AI-driven hazard prioritization of plastic additives using Tox21 bioassays and self-supervised graph transformers

Area of Science:

• Environmental Chemistry
• Toxicology
• Artificial Intelligence

Background:

• Plastic additives can leach into the environment, posing potential toxicity risks.
• Over 400 high-volume plastic additives are listed by the ECHA Plastic Additives Initiative.
• Assessing the toxicity of these additives is crucial for environmental and human health.

Purpose of the Study:

• To screen the potential toxicity of numerous plastic additives.
• To apply deep learning models and Tox21 bioassay data for chemical hazard assessment.
• To identify highly active plastic additives and investigate their existing hazard data.

Main Methods:

• Collected and utilized the Tox21 dataset with extensive bioactivity profiles.
• Trained and fine-tuned the GROVER deep learning algorithm on Tox21 bioassay data.
• Evaluated model performance using the F1 score and compared with baseline algorithms.
• Investigated hazard information and GHS classifications for identified active chemicals.

Main Results:

• The GROVER model demonstrated superior performance compared to traditional machine learning algorithms.
• Identified 78 highly active chemicals among 171 screened plastic additives.
• Revealed significant data gaps in hazard information for potentially toxic plastic additives.
• Highlighted the utility of AI models as New Approach Methodologies (NAMs) for hazard assessment.

Conclusions:

• Advanced AI models like GROVER can effectively screen plastic additives for toxicity.
• The study identified key plastic additives requiring further toxicological investigation.
• This approach supports regulatory decision-making and aligns with the 3Rs principle for animal testing.
• Modernizing chemical hazard assessment through AI is feasible and beneficial.