Hepatotoxicity is a leading cause of drug failure and post-market withdrawal, arising from complex interactions among chemical structure, metabolism, cellular stress, and host susceptibility, which makes predictive toxicology particularly challenging. Traditional computational models often focus on either chemical structure or assay-derived toxicity signatures, yet fail to integrate these sources into a single interpretable framework, limiting their utility for medicinal chemists, especially when models flag hepatotoxicity without identifying the responsible molecular substructures or ignore cytotoxicity profiles that reveal mitochondrial dysfunction, oxidative stress, or membrane damage. To address this, a SHAP-guided graph neural network has been proposed, combining molecular graph representations with in vitro cytotoxicity assay endpoints—including viability loss, ATP depletion, reactive oxygen species generation, and mitochondrial membrane potential disruption—while decomposing each prediction into atom- and substructure-level contributions. The model encodes molecular graphs through a graph attention network and cytotoxicity endpoints via a fully connected network, fusing these representations before generating a hepatotoxicity risk estimate. SHAP values are then computed over graph nodes and aggregated into chemically meaningful substructures, allowing local explanations to highlight reactive or stress-associated motifs and clarify whether cytotoxicity assay signals reinforce the structural alert. This approach not only predicts hepatotoxicity but also provides mechanistic insights, linking molecular substructures to cellular stress mechanisms, thereby supporting safer molecular design and offering toxicologists a transparent, actionable basis for evaluating liver injury risks.