RFdiffusion3

Basics

• Uploads mount at /inputs; write any results to /outputs for download.
• The service sets runtime paths for you; you do not need to include them.
• Your arguments are passed after rfd3 design inside the container.
• SubSeq controls structure output type with subseq_output_format=cif|pdb|cif.gz. The default is plain .cif; choose pdb for PDB structures or cif.gz to keep RFdiffusion3's upstream compressed outputs.
• To request multiple designs, set n_batches and diffusion_batch_size (total designs = n_batches × diffusion_batch_size). Prefer larger diffusion_batch_size with fewer n_batches when memory allows. Do not use inference.num_designs for RFdiffusion3.
• RFdiffusion3 supports symmetry via inference_sampler.kind=symmetry; SubSeq supports cyclic, tetrahedral (T), octahedral (O), and icosahedral (I) frame generation. Dihedral support exists but is buggy, so guided mode omits and rejects it. In symmetry specs, length is the asymmetric-unit length, not the full complex length.

Example 1: unconditional monomer

Design a de novo monomer of length 120–130 residues with no structural input.

1. Create or upload a JSON file under /inputs, for example /inputs/uncond_monomer.json. The New Job page may instead prefill this as an inline aux file at /aux/uncond_monomer.json; use the path shown in the arguments box.

{
  "uncond_monomer": {
    "length": "120-130"
  }
}

2. Submit a job with the following arguments (New Job → RFdiffusion3):

inputs=/inputs/uncond_monomer.json
n_batches=1
diffusion_batch_size=8
low_memory_mode=False

RFdiffusion3 will sample one or more all-atom monomers in the requested length range and write the selected structure output type under /outputs.

Example 2: simple protein binder

Design a small binder against a protein target using a minimal hotspot-based configuration.

1. Upload your target structure to /inputs, for example /inputs/target.pdb or /inputs/target.cif.
2. Create a JSON spec next to it, for example /inputs/protein_binder_simple.json:

{
  "simple_binder": {
    "dialect": 2,
    "input": "/inputs/target.pdb",
    "contig": "50-80,/0,A40-120",
    "infer_ori_strategy": "hotspots",
    "select_hotspots": {
      "A64": "CG,CD1",
      "A88": "CG,CD1"
    }
  }
}

3. Submit a job with:

inputs=/inputs/protein_binder_simple.json

In this setup:

• input is the target (antigen) structure you uploaded.
• contig requests a new binder segment of length 50–80 residues and treats residues A40–A120 of the target as the region of interest.
• infer_ori_strategy=hotspots and select_hotspots tell RFD3 to position the binder using atoms on residues A64 and A88 as hotspots.

Adjust the target residue window, hotspot residues/atoms, and binder length range to match your system.

Example 3: symmetric unconditional design

Design a small cyclic oligomer using the Foundry symmetry sampler.

1. Create a JSON file under /inputs, for example /inputs/sym_c3.json:

{
  "sym_c3": {
    "length": "60",
    "is_non_loopy": true,
    "symmetry": {
      "id": "C3"
    }
  }
}

2. Submit with:

inputs=/inputs/sym_c3.json
inference_sampler.kind=symmetry
n_batches=1
diffusion_batch_size=1
inference_sampler.num_timesteps=200
low_memory_mode=True

SubSeq supports cyclic, T, O, and I symmetry in guided mode. Cyclic symmetry uses a separate order field; total protein residues are approximately ASU length times symmetry order: C<n>=n, T=12, O=24, I=60. For example, matched 480-residue designs can use ASU lengths T:40, O:20, and I:8. High-order symmetry can still produce clashing samples, so use 100-200 timesteps for inspection and filter outputs.

Symmetric motif scaffolding expects pre-symmetrized motifs centered at the origin. The motif copies must match the requested multiplicity, e.g. 12/24/60 copies for T/O/I. Asymmetric motif inputs and classifier-free-guidance conditioning should be treated as unsupported in symmetry mode.

Notes

• Multiple design specs can be placed in a single JSON file; each key is treated as an independent configuration.
• Structure outputs are named like <json_basename>_<spec_key>_<batch>_model_<i>.<ext> with matching .json metadata. The extension is selected by subseq_output_format: .cif, .pdb, or .cif.gz.
• PDB is a legacy fixed-column format with chain, residue, and atom-count limits; use cif or cif.gz for large or many-chain designs.
• Standard Foundry examples for nucleic acid binders, small-molecule binders, and enzymes can be adapted by pointing paths into /inputs and keeping outputs in /outputs.
• For full input-spec details, see the RFdiffusion3 documentation in the Foundry repo: models/rfd3/docs/input.md.

Submit

Launch from New Job → RFdiffusion3. Keep uploaded input data under /inputs; small inline or generated config files may be placed under /aux. Outputs will appear under /outputs on completion.