Rfdiffusion

RFdiffusion Backbone Generation

Prerequisites

How to run

First time? See Installation Guide to set up Modal and biomodals.

Option 1: Modal (recommended)

# Clone biomodals
git clone https://github.com/hgbrian/biomodals && cd biomodals

# Basic binder design
modal run modal_rfdiffusion.py \
  --pdb target.pdb \
  --contigs "A1-150/0 70-100" \
  --hotspot "A45,A67,A89" \
  --num-designs 100

# With custom GPU/timeout
GPU=A100 TIMEOUT=60 modal run modal_rfdiffusion.py \
  --pdb target.pdb \
  --contigs "A1-150/0 70-100" \
  --num-designs 100

GPU: A10G (24GB) | Timeout: 30min default

Option 2: Local installation

# Clone and install
git clone https://github.com/RosettaCommons/RFdiffusion.git
cd RFdiffusion && pip install -e .

# Download weights
wget http://files.ipd.uw.edu/pub/RFdiffusion/models/Complex_base_ckpt.pt

# Run inference
python run_inference.py \
  inference.input_pdb=target.pdb \
  contigmap.contigs=[A1-150/0 70-100] \
  ppi.hotspot_res=[A45,A67,A89] \
  inference.num_designs=100

Config Schema (Hydra)

Contigmap Syntax

# De novo single chain (50-100 residues)
contigmap.contigs=[50-100]

# Binder + target (A = target chain, fixed with /0)
contigmap.contigs=[A1-150/0 70-100]

# Motif scaffolding (preserve residues, /0 = fixed)
contigmap.contigs=[20-40/0 A10-30/0 20-40]

# Multi-chain binder
contigmap.contigs=[A1-100/0 B1-100/0 60-80]

# Variable length ranges
contigmap.contigs=[A1-150/0 50-100]  # Binder 50-100 AA

Hotspot Specification

# Residues for interface (chain + resnum, no spaces)
ppi.hotspot_res=[A45,A67,A89]

Common mistakes

Contig Syntax

✅ Correct:

contigmap.contigs=[A1-150/0 70-100]  # Target fixed (/0), binder variable

❌ Wrong:

contigmap.contigs=[A1-150 70-100]    # Missing /0 - target will move!
contigmap.contigs="A1-150/0 70-100"  # Quotes break parsing
contigmap.contigs=[A1-150/0, 70-100] # Comma breaks parsing

Hotspot Residues

✅ Correct:

ppi.hotspot_res=[A45,A67,A89]        # Chain letter + residue number

❌ Wrong:

ppi.hotspot_res=[45,67,89]           # Missing chain letter
ppi.hotspot_res=[A45, A67, A89]      # Spaces break parsing
ppi.hotspot_res="A45,A67,A89"        # Quotes break parsing

Complete Parameter Reference

Core Parameters

Advanced Parameters

Symmetry Parameters

Fold Conditioning

Model Checkpoints

Common workflows

Binder Design

1. Prepare target PDB (trim to binding region + 10A buffer)
2. Identify 3-6 hotspot residues (exposed, conserved)
3. Generate 100-500 backbones
4. Pass to proteinmpnn for sequence design

Motif Scaffolding

1. Extract motif coordinates
2. Use /0 to fix motif in contigmap
3. Generate surrounding scaffold
4. Validate motif preservation (RMSD < 1.5A)

Symmetric Oligomers

# C3 symmetric trimer
python run_inference.py \
  symmetry.symmetry=C3 \
  contigmap.contigs=[100-150] \
  inference.num_designs=50

# D2 symmetric tetramer
python run_inference.py \
  symmetry.symmetry=D2 \
  contigmap.contigs=[80-120] \
  symmetry.radius=25

# Supported symmetries: C2, C3, C4, C5, C6, D2, D3, D4, tetrahedral, octahedral

Partial Diffusion (Refinement)

# Start from existing structure, diffuse from timestep 10
python run_inference.py \
  inference.input_pdb=initial.pdb \
  diffuser.partial_T=10 \
  contigmap.contigs=[A1-100]

Output format

output/
├── output_0.pdb       # Generated backbone
├── output_1.pdb
├── ...
└── output_99.pdb

Each PDB contains polyalanine backbone - use proteinmpnn for sequence.

Sample output

Successful run

$ python run_inference.py inference.input_pdb=target.pdb contigmap.contigs=[A1-150/0 70-100] inference.num_designs=100
[INFO] Loading model from Complex_base_ckpt.pt
[INFO] Generating design 1/100...
[INFO] Generating design 50/100...
[INFO] Generating design 100/100...
[INFO] Saved 100 designs to output/

Generated:
output/output_0.pdb (85 residues)
output/output_1.pdb (92 residues)
...

What good output looks like:

• File size: 3-8 KB per PDB (backbone only)
• Residue count within specified range
• Secondary structure visible in PyMOL (helices/sheets, not random coil)

Decision tree

Should I use RFdiffusion?
│
├─ Need to generate protein backbone?
│  ├─ Yes → Continue below
│  └─ No, already have backbone → Use ProteinMPNN
│
├─ What type of design?
│  ├─ Binder for protein target → RFdiffusion ✓
│  ├─ De novo monomer → RFdiffusion ✓
│  ├─ Motif scaffolding → RFdiffusion ✓
│  └─ Symmetric assembly → RFdiffusion ✓
│
└─ Priority?
   ├─ Need highest success rate → Consider BindCraft
   ├─ Need diversity/exploration → RFdiffusion ✓
   └─ Need all-atom precision → Consider BoltzGen

Typical performance

Expected downstream yield: ~10-15% of backbones pass full QC after sequence design + validation.

Verify

ls output/*.pdb | wc -l  # Should match num_designs

Troubleshooting

Designs lack secondary structure: Decrease noise_scale to 0.5-0.8 Binder not contacting hotspots: Verify residue numbering, increase num_designs OOM errors: Reduce batch size or use A100 GPU Slow generation: Reduce diffuser.T to 25-35

Error interpretation

Next: proteinmpnn for sequence design → structure prediction for validation → protein-design-qc for filtering.

Templates and Demo

• Starter flags: templates/rfdiffusion/binder_flags.txt
• Planning context: examples/minimal-binder-campaign/target-brief.md

Inputs

• A prepared target structure plus contig and hotspot definitions for the design task.
• A clear generation objective such as binder design, motif scaffolding, or symmetry.
• Sampling settings such as design count, checkpoint choice, and noise scale.

Outputs

• Generated backbone structures that satisfy the requested geometric constraints.
• A candidate backbone library for sequence design with proteinmpnn or solublempnn.
• Run metadata needed to reproduce successful generation settings.

Next Step

Hand the generated backbones to proteinmpnn for sequence design, then validate with structure prediction and protein-design-qc.