ScRAT: Single-cell RNA (scRNA-seq) ATtention Network for clinical phenotype prediction

The official implementation of Phenotype prediction from single-cell RNA-seq data using Attention-Based neural networks (accepted at Bioinformatics).

Table of contents

1. Setup
2. Input Data Format
3. Test Run
4. Contact

Setup (python>=3.7)

git clone https://github.com/yuzhenmao/ScRAT
cd ScRAT
python -m venv scrat
source scrat/bin/activate
pip install -r requirements.txt

Input Data Format

ScRAT requires the following information. 1) scRNA-seq sample Matrices (row major) 2) Metadata for cells: patient id, phenotype labels, cell population

• For dataset mentioned in our paper:

  Please download datasets from: https://figshare.com/projects/ScRAT_Early_Phenotype_Prediction_From_Single-cell_RNA-seq_Data_using_Attention-Based_Neural_Networks/151659, and put the datasets under ScRAT/data. The folder should have the following content:

  
  ScRAT/data/
  ├── SC4
  │   ├── cell_type_large.pkl
  │   ├── cell_type.pkl
  │   ├── covid_pca.npy
  │   ├── patient_id.pkl
  │   ├── severity_label.pkl
  │   └── stage_label.pkl
  ├── Haniffa
  │   ├── cell_type_large.pkl
  │   ├── cell_type.pkl
  │   ├── Haniffa_X_pca.npy
  │   ├── patient_id.pkl
  │   └── labels.pkl
  ├── COMBAT
  │   ├── cell_type_large.pkl
  │   ├── cell_type.pkl
  │   ├── COMBAT_X_pca.npy
  │   ├── patient_id.pkl
  │   └── labels.pkl

(You can also get the raw data from the original papers listed in our paper.)

* ###  For customized dataset:
Please pack the dataset in the h5ad format and set the value of args.dataset to the path of the dataset. Also, please set args.task to 'custom'.

Furthermore, in the dataloader.py file, please modify the following lines: 
  1) line 178 for label dictionary to map string to integer (default: {})
  2) line 185 for patient id (default: data.obs['patient_id'])
  3) line 187 for label, which is clinical phenotype for prediction (default: data.obs['Outcome']) 
  4) line 189 for cell type, which assist for mixup (default: data.obs['cell_type'])

Also in the main.py file, please add `elif args.task == 'custom': label_dict = {0: 'XXX', 1: 'XXX'}` in line 83.

## Test Run
### Demo

bash run.sh

You can modify the hyper-parameters in `run.sh`.
- task: `stage`, `severity`, `haniffa`, `combat`, `custom`

### Output Template
`out.txt`

# Reference

Please cite the following paper if you found this library useful in your research:

### [Phenotype prediction from single-cell RNA-seq data using Attention-Based neural networks](https://academic.oup.com/bioinformatics/article/40/2/btae067/7613064)
[Yuzhen Mao](https://scholar.google.com/citations?user=9wKn1A0AAAAJ&hl=en), Yen-Yi Lin, Nelson KY Wong, Stanislav Volik, Funda Sar, [Colin Collins](https://collinslab.med.ubc.ca), [Martin Ester](https://sites.google.com/view/esterlab)\
*Bioinformatics*, 2024

@article{mao2024phenotype, title={Phenotype prediction from single-cell RNA-seq data using attention-based neural networks}, author={Mao, Yuzhen and Lin, Yen-Yi and Wong, Nelson KY and Volik, Stanislav and Sar, Funda and Collins, Colin and Ester, Martin}, journal={Bioinformatics}, pages={btae067}, year={2024}, publisher={Oxford University Press} }