NegBioDB / scripts /generate_dataset_card.py
jang1563's picture
NegBioDB final: 4 domains, fully audited
6d1bbc7
#!/usr/bin/env python3
"""Generate HuggingFace dataset card and Croissant JSON-LD metadata.
Reads parquet/JSONL export files, extracts schema and statistics,
and generates:
- exports/README.md (HuggingFace dataset card)
- exports/croissant.json (MLCommons Croissant 1.0 metadata)
Usage:
python scripts/generate_dataset_card.py
"""
import json
from pathlib import Path
PROJECT_ROOT = Path(__file__).resolve().parent.parent
EXPORTS_DIR = PROJECT_ROOT / "exports"
# Try importing pyarrow for parquet schema extraction
try:
import pyarrow.parquet as pq
HAS_PYARROW = True
except ImportError:
HAS_PYARROW = False
def get_parquet_info(path: Path) -> dict:
"""Extract row count and schema from a parquet file."""
if not HAS_PYARROW or not path.exists():
return {"rows": "N/A", "columns": []}
pf = pq.ParquetFile(path)
schema = pf.schema_arrow
return {
"rows": pf.metadata.num_rows,
"columns": [
{"name": field.name, "type": str(field.type)}
for field in schema
],
}
def count_jsonl(path: Path) -> int:
"""Count lines in a JSONL file."""
if not path.exists():
return 0
with open(path) as f:
return sum(1 for _ in f)
def format_size(path: Path) -> str:
"""Human-readable file size."""
if not path.exists():
return "N/A"
size = path.stat().st_size
for unit in ["B", "KB", "MB", "GB"]:
if size < 1024:
return f"{size:.1f} {unit}"
size /= 1024
return f"{size:.1f} TB"
def generate_readme() -> str:
"""Generate HuggingFace dataset card content."""
sections = []
# YAML frontmatter
sections.append("""---
license: cc-by-sa-4.0
language:
- en
tags:
- drug-target-interaction
- clinical-trials
- protein-protein-interaction
- negative-results
- benchmark
- bioinformatics
task_categories:
- text-classification
- question-answering
- text-generation
size_categories:
- 10M<n<100M
---
""")
sections.append("# NegBioDB: A Negative Results Database for Biomedical Sciences\n")
sections.append(
"NegBioDB is a large-scale database of experimentally confirmed negative "
"results across three biomedical domains, paired with dual ML/LLM benchmarks "
"for evaluating how well computational methods handle negative evidence.\n"
)
# Overview
sections.append("## Overview\n")
sections.append("| Domain | Negative Results | Entities | Benchmark Tasks |")
sections.append("|--------|-----------------|----------|-----------------|")
# DTI stats
dti_pairs = EXPORTS_DIR / "negbiodb_dti_pairs.parquet"
dti_info = get_parquet_info(dti_pairs)
sections.append(
f"| DTI (Drug-Target Interaction) | {dti_info.get('rows', '30.5M'):,} pairs "
f"| 919K compounds, 3.7K targets | ML (M1) + LLM (L1-L4) |"
)
# CT stats
ct_pairs = EXPORTS_DIR / "ct" / "negbiodb_ct_pairs.parquet"
ct_info = get_parquet_info(ct_pairs)
sections.append(
f"| CT (Clinical Trial Failure) | {ct_info.get('rows', '132K'):,} results "
f"| 216K trials, 176K interventions | ML (M1-M2) + LLM (L1-L4) |"
)
# PPI stats
ppi_pairs = EXPORTS_DIR / "ppi" / "negbiodb_ppi_pairs.parquet"
ppi_info = get_parquet_info(ppi_pairs)
sections.append(
f"| PPI (Protein-Protein Interaction) | {ppi_info.get('rows', '2.2M'):,} pairs "
f"| 18.4K proteins | ML (M1) |"
)
sections.append("")
# Data Sources
sections.append("## Data Sources\n")
sections.append("### DTI Domain")
sections.append("| Source | License | Contribution |")
sections.append("|--------|---------|-------------|")
sections.append("| ChEMBL 34 | CC BY-SA 3.0 | Curated bioactivity data |")
sections.append("| PubChem BioAssay | Public domain | HTS screening results |")
sections.append("| BindingDB | CC BY-SA 3.0 | Binding measurements |")
sections.append("| DAVIS | CC BY 4.0 | Kinase selectivity panel |")
sections.append("")
sections.append("### CT Domain")
sections.append("| Source | License | Contribution |")
sections.append("|--------|---------|-------------|")
sections.append("| AACT (ClinicalTrials.gov) | Public domain | Trial metadata |")
sections.append("| Open Targets | Apache 2.0 | Drug-target mappings |")
sections.append("| CTO | MIT | Clinical trial outcomes |")
sections.append("| Shi & Du 2024 | CC BY 4.0 | Safety/efficacy data |")
sections.append("")
sections.append("### PPI Domain")
sections.append("| Source | License | Contribution |")
sections.append("|--------|---------|-------------|")
sections.append("| IntAct | CC BY 4.0 | Curated non-interactions |")
sections.append("| HuRI | CC BY 4.0 | Y2H systematic negatives |")
sections.append("| hu.MAP 3.0 | CC BY 4.0 | Complex-derived negatives |")
sections.append("| STRING v12.0 | CC BY 4.0 | Zero-score protein pairs |")
sections.append("")
# File Structure
sections.append("## File Structure\n")
sections.append("### DTI Files")
sections.append("| File | Size | Rows | Description |")
sections.append("|------|------|------|-------------|")
dti_files = [
("negbiodb_dti_pairs.parquet", "All negative DTI pairs with metadata"),
("negbiodb_m1_balanced.parquet", "M1 balanced dataset (1:1 pos:neg)"),
("negbiodb_m1_realistic.parquet", "M1 realistic dataset (1:10 pos:neg)"),
("negbiodb_m1_uniform_random.parquet", "Control: uniform random negatives"),
("negbiodb_m1_degree_matched.parquet", "Control: degree-matched negatives"),
("negbiodb_m1_balanced_ddb.parquet", "M1 balanced with degree-balanced split"),
]
for fname, desc in dti_files:
fpath = EXPORTS_DIR / fname
info = get_parquet_info(fpath)
sections.append(
f"| `{fname}` | {format_size(fpath)} | {info.get('rows', 'N/A'):,} | {desc} |"
)
sections.append("")
sections.append("### CT Files")
sections.append("| File | Size | Rows | Description |")
sections.append("|------|------|------|-------------|")
ct_files = [
("ct/negbiodb_ct_pairs.parquet", "All CT failure pairs"),
("ct/negbiodb_ct_m1_balanced.parquet", "CT-M1 balanced (success vs failure)"),
("ct/negbiodb_ct_m1_realistic.parquet", "CT-M1 realistic ratio"),
("ct/negbiodb_ct_m2.parquet", "CT-M2 multiclass failure category"),
]
for fname, desc in ct_files:
fpath = EXPORTS_DIR / fname
info = get_parquet_info(fpath)
rows = info.get("rows", "N/A")
rows_str = f"{rows:,}" if isinstance(rows, int) else rows
sections.append(f"| `{fname}` | {format_size(fpath)} | {rows_str} | {desc} |")
sections.append("")
sections.append("### PPI Files")
sections.append("| File | Size | Rows | Description |")
sections.append("|------|------|------|-------------|")
ppi_files = [
("ppi/negbiodb_ppi_pairs.parquet", "All negative PPI pairs"),
("ppi/ppi_m1_balanced.parquet", "PPI-M1 balanced (1:1 pos:neg)"),
("ppi/ppi_m1_realistic.parquet", "PPI-M1 realistic (1:10 ratio)"),
]
for fname, desc in ppi_files:
fpath = EXPORTS_DIR / fname
info = get_parquet_info(fpath)
rows = info.get("rows", "N/A")
rows_str = f"{rows:,}" if isinstance(rows, int) else rows
sections.append(f"| `{fname}` | {format_size(fpath)} | {rows_str} | {desc} |")
sections.append("")
sections.append("### LLM Benchmark Files (DTI)")
sections.append("| File | Items | Description |")
sections.append("|------|-------|-------------|")
llm_files = [
("llm_benchmarks/l1_mcq.jsonl", "L1: 4-class activity MCQ"),
("llm_benchmarks/l3_reasoning_pilot.jsonl", "L3: Scientific reasoning (pilot)"),
("llm_benchmarks/l4_tested_untested.jsonl", "L4: Tested vs untested discrimination"),
]
for fname, desc in llm_files:
fpath = EXPORTS_DIR / fname
n = count_jsonl(fpath)
sections.append(f"| `{fname}` | {n:,} | {desc} |")
sections.append("")
# Benchmark Tasks
sections.append("## Benchmark Tasks\n")
sections.append("### ML Benchmarks")
sections.append("| Task | Domain | Type | Splits |")
sections.append("|------|--------|------|--------|")
sections.append("| M1 | DTI | Binary (active/inactive) | random, cold_compound, cold_target, degree_balanced |")
sections.append("| CT-M1 | CT | Binary (success/failure) | random, cold_drug, cold_condition, temporal, scaffold, cold_both |")
sections.append("| CT-M2 | CT | 7-way failure category | Same as CT-M1 |")
sections.append("| PPI-M1 | PPI | Binary (interact/non-interact) | random, cold_protein, cold_both, degree_balanced |")
sections.append("")
sections.append("### LLM Benchmarks (DTI)")
sections.append("| Task | Type | Size | Description |")
sections.append("|------|------|------|-------------|")
sections.append("| L1 | MCQ classification | 1,600 | 4-class activity level |")
sections.append("| L2 | Structured extraction | ~100 | Extract results from abstracts |")
sections.append("| L3 | Reasoning | 50 | Explain compound-target inactivity |")
sections.append("| L4 | Discrimination | 400 | Tested vs untested pair |")
sections.append("")
# License
sections.append("## License\n")
sections.append(
"This dataset is released under **CC BY-SA 4.0**, due to the viral "
"clause of ChEMBL's CC BY-SA 3.0 license. See the LICENSE file for details.\n"
)
# Citation
sections.append("## Citation\n")
sections.append("```bibtex")
sections.append("@dataset{negbiodb2026,")
sections.append(" title={NegBioDB: A Negative Results Database for Biomedical Sciences},")
sections.append(" author={Jang, Jungwon},")
sections.append(" year={2026},")
sections.append(" url={https://github.com/jang1563/NegBioDB}")
sections.append("}")
sections.append("```")
return "\n".join(sections)
def generate_croissant() -> dict:
"""Generate MLCommons Croissant JSON-LD metadata."""
croissant = {
"@context": {
"@vocab": "https://schema.org/",
"sc": "https://schema.org/",
"cr": "http://mlcommons.org/croissant/",
"rai": "http://mlcommons.org/croissant/RAI/",
},
"@type": "sc:Dataset",
"name": "NegBioDB",
"description": (
"A large-scale database of experimentally confirmed negative results "
"across three biomedical domains (DTI, Clinical Trials, PPI), "
"with dual ML/LLM benchmarks."
),
"license": "https://creativecommons.org/licenses/by-sa/4.0/",
"url": "https://github.com/jang1563/NegBioDB",
"version": "1.0.0",
"datePublished": "2026",
"creator": {
"@type": "sc:Person",
"name": "Jungwon Jang",
},
"distribution": [],
"recordSet": [],
}
# File objects (distribution)
file_defs = [
{
"name": "dti_pairs",
"contentUrl": "negbiodb_dti_pairs.parquet",
"encodingFormat": "application/x-parquet",
"description": "All negative DTI pairs with source, tier, and activity data",
},
{
"name": "dti_m1_balanced",
"contentUrl": "negbiodb_m1_balanced.parquet",
"encodingFormat": "application/x-parquet",
"description": "DTI M1 balanced benchmark dataset (1:1 positive:negative)",
},
{
"name": "ct_pairs",
"contentUrl": "ct/negbiodb_ct_pairs.parquet",
"encodingFormat": "application/x-parquet",
"description": "All clinical trial failure pairs",
},
{
"name": "ppi_pairs",
"contentUrl": "ppi/negbiodb_ppi_pairs.parquet",
"encodingFormat": "application/x-parquet",
"description": "All negative PPI pairs",
},
{
"name": "llm_l1",
"contentUrl": "llm_benchmarks/l1_mcq.jsonl",
"encodingFormat": "application/jsonl",
"description": "L1 MCQ classification benchmark",
},
{
"name": "llm_l4",
"contentUrl": "llm_benchmarks/l4_tested_untested.jsonl",
"encodingFormat": "application/jsonl",
"description": "L4 tested/untested discrimination benchmark",
},
]
for fd in file_defs:
croissant["distribution"].append({
"@type": "cr:FileObject",
"name": fd["name"],
"contentUrl": fd["contentUrl"],
"encodingFormat": fd["encodingFormat"],
"description": fd["description"],
})
# Record sets (key columns)
record_defs = [
{
"name": "dti_pairs_record",
"source": "dti_pairs",
"fields": [
{"name": "inchikey_connectivity", "dataType": "sc:Text"},
{"name": "uniprot_id", "dataType": "sc:Text"},
{"name": "activity_type", "dataType": "sc:Text"},
{"name": "pchembl_value", "dataType": "sc:Float"},
{"name": "source", "dataType": "sc:Text"},
{"name": "tier", "dataType": "sc:Text"},
],
},
{
"name": "ct_pairs_record",
"source": "ct_pairs",
"fields": [
{"name": "nct_id", "dataType": "sc:Text"},
{"name": "intervention_name", "dataType": "sc:Text"},
{"name": "failure_category", "dataType": "sc:Text"},
{"name": "tier", "dataType": "sc:Text"},
{"name": "highest_phase_reached", "dataType": "sc:Text"},
],
},
{
"name": "ppi_pairs_record",
"source": "ppi_pairs",
"fields": [
{"name": "protein_a", "dataType": "sc:Text"},
{"name": "protein_b", "dataType": "sc:Text"},
{"name": "source", "dataType": "sc:Text"},
{"name": "tier", "dataType": "sc:Text"},
],
},
]
for rd in record_defs:
croissant["recordSet"].append({
"@type": "cr:RecordSet",
"name": rd["name"],
"source": rd["source"],
"field": [
{
"@type": "cr:Field",
"name": f["name"],
"dataType": f["dataType"],
"description": f["name"].replace("_", " "),
}
for f in rd["fields"]
],
})
return croissant
def main():
# Generate README
readme_text = generate_readme()
readme_path = EXPORTS_DIR / "README.md"
readme_path.write_text(readme_text)
print(f"Written: {readme_path}")
print(f" Lines: {len(readme_text.splitlines())}")
# Generate Croissant
croissant = generate_croissant()
croissant_path = EXPORTS_DIR / "croissant.json"
with open(croissant_path, "w") as f:
json.dump(croissant, f, indent=2)
print(f"Written: {croissant_path}")
# Verify: check all referenced files exist
missing = []
for dist in croissant["distribution"]:
fpath = EXPORTS_DIR / dist["contentUrl"]
if not fpath.exists():
missing.append(dist["contentUrl"])
if missing:
print(f"\n WARNING: {len(missing)} referenced files not found:")
for m in missing:
print(f" {m}")
else:
print("\n All referenced files exist.")
if __name__ == "__main__":
main()