anon0416/bgc-submission

BGC-Bench Data Bundle (v0.1.0)

This is the data half of the BGC-Bench reproducibility release. It pairs with an anonymous code repository linked from the paper. Reviewers reach the code via OpenReview during review; the public code repository URL will be filled in at camera-ready.

BGC-Bench is a benchmark for biosynthetic gene cluster (BGC) representation learning, built around two evaluation axes — out-of-distribution generalisation across genotype/phenotype shifts and downstream functionality across classification, regression, retrieval, and active learning — using curated splits from MIBiG 4.0.

What's in this dataset

dir / file	contents	size (approx)	tier consumes
splits/	per-task train/val/test CSVs (T1, T2-class, T2-fewshot, T3, T_chem, AL grids)	~3 MB	B, C
processed/	curated metadata, NPClassifier labels, Bemis–Murcko scaffolds, RDKit descriptors, GCF assignments	~2 MB	B, C
embeddings/	per-encoder NPZ embeddings (8 encoders + per-CDS variants)	~534 MB	B, C
results/	per-cell evaluation JSONs, summary.json, AL CSVs, descriptor analyses	~54 MB	A, B, C
manifest.json	sha256 + size for every file in this dataset	<1 MB	all
DATASHEET.md	datasheet (Gebru et al. format)	<1 MB	reference
LICENSE	Apache-2.0 (code-derived results)	-	-
DATA_LICENSE	CC-BY-4.0 (label artefacts inheriting from MIBiG)	-	-

Not included on Hugging Face: raw/ (~5.5 GB MIBiG 4.0 GenBank + Pfam HMM + BiG-SCAPE outputs). Per Appendix H4 of the paper, raw upstream sequence data is not redistributed; users running the end-to-end Tier-C pipeline fetch from MIBiG / Pfam / BiG-SCAPE upstream.

How to load

The accompanying code repository ships a fetch wrapper (scripts/00_fetch_data/04_fetch_bundle.sh --hf) that downloads this dataset and places each subtree at the canonical path the evaluation pipeline expects (data/embeddings/, data/splits/, data/processed/, results/).

If you only need raw access:

from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="anon0416/bgc-submission",
    repo_type="dataset",
    local_dir="./bgc-bench-release",
)

The dataset viewer exposes a representative subset of the tabular tables (core MIBiG metadata + one canonical split per task family). The full collection — all 45 split files, all encoder embeddings (NPZ), per-cell evaluation results, and the Datasheet — is shipped as files; manifest.json is the canonical index with sha256 for every artefact.

Integrity verification

After download, verify every file's sha256 against manifest.json:

import hashlib, json, sys
from pathlib import Path
m = json.loads(Path("manifest.json").read_text())
fail = [r for r, i in m.items()
        if hashlib.sha256(Path(r).read_bytes()).hexdigest() != i["sha256"]]
print("OK" if not fail else f"{len(fail)} mismatches"); sys.exit(len(fail))

A zero exit status confirms every file matches its recorded hash.

License

• Code-derived results (results/, embeddings/, processed/, splits/): Apache-2.0 — see LICENSE.
• Label artefacts inheriting MIBiG provenance: CC-BY-4.0 — see DATA_LICENSE. Attribute MIBiG 4.0 (Terlouw et al., 2023) when redistributing labels or splits derived from MIBiG accessions.
• Encoder weights (referenced by embeddings/) are NOT re-hosted; their upstream licences govern reuse.

Citation

@article{bgcbench2026,
  title  = {BGC-Bench: A Biologically and Chemically Informed Evaluation Framework for Biosynthetic Gene Cluster Models},
  author = {Anonymous},
  year   = {2026}
}

The companion CITATION.cff lives in the code repository and will be filled in at camera-ready.