Species-aware prediction of fish acute toxicity improves hazard screening and prioritization of environmentally detected pharmaceuticals.

Widespread contamination of surface waters by active pharmaceutical ingredients (APIs) leads to ubiquitous exposure in fish, critical sentinels in aquatic food webs. However, ecotoxicity assessment remains constrained by pronounced interspecies differences in sensitivity and by the scarcity of toxicity data for fish species representative of distinct ecoregions. To support hazard screening of environmentally detected APIs, we developed an ensemble model integrating Deep Neural Networks (DNN) with six other machine learning models. Trained on 10,054 acute toxicity data points (4190 groups, 1651 organic compounds, 24 freshwater fish species), the model achieved a test-set R2 of 0.72 and an RMSE of 0.83. Ablation analysis showed that the model trained on a feature space integrating chemical, biological, and exposure-related descriptors outperformed the model based solely on chemical descriptors, indicating that biological and experimental factors provide complementary information for predicting fish acute toxicity. Feature interaction analysis revealed that Salmoniformes exhibited heightened sensitivity to lower-MW compounds, whereas under static exposure conditions, higher MW was associated with attenuated toxicity. Applied to 714 environmentally detected APIs lacking experimental fish acute toxicity data, the framework showed that applicability-domain definitions strongly influenced scenario coverage and downstream species sensitivity distribution construction. This work provides a predictive strategy that helps address the limitations of disparate and ecologically unrepresentative testing data that compromise regional relevance, thereby enabling proactive and ecologically grounded chemical prioritization in freshwater ecosystems.