Repository files navigation

English |简体中文

CAME is a tool forCell-type Assignment and Module Extraction, based on a heterogeneous graph neural network.

For detailed usage, please refer toCAME-Documentation.

CAME outputs the quantitative cell-type assignment for each query cell, that is,the probabilities of cell types that exist in the reference species, whichenables the identification of the unresolved cell states in the query data.

Besides, CAME gives the aligned cell and gene embeddings across species, whichfacilitates low-dimensional visualization and joint gene-module extraction.

It's recommended to create a conda environment for running CAME:

conda create -n env_came python=3.8conda activate env_came

Install required packages:

# on CPUpip install"scanpy[leiden]"pip install torch# >=1.8pip install dgl# tested on 0.7.2, better below 1.0.*

SeeScanpy,PyTorch andDGLfor detailed installation guide (especially for GPU version).

Install CAME by PyPI:

pip install came

Install the developmental version of CAME from source code:

git clone https://github.com/XingyanLiu/CAME.gitcd CAMEpython setup.py install

The test code is based on the sample data attached to the CAME package.It is initially saved in compressed form (CAME/came/sample_data.zip),and will be automatically decompressed to the default directory(CAME/came/sample_data/) when necessary, which contains the following files:

• gene_matches_1v1_human2mouse.csv (optional)
• gene_matches_1v1_mouse2human.csv (optional)
• gene_matches_human2mouse.csv
• gene_matches_mouse2human.csv
• raw-Baron_mouse.h5ad
• raw-Baron_human.h5ad

You can access these data bycame.load_example_data().

If you tend to apply CAME to analyze your own datasets, you need toprepare at least the last two files for the same species (e.g., cross-datasetintegration);

For cross-species analysis, you need to provide another.csvfile where the first column contains the genes in the reference species and thesecond contains the corresponding query homologous genes.

NOTE:the fileraw-Baron_human.h5ad is a subsample from the original datafor code testing. The resulting annotation accuracy may not be as good asusing the full dataset as the reference.

Suggestions

If you have sufficient GPU memory, setting the hidden-sizeh_dim=512in "came/PARAMETERS.py" may result in a more accurate cell-type transfer.

To test the package, run the python filetest_pipeline.py:

# test_pipeline.pyimportcameif__name__=='__main__':came.__test1__(6,batch_size=2048)came.__test2__(6,batch_size=None)

python test_pipeline.py

• Issue Tracker:https://github.com/XingyanLiu/CAME/issues
• Source Code:
  • https://github.com/zhanglabtools/CAME
  • https://github.com/XingyanLiu/CAME (the developmental version)

If you are having issues, please let us know. We have a mailing list located at:

• 544568643@qq.com
• xingyan@amss.ac.cn

If CAME is useful for your research, consider citing our work:

Liu X, Shen Q, Zhang S. Cross-species cell-type assignment of single-cell RNA-seq by a heterogeneous graph neural network[J]. Genome Research, 2022: gr. 276868.122.

Preprint:https://doi.org/10.1101/2021.09.25.461790