Legacy and novel brominated flame retardants (BFRs) are highly bioaccumulative contaminants for which risk assessment based on waterborne toxicity data is subject to substantial uncertainty. To address this limitation, we innovatively integrated regression- and classification-based machine learning models to predict tissue-residue-based chronic developmental toxicity to fish, leveraging the complementary strengths of both approaches and improving predictive reliability through cross-validation and application domain (AD) analysis. Among the evaluated models, random forest (RF) and eXtreme Gradient Boosting (XGBoost) exhibited optimal performance in classification and regression tasks, achieving a test-set accuracy of 0.75 and an R2 of 0.94, respectively. Feature analysis identified body weight as a key determinant of toxicity, while structural interpretation revealed that highly halogenated multiring aromatic scaffolds play a critical role in determining toxic potency. Across both common and threatened and endangered (T&E) fish species, smaller-bodied fish exhibited greater sensitivity to the developmental toxicity of BFRs, with 2,2′,3,3′,4,4′,5,5′,6,6′-decabromodiphenyl ether (BDE-209), α-hexabromocyclododecane, and 1,2-bis(2,3,4,5,6-pentabromophenyl) ethane (DBDPE) exhibiting particularly high toxicity. Risk assessment further identified the Pearl River Basin as a high-risk region, predominantly driven by legacy BDE-209. Among novel BFRs, DBDPE contributed substantially to ecological risk across seven representative freshwater basins worldwide. Overall, this validated machine learning (ML) framework enhances the reliability of toxicity prediction and enables pollutant prioritization for both common and T&E fish species, providing robust scientific support for aquatic biodiversity conservation.
Wang et al. (Sun,) studied this question.