2026 Sang - AF3-TurboAb antibody structural prediction

AF3-TurboAb is a practical repertoire-scale wrapper around AF3 for antibody-antigen complexes. It removes the CPU preprocessing bottleneck by using antibody-specific MSA/template libraries and reusing antigen features, then validates ipTM as an operational confidence metric for antibody/nanobody interfaces.

For STRC, this is a method precedent for high-throughput interface triage, decoy specificity checks, and atlas-style epitope mapping. It is not direct evidence that AF3 ipTM predicts affinity for arbitrary STRC protein complexes, peptides, or small molecules.

Citation

Sang Z, Xiang Y, Huang W, Sargunas PR, Kim YJJ, Feng Z, Taylor DJ, Shi Y. “Repertoire-scale antibody structural prediction informs therapeutic design.” Science Advances 12(17), eaef7163, 2026. DOI: 10.1126/sciadv.aef7163.

Numbers that matter

ClaimValueSource in paperSTRC use
Standard AF3 complex runtime~17 min per complexFig. 1A textShows preprocessing, not inference, is the scale blocker
CPU MSA/template share97.2% of runtimeFig. 1A captionReuse/precompute features before any large STRC AF3 sweep
Nb MSA runtime~670 s to 1.6 sResults / Fig. 1C418x speedup from antibody-specific search space
End-to-end TurboAb runtime~35 s per seedResults / fig. S1APractical million-scale screening with GPU parallelism
1M Nb screen, single GPU8055 h vs 294,722 h standard AF3ResultsSpeedup changes the experimental design space
Parallelized 1M Nb screen8.4 days on 40 GPUsResultsAcademic-cluster feasible
Posttraining Nb benchmark243 complexes, 4 A, post-2022Fig. 1C/E; MethodsLeakage-reduced validation set
TurboAb vs standard AF3 ipTMPearson r = 0.95Fig. 1EOptimization preserves AF3 confidence behavior
High-confidence thresholdmean ipTM >= 0.8ResultsUse only as class-calibrated antibody interface threshold
High-confidence structural qualityDockQ > 0.49, median epitope overlap 90%ResultsipTM tracks near-experimental antibody interfaces
Cognate vs decoy hit rates~23% vs 0.2% and 0.14%Fig. 1JDecoy panels are mandatory for specificity claims
In-house affinity panelhigh 43.8%, medium 16.7%, low 3.2% high-confidencefig. S2/table S1ipTM enriches affinity class, but is not a Kd predictor
Repertoire atlas size275,371 complexes, 3 seeds eachResultsDemonstrates atlas-scale structural decoding
High-confidence atlas models39,973 total; 28,013 distinct CDR sequencesResultsExpands structural coverage beyond PDB
Training-overlap check94.4% have <60% CDR3 identity to PDB CDR3sResultsUseful leakage-control pattern
Cryo-EM validation12 HSA-Nb complexes at 3.1 to 3.8 AFig. 2Multi-epitope experimental validation
Cryo-EM agreementall epitopes correct; 10/12 close by RMSDResultsConfirms epitope identity more robustly than full antigen conformation
HSA induced fit~19 deg / 12.7 A and ~15 deg / 8.5 A shiftsResultsAF3 can miss antigen conformational rearrangement
Predicted epitope surface coverage~60 to 100% of solvent-accessible surfaceFig. 3Atlas exposes unmapped antigen surfaces
Hotspot architecture95% of interfaces contain 1 to 5 hotspotsFig. 3K/LCompact residue clusters dominate contacts
Hotspot contact share31% average; up to 86% with 5 hotspotsFig. 3LDesign should target compact contact pillars
Hotspot chemistry83.6% top seven residues, aromatic or charged/polar enrichedFig. 3NMatches salt-bridge/H-bond/aromatic interface intuition
SARS-CoV-2 RBD triparatopic designscover 41.6 to 44.4% of RBD surfaceFig. 5HMultiparatopic design can combine breadth and potency
JN.1 neutralization gainup to 2716x EC50 improvement over matched monomersFig. 5LAvidity plus epitope complementarity can be very large
PD-L1 screen700 Nbs modeled; 14 high-confidence hitsMethodsGlycan-aware AF3 inputs can select domain-specific binders
PD-L1 P10 binding6 nM human, 19 nM mouse cell EC50ResultsCross-species binders can emerge from epitope selection
PD-L1 glycoform effectdeglycosylation weakens ELISA 4.9 to 27 nMResultsGlycoform modeling can matter structurally
MERS virtual screen20,000 Nbs to 25 seeds, expanded to 6290, then 566 hitsResults/MethodsTwo-stage seed plus neighbor expansion workflow
MERS prospective hit rate5/19 expressed binders, 26.3%ResultsipTM-only triage can enrich de novo binders

Method takeaway

  • The decisive engineering move is not a new neural model. It is replacing repeated full-database MSA/template searches with antibody-focused databases plus precomputed antigen features.
  • Mean ipTM >= 0.8 is useful in the validated antibody-antigen regime. Do not import this cutoff as an affinity threshold for STRC/TMEM145, WH2-actin, homodimers, or small-molecule complexes without a local benchmark and decoys.
  • The paper’s specificity logic is the part STRC should copy: cognate sets, biologically implausible decoys, repeated seeds, and explicit leakage checks.
  • The HSA cryo-EM cases are a good caution: epitope identity can be right while antigen subdomain conformation shifts after binding. For STRC, AF3 interface success still needs geometry/MD/experiment when induced fit is plausible.
  • The hotspot analysis is design-relevant beyond antibodies: compact aromatic/charged/polar patches can dominate contact counts, but translating this to STRC requires target-specific validation.

Relevance to STRC

h03 mini-STRC: supports continued use of AF3 as a structural gate only when paired with leakage control, decoys, and experimental follow-up. It does not change the h03 rank because h03 evidence already rests on STRC/TMEM145-specific AF3 and vector audit results.

h09 hydrogel: reinforces the need for decoy and threshold discipline in any AF3 peptide-interface gate. It does not rescue weak confidence-supported WH2-actin contacts, because the validated regime here is antibody-antigen, not short peptide-actin.

h26 engineered homodimer: does not reopen the paused branch. h26 failed its own AF3 G1 structural confidence gate; antibody repertoire screening does not provide a new homodimer scaffold or interface.

Contrast with Harvey 2026

2026-harvey-afm-nanobody-gpcr showed AF-M nanobody screening can enrich GPCR binders but transfers poorly to soluble and non-GPCR targets. Sang 2026 is stronger for antibody-antigen atlas work because it has posttraining benchmarks, decoys, HSA cryo-EM, and prospective MERS validation. The shared STRC lesson is still conservative: structural-confidence enrichment is target-class-specific.

STRC ranking impact

No ranking change. This paper improves method discipline and supports AF3-scale screening workflows, but it does not directly validate or falsify any STRC therapeutic hypothesis.

Follow-up

  • If a future STRC binder-library idea emerges, copy the paper’s decoy and leakage-control design before trusting ipTM.
  • If using AF3 at scale for STRC proteins, first identify the reusable feature layer analogous to antigen precomputation.
  • If citing this paper in a phase proof, cite the specific validated regime: antibody/nanobody-antigen complex prediction with ipTM >= 0.8, not general protein-protein affinity.
  • Future AF3/interface screens should start from STRC AF3 Interface Triage Protocol.

Connections