Large language models complement established workflows for mechanistic model construction and refinement in systems biology

Mechanistic models provide a framework for describing, understanding, and predicting biological processes, but their development requires an interdisciplinary team with expertise in biology, mathematics, and computation. Here we present Octomod, a multi-agent workflow based on large language models (LLMs) that integrates biological knowledge retrieval, model construction, parameter estimation, and hypothesis testing. Built on established software tools, including AMICI and pyPESTO, and community standards, including SBML and PEtab, Octomod supports model development and refinement through natural-language interaction with a user. The agents' proposals must pass defined checks and the user's decision before they are used. Every input, model, check, and decision is recorded, so that model building remains transparent and reproducible. We applied Octomod to four problems. Three used published data and differ in model size and in how much the user guided the run: Octomod constructed a model that matched a published model without user guidance, supported the interactive refinement of a gene-regulatory model, and built models of cell-cell communication that explained the responses of five immune cell types to different stimuli. In the fourth application, on newly generated and unpublished lipidomics data, Octomod helped to determine which processes the measurements could inform, identify unresolved label fluxes, and reduce a supplied model to those processes. Together, these results show how LLMs can strengthen and complement established workflows for constructing and refining biologically informed models.