oteam/rfdiffusion-inpaint-seq

RFdiffusion

Denoising-diffusion generation for de novo protein backbones, motif scaffolding, inpainting, interfaces, and binder design.

Disclaimer

This is an unofficial OFoldX implementation of De novo design of protein structure and function with RFdiffusion by Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, and Woody Ahern. The upstream authors did not write, review, or endorse this implementation or model card. The official repository is RosettaCommons/RFdiffusion.

Model Details

RFdiffusion generates protein backbones by reversing diffusion over residue frames. A RoseTTAFold-style denoiser combines sequence, MSA, pair, template, and timestep features with SE(3)-equivariant structure updates. Its sampler couples Gaussian C-alpha translation diffusion with IGSO(3) rotation diffusion. The same architecture supports de novo generation, motif scaffolding, sequence inpainting, symmetric oligomers, interfaces, and target-conditioned binder design through model-specific conditioning. The family generates protein backbones, not all-atom sequence-complete designs; pair it with a sequence design family such as ProteinMPNN to obtain a complete sequence.

Variants

Released checkpoint	OFoldX repository	Scientific use
Base_ckpt.pt	oteam/rfdiffusion-base	De novo generation and motif scaffolding
Base_epoch8_ckpt.pt	oteam/rfdiffusion-base-epoch8	Earlier released base checkpoint
Complex_base_ckpt.pt	oteam/rfdiffusion-complex-base	Complex and binder design
Complex_beta_ckpt.pt	oteam/rfdiffusion-complex-beta	Topology-diverse, beta-rich binder design
Complex_Fold_base_ckpt.pt	oteam/rfdiffusion-complex-fold-base	Fold-conditioned complex design
InpaintSeq_ckpt.pt	oteam/rfdiffusion-inpaint-seq	Sequence-conditioned inpainting
InpaintSeq_Fold_ckpt.pt	oteam/rfdiffusion-inpaint-seq-fold	Fold-conditioned sequence inpainting
ActiveSite_ckpt.pt	oteam/rfdiffusion-active-site	Active-site motif scaffolding

The RoseTTAFold pretraining file RF_structure_prediction_weights.pt is not an RFdiffusion denoising checkpoint and is not supported by this converter.

Model Specification

Main Blocks	Extra Blocks	Refinement Blocks	MSA Size	Pair Size	Diffusion Timesteps	Num Parameters (M)
32	4	4	256	128	50	59.808

Links

• Code: ofoldx.models.rfdiffusion
• Paper: De novo design of protein structure and function with RFdiffusion
• Developed by: Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim, Helen E. Eisenach, Woody Ahern, Andrew J. Borst, Robert J. Ragotte, Lukas F. Milles, Basile I. M. Wicky, Nikita Hanikel, Samuel J. Pellock, Alexis Courbet, William Sheffler, Jue Wang, Preetham Venkatesh, Isaac Sappington, Susana Vázquez Torres, Anna Lauko, Valentin De Bortoli, Emile Mathieu, Sergey Ovchinnikov, Regina Barzilay, Tommi S. Jaakkola, Frank DiMaio, Minkyung Baek, and David Baker
• Model type: SE(3)-equivariant protein backbone diffusion model
• Original Repository: RosettaCommons/RFdiffusion

Construct and output contract

One candidate is one realised output topology. Each declaration group becomes an output protein chain whose rows are either copies of source components (existing spans) or generated rows of the unknown amino acid UNK: the model predicts no sequence, so a generated identity stays unknown. A copied row keeps the coordinates, the coordinate evidence the source states — a motif copied from a predicted model stays predicted — and the source frame; a generated row starts without coordinates. Under point-group symmetry the output instances are the symmetry units of one entity, so every unit holds the same sequence.

Generation regenerates the generated and diffused rows only. The fixed motif keeps the coordinates it was copied with, and decode writes the predicted N, CA, C and O of every regenerated row; side-chain atoms of such a row are left unobserved and counted, because a reconstructed side chain of a masked identity is not a prediction. Rows whose coordinates the model reads as placed need complete N, CA and C — the frame they are diffused around.

Identities and positions are addressed by the canonical target: a source position by its ComponentRef (or, in the released contig grammar, by the author numbering the source states), a generated row by its placement and offset. A new SymbolicBond selects one occurrence per endpoint; a source AtomSite with several output copies is refused, and a SymbolicAtomSite names the copy. Junctions between neighbouring rows are those the output declares: a generated row or a placement boundary always joins the next row, two copied rows of one placement join exactly when the source joins them.

Usage

The model depends on the ofoldx library. Install it with pip:

pip install ofoldx

Backbone Design

Load the RFdiffusion design pipeline:

import ofoldx
from ofoldx.models.conditions import GeneratedSegment
from ofoldx.models.rfdiffusion import RfDiffusionPlacements

placement = RfDiffusionPlacements({"binder": {"binder": GeneratedSegment("protein", 100, 100)}})
pipeline = ofoldx.pipeline("design", model="oteam/rfdiffusion-base")
output = pipeline.design(None, placement=placement)

Training Details

Training Data

RFdiffusion fine-tunes a RoseTTAFold-style structure network as a denoising diffusion model over PDB protein structures. Refer to the RFdiffusion paper for the authoritative upstream data construction. OFoldX does not redistribute the training set.

Training Procedure

The training procedure adds Gaussian C-alpha translation noise and rotational Brownian motion to residue frames, then trains the network to denoise toward the true frames with self-conditioning across timesteps. OFoldX exposes the differentiable denoiser for training and fine-tuning; it does not by itself reproduce the coupled translation, rotation, timestep, and self-conditioning training procedure.

Citation

@article{watson2023denovo,
  author = {Watson, Joseph L. and Juergens, David and Bennett, Nathaniel R. and others},
  title = {De novo design of protein structure and function with RFdiffusion},
  journal = {Nature},
  volume = {620},
  number = {7976},
  pages = {1089--1100},
  year = {2023},
  doi = {10.1038/s41586-023-06415-8}
}

If OFoldX supports your research, also cite the OFoldX software:

@software{chen2026ofoldx,
  author = {Chen, Zhiyuan},
  title  = {OFoldX},
  year   = {2026},
  url    = {https://github.com/OTeam-AI4S/OFoldX},
}

Contact

Please use OFoldX GitHub issues for questions about this implementation or model card. Questions about the original model should be directed to the RFdiffusion authors.

License

This model implementation is licensed under the GNU Affero General Public License.

SPDX-License-Identifier: AGPL-3.0-or-later

The Pipeline is callable and supports staged execution with prepare, forward, and postprocess. Prepared values retain the source and correspondence needed by the scientific result; the concrete signatures describe each stage's controls.

Converted checkpoint

This repository contains converted model parameters under bsd-3-clause; the original terms and notices are preserved in LICENSE. OFoldX implementation licensing is described separately in license.md and license-faq.md. Conversion changes parameter names and native tensor layout; the upstream authors do not endorse this implementation.

• https://github.com/RosettaCommons/RFdiffusion/blob/main/LICENSE