DeorphaNN: Virtual screening of GPCR peptide agonists using AlphaFold-predicted active-state complexes and deep...

Area of Science:

• Biochemistry
• Computational Biology
• Pharmacology

Background:

• Peptide-activated G protein-coupled receptors (GPCRs) are crucial for physiological regulation.
• Identifying endogenous peptide agonists for GPCRs is challenging, especially for orphan GPCRs lacking characterized homologs.
• Existing methods offer limited predictive power for novel peptide-GPCR pairings.

Purpose of the Study:

• To develop a computational method for predicting peptide agonists of GPCRs.
• To leverage structural predictions and deep learning for prioritizing experimental screening.
• To accelerate the deorphanization of orphan GPCRs.

Main Methods:

• Utilized a dataset of experimentally validated peptide-GPCR interactions from *Caenorhabditis elegans*.
• Employed AlphaFold-multimer (AF-multimer) confidence metrics and active-state templates for discriminating interactions.
• Developed DeorphaNN, a graph neural network integrating structural predictions, interatomic interactions, and deep learning embeddings.

Main Results:

• AF-multimer confidence metrics showed partial discrimination of agonist complexes, improved by active-state templates.
• Feature analysis indicated that pair representations in AF-multimer provided superior signals compared to single representations.
• DeorphaNN successfully generalized across species, demonstrating strong performance on annelid and human benchmarks.
• Experimental validation confirmed predicted agonists for two orphan GPCRs.

Conclusions:

• DeorphaNN effectively prioritizes putative peptide agonists for GPCRs.
• The developed method accelerates the challenging process of peptide-GPCR deorphanization.
• This approach holds significant potential for discovering novel peptide-GPCR interactions across diverse species.