sigGCN

1. Generate the gene adjacency network

If the gene-gene network is not available, use the getADJ.py to generate the sparse gene adjacency network.

Five varables in the function:

@dataPath: directory of input expr matrix

@networkPath：directory of STRING or BioGrid network

@dataset：which dataset to load, get a list of gene official names

@net：which network to use, STRING or BioGrid

@pathToSave：directory to save the adj matrix

Run the file getADJ.py:

python getADJ.py --dataPath="data" --networkPath="data" --dataset="Zhengsorted" net="String" --pathToSave="data/Zhengsorted"

2. Load data and train the model

Six variables should be provided.

@dirData, type=str, "directory of cell x gene matrix"

@dataset, type=str, "dataset to load"

@dirAdj, type = str, "directory of adj matrix"

@dirLabel, type = str, "directory of adj matrix"

@outputDir, type = str, "directory to save results"

@saveResults, type=int, default=0, "whether or not save the results, use 1 to save and 0 to not save"

To train the model, run the file:

python siggcn.py --dirData="data/" --dataset="Zhengsorted" --dirAdj="data/Zhengsorted/" --dirLabel="data/Zhengsorted/" --outputDir="data/output" --saveResults=0

2.1 load data

1. The data matrix (cell x gene expression matrix), adj matrix, labels, and a fixed shuffle index are loaded first.
2. Then use the "down_gene" choose the top N genes by variance.
3. Split the data into training, validation, and testing dataset.
4. Generate the data loader for training and evaluation.

2.2 Train the model

The dataset will be split into 80% as training, 10% as validaton, 10% as testing.

2.3 Save results

if use @saveResults=1, the results will be saved into the output folder.