Dynamic Graph Neural Networks for Physiological Based Pharmacokinetic Modeling: A Novel Data Driven Approach to Drug Concentration Prediction
Physiologically Based Pharmacokinetic (PBPK) modeling is a key tool in drug development for predicting drug concentration dynamics across organs. Traditional PBPK approaches rely on ordinary differential equations with simplifying assumptions that limit their ability to capture nonlinear and system-level physiological interactions. In this work, we investigate data-driven PBPK modeling using deep learning. We implement two baseline architectures -- a multilayer perceptron (MLP) and a long short-term memory (LSTM) network -- and propose a Dynamic Graph Neural Network (Dynamic GNN) that explicitly models inter-organ interactions through recurrent message passing on a physiological graph. Experiments on a multi-organ pharmacokinetic dataset show that the Dynamic GNN achieves the lowest mean absolute percentage error (MAPE) of 15.7% among all models, demonstrating improved relative accuracy despite slightly higher absolute error compared to the MLP baseline. The model attains an R2 of 0.9342 with more stable error behavior and better captures inter-organ pharmacokinetic relationships. These results highlight the importance of structure-aware modeling for PBPK applications and demonstrate that the proposed Dynamic GNN offers a scalable, equation-free alternative for data-driven pharmacokinetic prediction.