Machine Learning in Nanoparticles Toxicology: Mechanisms, Predictive Modelling, and Risk Assessment Strategies

The increasing application of engineered nanoparticles (ENPs) in industry and medicine will require economic and ethically acceptable procedures for toxicological assessment. This paper employed supervised machine learning (ML) to model a new set of physicochemical data, correlating the 312 ENP samples contained in the eNano-Mapper database with ENP cytotoxicity. The major characteristics were particle size (mean = 68.5 nm), zeta potential (=18.6 mV), hydrodynamic diameter (102.3 nm), surface area, as well as solubility. Four approaches of ML: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), have been tested, and XGBoost demonstrated 91% accuracy, 0.90 precision, 0.89 recall, 0.89 F1-score, and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) = 0.94. The Logistic Regression and SVM models reported lower scores (accuracy: 78% and 80%, AUC: 0.81 and 0.83, respectively), indicating a diminished capacity to identify non-linear toxicological trends. The SHAP analysis provided zeta potential (0.146), hydrodynamic diameter (0.131), and particle size (0.112) as the best predictive factors of toxicity. Particle size showed a significant correlation with hydrodynamic diameter (r = 0.78), suggesting that future testing procedures could be redundant. There were only five entries (misclassifications), with the majority between the high and moderate toxicity categories. The research presents a compelling, fact-based alternative to animal testing, thereby driving the development of safer-by-design nanoparticles and enhancing risk assessment.