Rfdiffusion

RFdiffusion

Use RFdiffusion for the backbone-generation stage of a protein-design workflow. Treat its output as a structural proposal, not as evidence that a sequence folds or binds. Design sequences afterward with proteinmpnn (or ligandmpnn when non-protein atoms must be visible), then validate both the isolated design and the target–design complex with an independent predictor such as boltz, chai1, or alphafold2.

The adapted recipe is MIT-licensed. The upstream RFdiffusion code and the model weights referenced by its README are BSD-3-Clause; record the exact upstream commit and checkpoint digest used by each run.

Preflight: freeze the design contract

Before inference:

1. Preserve the original chain IDs and residue numbers. Write an explicit map if the input is renumbered, cropped, or has insertion codes.
2. Audit missing residues/atoms, alternate locations, non-standard residues, and biological assembly choice. Do not silently remove cofactors or chains.
3. For binder design, confirm proposed hotspot residues are surface-accessible and belong to the intended target chain. A typical pilot uses 3–6 spatially coherent hotspots, but the biological interface determines the final set.
4. Freeze the target PDB digest, RFdiffusion commit, checkpoint digest, contig string, hotspot list, seed policy, number of designs, and output prefix in a machine-readable manifest.
5. Keep target-only coordinates separate from any withheld reference binder. Do not leak a reference complex into generation or validation.

Do not trim a target merely to make inference cheaper unless the retained construct is biologically justified and the residue map is preserved.

Install and invoke the official runner

RFdiffusion is a source repository rather than an OpenAI4S sidecar. Prepare a pinned GPU environment, clone a fixed revision of the official repository, install it there, and download the documented checkpoint. Verify every download before starting a campaign. Follow the upstream CUDA/PyTorch/DGL compatibility instructions for the selected revision; do not improvise a version matrix from this recipe.

Do not install RFdiffusion into OpenAI4S's shared struct environment. That environment intentionally targets portable Python 3.13 with a CPU PyTorch build, whereas upstream RFdiffusion publishes a Python 3.9, CUDA-specific PyTorch/DGL stack. Use a pinned dedicated conda environment or the official RFdiffusion Docker image pinned by digest. openai4s setup --profile full creates the shared struct environment; it does not provision RFdiffusion, GPU drivers, model weights, or a container image.

Run scripts from the RFdiffusion repository root. The official inference entry point is scripts/run_inference.py:

./scripts/run_inference.py \
  inference.input_pdb=target.pdb \
  'contigmap.contigs=[A1-150/0 70-100]' \
  'ppi.hotspot_res=[A45,A67,A89]' \
  inference.output_prefix=results/backbones/design \
  inference.num_designs=48

This example fixes target chain A residues 1–150, inserts a chain break, and generates a 70–100-residue binder chain. The output is backbone-only: designed residues are represented as glycine by default. That is expected and is the handoff point to inverse folding.

Quote the entire Hydra list override, including brackets. Shell tokenization otherwise splits a contig containing spaces before Hydra sees it. Inside the contig:

• /0 means a chain break: slash, zero, then a space.
• / without the zero joins segments in the same output chain.
• A10-30 selects fixed input coordinates; 20-40 requests a generated segment of variable length.
• The order of contig segments controls the output topology. Never infer output residue identity from the PDB alone; read the mapping in the .trb.
• Hotspots use input chain plus residue number, with no spaces between list items. Quote that whole list override too.

Common malformed variants are an unquoted contig with spaces, a comma between contig segments, a missing /0 at a desired chain break, or hotspot numbers without chain IDs.