input_format.md

Input format

OnixBind reads AlphaFold 3 input JSON. One file describes one complex.

{
  "dialect": "alphafold3",
  "version": 1,
  "name": "5S8I_A",
  "modelSeeds": [42],
  "sequences": [
    {
      "protein": {
        "id": ["A"],
        "sequence": "SMSYDIQAWKKQ...",
        "unpairedMsa": ">query\nSMSYDIQAWKKQ...\n>hit1\nSMSYEIQAWKKQ...\n",
        "pairedMsa": "",
        "templates": []
      }
    },
    { "ligand": { "id": ["B"], "ccdCodes": ["2LY"] } }
  ]
}

• modelSeeds — the record carries its own seeds and one prediction is written per seed. --seed 42 on the command line overrides them for every record.
• protein sequence — one entry per chain. A multi-chain complex lists several protein entries, each with its own id and its own MSA.
• unpairedMsa — A3M text inline. unpairedMsaPath takes a file path instead, and pairedMsa / pairedMsaPath are the paired equivalents for multi-chain complexes. Paths must be local; object-storage URLs are rejected. An MSA is required: these weights were fit with real alignments and a query-only input is not equivalent. Deeper alignments are cropped to 4096 rows, of which the model samples 2048 per recycle with the query row pinned first.
• ligand — either ccdCodes for a chemical component dictionary entry, or smiles for an arbitrary molecule:

{ "ligand": { "id": ["B"], "smiles": "C#C[C@H](NCCc1cocn1)c1nn(C)c2ccccc12" } }

A record whose token count would exceed the crop size is refused rather than scored, because a cropped complex is not the complex the file describes.

See src/examples/5S8I_A.json for a complete working record.