www.nature.

com/npjsba

PERSPECTIVE OPEN

Towards a widespread adoption of metabolic modeling tools in

biopharmaceutical industry: a process systems biology
engineering perspective
Anne Richelle 1 ✉, Blandine David 1, Didier Demaegd1, Marianne Dewerchin1, Romain Kinet1, Angelo Morreale1, Rui Portela 1
,
Quentin Zune1 and Moritz von Stosch1 ✉

In biotechnology, the emergence of high-throughput technologies challenges the interpretation of large datasets. One way to
identify meaningful outcomes impacting process and product attributes from large datasets is using systems biology tools such as
metabolic models. However, these tools are still not fully exploited for this purpose in industrial context due to gaps in our
knowledge and technical limitations. In this paper, key aspects restraining the routine implementation of these tools are
highlighted in three research fields: monitoring, network science and hybrid modeling. Advances in these fields could expand the
current state of systems biology applications in biopharmaceutical industry to address existing challenges in bioprocess
development and improvement.
npj Systems Biology and Applications (2020)6:6 ; https://doi.org/10.1038/s41540-020-0127-y
1234567890():,;

INTRODUCTION methods exist to describe and quantify the reactions occurring in

The emergence of high-throughput technologies is elevating the a cell under specific environments, from deterministic kinetic
biotechnology field to the era of big data1. This shifting paradigm models to stochastic and statistical models. While kinetic
created considerable challenges for the interpretation of large modeling has been widely used for small-scale metabolic models,
datasets and the generation of meaningful outcomes impacting these mechanistic models are for the moment not scalable and
process and product attributes. Among the numerous computa- suitable for genome-wide approaches due to challenges in
tional methods developed in this context, metabolic modeling acquiring the required kinetic parameters (i.e., rate constants,
tools allowed the coherent organization of large datasets into enzymes and intracellular metabolite concentrations) and com-
biological networks providing nonintuitive insights on biological putational problems associated with such complex parametric
systems that in vivo experiments alone cannot provide2. These systems (i.e., model nonlinearity, parameter identifiability, and
model-based approaches have been proven to be invaluable at computational tractability)10. On the other hand, while statistical
the level of preclinical research. For example, these approaches modeling is very useful for the manipulation of large complex
have been used for the design of new drugs by informing target datasets, these black-box approaches have limited capacity to
selection and for the engineering of cells by rewiring metabolism generate biologically relevant information for bioprocess improve-
towards the production of a product of interest3,4. ment. Therefore, the lack of clear workflows to set up specific
Pharmaceutical companies are already investing substantially in process models based on genome-wide information have limited
computational approaches to guide drug discovery and cell the potential scope of computational biology applications in
design. While these important applications are out of the scope of industry.
this paper, we refer the interested readers to detailed reviews on However, the systems biology field is in continuous develop-
the usage of metabolic engineering and synthetic biology in ment and genome-scale metabolic networks are now available for
industrial context5–8. However, model-based methods can be used most of the industrially relevant organisms. Therefore, we should
for much more than this. For instance, these methods could also be able to systematically predict cell metabolic behavior under
be applied at the industrial level in the field of process design, different specific complex media compositions and process
monitoring and control to lower the experimental effort and conditions. These types of predictions should be used to
increase the process robustness. Indeed, model-based methods objectively determine the optimal operating conditions with
would allow a more rational design of bioprocesses but also the respect to the desired cell phenotype and related production
objective identification of the variables to monitor and control. criteria. Unfortunately, while numerous studies have successfully
Therefore, the routine implementation of these methods can be used genome-scale models to rationally guide the culture process
expected to greatly facilitate the implementation of regulatory design11–13, their practical implementation in biopharmaceutical
requirements related to the Quality by Design paradigm (i.e., research and development facilities still seems in its infancy.
systematic approach to development that begins with predefined Actually, bioprocess improvement performance is still mainly
objectives and emphasizes product and process understanding achieved by semi-empirical media and bioprocess optimization
and process control) and Process Analytical Technologies initiative (i.e., screening different media and process conditions and analysis
(mechanism to design, analyze and control manufacturing using statistical design-of-experiment strategies)14. In this paper,
processes through the measurement of Critical Process Para- gaps in current knowledge and technical limitations restraining
meters that affect Critical Quality Attributes9). Numerous modeling the widespread usage of metabolic modeling tools for process

GSK, Rixensart, Belgium. ✉email: [email protected]; [email protected]

1

Published in partnership with the Systems Biology Institute

 A. Richelle et al.
2
development (i.e., design, optimization and control) in biophar- flow cytometry with a bioreactor and its use for effective
maceutical industry are highlighted. Specifically, research topics in bioreactor control can be found in the literature29.
three fields (real-time monitoring, biological network science and
hybrid modeling) are identified as key drivers for evolving the
current state of some systems biology tools in industrial process COMPLEXITY OF METABOLIC NETWORKS: REDUCE TO BETTER
engineering applications. Advances in these fields will be PREDICT
invaluable to address existing challenges in bioprocess develop- A genome-scale metabolic model is a network connecting all
ment and improvement. metabolic reactions that can occur in a specific organism with
their associated metabolites, proteins and genes. While these
models have been proven to be powerful tools for in silico
REAL-TIME BIOPROCESS MONITORING: TRACK TO BETTER simulation of cell metabolism12, the complexity of these large
CONTROL networks also hinders their utility in various practical applications.
Real-time monitoring is a key issue for effective bioprocess Actually, metabolic networks involving, most of the time, more
operation. Indeed, one needs first to recognize and analyze reactions than metabolites, the associated system presents,
cellular physiological state in function of environmental conditions generally, a multitude of solutions (i.e., underdetermined system).
to be further able to adequately manipulate operating conditions. Therefore, metabolic network structures are also too complex to
Therefore, real-time monitoring technologies development is a be handled for the development of utilitarian tools (e.g.,
critical aspect of the bioprocess automatization. Previously, only a optimization and control of a bioprocess based on cell phenotype
small number of so-called “process variables” were commonly prediction).
measured online in bioreactors (e.g., pH, temperature, OD, Different approaches have been proposed for tailoring these
stirring). The scope of available real-time probes has now been
metabolic networks based on a priori knowledge and/or available
significantly expanded thanks to the recent advances in analytical
experimental data. For example, algorithms like the one presented
technologies and the adoption of Process Analytical Technologies
initiative15,16. in Erdrich et al.30 or Ataman et al.31 allow an easy-to-implement
In the context of bioprocesses, real-time measurements of the systematic reduction of genome-scale networks into core models.
1234567890():,;

metabolites and cell concentrations in the cell culture medium are However, these methods rely on the definition by the user of the
key elements to enable the prediction of cell phenotype with part of the model that should be “protected” during the reduction
respect to operating conditions (e.g., by using constraint-based (i.e., specific metabolites, reactions, phenotypes that need to be
modeling approaches)17. Unfortunately, some of these experi- conserved) which is far from easy task. On the other hand,
mental measurements cannot currently be achieved online in situ numerous algorithms have been developed to tailor genome-
and, doing so, limit our capacity to have a live snapshot of the cell scale metabolic networks by integrating diverse types of -omics
metabolism during the culture process. However, continuous data, including transcriptomics, proteomics and metabolomics32–35.
improvements in analytical technologies and the emergence of Actually, genome-scale models are used to link genes, enzymes,
the single-cell experiments open new perspectives for gaining and metabolism through the use of gene-protein-reactions (GPR).
insights into cell metabolism. Therefore, the knowledge of GPR is used by researchers to explain
Spectroscopic methods (e.g., NIR, FTIR, mid-IR, Raman, Fluores- the metabolic state of a cell, based on the expression of metabolic
cence) have attracted an increasing interest for their potential genes and/or specific proteins36,37. The most common “integration
capacity of simultaneous in situ measurement of nutrients, algorithms” use gene expression data to recapitulate the
metabolites and cell concentrations. These techniques have been metabolism of an organism under a specific condition (e.g.,
successfully integrated into real-time pharmaceutical manufactur- specific cell-line or operating conditions) by only extracting the
ing to provide a “molecular fingerprint” of samples enabling in- subset of active reactions from the genome-scale model. These
process correction needed to ensure the acquisition of a specific methods have been proven to be valuable for enhancing the
quality product18,19. However, the capacity to systematically accuracy of model-predicted growth rates and gene essentiality37.
extract accurate quantitative data is, for the moment, limited to However, since currently no quantitative description of the GPR
a handful number of metabolites (typically, glucose, lactate, relationship exists in genome-scale metabolic models, the
glutamine, glutamate, and ammonium). We refer the reader to the integration of gene expression data requires the use of strong
publications of Li et al.20 and Ryder21 for detailed comparison of assumptions to link the GPR expression and the metabolic
the performance and limitations of existing spectroscopic reaction activity, which could lead to an oversimplification of
methods. In this context, chemometric modeling has been a the complex relation existing between fluxes, enzymes and
state-of-art technique to effectively extract the maximum of genes36,38.
information from spectroscopic spectra. It is worth to highlight Finally, it is important to highlight that these “network tailoring”
that spectroscopic sensors can in this sense be embedded in the
approaches generally do not completely leverage the problem of
concept of soft sensor22 as they are used to compensate the lack
system underdetermination related to network complexity. There-
of specific measurement by reconstructing the missing signals
based on available measurements using data analysis models23. fore, the choice of adequate strategy to solve the system (e.g.,
This type of hybrid approach will therefore continue to greatly linear optimization techniques such as Flux Balance Analysis39) will
benefit from the flourishing evolution of artificial intelligence tools always be required to achieve an instantaneous picture of the flux
and certainly allows to overcome the current limitations of distributions in the cell. A recent research in this field presented a
spectroscopic monitoring methods24. new approach using biochemical thermodynamic constraints (i.e.,
Parallelly, the rise of online single‐cell probing opens the door Gibbs energy) to shape the solution space of potential flux
for the specific investigation of cell metabolism and bioenergetics distributions40. The promising results obtained in this study pave a
at a definition level never captured before25. These “single-cell new way towards the understanding of the mechanisms that
approaches” in combination with adequate analytical methods governs the definition of metabolic pathway usage across
could help unravel the impact of operating conditions on the organisms and conditions. Further development of these compu-
cellular physiology. Specifically, the new generation of high- tational methods will be crucial to obtain a holistic and integrated
throughput-omics technologies allow to envision in a close future picture of the cell metabolism in function of specific environ-
a near real-time measurements of cell transcriptome and mental conditions while considering uncertainty associated with
proteome26–28. Interestingly, very few examples of interfacing experiments, measurements and modeling.

npj Systems Biology and Applications (2020) 6 Published in partnership with the Systems Biology Institute
 A. Richelle et al.
3
Genome scale
metabolic network

Tailoring networks
based on data

Real-time AI & ML
monitoring tools
Metabolome
Proteome
Transcriptome
Reduced
metabolic network

Hybrid modeling

Fig. 1 From experimental data to bioprocess improvement. Systematic workflow using data extracted from real-time monitoring to tailor
genome-scale biological networks to core metabolic models that can be combined with artificial intelligence and machine learning tools for
an effective implementation of control and optimization strategies.

MODELING WITH HYBRID APPROACHES: COMBINE TO Therefore, these approaches can be used to effectively establish
EFFECTIVELY IMPLEMENT the link between metabolism and operating conditions in a way
The application of artificial intelligence and machine learning in that can be used to implement control and optimization
bioprocess engineering and systems biology has seen a significant strategies.
increase in the last years thanks to the research advances in these Despite their benefits and their first applications stemming from
fields41–43 and their successful applications for e.g. in tumor 199246, the first application of hybrid modeling in systems biology
detection44. In particular, advances in image recognition enable only arose in 201047 and few have been reported ever since48–53.
the automation of so far manual tasks, such as the possibility to One reason might be the absence of dedicated software that
automate the counting of cell colonies grown on petri-dishes, would allow for a straightforward development of these type of
using digital imaging45. Surprisingly, the practical implementation models. Another reason might be the broad interdisciplinarity of
of these hybrid approaches combining machine learning and the competences required to implement these approaches (i.e.,
extended knowledge available about biological systems is still in knowledge from engineering fields and data sciences that need
its infancy in bioindustries. ideally to be complemented with a biological background). Finally,
Considering the availability of this prior knowledge, hybrid there is still several major open questions that need to be
modeling and artificial intelligence approaches have the potential investigated to allow a better combined usage of hybrid modeling
to significantly reduce the amount of experiments that need to be and metabolic modeling tools54. From the effective generation of
executed, while potentially increasing the region in which the sufficiently informative experimental data to the joint parameter
model can reliably generate predictions. Actually, hybrid models identification of the mechanistic and data-driven parts of the
seek to complement what is mechanistically known (e.g., about model, the assessment of these gaps will be critical for the
the metabolism and kinetics) with data-driven methods to development of a systematic workflow for integrating knowledge
describe the unknown parts. This type of models provides an from systems biology into model structures enabling a next-
attractive method for modeling biochemical processes as much of generation technology for the life-sciences and biotechnology
the underlying complexity can be lumped in the data-driven part. sectors, besides others (Fig. 1).

Published in partnership with the Systems Biology Institute npj Systems Biology and Applications (2020) 6
 A. Richelle et al.
4
BRIDGING THE GAPS TOWARDS EFFECTIVE CONTINUOUS processing in biopharmaceutical industry but also foresee needs
PROCESSING for improvements in emergent biomanufacturing platforms such
Process intensification through continuous processing has been as cell-free systems.
adopted for decades in industries such as chemical and
petrochemical. In the context of biopharmaceutical production, Received: 10 October 2019; Accepted: 12 February 2020;
the move from batch to continuous or semi-continuous mode of
production has been slowed down mainly due to the more
conservative approach of this industry resulting from the stringent
quality/regulatory requirements. Therefore, innovation in biotech-
nology has become more product- than process-centered55. While REFERENCES
these old drivers of bioprocess development are still valid, new 1. Sadowski, M. I., Grant, C. & Fell, T. S. Harnessing QbD, programming languages,
emerging challenges are forcing the change towards extended and automation for reproducible biology. Trends Biotechnol. 34, 214–227 (2016).
2. Chen, C., Le, H. & Goudar, C. T. Integration of systems biology in cell line and
manufacturing flexibility and transferability. The commitment to
process development for biopharmaceutical manufacturing. Biochem. Eng. J. 107,
deliver on them includes process development timeline decrease, 11–17 (2016).
productivity increase and product quality modulation (e.g., for 3. Kubinyi, H. In Computer Applications in Pharmaceutical Research and Development
biosimilarity). (eds Wang, B. & Ekins, S.) 377–424 (John Wiley & Sons, 2006).
The proper implementation of continuous manufacturing 4. Schmidt, B. J., Papin, J. A. & Musante, C. J. Mechanistic systems modeling to guide
requires a thorough understanding of the process to achieve drug discovery and development. Drug Discov. Today 18, 116–127 (2013).
and maintain steady-state conditions (i.e., the causalities between 5. Choi, K. R. et al. Systems metabolic engineering strategies: integrating systems
physiological state, bioprocess parameters, productivity and and synthetic biology with metabolic engineering. Trends Biotechnol. 37, 817–837
(2019).
product quality attributes). Furthermore, the in-process quality
6. Nielsen, J. & Keasling, J. D. Engineering cellular metabolism. Cell 164, 1185–1197
should be evaluated and ensured by the real-time analysis of (2016).
process data. To this end, in-line process monitoring system linked 7. Jain, K. K. Synthetic biology and personalized medicine. Med. Princ. Pract. 22,
to robust model-based control will be crucial technologies to 209–219 (2013).
support this real-time release testing56,57. 8. Trosset, J.-Y. & Carbonell, P. Synthetic biology for pharmaceutical drug discovery.
Despite their advantages, the potential of metabolic modeling Drug Des. Devel. Ther. 9, 6285–6302 (2015).
tools is still not fully exploited to address these challenges in 9. Rathore, A. S. & Winkle, H. Quality by design for biopharmaceuticals. Nat. Bio-
bioprocess development and improvement. This is mainly due to technol. 27, 26–34 (2009).
10. Penas, D. R., González, P., Egea, J. A., Doallo, R. & Banga, J. R. Parameter estimation
the upfront time and effort investment required to implement
in large-scale systems biology models: a parallel and self-adaptive cooperative
these approaches and gain in process understanding as critical strategy. BMC Bioinforma. 18, 52 (2017).
gaps remain to set these methods up in clear workflows. Empirical 11. Burgard, A. P., Vaidyaraman, S. & Maranas, C. D. Minimal reaction sets for
process development techniques will continue to play an Escherichia coli metabolism under different growth requirements and uptake.
important role in addressing these challenges. However, systems Environments 17, 791–797 (2001).
biology-driven approaches are likely to be more impactful to 12. Selvarasu, S. et al. Combined in silico modeling and metabolomics analysis to
provide insights into the cellular responses to process conditions characterize fed-batch CHO cell culture. Biotechnol. Bioeng. 109, 1415–1429 (2012).
changes2,14. 13. Zhang, C. & Hua, Q. Applications of genome-scale metabolic models in bio-
technology and systems medicine. Front. Physiol. https://doi.org/10.3389/
Finally, the research topics, highlighted in this review, in the
fphys.2015.00413 (2016).
field of monitoring, network science and hybrid modeling, will 14. Kroll, P., Hofer, A., Ulonska, S., Kager, J. & Herwig, C. J. P. R. Model-based methods
also be cutting-edge technologies to improve new emergent in the biopharmaceutical process lifecycle. Farm Res. 34, 2596–2613 (2017).
biopharmaceutical production platforms such as cell-free systems. 15. Gernaey, K. V. et al. Monitoring and control of microbioreactors: an expert opi-
Actually, cell-free protein synthesis has been emerging as a nion on development needs. Biotechnol. J. 7, 1308–1314 (2012).
flexible and powerful platform to address challenges in biomanu- 16. Guerra, A., von Stosch, M. & Glassey, J. Toward biotherapeutic product real-time
facturing by overcoming inherent limitations related to the use of quality monitoring. Crit. Rev. Biotechnol. 39, 289–305 (2019).
living cells. Cell-free systems offer the ability to design metabolic 17. Bordbar, A., Monk, J. M., King, Z. A. & Palsson, B. O. Constraint-based models
predict metabolic and associated cellular functions. Nat. Rev. Genet. 15, 107–120
pathways towards the production of desired products but also to
(2014).
synthesize complex proteins with unnatural amino acids and to 18. Baradez, M.-O., Biziato, D., Hassan, E. & Marshall, D. Application of Raman spec-
buildup artificial cells. Despite these promising features, chal- troscopy and univariate modelling as a process analytical technology for cell
lenges remain such as the proper control of post-translational therapy bioprocessing. Front. Med. (Lausanne) 5, 47–47 (2018).
modifications and the expansion of the genetic code for unnatural 19. Buckley, K. & Ryder, A. G. Applications of Raman spectroscopy in biopharma-
amino acids incorporation. To address these challenges, targeted ceutical manufacturing: a short review. Appl. Spectrosc. 71, 1085–1116 (2017).
gene editing and addition of adequate exogenous substances to 20. Li, M., Ebel, B., Chauchard, F., Guédon, E. & Marc, A. Parallel comparison of in situ
control reaction conditions will be needed to optimally regulate Raman and NIR spectroscopies to simultaneously measure multiple variables
toward real-time monitoring of CHO cell bioreactor cultures. Biochem. Eng. J. 137,
the transcription and translation58,59. In this context, computa- 205–213 (2018).
tional tools will be invaluable to systematically identify system 21. Ryder, A. G. Cell culture media analysis using rapid spectroscopic methods. Curr.
limitations and areas of improvement for production efficiency, as Opin. Chem. Eng. 22, 11–17 (2018).
demonstrated in the work of Vilkhovoy et al.60. 22. Luttmann, R. et al. Soft sensors in bioprocessing: a status report and recom-
mendations. Biotechnol. J. 7, 1040–1048 (2012).
23. Esmonde-White, K. A. et al. Raman spectroscopy as a process analytical tech-
DISCUSSION nology for pharmaceutical manufacturing and bioprocessing. Anal. Bioanal.
We highlighted gaps in our knowledge and technical limitations Chem. 409, 637–649 (2017).
24. Storey, E. E. & Helmy, A. S. Optimized preprocessing and machine learning
restraining the effective application of metabolic modeling tools
quantitative Raman spectroscopy in biology. J. Raman spectrosc. 50, 958–968
throughout the lifecycle of a biopharmaceutical process. Specifi- (2019).
cally, advances in real-time monitoring of bioprocesses, biological 25. Vasdekis, A. E. & Stephanopoulos, G. Review of methods to probe single cell
network modeling and their combination with data-driven metabolism and bioenergetics. Metab. Eng. 27, 115–135 (2015).
approaches will be the key drivers to lower the time and cost 26. Erickson, B. K. et al. Active instrument engagement combined with a real-time
associated with the development of new drugs4. Taken together, database search for improved performance of sample multiplexing workflows. J.
these technologies could facilitate the switch to continuous Proteome Res. 18, 1299–1306 (2019).

npj Systems Biology and Applications (2020) 6 Published in partnership with the Systems Biology Institute
 A. Richelle et al.
5
27. Doerr, A. Single-cell proteomics. Nat. Methods 16, 20–20 (2019). 54. von Stosch, M., Oliveira, R., Peres, J. & Feyo de Azevedo, S. Hybrid semi-parametric
28. Rhoads, A. & Au, K. F. PacBio Sequencing and its applications. Genom. Proteom. modeling in process systems engineering: past, present and future. Computers
Bioinforma. 13, 278–289 (2015). Chem. Eng. 60, 86–101 (2014).
29. Delvigne, F. & Goffin, P. Microbial heterogeneity affects bioprocess robustness: 55. Warikoo, V. et al. Integrated continuous production of recombinant therapeutic
dynamic single-cell analysis contributes to understanding of microbial popula- proteins. Biotechnol. Bioeng. 109, 3018–3029 (2012).
tions. Biotechnol. J. 9, 61–72 (2014). 56. Chaudary, R. S., Pazhayattil, A. & Spes, J. Continuous manufacturing: a generic
30. Erdrich, P., Steuer, R. & Klamt, S. An algorithm for the reduction of genome-scale industry perspective. Pharmaceut. Technol. http://www.pharmtech.com/
metabolic network models to meaningful core models. BMC Syst. Biol. 9, 48 continuous-manufacturing-generic-industry-perspective (2017).
(2015). 57. Hernandez, R. Continuous manufacturing: a changing processing paradigm. Bio-
31. Ataman, M., Hernandez Gardiol, D. F., Fengos, G. & Hatzimanikatis, V. redGEM: Pharm. Int. 28. http://www.biopharminternational.com/continuous-manufacturing-
systematic reduction and analysis of genome-scale metabolic reconstructions for changing-processing-paradigm (2015).
development of consistent core metabolic models. PLoS Comput. Biol. 13, 58. Jiang, L., Zhao, J., Lian, J. & Xu, Z. Cell-free protein synthesis enabled rapid pro-
e1005444 (2017). totyping for metabolic engineering and synthetic biology. Synth. Syst. Biotechnol.
32. Aurich, M. K., Fleming, R. M. T. & Thiele, I. MetaboTools: a comprehensive toolbox 3, 90–96 (2018).
for analysis of genome-scale metabolic models. Front. Physiol. 7, 327–327 (2016). 59. Lu, Y. Cell-free synthetic biology: engineering in an open world. Synth. Syst.
33. Berger, B., Peng, J. & Singh, M. Computational solutions for omics data. Nat. Rev. Biotechnol. 2, 23–27 (2017).
Genet. 14, 333–346 (2013). 60. Vilkhovoy, M. et al. Sequence specific modeling of E. coli cell-free protein
34. Bersanelli, M. et al. Methods for the integration of multi-omics data: mathematical synthesis. ACS Synth. Biol. 7, 1844–1857 (2018).
aspects. BMC Bioinforma. 17, S15 (2016).
35. García Martín, H. et al. A method to constrain genome-scale models with 13C
labeling data. PLoS Comput. Biol. 11, e1004363 (2015). ACKNOWLEDGEMENTS
36. Machado, D. & Herrgård, M. Systematic evaluation of methods for integration of The development of this article was sponsored by GlaxoSmithKline Biologicals SA. All
transcriptomic data into constraint-based models of metabolism. PLoS Comput. authors were employees of the GSK group of companies at the time of the
Biol. 10, e1003580 (2014). development of this article.
37. Opdam, S. et al. A systematic evaluation of methods for tailoring genome-scale
metabolic models. Cell Syst. 4, 318–329.e316 (2017).
38. Richelle, A., Joshi, C. & Lewis, N. E. Assessing key decisions for transcriptomic data
AUTHOR CONTRIBUTIONS
integration in biochemical networks. PLoS Comput. Biol. 15, e1007185 (2019).
39. Orth, J. D., Thiele, I. & Palsson, B. Ø. What is flux balance analysis? Nat. Biotechnol. A.R. and M.v.S. wrote and edited the manuscript. All authors were involved in drafting
28, 245–248 (2010). the content and structure of the manuscript and approved the final version.
40. Niebel, B., Leupold, S. & Heinemann, M. An upper limit on Gibbs energy dis-
sipation governs cellular metabolism. Nat. Metab. 1, 125–132 (2019).
41. Angermueller, C., Pärnamaa, T., Parts, L. & Stegle, O. Deep learning for compu- COMPETING INTERESTS
tational biology. Mol. Syst. Biol. 12, 878 (2016). The authors declare no competing interests.
42. Miotto, R., Wang, F., Wang, S., Jiang, X. & Dudley, J. T. Deep learning for healthcare:
review, opportunities and challenges. Brief. Bioinforma. 19, 1236–1246 (2018).
43. Oyetunde, T., Bao, F. S., Chen, J.-W., Martin, H. G. & Tang, Y. J. Leveraging ADDITIONAL INFORMATION
knowledge engineering and machine learning for microbial bio-manufacturing. Supplementary information is available for this paper at https://doi.org/10.1038/
Biotechnol. Adv. 36, 1308–1315 (2018). s41540-020-0127-y.
44. Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural
networks. Nature 542, 115–118 (2017). Correspondence and requests for materials should be addressed to A.R. or M.v.S.
45. Ferrari, A., Lombardi, S. & Signoroni, A. Bacterial colony counting with convolu-
tional neural networks in digital microbiology imaging. Pattern Recogn. 61, Reprints and permission information is available at http://www.nature.com/
629–640 (2017). reprints
46. Psichogios, D. C. & Ungar, L. H. A hybrid neural network-first principles approach
to process modeling. AIChE J. 38, 1499–1511 (1992). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims
47. von Stosch, M., Peres, J., de Azevedo, S. F. & Oliveira, R. Modelling biochemical in published maps and institutional affiliations.
networks with intrinsic time delays: a hybrid semi-parametric approach. BMC Syst.
Biol. 4, 131 (2010).
48. Carinhas, N. et al. Hybrid metabolic flux analysis: combining stoichiometric and
statistical constraints to model the formation of complex recombinant products.
BMC Syst. Biol. 5, 34–34 (2011). Open Access This article is licensed under a Creative Commons
49. Ferreira, A. R. et al. Projection to latent pathways (PLP): a constrained projection Attribution 4.0 International License, which permits use, sharing,
to latent variables (PLS) method for elementary flux modes discrimination. BMC adaptation, distribution and reproduction in any medium or format, as long as you give
Syst. Biol. 5, 181–181 (2011). appropriate credit to the original author(s) and the source, provide a link to the Creative
50. Hamilton, F., Lloyd, A. L. & Flores, K. B. Hybrid modeling and prediction of Commons license, and indicate if changes were made. The images or other third party
dynamical systems. PLoS Comput. Biol. 13, e1005655 (2017). material in this article are included in the article’s Creative Commons license, unless
51. Isidro, I. A. et al. Hybrid metabolic flux analysis and recombinant protein pre- indicated otherwise in a credit line to the material. If material is not included in the
diction in Pichia pastoris X-33 cultures expressing a single-chain antibody frag- article’s Creative Commons license and your intended use is not permitted by statutory
ment. Bioprocess. Biosyst. Eng. 39, 1351–1363 (2016). regulation or exceeds the permitted use, you will need to obtain permission directly
52. Teixeira, A. P. et al. Cell functional enviromics: unravelling the function of from the copyright holder. To view a copy of this license, visit http://creativecommons.
environmental factors. BMC Syst. Biol. 5, 92 (2011). org/licenses/by/4.0/.
53. von Stosch, M., Carinhas, N. & Oliveira, R. In Large-Scale Networks in Engineering
and Life Sciences (eds Benner, P. et al.) 367−388 (Springer International Pub-
lishing, 2014). © The Author(s) 2020

Published in partnership with the Systems Biology Institute npj Systems Biology and Applications (2020) 6