Graph-Based Classification with GNN-Explainer for Predicting Cardiac Toxicity Associated with Multi-Ion Channel Blockers

Cardiotoxicity remains a critical concern in drug development, often leading to late-stage attrition of promising compounds. While traditional assessments focus on Kv11.1 channel inhibition, the Comprehensive in Vitro Proarrhythmic Assay (CiPA) initiative emphasizes the importance of evaluating additional cardiac ion channels, notably Cav1.2 and Nav1.5. In this study, we address the limitations of existing machine learning (ML) models, which typically rely on Kv11.1-specific data, by developing a deep learning (DL) framework that integrates inhibition data across all three key ion channels. A large and diverse data set (Cardio-Tox) was curated by combining experimental data from the PubChem, CUPID, and CToxPred2 repositories, totaling 34,124 molecules for Kv11.1, 1564 for Cav1.2, and 3217 for Nav1.5. Using this data set, trained GNN models are capable of individual channel prediction. The developed CardiotoxPred method, which includes the Kv, Cav, and Nav models, achieved an average prediction accuracy of 86.7% on a test data set. In addition to robust predictive performance, GNNExplainer offers interpretable visualizations by highlighting atom- and bond-level contributions via colors. These insights support cardiac molecular severity estimation, optimization, and safety profiling. All the models are freely accessible via GitHub in a user-friendly Docker container, providing a practical tool for early-stage cardiotoxicity risk assessment in drug discovery pipelines.