Rfdiffusion

RFdiffusion: De Novo Protein Backbone Generation

Overview

RFdiffusion generates protein backbones via a denoising diffusion process. It starts from random noise and iteratively refines toward a plausible protein structure. It is the backbone generation half of the modern protein design pipeline:

RFdiffusion (backbone) → ProteinMPNN (sequence) → AlphaFold2 (validation)

Core design modes:

• Unconditional generation — design monomers of specified length
• Motif scaffolding — build a new protein around a fixed functional motif
• Binder design — create proteins that bind a target surface
• Symmetric oligomers — cyclic (Cn), dihedral (Dn), tetrahedral, octahedral, icosahedral
• Partial diffusion — diversify an existing backbone (keep fold, vary details)
• Active site scaffolding — design enzymes around catalytic residues
• Cyclic peptides — macrocyclic monomers and binders
• Sequence inpainting — redesign selected residues in context

Installation

git clone https://github.com/RosettaCommons/RFdiffusion.git
cd RFdiffusion

# Create environment
conda env create -f env/SE3nv.yml
conda activate SE3nv

# Install SE3-Transformer
cd env/SE3Transformer
pip install --no-cache-dir -r requirements.txt
python setup.py install
cd ../..

# Install RFdiffusion
pip install -e .

Download model weights

mkdir models && cd models
# Base (unconditional + motif scaffolding)
wget http://files.ipd.uw.edu/pub/RFdiffusion/6f5902ac237024bdd0c176cb93063dc4/Base_ckpt.pt
# Binder / PPI design
wget http://files.ipd.uw.edu/pub/RFdiffusion/e29311f6f1bf1af907f9ef9f44b8328b/Complex_base_ckpt.pt
# Active site scaffolding
wget http://files.ipd.uw.edu/pub/RFdiffusion/5532d2e1f3a4738decd58b19d633b3c3/ActiveSite_ckpt.pt
# Sequence inpainting
wget http://files.ipd.uw.edu/pub/RFdiffusion/74f51cfb8b440f50d70878e05361d8f0/InpaintSeq_ckpt.pt
# Epoch 8 (metal binding, symmetric oligomers)
wget http://files.ipd.uw.edu/pub/RFdiffusion/12fc204edeae5b57713c5ad7dcb97d39/Base_epoch8_ckpt.pt

Requirements: NVIDIA GPU with CUDA 11.1+, ~8 GB VRAM recommended.

Model Checkpoints

Checkpoint	Use case
Base_ckpt.pt	Unconditional monomers, motif scaffolding
Complex_base_ckpt.pt	Binder / PPI design
Complex_Fold_base_ckpt.pt	Binder design with fold conditioning
InpaintSeq_ckpt.pt	Sequence inpainting / redesign
InpaintSeq_Fold_ckpt.pt	Sequence inpainting with fold conditioning
ActiveSite_ckpt.pt	Enzyme active site scaffolding
Base_epoch8_ckpt.pt	Metal binding, symmetric oligomers

Contig Syntax (Critical)

The contig string tells RFdiffusion what to design and what to keep fixed. This is the most important concept.

• Bare numbers = residues to design (variable length range)
• Chain-prefixed residues = residues to keep from the input PDB
• / separates segments within a chain
• /0 separates different chains

[150-150]                    ← unconditional: design exactly 150 residues
[100-200]                    ← unconditional: random length 100–200
[10-40/A10-25/30-40]         ← motif scaffolding: design, keep A10-25, design
[B1-100/0 70-100]            ← binder: keep target B1-100, design 70-100aa binder
[480-480]                    ← symmetric: total residues for C6 (480/6 = 80 per subunit)

Core Workflows

1. Unconditional monomer

./scripts/run_inference.py \
  'contigmap.contigs=[100-200]' \
  inference.output_prefix=outputs/monomer \
  inference.num_designs=10

2. Motif scaffolding

Build a new protein around residues A163-181 from a known structure:

./scripts/run_inference.py \
  'contigmap.contigs=[10-40/A163-181/10-40]' \
  inference.input_pdb=input_pdbs/5TPN.pdb \
  inference.output_prefix=outputs/motif_scaffold \
  inference.num_designs=10

3. Binder design

Design a protein that binds to target chain A (keeping residues A1-150), with hotspot guidance:

./scripts/run_inference.py \
  'contigmap.contigs=[A1-150/0 70-100]' \
  inference.input_pdb=target.pdb \
  'ppi.hotspot_res=[A59,A83,A91]' \
  inference.output_prefix=outputs/binder \
  inference.num_designs=1000 \
  denoiser.noise_scale_ca=0 \
  denoiser.noise_scale_frame=0

Binder design tips:

• Generate 1,000–10,000 backbones
• Select 3–6 hotspot residues on the target surface
• Truncate large targets to ~200 residues around the interface
• Set noise scales to 0 for better quality
• Filter downstream with AF2: keep designs where pAE_interaction < 10

4. Partial diffusion (backbone diversification)

Add controlled noise to an existing structure, then denoise to generate variants:

./scripts/run_inference.py \
  'contigmap.contigs=[79-79]' \
  inference.input_pdb=input_pdbs/2KL8.pdb \
  diffuser.partial_T=10 \
  inference.output_prefix=outputs/partial_diff \
  inference.num_designs=10

Important: Contig length must exactly match the input PDB length for partial diffusion.

5. Symmetric oligomers

# C6 cyclic symmetry (80 residues per subunit × 6 = 480 total)
./scripts/run_inference.py --config-name=symmetry \
  inference.symmetry="C6" \
  'contigmap.contigs=[480-480]' \
  'potentials.guiding_potentials=["type:olig_contacts,weight_intra:1,weight_inter:0.1"]' \
  potentials.olig_intra_all=True \
  potentials.olig_inter_all=True \
  potentials.guide_scale=2.0 \
  potentials.guide_decay="quadratic" \
  inference.num_designs=10

Supported symmetries: C2–C6, D2–D6, tetrahedral, octahedral, icosahedral.

6. Active site scaffolding

./scripts/run_inference.py \
  'contigmap.contigs=[10-100/A1083/10-100/A1051/10-100/A1180/10-100]' \
  inference.input_pdb=input_pdbs/5an7.pdb \
  'potentials.guiding_potentials=["substrate_contacts:s=1,r_0=8,rep_r_0=5.0,rep_s=2,rep_r_min=1"]' \
  potentials.substrate=LLK \
  potentials.guide_scale=1 \
  inference.ckpt_override_path=models/ActiveSite_ckpt.pt \
  inference.num_designs=10

7. Cyclic peptides

./scripts/run_inference.py \
  'contigmap.contigs=[12-18]' \
  inference.cyclic=True \
  inference.cyc_chains=a \
  inference.output_prefix=outputs/cyclic_peptide \
  inference.num_designs=10

Key Parameters

Parameter	Default	Description
inference.num_designs	10	Number of backbones to generate
inference.input_pdb	null	Input PDB (null for unconditional)
inference.output_prefix	samples/design	Output path prefix
inference.ckpt_override_path	—	Override model checkpoint
diffuser.T	50	Diffusion timesteps (200 for max quality, 20 for speed)
diffuser.partial_T	null	Timesteps for partial diffusion
denoiser.noise_scale_ca	varies	Cα noise (0 for zero-noise inference)
denoiser.noise_scale_frame	varies	Frame noise (0 for zero-noise inference)
ppi.hotspot_res	null	Target residues for binding [A30,A33,A34]
potentials.guide_scale	1.0	Strength of guiding potentials
potentials.guide_decay	null	"quadratic" or linear decay
inference.symmetry	null	Symmetry type (C3, D2, tetrahedral, etc.)
inference.cyclic	false	Enable cyclization

Output Format

outputs/
├── design_0.pdb     ← backbone PDB (Cα, N, C, O atoms only — no sidechains)
├── design_0.trb     ← metadata pickle (config, contig mapping, scores)
├── design_1.pdb
└── ...

The output PDBs are backbone-only — they have no sequence or sidechains. The next step is always ProteinMPNN for sequence design.

The Full Design Pipeline

1. RFdiffusion    → backbone PDB (no sequence)
2. ProteinMPNN    → sequences for the backbone (FASTA)
3. AlphaFold2     → validate: does the designed sequence fold to the intended structure?

For binder design, filter AF2 predictions by pAE_interaction < 10. For monomers, check that AF2's predicted structure matches the RFdiffusion backbone (low RMSD).

Performance

• ~20 diffusion steps produces equivalent quality to 200 steps (10× speedup)
• Runtime scales O(N²) with residue count
• First run caches IGSO3 calculations (~30 min); subsequent runs are faster
• Generate many designs (100–10,000) and filter downstream — RFdiffusion is fast, validation is the bottleneck

Resources

• GitHub: https://github.com/RosettaCommons/RFdiffusion
• Paper: Watson et al., Nature 2023 — https://doi.org/10.1038/s41586-023-06415-8
• Binder design protocol: https://github.com/nrbennet/dl_binder_design
• Docker: rosettacommons/rfdiffusion on Docker Hub

References

• references/design-modes.md — all design modes with full command examples, contig syntax reference, guiding potentials, fold conditioning