A two-stage stacking machine learning framework for predicting metal oxide nanoparticle toxicity and biological outcomes.

Metal oxide nanoparticles (MONPs) raise growing cytotoxicity concerns, yet experimental assessment requires multiple time-consuming biological assays. We developed a leakage-free two-stage machine learning framework predicting MONP toxicity and biological endpoints from physicochemical features. In Stage 1, a stacking ensemble (ExtraTrees + GradientBoosting + RandomForest + HistGradientBoosting → LogisticRegression meta-learner) trained on 304 KONA dataset records using out-of-fold (OOF) probability generation achieved Accuracy = 0.87 and ROC-AUC = 0.91 on held-out data, and ROC-AUC = 0.76 on an independent external validation set. A three-way ablation study revealed that IC50-derived toxicity class substantially improves Stage 2 regression (ΔR2 = +0.22 for ROS; +0.12 for Membrane Damage), while out-of-fold-predicted class provides negligible improvement with toxicity classification. In Stage 2, ordinal classification (Low/Moderate/High) with literature-based thresholds replaced continuous regression. BaggingClassifier and GradientBoostingClassifier achieved AUC = 0.82-0.86 across four endpoints. Grouped permutation importance and LIME identified nanoparticle composition, surface chemistry, core size, and surface area as dominant predictors, consistent with ion dissolution and ROS-mediated mechanisms. This framework reduces experimental burden while providing interpretable, generalizable MONP cytotoxicity predictions.