Modeling Ethnic Differences via Drug-Predicted Clearance from Chemical Structures

Ethnic differences in pharmacokinetics can result in clinically relevant differences in drug efficacy and safety, even when identical doses are administered. Although mechanistic approaches such as physiologically based pharmacokinetic modeling can be used to evaluate population-specific pharmacokinetics, there remains a need for complementary screening approaches that can assess ethnic difference risk at very early stages of drug discovery with limited experimental information. The aim of this study was to develop a machine-learning-based approach to assess the risk of ethnic differences in oral clearance between Japanese and Western populations by integrating chemical structural features with in silico-predicted pharmacokinetic (PK) parameters. The models incorporating extended-connectivity fingerprints (ECFPs) and 17 DruMAP-derived PK parameters demonstrated improved predictive performance compared with the models using structural information alone, achieving a receiver operating characteristic area under the curve of 0.811 vs, 0.701 for the ECFP-only model. Model interpretation using SHapley Additive exPlanations indicated that biologically plausible factors, including the unbound fraction in the plasma and the predicted contributions of specific cytochrome P450 enzymes, were important contributors to the classification. External validation using an independent set of compounds (n = 21) achieved a recall of 75% for compounds exhibiting ethnic differences, thereby supporting the feasibility of the proposed approach. These results suggest that integrating chemical structural information with DruMAP-derived PK properties enables the early-stage assessment of ethnic differences in risk without relying on clinical pharmacokinetic data. The proposed framework may support rational decision-making for bridging strategies and efficient design of multiregional clinical trials in international drug development.