Drug-Induced Liver Injury Prediction and Structural Risk Factor Identification Using Multi-Source Molecular Feature Fusion and Explainable Machine Learning

Abstract

Drug-induced liver injury (DILI) is a critical issue in drug safety research, and its mechanism is rather complex, posing significant challenges for DILI risk prediction. Existing studies mostly employ single-molecule representation or fixed feature fusion methods, with less consideration given to the consistency among different molecular information, the reliability of prediction results, and the interpretability of the model. To address this issue, this study proposes RAFF-DILI, a multi-source learning framework, which comprehensively utilizes physical-chemical descriptors, Morgan fingerprints, and molecular graph representations to describe molecular structural features from different aspects. The model adopts sample adaptive gated fusion and combines methods such as cross-view consistency regularization, probability calibration, and consensus attribution to improve the stability and interpretability of DILI risk prediction. The FDA DILIst test results show that RAFF-DILI maintained excellent predictive performance in five independent runs. The AUROC, AUPRC, and MCC values were 0.892 ± 0.012, 0.918 ± 0.010, and 0.671 ± 0.021 respectively, where the AUROC being 0.019 higher than the strongest baseline model.In the DILIrank 2.0 structure repetition external evaluation, the model still achieved an AUROC of 0.841 ± 0.017, indicating its stable prediction performance across different datasets. Further analysis revealed that some physical-chemical features and substructure patterns repeatedly appeared in the prediction results of higher DILI risks. In conclusion, RAFF-DILI can combine molecular information from different sources, maintaining good prediction performance while enhancing the reliability and interpretability of the prediction results, and can provide computational methods for risk screening and priority ranking in drug safety evaluation.