Machine Learning Consensus To Predict the Binding to the Androgen Receptor within the CoMPARA Project.

Area of Science:

• Computational toxicology
• Endocrinology
• Machine learning

Background:

• Nuclear androgen receptor (AR) is a critical target for endocrine disrupting chemicals (EDCs).
• EDCs interfere with hormonal regulation, leading to adverse health effects.
• Prioritizing experimental testing for a large number of compounds is crucial for chemical safety assessment.

Purpose of the Study:

• To develop and describe novel in silico models for identifying organic AR modulators.
• To support the Collaborative Modeling Project of Androgen Receptor Activity (CoMPARA) in prioritizing compound testing.
• To provide data-driven insights into structural features influencing AR binding, agonism, and antagonism.

Main Methods:

• A consensus machine learning approach combining Naive Bayes, Random Forest, and N-Nearest Neighbor models.
• Model training on 1687 ToxCast molecules using 11 in vitro assays.
• External validation using a set of 3,882 compounds.

Main Results:

• The developed models demonstrated robust and reliable predictions for AR binding.
• The approach successfully prioritized experimental testing for approximately 40,000 compounds within the CoMPARA project.
• Novel insights into structure-activity relationships for AR modulators were generated.

Conclusions:

• The in silico models offer an effective strategy for identifying potential AR modulators.
• This computational approach aligns with OECD principles for chemical safety testing.
• The models contribute to a better understanding of EDC interactions with the androgen receptor.