Machine learning models for prediction of emerging pollutants' risk in Daphnia magna.

Abstract

Emerging pollutants may threaten aquatic ecosystems, yet the limited toxicity data for many compounds significantly restrict high-throughput screening of these contaminants. Because Daphnia magna is a sensitive standard test organism and an important link in aquatic food webs, reliable prediction of its acute toxicity response is critical for ecological risk assessment. Here, we curated a Daphnia magna acute toxicity (EC50) database for 170 compounds across several classes of emerging pollutants, including per- and polyfluoroalkyl substances (PFASs) and organophosphate esters (OPEs), using public databases and literature. We developed machine learning models to predict 1-4 day EC50 values and found that eXtreme Gradient Boosting (XGB) performed best (QLOO2 = 0.95, QEXT2 = 0.85), followed by Random Forest (RF; QLOO2 = 0.89, QEXT2 = 0.77). An eight-compound case-study validation showed that predicted and empirical Daphnia magna acute-effect values were generally within one order of magnitude, but the small sample size and high leverage of extreme values require cautious interpretation. Model interpretation identified molecular topology and electronic distribution as key toxicity drivers. The predicted Daphnia magna EC50 values were then combined with empirical toxicity data from other aquatic species to construct Species Sensitivity Distribution (SSD) curves. This framework provides a high-throughput tool for filling toxicity data gaps and supporting ecological risk assessment of emerging pollutants.