Backbone generation · Baker lab · 2023

RFdiffusion

Generates protein backbones from design constraints by fine-tuning RoseTTAFold to denoise residue positions and orientations. It supports binder design, symmetric assemblies and motif scaffolding. Sequence design is a separate step downstream.

In plain language

It starts from random noise and removes the noise step by step until a protein backbone appears. You can pin parts of it: keep this motif, bind that target, obey this symmetry.

One way to picture it

The same trick that turns static into a picture, applied to a protein. You hold a few pieces fixed and let the process invent everything around them.

Commonly misread as

What comes out is geometry, not a protein. It has no sequence yet, and the reported success rates already assume heavy filtering before anything is synthesised.

How it is put together

Select a component to read it

Conditioning

Noisy backbone

Backbone coordinates

Try it

Generate a backbone that scaffolds a motif

python scripts/run_inference.py \
  inference.output_prefix=out/motif \
  inference.input_pdb=motif.pdb \
  'contigmap.contigs=[20-20/A10-25/15-15]' \
  inference.num_designs=10

# then give the backbones a sequence, then check they fold back:
#   ProteinMPNN  ->  ESMFold  ->  keep what matches

These snippets have not been executed here. Versions move; check the model card before trusting a line of it.

01 / Why it is here

Standing

Joined a pretrained structure predictor to generative diffusion and tested diverse designs experimentally. Later catalytic-motif scaffolding methods build on it directly.

02 / What sets it apart

Distinctions

Diffusion runs on residue frames, so backbone geometry stays legal without extra constraints.
The denoiser starts from a network that already knows protein structure.
The output has no sequence. ProteinMPNN and a folding check complete the pipeline.

03 / Where it stops

Limits

Success is reported per tested design, and in silico filtering does much of the work. RFdiffusion2 removed the requirement to specify residue positions for a catalytic motif, scaffolding all 41 active sites in its benchmark against 16 for earlier methods.

Weights and code

Checked against the registry, not from memory

Repository	Size	Licence	Note
RosettaCommons/RFdiffusion	Open weights	BSD	—
dn6/RFDiffusion-3	All-atom successor	See repository	Community mirror

Sources

Each number above comes from one of these

Watson et al. 2023, Nature. De novo design of protein structure and function with RFdiffusionPeer reviewed Ahern et al. 2025, Nature Methods. Atom-level enzyme active site scaffolding using RFdiffusion2Peer reviewed

Same task, other answers

Backbone generation

Backbone generation2025

RFdiffusion2

Enzymes from a geometry, not from a residue numbering.

Designs scaffolds directly from the geometry of catalytic functional groups, without specifying which sequence positions those residues occupy and without inverse rotamer generation..

Read the entry

Back to the catalogueCompare it with its neighbours