Virtual Screening of the Orphan GPCR, GPR151, Using Boltz-2 Structure Prediction and DOCK6 Molecular Docking

1. Abstract

GPR151 is an orphan G protein-coupled receptor implicated in pain signaling and nicotine dependance, yet it remains poorly characterized pharmacologically, lacks an experimentally determined structure, its endogenous ligand(s) aren’t known, and has only a limited number of reported active compounds. We propose a hybrid in silico workflow integrating AI-based structural modeling with large-scale virtual screening to prospectively identify novel candidate ligands. A ligand-bound structural model of GPR151 will be generated using Boltz-2 and used as a template to screen compounds from the ZINC-22 database using the DOCK6 Hierarchical DataBase (HDB) protocol. Docked compounds will be ranked using a composite scoring strategy integrating the DOCK6 grid score, PLIP-derived interaction fingerprints, and scaffold diversity. Top-ranked compounds will be rescored using the Boltz-2 affinity module to provide an orthogonal machine learning-informed validation of the classical docking results. Final candidates will be reported with predicted binding poses, interaction fingerprints, and predicted ADMET profiles to support experimental follow-up. This workflow is designed to be generalizable, offering a reproducible framework for early-stage ligand discovery at structurally unresolved orphan protein targets, with GPR151 serving as a biologically relevant orphan GPCR case study.

2. Introduction

G protein-coupled receptors (GPCRs) are a major class of therapeutic targets yet remain pharmacologically underexplored. Approximately 35% of all approved therapeutic drugs act through GPCRs, yet only ~15% of the ~800 human GPCRs have been successfully targeted (Wingler and Lefkowitz, 2020; Yang et al., 2024). Orphan GPCRs, receptors whose endogenous ligands remain unidentified, comprise nearly one-third of this family and represent a significant opportunity for drug discovery (Michino et al., 2025).

GPR151 is an orphan class A GPCR with restricted expression in the habenula, spinal cord neurons, and dorsal root ganglia that is implicated in pain signaling and nicotine dependence (DePasquale et al., 2025). Emerging evidence also links GPR151 to hepatic glucose metabolism, with loss-of-function variants associated with reduced risk of type 2 diabetes (Bielczyk-Maczynska et al., 2022). Although GPR151 lacks an experimentally determined structure, recent work by Zhang et al. (2025) demonstrated that structure-based ligand discovery for GPR151 is feasible. Using the AuroBind platform, they screened a library of approximately 30 million compounds and experimentally validated the first reported small-molecule GPR151 agonists, C15 and C40, which exhibited BRET2 based Go recruitment EC₅₀ values of 1.1 µM and 4.2 µM, respectively. These findings demonstrate the feasibility of computational ligand discovery for GPR151 despite the absence of experimental structural information. Going beyond these hits, given the complexity of GPCR signaling and moderate potency of C15 and C40, identifying more diverse binders will be useful as chemical probes.

Recent advances in AI-based structure prediction enable modeling of receptor-ligand complexes in the absence of experimental structures. Boltz-2 supports receptor-ligand co-folding and provides an affinity prediction module within a unified framework, enabling both structural and energetic evaluation of candidate ligands (Passaro et al., 2025). In combination, DOCK6 enables large-scale virtual screening through its Hierarchical Database (HDB) search protocol, allowing efficient docking of millions to billions of compounds from ZINC-22 (Balius et al. 2024; Tingle et al. 2022). In this workflow, AI-based and physics-based approaches provide complementary strengths: Boltz-2 leverages learned structural priors to model receptor–ligand complexes, whereas DOCK6 applies physics-based scoring to evaluate molecular interactions within the binding pocket. Structure-based virtual screening has successfully identified novel GPCR ligands from ultra-large chemical libraries and is increasingly being applied to targets lacking experimental structures, supporting the use of predicted receptor models for prospective ligand discovery (Lyu et al., 2019).

This project proposes a hybrid virtual screening workflow integrating DOCK6 and Boltz-2 to identify novel candidate ligands for GPR151. Boltz-2 will be used to generate ligand-bound receptor models and to rescore top-ranked compounds, while DOCK6 will enable large-scale screening of chemical space. By combining physics-based docking with machine learning-based affinity prediction, this approach will define a reproducible strategy for ligand discovery in an orphan GPCR context that may also be applicable to other structurally unresolved orphan targets. Success will be evaluated by whether the workflow produces a stable binding-site definition, enriches known actives above decoys, avoids prioritization dominated by obvious scoring artifacts, and yields chemically diverse, commercially available compounds for experimental testing.

Hypotheses

Primary Outcome: The primary outcome of this study will be a ranked list of 200 commercially available candidate ligands for GPR151. Each compound will be annotated with predicted binding poses, composite docking scores, Boltz-2 affinity estimates, PLIP-derived interaction fingerprints, and ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) profiles. This candidate set will represent a chemically diverse collection of ligands spanning multiple interaction patterns within the predicted binding pocket and will be directly suitable for experimental validation using standard GPCR functional assays.

Secondary Outcome: In addition to the candidate list, this study will produce a fully specified and reproducible virtual screening workflow, including receptor model selection criteria, docking parameters, decoy and enrichment evaluation, interaction profiling methods, quality-control flags, and prioritization strategy, enabling application to other orphan GPCR systems and related structurally unresolved orphan targets. Failure would be indicated by an unstable binding-site definition, inability to enrich known actives above decoys, prioritization driven primarily by physicochemical artifacts rather than plausible binding interactions, or failure to identify a chemically diverse and commercially available candidate set.

3. Methods

Participants/Sample

Proposed virtual discovery pipeline sample sizes:

AIM 1: Receptor site preparation
1. Reference ligands: 2 experimentally validated hits from Zhang et al. (2025)
2. Reference decoys: 100 property matched using DUDE-Z
3. Docking site: Boltz-2 docking pose, Dock6 site prearation assessed with decoy-discrimination and molecular dynamics.
AIM 2: HDB Docking
1. Dock6 virtual screening of atleast 1 billion compounds from ZINC-22/Enamine
2. Filter for top 1% by score and chemical and interaction diversity => 200,000 hits
AIM 3: ML Rescoring and Hit-Selection
1. Counter screen for synthetic accessiblity and ADMET risk
2. Screen for Boltz-2 predicted affinity => 200 selected compounds

The virtual discovery pipeline is designed to efficiently and at low cost optimize the following target product profile:

DOCK6 Grid Score No-Go Threshold: > 1%. Target Threshold: < 0.1%
Boltz-2 BindingPred: No-Go Threshold: < 50% Target Thrsehold: > 75%
Chemical Diversity: No-Go Threshold: Tc > 0.6 Target Threshold: Tc < 0.4
ADMET Profile: No-Go Threshold: > 0 alerts Target Threshold: 0 alerts
Commercial Availability: No-Go Threshold: Not Available Target Threshold: Available

Materials and Procedures

Software Used:

Hardware Used (UMich HPC Cluster):

GPU cluster:
- 1 node x 4 H200 NVIDIA GPU
- 20 nodes x 2 V100 NVIDIA GPUs
- 28 nodes x 8 A40 NVIDIA GPUs
CPU cluster:
- 455 nodes x 36 CPUs per node

Planned Analyses

Aim 1. Refine ligand-bound GPR151 models and define the extracellular binding site

In the preliminary DOCK6 screen, a ligand-bound GPR151 model was selected based on the highest quality predicted Boltz-2 model and used for all docking calculations. In this aim, the selected receptor conformation will be further evaluated, and the extracellular binding pocket will be defined in a robust manner to support downstream screening and analysis. The robustness of the predicted binding pocket will be assessed using the existing ensemble of Boltz-2 predictions, including models generated with and without molecular dynamics (MD) refinement, as well as apo receptor models and affinity-module confidence metrics.

To further characterize conformational stability and binding-pocket heterogeneity, molecular dynamics simulations will be performed on both apo and C40-bound GPR151 models. These simulations will be used to evaluate the stability of predicted ligand interactions, characterize residue flexibility within the extracellular binding pocket, and identify interaction trends that may not be fully captured in static structure predictions. Analysis of MD trajectories will additionally be used to refine docking-site boundaries and define conformationally stable regions for downstream virtual screening.

Docking-site boundaries will be defined based on spatial clustering of ligand poses from preliminary docking, structural features conserved across the Boltz-2 ensemble, persistent interaction regions identified through MD simulations, and residues comprising the predicted orthosteric binding site. The Asp115-Trp270-Gln292 interaction network described by Zhang et al. (2025), together with PLIP-derived interaction fingerprints from preliminary models and docking results, will be used as initial reference points for evaluating ligand engagement. However, interaction patterns will not be constrained to a predefined network. Instead, PLIP-derived interaction fingerprints across prioritized docked ligands will be used to characterize recurrent ligand-binding modes and interaction diversity within the extracellular pocket. The final binding-site definition will therefore be based on reproducibility across modeling conditions, MD-supported pocket stability, and the ability to support diverse ligand interaction patterns, establishing a validated receptor framework for large-scale virtual screening and downstream prioritization.

Aim 2. Expand virtual screening of ZINC-22 chemical space

Building on the preliminary DOCK6 screening of ZINC-22 heavy-atom tranches H11–H13, large-scale virtual screening will be extended to larger and more chemically diverse regions of chemical space. Initially, screening efforts will focus on approximately 10 million compounds from ZINC-22’s H14–H16 tranches before reassessing docking score distributions, scaffold diversity, ligand efficiency, molecular weight, LogP, and hit quality. This first expanded screen will test whether larger heavy-atom tranches provide additional high-quality chemotypes beyond those observed in the preliminary H11–H13 screen. Practical quality-control considerations from large-scale docking workflows, including the use of decoy sets to evaluate scoring behavior, will be incorporated into this staged evaluation (Bender et al., 2021). Recent studies have demonstrated that screening larger virtual libraries can improve hit rates, scaffold diversity, and ligand potency by accessing rare chemotypes and underexplored regions of chemical space (Liu et al., 2024). Therefore, the long-term goal of this aim is to systematically expand virtual screening coverage across ZINC-22, evaluating up to one billion compounds. Library expansion will proceed iteratively across tranches spanning different heavy atom counts and physiochemical properties. As library size increases, docking distribution, scaffold diversity, molecular weight, and LogP will be monitored to evaluate the benefits of exploring additional chemical spaces. To reduce the bias of affinity-based scoring toward larger and more lipophilic molecules, docking scores and Boltz-2 affinity predictions will be evaluated alongside ligand-based affinity metrics that account for molecular weight and LogP (Hopkins et al., 2014). These analyses will identify regions of chemical space enriched for high-quality candidate ligands and establish practical saturation criteria based on diminishing improvements in predicted affinity, ligand efficiency, docking performance, and chemical diversity.

The receptor model and binding-site definition established in Aim 1 will be used for all docking calculations. Docking will be performed using the DOCK6 Hierarchical DataBase (HDB) protocol, enabling efficient evaluation of large compound libraries. Ligands will be sampled across multiple conformations and orientations within the defined binding pocket, and poses will be scored using the DOCK6 Grid Score. All successfully docked compounds will be ranked by Grid Score, and the top ~1% of compounds will be retained for downstream analysis. Docking outputs, including pose geometries and energy components, will be stored for subsequent interaction profiling and rescoring. The resulting subset of compounds serves as the input for interaction-driven prioritization in Aim 3.

Aim 3. Prioritize candidate ligands using integrated scoring and machine learning rescoring

Top-ranked compounds from Aim 2 will be analyzed using PLIP to generate residue-level interaction fingerprints, including hydrogen bonds, salt bridges, and hydrophobic contacts. Rather than using PLIP-derived features as a direct scoring function, these interaction fingerprints will be used to characterize ligand binding modes and guide diversification of prioritized compounds. Ligands exhibiting distinct interaction patterns within the predicted extracellular binding pocket will be preferentially retained to maximize mechanistic diversity among candidate binders. PLIP-derived interaction fingerprints will be compared using pairwise similarity metrics such as Tanimoto similarity, and compounds will be selected using a greedy diversification strategy that balances interaction fingerprint diversity with ligand structural diversity, similar to the approach used in BoltzGen (Stark et al., 2025). To ensure chemical diversity, compounds will be clustered by molecular scaffold, and representative candidates from major clusters will be retained. Compounds will then be screened using established PAINS (Pan-Assay Interference Compounds) substructure filters to flag molecules containing chemical motifs associated with assay interference and nonspecific biological activity. Flagged compounds will be deprioritized during downstream candidate selection, reducing the likelihood of prioritizing false-positive screening hits (Baell and Holloway, 2010). Initial prioritization will primarily be based on the DOCK6 Grid Score, resulting in a set of ~500 chemically and mechanistically diverse candidate ligands.

To further refine candidate selection, top compounds will be rescored using the Boltz-2 affinity module. For each ligand, Boltz-2 will generate a predicted receptor-ligand complex and a confidence-weighted affinity estimate, enabling structure-aware evaluation beyond classical docking scores. Additional filtering will be performed based on predicted pharmacokinetic properties (using tools such as SwissADME) as well as commercial availability. As an additional quality-control step, adversarial decoy or background compounds with extreme charge states or physicochemical properties enriched among top-ranked virtual-screening hits may be used to assess whether prioritization is driven by plausible binding features or by systematic scoring bias. Final prioritization will integrate docking scores, PLIP-derived interaction features, scaffold diversity, and Boltz-2 affinity predictions to generate a ranked list of 200 commercially available candidate ligands suitable for experimental validation.

Ethics and Data Management

The work in this proposal is purely in silico. The analysis pipeline and docking results generated will be shared open-source through GitHub and HuggingFace, and along with any pee-reviewed publications.

4. Pilot Data (Optional)

We used Boltz-2 modeling to generate structural models of GPR151 in complex with the two reported agonists and evaluated model confidence using both structural and affinity prediction metrics. Models were generated with and without the molecular dynamics (MD) refinement option in Boltz-2 to assess the robustness of receptor-ligand predictions and improve local structural optimization. The selected C40-bound receptor model adopted a closed, inactive-like conformation based on the 2x46–6x37 C-alpha distance of 11.2Å, which falls below the Class A GPCR open-state threshold (Aranda-García et al., 2025). Despite this closed-state geometry, the model showed moderate predicted structural quality (confidence_score = 0.63), a predicted binding probability of 52% (affinity_probability_binary), and a predicted affinity value of 7 µM, supporting the plausibility of the modeled receptor-ligand complex. Across generated models, a polar interaction network involving residues Asp115, Trp270, and Gln292, as reported by Zhang et al. (2025), is consistently observed, supporting the structural plausibility of the predicted binding region. To further evaluate whether the model captured meaningful ligand recognition rather than nonspecific binding, preliminary decoy discrimination analysis was performed by comparing predictions across 2 active compounds and 100 decoys matched based on cLogP and molecular weight from ZINC (Stein et al., 2021). The model achieved an area under the receiver operating characteristic (AUROC) curve of 0.84, indicating good separation between active and decoy compounds, and one of the known active compounds (C40) was recovered within the top 5% enrichment fraction (EF5%), demonstrating early enrichment of likely binders.

To further characterize conformational variability, extended sampling of the C40-bound system was performed (n = 200 models), along with generation of an apo receptor model. These models reveal slight variability in ligand geometry within the extracellular pocket while preserving key structural features of the binding region. PLIP (Protein-Ligand Interaction Profiler) analysis was performed on the C40-bound models to characterize residue-level interactions, including hydrogen bonding, hydrophobic contacts, and salt bridge formation (Salentin et al., 2015). This analysis indicates that, while the previously reported Asp115-Trp270-Gln292 interaction network is observed across models, the C40-ligand also form contacts with residues outside this network, including Glu172. These observations suggest that the predicted binding pocket can accommodate alternative interaction patterns rather than enforcing a single predefined binding mode.

Initial virtual screening using DOCK6 was performed on ~1.9 million ZINC-22 compounds containing between 11-13 heavy atoms (H11–H13). Docking was performed against a single selected C40-bound model, rather than an apo structure or multi-state model. The top 1,500 docked ligands from the initial ZINC-22 H11–H13 screen, ranked by DOCK6 Grid Score, localized predominantly to the predicted extracellular binding pocket in the context of the full receptor. A focused overlay of representative top-ranked molecules demonstrated dense pose clustering within this region. This clustering behavior provides structural support for the predicted binding pocket and indicates that the receptor model can accommodate a broad range of chemotypes identified through the preliminary screen. Docking outputs included Grid Score with van der Waals and electrostatic energy components, which enable downstream ranking and prioritization of ligands. Together, these results establish the feasibility of large-scale virtual screening against GPR151 and define the structural and computational framework in this study.

5. Budget

Compute credits on the University of Michigan high-performance computing cluster: 2,000$

Salary support for Annika Backer, a research technician in the O’Meara lab: 3,000$

6. References

Aranda-García, David, et al. “Large Scale Investigation of GPCR Molecular Dynamics Data Uncovers Allosteric Sites and Lateral Gateways.” Nature Communications, vol. 16, no. 1, Feb. 2025, p. 2020. DOI.org (Crossref), https://doi.org/10.1038/s41467-025-57034-y.

Baell, Jonathan B., and Georgina A. Holloway. “New Substructure Filters for Removal of Pan Assay Interference Compounds (PAINS) from Screening Libraries and for Their Exclusion in Bioassays.” Journal of Medicinal Chemistry, vol. 53, no. 7, Apr. 2010, pp. 2719–40, https://doi.org/10.1021/jm901137j.

Balius, Trent E., et al. “DOCK 6: Incorporating Hierarchical Traversal through Precomputed Ligand Conformations to Enable Large‐scale Docking.” Journal of Computational Chemistry, vol. 45, no. 1, Jan. 2024, pp. 47–63. DOI.org (Crossref), https://doi.org/10.1002/jcc.27218.

Bender, Brian J., et al. “A Practical Guide to Large-Scale Docking.” Nature Protocols, vol. 16, no. 10, Oct. 2021, pp. 4799–832. DOI.org (Crossref), https://doi.org/10.1038/s41596-021-00597-z.

Bielczyk-Maczynska, Ewa, et al. “G Protein-Coupled Receptor 151 Regulates Glucose Metabolism and Hepatic Gluconeogenesis.” Nature Communications, vol. 13, no. 1, Dec. 2022, p. 7408. DOI.org (Crossref), https://doi.org/10.1038/s41467-022-35069-9.

DePasquale, Olivia, et al. “The Orphan Receptor GPR151: Discovery, Expression, and Emerging Biological Significance.” ACS Chemical Neuroscience, vol. 16, no. 9, May 2025, pp. 1639–46. DOI.org (Crossref), https://doi.org/10.1021/acschemneuro.4c00780.

Hopkins, Andrew L., et al. “The Role of Ligand Efficiency Metrics in Drug Discovery.” Nature Reviews Drug Discovery, vol. 13, no. 2, Feb. 2014, pp. 105–21. DOI.org (Crossref), https://doi.org/10.1038/nrd4163.

Liu, Fangyu, et al. “The Impact of Library Size and Scale of Testing on Virtual Screening.” Biophysics, 11 July 2024. DOI.org (Crossref), https://doi.org/10.1101/2024.07.08.602536.

Michino, Mayako, et al. “AI Meets Physics in Computational Structure-Based Drug Discovery for GPCRs.” Npj Drug Discovery, vol. 2, no. 1, July 2025, p. 16. DOI.org (Crossref), https://doi.org/10.1038/s44386-025-00019-0.

Passaro, Saro, et al. “Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction.” Molecular Biology, 18 June 2025. DOI.org (Crossref), https://doi.org/10.1101/2025.06.14.659707.

Salentin, Sebastian, et al. “PLIP: Fully Automated Protein–Ligand Interaction Profiler.” Nucleic Acids Research, vol. 43, no. W1, July 2015, pp. W443–47. DOI.org (Crossref), https://doi.org/10.1093/nar/gkv315.

Stark, Hannes, et al. “BoltzGen: Toward Universal Binder Design.” Bioengineering, 24 Nov. 2025. DOI.org (Crossref), https://doi.org/10.1101/2025.11.20.689494.

Stein, Reed M., et al. “Property-Unmatched Decoys in Docking Benchmarks.” Journal of Chemical Information and Modeling, vol. 61, no. 2, Feb. 2021, pp. 699–714. DOI.org (Crossref), https://doi.org/10.1021/acs.jcim.0c00598.

Tingle, Benjamin, et al. “ZINC-22 - A Free Multi-Billion-Scale Database of Tangible Compounds for Ligand Discovery.” 19 Oct. 2022. DOI.org (Crossref), https://doi.org/10.26434/chemrxiv-2022-82czl.

Wingler, Laura M., and Robert J. Lefkowitz. “Conformational Basis of G Protein-Coupled Receptor Signaling Versatility.” Trends in Cell Biology, vol. 30, no. 9, Sept. 2020, pp. 736–47. DOI.org (Crossref), https://doi.org/10.1016/j.tcb.2020.06.002.

Yang, Hanting, et al. “Genome-Wide Pan-GPCR Cell Libraries Accelerate Drug Discovery.” Acta Pharmaceutica Sinica B, vol. 14, no. 10, Oct. 2024, pp. 4296–311. DOI.org (Crossref), https://doi.org/10.1016/j.apsb.2024.06.023.

Zhang, Zhongyue, et al. “Fitness Aligned Structural Modeling Enables Scalable Virtual Screening with AuroBind.” Version 1, arXiv, 2025. DOI.org (Datacite), https://doi.org/10.48550/ARXIV.2508.02137.

Virtual Screening of the Orphan GPCR, GPR151, Using Boltz-2 Structure Prediction and DOCK6 Molecular Docking

Virtual Screening of the Orphan GPCR, GPR151, Using Boltz-2 Structure Prediction and DOCK6 Molecular Docking

Funding Progress

Funders

Nonprofit Recipient

Topics

DOI