Wolfram Liebermeister

Institut National de Recherche pour l’Agriculture, l’Alimentation et l’Environnement (INRAE)
Unité MaIAGE, Domaine de Vilvert, Bâtiment 210, 78350 Jouy en Josas

wolfram.liebermeister at inrae.fr

Publications et révisions d'articles

1. Publications sur Google Scholar | ResearchGate | HAL INRAE
2. Publications et révisions d'articles sur Web of Science
3. ORCiD 0000-0002-2568-2381

Preprints

Articles de journaux

Commentaires

Articles de conférence

Documents sur les standards et formats de modèles

Livres

Chapitres de livre

1. RBA: Analyse du bilan des ressources
2. Proteomaps: visualisation des données protéomiques
3. SBtab: format de données pour la biologie des systèmes
4. ObjTables: schémas pour les fichiers de données
5. Parameter balancing
6. Model balancing
7. Modélisation structurelle-thermokinétique
8. Minimisation du coût des enzymes
9. Metabolic Network Toolbox
10. semanticSBML: Outils pour les modèles au format SBML
11. SB.OS: Outils pour la biologie des systèmes

1. Code sur github: https://github.com/liebermeister
2. Noor E. and Liebermeister W. (2023). Optimal enzyme profiles in unbranched metabolic pathways: Jupyter Notebooks. Zenodo

Thèses

Expérience professionelle

CV détaillé (pdf)

Posters récents

Projet PEPR B-BEST ``Multi-size Hybrid Cell Models'' (MuSiHC) (2024-2028)

The project Multi-Size Hybrid Cell Models (MuSiHC) aims at developing novel hybrid approaches to cell and bioreactor modeling for the production of added-value compounds. MuSiHC addresses gaps in current cell models used to simulate bioproduction: a dichotomy between linear genome-scale and small kinetic models with no realistic intermediate solutions that describe metabolism precisely; and a division between interpretable mechanistic models with a massive numerical effort, vs efficient but black-box Artificial Intelligence (AI)/Machine Learning (ML) models. As a proof of concept, the project will focus on Escherichia coli as a platform for the bioproduction of 1,3-propanediol (1,3-PDO), a high-value compound with vast applications in the chemical industry, ranging from solvents to antifreeze.

The project will develop a toolkit of hybrid models, combining mechanistic description and AI/ML of different size, to obtain more reliable cell and bioreactor simulations for E. coli producing 1,3-PDO. Models at different levels of detail will be connected with the help of ML. (i) To model metabolism at a genome scale, Artificial Metabolic Network (AMN) models will be adapted and trained on experimental data for simulating E. coli as a platform for bioproduction. The aim is a genome scale neural-mechanistic hybrid model, to accurately predict production rates for different media and gene deletion sets. Using the novel AMNs, it will be possible to predict both heterologous pathways and gene deletions, and validate the predictions with batch cultures. (ii) In parallel, medium-scale kinetic models will be analyzed with ML (white-box and black-box) to better understand the usage of different Elementary Flux Modes (EFMs) by cells, depending on external conditions (composition of the growth medium). We will obtain effective rules describing choices of metabolic strategies, including bioproduct production, depending on the environment, which can inform the construction of simple, yet realistic cell models. The latter will be used within bioreactor models, and experiments at the lab scale will validate, fit, and improve the predictions.

Addressing different levels of detail, the hybrid models will (i) develop approaches to switch between models of different resolution, creating well-justified methods for model reduction based on the concept of EFMs; (ii) predict metabolic behavior as a function of extracellular metabolite concentrations, taking into account discrete transitions between EFMs, and allowing the simulation of engineered cells, with enzymes or pathways added or removed; (iii) use small-sized models obtained from model reduction to numerically simulate a lab-scale bioreactor environment and explore whether these simulations can be used to dynamically optimize bioproduction, using reinforcement learning techniques and model-based control approaches on digital twins of the system. Experimental data will be collected for model development and validation, plus proof-of-concepts for bioproduction. The experimental plan is to (i) grow E. coli strains in a computer-controlled mini-bioreactor system to calibrate the reduced, small-sized models, by means of measurements of growth rate and uptake/secretion rates of key metabolites; (ii) carry out experiments pf 1,3-PDO production in connected fermentors to study the effects of spatial heterogeneity under controlled conditions, using cell populations models with distributions over their parameters; (iii) perform validation runs in larger, pilot-scale fermentors, optimizing bioproduction. Once the proof of concept on E. coli and 1,3-PDO will be validated, the established methodology will be formalized in a protocol easily adaptable to different organisms and bioproducts.

Projet ANR ``Artificial Metabolic Networks'' (AMN) (2021-2024)

Our main objective is to demonstrate that the metabolism of microorganisms can serve as a device for solving computational problems. The primary role of metabolism is to process food and it is not a priori viewed as an information processing apparatus. Yet, organisms have evolved to cope with different environments and metabolism necessarily played a role in processing and transducing environmental signals to the genetic layer. It is often assumed that biochemical networks serve either for chemical production (metabolism, synthesis of macromolecules) or signaling (MAPK pathways, gene regulation networks,..). This separation ignores the fact that metabolism is also involved in cell signaling and that metabolic fluctuations may allow signals to spread. There is indeed some evidence that metabolism contributes to quite sophisticated signal transmission and decision-making tasks in microorganisms [1][2] plants [3] and of course with the production of neurotransmitters involved in the chemical synapses of our brains.

Here we hypothesize that this ability - the processing and transduction of signals in metabolism - can be diverted to tackle problems that are typically solved by artificial intelligence and in particular by artificial neural networks (ANNs). Since metabolic networks link inputs (concentrations of external metabolites) to outputs (fluxes, concentrations, cell growth), they can be considered as "programs", comparable to ANNs. By adding entries or changing their dynamics (for example by gene deletions), these programs could be shaped to perform classification or regression tasks.

Our objective is to formalize this new paradigm and to establish a relationship between metabolic models and ANNs. A formal connection will help us (i) understand in WP1 the information processing capacities of metabolism (e.g. transducing information about external conditions or aggregating information about metabolic demands) and how evolution may create "informative” metabolites, such as flux sensors [4], (ii) investigate in WP2 metabolism-inspired artificial networks and their potential for machine learning, and (iii) engineer in WPs 3 and 4 in vivo "metabo-genetic devices" that can detect biochemical signals and convert these into informative outputs in the context of two biotechnologically relevant problems: bioprocess optimization and medical diagnostics.

Projet DFG "An economic network theory of cell metabolism" (LI 1676/2-2) (2016-2019)

An efficient use of resources such as metabolites, proteins, energy, membrane space, and time is vital for cells. Cells can be studied as economic systems and metabolism is a core component of this system. Mathematical models such as constrained-based, kinetic, and simplified whole-cell models combine knowledge about network structure, thermodynamics, kinetics, enzyme regulation, and allocation of cellular resources. However, these modelling approaches rely on simplifications which limit their accuracy and mutual compatibility. In the project "An economic network theory of cell metabolism", I will develop a unified mathematical theory to study optimal enzyme allocation in cells. The theory, called metabolic economics, extends existing modelling approaches and links them in a new way. The optimality conditions for protein allocation are formulated as economic balance equations, which relate the "economic values" of individual metabolites and enzymes in the network. Metabolic models studied by metabolic economics can be constraint-based or kinetic, and their fitness objectives can score fluxes, metabolite levels, enzyme levels, and cell growth. In addition, I develop a method for flux prediction that combines flux analysis, kinetic modelling, and cost-benefit models of growing cells. I will implement the theoretical concepts and methods in computational workflows, apply them to whole-cell models, and study three biological cases: selection pressures in engineered pathways, the optimal scheduling of enzyme expression in metabolic cycles, and general conditions for symbiosis in bacterial communities. Metabolic economics will extend metabolic modelling, clarify trade-offs between protein investments and metabolic fluxes, and deepen our understanding of enzyme usage in cells.

Projet DFG "Dynamics and function of enzyme regulation in large metabolic networks" (LI 1676/2-1) (2011-2015)

Kinetic models are essential to better understand the dynamics and function of enzyme regulation, and hence to predict the metabolic effects of differential enzyme expression and enzyme-inhibiting drugs. However, kinetic modelling is not yet applicable to large, genome-scale networks. Existing genome-scale modelling approaches, such as flux balance analysis, are based on stoichiometry only and therefore inherently limited in use. This project aims to fill the gap between genome-scale stoichiometric and small-scale kinetic models by the development of a novel kinetic modelling approach for large metabolic networks. The method combines metabolic control analysis with data integration and sampling techniques and accounts for thermodynamic constraints. The results will be probabilistic, reflecting the availability and quality of input data.

The applicability of the new method will be tested through its use to create large network models and to predict flux changes caused by enzyme regulation, to couple these models with detailed kinetic pathway models, to compute synergisms between enzyme-inhibiting drugs, to predict the dynamic effects of alternating enzyme levels, and to study the advantages of enzymatic regulation and alternating enzyme levels in fluctuating environments through a computational cost-benefit approach. Through external collaborations, model predictions will be validated with experimental omics data from bacterial and yeast cultures and from human hepatocytes. By extending dynamic modelling to large metabolic networks, the project will substantially improve the understanding of enzyme regulation, with potential future applications in cell simulation, prediction of drug interactions and side effects, and chronotherapies.