lamm-mit/skills/rfdiffusion

RFdiffusion Backbone Generation

Diffusion-based de novo protein backbone generation. Use for: binder scaffolds targeting specific hotspot residues, novel fold generation, motif scaffolding, and symmetric oligomer design.

Requirements

• Python 3.9+
• 16–24 GB GPU VRAM (A100 recommended for large designs)
• ~2 GB disk for model weights

Installation

# Via SE3-Diffusion environment (recommended)
conda create -n rfdiffusion python=3.9
conda activate rfdiffusion
pip install rfdiffusion

# Or from source
git clone https://github.com/RosettaCommons/RFdiffusion
cd RFdiffusion
pip install -e .

# Download model weights
wget http://files.ipd.uw.edu/pub/RFdiffusion/rfdiffusion_weights.tar.gz
tar -xzf rfdiffusion_weights.tar.gz

Core Use Cases

1. Binder Design (Most Common)

Design a protein that contacts specific hotspot residues on a target:

python3 scripts/run_inference.py \
    inference.output_prefix=outputs/binder \
    inference.input_pdb=target.pdb \
    'ppi.hotspot_res=[A45,A67,A89]' \
    contigmap.contigs='[A1-150/0 70-100]' \
    inference.num_designs=50

contigmap.contigs format:

• A1-150 = use chain A residues 1-150 (target, fixed)
• /0 = chain break
• 70-100 = design 70–100 residue binder (length range)

2. Motif Scaffolding

Scaffold a functional motif (active site, binding epitope) into a new protein:

python3 scripts/run_inference.py \
    inference.output_prefix=outputs/scaffold \
    inference.input_pdb=motif.pdb \
    'contigmap.contigs=[10-40/A25-35/10-40]' \
    inference.num_designs=100

10-40/A25-35/10-40 = 10–40 designed residues + motif A25-35 (fixed) + 10–40 designed residues

3. Unconditional Generation

Generate novel protein folds:

python3 scripts/run_inference.py \
    inference.output_prefix=outputs/novel \
    'contigmap.contigs=[100-150]' \
    inference.num_designs=100

4. Symmetric Oligomers

python3 scripts/run_inference.py \
    inference.output_prefix=outputs/trimer \
    'contigmap.contigs=[100]' \
    symmetry=C3 \
    inference.num_designs=50

Python API

from rfdiffusion.inference.utils import preprocess_pdb
from rfdiffusion import RFDiffusion

model = RFDiffusion.from_pretrained("rfdiffusion_weights/")

results = model.design_binder(
    target_pdb="target.pdb",
    hotspot_residues=["A45", "A67", "A89"],
    binder_length=(70, 100),
    num_designs=50,
    output_dir="outputs/"
)

Key Parameters

| Parameter | Description | Example | |-----------|-------------|---------| | inference.num_designs | Number of backbones to generate | 50–500 | | diffuser.T | Diffusion steps (quality vs speed) | 50 (fast) – 200 (quality) | | ppi.hotspot_res | Target residues binder must contact | [A45,A67,A102] | | contigmap.contigs | Design specification string | [A1-200/0 50-80] | | inference.ckpt_override_path | Use specific model checkpoint | — |

Partial Diffusion (Diversify Existing Design)

# Partially noise and re-denoise an existing backbone
python3 scripts/run_inference.py \
    inference.output_prefix=outputs/diversified \
    inference.input_pdb=good_design.pdb \
    diffuser.partial_T=20 \
    'contigmap.contigs=[A1-80]' \
    inference.num_designs=20

Downstream: Add Sequences with ProteinMPNN

RFdiffusion outputs backbone-only PDBs (no sequence). Use ProteinMPNN or LigandMPNN to design sequences:

# After RFdiffusion
python3 run_ProteinMPNN/protein_mpnn_run.py \
    --pdb_path outputs/binder_0.pdb \
    --out_folder mpnn_output/ \
    --num_seq_per_target 8 \
    --sampling_temp 0.1

Output Files

| File | Contents | |------|----------| | {prefix}_0.pdb | Generated backbone (N, CA, C, O only) | | {prefix}_0.trb | Trajectory info, contig mapping | | traj/{prefix}_0_pX0_traj.pdb | Diffusion trajectory |

Common Issues

| Problem | Fix | |---------|-----| | Binder doesn't contact hotspots | Increase ppi.hotspot_res weights or reduce binder length range | | All designs look similar | Increase diffuser.T or use partial diffusion | | OOM error | Reduce design length or use inference.chunk_size=64 | | Poor downstream AF2 ipTM | Try partial diffusion to diversify; check hotspot accessibility |