AI for Bio · Model

RF3

Open co-folding model from a protein design institute, released under BSD with the AtomWorks data framework

Structure & co-foldingAcademic labUnited StatesActive中文

Overview

NameRF3 · aliases: RF3, RoseTTAFold3, RoseTTAFold 3
Full nameRoseTTAFold3
Ambiguous names (not auto-linked)RoseTTAFold
FamilyNot a family entry. Single model; design models sharing the repository have their own entries
OrganizationUniversity of Washington Institute for Protein Design, RosettaCommons
Organization typeAcademic lab
CountryUnited States
Model classCo-folding (arbitrary biomolecular complexes)
DomainStructure & co-folding (also: Small molecules)
First release · preprint
Basis: bioRxiv preprint v1 (AtomWorks and RF3) source
Other milestones
  • First public evidence: · bioRxiv preprint v1 source
  • Preprint v1: · bioRxiv preprint v1 source
  • Code release: † · Announcement of open-source code in the Foundry repository source
  • Weights release: † · Weights public with the Foundry open-source announcement source
  • Official announcement: † · RosettaCommons announcement source
Latest update · maintenance only
Basis: Foundry repository commit log
Foundry repository restructure (shared infrastructure moved into model packages); latest tag remains v0.2.0 (2026-06-05); commits to 2026-11-04 are docs; no RF3 weight change
Note: No 'Foundry 1.0' tag exists source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Shared repository has a release and maintenance commits within 12 months
OpennessCode: Open · BSD
Weights: Open · BSD
Code, weights and data pipeline under BSD; exact BSD variant not recorded this round
Benchmarks reported1 counted (board 1)
Wet-lab validationNone recorded
Unverified fields† release_milestones.code_release, release_milestones.weights_release, open_status.code_license, open_status.weights_license
As of2026-11-16

What it does

Successor to RoseTTAFold All-Atom, retrained on the AtomWorks data framework to predict arbitrary biomolecular complexes with better chirality handling. AtomWorks serves both structure prediction and design and is also used by RFdiffusion2/3. Like OpenFold3 and Protenix it is an open reproduction route, but it comes from a design lab sharing one data pipeline with design models.

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
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboard (2024-01+ subset)BoardMaintainer leaderboard (2024-01+ subset) · evidence

Other evaluations (outside the 44)

  • Gap to AlphaFold 3 (Comparison against baselines only) Self-reported: The preprint says it narrows the gap between closed AF3 and open implementations; no values taken this round

Developers report: it narrows the gap between closed AlphaFold 3 and open implementations and improves chirality handling and atom-distance conditioning.

Independent evidence: Listed on the FoldBench maintainer leaderboard, run by the board.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB, processed through the AtomWorks pipeline source
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: The University of Washington Institute for Protein Design (IPD) is an academic protein design institute whose director shared the 2024 Nobel Prize in Chemistry; RF3 was developed by two of its labs with the Microsoft AI for Good lab.
  • Funding: Predecessor RoseTTAFold All-Atom acknowledged The Audacious Project, HHMI, the Gates Foundation, Open Philanthropy, Schmidt Futures, the Helmsley Trust and a Microsoft donation; RF3's full funding statement was not transcribed.
  • Products and partnerships: Product line: RoseTTAFold All-Atom (predecessor), the RFdiffusion series (design) and RF3 (open-weight co-folding), with code in the Foundry repository; collaboration with Microsoft.
  • Related models: RFdiffusion2 / RFdiffusion3, OpenFold3, Protenix, AlphaFold 3

Reading notes

  1. No verifiable self-reported numbers exist, so it cannot be called better or worse than AlphaFold 3 or Boltz.
  2. The point is AtomWorks: a unified data and training pipeline matters more for design models.
  3. Foundry has tags up to v0.2.0 only; the claim that 'Foundry 1.0 was released' is unsupported.
  4. Design-side enzyme active-site results belong to RFdiffusion2/3, not RF3 itself.

Caveats

  1. The date follows bioRxiv (2025-08-14); some second-hand sources say 2025-09.
  2. No PoseBusters result was found in two search rounds.
  3. No journal version and no RF3-specific new weights found.

Sources

  1. AtomWorks and RF3 bioRxiv preprint (preprint)
  2. RosettaCommons: AtomWorks and RF3 open-sourced in Foundry (official site)
  3. Foundry GitHub releases (repository)
  4. Foundry commit log (repository)
  5. Third-party service page: RoseTTAFold3 (second-hand) (second-hand) second-hand

Corrections

3 corrections
FieldWasNowBasisFound
latest_update2026-06-05 (v0.2.0)2026-10-16 (repository restructure commit)source2026-11-16
latest_update.descriptionFoundry 1.0 restructure / Foundry 1.0 releasedRemoved: no Foundry 1.0 tag exists; latest is v0.2.0 (2026-06-05)source2026-11-16
benchmarks_reportedNo verified site benchmark pairingListed on the FoldBench maintainer leaderboard (official participation)source2026-11-16

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