Metabolomics‑driven, data‑augmented machine learning for predicting toxicity of microplastic mixtures

Microplastics (MPs) occur as heterogeneous mixtures in real‑world environments, making one‑by‑one toxicity testing impractical. This study aims to use predictive models to quickly and effectively evaluate the toxicity of MPs. We explored three model frameworks: a quantitative structure-activity relationship (QSAR) model based on physicochemical descriptors; a quantitative bioactivity relationship (QBAR) model with biodescriptors screened by metabolomics data; and a quantitative structure-bioactivity relationship (QSBAR) model combining both physicochemical and biodescriptors. Under a simplex centroid design, six machine learning algorithms were trained using data augmentation strategies to predict the cytotoxicity of microplastic mixtures. The results showed that the QBAR-based eXtreme Gradient Boosting (XGB-qbar) model performed best (R²_tra = 0.9322, R²_test = 0.8923), outperforming the QSAR and the QSBAR frameworks. The three descriptor importance methods consistently identified key biological descriptors helpful for toxicity prediction. Moreover, metabolomics analysis indicated that mixed exposure to MPs may mediate toxic responses by reprogramming cellular energy metabolism pathways. The metabolomics-driven and data-augmented machine learning approach proposed in this study can efficiently predict toxicity and provide mechanistic clues in small sample and complex mixture scenarios, providing a feasible path for environmental exposure risk assessment.