Drug-induced liver injury (DILI) remains one of the most pressing challenges in drug development, contributing to 25-30% of late-stage clinical attrition and representing a leading cause of post-marketing drug withdrawals. Here, we introduce Oakuloid, a hybrid DILI prediction framework that combines a stacked ensemble of three gradient-boosted decision-tree models, including LightGBM, XGBoost, and CatBoost, with an inference-time rank-mean fusion module. This module integrates a structure-based prediction score with the experimental IC50/Cmax ratio measured using our previously reported 3D primary human hepatocyte iBAC platform. Oakuloid establishes a new state of the art among open structure-only DILI predictors, achieving an AUROC of 0.767 +/- 0.001 on a strict external benchmark and an AUROC of 0.821 on our internal DILIrank cohort. When wet-lab fusion is enabled, performance on the internal cohort further improves to an AUROC of 0.872, while per-class sensitivity for less-severe DILI compounds increases from 0.24 to 0.76. The fusion strategy correctly identifies paradigm DILI drugs, including Acetaminophen, Clozapine, and Simvastatin, that are systematically misclassified by either signal alone. The two channels provide mechanistically complementary information: structural features capture a priori reactive or scaffold-related liabilities, whereas the IC50/Cmax ratio reflects a posteriori cellular damage at clinically relevant exposure levels. Their integration is defined by a single equation that enables scoring of any new compound at single-compound inference time. Beyond prediction, we release the trained model bundle, a curated 122-compound DILIrank wet-lab benchmark, calibration analysis, and a Mitchell-style model card under the Apache License 2.0, supporting transparent and reproducible drug-safety screening.
Zhang, F., Zhou, Y., Ding, D., Zhang, F., Xiao, R., Ai, X.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 1
- Comments 0