vToxiNet: An interpretable deep learning framework for hepatotoxicity prediction and mechanism exploration

Hepatotoxicity remains a leading cause of drug attrition and post-marketing withdrawal, resulting from diverse and complex toxicity mechanisms. Traditional in vitro models can only capture a limited subset of toxicity pathways, and animal studies face translational and ethical limitations. Regulatory agencies have therefore promoted new approach methodologies, including human-relevant assays, omics technologies, and computational models to improve predictive toxicology and support evidence-based decision-making. However, most machine learning models for hepatotoxicity either rely solely on chemical structure or operate as black boxes, limiting mechanistic interpretability and broader applicability. Here, we introduce the virtual toxicity network (vToxiNet), a biologically constrained deep learning framework that embeds systems toxicology knowledge directly into neural network architecture for interpretable hepatotoxicity prediction. vToxiNet integrates chemical descriptors, high-throughput assay responses, transcriptomic signatures, and Reactome pathway hierarchy to construct a virtual adverse outcome pathway network. Across cross-validation and external validation, vToxiNet demonstrates promising predictive performance. Importantly, interpretation of vToxiNet enables gene and pathway-level attribution, supporting mechanism-informed hazard characterization and chemical prioritization. By integrating multimodal data and biological hierarchy as architectural constraints, this study provides a framework that combines prediction with mechanistic interpretation and supports further development of mechanism-informed approaches for chemical hazard assessment.