SkyWhal3/PMC-Corpus-Harvester

🧬 PMC Corpus Harvester

Build a disease-specific literature corpus in one evening!

Gene-agnostic PubMed Central corpus builder for rare disease research. Downloads HTML + metadata from PMC for building RAG (Retrieval-Augmented Generation) systems.

Originally developed for STXBP1-ARIA and SNAP25-ARIA projects.

✨ Features

Feature	Description
Gene-Agnostic	Configure for any gene or disease with YAML/JSON
NCBI API Support	10 req/sec with API key (vs 3/sec without)
Multi-Format	Downloads HTML (for multimodal) + XML + JSON metadata
Smart Filtering	Relevance scoring with wrong-gene exclusion
Resume Capable	Tracks progress, restart where you left off
Rate Limited	Built-in exponential backoff, respects NCBI limits
Built-in Configs	SNAP25, STXBP1 ready to use

🚀 Quick Start

# 1. Clone or download
git clone https://huggingface.co/datasets/SkyWhal3/PMC-Corpus-Harvester
cd PMC-Corpus-Harvester

# 2. Install dependencies
pip install -r requirements.txt

# 3. Set your NCBI API key (10x faster!)
export NCBI_API_KEY=your_key_here

# 4. Harvest!
python pmc_corpus_harvester.py --gene SNAP25

📋 Usage

Using Built-in Gene Configs

# List available genes
python pmc_corpus_harvester.py --list-genes

# Harvest SNAP25 literature
python pmc_corpus_harvester.py --gene SNAP25

# Harvest STXBP1 literature
python pmc_corpus_harvester.py --gene STXBP1

# Custom output directory
python pmc_corpus_harvester.py --gene SNAP25 --output /path/to/corpus

Using Custom Gene Config

# Export a template config
python pmc_corpus_harvester.py --gene SNAP25 --export-config my_gene.yaml

# Edit my_gene.yaml for your gene of interest

# Run with custom config
python pmc_corpus_harvester.py --config my_gene.yaml

Environment Variables

Variable	Description
NCBI_API_KEY	Your NCBI API key (get one here)
NCBI_EMAIL	Your email for NCBI API (recommended)

With API key: 10 requests/second
Without API key: 3 requests/second

📁 Output Structure

corpus/
└── SNAP25/
    ├── html/           # Full HTML articles (for multimodal RAG)
    │   ├── PMC1234567.html
    │   └── ...
    ├── json/           # Parsed metadata + extracted text
    │   ├── PMC1234567.json
    │   └── ...
    ├── xml/            # Raw PMC XML (backup)
    │   ├── PMC1234567.xml
    │   └── ...
    └── logs/
        ├── progress.json    # Resume state
        └── harvest_*.log    # Download logs

JSON Article Format

Each JSON file contains:

{
  "pmc_id": "PMC1234567",
  "pmid": "12345678",
  "doi": "10.1234/example",
  "title": "Article Title",
  "abstract": "Full abstract text...",
  "keywords": ["keyword1", "keyword2"],
  "main_text": "Full article body text...",
  "figures": [
    {"label": "Figure 1", "caption": "Description...", "href": "fig1.jpg"}
  ],
  "tables": [
    {"label": "Table 1", "caption": "Data summary..."}
  ],
  "authors": ["Author One", "Author Two"],
  "metadata": {
    "journal": "Journal Name",
    "pub_date": "2024-01-15",
    "source": "pmc_eutils"
  },
  "relevance": {
    "score": 45,
    "reason": "primary:snap25(3),snare:5,disease:2"
  },
  "has_html": true
}

⚙️ Custom Gene Configuration

Create a YAML config for your gene of interest:

# my_gene_config.yaml
gene_symbol: SCN1A
gene_name: "Sodium Voltage-Gated Channel Alpha Subunit 1"
aliases:
  - Nav1.1
  - sodium channel

primary_queries:
  - '"SCN1A"[Title/Abstract]'
  - '"Nav1.1"[Title/Abstract]'
  - 'SCN1A[Gene]'
  - '"SCN1A encephalopathy"'
  - '"Dravet syndrome"[Title/Abstract]'

secondary_queries:
  - '"sodium channel"[Title/Abstract] AND epilepsy'
  - '"voltage-gated sodium"[Title/Abstract] AND mutation'

primary_keywords:
  - scn1a
  - nav1.1
  - dravet syndrome
  - dravet

related_keywords:
  - sodium channel
  - voltage-gated
  - channelopathy

disease_keywords:
  - epileptic encephalopathy
  - febrile seizure
  - intractable epilepsy

therapy_keywords:
  - gene therapy
  - antisense oligonucleotide
  - sodium channel blocker

exclude_terms:
  - scn2a
  - scn8a
  - scn9a

relevance_threshold: 15

Then run:

python pmc_corpus_harvester.py --config my_gene_config.yaml

📊 Search Strategy

Primary Queries (High Specificity)

• Direct gene name searches ("SNAP25"[Title/Abstract])
• Gene symbol searches (SNAP25[Gene])
• Disease associations ("SNAP25 encephalopathy")
• Variant mentions ("SNAP25 mutation")

Secondary Queries (Related Concepts)

• Protein family/pathway searches
• Partner protein interactions
• Therapeutic approaches
• Related diseases

Relevance Scoring

Category	Points	Examples
Primary keywords	+20 each	Gene name, direct disease
Related keywords	+5 each	Protein family, pathway
Disease keywords	+3 each	Epilepsy, encephalopathy
Therapy keywords	+3 each	Gene therapy, CRISPR
Exclusion terms	-10 each	Wrong gene family members

Threshold: 15 points (configurable)

🎯 Built-in Gene Configs

SNAP25

• Full name: Synaptosomal-Associated Protein 25
• Disease: SNAP25-DEE (Developmental Epileptic Encephalopathy)
• Expected corpus: ~2,000-5,000 papers
• Key papers: Klöckner 2020, Rohena 2013, Bhumbra 2024

STXBP1

• Full name: Syntaxin Binding Protein 1 (Munc18-1)
• Disease: STXBP1-DEE, Ohtahara syndrome
• Expected corpus: ~15,000-20,000 papers
• Key papers: Stamberger 2016, Kovačević 2018

🔧 Requirements

requests>=2.28.0
tqdm>=4.65.0
pyyaml>=6.0

Optional for RAG building:

sentence-transformers>=2.2.0
chromadb>=0.4.0

📈 Performance

Metric	With API Key	Without API Key
Request rate	10/sec	3/sec
1000 papers	~2 min	~6 min
10000 papers	~20 min	~60 min

🔗 Related Projects

• STXBP1-RAG-Nemotron - RAG database for STXBP1 research
• SNAP25-RAG-Nemotron - RAG database for SNAP25 research
• STXBP1-Eval-Harness - Evaluation benchmark

🏗️ Next Steps After Harvesting

1. Build RAG Index: Use embeddings (Nemotron-1B recommended)
2. Deploy to HuggingFace: Share your corpus
3. Create Eval Benchmark: Build domain-specific test questions
4. Connect with Foundations: Share with patient communities

📜 License

MIT License - Use freely for research and clinical applications.

🙏 Acknowledgments

• NCBI/NLM for PubMed Central API
• STXBP1 Foundation for inspiration
• SNAP25 Foundation for collaboration

📬 Contact

• Author: Adam Freygang
• Project: NeuroSenpai / ARIA
• HuggingFace: @SkyWhal3

---

Built with ❤️ for the rare disease research community

"Let's freak people out by building this in one evening!"