Rfdiffusion3

RFdiffusion3

Defaults

Use these paths unless the user gives different ones:

textCopy code
env_prefix=/path/to/conda-envs/foundry-py312
checkpoint_dir=/path/to/foundry/checkpoints
workspace=/path/to/foundry/workspace
run_root=/path/to/foundry/workspace/rfd3_runs
conda_pkgs_dir=/path/to/conda-pkgs
pip_cache_dir=/path/to/pip-cache

Prefer a dedicated Python 3.12 Foundry environment unless the user provides a different supported installation.

First Step

For any install/debug/run request, inspect the live state before making claims:

bashCopy code
conda info --envs
nvidia-smi
test -d /path/to/conda-envs/foundry-py312
test -d /path/to/foundry/checkpoints

When running commands:

bashCopy code
source ${CONDA_BASE:-$HOME/miniconda3}/etc/profile.d/conda.sh
conda activate /path/to/conda-envs/foundry-py312
export FOUNDRY_CHECKPOINT_DIRS=/path/to/foundry/checkpoints
mkdir -p /path/to/foundry/workspace/rfd3_runs

Validate

Use the bundled script before expensive GPU work:

bashCopy code
python /path/to/foundry/workspace/skills/rfdiffusion3/scripts/check_foundry_env.py \
  --checkpoint-dir /path/to/foundry/checkpoints

For official demo smoke tests:

bashCopy code
python /path/to/foundry/workspace/skills/rfdiffusion3/scripts/prepare_rfd3_demo.py \
  --workdir /path/to/foundry/workspace/rfd3_demo \
  --checkpoint-dir /path/to/foundry/checkpoints
cd /path/to/foundry/workspace/rfd3_demo/verify_installation
FOUNDRY_CHECKPOINT_DIRS=/path/to/foundry/checkpoints rfd3 design out_dir=demo_output inputs=demo_compat.json skip_existing=False prevalidate_inputs=True diffusion_batch_size=1 n_batches=1 low_memory_mode=True dump_trajectories=False

Use demo_compat.json for this machine's smoke test; the upstream demo file included an unsupported allow_ligand_on_existing_chain field for the installed package version.

Run RFD3

Minimum command:

bashCopy code
FOUNDRY_CHECKPOINT_DIRS=/path/to/foundry/checkpoints rfd3 design out_dir=<out_dir> inputs=<input.json>

First-pass low-memory settings:

bashCopy code
rfd3 design out_dir=<out_dir> inputs=<input.json> prevalidate_inputs=True skip_existing=False diffusion_batch_size=1 n_batches=1 low_memory_mode=True dump_trajectories=False

Before writing an input file, inspect the user's PDB/CIF for chain IDs, residue numbers, and ligand residue names. Prefer absolute input paths or paths relative to the run directory.

References

Read only the reference needed for the task:

• references/foundry-rfd3-workflow.md: installation, local path policy, checkpoints, and environment validation.
• references/rfd3-usage-workflows.md: official tutorial distilled workflows, input fields, templates for PPI, NA, small molecule, enzyme scaffold, and partial diffusion.
• references/foundry-postprocess-workflow.md: RFD3 outputs into ProteinMPNN/LigandMPNN, RF3 fold/review, Python API in-memory pipeline, and RMSD/QC.

User-facing tutorials live outside the skill:

• /path/to/foundry/workspace/RFdiffusion3_Foundry_安装教程.md
• /path/to/foundry/workspace/RFdiffusion3_使用教程.md
• /path/to/foundry/workspace/Foundry_RFD3_MPNN_RF3_后处理教程.md

Bundled scripts:

• scripts/check_foundry_env.py: validate Python, torch/CUDA, Foundry CLIs, and checkpoints.
• scripts/prepare_rfd3_demo.py: download official demo inputs and create demo_compat.json.
• scripts/prepare_rfd3_tutorial_assets.py: download official tutorial/example assets by profile when available.
• scripts/make_foundry_postprocess_case.py: generate an MPNN JSON config plus runnable MPNN/RF3 shell scripts for a chosen RFD3 structure.
• scripts/foundry_python_api_pipeline_demo.py: template for notebook-style in-memory RFD3 -> MPNN -> RF3 -> RMSD workflow.
• scripts/qc_foundry_backbone_rmsd.py: file-based backbone RMSD/QC between an RFD3 design and RF3 refolded structure.