Structure‐Based Prediction of QT Prolongation Risk Using Graph Neural Networks: An Integrative Approach Combining In Vitro hERG Assays and Pharmacovigilance Data

Drug-induced QT-interval prolongation, a non-specific biomarker of increased risk for Torsades de Pointes (TdP), is a major safety concern in drug development. While in vitro hERG inhibition assays are required for early-phase screening, pharmacovigilance data from sources like the FDA Adverse Event Reporting System (FAERS) provide complementary insights. Integrating these heterogeneous data with molecular structure offers a promising, yet underutilized approach to predict proarrhythmic risk. We developed an interpretable graph neural network (GNN) framework integrating in vitro hERG inhibition data (PubChem AID 588834), FAERS-derived pharmacovigilance signals, and molecular structure information. Canonical SMILES were converted into molecular graphs using RDKit, and atom- and bond-level features were encoded. Four GNN architectures (GINE, GCN, GraphSAGE, and GATv2) were compared via stratified five-fold cross-validation. The best-performing model, GATv2, was further interpreted using Integrated Gradients to identify structural features contributing to QT liability. The final data set comprised 4,808 small molecules with binary QT-risk labels. GATv2 achieved a cross-validated ROC-AUC of 0.838, PR-AUC of 0.830, and F1-score of 0.756. Retraining on the full data set yielded ROC-AUC 0.918, PR-AUC 0.908, and F1-score 0.847. External validation on an independent hERG assay (AID 1671200, n = 2,405) confirmed strong performance (ROC-AUC 0.859, sensitivity 0.80, specificity 0.82). Atomic degree and hydrogen count were dominant predictors, consistent with known SARs. This GNN-based framework integrates structural and pharmacological data to predict QT risk, providing a transparent, structure-based decision-support tool aligned with ICH S7B/E14 and CiPA guidelines.