Skip to content

DATA.md

Latest commit

 

History

History
55 lines (44 loc) · 2.72 KB

DATA.md

File metadata and controls

55 lines (44 loc) · 2.72 KB
Raw

CATH benchmark data

Supply the exact preprocessed CATH 4.2 or CATH 4.3 benchmark files used for your experiment. The paper points to https://github.com/A4Bio/ProteinInvBench/ for code. This repository does not bundle the benchmark, rebuild the official split, or claim a particular download is identical to the authors' local data.

Verified download locations and the download helper are documented in PUBLIC_ASSETS.md. Data are kept outside the code release.

cath4.2/                   # or cath4.3/
  chain_set.jsonl
  chain_set_splits.json
  test_split_L100.json      # needed for --subset short
  test_split_sc.json        # needed for --subset single-chain

Each JSONL line has the form:

{"name":"chain_name","seq":"ACD","coords":{"N":[[0,0,0],[1,0,0],[2,0,0]],"CA":[[0,1,0],[1,1,0],[2,1,0]],"C":[[0,2,0],[1,2,0],[2,2,0]],"O":[[0,3,0],[1,3,0],[2,3,0]]}}

The numbers above illustrate the schema only. Each atom array must have exactly the same residue count as seq. Use canonical amino acid letters. Missing coordinates may be null or nonfinite values in a locally produced JSONL file; the loader drops the affected residue from both the sequence and coordinates. The backbone atom order passed to ESM3 is N, CA, C, O. Conversion to ESM's atom37 representation places oxygen at atom37 index 4.

The main split JSON must include a test list of chain names; training and validation lists are allowed but are not loaded for evaluation. Optional subset files use the same {"test": [...]} format and must be contained in the main test split. Single-chain membership is read from the supplied subset file rather than guessed from chain names. Short membership likewise uses the supplied L100 file, preserving the intended benchmark split.

Filtering follows the original source's test-data behavior:

  1. Keep chains in the selected official test split, in JSONL file order.
  2. Exclude chains with noncanonical residues or raw length above --max-length.
  3. Remove residues with missing N/CA/C/O coordinates, synchronously in all arrays.
  4. Exclude chains shorter than --min-length after filtering.
  5. Apply the requested start/end range to this filtered order.

The default maximum length is 500, matching the supplied CATH loader. The default minimum is 1: the old inference script exposed a minimum of 40 but never passed it to that loader. Set --min-length 40 explicitly if desired, and report this change in the evaluation protocol. No random cropping is applied.

The manifest records skipped-chain counts, removed residues, retained indices, selected names, and input hashes. Missing split members, duplicate selected chain names, and mismatched coordinate lengths produce errors instead of a silently smaller test set.