AI for Bio · Model

BindCraft

Open binder design pipeline that backpropagates through a co-folding network, with no extra training

ProteinsAcademic labSwitzerlandActive中文

Overview

NameBindCraft · aliases: BindCraft
Full nameSources use BindCraft only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single pipeline with version iterations; no family members
OrganizationEPFL Laboratory of Protein Design and Immunoengineering, MIT
Organization typeAcademic lab
CountrySwitzerland
Model classDesign (hallucination via backpropagation)
DomainProteins (also: Molecule generation)
First release · preprint
Basis: bioRxiv v1 (later v2 2024-12-07, v3 2025-04-25) source
Other milestones
  • First public evidence: · bioRxiv v1 (later v2 2024-12-07, v3 2025-04-25) source
  • Preprint v1: · bioRxiv v1 (later v2 2024-12-07, v3 2025-04-25) source
  • Journal online: · Journal version online (Crossref; print issue 2025-10-09) source
  • Code release: No sourced date† · The repository is public but its first release date was not retrieved this round source
Latest update · patch
Basis: Tag v1.5.3 (small fixes)
v1.5.3 small fixes; the journal version online 2025-08-27 is a separate milestone; commits to 2026-09-21 are install and notebook maintenance.
Note: The earlier record treated this date as the journal publication; corrected to a patch source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Tagged release and ongoing commits within 12 months
OpennessCode: Open
Weights: Unclear
Code is open; it has no weights of its own and depends on a third-party co-folding model's weights and licence
Benchmarks reported1 counted (board 1)
Wet-lab validation1 recorded campaign
Unverified fields† release_milestones.code_release, benchmarks_reported[0], open_status.weights
As of2026-11-16

What it does

It folds target and binder together in a co-folding network and iteratively edits the binder sequence to maximise predicted confidence, allowing some induced fit; there is no generative model and no training. Compared with RFdiffusion2/3, BoltzGen and Germinal it is the lightest and easiest to reproduce.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
28 Adaptyv competitions / BenchBBOfficial leaderboard or resultsNipah and EGFR roundsBoardThe developers took part; an early-round top-10 placement comes from an organiser post (second-hand) †partly verified · evidence

Other evaluations (outside the 44)

  • Own 12-target wet-lab set (Wet-lab validation) Self-reported: Self-chosen targets with the share of tested designs that bound; per-target denominators are very small (self-reported)
  • Application demonstrations (Case study) Self-reported: Binders blocking a pollen allergen and peptides improving a gene-therapy vector; the authors say they are extending to peptides

Developers report: Developers report that with no extra training most targets yield binders in one design round, with about 46.3% of tested designs binding (journal body text, self-reported).

Independent evidence: No independent evidence found. The competition publishes no per-tool hit rates, so performance there rests on entrants' own accounts (second-hand)

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesNo training of its own; it uses a third-party co-folding model's off-the-shelf weights, trained on PDB and other public data source
b. Private or licensed dataNoNo private or licensed data source
c. Own wet-lab data used for trainingNoNo own training data source
d. Wet lab used only for validationYesYes: the authors ran the wet-lab validation themselves source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Experimental binder success across 12 targets (self-reported) (journal version)
Targets 12 · Designs tested — · Hits — · Success 10-100% per target, average 46.3% (journal body text) · Best affinity —

Denominator basis: Twelve targets; the rate is the share of tested designs that bound, and per-target denominators are very small
The 46.3% figure is stated directly in the journal body text, so the earlier 'second-hand summary' label is dropped; the authors ran the experiments themselves, so this is not independent replication

Self-reported · · verified · source

Background

  • Organization: An academic project of a protein design and immunoengineering laboratory in the engineering school of a Swiss institute of technology, with no company entity; the university's news release describes collaboration with a US university and researchers from Switzerland, the US and the Netherlands.
  • Funding: Grants are listed only in the journal acknowledgements, not in the news release
  • Products and partnerships: Academic project; no product and no pharma partnership
  • Related models: RFdiffusion2 / RFdiffusion3, BoltzGen, Germinal, Mosaic

Reading notes

  1. The 10-100% span is wide and its average depends on target choice, so it is not an expected rate for an arbitrary target.
  2. Optimising a co-folding confidence score as the design objective disqualifies that score as an independent filter, a known issue for this family.
  3. The Adaptyv competition does not report per tool, so performance there rests on entrants' own accounts.
  4. Being open and training-free is why it is so widely used as a baseline.

Caveats

  1. The success rate is the share of tested designs that bound, with very small per-target denominators.
  2. It reports no site benchmark score; its only pairing is a competition entry record (second-hand).
  3. It has no weights of its own, so design quality inherits the biases of the co-folding model it depends on.
  4. The repository's first release date is unverified.

Sources

  1. Journal paper DOI (journal) ·
  2. Journal article page (journal)
  3. Official repository releases (repository)
  4. University news release (press)
  5. Competition results blog (second-hand) (second-hand) second-hand

Corrections

2 corrections
FieldWasNowBasisFound
latest_update.event_kind2026-04-17 was recorded as the journal publicationThat date is the v1.5.3 small fix, so the kind becomes patch; the journal version online 2025-08-27 is listed separatelysource2026-11-16
wetlab_validationThe 46.3% average was labelled a second-hand summaryThe figure is stated directly in the journal body text, so the second-hand label is dropped; the rate is the share of tested designs that boundsource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.