We propose a novel method, SANGO, for accurate single cell annotation by integrating genome sequences around the accessibility peaks within scATAC data.

The official implementation for "SANGO".

Table of Contents

• Datasets
• Installation
• Usage
• Tutorial
• Citation

We provide an easy access to the used datasets in the synapse.

To reproduce SANGO, we suggest first create a conda environment by:

conda create -n SANGO python=3.8
conda activate SANGO

and then run the following code to install the required package:

pip install -r requirements.txt

and then install PyG according to the CUDA version, take torch-1.13.1+cu117 (Ubuntu 20.04.4 LTS) as an example:

pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.13.1+cu117.html

In order to run SANGO, we need to first create anndata from the raw data.

The h5ad file should have cells as obs and peaks as var. There should be at least three columns in var: chr, start, end that indicate the genomic region of each peak. The h5ad file should also contain two columns in the obs: Batch and CellType （reference data）, where Batch is used to distinguish between reference and query data, and CellType indicates the true label of the cell.

Notice that we filter out peaks accessible in < 1% cells for optimal performance.

The processed data are used as input to CACNN and a reference genome is provided to extract the embedding incorporating sequence information:

# Stage 1: embeddings extraction
cd SANGO/CACNN

python main.py -i ../../preprocessed_data/reference_query_example.h5ad \ # input data(after data preprocessing)
               -g mm9 \ # reference genome
               -o ../../output/reference_query_example # output path

Running the above command will generate three output files in the output path:

• CACNN_train.log: recording logs during training
• CACNN_best_model.pt: storing the model weights with the best AUC score during training
• CACNN_output.h5ad: an anndata file storing the embedding extracted by CACNN.

# Stage 2: cell type prediction
cd ../GraphTransformer

python main.py  --data_dir ../../output/reference_query_example/CACNN_output.h5ad \ # input data
                --train_name_list reference --test_name query \
                --save_path ../../output \
                --save_name reference_query_example

Running the above command will generate three output files in the output path:

• model.pkl: storing the model weights with the best valid loss during training.
• embedding.h5ad: an anndata file storing the embedding extracted by GraphTransformer. And .obs['Pred'] saves the results of the prediction.

1. Install the required environment according to Installation.
2. Create a data folder in the same directory as the 'SANGO' folder and download datasets from LargeIntestineA_LargeIntestineB.h5ad.
3. Create a folder genome in the ./SANGO/CACNN/ directory and download mm9.fa.h5.
4. For more detailed information, run the tutorial LargeIntestineB_LargeIntestineA.ipynb for how to do data preprocessing and training.

1. Install the required environment according to Installation.
2. Create a data folder in the same directory as the 'SANGO' folder and download datasets from MosP1_Cerebellum.h5ad.
3. Create a folder genome in the ./SANGO/CACNN/ directory and download mm10.fa.h5.
4. For more detailed information, run the tutorial MosP1_Cerebellum.ipynb for how to do data preprocessing and training.

1. Install the required environment according to Installation.
2. Create a data folder in the same directory as the 'SANGO' folder and download datasets from BoneMarrowB_Liver.h5ad.
3. Create a folder genome in the ./SANGO/CACNN/ directory and download mm9.fa.h5.
4. For more detailed information, run the tutorial BoneMarrowB_Liver.ipynb for how to do data preprocessing and training.

1. Install the required environment according to Installation.
2. Create a data folder in the same directory as the 'SANGO' folder and download datasets from BCC_TIL_atlas.h5ad, BCC_samples.zip, HHLA_atlas.h5ad.
3. Create a genome folder in the same directory as the 'SANGO' folder and download GRCh38.primary_assembly.genome.fa.h5.
4. For more detailed information, run the tutorial tumor_example.ipynb for how to do data preprocessing and training.

If you find our codes useful, please consider citing our work:

@article{zengSANGO,
  title={Deciphering Cell Types by Integrating scATAC-seq Data with Genome Sequences},
  author={Yuansong Zeng, Mai Luo, Ningyuan Shangguan, Peiyu Shi, Junxi Feng, Jin Xu, Weijiang Yu, and Yuedong Yang},
  journal={},
  year={2023},
}