Predicting the protein targets for athletic performance-enhancing substances.

Area of Science:

• Computational chemistry
• Pharmacology
• Anti-doping science

Background:

• The World Anti-Doping Agency (WADA) maintains a Prohibited List of banned substances and methods.
• Automated, fast, and cost-effective identification of performance-enhancing substances is needed.
• Existing methods for identifying doping agents can be improved through advanced computational approaches.

Purpose of the Study:

• To build a database model using ChEMBL experimental data (structure and activity) for predicting doping-relevant substance interactions.
• To identify on-target and off-target interactions of compounds with targets relevant to sport doping.
• To develop a computational tool for the automated detection of potential doping agents.

Main Methods:

• Screening the ChEMBL database and filtering activity records (Ki, Kd, EC50, etc.) to define "active" and "inactive" compounds.
• Applying structure-based clustering to group compounds with similar scaffolds and bioactivities.
• Utilizing the Parzen-Rosenblatt machine learning approach for predictive modeling and validation.

Main Results:

• Refined families of compounds were created, sharing common scaffolds and bioactivities against specific targets.
• The machine learning model demonstrated significant increases in prediction accuracy compared to baseline methods.
• Validation showed high accuracy, with 66.98% of queries correctly predicting the parent family and 87.25% within the top four predictions.

Conclusions:

• The validated machine learning approach effectively predicts compound families and their bioactivities.
• The model was successfully applied to identify protein targets associated with WADA prohibited classes.
• The study provides a framework for predicting bioactivity for compounds lacking experimental data, encouraging future experimental validation.