Machine Learning-Driven Cross-Species Toxicity Prediction for Advancing Ecologically Relevant PFAS Water Quality Criteria

Traditional toxicity testing cannot keep pace with the
rapid growth
of synthetic chemicals, creating major data gaps that hinder the development
of water quality criteria (WQC) for emerging contaminants. This study
developed a machine learning model integrating compound- and organism-related
features to enable cross-compound and cross-species toxicity prediction.
The model demonstrated strong robustness and generalization, outperforming
the Interspecies Correlation Estimation application in cross-species
prediction, particularly across large taxonomic distances. SHAP analysis
identified water solubility and lipophilicity as dominant predictors,
with organism-related features also contributing substantially. The
model predicted the acute toxicity of 30 representative per- and polyfluoroalkyl
substances (PFAS) across 181 aquatic species. Habitat-informed species
selection was then used to derive ecologically relevant 5% hazardous
concentrations (HC5), which were generally higher in saltwater
than in freshwater. Cross-regional comparisons further indicated that
salinity may modulate fish sensitivity to PFAS. HC5 estimates
for China were higher than those for North America and Europe, potentially
reflecting inter-regional differences in species sensitivity, with
Chinese species appearing comparatively more tolerant. Finally, site-specific
WQC for perfluorooctanoic acid (PFOA) and perfluorooctanesulfonic
acid (PFOS) were derived for the Great Lakes using predicted sensitivities
of 76 dominant native species, providing greater ecological relevance
than existing criteria.