OpenFold3

How SubSeq Runs It

• Entrypoint is fixed to run_openfold predict.
• The platform forces local-only output/cache paths under /outputs and /ref.
• /ref is writable and used for OPENFOLD_CACHE=/ref/.openfold3.
• SubSeq does not expose hosted third-party MSA services in this profile.

Required Input

OpenFold3 needs a query JSON file. You can upload it under /inputs or use the UI prefill, which places a starter file at /aux/query.json.

{
  "queries": {
    "ubiquitin": {
      "chains": [
        {
          "molecule_type": "protein",
          "chain_ids": ["A"],
          "sequence": "MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG"
        }
      ]
    }
  }
}

Example Arguments

--query-json=/aux/query.json
--use-templates=true
--num-diffusion-samples=1
--num-model-seeds=1

If you already staged a checkpoint under /ref/.openfold3, you can also pass --inference-ckpt-path=/ref/.openfold3/<checkpoint>.pt.

Notes

• Stage checkpoints and any other required local assets under /ref.
• This SubSeq image tracks the upstream OpenFold3 CLI, but avoids the upstream Dockerfile’s hardcoded CUDA arch list that skips V100 GPUs.

Submit

Queue a run from New Job -> OpenFold3.