TODO: COMPLETE

Setting Up RFdiffusion with Docker

RFdiffusion is an open‑source diffusion model for protein structure generation and design. It supports:

• Unconditional monomer generation
• Motif scaffolding
• Symmetric oligomer design (Cyclic, Dihedral, Tetrahedral)
• Binder (PPI) design with hotspot control
• Partial diffusion & design diversification

This guide walks through:

1. Prerequisites
2. Cloning and downloading model weights
3. Writing a Dockerfile
4. Building the Docker image
5. Running RFdiffusion via Docker
6. Example: Motif scaffolding
7. Tips & Best Practices

1. Prerequisites

• Docker (v20.10+) with GPU support (optional but recommended)
• nvidia-docker2 if using NVIDIA GPUs
• ~10–20 GB disk space for model weights
• Familiarity with the command line and basic protein files (FASTA/PDB)

2. Clone & Download Model Weights

# Clone the RFdiffusion repo
git clone https://github.com/RosettaCommons/RFdiffusion.git
cd RFdiffusion

# Create a folder for weights
mkdir -p models
cd models

# Download core checkpoints
wget http://files.ipd.uw.edu/pub/RFdiffusion/6f5902ac237024bdd0c176cb93063dc4/Base_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/e29311f6f1bf1af907f9ef9f44b8328b/Complex_base_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/60f09a193fb5e5ccdc4980417708dbab/Complex_Fold_base_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/74f51cfb8b440f50d70878e05361d8f0/InpaintSeq_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/76d00716416567174cdb7ca96e208296/InpaintSeq_Fold_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/5532d2e1f3a4738decd58b19d633b3c3/ActiveSite_ckpt.pt
wget http://files.ipd.uw.edu/pub/RFdiffusion/12fc204edeae5b57713c5ad7dcb97d39/Base_epoch8_ckpt.pt

# Optional beta binder model
wget http://files.ipd.uw.edu/pub/RFdiffusion/f572d396fae9206628714fb2ce00f72e/Complex_beta_ckpt.pt

cd ../

3. Dockerfile

Create docker/Dockerfile with the following contents:

# Base image with Python and CUDA
FROM nvidia/cuda:11.6.2-cudnn8-runtime-ubuntu20.04

# Install essentials
RUN apt-get update && apt-get install -y --no-install-recommends \
    git python3 python3-pip python3-venv && \
    rm -rf /var/lib/apt/lists/*

# Copy RFdiffusion code and weights
WORKDIR /app
COPY . /app

# Install RFdiffusion and dependencies
RUN python3 -m venv venv && \
    . venv/bin/activate && \
    pip install --upgrade pip && \
    pip install -e . && \
    pip install hydra-core==1.1.1

# Default entrypoint
ENTRYPOINT ["/app/venv/bin/python3", "scripts/run_inference.py"]

4. Build the Docker Image

From the repository root:

docker build -f docker/Dockerfile -t rfdiffusion:latest .

Use --build-arg if you need to pass custom CUDA or Python versions.

5. Running RFdiffusion via Docker

Mount your weights, inputs, and outputs into the container.

# Prepare directories
dkdir -p $HOME/rfd_models $HOME/rfd_inputs $HOME/rfd_outputs

# Copy downloaded weights into $HOME/rfd_models
# Place any PDB/FASTA inputs in $HOME/rfd_inputs

# Run an unconditional monomer design of length 150 (10 designs)

docker run --rm --gpus all \
  -v $HOME/rfd_models:/app/models \
  -v $HOME/rfd_inputs:/app/inputs \
  -v $HOME/rfd_outputs:/app/outputs \
  rfdiffusion:latest \
  'contigmap.contigs=[150-150]' \
  inference.output_prefix=/app/outputs/monomers \
  inference.num_designs=10

• contigmap.contigs=[L-L]: length range for chain
• inference.output_prefix: output directory + filename prefix
• inference.num_designs: how many designs to sample

6. Example: Motif Scaffolding

Assume you have a motif in inputs/motif.pdb, residues A30–A45.

docker run --rm --gpus all \
  -v $HOME/rfd_models:/app/models \
  -v $HOME/rfd_inputs:/app/inputs \
  -v $HOME/rfd_outputs:/app/outputs \
  rfdiffusion:latest \
  inference.input_pdb=/app/inputs/motif.pdb \
  'contigmap.contigs=[20-30/A30-45/20-30]' \
  inference.output_prefix=/app/outputs/scaffolded \
  inference.num_designs=5

This builds 20–30 AA on both sides of your motif, sampling varied loop lengths.

7. Tips & Best Practices

• Cache IGSO3: the first run computes geometric caches—subsequent runs are faster.
• Symmetric designs: use --config-name symmetry and inference.symmetry=c4|d2|tetrahedral.
• Hotspots for binders: set ppi.hotspot_res=[A45,A47,A50] for targeted PPI.
• Partial diffusion: add diffuser.partial_T=<steps> to explore around a seed structure.
• Checkpoint overrides: e.g. use inference.ckpt_override_path=models/ActiveSite_ckpt.pt for small motifs.
• Output artifacts: look in /traj/ for per‑step PDBs and .trb for metadata.

Happy protein designing with RFdiffusion!