PINNs, Pills, and Paradigms: Physics-Informed AI for Pathways, Patients, and Predictions

Data Science Seminar

Nazanin Ahmadi
Brown University

Abstract

Understanding and predicting complex biological and pharmacological processes remains a central challenge in quantitative systems medicine. Classical mechanistic models often struggle to capture nonlinear, time-varying, or partially observed dynamics, while purely data-driven approaches sacrifice interpretability and require large datasets that are not always available. In many biomedical contexts, the governing mechanisms are partially known but data are limited, motivating hybrid strategies that reduce data demands while still uncovering hidden dynamics. Moreover, parameters that are ideally constant at steady state frequently evolve over time, further complicating predictive modeling.

To address these challenges, we present a suite of physics-informed networks (PINs) that integrate mechanistic knowledge with modern architectures, including AI- Aristotle[1] for gray-box discovery, Compartment Model–Informed Neural Networks (CMINNs)[2] for PK/PD modeling, and physics-informed state-space models (MAMBA)[3] for latent dynamics inference.

Applications span multiple scales: in oncology, deep sequence models are coupled with PK/PD tumor growth dynamics to predict drug efficacy and explore alternative treatment strategies; in immune signaling, mechanistic ODEs are combined with dissipative particle dynamics (DPD) data to uncover unknown parameters in erythrophagocytosis, validated through identifiability analysis. Across case studies—including glucose–insulin regulation, anomalous drug diffusion, chemotherapy resistance[4], and multi-dose regimens—these PINs demonstrate how  embedding physics into AI architectures enables robust parameter discovery, trajectory reconstruction, and interpretable gray- box learning of hidden dynamics.