Beyond Molecular Structures: Investigating Demographic Factors in Drug-Induced Cardiotoxicity Prediction Models

Predicting drug-induced cardiotoxicity remains one of the most important challenges in drug safety, contributing to a substantial share of clinical trial failures and postmarket withdrawals. While clinical evidence shows differences in adverse responses across sex, age, and body mass, incorporating demographic factors into in silico prediction models remains challenging. We developed CARBIDE (CARdiotoxicity Based on Integrated Demographic Evidence), a collection of 27 dataset variants derived from the FAERS pharmacovigilance database, to systematically evaluate whether meaningful structure-demographic interactions could be learned from spontaneous reporting data. Through systematic evaluation of different FAERS filtering criteria, cardiotoxicity definitions, and statistical methods, together with comprehensive ablation studies, we found that machine learning models failed to extract useful structure-demographic relationships. The models either learned population-level statistics or relied solely on structural information, with demographic features derived using our approach providing little additional predictive value. While these findings reveal fundamental limitations in using pharmacovigilance data for demographic-aware toxicity prediction, CARBIDE's systematic evaluation provides important insights for the field, helping guide future efforts toward more effective approaches in personalized cardiotoxicity prediction.