Artificial Intelligence-Powered Prediction of Microplastic–Pollutant Adsorption Coefficients Enables Scalable Risk Prioritization in Aquatic Environments

Microplastics (MPs), prevalent in water bodies, soil, and the atmosphere, pose significant risks to environmental and ecological health. By adsorbing hazardous compounds, such as organic pollutants, MPs can alter the transport and fate of these pollutants. To address this, we developed a multimodal Siamese neural network named microplastic-pollutant adsorption prediction (MPAP), trained on 1101 adsorption records covering 403 compounds and six MP types. Unlike previous models that rely on single-feature representations, MPAP leverages a multimodal architecture that integrates molecular fingerprints and graph embeddings to capture chemical structures, along with microplastic morphological features such as the MP type and particle size, as well as water chemistry parameters, enabling a more comprehensive characterization of sorption behavior. The model outperforms baseline models with R2 = 0.869 on the validation set and 0.863 on the test set. Experimental validation using batch adsorption experiments with six previously untested pollutants, quantified via liquid chromatography-mass spectrometry or microwave plasma torch ionization-mass spectrometry in different environments, confirmed a strong predictive performance. To support broad application, we provide an open-access web platform (http://mpap.envwind.site:8004/) for rapid, high-throughput prediction across diverse MP-pollutant-water environment scenarios.