Comparison of Machine Learning Models for the Androgen Receptor

Area of Science:

• Endocrinology
• Toxicology
• Computational Chemistry

Background:

• The androgen receptor (AR) is crucial for reproductive and neurological development.
• Altered AR signaling is linked to endocrine disruption across generations.
• Existing U.S. Environmental Protection Agency (EPA) models require in vitro data for AR bioactivity assessment.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting AR bioactivity.
• To explore prospective prediction of AR agonist and antagonist activity from molecular structure.
• To assess the utility of Bayesian machine learning for endocrine disruption research.

Main Methods:

• Applied Bayesian machine learning algorithms to AR signaling pathway data.
• Utilized proprietary software, Assay Central, for data analysis.
• Evaluated model performance using 5-fold cross-validation and external reference chemical predictions.

Main Results:

• Bayesian machine learning models demonstrated predictive accuracy for AR bioactivity.
• Models were trained on continuous AC50 data from the ToxCast/Tox21 database (February 2019 release).
• Machine learning predictions were comparable to published EPA results.

Conclusions:

• Machine learning offers a viable method for predicting AR-mediated bioactivity.
• This approach can prioritize compounds for endocrine disruption research.
• The methodology is applicable to other endocrine disruption targets.