the ISIs are smaller than 15 ms with an exponential-like decay. Then, I performed a
bursting analysis using the algorithm described in Burst Analysis. Bursting features
are characterized by means of the IBIH (Fig. 3.2b) and the mean burst duration
histogram (Fig. 3.2c). As the table of the panel of Fig. 3.2d shows, bursting activity
presents a mean BD of 139 ± 89 ms, a MBR of 1.74 ± 3.31 bursts/s and a percentage
of random spiking of 51.26 ± 31.55%.
3.6.12 Connectivity Characterization
The connectivity characterization is reported in section DTE application to experi-

mental data.
References
1. Fawcett T (2006) An introduction to ROC analysis. Pattern Recogn Lett 27:861–874

2. Ito S et al (2011) Extending transfer entropy improves identification of effective connectivity in
a spiking cortical network model. PLoS ONE 6:e27431. https://doi.org/10.1371/journal.pone.
0027431
3. Poli D, Pastore VP, Martinoia S, Massobrio P (2016) From functional to structural connectivity
using partial correlation in neuronal assemblies. J Neural Eng 13:026023
4. Pastore VP, Godjoski A, Martinoia S, Massobrio P (2017) SPICODYN: a toolbox for the
analysis of neuronal network dynamics and connectivity from multi-site spike signal recordings.
Neuroinformatics 16:15–30. https://doi.org/10.1007/s12021-017-9343-z
5. van den Heuvel MP, Sporns O (2013) Network hubs in the human brain. Trends Cogn Sci
17:683–696. https://doi.org/10.1016/j.tics.2013.09.012
78 3 Results
6. Downes JH et al (2012) Emergence of a small-world functional network in cultured neurons.

PLoS Comput Biol 8:e1002522. https://doi.org/10.1371/journal.pcbi.1002522
7. Langer N, Pedroni A, Jäncke L (2013) The problem of thresholding in small-world network
analysis. PLoS ONE 8:e53199. https://doi.org/10.1371/journal.pone.0053199
10. Kreitzer AC (2009) Physiology and pharmacology of striatal neurons. Annu Rev Neurosci
32:127–147. https://doi.org/10.1146/annurev.neuro.051508.135422
11. Kaufman AM et al (2012) Opposing roles of synaptic and extrasynaptic NMDA receptor
signaling in cocultured striatal and cortical neurons. J Neurosci 32:3992–4003. https://doi.org/
10.1523/jneurosci.4129-11.2012
12. Marom S, Shahaf G (2002) Development, learning and memory in large random networks of
cortical neurons: lessons beyond anatomy. Q Rev Biophys 35:63–87
13. Schröter M, Paulsen O, Bullmore ET (2017) Micro-connectomics: probing the organization of
neuronal networks at the cellular scale. Nat Rev Neurosci 18:131. https://doi.org/10.1038/nrn.
2016.182
14. Barabasi AL, Albert R (1999) Emergence of scaling in random networks. Science 286:509–512
Front Neural Circ 9:57. https://doi.org/10.3389/fncir.2015.00057
16. P Massobrio L Arcangelis de V Pasquale HJ Jensen D Plenz 2015 Criticality as a signature of
healthy neural systems Front Syst Neurosci 9 22https://doi.org/10.3389/fnsys.2015.00022
17. Schroeter MS, Charlesworth P, Kitzbichler MG, Paulsen O, Bullmore ET (2015) Emergence
of rich-club topology and coordinated dynamics in development of hippocampal functional
networks in vitro. J Neurosci 35:5459–5470. https://doi.org/10.1523/JNEUROSCI.4259-14.
2015
18. Maccione A et al (2010) Experimental investigation on spontaneously active hippocampal
cultures recorded by means of high-density MEAs: analysis of the spatial resolution effects.
Front Neuroeng 3:4. https://doi.org/10.3389/fneng.2010.00004
19. Gerstein GL, Perkel DH (1972) Mutual temporal relationships among neuronal spike trains.
Biophys J 12:453–473. https://doi.org/10.1016/S0006-3495(72)86097-1
20. Rosenberg JR, Amjad AM, Breeze P, Brillinger DR, Halliday DM (1989) The Fourier approach
to the identification of functional coupling between neuronal spike trains. Prog Biophys Mol
Biol 53:1–31
10.1046/j.1460-9568.1999.00770.x
23. Quian Quiroga R, Kreuz T, Grassberger P (2002) Event synchronization: a simple and fast
method to measure synchronicity and time delay patterns. Phys Rev 66:041904
24. Bedenbaugh P, Gerstein GL (1997) Multiunit normalized cross correlation differs from the
average single-unit normalized correlation. Neural Comput 9:1265–1275
26. Aertsen AMHJ, Gerstein GL (1985) Evaluation of neuronal connectivity: sensitivity of cross-
correlation. Brain Res 340:341–354. https://doi.org/10.1016/0006-8993(85)90931-X
27. Melssen WJ, Epping WJ (1987) Detection and estimation of neural connectivity based on
crosscorrelation analysis. Biol Cybern 57:403–414
28. Dunn B, Mørreaunet M, Roudi Y (2015) Correlations and functional connections in a population
of grid cells. PLoS Comput Biol 11:e1004052. https://doi.org/10.1371/journal.pcbi.1004052
References 79
29. Aertsen A et al (1991) Neural interactions in the frontal cortex of a behaving monkey: signs of
dependence on stimulus context and behavioral state. J Hirnforsch 32:735–743
30. Schneidman E, Berry MJ, Segev R, Bialek W (2006) Weak pairwise correlations imply strongly
correlated network states in a neural population. Nature 440:1007–1012. http://www.nature.
com/nature/journal/v440/n7087/suppinfo/nature04701_S1.html
31. Cutts CS, Eglen SJ (2014) Detecting pairwise correlations in spike trains: an objective compar-
ison of methods and application to the study of retinal waves. J Neurosci 34:14288–14303.
https://doi.org/10.1523/jneurosci.2767-14.2014
32. Salinas E, Sejnowski TJ (2000) Impact of correlated synaptic input on output firing rate and
variability in simple neuronal models. J Neurosci 20:6193
35. Newman ME (2006) Modularity and community structure in networks. Proc Natl Acad Sci
USA 103:8577–8582
36. Kanagasabapathi TT et al (2012) Functional connectivity and dynamics of cortical-thalamic
networks co-cultured in a dual compartment device. J Neural Eng 9:036010. https://doi.org/
10.1088/1741-2560/9/3/036010
9:2644–2651
38. Buchs P-A, Muller D (1996) Induction of long-term potentiation is associated with major
ultrastructural changes of activated synapses. Proc Natl Acad Sci USA 93:8040–8045
39. Eversmann B et al (2003) A 128 x 128 CMOS biosensor array for extracellular recording of
neural activity. IEEE J Solid-State Circ 38:2306–2317
40. Muller J et al (2015) High-resolution CMOS MEA platform to study neurons at subcellular,
cellular, and network levels. Lab Chip 15:2767–2780. https://doi.org/10.1039/c5lc00133a
41. Viswam V, Dragas J, Muller J, Hierlemann A In: IEEE international solid-state circuits
conference. IEEE, pp 394–396
connectivity toolbox for in vitro networks. Front Neuroinformatics 10:13. https://doi.org/10.
3389/fninf.2016.00013
43. Overbey LA, Todd MD (2009) Dynamic system change detection using a modification of the
transfer entropy. J Sound Vib 322:438–453. https://doi.org/10.1016/j.jsv.2008.11.025
44. HDFGroup (2013) Hierarchical data format
45. Bennett K, Robertson J (2011) In: Ionescu CM (ed) MATLAB—a ubiquitous tool for the
practical engineer
46. Bologna LL et al (2010) Investigating neuronal activity by SPYCODE multi-channel data
analyzer. Neural Netw 23:685–697. https://doi.org/10.1016/j.neunet.2010.05.002
47. Vato A et al (2004) Spike manager: a new tool for spontaneous and evoked neuronal networks
activity characterization. Neurocomputing 58:1153–1161
48. Maccione A et al (2009) A novel algorithm for precise identification of spikes in extracellularly
recorded neuronal signals. J Neurosci Methods 177:241–249
49. Tam DC (2002) An alternate burst analysis for detecting intra-burst firings based on inter-burst
periods. Neurocomputing 44–46:1155–1159
50. Kapucu FE et al (2012) Burst analysis tool for developing neuronal networks exhibiting
highly varying action potential dynamics. Front Comput Neurosci 6:38. https://doi.org/10.
3389/fncom.2012.00038
51. Mazzoni A et al (2007) On the dynamics of the spontaneous activity in neuronal networks.
PLoS ONE 2:e439. https://doi.org/10.1371/journal.pone.0000439
52. Pasquale V, Martinoia S, Chiappalone M (2009) A self-adapting approach for the detection
of bursts and network bursts in neuronal cultures. J Comput Neurosci 29:213–229. https://doi.
org/10.1007/s10827-009-0175-1
80 3 Results
53. Chiappalone M et al (2005) Burst detection algorithms for the analysis of spatio-temporal
patterns in cortical networks of neurons. Neurocomputing 65–66:653–662
54. Bal-Price AK et al (2010) In vitro developmental neurotoxicity (DNT) testing: relevant models
and endpoints. NeuroToxicology 31:545–554. https://doi.org/10.1016/j.neuro.2009.11.006
55. Hogberg HT et al (2011) Application of micro-electrode arrays (MEAs) as an emerging tech-
nology for developmental neurotoxicity: evaluation of domoic acid-induced effects in primary
cultures of rat cortical neurons. NeuroToxicology 32:158–168. https://doi.org/10.1016/j.neuro.
2010.10.007
56. Grace AA, Bunney BS (1984) The control of firing pattern in nigral dopamine neurons: burst
firing. J Neurosci 4:2877
Conclusion
It is fundamental to elaborate research strategies aimed to a comprehensive struc-

tural description of neuronal interconnections as well as the networks’ elements
forming the human connectome [1]. The connectome will significantly increase our
understanding of how functional brain states emerge from their underlying struc-
tural substrate, and will provide new mechanistic insights into how brain function is
affected if this structural substrate is disrupted. The final goal of a connectivity anal-
ysis is the reconstruction of the human connectome, thus, the application of statis-
tical measures to the in vivo model in both physiological and pathological states.
Since the system under study (i.e. brain areas, cell assemblies) is highly complex, to
achieve the purpose described above, it is useful to adopt a reductionist approach.
During my Ph.D. work, I focused on a reduced and simplified model, represented
by neural networks chronically coupled to MEAs. I developed and optimized statis-
tical methods to infer functional connectivity from spike train data. In particular,
I worked on correlation-based methods: cross-correlation and partial correlation,
and information-theory based methods: Transfer Entropy (TE) and Joint Entropy
(JE). More in detail, my aim was to reconstruct the functional-effective connectivity,
trying to establish a relationship between dynamics and connectivity, in order to
understand if and in which way these parameters influence the topological (func-
tional) organization of a neural network. This information could be very useful to
help in understanding the mechanisms underlying the superior cognitive functions
(e.g., learning, memory, movement) in both physiological and pathological states. In
order to topologically characterize neural networks, it is important to use high reso-
lution acquisition system, having few neurons (1–3) per electrode. Thus, my PhD’s
aim has been applying functional connectivity methods to neural networks coupled to
MEA-4K (3Brain APS with 4096 electrodes). To reliably reconstruct the functional
topological organization of neural networks it is fundamental to detect and analyze
both excitatory and inhibitory links. In fact, inhibitory links not only represent a
consistent part of the total functional links (~25% [2]), but they can be fundamental
in orchestrating the synchrony and shaping the connectivity [3]. No published work
Neuronal Assemblies, Springer Theses, https://doi.org/10.1007/978-3-030-59042-0
82 Conclusion
about functional-effective connectivity estimation on multiple neuronal spike trains

provides a well-defined computational method to identify inhibitory connections.
Starting from the standard definition of the cross-correlation [4] (cf., Methods), I
adopted the normalization approach described in [5, 6], to obtain the “raw” Normal-
ized Cross-Correlation Histogram (NCCH). I formalized my hypothesis that, the
extraction of negative peaks (rather than troughs) obtained through a simple filtering
operation on the NCCH and followed by distinct thresholding operations for exci-
tatory and inhibitory connections, permits to identify a significant percentage of
inhibitory connections with a high level of accuracy. In fact, theoretically, cross-
correlation is able to detect both an increase and a decrease of the synchrony between
the spike trains relative to putative interconnected neurons. However, in real exper-
imental data, the cross-correlogram is very jagged making difficult the detection of
small peaks and troughs, and, apart specific conditions (i.e., high and tonic firing
rate) [7] hindering the detection of inhibition. My approach has consisted in a simple
post processing of the cross-correlation histogram, obtaining what I called Filtered
and Normalized Cross-Correlation Histogram. The FNCCH was able to detect with
very high precision both the excitatory and the inhibitory links in in silico neural
networks. I applied the FNCCH algorithm to neural networks coupled to the MEA-
4k, trying to extract the topological organization. I found that for cortical networks
coupled to MEA-4k devices it is possible to see the emergence of a clear small-world
(SW) topology. With the measurements performed with MEA-4k’s I can state that,
both inhibitory and excitatory links with their small world index, SWI >> 1 (9.2 ±
3.5 for the inhibitory links and 5.2 ± 2 for the excitatory ones) contribute to ‘segre-
gation’ with the emergence of small world networks. Moreover, inhibitory links,
with their somehow longer connections, might contribute also to network ‘integra-
tion’ (i.e., communication among the SWs). Such a result confirms the importance to
perform functional connectivity analysis using a method able to detect both excita-
tory and inhibitory links, above all when the final aim is to characterize the topology
of a neural network. When analyzing similar cortical networks but coupled to the
MEA-60 devices, no SW topology is identified; such networks seem to be charac-
terized by a sub-random topology with a SWI of 0.4 ± 0.1 for the excitatory and
0.2 ± 0.2 for the inhibitory links. These cortical networks are of the same type as
the ones coupled to the MEA-4k (i.e., similar density of neurons, same age, same
culture medium) and the apparent estimated random topology should be attributed
to the low number of recording sites (i.e., 60 channels) that are not enough to reli-
ably infer topological features (see section FNCCH results). Such a result confirms
the importance to consider neural networks coupled to high resolution acquisition
systems (MEA-4k). Both the results should be confirmed by applying the FNCCH
algorithm to process more neural networks not only in physiological state, but also
in a pathological state that could alter the topological organization. However, the
use of MEA-4k produces a huge amount of data, requiring the development of ad-
hoc computational efficient software platforms to process and analyze such data
[8]. For this reason, I developed the software package SpiCoDyn, aiming to be
a possible solution of the aforementioned big data problem, by introducing new
Conclusion 83
software tools, independent from the acquisition systems, able to process in reason-
able computational times, discrete time series coming from multi-site spike signal
recordings. The philosophy under this project was to develop and share with the
scientific community an open source platform that allows performing spike analysis
by including new tools, to enhance the repertoire of algorithms. The current version
of SpiCoDyn embeds a collection of functions (correlation and information-theory
based) to estimate functional-effective connectivity, part of which directly come from
the software ToolConnect [9], (https://www.nitrc.org/projects/toolconnect/). The
connectivity analysis section has been revised and expanded introducing the imple-
mentation of a customized fully optimized version of the TE algorithm, tailored to be
used on time stamps data extracted from multi-unit spike train recordings. In addition,
SpiCoDyn implements an extension of the TE method to deal with multiple time
delays (temporal extension) and with multiple binary pattern (high order extension),
the FNCCH and the JE. Another feature of the developed software is the possibility
to characterize the spiking and bursting dynamics, starting from raw data coming
from different acquisition systems. By exploiting the fact that several commercial
acquisition systems (e.g., 3Brain, Multi-Channel Systems) use the HDF5 file format
for storing and managing data, I realized a tool that, starting from such metafiles,
performs spike detection and then analysis of spiking and bursting activity. Nonethe-
less, also a specific analysis of the topological properties (i.e., connectivity proper-
ties) can be performed by exploiting the functional connectivity maps derived from
the available algorithms for inferring functional-effective connectivity (e.g., cross-
correlation, partial-correlation, joint entropy, transfer entropy). The current release
of the software presents also two main limitations: the first one is the lack of algo-
rithms to process the stimulus-evoked activity (e.g., Post-Stimulus Time Histogram,
PSTH). The second one is that all the algorithms work only on spike trains data (i.e.,
point process). On the other hand, it is worth underlining that most of the funda-
mental optimization strategies applied in the implementation of the connectivity
algorithms are based on the use of spike trains data, and therefore applicable only
because of working with this kind of input data. Finally, it is important to underline
that SpiCoDyn is available to the scientific community and it has been programmed
to be adapted, modified and extended by the interested researchers.
Appendix
Computational Model
I used computational model to assess the connectivity methods accuracies (cf.,

Fig. A1). The network model is made up of 1000 neurons randomly connected.
The dynamics of each neuron is described by the Izhikevich equations [10]. In the
actual model, two families of neurons have been taken into account: regular spiking
and fast spiking neurons for excitatory and inhibitory populations, respectively. The
ratio of excitation and inhibition has been set to 4:1 as experimentally founded in
cortical cultures [2]. In the model, each excitatory neuron receives 100 connections
from excitatory and/or inhibitory neurons, while each inhibitory neuron receives 100
input only from excitatory neurons). Autapses are not allowed. All the inhibitory
connections introduce a delay equal to 1 ms, while excitatory ones range from 1 to
20 ms. Synaptic weights are extracted from a Gaussian distribution with mean equal
to 6 and -5 for excitatory and inhibitory weights. Standard deviations have been
set to 1. Excitatory weights evolve following the spike timing dependent plasticity
(STDP) rule with a time constant equal to 20 ms [11]. The spontaneous activity of the
neuronal network has been generated by stimulating a randomly chosen neuron at
each time stamp injecting a current pulse extracted from a normal distribution (I stm,exc
= 11 ± 2; I stm,inh = 7 ± 2). The network model has been implemented in Matlab
(The Mathworks, Natick, MA, USA), and each run simulates 1 h of spontaneous
activity.
86 Appendix
Fig. A1 Computational model features and simulation results. a electrophysiological patterns of

excitatory (top) and inhibitory (bottom) neurons. b excitatory synaptic weights distribution at t = 0
(left side) and at the end of the simulation (right side). c sketch of the permitted connections among
the excitatory and inhibitory populations. d MFR distributions. e IBI distributions. Adapted from
[12]
References
1. Sporns O, Tononi G (2002) Classes of network connetivity and dynamics. Complexity 7:28–38
2. Marom S, Shahaf G (2002) Development, learning and memory in large random networks of
cortical neurons: lessons beyond anatomy. Q Rev Biophys 35:63–87
Nat Rev Neurosc 2:539–550
10.1046/j.1460-9568.1999.00770.x
7. Aertsen AMHJ, Gerstein GL (1985) Evaluation of neuronal connectivity: sensitivity of cross-
correlation. Brain Res 340:341–354. https://doi.org/10.1016/0006-8993(85)90931-X
8. Buzsaki G (2004) Large-scale recording of neuronal ensembles. Nat Neurosci 7:446–451
connectivity toolbox for in vitro networks. Front Neuroinform 10. https://doi.org/10.3389/fninf.
2016.00013
10. Izhikevich EM (2003) Simple model of spiking neurons. IEEE Trans Neural Netw 14:1569–
1572
11. Song S, Miller KD, Abbott LF (2000) Competitive Hebbian learning through spike-timing-
dependent synaptic plasticity. Nat Neurosci 3:919–926

Estimating Functional Connectivity and Topology in Large-Scale Neuronal Assemblies

Uploaded by

Copyright:

Available Formats

Estimating Functional Connectivity and Topology in Large-Scale Neuronal Assemblies

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

Estimating Functional Connectivity and Topology in Large-Scale Neuronal Assemblies

Uploaded by

Copyright:

Available Formats

Springer Theses

Recognizing Outstanding Ph.D. Research

Vito Paolo Pastore

Recognizing Outstanding Ph.D. Research

Theses are accepted into the series by invited nomination only

More information about this series at http://www.springer.com/series/8790

ISSN 2190-5053 ISSN 2190-5061 (electronic)

Genoa, Italy Prof. Sergio Martinoia

Pastore, V. P., Massobrio, P., Godjoski, A. & Martinoia, S. Excitatory-inhibitory

2.4 Metrics to Evaluate the Connectivity Methods’ Performances . . . . 24

3.6.4 Input Data Processing and Conversion Procedures . . . . . . 69

APS Active Pixel Sensor

1.1 Bridging Structure and Function of Neural Circuits

1.2 Different Types of Connectivity to Describe Neuronal

interactions can contribute to the shaping of the underlying anatomical substrate

1.3 The Relevance of the In Vitro Model

1.4 Micro Electrode Arrays

high-density configurations with thousands of microelectrodes (4000–10,000) with

1.5 How to Infer Functional Connectivity in In Vitro

In vitro networks represent an excellent benchmark for the validation of functional-

and to extract topological characteristics, it is necessary to monitor, in parallel, the

1.6 Functional Connectivity Analysis Requires Ad-Hoc

The analysis of the interactions in large-scale neuronal networks to infer functional

1.7 Goal of the Thesis

1.8 Structure of the Thesis

2.1 Correlation-Based Connectivity Methods

2.1.2 Cross-Correlation Histogram

2.1.3 Filtered and Normalized Cross-Correlation Histogram

2.1.3.1 Post Computation FNCCH Filtering

2.1.4 Partial Correlation

where Sx x (ω) and S yy (ω) are the Fourier Transform of Rx x (τ ) and R yy (τ ),

Fig. 2.4 Schematic

Sx y|Z (ω) = Sx y (ω) − Sx Z (ω)S Z−1Z (ω)S Z y (ω) (2.5)

where S Z Z (ω) is the auto-correlation of Z in the frequency domain and the

G x y (ω) C x y|Z (ω)

2.2 Information-Theory Based Connectivity Methods

2.2.1 Transfer Entropy

Transfer Entropy (TE) is an asymmetric directed information theoretic measure

2.2.2 Delayed Transfer Entropy

2.2.3 Delayed High Order Transfer Entropy

D H O T E y≥x (d) = p xt+1 , xt(k) , yt+1−d

2.2.4 Joint Entropy

The JE algorithm is based on the computation of the cross-Inter Spike Intervals

2.3 Thresholding Approaches

The statistical measures described in the previous paragraphs enable building a

full matrix of n [14] elements, and a thresholding procedure is required to throw

2.3.1 Hard Threshold

2.3.2 Shuffling Approach

Fig. 2.9 Schematic

Distribuon Dist of the connecons

Delta = mean(Dist) + n*std(Dist), n

Shuﬄing test with the desired

specified width. Thus, each spike at time t is randomly substituted by a spike at a

2.4 Metrics to Evaluate the Connectivity Methods’

2.4.1 Receiver Operating Characteristic (ROC) Curve

2.4.2 Matthews Correlation Coefficient (MCC) Curve

2.5 Graph Theory

The Average Cluster Coefficient, obtained by averaging the cluster coefficient of

This topological parameter is commonly used to evaluate the network’s level of