Toxicity-guided assessment of disinfection by-products via machine learning and network toxicology.
Disinfection by-products (DBPs) in drinking water pose significant health risks due to their cytotoxicity, bioaccumulation potential, and interactions with biological targets. Here, we develop a toxicity-guided and interpretable framework integrating molecular modeling, machine learning, network toxicology, and experimental characterization to systematically predict and prioritize potentially high-risk DBPs. A curated dataset of 271 DBPs was assembled, including cytotoxicity (pLC50) and bioconcentration factors (BCF), along with binding affinities to 15 bladder-related proteins obtained through molecular modeling. Machine learning models achieved strong predictive performance (R2 > 0.83), and SHAP analysis identified lipophilicity, electronic structure, and molecular connectivity as key determinants of toxicity. Prioritization of over 271 DBPs revealed that halobenzoquinones (HBQs), particularly halogenated naphthoquinone derivatives, rank among the highest-risk classes. Our results suggest that redox-active structural features, including electrophilicity and polarizability, may contribute to biological interactions and oxidative stress-related responses. Experimental validation across structurally diverse DBPs supported the predicted toxicity prioritization, with concentration-dependent cytotoxicity and Caspase-3-associated apoptosis observed for selected compounds, while network toxicology identified potential associations with oxidative stress, apoptosis, and inflammation-related pathways. Overall, this study establishes a cross-scale, toxicity-guided assessment framework for relating molecular properties to potential biological responses and prioritizing emerging DBPs, supporting evidence-based water quality management.