Artificial Metabolic Networks – AMN

The project follows a hybrid mechanistic–learning strategy organized into four scientific work packages (WP1–WP4) plus management, spanning metabolic model development, translation into AMNs, and demonstrators in bioproduction and diagnostics.

WP1 – Mechanistic modeling. We build large-scale cell and metabolic models adapted to the targeted Escherichia coli strains: mcFBA (molecular crowding Flux Balance Analysis) and/or RBA (Resource Balance Analysis) to predict fluxes, protein allocation and growth as functions of medium composition and gene deletions. These models are used to generate input/output datasets and multi-dimensional “Monod-like” landscapes, and to fit key parameters from experimental data (e.g., uptake efficiencies). In parallel, medium-scale Structural Kinetic Modeling (SKM) models of central metabolism are developed and analyzed using Metabolic Control Analysis (MCA) to study the propagation of static and dynamic perturbations, quantify uncertainty, and assess the information content of metabolic signals.

WP2 – Translation into AMNs and benchmarking. Stoichiometric AMNs are constructed to reproduce WP1 model behavior and are trained on large simulated datasets. Several architectures are evaluated (including unrolling-based schemes, dedicated activation functions, and RNN-LP-like architectures for linear/quadratic programs). AMNs are then used as reservoirs to solve standard machine-learning tasks (regression/classification) by developing methods that transform problem features into “nutrients” fed to strain models/AMNs, and by training readout layers. Training relies on stochastic gradient descent (SGD) variants and evolutionary optimization approaches.

WP3/WP4 – Bioproduction and diagnostics. Experimental methods are developed to generate training spaces for AMNs using robotic equipment and plate readers to prepare diverse culture media and measure strain growth. Gene-deletion strains are produced and characterized. In diagnostics, protocols are developed to grow E. coli on clinical plasma samples, and clinical datasets from CHU Grenoble-Alpes are used to classify samples (negative vs positive; mild vs severe).

Additional Developments. AMNs extend mechanistic, differentiable learning approaches by relying on differentiable optimization (LP/QP). A dynamic version (dAMN) combining neural networks and dFBA (dynamic flux balance analysis) was developed to predict full growth curves and explicitly capture the lag phase.

The work consolidated a coherent framework linking metabolism, information processing and learning.

Faure, Mollet et al. (Nature Communications, 2023) introduced AMNs as mechanistic–neural architectures embedding stoichiometric constraints from genome-scale metabolic models (GEMs) formulated in FBA through differentiable solvers. AMNs improve predictions of growth/fluxes and knockout effects (e.g., E. coli, Pseudomonas putida) with fewer data than black-box ML approaches. dAMN (Faulon et al., submitted to Bioinformatics, 2026) extends this principle to dynamics by integrating a trainable dFBA module to predict complete growth curves (including lag) and substrate trajectories, outperforming baseline dFBA and Physics-Informed Neural Networks (PINNs).

Mollet et al. (IDA / Lecture Notes in Computer Science, 2025) proposed a protocol that transforms supervised tasks into “nutrient-medium” inputs (AMN as a digital twin) and benchmarks AMNs on regression/classification tasks; performance is competitive with strong baselines (e.g., boosted trees), and ablation suggests that the biological component is a major contributor. The manuscript “Living Bacterial Reservoir Computers” (Ahavi et al., under revision at Cell Systems, 2026) experimentally demonstrates that E. coli, without genetic modification, can serve as a physical reservoir in reservoir computing (RC): growth dynamics can classify early COVID plasmas (mild vs severe) and perform well on standard ML tasks; multi-species simulations further connect phenotypic diversity to computational capacity.

Finally, Lequertier’s PhD thesis (Université Paris-Saclay, 2025) and Lequertier et al., GECCO’25 developed a theory of uncertainty and information propagation in metabolic networks based on linearization around steady-state (SKM/MCA) and a probabilistic description of perturbations. Covariance analysis and Shannon mutual information quantify dependencies and identify candidate metabolites as sensors (flux sensors/biosensor candidates), with extensions to a protein/regulatory layer and multi-objective evolutionary estimation of elasticities.

Dissemination highlights: two workshops were organized — “AI Methods and Models for (Bio)Catalysis and Synthetic Biology” (Université de Montpellier, June 5, 2024; ~150 participants) and “Digital Tools in PEPR B-BEST” (Institut des Systèmes complexes, Paris, November 17, 2025; ~50 participants).

Scientific impact: 5 publications (Faure et al., Nat. Commun. 2023; Ahavi et al., Biotechnol. Adv. 2025; Duigou et al., NAR 2026; + 2 under revision: Ahavi et al., Cell Systems; Faulon et al., Bioinformatics), 2 peer-reviewed international conference papers (Mollet et al., LNCS 2025; Lequertier et al., GECCO’25) and multiple oral/poster presentations. The project supported integration into the European BIOS consortium (101070281), with MICALIS leading the “hybrid learning” WP, and continued AMN/dAMN developments within PEPR Galaxy-BioProd and MUSICH. Economic/environmental benefits: sparse, structured models reducing compute time and experiments. Societal benefits: foundations for microbial biosensors and more accessible diagnostics.

Perpectives: the next step targets temporal challenges explicitly: leveraging bioreactors as a physical reservoir in RC for NARMA-like tasks. A controlled bacterial culture (turbidostat/chemostat) receives time-dependent inputs (programmed dilution/substrate/OD setpoints, with OD = optical density), and its dynamic response (OD and other online signals) forms the reservoir state. A simple readout layer is trained for time-series prediction (NARMA) or classification.

The scale-up is planned in stages: (i) establish stationarity and repeatability in turbidostat, (ii) generate controlled trajectories (OD up/down, input sequences), then (iii) deploy multi-bioreactor arrays (diverse conditions/strains) to increase reservoir dimension and memory. In parallel, this strengthens the biosensor axis by exploiting dynamic growth signatures (potentially with controlled perturbations) to detect and discriminate other diseases or environmental conditions. Two PhD students were recruited and a partnership with Inria was established to support this evolution.

While the primary role of metabolism is chemical conversions, can it also serve as an information processing device? To answer this question, we propose to encode various microbial metabolic models into Artificial Metabolic Networks (AMNs), which can be trained on experimental data or model simulations. Unlike “black box” artificial neural networks, our AMNs will be sparse and will reflect faithfully the structure and dynamics of metabolic networks.
Our AMNs will be benchmarked on classical machine learning problems to assess what level of computational sophistication metabolism is able to handle.
In the context of biotechnology, our AMNs will be applied to the design of experiments to (i) optimize the productivity of an added-value chemical (lycopene) E. coli producing strain defining nutrient compositions and gene deletions and (ii) classify infectious disease severity by engineering an E. coli biosensing strain detecting metabolic biomarkers in COVID-19 clinical samples.