Integrating Multi-Omics and Machine Learning to Predict Microplastic Cytotoxicity Under Leave-One-Polymer-Out Validation.

Microplastic (MP) cytotoxicity prediction remains challenging because available experimental datasets are limited and biological responses vary substantially among polymers. This study integrated multi-omics, physicochemical descriptors, and machine learning to predict MP cytotoxicity in BEAS-2B (a human bronchial epithelial cell line) cells. Biological descriptors were derived from multi-omics screening and concentration-dependent validation, while Z-Ave (hydrodynamic average particle size), zeta potential, and concentration were used as physicochemical descriptors. Using 50 original polymer-concentration observations from five polymers, 18 combinations of six machine-learning algorithms and three descriptor sets (QSAR, QBAR, and QSBAR) were evaluated by nested leave-one-polymer-out (LOPO) cross-validation. Gaussian perturbation, applicability-domain analysis, and SHAP (SHapley Additive exPlanations) were used for sensitivity and model interpretation. RF-QSBAR achieved the best overall out-of-fold performance (R2=0.794, RMSE=0.072, and CCC=0.858), followed by GBDT-QSBAR (R2=0.783, RMSE=0.074, and CCC=0.865). The polymer-level bootstrap 95% CI for the RF-QSBAR R2 was 0.428-0.818, reflecting uncertainty associated with the limited number of polymers. RF-QSBAR showed stable performance under training-set Gaussian perturbation, with R2 values ranging from 0.809 to 0.860. Applicability-domain analysis identified extrapolation risk for some held-out polymers, and SHAP analysis identified Z-Ave as the most influential predictive feature, followed by PLPP4 (phospholipid phosphatase 4) and GPD2 (mitochondrial glycerol-3-phosphate dehydrogenase 2). Integrating physicochemical and biological descriptors showed potential for cross-polymer MP cytotoxicity prediction; however, Z-Ave should be interpreted as a proxy for correlated between-polymer differences rather than an independent causal size effect, and further external validation is required. This framework may provide a potential tool for supporting risk assessment, waste management, and future regulatory decision-making.