Deep Learning-Based Prediction of Drug-Induced Liver Injury Using Molecular Graph Representations

• DILI-Graph, a graph convolutional network with attention, achieves AUC 0.92 for DILI prediction, outperforming baseline methods. • Molecular graph representations capture both local and global structural features, improving predictive accuracy. • Feature importance analysis reveals key substructures (e.g., aromatic rings, halogenated groups) associated with DILI risk. • The model shows strong generalizability on external validation sets, supporting its utility in drug safety screening.