Mechanism Based Hierarchical Machine Learning for High-Throughput Quantitative Prediction of Estrogenic, Androgenic,...

Area of Science:

• Environmental Chemistry
• Toxicology
• Computational Chemistry

Background:

• Quantitative high-throughput models for endocrine-disrupting chemicals (EDCs) are limited by data issues and complexity.
• Existing models struggle with predicting estrogenic, androgenic, and thyroid (EAT) disruption accurately.

Purpose of the Study:

• To develop a mechanistically informed hierarchical framework for quantitative prediction of EAT-disruption activities.
• To improve the accuracy and interpretability of endocrine potency predictions.

Main Methods:

• Data refinement: Five-step curation to create a high-confidence dataset, removing false positives and negatives.
• Mechanistic clustering: Fragment-based approach to classify EAT activity modes.
• Quantitative modeling: Cluster-specific ensemble regressor for potency estimation, informed by molecular simulations.

Main Results:

• Achieved improved model performance (R² = 0.72-0.78, RMSE = 0.22-0.48 log₁₀(μM)) compared to conventional methods.
• Identified key structural features of potent EAT agonists (aromatic cores, polar substituents) and antagonists (flexible chains, rigid scaffolds).
• Molecular simulations revealed mechanisms of receptor activation and disruption by agonists and antagonists.

Conclusions:

• The developed framework offers a next-generation strategy for accurate and interpretable endocrine potency prediction.
• Mechanistic insights enhance understanding of EDC-receptor interactions.
• This approach can aid in prioritizing chemicals for further testing and risk assessment.