Machine learning frameworks for predicting pulmonary cell toxicities induced by metal ions in the atmosphere

Metals in PM2.5 are closely associated with cardiopulmonary disease endpoints, potentially attributable to ionic species-induced oxidative stress effects. Computational prediction of metal ion respiratory toxicity remains challenging due to complex valence-dependent interactions. In this study, we constructed a predictive framework for high-throughput prediction of multiple metal ions toxicity endpoints. This framework leverages a 650-data-points multidimensional experimental dataset capturing dual pulmonary toxicological endpoints in cellular models, characterized by 31 valence-specific features, to identify key toxicological drivers simultaneously. Robust machine learning models established here accurately predicted metal ion respiratory toxicity (R²=0.89 for cell viability, ACC=0.86 for oxidative stress), validated externally (R²=0.74, ACC=0.70) by conducting Experimental validation on 10 new metal ions independent of the modeling dataset. Machine learning-driven feature selection ranked solubility, electronegativity, and cation charge of metal ions as top cell viability predictors, while the Pearson softness coefficient, molar mass, and density as vital parameters of oxidative stress. The predictive framework establishes property-toxicity relationships for PM2.5 metal ions, providing an effective tool for air pollution risk prioritization and mechanistic insights, thereby reducing unnecessary experimental burden.