Overview
| Name | RFantibody · aliases: RFantibody |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. Single-model entry, not a family |
| Organization | University of Washington Institute for Protein Design |
| Organization type | Nonprofit institute |
| Country | United States |
| Model class | De novo antibody design pipeline |
| Domain | Antibodies |
| First release | · preprint Basis: bioRxiv v1 preprint (single-domain antibody version); the code repository creation date is not used source |
| Other milestones | |
| Latest update | † partly verified · code release Basis: GitHub code release adding JSON input and TCR-MHC support Later items are a docs site (2026-09-23) and a licence clarification covering weights (2026-11-05); whether these count as updates is a policy call Note: The later docs and licence edits are not folded into this field; marked unverified source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Code release within 12 months |
| Openness | Code: Open · MIT Weights: Open · MIT MIT licence covers code and weights; no training code provided |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 3 recorded campaigns |
| Unverified fields | † latest_update |
| As of | 2026-11-16 |
What it does
A fine-tuned diffusion model generates backbones for a specified epitope, a sequence-design tool supplies sequences, and a fine-tuned folding model filters; it covers VHH, scFv and full-length antibody formats.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Yeast display of VHH designs; Affinity maturation (OrthoRep); Cryo-EM validation; scFv and full-length IgG designs
Other evaluations (outside the 44)
- Affinity maturation (Wet-lab validation) Self-reported: Developers report initial binders matured to single-digit nM while keeping epitope selectivity
- scFv and full-length IgG designs (Wet-lab validation) Self-reported: Developers extend the pipeline to scFv and IgG formats with yeast-display confirmation
Developers report: Developers report the first atomically accurate de novo antibody design to a specified epitope, validated by cryo-EM
Independent evidence: A competitor's preprint uses it as a baseline, and an independent compilation shows a low structure-filter pass rate (computational, not wet lab)
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Diffusion and folding models fine-tuned on public PDB antibody complexes |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | Yes | Yeast display, SPR, cryo-EM and affinity maturation used for validation only |
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.
Denominator basis: Four epitopes with about 9,000 designs per target; the developers give a 0%–2% success range and 28 VHHs confirmed by SPR; per-target percentages sit in supplementary material and are not cited
The 9,000 figure and the 0%–2% range are verified in the published paper, lifting the earlier unverified flag; developer-run
Denominator basis: Three further targets with 95 designs each, measured by single-concentration SPR
Recorded separately from the yeast-display arm; developer-run
Denominator basis: Structures solved for selected complexes: one VHH complex at 3.0 Å and two scFv structures at 3.6 Å and 6.1 Å
Cryo-EM, not negative stain; the institutional news page's looser wording must not replace the paper's resolutions
Background
- Organization: The developer is a nonprofit protein-design institute at the University of Washington
- Funding: Funders include NIH, Microsoft, Open Philanthropy, the Gates Foundation and HHMI
- Products and partnerships: Related technology is licensed to an AI-native biotech company; some authors are co-inventors on a patent
- Related models: RFdiffusion2 / RFdiffusion3, RF3, IgGM / IgGM-FM, tFold, Origin-1
Reading notes
- Only the 0%–2% range may be used; per-target percentages in the supplement cannot be cited
- The two arms (9,000 and 95 designs per target) differ greatly in scale and must be read separately
- No results reported on any of the 44 benchmarks
Caveats
- Whether the docs site and licence edits count as updates is a policy call; marked unverified
- All experiments were developer-run, not independent replications
Sources
- PubMed/PMC record (journal)
- bioRxiv preprint (preprint)
- Crossref record 10.1038/s41586-025-09721-5 (journal)
- GitHub repository RosettaCommons/RFantibody (repository)
- Official page ipd.uw.edu (company blog)
Corrections
2 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| wetlab_validation | Electron microscopy validation, no resolution given | Cryo-EM at 3.0 Å / 3.6 Å / 6.1 Å | source | 2026-11-16 |
| wetlab_validation | The 9,000-per-target figure and 0%–2% range marked unverified | Verified in the published paper and citable; the 95-per-target expression arm is added | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.