README.md

---
task_categories:
- text-classification
tags:
- biology
- chemistry
- peptide-ml
---

# PepBenchData_raw

This repository provides the **raw dataset** collected for **PepBenchmark: A Standardized Benchmark for Peptide Machine Learning**.

[**Paper**](https://huggingface.co/papers/2604.10531) | [**GitHub**](https://github.com/ZGCI-AI4S-Pep/PepBenchmark/) | [**Processed Dataset (PepBenchData)**](https://huggingface.co/datasets/jiahuizhang/PepBenchData)

---

## Difference Between `PepBenchData_raw` and `PepBenchData`

We provide two related but distinct datasets:

### 1. PepBenchData_raw (This Repository)

- Contains **raw peptide sequences** collected from various databases and published literature.
- Most samples are **experimentally validated**, ensuring high reliability.
- No extensive preprocessing or feature engineering is applied.
- Recommended for:
  - Generative modeling
  - Pretraining tasks
  - Studies requiring high-confidence biological sequences

---

### 2. PepBenchData

- A **processed dataset** designed for machine learning tasks.
- Includes:
  - Feature representations
  - Standardized train/validation/test splits
- Construction involves several preprocessing steps:
  - Data filtering
  - Redundancy removal among positive samples (e.g., via MMseqs2)
  - Negative sample construction and sampling

**Important Note:**  
During negative sample construction, some sequences are sampled from **UniRef50**.  
As a result, **not all negative samples are experimentally validated**.

---

## Sample Usage

You can use the `pepbenchmark` library to manage and load these datasets. Below is an example of how to load a dataset and its official splits:

```python
from pepbenchmark.dataset_manager.single_dataset import SinglePeptideDatasetManager

# Initialize the manager for a specific task (e.g., ACE inhibitory)
manager = SinglePeptideDatasetManager(
    "ace_inhibitory",
    official_feature_names=["fasta", "label"],
    dataset_dir="path/to/PepBenchData",
)

# Access sequences and labels
sequences = manager.get_feature("fasta")
labels = manager.get_feature("label")

# Load official splits
splits = manager.set_official_split_indices(
    split_type="hybrid_split",
    fold_seed=0
)

print(f"Train samples: {len(splits['train'])}")
print(f"Test samples: {len(splits['test'])}")
```

---

## Data Sources

The raw dataset is aggregated from multiple sources, including public biological databases and published scientific literature.

![Data Sources](https://huggingface.co/datasets/jiahuizhang/PepBenchData_raw/resolve/main/data_source.png)

---

## Dataset Statistics

### Classification Tasks

- `pos.csv`: Experimentally validated **positive samples**
- `neg.csv`: Experimentally validated **negative samples** (Note: For many datasets, `neg.csv` is not available.)
- `pos_filter_id_0.9_cov_f0.9.csv`: Positive samples after redundancy removal using **MMseqs2** (identity ≥ 0.9, coverage ≥ 0.9).

![Classification Statistics](https://huggingface.co/datasets/jiahuizhang/PepBenchData_raw/resolve/main/cls.png)

### Regression Tasks

Contains datasets with continuous labels for peptide-related properties.

![Regression Statistics](https://huggingface.co/datasets/jiahuizhang/PepBenchData_raw/resolve/main/reg.png)

---

## Directory Structure (Classification Datasets)

```text
dataset_name/
├── pos.csv
├── neg.csv (optional)
└── pos_filter_id_0.9_cov_f0.9.csv
```

## Citation

If you use PepBenchmark, please cite:

```bibtex
@inproceedings{zhang2026pepbenchmark,
  title={PepBenchmark: A Standardized Benchmark for Peptide Machine Learning},
  author={Zhang, Jiahui and Wang, Rouyi and Zhou, Kuangqi and Xiao, Tianshu and Zhu, Lingyan and Min, Yaosen and Wang, Yang},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026},
  url={https://openreview.net/forum?id=NskQgtSdll}
}
```