Towards a generalized toxicity prediction model for oxide nanomaterials using integrated data from different sources.

Area of Science:

• Nanomaterial science
• Toxicology
• Computational chemistry

Background:

• Assessing the toxicity of nanomaterials is crucial for safe application.
• Existing quantitative structure-activity relationship (QSAR) models have limitations in applicability.
• A generalized model is needed for diverse oxide nanomaterials.

Purpose of the Study:

• To develop a generalized toxicity classification model for seven oxide nanomaterials.
• To compare the performance of different machine learning algorithms and preprocessing techniques.
• To identify key physicochemical properties influencing nanomaterial toxicity.

Main Methods:

• Data extraction from literature and quality screening based on physicochemical properties.
• Application of preprocessing techniques including synthetic minority over-sampling technique (SMOTE) for class imbalance.
• Development and comparison of classification models: generalized linear model, support vector machine, random forest, and neural network.
• Applicability domain assessment using k-nearest neighbours (k-NN) algorithm.

Main Results:

• The neural network model, trained on a balanced dataset, demonstrated the highest predictive performance.
• Key attributes influencing toxicity were identified as dose, formation enthalpy, exposure time, and hydrodynamic size.
• The k-NN algorithm effectively defined the applicability domain of the developed model.

Conclusions:

• A generalized toxicity classification model for oxide nanomaterials was successfully developed.
• The neural network model offers superior predictive performance and a broader applicability domain compared to traditional QSAR models.
• The model provides valuable insights into the relationship between physicochemical properties and nanomaterial toxicity, aiding in risk assessment.