README.md

---
license: mit
task_categories:
- tabular-classification
- biology
tags:
- single-cell
- scRNA-seq
- deep-learning
pretty_name: scLightGAT Data
---

# scLightGAT Data

This repository contains the training and testing datasets for **scLightGAT**: A range-constrained Graph Attention Network for single-cell clustering and annotation.

These files are structured to be compatible with the [scLightGAT project](https://github.com/TBD).

## Dataset Structure

The dataset contains processed `.h5ad` files organized for the scLightGAT pipelines.

- **`Integrated_training/`**: Contains `train.h5ad`, the large-scale reference training set used for the DVAE and GAT models.
- **`Independent_testing/`**: Contains independent datasets used for benchmarking and inference (e.g., `sapiens_full`, `lung_full`, `GSE115978`, etc.).
- **`caf.data/`**: Additional data specific to Cancer-Associated Fibroblasts (CAF) experiments.

### Directory Layout
When downloaded, the data should follow this structure to work with `run_sclight.gat.sh`:

```
scLightGAT_data/
├── Integrated_training/
│   └── train.h5ad
├── Independent_testing/
│   ├── GSE115978.h5ad
│   ├── GSE123139.h5ad
│   ├── GSE153935.h5ad
│   ├── GSE166555.h5ad
│   ├── Zhengsorted.h5ad
│   ├── lung_full.h5ad
│   └── sapiens_full.h5ad
└── caf.data/
    ├── caf_train.h5ad
    └── caf_test.h5ad
```

## How to Use

### 1. Automated Download (Recommended)
You can use the `download_hf_data.sh` script provided in the scLightGAT repository to automatically fetch and place this data.

### 2. Manual Download
If you are manually setting up the project, download all files from this repository and place them in a directory named `scLightGAT_data` inside your project's `data/` folder.

**Project Structure Example:**
```
scLightGAT_Project/
├── scLightGAT.main/       # Code repository
│   ├── run_sclight.gat.sh
│   └── ...
└── data/
    └── scLightGAT_data/   # This dataset (Downloaded here)
        ├── Integrated_training/
        ├── Independent_testing/
        └── caf.data/
```

### Python Access
You can also access the files directly via `huggingface_hub`:

```python
from huggingface_hub import hf_hub_download
import scanpy as sc

# Example: Load the training data
file_path = hf_hub_download(
    repo_id="Alfiechuang/scLightGAT", 
    filename="Integrated_training/train.h5ad",
    repo_type="dataset"
)
adata = sc.read_h5ad(file_path)
```