GPCR-Filter: a deep learning framework for efficient and precise GPCR modulator discovery

TL;DR

GPCRs are central drug targets but challenging to modulate due to complex allosteric signaling. GPCR-Filter integrates ESM-3-based GPCR sequence embeddings with ligand graph features through a cross-attention decoder to predict receptor–ligand modulatory interactions, trained on a large GPCR–ligand dataset. It demonstrates strong predictive performance across in-distribution and out-of-distribution splits and provides mechanistic interpretability via attention-based pocket localization, with wet-lab validation identifying micromolar agonists for 5-HT1A. This work offers a scalable, interpretable AI-assisted screening component to accelerate GPCR-targeted drug discovery, bridging sequence, structure-agnostic data with functional modulation.

Abstract

G protein-coupled receptors (GPCRs) govern diverse physiological processes and are central to modern pharmacology. Yet discovering GPCR modulators remains challenging because receptor activation often arises from complex allosteric effects rather than direct binding affinity, and conventional assays are slow, costly, and not optimized for capturing these dynamics. Here we present GPCR-Filter, a deep learning framework specifically developed for GPCR modulator discovery. We assembled a high-quality dataset of over 90,000 experimentally validated GPCR-ligand pairs, providing a robust foundation for training and evaluation. GPCR-Filter integrates the ESM-3 protein language model for high-fidelity GPCR sequence representations with graph neural networks that encode ligand structures, coupled through an attention-based fusion mechanism that learns receptor-ligand functional relationships. Across multiple evaluation settings, GPCR-Filter consistently outperforms state-of-the-art compound-protein interaction models and exhibits strong generalization to unseen receptors and ligands. Notably, the model successfully identified micromolar-level agonists of the 5-HT\textsubscript{1A} receptor with distinct chemical frameworks. These results establish GPCR-Filter as a scalable and effective computational approach for GPCR modulator discovery, advancing AI-assisted drug development for complex signaling systems.

Figures (6)

• Figure 1: Application of GPCR-Filter and overview of its architecture.(a) Workflow illustrating how GPCR-Filter integrates into GPCR modulator discovery. After ligand preparation and initial structure-based virtual screening, GPCR-Filter uses only GPCR sequences to further refine docking outputs and prioritize candidates for downstream activation assays. (b) Schematic of the GPCR-Filter architecture. GPCR sequences are embedded into per-residue representations using a pretrained protein language model, while ligand SMILES are encoded into per-atom features through a molecular graph neural network. A cross-attention module couples ligand and receptor representations, and the fused features are aggregated to produce a final interaction probability.
• Figure 2: Dataset split strategies and performance overview.(a) Random Split: Each GPCR (receptor icon) and its ligands are randomly assigned to training and testing sets. Both receptors and ligands can appear in both sets, simulating a fully mixed in-distribution evaluation. (b) Intra-Target Split: Each receptor is present in both training and testing, but its associated ligands are divided into mutually exclusive subsets. This evaluates generalization to unseen ligands for the same receptor. (c) Inter-Target Split: The receptor set is partitioned into disjoint training and held-out target subsets (e.g., 9:1). Validation examples are drawn from the held-out subset to keep target identities disjoint between training and evaluation. On the right, the ROC curves illustrate the discriminative performance of GPCR-Filter and baseline models under the three data partitioning protocols. Receptor and ligand icons are for schematic illustration only, with color denoting training or testing samples.
• Figure 3: Binding-pocket analysis across two GPCR complexes.(a–c) correspond to PDB 9bsb; (d–f) correspond to PDB 9jcl. (a,d) Model prediction scores under the three evaluation settings (Random, Intra-target, Inter-target). All predicted probabilities exceed 0.5, indicating consistent positive predictions across splits. (b,e) Top-20 attended residues ranked under the three evaluation settings. Each entry is reported as index–residue; highlighted cells indicate residues that are repeatedly prioritized across splits. (c,f) Spatial mapping of the Inter-Target model’s Top-20 attended residues onto the corresponding structure and binding pocket. Inset panels show the ligand in sticks and the pocket region enlarged (red box), illustrating which high-attention residues lie within the crystallographic binding site. All residue indices correspond to the numbering used in the PDB sequences.
• Figure 4: Experimental validation of GPCR-Filter-predicted agonists on the 5-HT$_{1\mathrm{A}}$ receptor.(a) Chemical structures of 5-HT and the four validated hits. (b) Single-concentration screening at 30 $\mu$M using a GloSensor-cAMP assay. Bars represent luminescence (LUM) values for individual compounds; dashed lines denote Forskolin (black, baseline) and 5-HT (red, canonical agonist). Four compounds prioritized by GPCR-Filter (D24, D29, D34, D47) exhibited reduced LUM, indicative of receptor activation. (c) Normalized concentration--response curves for 5-HT and the four hits, showing comparable or higher maximal effects (E$_\mathrm{max}$) but right-shifted potencies (higher EC$_{50}$), indicating increased efficacy but lower potency relative to 5-HT. Together, these results confirm that GPCR-Filter successfully identified four true agonists activating 5-HT$_{1\mathrm{A}}$ in a dose-dependent manner.
• Figure S1: Distribution of target occurrences per GPCR. Binned frequency-of-frequency histogram summarizing the curated GPCR--ligand dataset. The x-axis represents the number of ligand occurrences per GPCR (binned), and the y-axis represents the number of GPCRs (population) within each bin. The pronounced long-tail pattern indicates that while most GPCRs interact with only a few ligands, a small subset of receptors accounts for a disproportionately large number of interactions.