Metabolic Modeling

Metabolic modeling sits at the intersection of biochemistry, systems biology, and computational science, and at the intersection of essentially every research area in this library. The cost-driving questions in cellular agriculture: what minimal medium supports growth?, what flux distribution maximizes biomass while suppressing lactate?, what gene knockouts increase yield in suspension culture?, are all metabolic-modeling questions in disguise. Metabolic Modeling is a matrix research area, and this page is its deep dive. The column holds papers that model a metabolic network: reconstructing or reducing a genome-scale model, predicting flux, or designing pathways and strains against one. Much of the surrounding work described below sits in neighbouring columns instead, because the column records what a paper demonstrates rather than what it is about: a medium formulation reached with a metabolic model behind it is Media Optimization, flux-informed reactor operation is Bioprocess & Scale-Up, and metabolomics read out to predict an eating-quality attribute is Sensory Prediction. See Taxonomy.md for the column’s scope, which is the definition a placement is judged against; this page is an AI-assisted overview and is not a definition source.

Two papers sit in the column. What places them here rather than in Cellular Engineering is that each demonstrates the construction or search of a metabolic network rather than an intervention on a cell. Xiao et al. (2026, ref #145)‘s MetaGEM reconstructs genome-scale models bottom-up, casting the problem as enzyme–metabolite link prediction across a protein-language-model tower and a 3D metabolite tower, then mapping the predicted enzyme set onto BiGG reaction templates. Li et al. (2024, ref #68) mines metabolic-engineering strategies from the literature at scale and trains a hybrid graph model over the stoichiometric network to predict which gene modifications raise product yield. The column is small because the boundary is narrow, not because the subject is: the theme beside it tags many times more material that never enters the matrix.

Constraint-based modeling (FBA family)

The dominant paradigm for whole-cell metabolic analysis is constraint-based modeling: starting from a genome-scale metabolic reconstruction (a stoichiometric matrix of every known reaction in the organism), solve a linear program to find the flux distribution that maximizes biomass (or another objective) subject to thermodynamic and capacity constraints. Flux balance analysis (FBA), flux variability analysis (FVA), and Monte-Carlo sampling are the core algorithms; strain-design extensions (OptKnock, OptForce, OptCouple) layer mixed-integer programming on top to identify gene knockouts that redirect flux toward desired products.

The open-source stack for this work is mature and largely centered on the openCOBRA consortium: see COBRApy, COBRA Toolbox, Memote, Escher, BiGG Models, RAVEN Toolbox, StrainDesign, and CNApy in Software.md. The substrate for cell-ag work (species-specific genome-scale metabolic models (GEMs)) is fragmented and growing: each GEM is catalogued on its species page in the Datasets/ directory: the bovine BtaSBML2986, chicken iES1300, porcine pcPigMNet2025, salmon SALARECON, and CHO iCHO family reconstructions. Recent applied work in this area includes the CHO-focused data-driven GEM reduction approach (Antonakoudis & Richelle 2026, npj Sys Bio Apps) and the multi-omics cultivated-meat framework #60 (Mathieu et al. 2025, Trends in Food Science & Technology).

Kinetic and dynamic modeling

A complementary paradigm models metabolism as a system of ordinary differential equations (ODEs) with enzyme kinetic parameters, simulating concentration trajectories rather than steady-state fluxes. This approach captures regulatory dynamics, time-dependent responses to perturbations, and bioreactor behavior that constraint-based models cannot, at the cost of needing measured kinetic parameters that are often unavailable. The principal open-source tools are COPASI and Tellurium wrapping libRoadRunner; the BioModels repository at EMBL-EBI hosts thousands of curated SBML models. Kinetic parameters themselves are catalogued in BRENDA.

ML-based and hybrid approaches

A third paradigm (increasingly active since 2024) uses ML directly on omics or imaging data to predict metabolic phenotypes, often as a complement rather than replacement for mechanistic models. Single-cell foundation models (scGPT, Geneformer, scFoundation) capture pathway membership and co-expression structure but do not directly predict fluxes; bridging foundation-model embeddings to flux predictions is an active research frontier. The recent State (Adduri et al. 2025) and OmicsLM (Sypetkowski et al. 2026) papers represent two flavors of this trend, perturbation prediction and multimodal omics reasoning respectively. Bayesian-optimization approaches (e.g. Narayanan et al. 2025, Nat Comms, MIT Love lab) demonstrate that data-efficient ML can drive media-formulation work without requiring a full GEM.

The AI agent layer

The newest layer (and the one most rapidly evolving) wraps the above tools with large-language-model agents that can plan multi-step metabolic-reasoning workflows, call FBA / FVA / kinetic primitives as functions, and interpret results. Closest in spirit to a “ChatCOBRA” today are Talk2Biomodels (Wehling et al. 2025) for kinetic SBML models, Genesis (Tiukova et al. 2024, King group) for systems-biology automation, AutonoMS (Brunnsåker et al. 2025) for LLM + symbolic logic + automated cell culture, Lila (Singh et al. 2023, Carbonell group) for microbial strain design, and the Saez-Rodriguez MCP servers (Ruscone et al. 2025) for gene-regulatory and Boolean-network modeling. Adjacent infrastructure includes Biomni, TxAgent, ToolUniverse, PaperQA / PaperQA2, and the AI Scientist family from Sakana, none of which is metabolic-modeling-specific but all of which provide patterns transferable to cell-ag agentic workflows.

Open gaps for cell-ag

Despite the activity above, several gaps remain conspicuous for cellular agriculture specifically: no MCP server dedicated to COBRApy or BiGG Models yet exists; no benchmark suite scores AI agents on cell-ag-relevant metabolic-engineering tasks (the equivalent of Arc Institute’s Virtual Cell Challenge for FBA); no published bridge connects single-cell foundation-model embeddings to flux predictions; and species-specific GEMs for bovine, porcine, avian, and fish are scattered across preprints rather than centralized in a BiGG-style canonical home. The cultivated-meat industry has reported significant cost wins from internal AI stacks (Pythag Tech, DeepLife, Magic Valley) but these are proprietary: credible open competitors that close the gap between published GEMs, open-source FBA tools, and active-learning media formulation are still emerging. The repository tracks this evolving landscape; contributors are encouraged to add new GEMs, agentic tooling, and applied work as it appears.