PHYSICOCHEMICAL FEATURE-DRIVEN NANOTOXICITY PREDICTION USING SUPERVISED MACHINE LEARNING ALGORITHMS

Abstract

The widespread use of metal oxide nanoparticles across various industries has raised significant concerns regarding their potential toxicity. Conventional toxicological assessment methods remain time-intensive, costly, and limited in scalability. In this study, a machine learning–based framework was developed to classify nanoparticle toxicity using physicochemical descriptors. A dataset containing nine key features such as dosage, surface area, and core size of the nanoparticles was employed to train and evaluate six supervised learning algorithms: Decision Tree, Random Forest, Gradient Boosting, Logistic Regression, Support Vector Machine, and K-Nearest Neighbors. The analysis focused on five widely used metal oxide nanoparticles: Fe₂O₃, TiO₂, ZnO, CuO, and Al₂O₃.The Decision Tree model achieved the highest classification accuracy (96.05%) and was noted for its interpretability and transparent decision rules. Model performance was assessed using ROC curves, precision-recall analysis, confusion matrices, and residual distributions, all of which confirmed the model’s robustness and generalization capability. Feature importance analysis indicated that dosage, number of oxygen atoms, and electron affinity were the most significant predictors of toxicity.This approach enables accurate and interpretable nanotoxicity prediction and may serve as a valuable tool for risk-based assessment and the design of safer nanomaterials. The findings support the integration of data-driven methodologies into toxicological evaluation workflows.