Interpretable QSAR models for acute oral toxicity via tuned XGBoost and hybrid data sampling techniques

Accurate, interpretable prediction of acute oral toxicity (LD₅₀) is challenged by severe dataset imbalance and complex structure-activity relationships. This study develops a transparent QSAR framework by integrating 2D topological and 3D conformational descriptors with hybrid SMOTE-RUS resampling to classify rat oral LD₅₀ into Very Toxic vs. Not Very Toxic. Using 8,396 compounds from NICEATM-EPA NCCT, features were refined via sequential filtering, correlation pruning, and Random Forest ranking. Among seven optimized machine learning models, XGBoost achieved superior external performance (F₁=0.62, accuracy = 0.87). The model's interpretability was ensured via Explainable AI: Permutation Feature Importance highlighted global contributors like nP and TDB01m; SHAP analysis identified local determinants such as TPSA(Tot); and surrogate decision trees distilled the logic into rule-based thresholds with high fidelity (≥ 0.88). This pipeline aligns with OECD principles, offering a regulatory-grade, explainable QSAR model that balances predictive power with mechanistic transparency for chemical safety assessment.