A robust graph-based computational model for predicting drug-induced liver injury compounds

Drug-induced liver injury (DILI) is a leading cause of drug development failures and post-market withdrawals, presenting significant challenges to both pharmaceutical safety and clinical care. Conventional screening methods, such as laboratory-based cell assays and animal models, are limited in their ability to accurately predict human hepatotoxicity. To address this, we propose a robust graph-based deep learning model that combines graph attention networks (GAT) and graph convolutional networks (GCN) for the prediction of DILI-associated compounds. Molecular structures are represented as graphs, where atoms and bonds are encoded with rich chemical and topological features derived from molecular formulas. The model effectively learns complex interaction patterns within each compound and demonstrates superior performance over existing machine learning and deep learning baselines. It achieves an area under the receiver operating characteristic curve (ROC-AUC) of 0.8481, an area under the precision–recall curve (PR-AUC) of 0.9008, and an F1-score of 0.8017. Furthermore, evaluation across twenty independent trials confirms strong reproducibility. These findings highlight the model’s reliability and potential as a practical tool for early-stage hepatotoxicity screening in drug discovery.