Geometric-Evolutionary Deep Learning Decodes the Human GPCR-Metabolome Interactome and Enables Systematic De-Orphanization

Abstract

G protein-coupled receptors (GPCRs) are the largest class of drug targets, yet hundreds of orphan GPCRs lack known endogenous ligands, limiting our understanding of human physiology and therapeutic development. Existing computational approaches often fail to generalize to these unseen targets due to a reliance on target-specific priors and linear feature integration. Here, we present G-LEAP, a deep learning framework that learns generalizable principles of GPCR-ligand recognition by synergizing evolutionary protein language models with 3D-aware geometric molecular representations. By implementing a bilinear interaction module, G-LEAP explicitly models non-linear cross-modal interactions and achieves superior generalization on stringent benchmarks, outperforming state-of-the-art methods by 22.4% in error reduction. Crucially, G-LEAP demonstrates robust chemical discrimination by effectively distinguishing active ligands from property-matched physicochemical decoys in the DUDE-Z benchmark and correctly rejecting 67% of hard-negative artifacts prioritized by physics-based docking simulations. Leveraging this capacity, we constructed a comprehensive atlas of over 120 million predicted interactions between 217,000 human metabolites and the GPCR superfamily, which ranked the true endogenous ligand within the top 1% of candidates for 33.7% of known pairs and identified putative orphan ligands validated by significant tissue-specific co-expression with their biosynthetic enzymes. Furthermore, large-scale virtual screening retrieved potent hits with novel chemical scaffolds distinct from known GPCR ligands, demonstrating robust scaffold hopping. G-LEAP thus provides a systematic and biologically validated platform to accelerate de-orphanization and expand the therapeutic chemical space.