Integrating structure annotation and machine learning approaches to develop graphene toxicity models.

Area of Science:

• Nanotechnology
• Computational Toxicology
• Materials Science

Background:

• Nanomaterials (NMs) offer efficient and cost-effective applications but raise nanotoxicity concerns.
• Traditional animal testing for nanotoxicity is costly and time-consuming.
• Machine learning (ML) offers a promising alternative for nanotoxicity evaluation using nanostructure features.

Purpose of the Study:

• To develop a novel nanostructure annotation strategy for complex nanomaterials like graphenes.
• To construct a virtual graphene library for machine learning-based nanotoxicity modeling.
• To generate high-quality nanodescriptors for nanoinformatics studies.

Main Methods:

• Generated irregular graphene structures by modifying virtual nanosheets.
• Digitalized and annotated nanostructures for quantitative analysis.
• Computed geometrical nanodescriptors using Delaunay tessellation for ML modeling.
• Built and validated partial least square regression (PLSR) models using leave-one-out cross-validation (LOOCV).

Main Results:

• Developed predictive PLSR models for graphenes with good accuracy (R² ranging from 0.558 to 0.822).
• Demonstrated the models' predictivity across four toxicity-related endpoints.
• Validated the effectiveness of the nanostructure annotation strategy.

Conclusions:

• The novel nanostructure annotation strategy enables the generation of high-quality nanodescriptors.
• This approach facilitates the development of ML models for nanotoxicity prediction.
• The methodology is applicable to nanoinformatics studies of graphenes and other nanomaterials.