Development of hybrid models for genome-scale metabolic networks

Summary

Over the past two decades, the systems biology community has dedicated substantial efforts to constructing genome-scale metabolic models (GEMs), which offer detailed representations of an organism’s entire metabolism. GEMs represent metabolisms as networks, linking metabolic reactions and metabolites. Despite their wealth of information, GEMs come with notable limitations. They attempt to encompass all potential metabolic phenotypes, leading to an extensive solution space that can be challenging to explore efficiently. The predominant approach for exploiting GEMs, Flux Balance Analysis (FBA), relies on simplification and lacks the ability to generalize across diverse conditions. By contrast, Machine Learning (ML) techniques are of interest for metabolic modeling, notably by harnessing largescale-omics data to predict biological behaviors in various environments. While many approaches combine GEMs and ML, they still separate the processes of metabolic modeling and ML, limiting their adaptability and reusability. Within this Ph.D. thesis, I use an innovative approach that tackles this limitation: a hybrid neural-mechanistic model for GEMs known as an Artificial Metabolic Network (AMN). This entails the development of FBA surrogate methods compatible with gradient backpropagation and the creation of a mechanistic loss function to align AMN predictions with the constraints of GEMs. This dissertation delves into the biological phenomena addressed by AMNs and surveys state-of-the-art GEM utilization methods. Then, it demonstrates how AMNs outperform FBA in predicting E. coli growth rates across diverse media and genetic conditions, without requiring additional experimental data. The capabilities and limitations of AMNs are then thoroughly examined. Finally, I summarize the findings and offer insights into ways to pursue the development of hybrid models for GEMs, that may help in building high-performance, insightful whole-cell models—an ambitious goal in the realm of systems biology.

Contact

• Leon Faure (SDSV, UMR MICALIS)
• Linkedin
• ResearchGate

Publications

Thesis

Léon Faure. Development of hybrid models for genome-scale metabolic networks. Quantitative Methods [q-bio.QM]. Université Paris-Saclay, 2023. English. ⟨NNT : 2023UPASL103⟩. ⟨tel-04562281⟩

journal articles

• Cell-Free Biosensors and AI Integration
• In silico, in vitro, and in vivo machine learning in synthetic biology and metabolic engineering (Current Opinion in Chemical Biology)
•  Faure, L., Mollet, B., Liebermeister, W. et al. A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models. Nat Commun 14, 4669 (2023). https://doi.org/10.1038/s41467-023-40380-0
• Pandi, A., Diehl, C., Yazdizadeh Kharrazi, A. et al. A versatile active learning workflow for optimization of genetic and metabolic networks. Nat Commun 13, 3876 (2022). https://doi.org/10.1038/s41467-022-31245-z

Conference paper 

• Léon Faure, Manish Kushwaha, Thomas Duigou, Jean-Loup Faulon. Biological Design and Engineering By and For machine Learning. Engineering, Evolution & Design (SEED), AICHE, Jun 2021, Vitual, United States. ⟨hal-04275493⟩
• 2022 AAAI Spring Symposium: Artificial Intelligence for Synthetic Biology, March 21-23, 2002, Stanford University, Palo Alto, California, USA,  AI-4-SynBio Symposium, Imperial College London, (slides) March 24, 2022, London UK.

Projects 

• Projet ANR Réseaux Métaboliques Artificiels (AMN) : démarrage mars 2022 (42 Mois). Aide ANR : 498 390 €. Coordinateur : Jean-Loup Faulon