Prediction of selective estrogen receptor beta agonist using open data and machine learning approach

Area of Science:

• Pharmacology and Cheminformatics
• Computational Drug Discovery

Background:

• Estrogen receptors (ERs) are key regulators of human physiological processes.
• ERs are validated drug targets for diseases like breast cancer and osteoporosis.
• Developing selective ER-beta (ER-β) ligands offers a promising therapeutic strategy with potentially fewer side effects.

Purpose of the Study:

• To develop in silico quantitative structure-activity relationship (QSAR) models for ER-β.
• To identify selective ER-β agonists using machine learning (ML) approaches.

Main Methods:

• Utilized public chemogenomics data for ER-β activity and chemical structures.
• Employed four fingerprinting methods (MACCS, PubChem, 2D atom pairs, CDK extended) as molecular descriptors.
• Trained four ML classifiers: Naïve Bayesian, k-nearest neighbor, random forest, and support vector machine.
• Validated models using 5-fold cross-validation.

Main Results:

• Achieved classification accuracies ranging from 77.10% to 88.34%.
• Obtained Area Under the ROC Curve values between 0.8151 and 0.9475.
• Identified random forest and support vector machine as superior classifiers for ER-β agonists.
• Found Chemistry Development Kit extended and MACCS fingerprints effective for structural representation.

Conclusions:

• The integration of fingerprinting methods and ML provides robust predictive models for ER-β agonists.
• These models show potential for identifying novel selective ER-β agonists in drug discovery efforts.