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

Hepatotoxicity evaluation is vital in drug development and chemical safety assessment. With increased human exposure to diverse chemicals from global industrialization, regulatory authorities seek computational alternatives to animal testing for efficient toxicity assessment. However, existing prediction models face challenges including activity cliffs, interpretability and limited applicability. Here, we present ChemBioHepatox, a novel multimodal deep-learning framework integrating chemical structures with biological assay responses for accurate hepatotoxicity prediction. We constructed the first comprehensive hepatotoxicity assay response spectrum through rigorous assay selection and fusion. ChemBioHepatox employs a multimodal architecture bridging structural and biological information within a shared embedding space, with a linear output layer strengthening mechanistic interpretability. Our model demonstrates superior predictive performance (AUC 0.92, precision 0.88, recall 0.87) while mitigating activity cliffs. Validation across pharmaceuticals, pesticides, and food additives confirms broad applicability. By coupling accuracy with mechanistic insight, ChemBioHepatox offers a new paradigm for toxicological evaluation and regulatory decision-making, available at http://exposomex.cn:58080/.