In silico prediction of drug-induced cardiotoxicity with ensemble machine learning and structural pattern recognition

Drug-induced cardiotoxicity poses a significant risk to human health, and reliable predictive models are needed for safety assessment. In this study, a range of machine and deep learning models were developed for five cardiotoxicity end points, including heart failure (HF), arrhythmia (ARR), heart block (HB), hypertension (HP), and heart attack (HA). A total of 110 predictive models were constructed for each cardiotoxicity endpoint using various algorithms and molecular descriptors, and consensus models were developed based on the best-performing individual classifiers. The consensus models consistently outperformed individual models in cross-validation and external validation. Further molecular property analysis revealed that cardiotoxic compounds tend to exhibit higher molecular weight, increased lipophilicity (logP), lower hydrogen bonding capacity (HBD and HBA), and reduced topological polar surface area (TPSA). Additionally, key structural alerts (SAs), including secondary amines, benzene derivatives, sulfonamide/sulfonylurea groups, and heterocyclic structures, were identified. These SAs may mediate cardiotoxicity through ion channel inhibition, oxidative stress induction, and calcium homeostasis disruption. This study provides an integrated machine learning and deep learning computational framework for drug cardiotoxicity assessment and provides an exploration of the structural characteristics of cardiotoxic compounds, which is helpful for the discovery of safer drugs and chemical risk assessment.