RenalTox-XAI: Applicability-Domain-Constrained and Descriptor-Guided QSAR Framework for Early Screening and Safety Prioritization of Drug-Induced Renal Injury.

Drug-induced renal injury (DIRI) remains a major cause of drug attrition and post-marketing safety concerns, highlighting the need for reliable early-stage predictive approaches. Conventional biomarkers, including serum creatinine and blood urea nitrogen, are widely used in clinical monitoring; however, their limited sensitivity for early-stage injury restricts their utility for proactive drug safety assessment. In this study, we introduce RenalTox-XAI, an applicability-domain-constrained and descriptor-guided explainable QSAR framework for the early screening and safety prioritization of DIRI. A curated dataset of 316 orally administered drugs was used to develop classification models directly from chemical structure information. Starting with 1444 molecular descriptors, a multi-step refinement strategy identified 11 informative, non-redundant descriptors representing toxicologically relevant steric, electronic, hydrogen-bonding, topological, and connectivity-related molecular features. Multiple machine learning algorithms were evaluated using a five-times repeated 10-fold cross-validation scheme, and predictive performance was further assessed on an independent test set and an independent external validation dataset assembled from publicly available drug safety resources, consisting of 85 drugs (43 nephrotoxic and 42 reference-negative drugs). Among the investigated models, XGBoost showed the best overall performance, achieving an accuracy of 0.7538, an AUC of 0.7862, an F-score of 0.7647, and a Matthews correlation coefficient of 0.5081 on the independent test set. The final model further demonstrated consistent performance on the independent external validation dataset, achieving an accuracy of 0.7529, an AUC of 0.7654, an F-score of 0.7586, and a Matthews correlation coefficient of 0.5058, supporting its robustness and generalizability beyond the DIRIL modeling dataset. To move beyond black-box classification, SHapley Additive exPlanations provided both global and compound-specific interpretations of renal toxicity predictions. SHAP analysis identified AATS5v, minHBint2, AATSC0i, AATS1v, and MLFER_BH as the most influential descriptors, suggesting that steric organization, hydrogen-bonding interactions, electronic environment, and molecular connectivity collectively shape model-predicted renal toxicity profiles. Applicability domain analysis further established reliability boundaries for future screening applications. RenalTox-XAI supports DIRI-risk prediction, descriptor-guided interpretation, and renal safety prioritization, enabling mechanism-aware screening and predictive toxicology aligned with established QSAR principles.