Protein Design · Baker Lab / Institute for Protein Design
RFDiffusion
Generates de novo protein backbones and targeted binders using SE(3) denoising diffusion.
What RFDiffusion is
RFDiffusion is a structure-generation method developed by the Baker Lab and the Institute for Protein Design. It generates de novo protein backbones from random noise, operating either unconditionally or conditioned on structural contexts such as target binding surfaces, functional motifs, or symmetry groups.
The model produces backbone coordinates where newly generated regions have poly-glycine sequences. It serves as the initial backbone-generation stage in computational protein design workflows, typically followed by sequence design models such as ProteinMPNN to place functional amino acid side chains.
RFDiffusion addresses diverse macromolecular engineering tasks, including protein binder design, functional site scaffolding, and structural diversification around existing folds, with experimental validation across multiple target proteins.
How it works
RFDiffusion generates structural coordinates using a denoising diffusion process implemented with SE(3)-Transformers. Although originally trained across 200 discrete timesteps, inference typically uses 50 steps, or approximately 20 steps for faster sampling without loss of in silico quality. Computation time scales quadratically, as O(N^2), with the total residue count of the system.
Conditioning is managed through contig specifications, which designate fixed target chains, scaffolded motifs, and variable residue length ranges to build. For targeted binder generation, user-defined hotspot residues guide the model towards specific interface sites within 10 Å C-beta distances. Specialised checkpoints are provided for distinct contexts, including fine-tuned models for minimalist active-site motifs, sequence-masked inpainting, secondary-structure and block-adjacency conditioning, and beta-sheet topologies.
Key facts
- Version
- 1.1.0
- Architecture
- SE(3)-Transformer denoising diffusion model
- Core tasks
- Backbone generation, binder design, motif scaffolding, design diversification
- Input formats
- Target PDB structure, contig map specifications
- Output formats
- All-atom protein coordinates (PDB format)
- Computational complexity
- O(N^2) with respect to residue count
- GPU memory requirement
- 40 GB
- Default diffusion steps
- 50 (reduced from 200)
What you can use it for
- De novo binder design
- Generates binding protein backbones directed against specific target proteins, guided by designated target hotspot residues.
- Motif and active-site scaffolding
- Builds new supporting scaffold structures around predefined functional motifs, using dedicated active-site checkpoints to hold small catalytic geometries fixed.
- Symmetric oligomer generation
- Generates symmetric protein architectures adhering to cyclic, dihedral, or tetrahedral symmetry groups.
- Design diversification via partial diffusion
- Applies intermediate noise and denoising cycles to existing structures to explore topological variants around a general protein fold.
- Flexible peptide binder design
- Designs binders against peptide targets using secondary structure conditioning without requiring fixed three-dimensional peptide coordinates.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on Managed GPU endpoint
Runs on managed compute. Availability and startup time depend on current capacity.Run this model
RFDiffusion
Strengths and limits
Good at
- Supports both unconditioned structure generation and precise conditioning on target surfaces, motifs, or symmetries.
- Inference timesteps can be reduced to 50 or approximately 20 steps, reducing computational runtime while retaining in silico design quality.
- Includes specialised fine-tuned model weights for delicate active sites, sequence-masked core repacking, and alternative beta topologies.
- Permits low-resolution topological guidance via secondary structure and block adjacency inputs.
Watch out for
- Generates only protein backbone coordinates with poly-glycine sequences for designed regions, necessitating downstream sequence design with external tools like ProteinMPNN.
- Runtime scales quadratically with residue count, requiring large targets to be structurally truncated to prevent excessive compute demands.
- The default model exhibits an architectural bias towards generating primarily alpha-helical structures.
- Binding design remains challenging for target sites lacking hydrophobic character or located near flexible glycans.
Inputs and outputs
You provide
Target structure (PDB)
Required. The contig map refers to this structure's chains.
Contig mapoptional
RFDiffusion contig specification, e.g. "A17-145/0 100-100".
Diffusion stepsoptional
More steps refine the backbone further and cost proportionally more time.
You get back
All-atom coordinates in PDB format, viewable and downloadable.
Licence
Academic and non-commercial use only · Non-commercial (BSD-derived)
Academic tier only. Commercial users are routed to BindCraft + ProteinMPNN for binder design.
Questions
Can I use RFDiffusion commercially?
Academic and non-commercial use only. Academic tier only. Commercial users are routed to BindCraft + ProteinMPNN for binder design.
How do I run RFDiffusion without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 2× A100, typically in 7.0min, and a run costs about 3,360 credits. Visitors without an account get 1,000 credits a month.
Why do designed regions in the output PDB file contain only glycine residues?
RFDiffusion is exclusively a backbone coordinate generation model. It outputs poly-glycine backbones for newly built segments, which must subsequently be assigned amino acid sequences using sequence design tools like ProteinMPNN.
How should target hotspot residues be chosen for binder design?
Users should select three to six residues within 10 Å C-beta distance of the desired interface, ideally containing at least three hydrophobic residues. Because the model was trained with masked hotspots, it expects to form additional interface contacts beyond those explicitly provided.
How does partial diffusion differ from standard generation?
Partial diffusion starts from a known structure and noises it up to an intermediate timestep before denoising, creating structural diversity around the starting fold without changing total residue length.
What in silico filtering metric is recommended for binder backbones?
Candidate complexes are typically evaluated using AlphaFold2 with initial guess and templating; designs achieving an interface predicted alignment error (pae_interaction) under 10 are prioritised for experimental testing.
What symmetries can be generated?
RFDiffusion supports symmetric generation and motif scaffolding under cyclic, dihedral, and tetrahedral symmetry configurations.
How to cite it
Written from RFDiffusion's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.