Abstract
Although the health impacts of fine particulate matter (PM(2.5)) are primarily attributed to its chemically diverse composition rather than to mass concentration, assessing the toxicity of PM(2.5) constituent compounds remains highly challenging due to chemical complexity and limited experimental scalability. This study introduces an interpretable machine learning (ML) framework using a curated A549 cytotoxicity data set (19,841 compounds) that integrates customized Molecular Access System (MACCS) fingerprints, six base models, and a stacking ensemble meta-model, all optimized with a biobjective strategy to assess the toxicity of organic compounds in PM(2.5). The stacking ensemble model demonstrated satisfactory performance (test AUC > 0.8), exhibiting good generalization, adaptability, and robustness. Nontargeted analysis generated and experimentally validated a prediction set from the Hong Kong PM(2.5) samples (51 of 387 compounds confirmed from 13 classes), demonstrating broad applicability on an independent Nanjing PM(2.5) set (572 compounds). Key substructures driving toxicity, identified through a SHapley Additive exPlanations (SHAP) analysis and cell experimental validation, revealed that PM(2.5) compounds with aromatic rings and nitrogen-based functional groups (e.g., aromatic α,β-unsaturated ketones, aromatic amines, aromatic nitro compounds, and tertiary amines) likely contribute to high toxicity. The derived "structure-toxicity" rules narrow the search space from thousands of compounds to those containing critical substructures, enabling efficient prioritization of toxic components and providing a foundation for improving model specificity and predictive accuracy in future studies.