GPCR-Filter: Deep Learning for Modulator Discovery

• GPCR-Filter is a deep learning framework that combines ESM-3 protein embeddings, ligand graph neural networks, and cross-attention to predict functional GPCR modulators.
• It re-ranks candidate compounds using only GPCR sequence and ligand SMILES, addressing the limitations of binding-centric virtual screening methods.
• The framework shows high performance across random, intra-target, and inter-target splits and is validated by experimental success in a 5-HT1A screening campaign.

GPCR-Filter is a deep learning framework for GPCR modulator discovery that functions as a computational filter layered on top of standard virtual screening. It is designed to re-rank candidate compounds using only GPCR sequence and ligand SMILES, with the stated objective of enriching for compounds that are functional modulators rather than merely binders. The framework was introduced on a curated corpus of 91,396 validated human GPCR–drug interactions spanning 527 distinct GPCRs and 72,177 distinct drug SMILES strings, and combines ESM-3 protein representations, ligand graph neural networks, and attention-based receptor–ligand fusion (Ning et al., 27 Jan 2026).

1. Scientific setting

GPCR-Filter is situated in a pharmacological regime where generic compound–protein interaction prediction is often inadequate. The framework paper explicitly identifies three difficulties: GPCRs are highly dynamic and allosteric, functional assays are slow and expensive, and standard compound–protein interaction models are not tailored to GPCR pharmacology (Ning et al., 27 Jan 2026). In this formulation, the key problem is not only whether a ligand binds, but whether the receptor–ligand combination is likely to produce modulation.

This problem statement is consistent with a broader GPCR literature that emphasizes receptor dynamics, state heterogeneity, and nontrivial signal propagation. Ligand exposure of NTSR1 in living HEK293T cells shifts receptor organization from a relatively homogeneous membrane pattern toward clustered, patchy, heterogeneous distributions, with decreased anisotropy consistent with homoFRET and oligomerization (Westphal et al., 2019). Network analysis of class A receptors identifies distributed allosteric pathways and rotameric microswitches rather than a purely local binding mechanism (Lee et al., 2013). A recent review on GPCR mechanosensitivity further frames these receptors as integrators of chemical and mechanical inputs, with conformational ensembles shaped by shear stress, stretch, osmotic swelling, and ligand binding (Shetty et al., 2 Mar 2025). Taken together, these results support the rationale for a GPCR-specific screening model that is optimized for functionally meaningful receptor–ligand relationships rather than generic affinity proxies.

The framework therefore addresses a specific mismatch in computational pharmacology: receptor activation can arise from complex allosteric effects rather than direct binding affinity, yet much of the older virtual-screening stack is optimized for binding-centric ranking. GPCR-Filter is presented as an attempt to close that gap at the sequence-and-ligand level (Ning et al., 27 Jan 2026).

2. Dataset construction and task definition

The training corpus is a curated human GPCR pharmacology dataset assembled from GPCRdb and GtoPdb, with human targets aligned to UniProt sequences (Ning et al., 27 Jan 2026).

Item	Value
Interactions	91,396 validated human GPCR–drug interactions
GPCRs	527 distinct GPCRs
Ligands	72,177 distinct drug SMILES strings
Sources	GPCRdb and GtoPdb

The dataset is described as long-tailed: a few GPCRs have many ligands, while most receptors have only a few. This long-tailed structure is important because it makes naive memorization less useful and makes receptor-level generalization a primary concern.

Because the curated resources contain experimentally verified positive interactions rather than explicit negatives, GPCR-Filter generates negative examples computationally. The procedure enumerates all receptor–ligand combinations, removes known positives, and samples an equal number of negatives to create a 1:1 balanced dataset (Ning et al., 27 Jan 2026). The supervised task is then defined on ligand–target pairs $(D,T)$, with 2-class logits $\mathbf{o}(D,T)\in\mathbb{R}^2$ and positive-class probability $p=\mathrm{softmax}(\mathbf{o})_{[1]}$.

Three evaluation regimes are used to separate in-distribution performance from harder transfer settings:

Evaluation regime	Split sizes
Random split	146,233 / 18,279 / 18,280
Intra-target split	145,613 / 18,110 / 18,225
Inter-target split	167,521 / 15,271 / 15,271

The random split is standard in-distribution evaluation. The intra-target split retains the same receptors across training and test but holds out ligands, so the test set contains unseen ligands for known receptors. The inter-target split holds out entire receptors and therefore tests cross-target generalization to unseen GPCRs (Ning et al., 27 Jan 2026).

3. Architecture and representation strategy

GPCR-Filter is organized around three components: an ESM-3 encoder for receptor sequences, a graph convolutional network for ligands, and an attention-based fusion module that models receptor–ligand functional relationships (Ning et al., 27 Jan 2026).

For the receptor branch, the framework starts from precomputed per-residue ESM-3 representations with initial embedding dimension $h_t = 1536$, then linearly projects them to a shared hidden size $d = 256$. This design allows the model to inherit large-scale protein sequence pretraining rather than learning receptor features from the pharmacology dataset alone.

For the ligand branch, SMILES strings are converted into molecular graphs using RDKit with atom and bond descriptors, then encoded by a graph convolutional network. The model prepends a learnable graph-level token to the ligand representation, and that token is later used for prediction (Ning et al., 27 Jan 2026).

The central fusion mechanism is a Transformer-style decoder with ligand-to-protein cross-attention. In the formulation given by the paper, the GPCR sequence acts as memory, providing keys and values, while ligand tokens act as target and query. After ligand self-attention, protein self-attention, and ligand-to-protein cross-attention, the model extracts the graph-level token at index $0$, $x''_d[0]$, and passes it through a multi-layer perceptron to obtain the final prediction.

This architecture is noteworthy because it does not require an experimentally determined GPCR structure at inference time. The framework paper explicitly motivates this choice by noting that fewer than 1,800 experimentally determined GPCR–ligand complex structures are available and that many receptors still lack reliable structural coverage (Ning et al., 27 Jan 2026). A plausible implication is that sequence-conditioned, chemistry-aware models are being used here to trade structural completeness for broader receptor coverage.

Training minimizes binary cross-entropy implemented as 2-class cross-entropy over logits (Ning et al., 27 Jan 2026). In effect, GPCR-Filter is a binary interaction classifier, but the paper consistently interprets it as a modulator-discovery filter because its intended use is to rank candidates after an initial screen.

4. Generalization regimes and benchmark performance

GPCR-Filter is benchmarked against two sequence-based compound–protein interaction baselines, ConPLex and TransformerCPI2.0, under matched sequence-plus-SMILES inputs (Ning et al., 27 Jan 2026). The reported AUC values are as follows:

Split	Baseline AUCs	GPCR-Filter AUC
Random	53.03, 65.94	99.02
Intra-target	52.89, 62.02	97.98
Inter-target	44.91, 66.31	80.35

These results support two distinct claims. First, GPCR-Filter is very strong in standard random and intra-target evaluation. Second, it remains materially effective in the harder inter-target regime, where entirely unseen receptors are held out. The framework paper also reports inter-target AP 70.99 and states that GPCR-Filter consistently leads in AP, accuracy, and precision across settings (Ning et al., 27 Jan 2026).

The significance of the inter-target split is methodological rather than merely numerical. GPCR pharmacology datasets are often dominated by a subset of heavily studied receptors, so strong random-split performance can be inflated by ligand or receptor recurrence. By contrast, the inter-target setting probes whether the model has learned receptor-level sequence determinants rather than only pairwise co-occurrence patterns. The paper interprets the retained performance under this split as evidence that the model does not merely memorize interactions (Ning et al., 27 Jan 2026).

The framework also reports additional stress tests through a Bemis–Murcko scaffold split and a GPCR-family hold-out protocol. The data block does not provide the corresponding numerical outcomes, but their inclusion indicates that scaffold novelty and receptor-family transfer were treated as explicit robustness questions.

5. Attention maps and experimental validation

A central interpretability analysis asks whether cross-attention is localized near actual binding pockets. Pocket residues are defined as residues within 5 Å of the ligand, and the model’s top-attended residues are then compared against those sites for DRD2 (PDB 9bsb) and a purinergic receptor complex (PDB 9jcl) (Ning et al., 27 Jan 2026). For 9bsb, the paper states that there are nine pocket residues within 5 Å, and the top-20 attention residues recover 6, 8, and 7 pocket residues across the random, intra-target, and inter-target models. This is presented as evidence that the attention mechanism is not arbitrary but is capturing receptor–ligand interaction patterns localized near the binding site.

The framework’s practical validation comes from a 5-HT$_{1A}$ screening campaign. In that workflow, 1,644,833 ChemDiv compounds were docked into the SEP-363856 binding pocket of 5-HT$_{1A}$ (PDB 8W8B), the top 8,705 docked compounds were passed to GPCR-Filter, 97 compounds with predicted probability $p>0.5$ were selected, and 52 were available and tested experimentally (Ning et al., 27 Jan 2026). The assay used was a GloSensor-cAMP assay in 293T cells, where lower luminescence indicates stronger receptor activation.

At 30 μM, four compounds—D24, D29, D34, and D47—were identified as promising because they produced stronger luminescence reductions than the 5-HT control. Dose–response experiments then showed sigmoidal response curves, maximal effects comparable to or exceeding 5-HT, micromolar EC$\mathbf{o}(D,T)\in\mathbb{R}^2$0 values, and distinct chemical scaffolds from 5-HT (Ning et al., 27 Jan 2026). This case study is important because it moves GPCR-Filter beyond retrospective benchmarking: the model is used as a second-stage prioritization layer after docking, and the selected compounds are experimentally validated as functional agonists.

6. Prehistory, scope, and limitations

GPCR-Filter belongs to a longer lineage of GPCR-focused computational filtering, but it is distinct in both target and granularity. Earlier work on sequence classification used an RBF-kernel SVM in Weka to distinguish human GPCR sequences from those of other species using N-terminal and extracellular-loop information, loop lengths, amino acid composition, and physicochemical descriptors; in a Class A rhodopsin case study, that system reported 94.44% accuracy, 100.00 sensitivity, 90.90 specificity, and $\mathbf{o}(D,T)\in\mathbb{R}^2$1 (Shrivastava et al., 2010). Reviews of automated GPCR ligand discovery later described broader screening stacks that combine docking, similarity, QSAR, random forests, GNN-derived fingerprints, active learning, and experimental feedback (Raschka, 2018, Raschka et al., 2020). This suggests an evolution from GPCR identification filters and general ligand-prioritization workflows toward a receptor-class-specific modulator-ranking model.

Within that trajectory, GPCR-Filter is notable for being explicitly GPCR-specific rather than a generic drug–target interaction model. The framework paper argues that AI-assisted drug discovery for GPCRs benefits from models trained on GPCR-specific biology rather than generic DTI data (Ning et al., 27 Jan 2026). That claim is consistent with earlier reviews emphasizing that pure similarity search is weak, docking alone is imperfect, and hybrid pipelines are strongest for GPCR ligand discovery (Raschka, 2018, Raschka et al., 2020).

The framework also states several limitations. Negative labels are sampled rather than experimentally confirmed, so some negatives may be unobserved actives. Ligands may recur across targets in the inter-target split, which could make the task easier for certain chemotypes. Stricter ligand or scaffold de-duplication is identified as a more demanding future test, and broader comparisons with deeper calibration or more structure-aware models are left for future work (Ning et al., 27 Jan 2026). These limitations matter because they qualify the meaning of the benchmark scores: the reported values demonstrate a strong and scalable GPCR-specific prioritization method, but not a complete solution to GPCR pharmacology.

In that sense, GPCR-Filter is best understood as a functional enrichment layer for screening campaigns. It does not replace docking, assays, or mechanistic studies of receptor activation. Rather, it operationalizes a particular hypothesis: sequence-conditioned receptor information and graph-based ligand information, fused through cross-attention and trained on validated GPCR pharmacology, can improve the ranking of candidate modulators at scale (Ning et al., 27 Jan 2026).