1. Download source codes of scDeepSort.
2. Download pretrained models from the release page and uncompress them.

tar -xzvf pretrained.tar.gz

After executing the above steps, the final scDeepSort tree should look like this:

 |- pretrained
     |- human
        |- graphs
        |- statistics
        |- models
     |- mouse
        |- graphs
        |- statistics
        |- models
 |- test
     |- human
     |- mouse
 |- train
    |- human
    |- mouse
 |- map
    |- human
        |- map.xlsx
    |- mouse
        |- map.xlsx
    |- celltype2subtype.xlsx
 |- models
    |- __init__.py
    |- gnn.py
 |- utils
    |- __init__.py
    |- preprocess.py
    |- preprocess_internal.py
 |- R
    |- example_data.rds
    |- geneinfo.rds
 |- pre-process.R
 |- predict.py
 |- train.py
 |- requirements.txt
 |- README.md
 |- LICENSE

• Dependencies can also be installed using pip install -r requirements.txt
• To use GPU, please install the gpu version of dgl, see Install DGL for more details.

1. The file name of test data should be named in this format: species_TissueNumber_data.csv. For example, human_Pancreas11_data.csv is a data file containing 11 human pancreas cells.

2. The test single-cell transcriptomics csv data file should be pre-processed by first revising gene symbols according to NCBI Gene database updated on Jan. 10, 2020, wherein unmatched genes and duplicated genes will be removed. Then the data should be normalized with the defalut LogNormalize method in Seurat (R package), detailed in pre-process.R, wherein the column represents each cell and the row represent each gene for final test data, as shown below.

   	Cell 1	Cell 2	Cell 3	...
   Gene 1	0	2.4	5.0	...
   Gene 2	0.8	1.1	4.3	...
   Gene 3	1.8	0	0	...
   ...	...	...	...	...

3. All the test data should be included under the test directory. Human datasets should be under ./test/human and mouse datasets should be under ./test/mouse

Use --evaluate to reproduce the results as shown in our paper. For example, to evaluate the data mouse_Testis199_data.csv, you should execute the following command:

python predict.py --species human --tissue Testis --test_dataset 199 --gpu -1 --evaluate --filetype gz --unsure_rate 2

• --species The species of cells, human or mouse.

• --tissue The tissue of cells. See wiki page

• --test_dataset The number of cells in the test data.

• --gpu Specify the GPU to use, 0 for gpu,-1 for cpu.

• --filetype The format of datafile, csv for .csv files and gz for .gz files. See pre-process.R

• --unsure_rate The threshold to define the unsure type, default is 2. Set it as 0 to exclude the unsure type.

Output: the output named as species_Tissue_Number.csv will be under the automatically generated result directory, which contains four columns, the first is the cell id, the second is the original cell type, the third is the predicted main type, the fourth is the predicted subtype if applicable.

Note: to evaluate all testing datasets in our paper, please download them in release page

Use --test to test your own datasets. For example, to test the data human_Pancreas11_data.csv, you should execute the following command:

python predict.py --species human --tissue Pancreas --test_dataset 11 --gpu -1 --test --filetype csv --unsure_rate 2

• --species The species of cells, human or mouse.

• --tissue The tissue of cells. See wiki page

• --test_dataset The number of cells in the test data.

• --gpu Specify the GPU to use, 0 for gpu, -1 for cpu.

• --filetype The format of datafile, csv for .csv files and gz for .gz files. See pre-process.R

• --unsure_rate The threshold to define the unsure type, default is 2. Set it as 0 to exclude the unsure type.

Output: the output named as species_Tissue_Number.csv will be under the automatically generated result directory, which contains three columns, the first is the cell id, the second is the predicted main type, the third is the predicted subtype if applicable.

To train your own model, you should prepare two files, i.e., a data file as descrived above, and a cell annotation file under the ./train directory as the example files. Then execute the following command:

python train.py --species human --tissue Adipose --gpu -1 --filetype gz

python train.py --species mouse --tissue Muscle --gpu -1 --filetype gz

• --species The species of cells, human or mouse.

• --tissue The tissue of cells.

• --gpu Specify the GPU to use, 0 for gpu, -1 for cpu.

• --filetype The format of datafile, csv for .csv files and gz for .gz files. See pre-process.R

Output: the trained model will be under the pretrained directory, which can be used to test new datasets on the same tissue using predict.py as described above.