Integrating multi-omics and machine learning to predict microplastic cytotoxicity under leave-one-polymer-out validation

Quantitative prediction of microplastics (MPs) cytotoxicity is constrained by the scarcity of experimental data and biological variations across different polymer types. To address this, this study proposes a multidimensional prediction framework that integrates multi-omics-derived biological descriptors with physicochemical descriptors, aiming to enhance the capability and interpretability of cross-polymer toxicity prediction. Based on raw polymer-concentration observational data for five MPs, we evaluated 18 model combinations using a nested leave-one-polymer-out (LOPO) cross-validation framework. To quantify model robustness and extrapolation risk, we incorporated Gaussian perturbation and applicability domain (AD) analysis, while utilizing the Shapley Additive exPlanations (SHAP) method to interpret feature contributions. The results demonstrate that multidimensional descriptors significantly improved cross-polymer prediction performance; the RF-QSBAR model achieved the best performance on the validation set (R² = 0.794, RMSE = 0.072, CCC = 0.858), followed by the GBDT-QSBAR model (R² = 0.783, RMSE = 0.074, CCC = 0.865). Polymer-level bootstrapping analysis revealed a 95% confidence interval of 0.428-0.818 for the RF-QSBAR model's R² value, reflecting uncertainty arising from the limited number of polymer types; however, model performance remained stable under Gaussian perturbation (R² = 0.809-0.860). AD analysis indicated potential extrapolation risks for some of the held-out polymers. SHAP analysis identified Z-Ave as the most important predictive feature, followed by PLPP4 and GPD2; however, Z-Ave should be regarded as a surrogate indicator reflecting relative differences between polymers rather than an independent causal size effect. Overall, the multidimensional descriptor approach shows promise for predicting MPs cytotoxicity, though further external validation is required to support its application in risk assessment and decision-making.