Multi-Endpoint Semisupervised Learning Identifies High-Priority Unregulated Disinfection Byproducts

Disinfection byproducts (DBPs) in drinking water pose significant health risks, while experimental toxicity data exist for fewer than one-third of them, limiting comprehensive risk assessment. Here, we develop a machine learning framework that integrates multiendpoint toxicity data with semisupervised learning to predict the toxicity of untested DBPs. We compiled four data sets encompassing cytotoxicity in CHO and HepG2 cells, developmental toxicity in zebrafish embryos, and genotoxicity in CHO cells, covering 227 DBPs with standardized endpoints (LC50, EC50). By unifying these data sets and incorporating endpoint-specific parameters (cell type, exposure duration), our model achieved robust predictive performance (R2 = 0.86, RMSE = 0.44), outperforming single-endpoint models (R2 = 0.41–0.73). To expand applicability, we implemented semisupervised learning with a hybrid acquisition function balancing structural similarity and prediction uncertainty, increasing coverage of structurally similar compounds by 18%. Predictions on untested DBPs revealed that several unregulated classes, particularly iodinated acetonitrile, halogenated diones, and halobenzoquinones, exhibit 2–10× higher toxicity than regulated trihalomethanes and haloacetic acids. Notably, haloacetonitriles dominate cytotoxicity and genotoxicity rankings, whereas halobenzoquinones show elevated developmental toxicity. This integrated framework, strengthened by external validation, provides a robust tool to prioritize toxic DBPs, guiding monitoring strategies and informing regulatory standards for safer drinking water.