ChemBioHepatox: Multimodal Integrating Chemical Structure and Biological Fingerprint for Robust and Interpretable Hepatotoxicity Prediction

Drug-induced liver injury (DILI) is a leading cause of clinical trial attrition and postmarketing withdrawal and a major contributor to acute liver failure. As regulators increasingly encourage human-relevant, nonanimal approaches, accurate and interpretable computational tools for liver safety are needed. Existing hepatotoxicity models are limited by activity cliffs, restricted applicability domains, accuracy, and poor interpretability. We developed ChemBioHepatox, a multimodal framework that couples chemical structure with biological assay responses to improve predictive performance and mechanistic transparency. The framework was pretrained on the DILIst data set (768 DILI-positive and 511 DILI-negative compounds) and fine-tuned on a multisource downstream data set compiled from seven studies. ChemBioHepatox achieved an AUC of 0.92 (precision = 0.88, recall = 0.87), and 5-fold cross-validation under random, scaffold-based, and cluster-based partitions further confirmed its robustness and generalizability to unseen chemotypes. A linear classifier operating on the concatenated structural embedding and 19 assay probabilities enables direct attribution of each assay's contribution via its learned weights. External validations (including LiverTox severity discrimination and targeted HepG2 CCK-8 assays) further confirmed model-flagged high-risk hepatotoxicants among pesticides and food additives. Activity-cliff analysis on the held-out test set further revealed complementary contributions of structural embeddings and mechanism-informed assay-response fingerprints. ChemBioHepatox advances mechanism-driven liver safety assessment, supports regulatory modernization, and provides openly available code, data, and a web interface (http://exposomex.cn:58080/).