Cardiosim-Tox: an interpretable multitask deep learning QSAR platform with multimodal feature fusion for predicting...

Drug-induced cardiotoxicity, mainly driven by cardiac ion-channel blockade, remains a leading cause of drug attrition and post-market withdrawal, highlighting the need for reliable early-stage screening tools. Existing computational methods, including QSAR models, largely focus on single ion channels, limiting their ability to assess multi-channel safety profiles. To address this gap, we developed Cardiosim-Tox, a modular multi-modal deep learning platform to simultaneously predicts blockade risk (binary classification) and potency (pIC₅₀) for hERG, Cav1.2, and Nav1.5. The framework integrates topological fingerprints, Mordred 2D/3D descriptors, and molecular graph encodings under both single-task (STL) and multi-task learning (MTL) configurations. Developed on 31,816 curated unique compounds, model performance was evaluated on three Tanimoto similarity-stratified external validation sets, with domain-shift metrics and applicability domain analyses supporting robustness. MTL models consistently outperformed STL, while the full modality combination (FP + MD + Graph) achieved AUCs of 0.93-0.98 for Cav1.2, 0.88-0.96 for Nav1.5, and 0.79-0.95 for hERG across validation sets. Regression tasks similarly favored MTL (R² = 0.94), with multimodal configurations outperforming single-modality baselines. Further comparisons with classical machine learning models and previously developed cardiotoxicity models demonstrated the superiority of Cardiosim-Tox across all endpoints. SHAP analysis confirmed that feature-activity relationships, both at global and mechanistic levels, align with known ion-channel pharmacophores, enhancing interpretability beyond predictive accuracy. Overall, Cardiosim-Tox provides a reliable and interpretable platform for integrated multi-channel cardiac safety assessment, supporting early-stage cardiotoxicity screening in drug development in line with the CiPA paradigm. The Cardiosim-Tox web server is accessible at https://metaheart.kr/login .