Toxic Alerts of Endocrine Disruption Revealed by Explainable Artificial Intelligence.

Area of Science:

• Environmental Toxicology
• Computational Chemistry
• Machine Learning

Background:

• Endocrine-disrupting chemicals (EDCs) pose risks to human health and ecosystems.
• Identifying specific chemical substructures responsible for endocrine disruption is challenging.
• Existing methods often lack the interpretability needed for regulatory acceptance.

Purpose of the Study:

• To apply machine learning for identifying toxic substructures causing endocrine disruption.
• To enhance the interpretability of predictive models for endocrine disruption.
• To provide specific toxic alerts for risk assessment of chemical compounds.

Main Methods:

• Utilized the local interpretable model-agnostic explanation (LIME) method.
• Employed a random forest classifier trained on curated TOX21 datasets.
• Applied explainable models to EDC and EDKB-FDA datasets for substructure identification.

Main Results:

• Identified specific substructures (toxic alerts) linked to endocrine disruption.
• Achieved stable, specific, and consistent explanations for model predictions.
• Unveiled substructures associated with five key endocrine receptors: androgen, estrogen, aryl hydrocarbon, aromatase, and peroxisome proliferator-activated receptors.

Conclusions:

• The LIME approach significantly improves the interpretability of machine learning models in toxicology.
• Identified toxic alerts like thiophosphate, sulfamate, anilide, carbamate, sulfamide, and thiocyanate.
• Provides a foundation for better understanding and mitigating risks associated with endocrine disruption.