AI Potluck
Back to Gap Map Model components / Scientific AI models

RFdiffusion

Rosetta Commons
open weights / Overall score: 2.0

RFdiffusion is a diffusion model for de novo protein design from the Institute for Protein Design, built on RoseTTAFold's architecture and trained parameters. It generates unconditional monomers, scaffolds functional motifs, designs binders against a target with specified hotspot residues, builds symmetric oligomers, and partially rediffuses an existing backbone to diversify it.

GitHub's license endpoint reports NOASSERTION, but the LICENSE body is BSD three-clause text under a University of Washington copyright which states that it covers the model weights as well as the source code. Verified 2026-09-10 via the GitHub API, the README, and the LICENSE body.

Openness

3 high confidence
3.0
weights
open(checkpoints downloaded from files.ipd.uw.edu and covered explicitly by the repository LICENSE)
data
closed(training set not released
code
partial(inference scripts and hydra configs only
license
BSD-3-Clause(OSI)

GitHub reports NOASSERTION, which would have left this product looking unlicensed. The LICENSE body is BSD three-clause text under a University of Washington copyright, and it says outright that it covers the model weights as well as the source, which is the rarer and more useful of the two statements.

Adoption

2 medium confidence
2.0

Weights are served from an institutional file host rather than the Hub, so there is no download count; 3,043 GitHub stars are the last-resort signal.

Capability

2 high confidence
2.0

One attribute, and it is the one most of this category lacks - RFdiffusion designs proteins that do not exist rather than predicting a structure for one that does. The conditioning inputs are varied but all derive from the same PDB structure, which the category's note says does not make them multiple instruments.

  • https://raw.githubusercontent.com/RosettaCommons/RFdiffusion/main/README.md recorded 2026-09-10

    Documents unconditional monomer generation, motif scaffolding, binder design against a target with hotspot residues, symmetric oligomer design and partial diffusion of an existing backbone; conditioning comes from an input .pdb with contigs, secondary structure and block adjacency, and the repository publishes inference only.

Verified 2026-09-10