Total: 1
G protein-coupled receptors (GPCRs) are important cell surface receptors involved in numerous physiological processes. Although peptides are the cognate ligands for many of these receptors, identifying endogenous peptide agonists for GPCRs remains a significant challenge. Deep learning-based protein structure prediction algorithms, such as AlphaFold (AF) have utility in non-structural tasks including protein-protein interaction prediction, suggesting they may be useful for predicting GPCR-peptide agonist interactions. Leveraging a dataset of experimentally validated agonist and non-agonist GPCR-peptide interactions from Caenorhabditis elegans, we show that AF-Multimer confidence metrics enable partial discrimination between GPCR-agonist and non-agonist complexes. To better reflect agonist-bound conformations, AF-Multistate templates are used to produce active-state GPCR-peptide complexes, improving discriminatory power. Embeddings from the final hidden layer of AF-Multimer's neural network, which capture structural and interaction patterns, were used to train random forest classifiers to assess whether AF-Multimer protein representations can distinguish agonist from non-agonist complexes. Feature performance analysis reveals that AF-Multimer's pair representations outperform single representations, with distinct subregions of the pair representation providing complementary predictive signals. Building on these findings, we developed DeorphaNN—a graph neural network that integrates active-state GPCR-peptide structural predictions, interatomic interactions, and pair representations to predict agonist identity. DeorphaNN's predictive utility generalizes to datasets outside of C. elegans, including annelids and humans, and experimental validation of predicted agonists for two orphan GPCRs uncovers their cognate agonists. Our approach offers a resource to accelerate GPCR deorphanization through the in silico identification of receptor-agonist candidates for AI-guided experimental validation.