Explainable Deep Learning–Based Prediction of hERG Channel Blockage: A Cardio-Oncology Perspective
TL;DR
This explainable QSAR approach offers a transparent, mechanistically interpretable decision-support tool for early-stage cardiotoxicity screening, bridging the gap between computational predictions and clinical cardio-oncology safety assessments.
Abstract
Cardiotoxicity induced by human Ether-a-go-go-Related Gene (hERG) channel blockade remains a critical safety liability, particularly in cardio-oncology, where antineoplastic agents pose severe proarrhythmic risks. This study developed an explainable deep learning-based QSAR framework for the multidimensional prediction of hERG-mediated cardiotoxicity. A primary dataset of 22,213 compounds was used to train a deep neural network (DNN), utilizing RDKit descriptors, MACCS keys, and Morgan fingerprints individually and in hybrid combinations. The integrated DNN model achieved an exceptional ROC-AUC of 0.9801 during the training phase. For external validation, an independent oncology test set of 348 antineoplastic agents was evaluated. The results revealed that the high structural complexity of chemotherapeutics limits generalizability, highlighting the multifactorial nature of oncology-related cardiotoxicity. Crucially, explainable artificial intelligence (XAI) analyses demonstrated that the model's decisions are firmly grounded in pharmacologically coherent principles, identifying lipophilicity (MolLogP), polar surface area, and specific topological motifs as primary toxicity determinants. Overall, this explainable QSAR approach offers a transparent, mechanistically interpretable decision-support tool for early-stage cardiotoxicity screening, bridging the gap between computational predictions and clinical cardio-oncology safety assessments.