2026 Harvey - AF-M nanobody binders to GPCR

AF-M plus interface confidence metrics can enrich true GPCR nanobody binders from a synthetic VHH library, but the paper’s own benchmark says this does not yet transfer cleanly to soluble proteins or non-GPCR membrane proteins. For STRC, this is therefore a method caution and possible secondary interface-scoring gate, not a new therapeutic hypothesis.

Citation

Harvey EP, Smith JS, Hurley JD, et al. “In silico discovery of nanobody binders to a G-protein coupled receptor using AlphaFold-Multimer.” Nature Communications, 2026. DOI: 10.1038/s41467-026-72093-5.

Numbers that matter

ParameterValueUnitsSourceSTRC use
GPCR benchmark positives32nanobody-target pairsResults, benchmarking setPositive-control scale for method confidence
GPCR benchmark negatives127permuted non-binder pairsResults, benchmarking setNegative-control construction pattern
Non-GPCR membrane positives17nanobody-target pairsResults, benchmarking setShows weak transfer to non-GPCR membrane proteins
Non-GPCR membrane negatives376permuted non-binder pairsResults, benchmarking setNegative-control construction pattern
Soluble-protein positives49nanobody-target pairsResults, benchmarking setShows weak transfer to soluble targets
Soluble-protein negatives1469permuted non-binder pairsResults, benchmarking setNegative-control construction pattern
Interface contact cutoff10A C-alpha distanceResults, metric parserReusable for AF3/AF-M interface feature extraction
GPCR AUROC, average pTM0.73AUROCResults, Fig. 2 / Supp. Table 2Best single GPCR discriminator reported
GPCR AUROC, LCF0.71AUROCResults, LCF paragraphComposite score is robustness-oriented, not best single metric
GPCR top-5% precision0.93-1.0AUC5%Results, Fig. 2Useful only in high-rank screening regime
Soluble/non-GPCR AUC5%0.22AUC5%Results, Fig. 2Direct warning against applying as STRC-binder discovery evidence
Prospective library size10000VHH sequencesResults, MRGPRX2 screenScale required for enrichment
Liability-filtered hit threshold179 / 10000candidatesResults, MRGPRX2 screen1.79% exceeded best GPCR negative-control LCF
Extracellular predicted hits177 / 179candidatesResults, MRGPRX2 screenTopology filter for receptor-targeting screens
Expressed candidates10nanobodiesResults, MRGPRX2 screenWet-lab validation scale after compute triage
Validated top binders3nanobodiesResults / Table 1Ranks 1, 5, 7 validated as binders
Sim8619 Kd200 +/- 20; 100 +/- 10nMTable 1, HEK293T; ROSANanomolar validation anchor
Sim9877 Kd160 +/- 30; 20 +/- 4nMTable 1, HEK293T; ROSANanomolar validation anchor
Sim4784 Kd80 +/- 30; 50 +/- 10nMTable 1, HEK293T; ROSANanomolar validation anchor

Method essentials

  • AF-M predictions used local ColabFold / AF-M v2 or v3, 3 recycles, no templates, no Amber relaxation.
  • Metrics combined into LCF: average pTM, best model pTM, average interface PAE, best model interface PAE, average interface pLDDT, best model interface pLDDT. Best pDockQ was reported but excluded from LCF as redundant with pLDDT.
  • Negative controls were generated by permuting nanobody-antigen pairs, then checking that antigens were not too closely related by BLAST.
  • Developability liabilities were filtered before expression: CDR glycosylation sites and predicted polyreactivity.
  • Code: kruselab/MRGPRX2-AF-M-screen, MIT license, main commit observed 2025-03-18.

Limitations

  1. Target-class limitation is explicit: GPCR works; soluble and non-GPCR membrane targets do not show useful enrichment in the paper’s benchmark.
  2. The GPCR advantage likely depends on the large and stereotyped GPCR-nanobody structure set in PDB.
  3. The method is a binder-enrichment/ranking workflow, not an affinity predictor for arbitrary protein-protein interfaces.
  4. It requires a large VHH design library and wet-lab expression/cell-binding assays to validate hits.

Relevance to STRC

  • h09: applicable only as a secondary scoring discipline for AF3/AF-M interface predictions and negative controls. It should not override the current Phase 4m/4l/4k gates because h09 is peptide/actin/TMEM145, not GPCR nanobody discovery.
  • h26: not actionable until a structural candidate passes the homodimer gate. It does not rescue current h26 failures because the paper says non-GPCR transfer is weak.
  • New VHH therapy hypothesis: not warranted. A nanobody against TMEM145/STRC would add delivery, immunogenicity, size, and validation burdens while relying on a method the paper does not validate for this target class.

Connections