Use RFdiffusion Online
Commercially Available RFdiffusion No-Code Web Server
RFdiffusion: Diffusion-Based De Novo Protein Design for Binders, Symmetric Assemblies and Motif Scaffolding
RFdiffusion (RoseTTAFold diffusion) is a generative model for de novo protein backbone design, obtained by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks. It was developed by Joseph L. Watson, David Juergens, Nathaniel R. Bennett, Brian L. Trippe, Jason Yim and colleagues, with senior authors Frank DiMaio, Minkyung Baek and David Baker, and published in Nature in 2023. The code is free for academic, personal and commercial use on GitHub.
Deep learning has advanced protein design, but the authors note that no general framework had yet covered a wide range of challenges, including de novo binder design and higher-order symmetric architectures. Diffusion models had succeeded for images and language but had limited success on proteins, probably because of the complexity of backbone geometry and sequence-structure relationships.
RFdiffusion addresses this by learning to reverse noise added to protein structures, so that designs can be generated from simple specifications such as a target, a motif or a symmetry, much as images are generated from user inputs.
Property | Detail |
|---|---|
Method | Denoising diffusion over residue frames, fine-tuned from RoseTTAFold |
Sequence design | ProteinMPNN, typically eight sequences per backbone |
Design tasks | Monomers, binders, symmetric oligomers, motif and active-site scaffolding |
In silico filter | AlphaFold2 (AF2) prediction compared with the design model |
Availability | GitHub and a Google Colab notebook |
How RFdiffusion Generates Protein Backbones
Noising and denoising: Calpha coordinates are perturbed with 3D Gaussian noise and residue orientations are noised on the rotation manifold; the network learns to reverse this process.
Self-conditioning: the model conditions on its own previous predictions between timesteps, which the authors found notably improved performance.
Conditional inputs: a target, hotspot residues, a functional motif, secondary-structure information or a symmetry can guide generation, and fine-tuned variants cover complexes and fold specification.
Sequence design and filtering: ProteinMPNN designs sequences for each backbone, and AF2 predictions are used to select designs.
A design counts as an in silico success when the AF2 prediction has a pAE below five, lies within 2 angstroms backbone r.m.s.d. of the design model, and, for scaffolding, reproduces the functional motif within 1 angstrom.
RFdiffusion Experimental Results for Binder, Symmetric and Metal-Binding Designs
Design task | Reported result |
|---|---|
p53 helix scaffolding against MDM2 | Of 96 designs, 55 showed detectable binding at 10 uM; 0.5 and 0.7 nM binders found (native peptide 600 nM) |
C4 Ni2+-binding assemblies | 44 designs expressed; 18 of 36 tested bound Ni2+, with nanomolar to micromolar affinity |
De novo binders, five targets | Overall experimental success rate 19%; hits for influenza HA, IL-7Ralpha, InsR, PD-L1 and TrkA |
Influenza HA binder HA_20 | KD of 28 nM; cryo-EM at 2.9 angstroms matches the design within 0.63 angstroms |
Monomers and symmetric oligomers
RFdiffusion generates diverse unconditional designs up to 600 residues that AF2 predicts accurately. Despite no training on symmetric inputs, it produces symmetric oligomers with high in silico success rates, especially with an auxiliary contact potential. Negative-stain EM data were consistent with several designs, including a tetrahedron and an icosahedral assembly.
Motif scaffolding
The method outperforms other methods across 25 benchmark motif-scaffolding problems and can scaffold enzyme active sites and symmetric motifs, such as C3-symmetric trimers presenting a SARS-CoV-2 binder.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with RFdiffusion on Tamarind Bio
De novo binder design: generate protein binders to a target with specified interface hotspot residues, with far fewer designs tested per target than earlier approaches.
Symmetric assemblies: design cyclic, dihedral, tetrahedral, octahedral and icosahedral oligomers.
Motif and active-site scaffolding: build proteins that hold a functional motif or catalytic site in a defined geometry.
Metal-binding proteins: design assemblies that coordinate metals such as Ni2+.
How to Use RFdiffusion on Tamarind Bio
Open the tool: log in to tamarind.bio and open RFdiffusion.
Choose the design task: if exposed, select unconditional, binder, motif scaffolding or symmetric design.
Provide inputs: upload a target or motif PDB structure where the task requires one.
Set contigs and hotspots: if exposed, specify the lengths of designed segments, the fixed motif residues and the interface hotspot residues.
Set symmetry and size: if exposed, choose a symmetry and the number of designs to generate.
Run the job: submit it and download the generated backbones.
Design sequences and filter: run ProteinMPNN on the backbones and filter with AF2 or a similar predictor, as the paper does.
Test experimentally: express and characterize the top designs.
Things to Keep in Mind
RFdiffusion generates backbones; sequences come from ProteinMPNN and in silico filtering is needed to select designs.
Designs need experimental validation; reported success rates vary by target and task.
The authors note the method has not yet been tested on more polar target sites or sites without native binding partners.
Ligand and nucleic-acid design were listed as future extensions of the original model.
Source: Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J. et al., "De novo design of protein structure and function with RFdiffusion", Nature, vol. 620, 31 August 2023, DOI 10.1038/s41586-023-06415-8. Code: https://github.com/Rosetta-Commons/RFdiffusion.