CardioSafe: An open-source three-branch fusion model for hERG cardiotoxicity prediction with explainable atom-level attribution

Background hERG potassium channel blockade is a leading cause of drug-induced cardiotoxicity and late-stage clinical attrition, accounting for up to one-third of Phase III trial failures. Existing free computational tools for hERG prediction lack atom-level mechanistic explainability, do not integrate complementary molecular representations, and require expensive hardware. This study presents CardioSafe, a fully open-source, zero-cost platform addressing all three gaps. Methods CardioSafe employs a novel three-branch fusion architecture combining a three-layer Graph Attention Network (GAT), 2048-bit Morgan circular fingerprints, and seven physicochemical descriptors, trained on 8316 curated hERG compounds from ChEMBL v33 using BCEWithLogitsLoss with class weighting and OneCycleLR scheduling. GNNExplainer generates atom-level attribution maps per prediction. Performance was evaluated on a scaffold-based held-out test set (n = 835) with bootstrap confidence intervals (n = 1000) and five-fold cross-validation. A systematic ablation study quantified the contribution of each branch. Results CardioSafe achieved AUC-ROC of 0.731 (95% CI: 0.701–0.759), sensitivity 64.01%, specificity 71.03%, precision 77.96%, F1 0.703, and MCC 0.341. The ablation study confirmed that all three branches contribute independently. Five-fold cross-validation yielded AUC 0.844 ± 0.017 (SD). GNNExplainer attribution consistently identified aromatic rings and basic nitrogen atoms as primary cardiotoxicity drivers. Conclusions CardioSafe provides the freely accessible hERG prediction tool with atom-level explainability. The complete training pipeline runs on Google Colaboratory free tier without API keys, paid licences, or specialised hardware, making early-stage computational hERG screening more accessible to researchers, without replacing experimental cardiac safety testing.