README.md

---
task_categories:
- text-generation
language:
- en
tags:
- proteins
- biology
- uniprot
size_categories:
- 100K<n<1M
license: mit
---

<h1 align="center">Dataset Card for PAIR</h1>

 <p align="center">
 <a href="https://www.biorxiv.org/content/10.1101/2024.07.22.604688v2.abstract"><img src="https://img.shields.io/badge/bioRxiv-2024.07.22.604688-red?style=for-the-badge&logo=bioRxiv" alt="bioRxiv"/></a>
</p>

This dataset contains all the text annotations we collected and parsed from UniProt Swiss-Prot February 2023 and used to train PAIR from the paper "Boosting the Predictive Power of Protein Representations with a Corpus of Text Annotations". You can read more details about PAIR [here](https://www.biorxiv.org/content/10.1101/2024.07.22.604688v2.abstract).

 <p align="center">
<img src="https://huggingface.co/datasets/mskrt/PAIR/resolve/main/pair_data_fig.png" alt="drawing" style="width:500px;">
</p>

# Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->


### Dataset Sources

- [**Repository**](https://github.com/h4duan/PAIR)
- [**Pre-print**](https://www.biorxiv.org/content/10.1101/2024.07.22.604688v2.abstract)
- [**Model checkpoints**](https://huggingface.co/h4duan)

## Uses

<!-- Address questions around how the dataset is intended to be used. -->

### Example usage

```
from datasets import load_dataset
data = load_dataset("mskrt/PAIR", annotation_type="function", trust_remote_code=True)
```

where `annotation_type` is one of the 19 annotation types we considered in our work. Here is a list of all the possible annotation types you can load: `['function', 'active_sites', 'activity_regulation'...]`

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->

This dataset contains text annotations from Swiss-Prot February 2023; our models were trained on all of them. Please be mindful about potential data leakage from time splits/identical protein sequences on any downstream tasks in your setup.

## Dataset Structure

<!-- This section provides a description of the dataset fields, and additional information about the dataset structure such as criteria used to create the splits, relationships between data points, etc. -->

[Coming soon]

## Dataset Creation

### Source Data

This data was collected from the Swiss-Prot checkpoint from February 2023, found [here](https://ftp.uniprot.org/pub/databases/uniprot/previous_major_releases/release-2023_02/knowledgebase/).

#### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

To see how we parsed our data, select an annotation type folder from [this link](https://github.com/h4duan/PAIR/tree/main/_fact) and open the `parser.py` script. 

#### Who are the source data producers?

<!-- This section describes the people or systems who originally created the data. It should also include self-reported demographic or identity information for the source data creators if this information is available. -->

The data was originally produced by the [Uniprot consortium](https://www.uniprot.org/).


#### Personal and Sensitive Information

To our knowledge, this dataset does not contain any private information.

## Bias, Risks, and Limitations

In general, the dataset is highly imbalanced in terms of how many and what protein sequences in Swiss-Prot have an annotation for a given annotation type. This dataset is sparse


## Citation

**BibTeX:**

```
@article{duan2024boosting,
  title={Boosting the Predictive Power of Protein Representations with a Corpus of Text Annotations},
  author={Duan, Haonan and Skreta, Marta and Cotta, Leonardo and Rajaonson, Ella Miray and Dhawan, Nikita and Aspuru-Guzik, Alán and Maddison, Chris J},
  journal={bioRxiv},
  pages={2024--07},
  year={2024},
  publisher={Cold Spring Harbor Laboratory}
}
```

## Dataset Card Contact

For any issues with this dataset, please contact `[email protected]` or `[email protected]`