Architecture of dAMN (figure from the original article). The model combines two neural networks — one predicting lag phase parameters, the other metabolic fluxes — within a mechanistic integration loop constrained by the full stoichiometry of the bacterial metabolic network. The system is trained by backpropagation, simultaneously minimizing the deviation from measured concentrations, stoichiometric constraints, and flux positivity.

Published in Bioinformatics — https://doi.org/10.1093/bioinformatics/btag230

We have developed dAMN (dynamic Artificial Metabolic Networks), a hybrid model combining artificial neural networks with genome-scale dynamic flux balance analysis (dFBA) to predict how bacteria grow across diverse nutritional environments, in collaboration with Wageningen University (The Netherlands).

A central challenge in biology is accurately modeling cell growth as a function of culture medium composition. Classical mechanistic approaches struggle to generalize to new media and fail to capture lag phases — the initial adaptation delay before bacteria begin actively dividing. dAMN overcomes these limitations by embedding neural networks within a ResNet-like residual update scheme constrained by the full stoichiometry of the organism’s genome-scale metabolic network.

Trained on growth data from Escherichia coli and Pseudomonas putida across hundreds of combinatorial media, dAMN accurately predicts temporal growth dynamics and generalizes to media not seen during training (mean R² ≥ 0.9). The model spontaneously reproduces biologically meaningful phenomena such as substrate depletion, acetate overflow, and diauxic shifts — without having been explicitly supervised on these behaviors.

The software, models, and data are freely available on GitHub and Zenodo (DOI: 10.5281/zenodo.17908125).

Ref: Faulon JL et al. dAMN: a genome scale neural-mechanistic hybrid model to predict bacterial growth dynamics. Bioinformatics, 2026. doi: 10.1093/bioinformatics/btag230