Octomod slots LLM agents into a systems-biology modelling workflow: an unaided run matches a published model, and a revision cuts AIC by 6.8
Synopsis
Mechanistic models in systems biology require biological knowledge to be written into reaction networks and fitted repeatedly, work that has depended on biology, mathematics and computation together. Octomod embeds large-language-model agents in an existing SBML, PEtab, AMICI and pyPESTO workflow, so agents only propose candidate models and revisions while deterministic modules run checks, parameter estimation and comparison, and the user decides at each interaction point. On published CTD phosphorylation data, one unaided run selected a candidate comparable to the published model, and one revision round reached an AIC 6.8 lower than the published model.
Interpretation
The most labour-intensive part of mechanistic model building, translating a biological description into an executable reaction network and proposing revisions from fit residuals, can now be drafted by large-language-model agents while checks, fitting and model comparison stay with established software. Earlier knowledge-driven modelling routes such as INDRA and BioMASS mainly assembled textual knowledge into model structure, and symbolic-regression methods inferred equations from trajectories; model construction and iterative revision remained manual. On the published expert-built P-TEFb model of phosphorylation across the eight CTD repeats, with the knowledge module disabled, no extra information or web access, and every interaction point approved at its default proposal, the first Octomod round selected an 8-parameter model with AIC −166.4 against the published 7-parameter model at AIC −166.0; the observation and noise models are the same, so the two are directly comparable.
The workflow splits one modelling task into eight modules for knowledge, model building, model expansion, data extraction, PEtab assembly, compilation and probing, parameter estimation and evaluation, coordinated by a central engine, with agents proposing content only inside modules. Each agent is a separate memoryless call that receives only a declared set of inputs; a reply that breaks its output contract or fails a check is returned for redraft, and after three failures the user decides how to proceed. In the PBMC mass-cytometry case, the knowledge module drew on 44 sources in two calls to assemble 73 reactions and 25 facts with evidence grades, and 17 models were built and fitted in total; the best model combined monocyte-secreted type-I interferon and IL-6, a shut-off in T cells and a direct route.
Every input, agent call, check and user decision of a run is stored in a single HDF5 run record, so the derivation of a model can be inspected afterwards. SBML and PEtab record the final model and estimation problem but not how they came about. The run record also holds prompts, replies, model versions and costs, and can be trimmed for dissemination or turned into an HTML report; on unpublished lipidomics data, six consecutive runs reduced a supplied 14-species, 21-reaction model to 9 species and 11 reactions, fitted to 6 of 11 measured lipid classes and 108 data points.
Perspective
The demonstrated scope is four problems: published CTD phosphorylation data, published macrophage transcriptome data, published PBMC mass-cytometry data, and SH-SY5Y sphingolipid pulse-chase data newly generated for this study; the model form is restricted to ordinary-differential-equation reaction networks, with models and estimation problems expressed in SBML and PEtab. The intended readers are systems-biology and computational-biology researchers who build or revise mechanistic models.
The authors note that the comparison with two agent-only analyses is not exhaustive: each was run only once and was not instructed to use the relevant standards and tools, so benchmarking across levels of structure is needed. Exact reproducibility is out of reach for any LLM-based method, since outputs differ between runs and commercial frontier models can change behaviour even within the same version; whether an agent draws on training information it was not explicitly given is also hard to judge. In the macrophage case several parameters remained at their search bounds, and the early IRF1 peak and some donor-associated residuals still fitted poorly; in the lipidomics case the route by which palmitate label enters the aldehydes stayed unresolved, and ceramide remained overestimated late in the chase.
