Overview
| Name | RFdiffusion2 / RFdiffusion3 · aliases: RFdiffusion2, RFdiffusion3, RFD2, RFD3 |
|---|---|
| Full name | Two models in one record, so there is no single full name |
| Ambiguous names (not auto-linked) | RFdiffusion |
| Family | Covers RFD2 and RFD3 only; the antibody fine-tuned line, the 2026-10 latent-reasoning successor branch and the sibling structure predictor are each recorded separately.
|
| Organization | University of Washington Institute for Protein Design, Rosetta Commons |
| Organization type | Academic lab |
| Country | United States |
| Model class | Design (diffusion, atomic level) |
| Domain | Proteins (also: Molecule generation, Structure & co-folding) |
| First release | † partly verified · preprint Basis: RFD2 preprint (date carried from the existing note); the RFD3 preprint 2025-09-18 and open-source release 2025-12-03 are listed separately Note: Unverified: the first release covers RFD2 only; see the family field for the boundary source |
| Other milestones |
|
| Latest update | · journal Basis: RFD3 journal publication (Crossref issued 2026-10-15) RFD3's journal version appeared; the Foundry's latest tag and package release is v0.2.0 (2026-06-05) with no 1.0 tag, and commits to 2026-11-04 are untagged. Note: The earlier 'Foundry 1.0 released' claim has no such tag and is removed source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Journal publication, a tagged release and ongoing commits within 12 months |
| Openness | Code: Open Weights: Open RFD3 ships weights and training code through a community foundation's Foundry; licence text was not retrieved this round, so the terms are left empty |
| Benchmarks reported | 1 counted (board 1) |
| Wet-lab validation | 3 recorded campaigns |
| Unverified fields | † first_release, other_evaluations[2], open_status.code_license, open_status.weights_license |
| As of | 2026-11-16 |
What it does
RFD2 generates enzyme backbones conditioned on catalytic-residue geometry; RFD3 moves to a co-folding-style architecture and generates all-atom designs that interact with proteins, DNA and small molecules, with self-reported lower compute cost. Unlike the backpropagation route of BindCraft it is a trained generative model, and compared with BoltzGen it leans toward enzymes and all-atom interfaces.
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.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 40 MotifBench | Official leaderboard or results | official board entry (RFdiffusion3, 2026-08 snapshot) | Board | The board lists RFdiffusion3 (43.52, 2026-08); the Chinese model page says no entry was found, which is wrong. · evidence |
Other evaluations (outside the 44)
- Own 41-active-site set (Own task set) Self-reported: An author-built set of enzyme active sites reporting how many could be scaffolded versus the comparator (self-reported)
- Metallohydrolase design (Wet-lab validation) Self-reported: RFD2-designed metallohydrolases were tested in two rounds and the crystal structures match the design models (journal paper)
- Successor-branch sampling-efficiency comparison (Comparison against baselines only) Third-party: A vendor blog claims the successor branch needs fewer sampling steps and less memory; not peer reviewed and unverified
Developers report: Developers report that RFD3 beats prior methods across in silico evaluations at about one tenth of the compute cost, with wet-lab characterisation of a DNA-binding protein and a hydrolase (self-reported).
Independent evidence: The MotifBench board lists an RFD3 entry (no score quoted here).
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Distilled PDB and structure-database data source |
| b. Private or licensed data | No | No private or licensed data source |
| c. Own wet-lab data used for training | No | No own training data source |
| d. Wet lab used only for validation | Yes | Yes: small-scale enzyme and DNA-binding-protein validation 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.
Denominator basis: Backbones were generated for all 41 active sites (in silico; the comparator managed 16), and for each of three catalytic mechanisms an active candidate appeared within fewer than 96 tested sequences
Verified at abstract level only, without reading full-text kinetics; scaffolding is not activity
Denominator basis: The preprint describes characterisation of a DNA-binding protein and a cysteine hydrolase at abstract level only, with no counts
Verified at abstract level with no figures published; developer-run, not independent replication
Denominator basis: 67 designs were tested; this belongs to the 2026-10 latent-reasoning successor branch, not to RFD2 or RFD3 themselves
Not peer reviewed, and the activity is catalytic conversion rather than binding; partner-lab executed, not independent replication
Background
- Organization: A protein design institute at a US public university, funded by several foundations, a large software company and NIH; RFD3 is released through a community foundation's open-source channel rather than run by one company.
- Funding: Institutionally funded; no funding rounds at model level
- Products and partnerships: It succeeds the institute's structure predictor and earlier diffusion design model; design companies spun out of the institute are not tied to this model as products.
- Related models: RF3, RFantibody, Proteina-Complexa, BindCraft, BoltzGen
Reading notes
- The active-site set is author-built, and scaffolding a backbone is not the same as catalytic activity.
- The wet-lab volume is small and cannot support a claim that enzyme design is mature.
- RFD3 is open and ships training code, making it one of the few reproducible all-atom design routes.
- MotifBench is its only site board relation; every other evaluation sits on author-built sets.
Caveats
- One record covers two models, so RFD2 and RFD3 must be told apart when reading figures.
- The successor branch's enzyme test comes only from an institutional press release and is not peer reviewed.
- The earlier 'Foundry 1.0 released' claim has no such tag and is removed; the latest tag is v0.2.0 (2026-06-05).
- The first-release date covers RFD2 only and is marked unverified.
Sources
- Institute release page: RFD3 now available (official site) ·
- RFD3 preprint (preprint)
- RFD2 journal paper (journal)
- RFD3 journal record (Crossref) (archive)
- Foundry repository (tags and versions) (repository)
- MotifBench official repository board (leaderboard)
- Partner university news release (successor branch) (press)
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update.description | The record stated that Foundry 1.0 had been released | No such tag exists and the claim is removed; the latest tag and package release is v0.2.0 (2026-06-05) | source | 2026-11-16 |
| benchmarks_reported | The Chinese model page said no MotifBench entry was found | RFD3 is an official entry on that board (2026-08 snapshot) and the relation is official participation | source | 2026-11-16 |
| release_milestones.journal_online | The site said RFD3 was still preprint-only as of 2026-11-16 | RFD3's journal version appeared 2026-10-15 and is listed as its own milestone | 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.