Graph-Based Machine Learning Models Improve Prediction of Drug-Induced Liver Injury and Cellular Toxicity

A team of researchers published a preprint on bioRxiv describing a comprehensive quantitative structure-activity relationship (QSAR) modeling pipeline designed to predict in vitro and in vivo chemical toxicity endpoints. The pipeline integrates molecular fingerprints and graph-based representations with a range of machine learning models—including classical approaches, graph neural networks (GNNs), and graph transformers (GTs)—benchmarked against PubChem assay data for caspase-3/7 activation and mitochondrial membrane potential (MMP) disruption. For FDA Drug-Induced Liver Injury (DILI) prediction, the full consensus model achieved an AUC of 0.69 and the Graphormer model reached an F1 score of 0.79, both substantially exceeding the previous best reported AUC of 0.63 and F1 of 0.65. The study found that graph-based models outperformed classical ML when active compound counts were large, while classical models remained competitive for highly imbalanced datasets with few active compounds. Mechanistic analysis identified phenolic compounds with a para-hydroxyphenyl motif and lipophilic long-chain alkyl compounds as capable of collapsing mitochondrial membrane potential and subsequently activating caspases-3 and -7. Cell-line-specific analysis revealed distinct structural motifs: 1,1-dichloroethane and chlorobenzene for HEK293 cells, an epoxide fragment for SK-N-SH neuroblastoma cells, and a tetramethylcyclohexene motif with an acetaldehyde fragment for H-4-II-E rat hepatoma cells. The authors plan to incorporate large language models, agentic AI, and existing toxicology literature to further refine predictive performance.