A Stratified Machine Learning Framework for Acute Oral Toxicity Hazard Prediction of Chlorobenzene Compounds

Chlorobenzene compounds are environmentally relevant chlorinated chemicals for which efficient acute-hazard screening can complement experimental assessment. Reliable prediction of acute oral toxicity can support chemical prioritization, while conventional in vivo testing is costly and ethically constrained. Here, we compare global and structure-based stratified machine learning strategies under a common validation protocol. The fixed modeling dataset contained 540 mouse oral LD50 records, including 225 carbon-only and 315 heteroatom-containing compounds. We evaluated conventional RDKit descriptors (F1, 209 dimensions), KPGT-derived representations (F2, 2304 dimensions), and their combination (F3, 2513 dimensions) across Random Forest, LightGBM, CatBoost, and voting ensembles. A toxicity-stratified global hold-out test set comprising 20% of the data was reserved for final evaluation, while subgroup models were trained only on the corresponding subsets of the training partition. On this common test set, the global LightGBM model with F1 descriptors achieved the highest reported point estimates (accuracy = 0.889, balanced accuracy = 0.877, MCC = 0.753). Stratification did not establish a statistically significant improvement over the global strategy. SHAP analysis showed stable feature-importance rankings and subgroup-specific predictive associations, but these patterns do not establish distinct toxicity mechanisms. A Mahalanobis-distance applicability domain was used as a conservative extrapolation warning. The selected model is deployed through a public web platform for rapid screening of the study-specific higher acute-toxicity group (LD50 < 500 mg/kg) versus the at-or-above-threshold group (LD50 ≥ 500 mg/kg). The framework supports murine acute-hazard screening and prioritization; it does not constitute a complete environmental or human-health risk assessment.