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

Jingjie Ning; Xiangzhen Shen; Li Hou; Shiyi Shen; Jiahao Yang; Junrui Li; Hong Shan; Sanan Wu; Sihan Gao; H. Eric Xu; Xinheng He

Source-linked AI summary

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

Jingjie Ning, Xiangzhen Shen, Li Hou, Shiyi Shen, Jiahao Yang, Junrui Li, Hong Shan, Sanan Wu, Sihan Gao, H. Eric Xu, Xinheng He

arXiv:2601.19149v2Jan 27, 2026cs.LGq-bio.QM

TL;DR

GPCR modulator discovery is limited by dynamic allostery, incomplete binding-function correspondence, and costly, low-throughput assays. GPCR-Filter combines GPCR sequence modeling, ligand graph representations, and attention-based fusion, and it outperforms state-of-the-art DTI models across evaluation settings while identifying validated 5-HT1A agonists. The study presents a scalable and interpretable virtual-screening framework, although negative sampling and ligand reuse leave important evaluation boundaries.

Problem
Dynamic GPCR conformations and allosteric mechanisms can decouple ligand binding from signaling efficacy, while conventional assays are costly, low-throughput, and labor-intensive.
Method
GPCR-Filter uses ESM-3 GPCR sequence embeddings, ligand graph neural networks, and attention-based fusion to predict modulatory potential.
Results
Across multiple evaluation scenarios, GPCR-Filter surpasses state-of-the-art DTI models in accuracy and generalization and identifies micromolar-level 5-HT1A agonists.
Takeaways & Limitations
Combining curated GPCR-specific data with protein-language and ligand-graph fusion yields a scalable, interpretable virtual-screening model with strong out-of-distribution performance.
Takeaways & Limitations
Negative labels from 1:1 sampling may include unobserved actives, and inter-target splits allow ligand reuse across targets, potentially easing evaluation for some chemotypes.

Abstract

from arXiv · show

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.

1 Introduction

GPCR modulator discovery is difficult because dynamic allostery can separate ligand binding from functional signaling, while conventional assays are costly and low-throughput. GPCR-Filter addresses this challenge with sequence- and structure-informed deep learning designed to predict modulatory potential.

GPCR significance: GPCRs regulate major nervous, endocrine, and immune pathways and represent an important class of modern drug targets.The passage states that more than 800 GPCRs orchestrate diverse physiological and pathological processes and that up to 36% of FDA-approved drugs act on GPCRs.
Discovery challenge: Dynamic conformational landscapes and allosteric mechanisms can decouple GPCR ligand binding from downstream signaling efficacy.This makes binding affinity an incomplete indicator of functional modulation.
Discovery challenge: Traditional computational screening and experimental assays struggle with GPCR modulation because they may miss functional relationships and are costly, low-throughput, or labor-intensive.These limitations contribute to uncertain or inactive hits and restrict the discovery rate of novel modulators.
AI approaches: Structure-based AI screening is constrained by fewer than 1,800 experimentally determined GPCR-ligand complex structures and inaccuracies in predicted structures.The available experimental structures are also described as largely conserved, limiting effective model fine-tuning.
Proposed framework: GPCR-Filter combines ESM-3 GPCR sequence embeddings, ligand graph neural networks, and attention-based fusion to learn receptor-ligand functional correspondence.Its design aims to predict modulatory potential beyond simple binding.
Reported results: Across multiple evaluation scenarios, GPCR-Filter surpasses state-of-the-art DTI models in accuracy and generalization and identifies micromolar-level 5-HT1A agonists with distinct scaffolds.The reported experimental validation supports its potential for functional ligand discovery.

2 Results

GPCR-Filter combines sequence- and structure-based ligand representations with attention to predict GPCR–ligand interactions and refine virtual-screening candidates. It outperformed baseline models across increasingly challenging splits, produced pocket-relevant attention patterns, and identified four validated 5-HT1A agonists.

Model and workflow: GPCR-Filter processes GPCR sequences and ligand SMILES with pretrained sequence embeddings, graph neural networks, and cross-attention to generate interaction probabilities.In the discovery workflow, it follows initial structure-based virtual screening and prioritizes candidates for downstream activation assays.
Dataset: 91,396 validated human GPCR–drug interactions covering 527 GPCRs and 72,177 ligands formed the training and evaluation dataset.The dataset integrates GPCRdb and GtoPdb records and standardizes ligands as canonical SMILES.
Predictive performance: GPCR-Filter consistently outperformed ConPLex and TransformerCPI2.0 across random, intra-target, and inter-target evaluation regimes.These splits assess in-distribution performance, unseen ligands for known receptors, and transfer to entirely unseen receptors.
Predictive performance: 73.44% AUC and 64.04% AP were achieved by GPCR-Filter in the inter-target setting, while ConPLex had an AUC below 50%.Inter-target evaluation withheld receptor identities from training, making it the most stringent generalization task.
Interpretability: Top-20 attention rankings recovered 6, 8, and 7 crystallographic pocket residues for random, intra-target, and inter-target models in the DRD2 complex.A second GPCR complex showed similar enrichment, and the authors interpret these patterns as evidence of pocket-relevant interaction signals.
Experimental validation: Four GPCR-Filter-prioritized compounds—D24, D29, D34, and D47—activated 5-HT1A and showed maximal effects approaching or exceeding 5-HT with micromolar-level right-shifted EC50 values.The compounds were selected after screening 52 prioritized molecules at 30 µM and then characterized using concentration–response curves.

3 Discussion

GPCR-Filter combines curated GPCR-specific data with protein-language-model and ligand-graph representations to support interpretable screening under challenging generalization settings. Validation included stringent receptor-disjoint evaluation, wet-lab testing, and analyses of attention and dataset structure, while several sampling and evaluation limitations remain.

Evaluation and validation: GPCR-Filter maintained strong performance across random, intra-target, and inter-target splits, including a clear margin under the most stringent inter-target setting.The inter-target protocol evaluates receptors absent from training.
Evaluation and validation: Four agonists were validated among the top-ranked 5-HT1A predictions in wet-lab assays.
Interpretability and mechanism: Top-20 attention-ranked residues were enriched near crystallographic pockets across two newly solved complexes and all split types.
Interpretability and mechanism: Structured ligand-profile organization provides a plausible context for transferring learned chemical patterns to unseen receptors with similar ligand chemistry.
Limitations and future work: Negative sampling may include unobserved actives, and inter-target ligand reuse may provide an easier signal for some chemotypes.The authors propose alternative decoy strategies and stricter ligand or scaffold-level deduplication for future testing.
Limitations and future work: Future work will expand validation to additional protein families, refine negative sampling and calibration, and incorporate structural or pocket-level constraints.

4 Methods

GPCR-Filter predicts binary GPCR–ligand interactions from amino-acid sequences and ligand SMILES by encoding receptor residues and ligand atoms, then combining them through cross-attention. Its training objective is two-class cross-entropy, with threshold-dependent metrics evaluated at probability 0.5 unless otherwise specified.

Model inputs and outputs: GPCR-Filter takes a GPCR amino-acid sequence and ligand SMILES string as inputs and produces a binary interaction score.
Objective and evaluation: Training minimizes two-class cross-entropy over logits, while ACC, Precision, Recall, and F1 use a fixed probability threshold of 0.5.AUC and AUPR are threshold-free.
Model inputs and outputs: ESM-3-derived per-residue GPCR embeddings are projected to a shared hidden size and encoded by a Transformer.The implementation uses ht=1536, d=256, and a default two-layer encoder.
Ligand embedding: Ligand SMILES are converted into molecular graphs, whose atom features are projected and processed by a single graph convolutional layer.
Ligand embedding: A learnable graph-level token is prepended to the atom sequence, which is padded with a corresponding target-side mask for decoder fusion.

4.5 Decoder and attention-based fusion

GPCR-Filter fuses ligand and receptor representations through Transformer-based attention, using ligand features as queries and GPCR sequence features as memory to generate interaction predictions.

Cross-attention: The decoder uses ligand graph features as the target/query and encoded GPCR residue features as memory key/value in cross-attention.Multi-head scaled dot-product attention includes residual connections, LayerNorm, dropout, and padding masks.
Self-attention: Self-attention first refines ligand representations and independently refines the GPCR sequence representations.
Feature fusion: Attention weights dynamically aggregate information from relevant ligand and receptor regions to form fused interaction features.The ligand acts as query while the protein supplies key/value information during cross-attention.
Prediction and interpretation: The graph-level ligand token at index 0 is passed through an MLP to produce two-class logits and the scalar interaction score.Last-layer cross-attention from this token to residues supplies residue-level importance scores for interpretability.

4.6 Dataset curation

The study curated a large experimentally grounded dataset of human GPCR–drug interactions from GPCRdb and GtoPdb, then standardized receptor and ligand identities.

Data sources: 91,396 validated drug–GPCR interaction records were collected from GPCRdb and GtoPdb after deduplication.GPCR identifiers were aligned to UniProt, and drug identifiers were re-encoded for standardized analysis.
Dataset scope: The curated records represent human GPCRs and drugs for supervised learning of GPCR–ligand interactions.

4.7 Dataset distribution analysis

The curated interaction dataset is strongly long-tailed: a small subset of GPCRs accounts for many ligand associations, while most receptors have sparse coverage.

Long-tailed distribution: Most GPCRs interact with only a handful of ligands, whereas a relatively small subset accounts for a large fraction of interactions.The distribution was examined using a logarithmically binned frequency-of-frequency histogram.
Receptor imbalance: The Top-10 most frequent receptors dominate the interaction space, while the vast majority of GPCRs have far fewer associated ligands.These highly active receptors form the head of the observed interaction-count distribution.

4.8 Evaluation protocols and negative sampling

GPCR-Filter was evaluated under increasingly difficult random, intra-target, and inter-target splits, with independently sampled balanced negatives designed to prevent information leakage.

Evaluation scenarios: Three evaluation scenarios measure in-distribution performance, unseen ligand combinations for known receptors, and generalization to entirely unseen GPCRs.Task difficulty increases from the random split to the inter-target split.
Negative sampling: Negative examples were created by removing known positives from all drug–target combinations and randomly sampling an equal number of negatives for each split.Sampling was performed independently across splits to maintain a 1:1 class balance and avoid information leakage.
Random split: The random split partitions pooled positive and sampled negative pairs into training, validation, and test sets using an 80/10/10 ratio.This setting measures in-distribution performance under shared global ligand and target distributions.
Intra-target split: The intra-target split tests unseen ligand combinations for receptors represented during training, while receptors with fewer than ten positives remain exclusively in training.
Inter-target split: The inter-target split holds out disjoint GPCR identities at a 9:1 training-to-held-out ratio and draws validation and test examples only from held-out targets.Negative samples are generated separately within the training and held-out receptor pools.

4.9 Baselines and evaluation settings

GPCR-Filter was evaluated against ConPLex and TransformerCPI2.0 using matched inputs and increasingly demanding generalization splits, with interpretability assessed through ligand-to-protein attention and pocket overlap.

ConPLex and TransformerCPI2.0 served as competitive sequence-based DTI baselines under matched hyperparameters, optimization, and thresholding procedures.
The evaluation used random, intra-target, and inter-target splits to assess in-distribution performance, unseen ligand combinations, and transfer to unseen receptors.
Interpretability analysis extracted ligand-to-protein cross-attention, ranked residues, and compared Top-20 predictions with binding-pocket residues.Pocket hits@20 counts pocket residues within the Top-20 list, with enrichment reported against a random Top-20 baseline.

4.11 Virtual screening with GPCR-Filter

Virtual screening combined molecular docking with GPCR-Filter prioritization before experimental testing of compounds at the 5-HT1A receptor. Four candidates showing significant luminescence reductions advanced to dose-response analysis.

1,644,833 ChemDiv structures were prepared with LigPrep, docked using Schrödinger Glide, and ranked compounds were selected for GPCR-Filter input.The 5-HT1A-SEP-363856 complex with PDB ID 8W8B served as the docking starting structure.
The assay used full-length 5-HT1A and GloSensor-22F cAMP biosensor constructs expressed in 293T cells.Cells were cultured, transfected, washed, detached, starved, and prepared in media before measurement.
52 available candidates were tested at 30 µM, with luminescence benchmarked against 5-HT and Forskolin controls.
Four ligands—D24, D29, D34, and D47—showing significant luminescence reductions were selected for subsequent dose-response analysis.

Appendix

The appendix characterizes the dataset’s receptor-frequency distribution, ligand-similarity relationships among GPCRs, and its most frequent receptors.

Most GPCRs interact with few ligands, while a small receptor subset accounts for a disproportionately large number of interactions.Figure S1 uses ligand occurrences per GPCR on the x-axis and GPCR population per bin on the y-axis.
Hierarchical clustering groups GPCRs using ligand fingerprint similarity profiles aggregated across associated ligands.Heatmap cells represent pairwise ligand similarity, while dendrograms show hierarchical relationships among targets.
Table S1 lists the ten most frequent GPCRs in the dataset.