Machine Learning Models for Estrogen Receptor Bioactivity and Endocrine Disruption Prediction

Area of Science:

• Computational toxicology
• Endocrine disruption
• Machine learning applications

Background:

• The U.S. Environmental Protection Agency (EPA) uses mathematical models for endocrine disruptor testing, relying on in vitro data.
• Predictive models for estrogen receptor (ER) activity are crucial for chemical prioritization.
• Machine learning offers prospective prediction capabilities directly from molecular structure, bypassing the need for in vitro data.

Purpose of the Study:

• To develop and evaluate Bayesian machine learning models for predicting ER agonism.
• To compare the performance of these machine learning models against EPA's existing mathematical models.
• To assess the utility of machine learning in prioritizing chemicals for in vitro and in vivo endocrine disruption testing.

Main Methods:

• Generation and evaluation of Bayesian machine learning models using Assay Central software.
• Utilized multiple data types, grouped by EPA's ER agonist pathway model.
• Compared external predictions on in vitro and in vivo reference chemicals with previous EPA model publications.

Main Results:

• External predictions from the machine learning models were comparable or superior to previous EPA mathematical model studies.
• Assay Central demonstrated similar performance to six additional algorithms on training data, but with reduced computational cost.
• The study validated the capability of machine learning for prioritizing chemicals for ER agonism testing.

Conclusions:

• Machine learning models provide a powerful and efficient alternative for predicting ER agonism.
• These models can effectively prioritize chemicals for subsequent in vitro and in vivo endocrine disruption assays.
• The findings support the broader application of machine learning in regulatory toxicology and chemical safety assessment.