Available online at www.sciencedirect.

com

ScienceDirect

Hybrid modeling — a key enabler towards realizing

digital twins in biopharma?
Michael Sokolov1, Moritz von Stosch1, Harini Narayanan2, Fabian
Feidl1 and Alessandro Butté1

Digital twins (DTs) represent a vividly emerging technology in objectives can be targeted [1], from in silico process
the manufacturing industry strongly motivated by the goals of development and optimization [2], through risk mitiga-
industry 4.0. It strives for smart factories with completely tion in operations and maintenance [3], up to predictive
interconnected devices and data archives as well as their quality and quality by release [4].
automated and adaptive operation based on autonomous
decisions of smart algorithms. In this article, we want to provide The DTs can serve as a supportive digital stakeholder in
our vision on the broad potential, critical components as well as the daily workstream of several user groups, who can
current technological and organizational challenges of DTs in benefit from it in diverse directions. Scientists can ben-
biopharma. We put a profound focus on the capability of hybrid efit from DTs as a versatile counterpart in managing
modeling as a core technology to realize the capabilities of DTs process data and knowledge, supporting them to derive
in the biopharma environment, which is governed by novel insights in near-real-time, and interactively sug-
considerable levels of heterogeneity, complexity and gesting dynamic possibilities to validate them. Process
uncertainty. development teams can interact with DTs to efficiently
plan, monitor and analyze iterative campaigns for process
Addresses screening, characterization and validation. Here process
1
DataHow, Zürichstrasse 137, 8600 Dübendorf, Switzerland scientists can not only foresee to reduce timelines for
2
ETH Zurich, Institute for Chemical and Bioengineering, Vladimir-Prelog- process design, but also the ability to define better
Weg 1, 8093 Zurich, Switzerland performing and more robust processes, based on efficient
learning from all possible data sources as well as the
Corresponding author: Sokolov, Michael ([email protected])
ability to re-design this know-how into risk-based,
dynamic experimental scenarios. For scale-up and tech
Current Opinion in Chemical Engineering 2021, VOL:xx–yy transfer considerations, DTs have a unique ability to
This review comes from a themed issue on Biotechnology and bio- assess the dominant scaling challenges in an environ-
process engineering: mechanistic and data-driven modelling of ment with many technological and data source hetero-
bioprocesses geneities, so to ensure reproducible performance and
Edited by Cleo Kontoravdi and Colin Clarke robust operation at manufacturing scale. In this regard, it
is very difficult to foresee that the ultimate goal of
biopharma of real time release can be achieved without
the presence of a DT. As for process operations, alarms
https://doi.org/10.1016/j.coche.2021.100715 cannot only inform the operator about current problems,
2211-3398/ã 2021 Elsevier Ltd. All rights reserved. but DTs can be used for predictive identification of
abnormalities and inefficiencies. Further, an active man-
agement of the various controls via DTs helps to achieve
the desired process outcome and aids scheduling and
capacity planning assistance across the entire plant. The
ability to support in managing all data, knowledge and
Our motivation for DTs operations, brings DTs in a central reporting role in the
Before deep diving into technological building blocks and regulated environment providing a transparent basis for
conceptual definitions, let us start with our ambitious risk assessment. From the business landscape perspec-
vision of a digital twin (DT) in biopharma: we aim at tive, DTs are therefore connecting the end-users and
the establishment of a digital technology that can access regulatory authorities in pharma to the vendors of hard-
all available information in the process data archives ware equipment as well as software and IT solutions.
across scales and sites, and allow for real-time information This puts this technology into a central position with
exchange with the complete process control system as regards to the support of unifying and aligning data
well as all the human stakeholders involved. The DT can standards, allowing to exploit synergies between differ-
transform these data independently of their format into a ent hardware and software technologies, as well as to
relevant context, while capturing the required level of manage expectations to the potential and risks associ-
complexity and simulating various future scenarios to ated with the digital transformation of a development or
derive relevant decisions. Thereby, absolutely different manufacturing facility.

www.sciencedirect.com Current Opinion in Chemical Engineering 2019, 1:100715

2 Biotechnology and bioprocess engineering: mechanistic and data-driven modelling of bioprocesses

Figure 1

Hybrid Models Digital Twins

Architecture

Centered on process understanding Predictive capability for each process unit

Benefits

Adaptive to characteristics of process and Centered on practical application purpose such as

data heterogeneities monitoring, control and maintenance
Possible to train with limited amount of Two-way integration with physical system
know-ledge and/or data Comprehensible design to support humans
Extrapolation and transfer learning capability Transferrable architecture across scales and sites
Current Opinion in Chemical Engineering

Main building blocks and benefits of hybrid models (as digital model) and digital twins.

The economic potential of DTs can be interpreted from costly therapeutic areas such as cell and gene therapies
several attractive business cases. One can foresee [7], which are very complex with regards to the potentially
increased revenues due to improvements in productivity influential factors, could fail to be established if not
as well as clearly reduced revenue losses from failures due directly supported by an integrated smart predictive
to more robust and proactive controls. Faster develop- technology. Compared to the engagement into a new
ment cycles allow to reduce capacity utilization and to drug candidate facing a considerable attrition rate, the
perform more parallel developments of portfolio drugs. investment into such technology is likely to come at lower
With digital transformation not only a technological but financial risk [8], as it offers added value and direct
also an organizational challenge, being a tangible support- returns even if implemented for simpler objectives and
ive solution in the hands of numerous stakeholders, DTs for not completely integrated activities.
have also a strong ability to support a broad mind set shift
towards consistent value creation from smart digital solu- In order to define a DT, we want to stress three important
tions [5]. One can clearly argue that several of the listed perspectives: First, in order to predict the process evo-
business potentials can already be strived for with some of lution, a predictive component is required simulating the
the available building blocks of DTs. However, it is process behavior in the future or in unexplored regions. In
important to highlight that today only few of these other words, a DT is not singly a data management and
components are implemented consistently, often over- visualization system of the different available information
rating the technological implementation complexity and sources, but processes this information sufficiently fast to
risks compared to the direct economic benefits. There- provide the basis for decision support and control.
fore, the business model around DTs needs to be con- Hereby, it shall be stressed that complex model re-train-
sidered from a more holistic perspective, where, based on ings can be performed between process runs, while during
successful pilots, early movers will demonstrate the huge a process runs the model support must be realized fol-
synergistic potential of all integrated technologies, which lowing the dynamic needs of the considered unit opera-
is likely to prevail compared to the associated investment tion and It infrastructure. Second, DTs are designed
and implementation risks [6]. Moreover, new and very around well-defined purposes such as predictive

Current Opinion in Chemical Engineering 2019, 1:100715 www.sciencedirect.com

 Hybrid Modeling – a key enabler towards digital twins in biopharma? Sokolov et al. 3

monitoring, supervisory plant control [9], active learning In contrast to other processes, biomanufacturing is inher-
and predictive maintenance, all of which go along with ently complex due to the nature of the molecule and the
different requirements on the data sources, level of involvement of living organisms in their production.
integration with the physical system as well as predictive Subsequently, there is a lot of uncertainty (or variability)
methodology utilized. Third, DTs are built to support associated with the process which makes the develop-
humans and have to be designed as a comprehensible, ment of high-fidelity DTs an extremely complex task.
operatable and transferrable technology. Summarizing, Especially, specialized digital models capable of manag-
our pragmatic definition of DT is as follows: A DT is a ing the variability must be developed. Subsequently, it
purpose-centered interconnected digital representation should suitably account for the uncertainty while per-
of the process with all relevant specifications, user forming various decision support tasks of the DTs, be it
requirements and information sources, that supports asso- predictive alarming, optimization or maintenance.
ciated stakeholders in their decision taking and enables
direct control of the physical system. Subsequently, as Furthermore, there is no universally accepted technolog-
demonstrated in Figure 1, the central building blocks ical protocol for biomanufacturing. In other words, there
required to implement DTs are the digital model (virtual is a large diversity of operational modes and control
representation of the physical system), simulation meth- strategies used for a possibly varying sequence of pro-
odology of scenarios, near-real-time and historical data duction and purification unit operations sequence. Thus,
access, integration architecture and communication pro- the DT for biopharmaceutical industry should possess
tocols [10]. features for adaptability and transferability across various
systems.

Challenges for the realization of DTs in Additionally in a laboratory one will observe a large
biopharma diversity of apparatus and corresponding IT infrastruc-
Embracing the realization of the standards of industry ture utilized to conduct different operations. Each appa-
4.0 towards smart factories, the manufacturing industry ratus is equipped with different sensors producing data in
demonstrated a growing interest in DTs [11,12]. Since different formats, frequencies and accuracies, additional
2016 there has been an exponential rise in the publica- to the offline data quantified by the quality control
tions related to DTs [12], with the Gartner report pre- department from samples [24]. Furthermore, additional
dicting that by 2021 most of the large industrial compa- heterogeneity can be introduced when such landscape is
nies will be using DTs.3 extended to different sites, experimental and
manufacturing scales, data archives of previous develop-
Successful implementation of DTs has been demon- ments of similar molecules as well as different user groups
strated in several applications such as the aerospace such as process engineers and data scientists. From the
(NASA) [13–15],4 automotive industry [16],5,6 oil and scientific perspective, the different considered unit
gas industry [17], other manufacturing industry [18], city operations go along with different levels of prior knowl-
planning [19],7 medical and healthcare [20],8 and gastron- edge, limiting to completely understand the generated
omy [21]. The DTs are used in other areas to perform data and store them in perspective of deriving decisions
near-real-time monitoring and control, simulation, opti- for subsequent planning, operation and control activities
mization, risk analysis, maintenance and efficiency [5]. In the organizational domain, heterogeneities occur in
improvement [4]. The increasing competitive demand deviations across operators, data analysis methods and
for innovative technologies to reduce time to market and interpretations by process experts.
process development and manufacturing costs, and their
success stories in the aforementioned areas make it a Thus, DT for biopharmaceutical industry require an
promising digital technology also in the biopharmaceuti- extensive preprocessing pipeline to address data hetero-
cal industry. However, this sector faces several sources for geneity and anomaly. Some examples of tools in this
uncertainty and heterogeneity as the result of manifold pipeline include: (i) standardized nomenclature and
technological, scientific and organizational challenges units, (ii) error detection in data, (iii) outlier detection,
[22,23]. (iv) missing value handling, (v) measurement frequency
3
alignment representative for each variable, (vi) account-
https://www.gartner.com/smarterwithgartner/
ing of measurement accuracies, (vii) consistency checks
prepare-for-the-impact-of-digital-twins/.
4
https://www.thefuturefactory.com/blog/24. with available knowledge and calculated expectations.
5
https://www.xapix.io/post/digital-twinning-automotive.
6
https://www.ey.com/en_gl/advanced-manufacturing/ Additionally, unlike many other manufacturing indus-
how-digital-twins-give-automotive-companies-a-real-world-advantage. tries, biopharma does not yet fulfill the requirements of
7
https://govinsider.asia/digital-gov/
meet-virtual-singapore-citys-3d-digital-twin/.
Industry 3.0, that is, moving from manual to automated
8
https://www.dr-hempel-network.com/digital-health-technolgy/ systems [25]. This manual fingerprint can still be found
digital-twins-in-healthcare/. from sampling and data documentation in analytics to

www.sciencedirect.com Current Opinion in Chemical Engineering 2019, 1:100715

4 Biotechnology and bioprocess engineering: mechanistic and data-driven modelling of bioprocesses

control and maintenance in operations. It represents a organisms, products, and so on [33]. Model development
conservative approach to manage a complex system in a is largely either knowledge-centric (i.e. based on funda-
regulated environment without clear regulations to imple- mental, mechanistic or empirical considerations) or data-
ment DT technology [26] and highlights different levels of centric. In industry, the data centric approach prevails,
preference from the stakeholder in embracing procedural potential reasons being that (1) it is very systematic and
changes and technological advancements [27]. Moreover, delivers results in clear timelines, and (2) it is associated
the mode of data quantification for process variables and with the Quality by Design (QbD) paradigm that was
CQAs are still based on offline analytics. However, to promoted as part of FDA’s Process Analytical Technol-
ensure real-time inflow of data to the DT, online quantifi- ogy (PAT) initiative [7,34,35]. Though widely adopted,
cation methods such as spectroscopy should be embraced this approach has two distinct disadvantages: First, the
by the industry. Decreasing margins on the products as well number of experiments required to understand the
as increasing competition in various market fields must be impact of the process parameters on the system response,
met by convincing success stories and scalable business increases disproportionally sometimes even exponen-
cases to demonstrate the necessity of near-real-time moni- tially with the increase in number of process parameters,
toring, process automation as well as the corresponding that is, the curse of dimensionality. Second, the derived
management of all available data. understanding is mostly specific to the analyzed data,
limiting the transferability of knowledge.
However, the most crucial challenges, is the existence of a
digital model of the real system which is the basic In contrast the knowledge-centric approach provides a
requirements to implement DTs, first at a unit operation much better understanding of the system and many of the
level and subsequently integrating different units. Com- equations have been adopted and adapted during dec-
pared to many other process industries, it is only recently ades. Hence, much of this knowledge can be transferred
that predictive models are emerging in the biopharma- between projects, as physical laws remain valid across
ceutical industry [25], and there is not yet a standard projects. However, modeling the system in a knowledge-
guideline for their development and implementation. An centric way requires a high degree of understanding and
important scientific challenge in biomanufacturing is the experience, many times rendering the development and
variable degree of available knowledge about the under- application of these models cumbersome. This is even
lying process, which limits a purely mechanistic assess- more true the greater the number of process parameters
ment of the process in many cases. Combined with the that should be incorporated in the model, as they might
ability to manage the previously discussed challenges of interact making model development very complex [36].
process complexity and data heterogeneity [28–30], our
technological vision is the use of hybrid models for Hybrid modeling has been proposed as a pragmatic
successful implementation of DTs. approach that comes as an alternative to the knowledge
and data-centric approaches [37]. Its benefits for a QbD-
Our technological vision towards successful centric implementation of PAT has been outlined [28,38]
implementation of DTs and studied [39,40] and its role for digitalization has been
The business case of implementing DTs in biopharma reinstated [6,25,31,33].
considers hypothetically, future scenarios under which value
is added [31]. Most of these scenarios, such as what-if Hybrid models complement the knowledge backbone
scenario analysis, control, planning or scheduling, require stemming from the knowledge centric approach with
for the DT to predict the behavior of the system for certain data-driven methods, where needed. It overcomes the
conditions. In turn, the DT relies on mathematical models to shortcomings of the two original approaches by combin-
perform these predictions. Evidently for the predictions to ing their advantages, that is, the understanding and
resemble the real behavior, the models need to be faithful knowledge from the knowledge-centric approach and
representations of the system. Hence, the pertaining ques- straightforwardness of the data-centric approach. By using
tion is, how one can obtain these models [6]. the available knowledge, the behavior of essential parts of
the system can be understood. This combination of
Models are developed with a goal in mind [32], represent- knowledge sources also renders the development of
ing that part of the reality which represents their utilization hybrid models cost-effective and modular, supporting
purpose. Consequently, it is worthwhile considering the added-value guided implementations of DTs. More
whether the models can be predictive for the considered importantly though, as demonstrated in Figure 1 hybrid
scenario and how to obtain models that are predictive. modeling exhibits three key features that allow overcom-
ing the outlined challenges.
For most of the process unit operations that are comprised
in bioprocess trains, several different mathematical mod- (1) Combatting curse of dimensionality: By providing
els have been developed with varying degree of detail, structure and/or form to the interaction of system
different applications in mind and for a manifold of components, such as process parameters, the curse

Current Opinion in Chemical Engineering 2019, 1:100715 www.sciencedirect.com

 Hybrid Modeling – a key enabler towards digital twins in biopharma? Sokolov et al. 5

of dimensionality can be combatted [29,41]. Hence, horizontally (from project to project) as naturally the
the number of experiments to develop hybrid models models are modular allowing a targeted re-develop-
reduces as compared to a pure data-centric approach. ment of certain model components [31]. Conse-
This also seems particularly relevant with respect to quently, ‘generic’ digital models can be developed
the exorbitant number of process parameters and for process unit operations in which only certain
design choices whose impact on the Critical Quality specific parts, depending on the product, cell-line,
Attributes should potentially be captured. Even with process, and so on, are ‘updated’. For instance, in a
high-throughput experiment platforms it seems prac- model of a chromatographic column, taking all but the
tically infeasible to study all possible parameter com- loading material constant and knowing about the fluid
binations in a pure statistical fashion and the system- dynamic properties of the column only requires fitting
atic integration of knowledge via hybrid modeling the thermodynamic part [43]. Hence, experiments
constitutes an elegant approach to map knowledge can focus on elucidating insights into this part only.
against a generally valid backbone. Note that for a Knowledge consolidation is also furthered when
more ‘mechanistic’ modeling approach i) in light of generic modeling is performed across different uti-
the functional changes from project to project the lized apparatus and sensors independent of different
same number of experiments as for hybrid models data formats. Evidently, this should accelerate pro-
would be required to understand the changes in the cess development timelines, increase insights into the
mechanistic part, whereas ii) the effort for deriving manufacturing process and facilitate the engagement
the mechanistic model is significantly greater, except and exchange with regulators.
perhaps when symbolic regression approaches are
used [42]. These three key-features position hybrid models to
(2) Predictive capability: An additional benefit that is become extremely invaluable for personalized medicine,
indispensable for DT applications are the much- such as cell therapy, where few experiments can be
enhanced extrapolation capabilities, namely hybrid accomplished to develop the process, or continuous
models can predict beyond the tested experimental manufacturing, where the adaptive control of the process
settings when using the incorporated knowledge [29]. dynamics is of utmost importance. While in the latter
Only if the DT embedded models have good extrap- case, the typically dynamic nature of hybrid models
olation capabilities [28,43] can they be used for provides inherent frequency extrapolation properties
meaningful what-if analysis, process optimization or [29] that can further be improved by appropriate designs
process control [30]. Potential knock-on effects of [45], the case of cell therapy requires transferring the
miss-predictions from one digital unit-operation to knowledge from both the mechanistic and data-driven
the next underlines this importance, model quality or part. This could perhaps be accomplished by using Quan-
the absence of which therefore being critical to the titative structure–activity relationship (QSAR) inspired
success of every DT application. Further, a large models, here deriving the process from donor character-
number of the hybrid models published in literature istics. Alternatively, approaches that are based on the
are based on dynamic material balances wherefore similarity of process behavior [46]. In either case, hybrid
they are per se of dynamic nature [29], that is, by modeling provides a coherent conceptual backbone,
default they can predict the process evolution. The which when integrated into a digital twin can become
inherent capability to model dynamics renders the an enabling to define manufacturing strategies for such
same hybrid model to be applicable to perform dif- therapies.
ferent tasks. For instance, using the predicted
dynamic evolution of the process, advanced process Conclusion
control algorithms can be exploited to bring the For the successful implementation of digital twins in
process back on track. At the same, the very same biopharma several parties including the end users, sensor
hybrid model could be used for process optimization technology vendors, software providers and regulators
of all involved unit operations or even their combina- need to work together in order to establish a trustworthy
tion, experimental design (when to take samples to and scalable basis. Broad accessibility of all monitored
learn more, what scenarios to investigate to make the data and validation of all digital components in the
process more robust), what-if analysis or design space regulated environment are likely to be the largest chal-
characterization. It also renders hybrid modeling solu- lenges to be overcome. However, the first partners mov-
tions less prone to difficulties stemming from differ- ing in such advanced technological direction and consis-
ent experimental length, different and sampling tently implementing and utilizing such technology, are
times, and so on since the time domain is explicitly likely to receive a pronounced return on their investment,
build in the model. both from the technological as well as the organizational
(3) Knowledge transfer: Hybrid models exhibit the pos- perspectives. We believe that hybrid models are a central
sibility to transfer knowledge vertically (scale-up), as technology to make an essential step towards such broad
the models are generally scale independent [44], and realization, due to their technological ability to remain

www.sciencedirect.com Current Opinion in Chemical Engineering 2019, 1:100715

6 Biotechnology and bioprocess engineering: mechanistic and data-driven modelling of bioprocesses

robust in environments with little and heterogeneous 13. Rı́os J, Carlos J, Oliva M, Mas F: Product avatar as digital
counterpart of a physical individual product: literature review
data, as well as their pragmatic engineering knowledge and implications in an aircraft. Transdiscipl Lifecycle Anal Syst
backbone allowing to improve the human utilization 2015, 2:657-666 http://dx.doi.org/10.3233/978-1-61499-544-9-657.
readiness of such smart digital technology. 14. Fischer J, Boschert S: Next generation digital twin: an
ecosystem for mechatronic systems? IFAC-PapersOnLine
2019, 52:265-270 http://dx.doi.org/10.1016/j.ifacol.2019.11.685.
Conflict of interest statement 15. Shafto M, Conroy M, Doyle R, Glaessgen E, Kemp C, LeMoigne J,
Nothing declared. Wang L: DRAFT modeling, simulation, information technology
& processing roadmap - technology area 11. Natl Aeronaut Sp
Adm 2010:27.
References and recommended reading 16. Sharma M, George JP: Digital twin in the automotive industry:
Papers of particular interest, published within the period of review,
driving physical-digital convergence. White Pap (TATA Consult
have been highlighted as:
Serv) 2018:8.
 of special interest 17. Wanasinghe TR, Wroblewski L, Petersen BUIK, Gosine RG,
 of outstanding interest James LA, Silva ODE, Mann GKI, Warrian PJ: Digital twin for the
oil and gas industry: overview. Res Trends Oppor Challenges
1. Mabkhot MM, Al-Ahmari AM, Salah B, Alkhalefah H: 2020, 8:104175-104197 http://dx.doi.org/10.1109/
Requirements of the smart factory system: a survey and ACCESS.2020.2998723.
perspective. Machines 2018, 6 http://dx.doi.org/10.3390/
MACHINES6020023. 18. Lu Y, Liu C, Wang KI, Huang H, Xu X: Digital Twin-driven smart
manufacturing: connotation, reference model, applications
2. Liu S, Farid SS, Papageorgiou LG: Integrated optimization of and research issues. Robot Comput Manuf 2020, 61:101837
upstream and downstream processing in biopharmaceutical http://dx.doi.org/10.1016/j.rcim.2019.101837.
manufacturing under uncertainty: a chance constrained
programming approach. Ind Eng Chem Res 2016, 55:4599-4612 19. Dembski F, Wössner U, Letzgus M, Ruddat M, Yamu C: Urban
http://dx.doi.org/10.1021/acs.iecr.5b04403. digital twins for smart cities and citizens: the case study of
herrenberg, germany. Sustain 2020, 12:1-17 http://dx.doi.org/
3. Zalai D, Dietzsch C, Herwig C: Risk-based process development 10.3390/su12062307.
of biosimilars as part of the quality by design paradigm. PDA J
Pharm Sci Technol 2013, 67:569-580 http://dx.doi.org/10.5731/ 20. Tao F, Zhang M, Liu Y, Nee AYC: CIRP annals - manufacturing
pdajpst.2013.00943. technology digital twin driven prognostics and health
management for complex equipment. CIRP Ann Manuf Technol
4. Chen Y, Yang O, Sampat C, Bhalode P, Ramachandran R, 2018, 67:169-172 http://dx.doi.org/10.1016/j.cirp.2018.04.055.
 Ierapetritou M: Digital twins in pharmaceutical and
biopharmaceutical manufacturing: a literature review. 21. Karadeniz AM, Arif I, Kanak A, Ergun S: Digital twin of
Processes 2020, 8:1-33 http://dx.doi.org/10.3390/pr8091088 eGastronomic things: a case study for ice cream machines.
Provides the most up-to-date and broadest review of attempts to define 2019 IEEE Int. Symp. Circuits Syst.; IEEE: 2019:1-4 http://dx.doi.
digital twin strategies and implementation examples in biopharma. org/10.1109/ISCAS.2019.8702679.

5. Sokolov M, Feidl F, Morbidelli M, Butté A: Big data in 22. Papathanasiou MM, Kontoravdi C: Engineering challenges in
biopharmaceutical process development: vice or virtue? Chim therapeutic protein product and process design. Curr Opin
Oggi Chem Today 2018, 36:26-29 https://www.teknoscienze. Chem Eng 2020, 27:81-88 http://dx.doi.org/10.1016/j.
com/tks_article/big-data-in-biopharmaceutical-process- coche.2019.11.010.
development-vice-or-virtue/. (Accessed 30 May 2020).
23. Shukla AA, Wolfe LS, Mostafa SS, Norman C: Evolving trends in
6. Smiatek J, Jung A, Bluhmki E: Towards a digital bioprocess mAb production processes. Bioeng Transl Med 2017, 2:58-69
 replica: computational approaches in biopharmaceutical http://dx.doi.org/10.1002/btm2.10061.
development and manufacturing. Trends Biotechnol 2020,
38:1141-1153 http://dx.doi.org/10.1016/j.tibtech.2020.05.008 24. Steinwandter V, Borchert D, Herwig C: Data science tools and
Shows a very pragmatic perspective for the value created by a holistic applications on the way to Pharma 4.0. Drug Discov Today 2019,
modeling approach across all integrated biopharma unit operations to 24:1795-1805 http://dx.doi.org/10.1016/j.drudis.2019.06.005.
manage the uncertainty in drug manufacturing. 25. Narayanan H, Luna MF, von Stosch M, Cruz Bournazou MN,
7. Zobel-Roos S, Schmidt A, Uhlenbrock L, Ditz R, Köster D,  Polotti G, Morbidelli M, Butté A, Sokolov M: Bioprocessing in the
Strube J: Digital Twins in Biomanufacturing. 2020 http://dx.doi. digital age: the role of process models. Biotechnol J 2020, 15
org/10.1007/10_2020_146. http://dx.doi.org/10.1002/biot.201900172
Elaborate review of possibilities borne in consistent utilization of process
8. Varsakelis C, Dessoy S, von Stosch M, Pysik A: Show me the models for monitoring, control, process optimization and scale-up.
 money! Process modeling in pharma from the investor’s point
of view. Processes 2019, 7 http://dx.doi.org/10.3390/pr7090596 26. Bequette BW: 110th anniversary: commentary: the smart
Return on investment analysis of process modeling in biopharma reflect- human in smart manufacturing. Ind Eng Chem Res 2019,
ing monetary, innovation and organizational perspectives. 58:19317-19321 http://dx.doi.org/10.1021/acs.iecr.9b03544.

9. Sommeregger W, Sissolak B, Kandra K, von Stosch M, Mayer M, 27. Zobel-Roos S, Schmidt A, Mestmäcker F, Mouellef M, Huter M,
Striedner G: Quality by control: towards model predictive  Uhlenbrock L, Kornecki M, Lohmann L, Ditz R, Strube J:
control of mammalian cell culture bioprocesses. Biotechnol J Accelerating biologics manufacturing by modeling or: is
2017, 12:1-7 http://dx.doi.org/10.1002/biot.201600546. approval under the QbD and PAT approaches demanded by
authorities acceptable without a digital-twin? Processes 2019,
10. Kritzinger W, Karner M, Traar G, Henjes J, Sihn W: Digital twin in 7:1-28 http://dx.doi.org/10.3390/pr7020094
manufacturing: a categorical literature review and Based on a detailed conceptual investigation and a deep-dive mini-plant
classification. IFAC-PapersOnLine 2018, 51:1016-1022 http://dx. example the central role of a digital twin for technology management and
doi.org/10.1016/j.ifacol.2018.08.474. validation in biopharma is presented.
11. Fernando L, Dur CS, Haag S, Anderl R, Sch K, Zancul E: Digital 28. Narayanan H, Sokolov M, Morbidelli M, Butté A: A new generation
Twin Requirements in the Context of Industry 4 .0, 1. 2018:204-214 of predictive models: the added value of hybrid models for
http://dx.doi.org/10.1007/978-3-030-01614-2. manufacturing processes of therapeutic proteins. Biotechnol
Bioeng 2019, 116:2540-2549 http://dx.doi.org/10.1002/bit.27097.
12. Pires F, Cachada A, Barbosa J, Moreira AP, Leitão P: Digital twin
in industry 4.0: technologies, applications and challenges. 29. Glassey J, von Stosch M (Eds): Hybrid Modeling in Process
2019 IEEE 17th Int. Conf. Ind. Informatics 2019:721-726 http://dx. Industries, edn 1. CRC Press; 2018 http://dx.doi.org/10.1201/
doi.org/10.1109/INDIN41052.2019.8972134. 9781351184373.

Current Opinion in Chemical Engineering 2019, 1:100715 www.sciencedirect.com

 Hybrid Modeling – a key enabler towards digital twins in biopharma? Sokolov et al. 7

30. Narayanan H, Behle L, Luna MF, Sokolov M, Guillén-Gosálbez G, 38. von Stosch M, Davy S, Francois K, Galvanauskas V, Hamelink J-M,
Morbidelli M, Butté A: Hybrid-EKF: hybrid model coupled with Luebbert A, Mayer M, Oliveira R, O’Kennedy R, Rice P, Glassey J:
extended Kalman filter for real-time monitoring and control of Hybrid modeling for quality by design and PAT-benefits and
mammalian cell culture. Biotechnol Bioeng 2020, 117:2703-2714 challenges of applications in biopharmaceutical industry.
http://dx.doi.org/10.1002/bit.27437. Biotechnol J 2014, 9 http://dx.doi.org/10.1002/biot.201300385.
31. Sokolov M: Decision making and risk management in 39. von Stosch M, Hamelink J-M, Oliveira R: Hybrid modeling as a
biopharmaceutical engineering - opportunities in the age of QbD/PAT tool in process development: an industrial E. coli
Covid-19 and digitalization. Ind Eng Chem Res 2020, 59:17587- case study. Bioprocess Biosyst Eng 2016, 39:773-784 http://dx.
17592 http://dx.doi.org/10.1021/acs.iecr.0c02994. doi.org/10.1007/s00449-016-1557-1.

32. Bonvin D, Georgakis C, Pantelides CC, Barolo M, Grover MA, 40. Campa C, Khan MA (Eds). Quality by Design - An Indispensable
Rodrigues D, Schneider R, Dochain D: Linking models and Approach to Accelerate Biopharmaceutical Product Development
experiments. Ind Eng Chem Res 2016, 55:6891-6903 http://dx. (Single User Digital Version). 2021.
doi.org/10.1021/acs.iecr.5b04801.
41. Fiedler B, Schuppert A: Local identification of scalar hybrid
33. Cardillo AG, Castellanos MM, Desailly B, Dessoy S, Mariti M, models with tree structure. IMA J Appl Math (Institute Math Appl)
Portela RMC, Scutella B, von Stosch M, Tomba E, Varsakelis C: 2008, 73:449-476 http://dx.doi.org/10.1093/imamat/hxn011.
Towards in silico process modeling for vaccines. Trends 42. Willis MJ, von Stosch M: Simultaneous parameter identification
Biotechnol 2021 http://dx.doi.org/10.1016/j.tibtech.2021.02.004. and discrimination of the nonparametric structure of hybrid
in press. semi-parametric models. Comput Chem Eng 2017, 104:366-376
34. Mercier SM, Diepenbroek B, Wijffels RH, Streefland M: http://dx.doi.org/10.1016/J.COMPCHEMENG.2017.05.005.
Multivariate PAT solutions for biopharmaceutical cultivation: 43. Narayanan H, Seidler T, Luna MF, Sokolov M, Morbidelli M,
current progress and limitations. Trends Biotechnol 2014, Butté A: Hybrid models for the simulation and prediction of
32:329-336 http://dx.doi.org/10.1016/j.tibtech.2014.03.008. chromatographic processes for protein capture. J Chromatogr
A (Under Rev) 2021, 1650:462248.
35. Simon LL, Pataki H, Marosi G, Meemken F, Hungerbuehler K,
Baiker A, Tummala S, Glennon B, Kuentz M, Steele G et al.: 44. Richelle A, Lee BW, Portela RMC, Raley J, Stosch M: Analysis of
Assessment of recent process analytical technology (PAT) transformed upstream bioprocess data provides insights into
trends: a multiauthor review. Org Process Res Dev 2015, 19:3-62 biological system variation. Biotechnol J 2020, 15:2000113
http://dx.doi.org/10.1021/op500261y. http://dx.doi.org/10.1002/biot.202000113.
36. Shahriari B, Swersky K, Wang Z, Adams RP, De Freitas N: Taking 45. Bayer B, Striedner G, Duerkop M: Hybrid modeling and
the human out of the loop: a review of Bayesian optimization. intensified DoE: an approach to accelerate upstream process
Proc IEEE 2016, 104:148-175 http://dx.doi.org/10.1109/ characterization. Biotechnol J 2020:1-10 http://dx.doi.org/
JPROC.2015.2494218. 10.1002/biot.202000121. 2000121.
37. Schubert J, Simutis R, Dors M, Havlik I, Lübbert A: Bioprocess 46. Hutter C, von Stosch M, Bournazou MNC, Butté A: Knowledge
optimization and control: application of hybrid modelling. J Transfer Across Cell Lines Using Hybrid Gaussian Process Models
Biotechnol 1994, 35:51-68 http://dx.doi.org/10.1016/0168-1656 With Entity Embedding Vectors. 2020. https://arxiv.org/abs/
(94)90189-9. 2011.13863v1. (Accessed 11 July 2021).

www.sciencedirect.com Current Opinion in Chemical Engineering 2019, 1:100715