MACHINE LEARNING-BASED PREDICTION OF METAL OXIDE NANOPARTICLE TOXICITY USING PHYSICOCHEMICAL AND EXPOSUREDEPENDENT PARAMETERS

Abstract

Metal oxide nanoparticles are widely used in pharmaceutical, biomedical, environmental, and industrial applications; however, their potential toxicity remains an important safety concern. This study evaluated the toxicity of selected metal oxide nanoparticles using physicochemical, chemical, exposure-dependent, and machine learning-based approaches. The analysis included five nanoparticles: Al₂O₃, CuO, Fe₂O₃, TiO₂, and ZnO. Variables considered included nanoparticle type, core size, hydrodynamic size, surface charge, surface area, exposure time, dosage, chemical/electronic descriptors, number of oxygen atoms, and toxicity class. A total of 881 observations were analyzed, of which 476 were toxic and 405 were non-toxic. Nanoparticle-wise analysis showed that CuO and ZnO had stronger toxic tendencies, while TiO₂, Al₂O₃, and Fe₂O₃ were mostly or completely associated with non-toxic outcomes. Toxic observations were generally associated with higher core size, positive surface charge, higher dosage, and longer exposure time. Machine learning models, including Logistic Regression, Random Forest, and Gradient Boosting, were applied for toxicity classification. Gradient Boosting achieved the best performance with 91.6% accuracy, 83.5% F1-score, and 0.976 ROC-AUC, followed closely by Random Forest. Feature importance analysis identified dosage as the most influential predictor of toxicity. Overall, the findings suggest that ensemble-based machine learning models can support preliminary in silico nanotoxicity screening and safer metal oxide nanoparticle design.