An integrated chemical and ecological framework for deriving freshwater contaminant protection thresholds using machine learning.

Current ecological risk assessments often rely on generic protection thresholds that overlook regional assemblage composition because toxicity data cover only a small proportion of indigenous species. We developed an interpretable, transferable machine-learning framework integrating chemical descriptors and ecological traits to derive assemblage-specific thresholds. Using polycyclic aromatic hydrocarbons (PAHs) and freshwater macroinvertebrates as a case study, Bayesian-optimised XGBoost achieved the highest cross-validated R2 among the tested models. SHAP analysis identified water solubility and log Kow as key chemical descriptors and maximum body size and feeding group as important ecological traits. A streamlined four-predictor model further improved cross-validated performance (R2 = 0.715), supporting toxicity prediction across chemicals and taxa. The framework was applied to freshwater assemblages in Jiangsu, China, and RIVPACS reference assemblages in England. Acute hazardous concentrations for 5% of species (HC5) were derived using probabilistic species sensitivity distributions and 10,000 Monte Carlo simulations, and chronic values were extrapolated using chemical-specific acute-to-chronic ratios. HC5 values were broadly comparable between regions, supporting cross-regional transferability. In Jiangsu, acute HC5 values ranged from 1.085 to 213.2 μg/L, with several high-molecular-weight PAHs exceeding solubility limits, whereas chronic values ranged from 0.007 to 17.0 μg/L and remained below these limits. Derived thresholds generally exceeded existing benchmarks, reflecting differences in protection objectives and assemblage composition. Individual PAH risks were generally negligible to low, whereas mixture risks reached moderate to high levels in northwestern and southeastern Jiangsu. Overall, the framework provides a transparent approach for deriving region-specific protection thresholds under limited toxicity data while explicitly propagating uncertainty.