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

Peptide-activated G protein-coupled receptors (GPCRs) regulate physiological processes through interaction with neuropeptides and peptide hormones. Identifying endogenous peptide agonists remains challenging, as peptide-GPCR pairings often follow gene-family relationships that offer limited predictive insight for orphan GPCRs without characterized homologs. Using a dataset of experimentally validated peptide-GPCR interactions from Caenorhabditis elegans, we demonstrate that AF-multimer confidence metrics partially discriminate agonist from non-agonist complexes, with improved discrimination using AF-Multistate-derived active-state templates. Feature analysis revealed that AF-multimer's pair representations outperform single representations, with distinct subregions providing complementary signals. Leveraging these insights, we developed DeorphaNN, a graph neural network integrating active-state GPCR-peptide structural predictions, interatomic interactions, and deep learning embeddings to prioritize putative peptide agonists for experimental screening. DeorphaNN generalized across diverse species, as shown by performance on annelid and human retrospective benchmarks. Experimental validation confirmed predicted agonists for two orphan GPCRs, demonstrating its utility for accelerating peptide-GPCR deorphanization.