TY - JOUR
T1 - AI-driven hazard prioritization of plastic additives using Tox21 bioassays and self-supervised graph transformers
AU - Kim, Donghyeon
AU - Lee, Eungyeong
AU - Yi, Youngmin
AU - Choi, Jinhee
N1 - Publisher Copyright:
© The Author(s) 2025. Published by Oxford University Press on behalf of the Society of Environmental Toxicology and Chemistry. All rights reserved. For commercial re-use, please contact [email protected] for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact [email protected]. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/pages/standard-publication-reuse-rights)
PY - 2026/9
Y1 - 2026/9
N2 - As plastics degrade into micro- and nano-sized particles, they can leach additive chemicals into the environment, potentially exerting greater toxicity than the polymer matrix itself. The European Chemicals Agency Plastic Additives Initiative has compiled a list of more than 400 plastic additives that are used in high volumes. This study aimed to screen the potential toxicity of these chemicals using Tox21 bioassays and deep learning models. To this end, we collected the Tox21 dataset, which provides extensive bioactivity profiles for more than 7,000 chemicals across various endpoints, including human nuclear receptor signaling and stress response pathways. We then trained deep learning models using experimental data from Tox21 bioassays. Specifically, we employed the GROVER algorithm, which was designed to overcome typical limitations of traditional graph neural networks by leveraging transformers and self-supervised pretraining. We fine-tuned the model on twelve Tox21 bioassay datasets, using the F1 score as the primary evaluation metric. As a result, the GROVER model outperformed baseline algorithms, including graph convolutional networks, random forest, support vector machines, and logistic regression. Using the fine-tuned GROVER models, we identified 78 highly active chemicals among 171 additives. For these active plastic additive chemicals, we also investigated existing hazard information (minimal oral point-of-departure) from the CompTox Chemical Dashboard and their Globally Harmonized System of Classification and Labelling of Chemicals information from PubChem database. This approach revealed significant data gaps for plastic additive chemicals with potential toxicity and can support regulatory decision-making. Collectively, this study provides a practical use case for applying cutting-edge AI models as new approach methodologies to modernize hazard assessment, in alignment with the 3Rs (Replacement, Reduction, Refinement) principle for animal testing.
AB - As plastics degrade into micro- and nano-sized particles, they can leach additive chemicals into the environment, potentially exerting greater toxicity than the polymer matrix itself. The European Chemicals Agency Plastic Additives Initiative has compiled a list of more than 400 plastic additives that are used in high volumes. This study aimed to screen the potential toxicity of these chemicals using Tox21 bioassays and deep learning models. To this end, we collected the Tox21 dataset, which provides extensive bioactivity profiles for more than 7,000 chemicals across various endpoints, including human nuclear receptor signaling and stress response pathways. We then trained deep learning models using experimental data from Tox21 bioassays. Specifically, we employed the GROVER algorithm, which was designed to overcome typical limitations of traditional graph neural networks by leveraging transformers and self-supervised pretraining. We fine-tuned the model on twelve Tox21 bioassay datasets, using the F1 score as the primary evaluation metric. As a result, the GROVER model outperformed baseline algorithms, including graph convolutional networks, random forest, support vector machines, and logistic regression. Using the fine-tuned GROVER models, we identified 78 highly active chemicals among 171 additives. For these active plastic additive chemicals, we also investigated existing hazard information (minimal oral point-of-departure) from the CompTox Chemical Dashboard and their Globally Harmonized System of Classification and Labelling of Chemicals information from PubChem database. This approach revealed significant data gaps for plastic additive chemicals with potential toxicity and can support regulatory decision-making. Collectively, this study provides a practical use case for applying cutting-edge AI models as new approach methodologies to modernize hazard assessment, in alignment with the 3Rs (Replacement, Reduction, Refinement) principle for animal testing.
KW - Artificial intelligence
KW - Deep learning
KW - New approach methodologies
KW - Plastic additive chemicals
KW - Toxicity Forecasting (ToxCast)
UR - https://www.scopus.com/pages/publications/105049598309
U2 - 10.1093/etojnl/vgaf228
DO - 10.1093/etojnl/vgaf228
M3 - Article
C2 - 40971996
AN - SCOPUS:105049598309
SN - 0730-7268
VL - 45
SP - 2447
EP - 2457
JO - Environmental Toxicology and Chemistry
JF - Environmental Toxicology and Chemistry
IS - 9
ER -