Predicting and investigating cytotoxicity of nanoparticles by translucent machine learning

Area of Science:

• Environmental Science
• Toxicology
• Computational Science

Background:

• Engineered nanoparticles (ENPs) have potential applications but face safety concerns hindering commercialization.
• Understanding the complex relationships between ENPs, organisms, and the environment is crucial for risk assessment.
• Existing machine learning models for nanotoxicity often act as 'black boxes,' limiting trust and scientific interpretability.

Purpose of the Study:

• To develop a transparent approach for uncovering causal structures in nanotoxicity datasets.
• To enhance trust in machine learning models by providing interpretable predictions.
• To facilitate a deeper understanding of the factors influencing ENP toxicity.

Main Methods:

• Utilized mutually validated and model-agnostic interpretation methods to analyze nanotoxicity data.
• Employed feature importance, feature effects, and feature interactions to explain model predictions.
• Demonstrated the approach through case studies on cadmium-containing quantum dots and metal oxide nanoparticles.

Main Results:

• Identified key features correlating with cytotoxicity for specific ENPs.
• Explained the influence of these features on model predictions and their interactions.
• Showcased how interpretation methods can aid in model validation and dataset quality assessment.

Conclusions:

• The integrated approach of machine learning and interpretation methods offers a transparent roadmap for predicting ENP toxicity.
• This method enhances understanding of ENP-organism-environment interactions, fostering trust and safer applications.
• The approach aids in identifying critical toxicity drivers and validating predictive models in nanotoxicology.