Evolutionary-guided advanced deep-learning architecture powers mammalian GPCRome agonist predictions

G-protein-coupled receptors constitute a highly conserved superfamily that orchestrates essential signaling processes across species. Contemporary computational approaches for predicting ligand-receptor interactions are hindered by restricted receptor coverage, limited incorporation of odorant receptors, and datasets that insufficiently capture cross-species diversity. Here, we present EvOlf, an evolutionary-guided deep-learning framework that integrates odorant and non-odorant receptors through an expansive dataset comprising 105,235 experimentally validated interactions across 24 mammalian species. EvOlf leverages hierarchical self-attention, multi-resolution feature encoding, and a fusion transformer to construct biologically grounded representations of ligand-receptor interactions, achieving state-of-the-art performance. EvOlf substantially outperformed existing prediction methods across an independent test dataset. Experimental validation identified three structurally diverse allosteric co-activators of the β1‑adrenergic receptor that significantly potentiated agonist-evoked signaling in human and rat cardiomyocytes. EvOlf provides a robust and scalable platform for receptor deorphanization and ligand discovery.