danielquillanroxas/pubmed-graphrag-data

PubMed-GraphRAG Dataset

A large-scale biomedical knowledge graph dataset for Graph-based Retrieval-Augmented Generation (GraphRAG).

Dataset Overview

Component	Size	Description
Papers	12.5M	PubMed abstracts (2000-2024) with title, abstract, DOI
Entities	7.5M	Genes, Diseases, Chemicals, Species, Mutations, Cell Lines
Relationships	266M+	MENTIONED_IN, TREAT, ASSOCIATE, CAUSE, INHIBIT, etc.
Embeddings	12.5M vectors	MedCPT, BiCA, MedTE (768-dim)

Dataset Structure

pubmed-graphrag-data/
├── graph_ready/                    # Papers & major MENTIONED_IN relationships
│   ├── papers.csv                  # 12.5M papers (pmid, title, abstract, year, doi)
│   ├── mentioned_in_gene.csv       # Gene-Paper relationships
│   ├── mentioned_in_disease.csv    # Disease-Paper relationships
│   ├── mentioned_in_chemical.csv   # Chemical-Paper relationships
│   ├── entity_relations.csv        # Entity-Entity semantic relationships
│   └── complete_papers_with_metadata.jsonl  # Full data in JSONL format
│
├── neo4j_csv/                      # Entity definitions & additional relationships
│   ├── gene_entities.csv           # 3.8M genes (entity_id, mention)
│   ├── disease_entities.csv        # 12K diseases
│   ├── chemical_entities.csv       # 120K chemicals
│   ├── species_entities.csv        # 359K species
│   ├── mutation_entities.csv       # 3.2M mutations
│   ├── cellline_entities.csv       # 63K cell lines
│   ├── mentioned_in_species.csv    # Species-Paper relationships
│   ├── mentioned_in_mutation.csv   # Mutation-Paper relationships
│   ├── mentioned_in_cellline.csv   # CellLine-Paper relationships
│   └── rel_*.csv                   # 32 semantic relationship files (TREAT, CAUSE, etc.)
│
├── medcpt/                         # MedCPT embeddings (recommended)
│   ├── embeddings_000000.npz       # Each NPZ contains 'embeddings' and 'pmids' arrays
│   ├── ...
│   └── embeddings_000251.npz
│
├── bica_base/                      # BiCA embeddings
│   └── embeddings_*.npz
│
└── medte/                          # MedTE embeddings
    └── embeddings_*.npz

Quick Start: Building the Knowledge Graph

Prerequisites

• Neo4j 5.x (Community Edition works)
• Python 3.8+
• ~50GB disk space
• 16GB+ RAM recommended

Step 1: Download all CSV files

from huggingface_hub import snapshot_download

# Download all CSVs
snapshot_download(
    repo_id="dannyroxas/pubmed-graphrag-data",
    repo_type="dataset",
    local_dir="./pubmed-data",
    allow_patterns=["graph_ready/*.csv", "neo4j_csv/*.csv"]
)

Step 2: Import to Neo4j

# Stop Neo4j
sudo systemctl stop neo4j

# Import (adjust paths as needed)
cd ./pubmed-data

sudo neo4j-admin database import full \
  --nodes=Paper="graph_ready/papers.csv" \
  --nodes=Gene="neo4j_csv/gene_entities.csv" \
  --nodes=Disease="neo4j_csv/disease_entities.csv" \
  --nodes=Chemical="neo4j_csv/chemical_entities.csv" \
  --nodes=Species="neo4j_csv/species_entities.csv" \
  --nodes=Mutation="neo4j_csv/mutation_entities.csv" \
  --nodes=CellLine="neo4j_csv/cellline_entities.csv" \
  --relationships=MENTIONED_IN="graph_ready/mentioned_in_gene.csv" \
  --relationships=MENTIONED_IN="graph_ready/mentioned_in_disease.csv" \
  --relationships=MENTIONED_IN="graph_ready/mentioned_in_chemical.csv" \
  --relationships=MENTIONED_IN="neo4j_csv/mentioned_in_species.csv" \
  --relationships=MENTIONED_IN="neo4j_csv/mentioned_in_mutation.csv" \
  --relationships=MENTIONED_IN="neo4j_csv/mentioned_in_cellline.csv" \
  --relationships=ASSOCIATE="neo4j_csv/rel_*_ASSOCIATE.csv" \
  --relationships=TREAT="neo4j_csv/rel_*_TREAT.csv" \
  --relationships=CAUSE="neo4j_csv/rel_*_CAUSE.csv" \
  --relationships=INHIBIT="neo4j_csv/rel_*_INHIBIT.csv" \
  --relationships=STIMULATE="neo4j_csv/rel_*_STIMULATE.csv" \
  --relationships=INTERACT="neo4j_csv/rel_*_INTERACT.csv" \
  --relationships=POSITIVE_CORRELATE="neo4j_csv/rel_*_POSITIVE_CORRELATE.csv" \
  --relationships=NEGATIVE_CORRELATE="neo4j_csv/rel_*_NEGATIVE_CORRELATE.csv" \
  --relationships=COTREAT="neo4j_csv/rel_*_COTREAT.csv" \
  --relationships=COMPARE="neo4j_csv/rel_*_COMPARE.csv" \
  --relationships=PREVENT="neo4j_csv/rel_*_PREVENT.csv" \
  --relationships=DRUG_INTERACT="neo4j_csv/rel_*_DRUG_INTERACT.csv" \
  --skip-bad-relationships \
  --skip-duplicate-nodes \
  neo4j

# Fix permissions and start
sudo chown -R neo4j:neo4j /var/lib/neo4j/data
sudo systemctl start neo4j

Step 3: Create indexes

CREATE INDEX paper_pmid FOR (p:Paper) ON (p.pmid);
CREATE INDEX gene_mention FOR (g:Gene) ON (g.mention);
CREATE INDEX gene_id FOR (g:Gene) ON (g.entity_id);
CREATE INDEX disease_mention FOR (d:Disease) ON (d.mention);
CREATE INDEX chemical_mention FOR (c:Chemical) ON (c.mention);
CREATE INDEX species_mention FOR (s:Species) ON (s.mention);
CREATE INDEX mutation_mention FOR (m:Mutation) ON (m.mention);

Quick Start: Building FAISS Index

Step 1: Download embeddings

from huggingface_hub import snapshot_download

# Download MedCPT embeddings (recommended for biomedical)
snapshot_download(
    repo_id="dannyroxas/pubmed-graphrag-data",
    repo_type="dataset",
    local_dir="./pubmed-data",
    allow_patterns="medcpt/*.npz"
)

Step 2: Build IVF-PQ index

import numpy as np
import faiss
from pathlib import Path
from tqdm import tqdm

# Configuration
NPZ_DIR = Path("./pubmed-data/medcpt")
OUTPUT_DIR = Path("./faiss_index")
OUTPUT_DIR.mkdir(exist_ok=True)

NLIST = 4096      # Number of clusters
M = 96            # Subquantizers (768/96 = 8 dims each)
NBITS = 8         # Bits per code
TRAIN_SIZE = 500000

# Get all NPZ files
npz_files = sorted(NPZ_DIR.glob("embeddings_*.npz"))
print(f"Found {len(npz_files)} embedding files")

# Step 1: Sample vectors for training
print("Sampling vectors for training...")
train_vectors = []
for npz_file in tqdm(npz_files[:50], desc="Sampling"):
    data = np.load(npz_file)
    emb = data["embeddings"].astype(np.float32)
    idx = np.random.choice(len(emb), min(10000, len(emb)), replace=False)
    train_vectors.append(emb[idx])

train_vectors = np.vstack(train_vectors)[:TRAIN_SIZE]
faiss.normalize_L2(train_vectors)
dim = train_vectors.shape[1]
print(f"Training set: {len(train_vectors):,} vectors, dim={dim}")

# Step 2: Create and train index
print("Training IVF-PQ index...")
quantizer = faiss.IndexFlatIP(dim)
index = faiss.IndexIVFPQ(quantizer, dim, NLIST, M, NBITS, faiss.METRIC_INNER_PRODUCT)
index.train(train_vectors)
index.nprobe = 128
del train_vectors

# Step 3: Add all vectors
print("Adding all vectors...")
all_pmids = []
for npz_file in tqdm(npz_files, desc="Adding"):
    data = np.load(npz_file)
    emb = data["embeddings"].astype(np.float32)
    pmids = data["pmids"]
    faiss.normalize_L2(emb)
    index.add(emb)
    all_pmids.extend(pmids.tolist())

# Step 4: Save
print("Saving index...")
faiss.write_index(index, str(OUTPUT_DIR / "index.faiss"))
np.save(OUTPUT_DIR / "pmids.npy", np.array(all_pmids))

print(f"\nDone!")
print(f"  Index: {index.ntotal:,} vectors")
print(f"  Size: ~{index.ntotal * M / 1e9:.1f} GB (compressed)")

NPZ File Format

Each .npz file contains:

• embeddings: numpy array of shape (N, 768) - float32 vectors
• pmids: numpy array of shape (N,) - corresponding PubMed IDs

import numpy as np

data = np.load("medcpt/embeddings_000000.npz")
print(data["embeddings"].shape)  # (50000, 768)
print(data["pmids"].shape)       # (50000,)

Entity Types

Type	Count	ID Format	Example
Gene	3.8M	NCBI Gene ID	348 (APOE)
Disease	12K	MeSH ID	MESH:D000544 (Alzheimer Disease)
Chemical	120K	MeSH ID	MESH:D001241 (Aspirin)
Species	359K	NCBI Taxonomy	9606 (Homo sapiens)
Mutation	3.2M	dbSNP/HGVS	rs121912438
Cell Line	63K	Cellosaurus	CVCL_0030

Relationship Types

Type	Count	Example
MENTIONED_IN	248M	(Gene)-[:MENTIONED_IN]->(Paper)
ASSOCIATE	9.2M	(Gene)-[:ASSOCIATE]->(Disease)
TREAT	3.1M	(Chemical)-[:TREAT]->(Disease)
POSITIVE_CORRELATE	1.8M	(Chemical)-[:POSITIVE_CORRELATE]->(Gene)
NEGATIVE_CORRELATE	1.8M	(Chemical)-[:NEGATIVE_CORRELATE]->(Gene)
CAUSE	1.3M	(Mutation)-[:CAUSE]->(Disease)
STIMULATE	388K	(Chemical)-[:STIMULATE]->(Gene)
INHIBIT	307K	(Chemical)-[:INHIBIT]->(Gene)
COTREAT	237K	(Chemical)-[:COTREAT]->(Chemical)
COMPARE	208K	(Chemical)-[:COMPARE]->(Chemical)
INTERACT	123K	(Gene)-[:INTERACT]->(Gene)
PREVENT	~5K	(Mutation)-[:PREVENT]->(Disease)

Embedding Models

Model	HuggingFace	Dimensions	Recommended For
MedCPT	ncbi/MedCPT-Article-Encoder	768	Medical literature search
BiCA	bisectgroup/BiCA-base	768	Biomedical contrastive learning
MedTE	MohammadKhodadad/MedTE	768	Medical text embeddings

Sample Neo4j Queries

Find genes associated with Alzheimer's disease

MATCH (d:Disease)-[:ASSOCIATE]-(g:Gene)
WHERE d.mention CONTAINS "Alzheimer"
RETURN g.mention AS gene, g.entity_id AS ncbi_id
LIMIT 20

Find drugs that treat diabetes

MATCH (c:Chemical)-[:TREAT]->(d:Disease)
WHERE d.mention CONTAINS "Diabetes"
RETURN c.mention AS drug, count(*) AS evidence_count
ORDER BY evidence_count DESC
LIMIT 10

Multi-hop: Find genes related to drugs treating a disease

MATCH (c:Chemical)-[:TREAT]->(d:Disease)
WHERE d.mention CONTAINS "Breast"
MATCH (c)-[r:ASSOCIATE|INHIBIT|STIMULATE]-(g:Gene)
RETURN DISTINCT g.mention AS gene, c.mention AS drug, type(r) AS relationship
LIMIT 25

Find papers discussing a specific gene

MATCH (g:Gene {mention: "BRCA1"})-[:MENTIONED_IN]->(p:Paper)
RETURN p.pmid, p.title, p.year
ORDER BY p.year DESC
LIMIT 10

Data Sources

Source	Description	Link
PubMed	12.5M biomedical abstracts (2000-2024)	pubmed.ncbi.nlm.nih.gov
PubTator3	Pre-extracted entities & relationships	ncbi.nlm.nih.gov/research/pubtator3
NCBI Gene	Gene symbols and metadata	ncbi.nlm.nih.gov/gene
MeSH	Medical Subject Headings (diseases, chemicals)	meshb.nlm.nih.gov
NCBI Taxonomy	Species classification	ncbi.nlm.nih.gov/taxonomy
Cellosaurus	Cell line database	cellosaurus.org
dbSNP	Mutation/variant identifiers	ncbi.nlm.nih.gov/snp

Citation

@dataset{pubmed_graphrag_2025,
  author = {Roxas, Danny},
  title = {PubMed-GraphRAG: A Large-scale Biomedical Knowledge Graph Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/dannyroxas/pubmed-graphrag-data}}
}

License

This dataset is derived from publicly available biomedical databases:

• PubMed abstracts (NLM/NCBI)
• PubTator3 annotations (NCBI)
• MeSH descriptors (NLM)
• NCBI Gene and Taxonomy databases

Released under CC-BY-4.0 for academic research purposes.

Acknowledgments

• National Center for Biotechnology Information (NCBI)
• National Library of Medicine (NLM)
• Meta AI - FAISS
• Neo4j