StackDILI: Enhancing\nDrug-Induced Liver Injury Prediction\nthrough Stacking Strategy with Effective Molecular Representations

Drug-induced liver injury (DILI) is a major challenge\nin drug development,\noften leading to clinical trial failures and market withdrawals due\nto liver toxicity. This study presents StackDILI, a computational\nframework designed to accelerate toxicity assessment by predicting\nDILI risk. StackDILI integrates multiple molecular descriptors to\nextract structural and physicochemical features, including the constitution,\npharmacophore, MACCS, and E-state descriptors. Additionally, a genetic\nalgorithm is employed for feature selection and optimization, ensuring\nthat the most relevant features are used. These optimized features\nare processed through a stacking ensemble model comprising multiple\ntree-based machine learning models, improving prediction accuracy\nand interpretability. Notably, StackDILI demonstrates a strong performance\non the DILIrank test set and maintains robustness across cross-validation.\nMoreover, interpretability analysis reveals key molecular features\nassociated with DILI risks, providing valuable insights into toxicity\nprediction. To further improve accessibility, a user-friendly web\ninterface is developed, allowing users to input SMILES strings and\nreceive rapid predictions with ease. The StackDILI model provides\na powerful tool for efficient DILI assessment, supporting safer drug\ndevelopment. The web interface is accessible at https://awi.cuhk.edu.cn/biosequence/StackDILI/.