NanoTox

NanoTox predicts cytotoxicity and ecotoxicity of metal-oxide nanoparticles (MeOx NPs) using machine-learning models trained on intrinsic and extrinsic physico-chemical properties to support hazard assessment.

Key Features:

• Feature Space Formulation: Constructs a feature space from intrinsic and extrinsic physico-chemical properties of nanoparticles while excluding in vitro characteristics such as cell line, cell type, and assay method to promote generalizability across experimental setups.
• Feature Optimization: Applies variance inflation analysis to refine the correlation structure and produce a minimal, effective set of predictive features for nanoparticle cytotoxicity.
• Machine Learning Models: Implements hyperparameter-tuned random forest, neural networks with one hidden layer, support vector machines, and logistic regression to map a normalized feature space to toxicity classes, achieving over 96% balanced accuracy on unseen test sets.
• Interpretability and Applicability: Identifies key predictor variables that influence MeOx NP cytotoxicity and performs an applicability check to ensure predictions lie within the training-data domain.
• Ensemble Classifier - NanoTox Pipeline: Integrates multiple model outputs via a majority-voting ensemble classifier named NanoTox to enhance prediction reliability for new, untested metal-oxide nanoparticles.

Scientific Applications:

• Hazard assessment of MeOx NPs: Provides predictive cytotoxicity and ecotoxicity information for metal-oxide nanoparticles to inform hazard characterization.
• Toxicity prediction for untested oxides: Enables prediction of toxicity for previously untested metal-oxide nanoparticles based on physico-chemical properties.
• Support for research and regulatory decision-making: Supplies model-based evidence to assist researchers and regulators in evaluating nanoparticle risks and guiding development choices.

Methodology:

Constructs a feature space from intrinsic and extrinsic physico-chemical properties excluding in vitro variables; applies variance inflation analysis for feature selection; normalizes the feature space; trains hyperparameter-tuned random forest, one-hidden-layer neural networks, support vector machines, and logistic regression models to map features to toxicity classes; combines model outputs via a majority-voting ensemble named NanoTox; identifies key predictors and applies an applicability-domain check (reported >96% balanced accuracy on unseen test sets).