README.md

---
license: mit
task_categories:
- named-entity-recognition
- named-entity-linking
- text-classification
language:
- en
tags:
- biology
- proteins
- genes
- uniprot
- named-entity-linking
- biomedical
- life-sciences
size_categories:
- 10K<n<100K
---

# SODA-SPROUT: Role-Filtered Named Entity Linking Dataset

## Dataset Description

This dataset is an improved version of the SODA-SPROUT NEL dataset, specifically filtered to include only the most relevant biological entities for Named Entity Linking tasks. The dataset focuses on proteins and genes with roles of **'assayed'** and **'intervention'**, which represent the core biological entities that are actually measured or manipulated in scientific experiments.

### Key Improvements

- **Role Filtering**: Only includes entities with roles 'assayed' (104,286 entities) and 'intervention' (80,127 entities)
- **Higher Quality**: Removes less relevant entities like 'reporter', 'component', 'normalizing', and 'experiment' roles
- **Better NEL Performance**: Focuses on the most important biological entities for linking tasks
- **Reduced Noise**: 16,532 figure-level records (vs. 21,000+ in unfiltered version)

## Dataset Structure

### Statistics
- **Total Records**: 16,532 figure-level records
- **Training Set**: 14,880 records (90%)
- **Validation Set**: 825 records (5%)
- **Test Set**: 827 records (5%)
- **Total Entities**: 184,413 (104,286 assayed + 80,127 intervention)
- **Unique Papers**: 3,101 papers

### Data Format

Each record contains:
- **Paper Context**: PMID, DOI, title, abstract, year
- **Figure Context**: Figure label, caption, figure URL
- **Entities**: List of gene/protein entities with:
  - `text`: Entity mention text
  - `type`: "geneprod" (for NER compatibility)
  - `original_type`: "gene" or "protein" (from SourceData)
  - `role`: "assayed" or "intervention"
  - `ext_tax_names`: Organism taxonomy name
  - `uniprot_ids`: Ground truth UniProt accessions
  - `mapping_status`: "mapped" or "unmapped"

## Usage

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("EMBO/soda-nel")

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

# Example record
record = train_data[0]
print(f"Paper: {record['title']}")
print(f"Figure: {record['fig_label']}")
print(f"Caption: {record['caption']}")
print(f"Entities: {len(record['entities'])}")
```

## Role Definitions

- **Assayed**: Proteins/genes that were actually measured, quantified, or tested in the experiment
- **Intervention**: Proteins/genes that were manipulated, targeted, or used as experimental interventions

## Data Sources

- **Source Dataset**: [EMBO/SourceData](https://huggingface.co/datasets/EMBO/SourceData)
- **Original Paper**: [SourceData: A platform for the large-scale curation of unstructured data from the scientific literature](https://arxiv.org/abs/2310.20440)
- **UniProt Mapping**: NCBI Gene IDs mapped to UniProt accessions via UniProt REST API

## Citation

If you use this dataset, please cite both the original SourceData paper and this improved version:

```bibtex
@article{source_data_2023,
  title={SourceData: A platform for the large-scale curation of unstructured data from the scientific literature},
  author={...},
  journal={...},
  year={2023},
  url={https://arxiv.org/abs/2310.20440}
}
```

## License

This dataset is released under the MIT License. The underlying SourceData is available under its original license.

## Updates

- **v1.1.0**: Added role filtering for 'assayed' and 'intervention' entities only
- **v1.0.0**: Initial release with all entity roles