RFdiffusion3 on the cluster

RFdiffusion3 (RFD3) is an all-atom biomolecular diffusion design model. It supports designs constrained by structures, sequences, and ligands, including protein binders, protein–nucleic-acid interfaces, small-molecule binders, enzymes, symmetric assemblies, and partially fixed designs. The CHESS application submits GPU jobs from JSON configurations and referenced PDB/CIF files.

References: official repository, input specification, and examples.

Prepare inputs

Sign in to CHESS:

Open Application Center (应用中心):

Create input and output directories (用户名 means your username):

/hpcfs/fhome/用户名/rfdiffusion3/input/
/hpcfs/fhome/用户名/rfdiffusion3/output/

Upload JSON and referenced structures. Ensure readable inputs and writable outputs. Example JSON syntax:

{
  "spec-1": {
    "input": "/hpcfs/fhome/用户名/rfdiffusion3/input/example.pdb",
    "contig": "50-80,/0,A1-100",
    "select_unfixed_sequence": "A20-35",
    "ligand": "HAX,OAA"
  }
}

Field	Meaning
spec-1	Design name; multiple top-level designs are allowed
input	Full cluster path to a PDB/CIF structure
contig	Generated segments, fixed structural segments, and chain relationships
select_unfixed_sequence	Input residues whose sequence may change
ligand	Comma-separated ligand residue names from the input structure

Replace paths, residue numbers, and ligands with values actually present in your inputs. Consult the official examples for your design type. Standard JSON does not permit comments or trailing commas; validate syntax before submission.

Submit

Find RFdiffusion3 in Application Download (应用下载) and add it to the desktop:

Open it and choose Submit (提交). Enter a descriptive job name, uploaded JSON file, and result directory:

The documented template requests one GPU on qgpu_3090. Busy resources can cause queuing.

Monitor

Open List (列表) for the job ID, state, resources, and node:

Job details provide Output (输出), Data (数据), Input (输入), Details (详情), and Performance (性能):

Use the logged start/end times and input/output paths to confirm execution. The error-output pane can also contain normal program logs; judge success from job state and actual results, not the pane's name alone.

Demonstration:

Results

Outputs are organized by input filename, design name, batch, and model number. Each design normally produces a compressed structure (.cif.gz) and JSON configuration/result information. The original filename pattern is:

输入文件名_配置名_批次号_model_模型号.cif.gz
输入文件名_配置名_批次号_model_模型号.json

Here the Chinese terms stand for input filename, configuration name, batch number, and model number. Example:

demo_partial_diffusion_0_model_0.cif.gz
demo_partial_diffusion_0_model_0.json

Download via CHESS or inspect files on the cluster:

cd /hpcfs/fhome/用户名/rfdiffusion3/output
find . -maxdepth 2 -type f | sort

Decompress .cif.gz files and view them in a structural viewer such as PyMOL or ChimeraX.

Troubleshooting

Validate JSON quotes, commas, and brackets, and remove comments. Use full cluster paths for PDB/CIF inputs and check read permission. Match chain IDs, residue numbers, and ligand names exactly. Avoid duplicate submissions while waiting. If a completed job has no visible results, check its submitted output path and logs for actual model paths. Use separate directories for separate runs:

/hpcfs/fhome/用户名/rfdiffusion3/output/20260907_test/