AI for Bio · Model

HelixFold-S1

A contact-guided sampling method added on top of HelixFold3, not a new base model

Structure & co-foldingBig-tech labChinaActive中文

Overview

NameHelixFold-S1 · aliases: HelixFold-S1
Full nameNo separate expanded name
Ambiguous names (not auto-linked)HelixFold
FamilyNot a family entry. Single-model entry, not a family
OrganizationBaidu
Organization typeBig-tech lab
CountryChina
Model classContact-guided sampling method
DomainStructure & co-folding
First release† partly verified · preprint
Basis: arXiv 2507.09251 v1 (per the earlier record; only the v3 date was verified) source
Other milestones
  • Preprint v1: † · arXiv v1 (date not re-verified) source
  • Journal online: † · Nature Machine Intelligence online (Crossref) source
Latest update · paper revision
Basis: arXiv v3 revision
arXiv v3 revision, after journal online publication on 2026-07-02; code and weights still not confirmed released source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Journal publication and paper revision within 12 months
OpennessCode: Closed
Weights: Closed
Code and weights promised after acceptance but not seen released; an online service is available
Benchmarks reported0 counted · none of the 44
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.journal_online
As of2026-11-16

What it does

Not a new base model but a sampling method on top of HelixFold3: it first predicts inter-chain contacts, then samples likely interfaces, reaching comparable accuracy with far fewer samples.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Self-built PDB interface sets; Sampling efficiency

Other evaluations (outside the 44)

  • Self-built PDB interface sets (Own task set) Self-reported: Post-2021 PDB complexes: antibody-protein 221, protein-protein 198, ligand 238, RNA 177, DNA 254
  • Sampling efficiency (Comparison against baselines only) Self-reported: Compared with baselines on 74 antibody complexes at about one-tenth the sampling

Developers report: Developers report contact-guided sampling gives higher interface accuracy at about one-tenth the sampling, with the largest gain on antibody-antigen interfaces

Independent evidence: No independent evidence found. No independent evidence found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB entries before 2021-09-30 plus structure-database distillation data; antibody data from SAbDab
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: A biocomputing team at Baidu, built on its own deep-learning framework, internally funded; headcount not disclosed
  • Funding: Not disclosed or not applicable
  • Products and partnerships: The same team also built HelixFold3 and two design platforms; no HelixFold4 found
  • Related models: AlphaFold 3

Reading notes

  1. All evaluations are self-built PDB subsets outside the 44 benchmarks; on the zero-pairing list under this round's rules
  2. Code and weights are unreleased, so the sampling-efficiency claim cannot be reproduced externally

Caveats

  1. Neither the arXiv v1 date nor the journal online date is fully verified
  2. It is a method, not a base model; not directly comparable with co-folding models

Sources

  1. arXiv preprint 2507.09251 (preprint)
  2. Journal article DOI 10.1038/s42256-026-01264-2 (journal)

Corrections

1 correction
FieldWasNowBasisFound
latest_update.event_kindJournal publicationPaper revision; the journal event is listed separately under 2026-07-02source2026-11-16

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