X-DILiver: an ensemble learning framework for predicting drug-induced liver injury

Drug-induced liver injury (DILI) is a primary cause of drug attrition, associated with over 1000 medications and accounting for 32% of marketed drugs withdrawn due to toxicity. This has created an urgent demand for high-accuracy computational methods predicting DILI risk early in the drug development pipeline. We developed X-DILiver, a predictive framework built on the largest DILI-annotated dataset to date. We used data augmentation to improve model robustness and address class imbalance. A library of 312 machine learning models was created using various algorithms and molecular features, and then an optimized ensemble strategy was applied. The final model is an ensemble of two extreme gradient boosting models and seven recurrent neural networks. It achieved an accuracy (ACC) of 0.64 and a Matthew's correlation coefficient (MCC) of 0.33 on the Tanimoto-filtered external test dataset, and an ACC of 0.64 and an MCC of 0.38 on the scaffold-unique external test dataset, outperforming all other publicly available DILI prediction models. While this performance improvement is significant, this accuracy reflects ongoing challenges in DILI prediction. X-DILiver represents a valuable advancement that can assist in DILI risk assessment during drug discovery, serving as a reliable tool to predict DILI potential, thereby accelerating and improving drug discovery safety. To facilitate broad access, X-DILiver is accessible via http://xdiliver.lile.bio.