← All tools
Protein Design · Backbone design

RFdiffusion3

Generates protein backbone coordinates around proteins, small molecules, nucleic acids, and metals. Given geometric and other constraints, specifies backbone geometry. A useful first step in a protein design pipeline. Very well documented!

RFdiffusion3 is a practical first step in protein design workflows: It generates coordinates of the backbone atoms for a protein based on contraints (For example, to spacially deconflict with specific molecules). Its output can be fed into ProteinMPNN or LigandMPNN to generate an amino acid sequence.

No jobs before this; it will run immediately.

Ready
Choose an example to fill in its settings and input structures.
The path and file name of the JSON or YAML file where you have defined your inference constraints. In this form: Enter the JSON or YAML contents here. Use "input": "uploaded" to reference the structure upload. Presets may use a bundled bio-tools://rfd3/ reference or an official ../input_pdbs/ example path. Enter null with specification overrides for an override-only design.
Optional PDB or mmCIF for configurations with "input": "uploaded".
The path and file name of the JSON or YAML file where you have defined your inference constraints. In this form: Choose a JSON or YAML file. Use "input": "uploaded" to reference the structure upload, or keep a bundled bio-tools://rfd3/ reference from a preset.

InputSpecification fields

Official documentation ↗
Path to and file name of PDB/CIF. Required if you provide contig+length. In this form: Choose a PDB or mmCIF file. Presets include their input structures; a new upload replaces the included file.
(Can only pass a contig string.) Indexed motif specification, e.g., "A1-80,10,/0,B5-12". Residues whose position/order in the final sequence is specified.
(Can only pass a contig string or dictionary.) Unindexed motif components, the specified residues can be anywhere in the final sequence. See Unindexing Specifics for more information. Which residues should have their index be inferred by the model instead of prespecified
Total design length constraint; "min-max" or int for specified length.
Ligand(s) by chemical component name (from RSCB PDB) or index. These are ligands built in to the source mmCIF file, using the identifiers specified there
Optional args to CIF loader. See CIF parser options for more information. In this form: Enter a JSON object. The local cache_dir option is not available, and extra_fields is accepted only alongside an mmCIF structure because the PDB reader ignores it.
Extra metadata (e.g., logs). Current options include sampled_contig.
2=new (default), 1=legacy, Learn more about the legacy parsing system by looking at input_parsing.py.
Atoms with fixed coordinates. See the Select Fixed Atoms subsection for more information. Dictionary that indicated which atoms should be fixed (can use ALL, BKBN, or TIP for all atoms in the residue, backbone atoms only and tip atoms only In this form: Name a ligand by its chemical component code, such as NAI, rather than by chain and residue number. A ligand written into a chain that also holds polymer residues is moved onto a chain of its own while the structure is parsed, so a chain-and-number reference to it no longer matches.
Where sequence can change. Default is True - all input regions have fixed sequences. Contig string input specifies components to unfix the sequence for. Dictionary inputs are allowed but not recommended.
Selection of RASA (Relatively Accessible Surface Area) for buried, partially buried, and exposed conditioning, respectively. Only contig string and dictionary are acceptable inputs.
Selection of RASA (Relatively Accessible Surface Area) for buried, partially buried, and exposed conditioning, respectively. Only contig string and dictionary are acceptable inputs.
Selection of RASA (Relatively Accessible Surface Area) for buried, partially buried, and exposed conditioning, respectively. Only contig string and dictionary are acceptable inputs.
Atom-wise donor/acceptor flags. Atom-wise selection of hydrogen bond donors and acceptors, respectively. Only dictionary inputs allowed. See RFdiffusion3 — Nucleic acid binder design examples for an example.
Atom-wise donor/acceptor flags. Atom-wise selection of hydrogen bond donors and acceptors, respectively. Only dictionary inputs allowed. See RFdiffusion3 — Nucleic acid binder design examples for an example.
Atom-level or residue-level hotspots. Hotspots will typically be at most 4.5 Å to any heavy atom in the designed structure. Typically used for designing binders.
See Overview of Symmetry in RFD3. In this form: Enter SymmetryConfig as JSON, for example {"id": "C3", "is_symmetric_motif": true}.
[x,y,z] origin override to control COM (center of mass) placement of designed structure.
"com" or "hotspots". The center of mass of the diffused region will typically be within 5Å of the ORI token. Using hotspots will place the ORI token 10Å outward from the center of mass of the specified hotspots. Using com will place the token at the center of mass of the input structure.
Default None. If True/False, produces output structures with fewer/more loops.
Noise (Å) for partial diffusion, enables partial diffusion (sets the noise level.) Recommended values are 5.0-15.0 Å. See Partial Diffusion for more information.

Most useful CLI arguments

Official documentation ↗
number of batches to generate per input key (default: 1).
number of diffusion samples (designs) per batch (default: 8). If n_batches=1 and diffusion_batch_size=8 then 8 designs will be generated from the inference run.
diffusion timesteps for sampling (default: 200).
scales diffusion step size; higher → less diverse, more designable (default: 1.5).
JSON overrides for the per-example InputSpecification (default: {}). For example, you can run rfd3 design inputs=null specification.length=200 for a quick debug of creating a 200-length protein.
This setting allows you to change the beginning of the name of the output files from the name of the input JSON or YAML file to your own string (default: null). In this form: Leave blank for the default. Enter "" to omit the prefix. Use letters, digits, dots, underscores or hyphens.

Other CLI options

Official documentation ↗
This value (default: 0.6) influences the diversity of the designs from RFD3. A lower value increases designability but decreases diversity.
Allows the user to extract only a subset of the JSON keys provided in the inputs file (default: null).
Number of recycling iterations per diffusion step (default: null, uses the model checkpoint default of 2). Recycling allows the network to refine its predictions iteratively within each denoising step.
Change this value to symmetry (default: default) to turn on symmetry mode for the inference sampler. In this form: Automatic selects the sampler from the input specification.
The values specified (options are active_donor, active_acceptor, or ref_atomwise_rasa) are set to 0 for classifier-free guidance. Classifier-free guidance is how the diffusion model can steer the calculation towards a condition without training a separate classifier.
The maximum time to apply classifier-free guidance to the inference run (default: null).
Controls the influence of the classifier-free guidance adjustment (default: 1.5).
Specifies how to center the coordinates during the inference run to ensure that structures are alined around a specific point. Options include: all — (default) Uses the center of mass (COM) of all atoms motif — Uses the COM of the motif atoms with fixed coordinates diffuse — Uses the COM of all fixed coordinates that are not part of motif atoms
Translational noise scale for augmentation during inference (default: 1.0).
This parameter sets the scaling for the noise during inference (default 1.003). A smaller value will lead to less noise in your system leading to less diversity in the outputs.
Determines the ‘shape’ of the noise schedule (default: 7).
Controls when gamma_0 is used, if t>gamma_min, gamma_0 is used as the value of gamma, which influences the diversity of the designs from RFD3.
Controls the standard deviation of the Gaussian distribution that is used to ‘jitter’ the motif offset (default: 0.0, no jitter).

These local CLI options are not applicable to Bio Web and are omitted from its inputs:

  • out_dir: Bio Web creates a separate output directory for each run and collects its result files.
  • ckpt_path: Bio Web uses the installed RFD3 checkpoint configured on the server.
  • skip_existing: Each web run starts with a fresh output directory, so there are no existing designs to skip.
Ready