---
license: mit
tags:
- biology
---

# SwissIPG
> SwissProt-based Protein Function (InterPro and Gene Ontology) Benchmark Datasets

This repository provides benchmark datasets for protein function evaluation, constructed from **UniProt/SwissProt**, **InterPro (IPR)**, and **Gene Ontology (GO)**.  
It includes both test sets and training sets, supporting **keyword-guided** and **description-guided** protein design/evaluation tasks.


## 📖 Dataset Overview

**Keyword-guided protein tasks** previously lacked a **publicly available, high-quality evaluation dataset**.  To address this gap, we curated a novel dataset from SwissProt, together with their **IPR entries** and **GO terms**.  Note that the GO terms are restricted to **Molecular Function (MF)**.

| Subset       | Num of Proteins | Num of IPR entries | Num of GO terms |
| ------------ | --------------- | ------------------ | --------------- |
| Test Set     | 1,057           | 1,297              | 380             |
| Train Set    | 555543          | 39431              | 8110            |

## 🔧 Dataset Construction

1. Protein selection
Extracted proteins from UniProt/SwissProt released between `2025-01-01` and `2025-08-25` for testing. The remaining proteins are used for training.

2. Keyword collection
Retrieved corresponding protein sequences, InterPro IDs, and GO terms.  

3. Instruction construction for description-guided task
Following [Mol-Instructions](https://github.com/zjunlp/Mol-Instructions), we concatenated text descriptions of InterPro entries and GO terms to form natural-language prompts.  

| Keyword                               | Description                                                                                     |
|---------------------------------------|-------------------------------------------------------------------------------------------------|
| InterPro (Domain)                     | The protein should contain one or more *{}* that are essential for its biological function       |
| InterPro (Family)                     | The protein should belong to *{}* that shares evolutionary origin and functional similarity      |
| InterPro (Homologous_Superfamily)     | The protein should be classified within *{}* sharing conserved structural features               |
| InterPro (Repeat)                     | The protein should include one or more *{}* that provide structural or functional support        |
| InterPro (Conserved_Site)             | The protein should contain *{}* that is preserved across related proteins                        |
| InterPro (Active_Site)                | The protein must have *{}* that is conserved among related catalytic enzymes                     |
| InterPro (Binding_Site)               | The protein should include a *{}* that enables ligand binding under diverse conditions           |
| InterPro (PTM)                        | The protein should contain *{}* that allow regulation through chemical modifications             |
| Gene Ontology (molecular function)    | The protein must be able to perform the *{}* required for its activity.                          |

4. Final curation
Curated into **test** and **train** sets with consistent formats.  


## 📂 Dataset Structure

| File                | Func-Seq Pairs | Description                |
| ------------------- | -------------- | -------------------------- |
| `go_test.json`      | 693            | Test set for GO task.      |
| `go_train.json`     | 492124         | Train set for GO task.     |
| `ipr_test.json`     | 870            | Test set for IPR task.     |
| `ipr_train.json`    | 555543         | Train set for IPR task.    |
| `ipr_go_test.json`  | 674            | Test set for IPR&GO task.  |
| `ipr_go_train.json` | 487848         | Train set for IPR&GO task. |
| `SwissIPG.py`       | -              | Building Class             |

## 🎯 Intended Use
- **Benchmarking protein design models** under keyword-guided and description-guided settings
- **Training and fine-tuning** models using consistent annotation standards
- Facilitating **fair comparison** across models through a unified dataset 

## 🔍 Evaluation Metrics

Our work, [PDFBench](https://github.com/PDFBench/PDFBench), provides a comprehensive evaluation for both tasks. For further details, please refer to the repository.

## 📑 Citation

If you use this dataset, please cite:

```bibtex
@misc{kuang2025pdfbenchbenchmarknovoprotein,
      title={PDFBench: A Benchmark for De novo Protein Design from Function}, 
      author={Jiahao Kuang and Nuowei Liu and Changzhi Sun and Tao Ji and Yuanbin Wu},
      year={2025},
      eprint={2505.20346},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.20346}, 
}
```