Machine Learning-Based Prediction of NOECs for over 10,000 PFAS across Multiple Fish Species Enables Large-Scale Toxicological Analysis

Per- and polyfluoroalkyl substances (PFASs), comprising over 10,000 persistent chemicals, are prevalent in aquatic ecosystems and threaten ecological health. Fish no-observed-effect concentrations (NOECs) are a key ecological safety indicator; however, NOECs remain undefined for most PFASs. Here, we develop a machine learning model to predict mortality-based NOECs for 10,863 PFAS in Danio rerio using compiled toxicity data and mechanistically meaningful molecular structural descriptors. The model achieved robust predictive performance (Rtest2 = 0.7096), with predicted NOECs ranging from 0.239 (0.131–1.129) to 163.452 (77.207–341.332) mg/L, highlighting a wide toxicity variation that suggests further structural investigation would be beneficial. Analysis reveals a “U-shaped” toxicity chain-length trend, with higher toxicity observed for C8–C12 PFAS and polar functional groups (e.g., sulfonates and carboxylates). We further extrapolate these toxicity profiles to 12 additional fish species through trait-based inference and found that warm water and omnivorous species are more sensitive, reflecting trait-mediated differences in bioaccumulation potential and metabolic vulnerability. Structure–species interactions jointly shape toxicity patterns across species. Together, these results extend the mechanistic understanding of the toxicological responses of fish to PFASs. Integrating chemical structures with ecological characteristics enables high-throughput toxicity predictions using limited data, provides mechanistic insights into PFAS toxicity, and establishes a transferable framework for ecological risk screening and prioritization.