Machine-Learning-Accelerated Prediction of Water Quality Criteria for Microplastics

Abstract Microplastics (MPs) are widely distributed in aquatic environments, raising global concerns. However, determining their toxic effects on aquatic organisms and deriving water quality criteria (WQC) for hazardous substances remain challenging due to the heterogeneity of existing toxicity data. Additionally, acquiring sufficient data requires substantial resources. This study proposes a machine learning framework to predict the aquatic toxicity of five types of MPs. Three machine learning algorithms, k-nearest neighbors (kNN), support vector machine (SVM), and random forest (RF), were used to develop quantitative structure–toxicity relationships and derive site-specific WQC from predicted toxic end points. The RF model outperformed kNN and SVM in predictive accuracy after both internal and external validation. SHAP analysis revealed that particle size, density, and aquatic group accounted for 72% of the variability in the predictions. Polystyrene and polyethylene terephthalate exhibited significant toxicity in both freshwater and saltwater, with MPs being more toxic in freshwater. These findings highlight the need for site-specific WQC to protect aquatic ecosystems and improve ecological risk assessments of emerging contaminants.