MEDA System Discovers ODE Models for Biological Systems.

The long-standing goal of automatic scientific discovery, where machines independently generate and refine mechanistic models, is being advanced by systems like MEDA. This new framework leverages Large Language Models (LLMs) and symbolic regression (SR) within an agentic architecture to discover Ordinary Differential Equation (ODE) models specifically for biological and biologically inspired dynamical systems. MEDA operates by retrieving relevant background knowledge, defining admissible variables, generating mechanistic constraints, proposing candidate ODEs, and then fitting and evaluating these models. The system has been tested across various scenarios, including canonical model retrieval, reasoning-based extrapolation to novel variants, and open-ended discovery, both with and without experimental data. Results show MEDA successfully recovers correct state variables, achieves high structural accuracy in retrieval and extrapolation, and produces biologically plausible models. Ablation studies highlight the critical role of knowledge-guided formalization and mechanistic constraints, demonstrating their superiority over purely numerical fitting approaches.