ker2xu / igsimpute

IGSimpute

IGSimpute is an accurate and interpretable imputation method for recovering missing values in scRNA-seq data with an interpretable instance-wise gene selection layer.

Requirements

Download with

$ git clone https://github.com/ker2xu/IGSimpute

You need to create an enviroment named IGSimpute using conda from environment.yml.

$ conda env create -f environment.yml

Usage

You need to first activate the environment by:

conda activate IGSimpute

You can run the command below to perform imputation on the heart-and-aorta tissue in the Tabula Muris atlas:

./run_IGSimpute.sh

If you want to perform imputation on your own dataset, you need to modify parameters defined in run_IGSimpute.sh.

Parameters

Name	Default value	Description
data_dir	-	Path to the directory that contains all datasets.
dataset_dir	'tm_droplet_Heart_and_Aorta'	Directory name of the dataset to be imputed.
exp_file_name	'X.csv'	Expression file name.
output_dir	'imputation_output'	Output directory name.
hg	'0.1'	The percentage or the number of used highly variable genes.
epochs	100	The maximum allowed epochs.
split_pct	"0.8"	The percentage of cells used as training dataset, and the left will be used for validtaion.
target_format	"count"	The expected output format.
ggl_loss_weight	"1"	The weight of $L_{gg}$.
gsl_L1_weight	"0.01	The weight of $L_{gs}$.
rec_loss_weight	"0.1	The weight of $L_{rec}$.
batch_size	256	Minibatch size.
dim	400	Size of the innermost embedding.
encoder_dropout_rate	"0.2"	Dropout rate of the dropout layer in the encoder part.
gpu_node	0	The index of GPU to use.
low_expression_percentage	0.80	If the percentage of low expression neighbor entries exceeds low_expression_percentage, the gene expression target entry will be changed to zero.
low_expression_threshold	0.20	All zero entries, and non-zero entries with expression less than low_expression_threshold quantile will be taken as low expression entires in the KNN post-processing.
lr	"1e-4"	Learning rate.
seed	0	Seed number.
sub_sampling_num	"None"	Randomly select sub_sampling_num cells for training and validation.
valid_dropout	"0.2"	The percentage of non-zero entries to be used for validation.

Expected input format

IGSimpute accepts expression profiles in h5ad or csv format. Each row should correspond to a cell and each column should correspond to a gene.

Results

The output will be put inside "data_dir/dataset_dir/imputation_output" directory. "IGSimpute.name.csv.gz" is the imputed expression matrix without KNN post-processing and "IGSimpute.KNN.name.csv" is the imputed expression matrix wit KNN post-processing.