Robust classification frameworks for oxide nanomaterials assisted by machine learning

Accurate prediction of oxide nanomaterial toxicity is vital for advancing material safety and applications. Machine learning (ML) methods provide robust tools to explore complex relationships between material properties and biological responses. This study utilized a comprehensive dataset integrating physicochemical (e.g., hydrodynamic size, core size, surface area, surface charge), quantum mechanical (e.g., valence/conduction band energies, electronegativity, formation enthalpy), and toxicity-related parameters (e.g., assay methods, cell species, cell type, exposure type, dose). Tree-based ML models, including Decision Tree, KNN, LightGBM, XGBoost, BaggingClassifier, and logistic regression variants, were applied to predict cell viability. The BaggingClassifier outperformed others, achieving high accuracy, precision, and recall. Hydrophobicity (Hsf, − 0.46), exposure time (− 0.26), and core size (− 0.21) negatively impacted viability, while surface area (0.23) and MeO (0.26) showed positive correlations. SHAP analysis highlighted dose and Hsf as key predictors. This study addresses data challenges, offering insights for safer nanomaterial development.