Comparative evaluation and selection of optimal QSAR-based machine learning model for liver toxicity prediction.

Drug induced liver toxicity remains the most common cause of acute liver failure. Conventional toxicity detection relies on resource-intensive in vivo and in vitro assays that are costly, time-consuming, and poorly scalable to the volumes of candidate compounds typical of modern drug discovery. Robust, interpretable computational tools capable of early stage hepatotoxicity prediction from molecular structure alone are therefore urgently required. A curated multi source dataset of 6,219 compounds was gathered from four publicly available datasets: FDA DILIst, LTKB Benchmark, Open TG-GATEs, and ToxRIC. Seven machine learning algorithms including linear and nonlinear, Logistic Regression, Support Vector Machine, k-Nearest Neighbours, Random Forest, Gradient Boosting, AdaBoost, and XGBoosts were systematically benchmarked within a unified, leak-free pre processing pipeline incorporating median imputation, variance-based feature filtering, SMOTE applied exclusively within stratified fivefold cross-validation folds, and Z-score scaling. Molecular features calculate by using RDKIT included physicochemical, topological, and electronic descriptors, combined with ECFP4 circular fingerprints. Hyper parameter optimisation done by using Grid Search CV and Randomized Search CV, with ROC-AUC as the scoring objective, with (n-estimators = 388, max-depth = 15) achieved a test set ROC-AUC of 0.728, a recall of 0.706, and an F1-score of 0.700. External validation on an independent 50 compound set yielded ROC-AUC of 0.741, confirming consistent generalisation. SHAP Tree-Explainer analysis identified ECFP4 circular fingerprint bits encoding reactive substructural environments and VSA-type descriptors (SlogP_VSA6/7, PEOE_VSA9/10) as the primary structural drivers of hepatotoxicity predictions, consistent with lipophilic accumulation and CYP enzyme inhibition as DILI mechanisms. Applicability domain analysis using the leverage-based hat matrix method confirmed substantially higher prediction reliability for within-domain compounds, providing confidence bounds essential for prospective screening applications. SHAP and LIME consensus analysis validated the importance of descriptors across two independent interpretability frameworks, strengthening the mechanistic credibility of the identified structural features. The developed model provides a reproducible, interpretable, and externally validated computational tool for early-stage hepatotoxicity screening, capable of correctly identifying approximately 71% of hepatotoxic compounds before experimental testing, thereby supporting more efficient and cost-effective prioritisation of drug candidates in the safety evaluation pipeline.