Application of Machine Learning in Nanotoxicology: A Critical Review and Perspective

Area of Science:

• Nanomaterial safety assessment
• Computational toxicology
• Predictive modeling

Background:

• Nanomaterials (NMs) production and use raise health and environmental concerns.
• Laboratory toxicity testing is costly, slow, and struggles to keep pace with new NMs.
• In silico methods using machine learning offer a promising alternative for predicting NM toxicity.

Purpose of the Study:

• To review and enhance the development of in silico predictive models for nanomaterial toxicity.
• To improve model representativeness and performance through data curation, descriptor selection, and algorithm choice.
• To provide a recommended workflow and best practices for developing reliable and interpretable NM toxicity models.

Main Methods:

• Statistical evaluation of existing literature on in silico NM toxicity models.
• Analysis of key aspects: data set curation, descriptor selection, task type, algorithm choice, and model evaluation.
• Identification of challenges and future perspectives in the field.

Main Results:

• Existing reviews and OECD guidance provide a foundation for building in silico NM models.
• Significant room for improvement exists in model representativeness and performance.
• Mechanistic interpretation is key to building stakeholder confidence in predictive models.

Conclusions:

• Enhanced in silico models are vital for effective risk assessment and safe-by-design development of NMs.
• A structured workflow and adherence to best practices will lead to more predictive and reliable models.
• Future work should focus on improving model interpretability and regulatory acceptance.