Prediction of the joint toxicity of microplastics and organic pollutants on algae based on machine learning.

Complex joint toxicity driven by microplastic (MP)-organic pollutant mixtures in aquatic ecosystems remains poorly captured by conventional models. Herein, a dataset covering 10 types of MPs and 6 organic pollutants was compiled from literature to investigate joint toxicity mechanisms. Employing 5-fold cross-validation, six machine learning models were developed to predict the algal growth inhibition ratio based on MP properties, physicochemical characteristics of organic pollutants, and experimental conditions. The best predictive performance was achieved using the CatBoost model (test-set AUC: 0.935 ± 0.021; F1: 0.832), which outperformed benchmarks. SHAP analysis revealed the significance of experimental conditions (exposure time and concentration), physicochemical properties (hydrophobicity and molecular substructure fingerprints), and their interaction terms (LogP × Time × Conc) on joint toxicity prediction, supporting the cumulative effect principle and quantitative structure-activity relationships. The phenolic groups (SHAP = 0.43) and polycyclic aromatics (FP_58) were identified as key toxic molecular substructures. Notably, a Synergy-Antagonism Index integrating structural similarity, mechanistic pathways, and SHAP weights was proposed to distinguish the joint effects. Results indicated that phenolic hydroxyls, polycyclic aromatics, and amides contributed to significant synergistic effects, whereas hydrophobic aliphatic chains (FP_47, interaction = -0.359) often drove antagonism. Overall, this work offers robust prediction for MP-organic pollutant joint toxicity and provides new insights into high-throughput risk assessment of co-exposures in aquatic environments.