Bayesian Neural Networks for Uncertainty‑Aware Toxicity Prediction of Metallic Nanoparticles

Nanoparticle cytotoxicity prediction remains challenging due to the complex interplay of physicochemical properties and experimental conditions, compounded by the critical need for uncertainty quantification in safety assessments. Here, we present a hybrid computational framework integrating Quantitative Structure-Activity Relationship (QSAR) feature engineering with Bayesian Neural Networks (BNNs) to predict metallic nanoparticle toxicity while explicitly quantifying predictive uncertainty. Using a curated dataset of 1,363 experimental records (25 nanoparticle formulations) from multiple in vitro studies, with every model evaluated on formulations unseen during training, we implemented a three-stage pipeline: (i) domain-guided preprocessing with multicollinearity assessment, (ii) benchmarking of deterministic classifiers (Random Forest, SVM, Multi-Layer Perceptron and three simpler models), and (iii) variational Bayesian inference for probabilistic classification. In repeated formulation-grouped cross-validation Random Forest discriminated best (ROC-AUC = 0.974), followed by logistic regression (0.967), which was better calibrated on held-out formulations (ECE 0.046 versus 0.130). The BNN discriminated well (ROC-AUC = 0.979) but underestimated toxicity at the 0.5 threshold (recall 0.520), while providing individual prediction uncertainty estimates from the variational posterior. Predictive uncertainty identified most misclassified predictions (AUROC = 0.92) but did not rise for samples outside the training descriptor space, which were detected by leverage. Dual-criterion applicability domain filtering (leverage and uncertainty thresholds) retained 44.2% of test samples (52 of the 125 toxic ones), achieving 0.942 recall and a 0.058 false negative rate, and flagged the remaining 55.8% for confirmatory testing. Impact statement: uncertainty-aware models evaluated on unseen formulations can separate nanoparticle cytotoxicity predictions usable for screening from those that require confirmatory testing.