Prediction of GPCR activity using machine learning.

Area of Science:

• Computational Biology and Cheminformatics
• Pharmacology and Drug Discovery

Background:

• G protein-coupled receptors (GPCRs) are crucial drug targets, yet their structure-activity-function relationships remain mechanistically unclear.
• Developing specific drugs with minimal side effects requires quantitative understanding of GPCR structural features and their link to receptor activation states.

Purpose of the Study:

• To develop and apply machine learning (ML) approaches for predicting GPCR conformation states.
• To correlate GPCR structure with predicted activity levels.
• To interpret ML models to identify key structural features influencing GPCR conformations.

Main Methods:

• Developed three distinct ML approaches: XGBoost for interpretability, 3D convolutional neural networks (CNNs) for minimal feature engineering, and graph neural networks (GNNs) for graph-based protein structure representation.
• Leveraged the unique strengths of each ML model to analyze GPCR structural data.

Main Results:

• Achieved high accuracy (91%-95%) in predicting GPCR activation states using the developed ML models.
• Predicted GPCR activity levels with a low mean absolute error (MAE) ranging from 7.15 to 10.58.
• Model interpretation identified critical structural features responsible for differentiating GPCR conformations.

Conclusions:

• The study successfully demonstrates the utility of multiple ML approaches in predicting GPCR conformation and activity.
• These findings provide a quantitative framework for understanding GPCR structure-activity relationships, aiding in the design of novel therapeutics.
• The interpretability of the ML models offers valuable insights into GPCR structural dynamics and activation mechanisms.