Machine learning-driven QSAR models for predicting the mixture toxicity of nanoparticles.

Area of Science:

• Environmental chemistry and toxicology
• Computational toxicology and cheminformatics
• Nanotechnology and risk assessment

Background:

• Predicting the mixture toxicity of engineered nanoparticles (ENPs) is complex.
• In silico methods, particularly machine learning (ML), offer a promising strategy for toxicity prediction.
• Existing component-based models have limitations in predicting complex mixture effects.

Purpose of the Study:

• To develop and compare ML-based quantitative structure-activity relationship (QSAR) models for predicting the combined toxicity of metallic ENPs.
• To evaluate the performance of ML models against traditional component-based mixture models.
• To provide a robust methodological framework for the ecological risk assessment of ENP mixtures.

Main Methods:

• Combined in-house and literature toxicity data for seven metallic ENPs against Escherichia coli.
• Applied machine learning techniques: Support Vector Machine (SVM) and Neural Network (NN).
• Developed and validated QSAR models, comparing them with Independent Action and Concentration Addition models.

Main Results:

• Developed several high-performing SVM-QSAR and NN-QSAR models.
• An NN-based QSAR model using specific molecular descriptors achieved excellent predictive power (R²test = 0.908–0.911).
• The developed QSAR models demonstrated superior performance compared to component-based mixture models and were within their applicability domain.

Conclusions:

• Machine learning, specifically QSAR modeling, is effective for predicting the mixture toxicity of ENPs.
• The developed models offer a reliable alternative to traditional methods for assessing ENP mixture risks.
• This approach provides a theoretical and methodological foundation for ecological risk assessment of ENP mixtures.