AlphaFold-Multimer Virtual Screening for GPCR Nanobody Discovery: From Proof of Concept to Prospective Application

Original Authors: Harvey EP, et al.
Published in: Nature Communications, 2026

Abstract

Antibody discovery traditionally relies on immunization or in vitro library screening, with high experimental costs and long timelines. A recent study by Harvey et al. published in Nature Communications demonstrates that AlphaFold-Multimer (AF-M) can be used for virtual screening of nanobodies against GPCR targets, prospectively identifying MRGPRX2-binding nanobodies with affinities of 20-200 nM from 10,000 simulated sequences, validated in cellular functional assays.

1. Background: Needs and Current Status of Computational Antibody Discovery

Antibody drugs represent the fastest-growing class of therapeutic proteins, with over 100 monoclonal antibodies approved and more than 800 in clinical development. Nanobodies (single-domain heavy chain antibody fragments) are emerging as a new therapeutic antibody class due to their small molecular weight, simple structure, and favorable biochemical properties.

Traditional antibody discovery relies on immunization techniques developed in the 1970s, but faces significant limitations: many potential drug targets are highly conserved in mammals, leading to immunization failure; yeast surface display and phage display can bypass immunization restrictions but require specialized equipment and often produce polyreactive binders.

In recent years, AlphaFold2 and its successors have achieved near-experimental accuracy in protein structure prediction by combining structural data from the Protein Data Bank (PDB) with co-evolutionary information from multiple sequence alignments (MSA). However, antibody-antigen pairs lack meaningful co-evolutionary relationships, requiring models to rely entirely on structural patterns from PDB.

2. Assessment of AF-M's Ability to Predict GPCR-Nanobody Binding

2.1 Prediction Accuracy for Known Complex Structures

The research team first evaluated AF-M's ability to predict known GPCR-nanobody complexes. Using the angiotensin II type 1 receptor (AT1R) in complex with synthetic nanobody AT118-H as an example, AF-M predictions showed high agreement with experimental structures:

• Overall RMSD: 2.9 Å
• Receptor RMSD: 2.9 Å
• Nanobody RMSD: 1.8 Å
• CDR loop RMSD: 1.52-2.65 Å

Notably, AT118-H induces AT1R to adopt a previously unseen conformation—active-like on the extracellular side and inactive-like on the intracellular side. AF-M correctly predicted this atypical state.

2.2 Benchmark Testing for Distinguishing Binders from Non-binders

To systematically evaluate AF-M's discrimination ability, the research team constructed three benchmark datasets:

Key Finding: In the GPCR-nanobody dataset, seven metrics achieved AUROC > 0.65, with mean pTM (AUROC = 0.73) and best model pTM (AUROC = 0.71) performing best. However, for non-GPCR membrane proteins and soluble proteins, these metrics showed AUROC near 0.5, indicating AF-M currently cannot effectively distinguish true binders from negative controls in these categories.

2.3 Precision for High-ranking Samples

In virtual screening scenarios, model precision for high-ranking samples is more critical than overall accuracy. The team calculated AUC5% (evaluating precision for only the top 5% ranked samples):

• GPCR dataset: AUC5% reached 0.93-1.00
• Soluble protein and membrane protein datasets: AUC5% ≤ 0.22

2.4 Comparison with AlphaFold3 and ESMFold

• AlphaFold3: GPCR dataset best model ipTM AUROC was 0.74, soluble proteins 0.66, but could not distinguish binding from non-binding pairs in non-GPCR membrane protein datasets
• ESMFold: Could not distinguish positive binding pairs from negative controls on GPCR datasets

Current Conclusion: AF-M outperforms ESMFold and performs comparably to AlphaFold3 (on GPCR targets) for predicting GPCR-nanobody binding.

3. Prospective Virtual Screening for MRGPRX2 Nanobodies

3.1 Screening Strategy Design

The team conducted prospective screening using MRGPRX2 (MAS-related G protein-coupled receptor X2) as the target. MRGPRX2 regulates IgE-independent mast cell degranulation and is a potential target for treating pseudo-allergic reactions and pruritus. Reasons for selecting this receptor include:

• Its shallow binding pocket implies a more permissive thermodynamic landscape
• Exists only in primates, with low sequence similarity to MRGPRX1/3/4
• No FDA-approved drugs currently target this receptor
• MRGPRX2 structure was deposited in PDB after AF-M training cutoff date

Based on previously published yeast display nanobody library design parameters, the team generated 10,000 simulated nanobody sequences with CDR3 lengths of 7, 11, and 15 residues, matching published library amino acid distributions.

3.2 Screening Pipeline and Filtering

Using AF-M (template-free mode) to predict 10,000 MRGPRX2-nanobody complex structures, ranked by LCF. Excluding sequences with potential defects (glycosylation sites in CDR regions, predicted polyreactivity), approximately 25% of sequences were filtered. Ultimately, 179 nanobodies (1.79%) had LCF exceeding the highest LCF of negative controls in the GPCR benchmark dataset.

The team selected the top 6 defect-free nanobodies for expression and purification, plus 4 lower-ranked nanobodies for broad sampling. Ten candidate nanobodies entered experimental validation.

3.3 Experimental Validation Results

Cell Binding Assays: In ROSA mast cell lines expressing MRGPRX2, nanobodies ranked 1 (Sim8619), 5 (Sim9877), 7 (Sim4784), and 90 (Sim4177) showed high-level binding, while negative control Nb60 showed no significant binding.

In HEK293T cells transfected with MRGPRX2, three nanobodies showed similar binding characteristics and dissociation constants to ROSA cells, confirming specific MRGPRX2 binding:

Functional Activity Validation:

• Agonist activity: Sim8619, Sim9877, and Sim4784 did not induce mast cell degranulation, indicating they are not functional agonists
• Antagonist activity: All three nanobodies attenuated compound 48/80-induced degranulation, suggesting they are functional antagonists
• G protein signaling assays: Sim8619 did not directly activate Gi protein, but pretreatment reduced 48/80 G protein activity Emax and right-shifted EC50

4. Structural Validation of Predicted Binding Modes

AF-M predicted that Sim8619, Sim9877, and Sim4784 all bind to the orthosteric pocket of MRGPRX2, overlapping with the compound 48/80 binding site. Arginine residues in CDR3 of Sim8619 and Sim4784 mimic 48/80's positively charged side chains interacting with receptor E164 and D184.

The team validated predicted binding modes through mutagenesis experiments:

• Nanobody mutations: Sim8619 R102A mutation (predicted to disrupt E164/D184 interactions) reduced binding affinity approximately 2-fold with decreased maximum binding
• Receptor mutations: MRGPRX2 E164A/D184A double mutation significantly reduced Sim8619 and Sim4784 binding but had no significant effect on Sim9877—consistent with AF-M predictions
• Additional CDR mutations: Multiple CDR mutations predicted to disrupt MRGPRX2 interactions all reduced affinity

These results support that AF-M can accurately predict nanobody binding poses.

5. Discussion and Limitations

5.1 Main Contributions

This study provides proof of concept for AF-M in GPCR nanobody virtual screening. From 10,000 simulated sequences, the team identified three nanomolar-affinity, functional antagonist MRGPRX2 nanobodies through computational screening and experimental validation. This workflow completely bypassed traditional immunization or in vitro display screening steps.

5.2 Key Limitations

• Target specificity limitation: AF-M currently shows reliable discrimination only for GPCR targets. For soluble proteins and non-GPCR membrane proteins, AF-M cannot effectively distinguish binders from non-binders
• Library design dependence: The simulated library used was based on published yeast display library parameters. More challenging GPCR targets may require larger computational libraries
• MRGPRX2特殊性: MRGPRX2 was selected partly due to its "promiscuous binder" characteristics—shallow binding pocket accommodating diverse charged ligands
• Lack of experimental structure validation: Due to MRGPRX2 purification difficulties, the team could not obtain experimental structures of nanobody-MRGPRX2 complexes
• Computational resource requirements: 10,000 AF-M predictions required substantial computational resources (study used NVIDIA A100 GPUs)

5.3 Competitive Landscape

This study is not the only advance in computational antibody discovery. Other recent methods have also reported success:

• Chai-2: Multimodal generative model successfully identifying nanobodies for multiple GPCRs, including functional agonists
• JAM: Generates VHHs against multi-transmembrane proteins including GPCRs
• RFdiffusion, BoltzGen, Germinal, mBER: Reported nanobody discovery for various target types

5.4 Future Directions

• Expanding computational libraries: Combining Bayesian optimization/active learning approaches to learn amino acid binding contributions from subset screening
• Training data growth: As antibody-protein structures in PDB continue to increase, AF-M and similar models' predictive capabilities are expected to further improve
• GPCR antibody therapeutics: Approximately 35% of FDA-approved drugs target GPCRs, but only 3 GPCR-targeting monoclonal antibodies are approved. Computational discovery methods may lower the technical barrier for GPCR antibody discovery

6. Conclusion

Harvey et al.'s study demonstrates that AlphaFold-Multimer can serve as an effective tool for GPCR nanobody virtual screening, prospectively identifying functional binders from large-scale computational libraries. This method is currently limited to GPCR targets and constrained by computational resources and target characteristics. Nevertheless, this proof of concept provides an important foundation for computational antibody discovery workflows, and its application scope is expected to expand with model improvements and training data growth.

References

Harvey, E.P. et al. In silico discovery of nanobody binders to a G-protein coupled receptor using AlphaFold-Multimer. Nat Commun (2026). https://doi.org/10.1038/s41467-026-72093-5

This article is based on Harvey et al.'s 2026 study published in Nature Communications, aiming to provide objective technical interpretation for professional practitioners.