Scientific Biomedical PubTator
PubTator3 API を活用したバイオメディカル文献エンティティ認識・ 関係抽出パイプラインを提供する。
When to Use
- •PubMed 論文から遺伝子・疾患・化合物のエンティティを自動抽出するとき
- •バイオメディカル NER (Named Entity Recognition) を実行するとき
- •遺伝子-疾患・薬物-標的の関係を文献から抽出するとき
- •大規模文献コーパスのバイオアノテーションを行うとき
- •テキストマイニング結果を知識グラフに統合するとき
Quick Start
1. PubTator3 エンティティアノテーション
python
import requests
import pandas as pd
import json
import time
PUBTATOR_API = "https://www.ncbi.nlm.nih.gov/research/pubtator3-api"
def annotate_pmids(pmids, concepts=None):
"""
PubTator3 — PMID リストのバイオメディカルエンティティアノテーション。
Parameters:
pmids: list — PMID リスト (e.g., [12345678, 23456789])
concepts: list — エンティティタイプ
"gene", "disease", "chemical", "mutation", "species", "cellline"
ToolUniverse:
PubTator_annotate(pmids=pmids, concepts=concepts)
PubTator_search(query=query)
"""
if concepts is None:
concepts = ["gene", "disease", "chemical", "mutation", "species"]
pmid_str = ",".join(str(p) for p in pmids)
params = {
"pmids": pmid_str,
"concepts": ",".join(concepts),
"format": "biocjson",
}
resp = requests.get(f"{PUBTATOR_API}/publications/export/biocjson", params=params)
resp.raise_for_status()
data = resp.json()
# Parse annotations
all_annotations = []
for doc in data.get("PubTator3", []) if isinstance(data, dict) else [data]:
pmid = doc.get("pmid", doc.get("id", ""))
for passage in doc.get("passages", []):
for annotation in passage.get("annotations", []):
infons = annotation.get("infons", {})
all_annotations.append({
"pmid": pmid,
"text": annotation.get("text", ""),
"type": infons.get("type", ""),
"identifier": infons.get("identifier", ""),
"offset": annotation.get("locations", [{}])[0].get("offset", ""),
"length": annotation.get("locations", [{}])[0].get("length", ""),
"passage_type": passage.get("infons", {}).get("type", ""),
})
df = pd.DataFrame(all_annotations)
type_counts = df["type"].value_counts().to_dict() if not df.empty else {}
print(f"PubTator annotation: {len(pmids)} PMIDs → "
f"{len(df)} entities {type_counts}")
return df
2. PubTator3 テキスト検索
python
def search_pubtator(query, max_results=100):
"""
PubTator3 テキスト検索 — バイオメディカルエンティティ付き論文検索。
Parameters:
query: str — 検索クエリ (遺伝子名、疾患名、化合物名)
max_results: int — 最大取得数
"""
params = {
"text": query,
"sort": "score",
"page_size": min(max_results, 100),
}
resp = requests.get(f"{PUBTATOR_API}/search/", params=params)
resp.raise_for_status()
data = resp.json()
results = []
for hit in data.get("results", []):
results.append({
"pmid": hit.get("pmid", ""),
"title": hit.get("title", ""),
"journal": hit.get("journal", ""),
"year": hit.get("year", ""),
"score": hit.get("score", 0),
"genes": hit.get("genes", []),
"diseases": hit.get("diseases", []),
"chemicals": hit.get("chemicals", []),
"mutations": hit.get("mutations", []),
})
df = pd.DataFrame(results)
total = data.get("count", 0)
print(f"PubTator search '{query}': {total} total, {len(df)} returned")
return df
3. エンティティ関係抽出
python
def extract_entity_relations(pmids, relation_types=None):
"""
PubTator3 — エンティティ間関係 (gene-disease, drug-target 等) 抽出。
Parameters:
pmids: list — PMID リスト
relation_types: list — 関係タイプフィルタ
"GDA" (gene-disease), "CDA" (chemical-disease),
"CGA" (chemical-gene), "PPI" (protein-protein)
"""
# Get annotations with relations
df_annotations = annotate_pmids(pmids)
# Extract co-occurrences within same passage
relations = []
for pmid in df_annotations["pmid"].unique():
pmid_df = df_annotations[df_annotations["pmid"] == pmid]
# Gene-Disease relations
genes = pmid_df[pmid_df["type"] == "Gene"]
diseases = pmid_df[pmid_df["type"] == "Disease"]
chemicals = pmid_df[pmid_df["type"] == "Chemical"]
if not relation_types or "GDA" in relation_types:
for _, gene in genes.iterrows():
for _, disease in diseases.iterrows():
relations.append({
"pmid": pmid,
"relation_type": "GDA",
"entity1_type": "Gene",
"entity1_text": gene["text"],
"entity1_id": gene["identifier"],
"entity2_type": "Disease",
"entity2_text": disease["text"],
"entity2_id": disease["identifier"],
})
if not relation_types or "CGA" in relation_types:
for _, chem in chemicals.iterrows():
for _, gene in genes.iterrows():
relations.append({
"pmid": pmid,
"relation_type": "CGA",
"entity1_type": "Chemical",
"entity1_text": chem["text"],
"entity1_id": chem["identifier"],
"entity2_type": "Gene",
"entity2_text": gene["text"],
"entity2_id": gene["identifier"],
})
if not relation_types or "CDA" in relation_types:
for _, chem in chemicals.iterrows():
for _, disease in diseases.iterrows():
relations.append({
"pmid": pmid,
"relation_type": "CDA",
"entity1_type": "Chemical",
"entity1_text": chem["text"],
"entity1_id": chem["identifier"],
"entity2_type": "Disease",
"entity2_text": disease["text"],
"entity2_id": disease["identifier"],
})
rel_df = pd.DataFrame(relations)
rel_counts = rel_df["relation_type"].value_counts().to_dict() if not rel_df.empty else {}
print(f"Entity relations: {len(rel_df)} total {rel_counts}")
return rel_df
4. バイオアノテーション集計ダッシュボード
python
def annotation_summary_dashboard(pmids, output_prefix="pubtator"):
"""
PubTator アノテーション集計・可視化。
Parameters:
pmids: list — PMID リスト
output_prefix: str — 出力ファイルプレフィックス
"""
import matplotlib.pyplot as plt
# Get annotations
df = annotate_pmids(pmids)
if df.empty:
print("No annotations found")
return {}
# Entity type distribution
fig, axes = plt.subplots(1, 3, figsize=(15, 5))
# 1. Entity type counts
type_counts = df["type"].value_counts()
type_counts.plot(kind="bar", ax=axes[0], color="#2196F3")
axes[0].set_title("Entity Type Distribution")
axes[0].set_ylabel("Count")
# 2. Top entities per type
for entity_type in ["Gene", "Disease", "Chemical"]:
sub = df[df["type"] == entity_type]
top = sub["text"].value_counts().head(10)
if not top.empty:
print(f"\nTop {entity_type}s: {top.to_dict()}")
# 3. Articles per entity count
per_article = df.groupby("pmid")["type"].count()
per_article.hist(ax=axes[1], bins=20, color="#4CAF50")
axes[1].set_title("Entities per Article")
axes[1].set_xlabel("Number of entities")
# Entity type per article
pivot = df.groupby(["pmid", "type"]).size().unstack(fill_value=0)
pivot.plot(kind="box", ax=axes[2])
axes[2].set_title("Entity Types per Article")
plt.tight_layout()
fig_path = f"figures/{output_prefix}_dashboard.png"
plt.savefig(fig_path, dpi=150, bbox_inches="tight")
plt.close()
# Save results
df.to_csv(f"results/{output_prefix}_annotations.csv", index=False)
summary = {
"total_pmids": df["pmid"].nunique(),
"total_annotations": len(df),
"entity_types": type_counts.to_dict(),
"unique_entities": df.groupby("type")["text"].nunique().to_dict(),
}
print(f"\nSummary: {summary}")
return summary
5. 知識グラフ構築用エンティティネットワーク
python
def build_entity_network(pmids, min_cooccurrence=2):
"""
PubTator エンティティ共起ネットワーク構築。
Parameters:
pmids: list — PMID リスト
min_cooccurrence: int — 最小共起回数
"""
import networkx as nx
from collections import Counter
rel_df = extract_entity_relations(pmids)
if rel_df.empty:
return nx.Graph()
# Count co-occurrences
edge_counter = Counter()
for _, row in rel_df.iterrows():
key = tuple(sorted([
f"{row['entity1_type']}:{row['entity1_text']}",
f"{row['entity2_type']}:{row['entity2_text']}",
]))
edge_counter[key] += 1
# Build network
G = nx.Graph()
for (node1, node2), count in edge_counter.items():
if count >= min_cooccurrence:
G.add_edge(node1, node2, weight=count)
print(f"Entity network: {G.number_of_nodes()} nodes, "
f"{G.number_of_edges()} edges "
f"(min cooccurrence = {min_cooccurrence})")
return G
利用可能ツール
| ToolUniverse カテゴリ | 主なツール |
|---|---|
pubtator | PubTator_annotate, PubTator_search |
パイプライン出力
| 出力ファイル | 説明 | 連携先スキル |
|---|---|---|
results/pubtator_annotations.csv | エンティティアノテーション | → text-mining-nlp, knowledge-graph |
results/entity_relations.csv | エンティティ間関係 | → network-analysis, disease-research |
results/entity_network.graphml | エンティティ共起ネットワーク | → graph-neural-networks |
figures/pubtator_dashboard.png | アノテーション集計 | → publication-figures |
パイプライン統合
code
literature-search ──→ biomedical-pubtator ──→ text-mining-nlp
(PubMed/OpenAlex) (PubTator NER) (KG 構築)
│
├──→ disease-research (GDA 関係)
├──→ drug-target-profiling (CGA 関係)
└──→ preprint-archive (プレプリント NER)