NephroTox AI: an explainable machine learning framework with uncertainty quantification for drug-induced nephrotoxicity prediction

Drug-induced kidney injury remains a major cause of clinical-stage drug attrition, yet computational nephrotoxicity prediction is hampered by fragmented datasets, opaque models lacking mechanistic interpretability, the absence of uncertainty quantification, and the lack of publicly available prediction platforms. Here, we curated a dataset of 2,157 compounds from three sources with deduplication and label harmonization, and benchmarked 28 model–descriptor combinations (7 algorithms × 4 descriptor types) to identify Random Forest with RDKit 2D descriptors as the best performer (validation AUC 90.7%, test AUC 89.6%). A four-tier uncertainty quantification system based on tree variance enabled confidence-aware filtering: restricting evaluation to high- and medium-confidence predictions (21.7% coverage, ~54 of 247 compounds) yielded an external AUC of 90.2% versus 72.7% on the full set, with 88.3% of false negatives and 94.5% of false positives concentrated in the discarded low-confidence tiers. SHapley Additive exPlanations analysis identified charge (MaxPartialCharge), molecular complexity (BertzCT), and molecular weight as the top predictive features—descriptors whose physicochemical relevance to renal handling has been established in prior pharmacokinetic studies, though their SHAP contributions reflect statistical associations rather than pathway-specific attributions. Applying the model to 11,093 DrugBank compounds yielded 59 high-confidence nephrotoxic candidates, including cisplatin, colistin, and methotrexate. Network pharmacology identified candidate convergent targets (e.g., OAT1, AKT1, and CASP3) along an accumulation–signaling–apoptosis cascade, explored by molecular docking. The complete computational pipeline is implemented as NephroTox AI, a public web application available at https://lishulab.mpu.edu.mo/nephrotox , integrating prediction, interpretable explanations, confidence labeling, and LLM-powered conversational assistance to facilitate safer drug design and risk assessment. Scientific contribution This work advances computational nephrotoxicity prediction through five integrated contributions: (1) a curated multi-source dataset of 2,157 compounds with traceable annotation provenance; (2) a systematic benchmark of 28 model–descriptor combinations establishing RF_RDKit as the optimal classifier; (3) a four-tier uncertainty quantification system enabling confidence-aware applicability domain definition (external AUC 90.2% at 21.7% coverage versus 72.7% unfiltered); (4) convergence of SHAP interpretability with network pharmacology and molecular docking to propose a mechanistic accumulation–signaling–apoptosis cascade hypothesis; and (5) NephroTox AI, to our knowledge the first publicly available nephrotoxicity-focused web platform to integrate ML prediction, SHAP-based interpretability, tier-based uncertainty quantification labeling, and LLM-powered conversational explanation within a single deployment.