European Conference on Mathematical and Theoretical Biology (ECMTB 2026)

LLM-Guided Mechanistic Model Discovery in Population Pharmacometrics

MS157-01

14 Jul 2026, 17:00

20m

15.04 - HS (University of Graz)

Minisymposium Talk Mathematical Oncology Integrating machine learning into mathematical oncology

Speaker

Sébastien Benzekry (1. COMPutational pharmacology and clinical Oncology, Centre Inria d'Université Côte d'Azur 2. Cancer Research Center of Marseille, Institut Paoli-Calmettes, Inserm UMR1068, CNRS UMR7258, Aix Marseille University UM105, Marseille, France)

Description

Structural model selection for pharmacokinetics (PK) and tumor dynamics (TD) is iterative and expert-driven, requiring ODE formulation, nonlinear mixed-effects fitting, and biological plausibility assessment. We present an LLM-agent framework for automated population ODE model discovery, fit locally via SAEM (Monolix).

The workflow iterates: a builder agent proposes candidate ODE systems in MLXTRAN; a diagnostic agent interprets goodness-of-fit metrics (BICc, RSEs, IWRES); a reflection agent selects the best structure. No patient data are sent to the LLM.

For PK discovery (synthetic benchmark, n=10), 7/10 ground-truth structures were identified, including 4/4 one-compartment models. On real clinical data (Theophylline, Warfarin, Docetaxel, Irinotecan), AI-selected models matched or outperformed published references.

For TD (synthetic and real preclinical/clinical datasets), canonical growth models (exponential, logistic, Gompertz) were recovered alongside drug-effect structures. Exact model recovery occurred in 50% of six PK-TD scenarios; biologically plausible alternatives emerged otherwise.

Runtimes remained under 20 minutes. This framework offers a tractable, reproducible assistant for mechanistic model discovery in mathematical oncology, complementing expert judgment.