Use RFdiffusion3 Online
Commercially Available RFdiffusion3 No-Code Web Server
RFdiffusion3: All-Atom Diffusion for De Novo Design of Binders, Enzymes and DNA-Binding Proteins
RFdiffusion3 (RFD3) is a diffusion model that generates protein structures in the context of ligands, nucleic acids and other non-protein atoms. It was developed at the Institute for Protein Design at the University of Washington, led by David Baker and Rohith Krishna, with collaborators at DTU, Cambridge, Oxford and NVIDIA, and is described in a bioRxiv preprint (CC BY 4.0).
Most protein design methods generate backbone coordinates only and treat other biomolecules as an afterthought. That makes atom-level constraints, such as the geometry of an enzyme active site, harder to specify and less effective. RFD3 models all polymer atoms explicitly, which the authors report makes such constraints simpler and more effective.
The authors report better in silico benchmark performance than earlier methods at about one tenth of the computational cost, and they experimentally characterize designed DNA-binding proteins and cysteine hydrolases.
Property | Detail |
|---|---|
Task | All-atom de novo design of proteins that interact with ligands, DNA and other proteins |
Architecture | Transformer-based U-Net similar to the AlphaFold3 diffusion module |
Size | 168M trainable parameters (versus about 350M for AF3) |
Representation | 4 backbone atoms and 10 side-chain atoms per residue, with virtual atoms padding smaller residues |
Sequence design | Done afterward with ProteinMPNN or LigandMPNN |
How RFdiffusion3 Designs Proteins Around Ligands, DNA and Active Sites
RFD3 uses sparse attention in which only residues and atoms that are geometrically close in the noisy structure attend to each other. The Pairformer is reduced from 48 to 2 layers, and triangle updates are omitted for efficiency.
Conditioning options
Motif coordinates: fix active-site or interface atoms to scaffold around them.
Hotspots and contact labels: specify hotspot atoms, hydrogen-bond donor and acceptor labels, and relative solvent-accessible surface area.
Symmetry and placement: request symmetric assemblies through symmetric noise, and set a center-of-mass position.
Classifier-free guidance: average conditional and unconditional passes at each denoising step.
Training
RFD3 was trained on PDB complexes (protein-protein, protein-small molecule and protein-DNA), scaffolding tasks and AlphaFold2 distillation structures, using PDB entries deposited through December 2024.
RFdiffusion3 Benchmark Results for Binders, Enzymes and Symmetric Scaffolds
Benchmark | Reported result |
|---|---|
Protein binders (5 targets, 400 designs each) | Mean unique successful clusters 8.2 for RFD3 versus 1.4 for RFdiffusion1 |
Enzyme active sites (41 cases) | RFD3 beat RFdiffusion2 on 37 of 41 |
Active sites with more than 4 residue islands | 15% passing designs versus 4% for RFdiffusion2 |
Small-molecule binders (4 ligands) | Outperformed RFdiffusionAA on all four rigid ligands |
DNA binders | Monomer pass rate 8.67%, dimer 6.67% |
Unconditional designs, length 100-200 | 98% had at least one of 8 sequences predicted within 1.5 A by AF3 |
In experiments, one of five synthesized DNA-binding designs bound its target sequence by yeast surface display, with an EC50 of 5.89 +/- 2.15 uM.
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 RFdiffusion3 on Tamarind Bio
Enzyme design: scaffold catalytic motifs with atom-level constraints, as in the 41-case active-site benchmark.
Protein binder design: generate binders against targets such as PD-L1, IL-7Ra and Tie2.
DNA-binding proteins: design proteins aimed at specified DNA sequences.
Small-molecule binders and symmetric assemblies: build pockets around ligands such as FAD and SAM, or D2, C3, C5 and C7 symmetric scaffolds.
How to Use RFdiffusion3 on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select RFdiffusion3.
Provide the design context: supply the target protein, ligand or DNA structure, or the motif to scaffold, as the form allows.
Set conditioning: if exposed, choose hotspots, hydrogen-bond donor and acceptor labels, symmetry and length.
Choose how many designs to generate: the authors sampled hundreds of designs per target, so generate more than a handful.
Run the job: submit and wait for backbones to be generated.
Add sequences: RFD3 does not design sequences, so use ProteinMPNN or LigandMPNN, if exposed.
Filter and download: rank designs with a structure predictor, using criteria such as predicted interface error, pTM and RMSD, as in the paper.
Things to Keep in Mind
Most success criteria rely on AF3 or Chai-1 predictions rather than experimental structures.
Sequences must be designed separately, and the authors note they are not always consistent with the backbone.
DNA-binder success rates were about 6-9% in silico, and only one of five tested designs bound, with a micromolar EC50.
Unconditional designs are biased toward alanine relative to native proteins.
Wet-lab validation covers a small number of designs.
Source: Butcher et al., De novo Design of All-atom Biomolecular Interactions with RFdiffusion3, bioRxiv preprint (v2, 2025), doi.org/10.1101/2025.09.18.676967. No code repository was listed on the fetched page.