RFdiffusion3
latestAll-atom generative diffusion for designing binders, enzymes, and symmetric protein assemblies
All-atom generative diffusion for designing binders, enzymes, and symmetric protein assemblies
RFdiffusion3 (RFD3) is a generative diffusion model for de novo protein design. Unlike structure prediction tools that fold existing sequences, RFD3 creates entirely new protein backbones through an iterative denoising process. The model can generate proteins from scratch or design new chains that interact with specified targets.
RFD3 extends the original RFdiffusion architecture with all-atom capabilities, enabling design tasks involving small molecules, metals, and other non-protein components. The model learns to reverse a diffusion process that gradually adds noise to protein structures, generating novel backbones that satisfy specified constraints.
For sequence design after generating backbones, pair RFD3 with LigandMPNN to optimize amino acid sequences. To validate designed structures, use RosettaFold3 or Boltz-2 for structure prediction.
RFD3 operates on a diffusion framework where protein structures are progressively corrupted with Gaussian noise during training. At inference time, the model reverses this process:
The number of diffusion steps controls the quality-speed tradeoff. More steps (100-200) produce higher quality structures but take longer, while fewer steps (10-50) enable rapid prototyping.
RFD3 can incorporate various constraints during generation:
Structural constraints: Fix specific residue positions from a template structure while designing new regions around them. The contig specification language allows precise control over which regions are fixed versus designed.
Hotspot targeting: For binder design, specify which target residue atoms should be contacted by the designed protein. The model orients the new chain to maximize interactions with these hotspots.
Length constraints: Control the size of designed proteins or protein regions using length ranges. The model samples within the specified range during generation.
Provide a target protein structure for constrained design tasks like binder design. Upload a PDB/ENT file or fetch directly from RCSB using a PDB ID (e.g., 4ZXB). The target structure defines the binding interface for binder design or provides structural motifs for scaffolding tasks.
To use only a particular chain, include it in the RCSB identifier, for example 1FJS:L. Only the selected chain is passed to input validation and the design job. For an uploaded PDB, open Trim structure beside the file and keep the required chains. Trimming preserves original residue numbers and insertion codes; validation still checks the retained structure before submission.
Insertion codes distinguish residues such as 61 and 61A during structure validation. The custom contig field uses numeric residue references and cannot explicitly address A61A; passing upload validation does not establish support for insertion-coded contigs or hotspots. If the intended target is a different chain, select that chain before configuring the design.
For unconditional generation (creating proteins from scratch), leave this empty and RFD3 will generate novel folds based solely on the specified length.
Select the type of protein design:
Generate multiple independent designs (1-50) in a single job. Each design samples a different trajectory through the diffusion process, producing structural diversity. More designs increase the chance of finding high-quality candidates but require more computation time.
Controls the number of denoising iterations (10-200). Higher values produce more refined structures:
| Steps | Use Case |
|---|---|
| 10-20 | Quick prototyping, initial exploration |
| 50 | Standard quality (default) |
| 100-200 | High-quality final designs |
Specify the length of designed protein regions as a range (e.g., 50-100). The model samples within this range during generation. For binder design, this controls the size of the designed binder chain.
For binder design, choose Target chains after uploading the structure. The first chain is selected by default. Selecting other chains builds a contig from their existing residue numbers, preserving gaps and chain breaks.
The contig specification provides fine-grained control over the design:
Turn on Use custom contig to reveal its field. A non-empty custom contig replaces the target-chain selector and design length range. Turning it off returns to automatic selection; your custom text is kept while the form stays open.
If your uploaded structure has missing residues or gaps, leave Custom contig blank unless you explicitly split the fixed residue ranges to match residues that actually exist in the structure.
Contigs are written as comma-separated pieces:
A1-95 or B5-12 references residues from the uploaded structure50 or 50-100 asks RFD3 to design a region of that length/0 creates a chain breakExamples:
50-100,/0,A1-95 - Design a 50-100 residue binder and place it on a separate chain from target residues A1-A95A40-60,70,A120-170,/0,B3-45,60-80 - Keep residues A40-A60, design a 70-residue segment, keep A120-A170, then start a new chain with B3-B45 followed by a designed 60-80 residue segment100 - Generate a 100-residue protein unconditionallyDo not put spaces anywhere in the contig. RFdiffusion3 reads it exactly as written, so 56-64,/0,E395-510 works while 56-64, /0, E395-510 fails inside the model. ProteinIQ refuses a spaced contig before the job runs, and the message shows the version without spaces. Use /0, not a bare /, for chain breaks.
Target specific atoms on the target protein for binder design. This focuses the designed interface on functionally important residues.
In the playground, enable Hotspot residues and choose Edit. Search by residue identifier or name within the popover, or scroll through the full list. New selections use ALL, and selected residues can be removed individually. Turn on Atom-specific hotspots to enter explicit atom names or edit existing selections. Switching back to the residue selector preserves atom-specific values. Disabling hotspots excludes them from the run and preserves your selection for when you enable them again.
Format: JSON object mapping residue IDs to atom names
{
"A64": "CD2,CZ",
"A88": "CG,CZ",
"A96": "CD1,CZ"
}Residue ID format: ChainResidueNumber (e.g., A64 = chain A, residue 64)
The advanced field and API accept a JSON object only. Residue IDs must match residues that are actually present in the uploaded target structure.
Each value can also be one of these uppercase shorthands:
| Value | Selected atoms |
|---|---|
ALL | All heavy atoms present in the residue. |
BKBN | Backbone atoms: N,CA,C,O. All four must be present. |
TIP | Foundry's residue-specific tip atoms. The residue must have a defined tip and all required tip atoms must be present. |
For example, {"A64":"ALL","A88":"BKBN","A96":"TIP"} selects different atom groups at three residues. Use a shorthand alone, not mixed with atom names such as "ALL,CA". Write atom lists with commas and no spaces, as in "CD2,CZ", and use ALL rather than * to select every heavy atom, because RFdiffusion3 reads each value exactly as written. TIP is not defined for alanine, glycine, or proline; use ALL, BKBN, or explicit atom names for those residues.
Common atom names:
N, CA, C, OCD1, CD2, CE1, CE2, CZCB, CG, CDNZ (Lys), OD1/OD2 (Asp), OE1/OE2 (Glu)Hotspot selection dramatically improves binder design success rates by ensuring the designed protein contacts critical interface residues.
Controls how the designed chain is positioned relative to the target:
com): Orient based on the target's center of mass. Use when no specific hotspots are known.Each design generates a structure file, normally CIF, named sequentially (for example, design_01.cif, design_02.cif). PDB output is also supported. The structures contain:
RFD3 outputs are backbone-only structures requiring sequence design:
The Data tab shows the metrics reported by RFdiffusion3 for each design, including chain breaks, backbone deviations, clashes, secondary-structure fractions, and radius of gyration when available. Metric availability depends on the design task. Fractions remain on their native 0 to 1 scale; distances and RMSDs are in angstroms.
The Files tab includes the complete per-design JSON report, with metrics, residue mappings, and inference metadata. Runs without recorded metrics, including older jobs, show Structure and Files without a Data tab. See the RFdiffusion3 output reference for metric definitions and conditions.
These metrics describe generated geometry and do not establish binding affinity or experimental success. Assess design quality with:
RFD3 excels at:
Scaffold enzyme active sites with atom-level control and ligand-aware protein design.
Design binders, enzymes and ligand pockets with all-atom flow matching
Generate protein sequences de novo, scaffold fixed motifs, or inpaint selected regions.
All-atom SE(3)-equivariant diffusion for unconditional generation, motif scaffolding, and binder design.
Design protein sequences around ligands, metals, and nucleotides for enzyme engineering and binding-site optimization.
Design all-atom protein binders with controlled target sites, lengths, and interaction parameters.
Generate novel drug-like molecules in protein binding pockets using AI-powered structure-based design.
Generate docking poses, drug-like molecules, and peptides inside protein binding pockets with PocketXMol.
Design proteins from a text description
Design antibody CDRs with Proteo-R1 reasoning, diffusion, and optional framework inpainting.