G Protein-Coupled Receptor Interaction Prediction Based on Deep Transfer Learning.

Area of Science:

• Pharmacology
• Computational Biology
• Drug Discovery

Background:

• G protein-coupled receptors (GPCRs) are crucial drug targets, implicated in numerous human diseases.
• Accurate prediction of GPCR interactions is vital for understanding their structural roles and designing effective therapeutics.
• Current machine learning methods for GPCR interaction prediction are limited by scarce labeled sample data.

Purpose of the Study:

• To develop an effective transfer learning method for predicting GPCR interactions despite limited labeled data.
• To improve the accuracy of GPCR interaction prediction for drug design and development.

Main Methods:

• A transfer learning framework utilizing sample similarity was developed.
• XGBoost was employed as a weak classifier, with TrAdaBoost and JS divergence for data weight initialization.
• A deep neural network with an attention mechanism was used for model training on constructed datasets.
• The method was validated using existing GPCR data for prediction tasks.

Main Results:

• The proposed transfer learning method significantly improved prediction accuracy.
• Specifically, in short-distance contact prediction, the method achieved an accuracy increase of 0.26 compared to similar existing approaches.
• Demonstrated the efficacy of transfer learning in addressing data scarcity for GPCR interaction prediction.

Conclusions:

• The developed transfer learning method offers a robust solution for predicting GPCR interactions with limited data.
• This approach facilitates more accurate drug design and enhances the understanding of GPCR roles in disease.
• The findings highlight the potential of transfer learning in computational drug discovery.