OpenMed/drugprot-parquet

DrugProt (Parquet)

A clean, ready-to-use Parquet version of the DrugProt corpus from BioCreative VII Track 1, converted for seamless use with the Hugging Face datasets library.

DrugProt is a gold-standard corpus of 4,250 PubMed abstracts annotated for drug/chemical–protein interactions, covering 13 fine-grained relation types and 3 entity types. It is designed for training and evaluating biomedical relation extraction systems.

Quick Start

from datasets import load_dataset

dataset = load_dataset("OpenMed/drugprot-parquet")

# Access splits
train = dataset["train"]       # 3,500 abstracts
val   = dataset["validation"]  # 750 abstracts

# Inspect a sample
example = train[0]
print(example["title"])
print(f"Entities: {len(example['entities'])}")
print(f"Relations: {len(example['relations'])}")

Dataset Description

Each example represents a PubMed abstract with expert-annotated entity spans and relation labels:

Field	Type	Description
pmid	string	PubMed article ID
title	string	Article title
abstract	string	Article abstract
text	string	Full text (title + abstract)
entities	list[dict]	Annotated entity spans
relations	list[dict]	Annotated drug–protein relations

Entity Schema

Each entity contains:

Field	Type	Description
id	string	Unique entity ID (e.g., T1, T2)
type	string	Entity type: CHEMICAL, GENE-Y, or GENE-N
text	string	Surface text of the entity mention
start	int	Character offset (start) in text field
end	int	Character offset (end) in text field

Entity types:

• CHEMICAL — Drugs, small molecules, metabolites, and other chemical compounds
• GENE-Y — Gene/protein mentions that are a normalized, valid gene/protein
• GENE-N — Gene/protein mentions that are NOT normalized (e.g., protein families, complexes)

Relation Schema

Each relation contains:

Field	Type	Description
type	string	One of 13 relation categories (see below)
arg1	string	Entity ID of the first argument
arg2	string	Entity ID of the second argument

Relation Types (13 Classes)

Relation	Description	Train	Val
INHIBITOR	Chemical inhibits the protein	5,388	1,150
DIRECT-REGULATOR	Chemical directly regulates protein (mechanism unspecified)	2,247	458
SUBSTRATE	Chemical is a substrate of the enzyme	2,003	494
ACTIVATOR	Chemical activates the protein	1,428	246
INDIRECT-UPREGULATOR	Chemical indirectly increases protein activity/expression	1,378	302
INDIRECT-DOWNREGULATOR	Chemical indirectly decreases protein activity/expression	1,329	332
ANTAGONIST	Chemical acts as antagonist of the receptor/protein	972	218
PRODUCT-OF	Chemical is a product of the enzyme	920	158
PART-OF	Chemical is part of the protein complex	885	257
AGONIST	Chemical acts as agonist of the receptor/protein	658	131
AGONIST-ACTIVATOR	Chemical is both agonist and activator	29	10
SUBSTRATE_PRODUCT-OF	Chemical is both substrate and product	24	3
AGONIST-INHIBITOR	Chemical is agonist but inhibits downstream effects	13	2
Total		17,274	3,761

Dataset Statistics

	Train	Validation	Total
Abstracts	3,500	750	4,250
Abstracts with relations	2,433	—	—
Total entities	89,529	—	—
• CHEMICAL	46,274 (51.7%)	—	—
• GENE-Y	28,421 (31.7%)	—	—
• GENE-N	14,834 (16.6%)	—	—
Total relations	17,274	3,761	21,035
Avg. entities / abstract	25.6	—	—
Avg. relations / abstract	4.9	—	—

Usage Examples

Relation Extraction

from datasets import load_dataset

ds = load_dataset("OpenMed/drugprot-parquet", split="train")

for example in ds:
    entities = {e["id"]: e for e in example["entities"]}
    for rel in example["relations"]:
        arg1 = entities[rel["arg1"]]
        arg2 = entities[rel["arg2"]]
        print(f"{arg1['text']} --[{rel['type']}]--> {arg2['text']}")

Named Entity Recognition (NER)

from datasets import load_dataset

ds = load_dataset("OpenMed/drugprot-parquet", split="train")

for example in ds:
    text = example["text"]
    for ent in example["entities"]:
        span = text[ent["start"]:ent["end"]]
        assert span == ent["text"], f"Offset mismatch: '{span}' != '{ent['text']}'"
        print(f"[{ent['type']}] {ent['text']} ({ent['start']}:{ent['end']})")

Convert to Token Classification Format

from datasets import load_dataset

ds = load_dataset("OpenMed/drugprot-parquet", split="train")

# Build BIO tags from character offsets
example = ds[0]
text = example["text"]
char_labels = ["O"] * len(text)
for ent in sorted(example["entities"], key=lambda e: e["start"]):
    tag = ent["type"]
    char_labels[ent["start"]] = f"B-{tag}"
    for i in range(ent["start"] + 1, ent["end"]):
        char_labels[i] = f"I-{tag}"

Source

This dataset is a Parquet conversion of the DrugProt BioCreative VII corpus. The original data was released under CC BY 4.0 for the BioCreative VII shared task.

Original paper:

M. Krallinger, O. Rabal, A. Lourenco, J. Oyarzabal, A. Valencia.
"Overview of the BioCreative VII Track 1 – DrugProt: Drug-Protein Relation Extraction."
Proceedings of the BioCreative VII Challenge Evaluation Workshop, 2021.

BibTeX:

@inproceedings{drugprot2021,
  title={Overview of the BioCreative VII Track 1 -- DrugProt: Drug-Protein Relation Extraction},
  author={Krallinger, Martin and Rabal, Obdulia and Lourenco, Analia and Oyarzabal, Julen and Valencia, Alfonso},
  booktitle={Proceedings of the BioCreative VII Challenge Evaluation Workshop},
  year={2021}
}

License

CC BY 4.0 — following the original DrugProt corpus license.

About OpenMed

OpenMed provides clean, standardized biomedical datasets and RL training environments for medical AI research.