Drug-induced liver injury (DILI) remains a significant safety concern in drug discovery and clinical development, often arising from mitochondrial dysfunction, impaired hepatobiliary transport, reactive metabolite formation, or combinations of these biological stressors. Many computational models for predicting liver injury rely primarily on chemical structure or clinical labels, providing limited insight into the underlying biological mechanisms. While neural networks can capture complex patterns, they often lack mechanistic transparency for safety pharmacologists. To address this gap, we propose an explainable deep kernel modeling framework that integrates mitochondrial toxicity endpoints, transporter inhibition profiles, and molecular structure features to generate interpretable, feature-attributed predictions. In this approach, a deep kernel learning model embeds a neural representation within a Gaussian process, linking mechanistic assay features to DILI risk labels, and explanation methods such as SHAP or integrated gradients are applied to decompose each prediction into contributions from specific assays, transporter signals, and structural features. Conceptually, the model would flag a drug candidate as higher risk when transporter inhibition and mitochondrial impairment jointly indicate hepatotoxicity, while the explanation layer identifies whether the risk is driven primarily by BSEP inhibition, reduced mitochondrial respiration, ATP depletion, or structural alerts. By connecting predicted liver injury risk to interpretable biological drivers, this explainable deep kernel framework can support transparent, mechanism-informed safety assessment and help medicinal chemists and safety scientists prioritize de-risking strategies earlier in development.