README.md

---
license: apache-2.0
task_categories:
- text-classification
- feature-extraction
language:
- en
tags:
- pharmacogenomics
- biomedical
- variant-drug-associations
- literature-mining
- genomics
size_categories:
- 1K<n<10K
source_datasets:
- original
multilinguality:
- monolingual
pretty_name: AutoGKB Annotation Benchmark
dataset_info:
  features:
  - name: pmcid
    dtype: string
  - name: article_title
    dtype: string
  - name: article_path
    dtype: string
  - name: article_text
    dtype: string
  - name: variant_annotation_id
    dtype: int64
  - name: variant_haplotypes
    dtype: string
  - name: gene
    dtype: string
  - name: drugs
    dtype: string
  - name: pmid
    dtype: int64
  - name: phenotype_category
    dtype: string
  - name: significance
    dtype: string
  - name: notes
    dtype: string
  - name: sentence
    dtype: string
  - name: alleles
    dtype: string
  - name: specialty_population
    dtype: string
  - name: metabolizer_types
    dtype: string
  - name: is_plural
    dtype: string
  - name: is_is_not_associated
    dtype: string
  - name: direction_of_effect
    dtype: string
  - name: pd_pk_terms
    dtype: string
  - name: multiple_drugs_and_or
    dtype: string
  - name: population_types
    dtype: string
  - name: population_phenotypes_or_diseases
    dtype: string
  - name: multiple_phenotypes_or_diseases_and_or
    dtype: string
  - name: comparison_alleles_or_genotypes
    dtype: string
  - name: comparison_metabolizer_types
    dtype: string
  splits:
  - name: train
    num_examples: 3124
  - name: validation
    num_examples: 796
  - name: test
    num_examples: 596
---

# AutoGKB Annotation Benchmark

## Dataset Description

The AutoGKB Annotation Benchmark is a comprehensive dataset designed to evaluate models' ability to extract pharmacogenomic variant-drug associations from scientific literature. This ground truth values for this data were compiled by reviewers from [PharmGKB](https://www.pharmgkb.org/). This benchmark addresses the critical need for automated systems that can identify genetic variants, associated drugs, and their clinical relationships from biomedical texts.

### Dataset Summary

This dataset contains **4,516 annotations** extracted from **1,431 unique scientific papers** (PMIDs), covering a wide range of pharmacogenomic relationships. Each annotation includes detailed information about genetic variants, drugs, phenotype categories, population specifics, and statistical associations. The goal of this dataset is to benchmark the process of extracting information a PubMed paper on pharmacogenomics. The annotation system should be able to understand and extract the key information that is represented in the dataset.

### Languages

The dataset is in English (en).

## Dataset Structure

### Data Instances

Each example contains:
- **Core annotation fields**: Variant, gene, drug, PMID, phenotype category
- **Association details**: Significance, direction of effect, comparison data
- **Population information**: Specialty populations, metabolizer types
- **Full text**: Complete scientific article in markdown format

Example:
```json
{
  "pmcid": "PMC6714673",
  "article_title": "Warfarin Dose Model for the Prediction of Stable Maintenance Dose in Indian Patients",
  "article_path": "articles/PMC6714673.md",
  "article_text": "# Warfarin Dose Model for the Prediction of Stable Maintenance Dose in Indian Patients\n\n## Abstract\n\nWarfarin is a commonly used anticoagulant...",
  "variant_annotation_id": 1449192282,
  "variant_haplotypes": "rs1799853",
  "gene": "CYP2C9",
  "drugs": "warfarin",
  "pmid": 28049362,
  "phenotype_category": "Dosage",
  "significance": "yes",
  "sentence": "Genotype CT is associated with decreased dose of warfarin as compared to genotype CC.",
  "alleles": "CT",
  "is_is_not_associated": "Associated with",
  "direction_of_effect": "decreased",
  "pd_pk_terms": "dose of",
  "comparison_alleles_or_genotypes": "CC"
}
```

### Data Fields

#### Core Fields
- `pmcid`: PubMed Central identifier of the source article
- `article_title`: Title of the source scientific article
- `article_path`: Relative path to the article file (markdown format)
- `article_text`: Full text of the scientific article in markdown format
- `variant_annotation_id`: Unique identifier for each annotation
- `variant_haplotypes`: Genetic variant identifier (e.g., rs numbers, haplotypes)
- `gene`: Gene symbol (e.g., CYP2D6, ABCB1)
- `drugs`: Drug name(s) involved in the association
- `pmid`: PubMed identifier of the source article

#### Phenotype Information
- `phenotype_category`: Type of effect (Efficacy, Toxicity, Dosage, Metabolism/PK, etc.)
- `significance`: Statistical significance (yes/no/not stated)
- `sentence`: Key sentence describing the association
- `notes`: Additional context or study details

#### Association Details
- `is_is_not_associated`: Whether variant is associated with outcome
- `direction_of_effect`: Direction of association (increased/decreased)
- `pd_pk_terms`: Pharmacodynamic/pharmacokinetic terms
- `alleles`: Specific alleles involved

#### Population Context
- `specialty_population`: Specific patient populations
- `population_types`: General population categories
- `population_phenotypes_or_diseases`: Diseases or conditions
- `metabolizer_types`: CYP metabolizer classifications

#### Comparison Data
- `comparison_alleles_or_genotypes`: Reference genotypes for comparison
- `comparison_metabolizer_types`: Reference metabolizer types


### Data Splits

| Split | Annotations | Unique Papers |
|-------|-------------|---------------|
| Train | 3,124 | 1,001 |
| Validation | 796 | 215 |
| Test | 596 | 215 |

**Total**: 4,516 annotations across 1,431 papers

## Dataset Creation

### Curation Rationale

This benchmark was created to address the growing need for automated pharmacogenomic knowledge extraction. With the rapid expansion of pharmacogenomic literature, manual curation becomes increasingly challenging. This dataset provides a standardized evaluation framework for developing and comparing automated extraction systems.

### Source Data

#### Initial Data Collection and Normalization

The dataset is derived from peer-reviewed scientific publications in the pharmacogenomics domain. Articles were selected based on their content related to genetic variant-drug associations and clinical outcomes.

#### Who are the source language producers?

The source texts are scientific articles authored by researchers in pharmacogenomics, clinical pharmacology, and related biomedical fields, published in peer-reviewed journals.

### Annotations

#### Annotation process

Annotations were created by domain experts following a comprehensive schema covering:
- Genetic variant identification and standardization
- Drug name normalization
- Phenotype categorization using controlled vocabularies
- Population and study context extraction
- Statistical association characterization

#### Who are the annotators?

The annotations were created by experts in pharmacogenomics and biomedical informatics with specialized knowledge in genetic variant-drug associations.

### Personal and Sensitive Information

The dataset contains only published scientific literature and does not include personal or sensitive information about individuals.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset supports the development of automated systems for pharmacogenomic knowledge extraction, which can:
- **Accelerate precision medicine**: Enable faster identification of clinically relevant variant-drug associations
- **Support clinical decision-making**: Facilitate evidence-based prescribing decisions
- **Advance research**: Enable large-scale analysis of pharmacogenomic literature

### Discussion of Biases

Potential biases in the dataset may include:
- **Publication bias**: Overrepresentation of statistically significant results
- **Population bias**: Uneven representation of different ethnic populations in source studies
- **Drug bias**: Focus on commonly studied drugs and variants
- **Temporal bias**: Emphasis on more recent research publications

### Other Known Limitations

- **Coverage**: Represents approximately 33.6% of original pharmacogenomic annotations from the source database
- **Language**: Limited to English-language publications
- **Domain scope**: Focused specifically on pharmacogenomics, may not generalize to other biomedical domains
- **Text quality**: Depends on the quality of PDF-to-text conversion for source articles

## Additional Information

### Dataset Curators

AutoGKB Team

### Licensing Information

This dataset is released under the Apache License 2.0.

### Citation Information

```bibtex
@misc{autogkb_annotation_benchmark_2025,
  title={AutoGKB Annotation Benchmark},
  author={Shlok Natarajan},
  year={2025},
  note={A benchmark for pharmacogenomic variant-drug annotation extraction from scientific literature}
}
```

### Contributions

This dataset contributes to the biomedical NLP community by providing:
- A standardized benchmark for pharmacogenomic information extraction
- High-quality annotations with detailed schema
- Full-text articles paired with structured annotations
- Evaluation metrics and baseline models for comparison

## Usage Examples

### Loading the Dataset

```python
from datasets import load_dataset

# Load the dataset from Hugging Face Hub
dataset = load_dataset("autogkb/autogkb-annotation-benchmark")

# Access different splits
train_data = dataset["train"]
val_data = dataset["validation"] 
test_data = dataset["test"]

# Example: Get all efficacy-related annotations
efficacy_examples = train_data.filter(
    lambda x: "Efficacy" in x["phenotype_category"]
)

# Example: Access article text for a specific annotation
first_example = train_data[0]
print(f"PMC ID: {first_example['pmcid']}")
print(f"Article Title: {first_example['article_title']}")
print(f"Gene: {first_example['gene']}")
print(f"Drug: {first_example['drugs']}")
print(f"Full Article Text: {first_example['article_text'][:500]}...")
```

### Evaluation

The dataset includes evaluation scripts for measuring:
- Field-level exact match accuracy
- Overall accuracy across core fields
- Phenotype-specific performance

```bash
# Run baseline model
python baseline_model.py val baseline_predictions.tsv

# Evaluate predictions
python evaluate.py baseline_predictions.tsv val/annotations.tsv --output results.json
```

## File Structure

```
autogkb/
├── articles/             # Full article texts in markdown format
│   ├── PMC10038974.md
│   ├── PMC10085626.md
│   └── ...              # 1,431 articles total
├── train.jsonl          # Training annotations (3,124 examples)
├── val.jsonl            # Validation annotations (796 examples)
├── test.jsonl           # Test annotations (596 examples)
├── autogkb_annotation_benchmark.py  # HuggingFace dataset script
├── dataset_infos.json   # Dataset metadata
├── dataset_statistics.json  # Dataset statistics
├── LICENSE              # Apache 2.0 license
└── README.md           # This file
```