1. Introduction

scRecover is an R package for imputation of single-cell RNA-seq (scRNA-seq) data. It will detect and impute dropout values in a scRNA-seq raw read counts matrix while keeping the real zeros unchanged.

Since there are both dropout zeros and real zeros in scRNA-seq data, imputation methods should not impute all the zeros to non-zero values. To distinguish dropout and real zeros, scRecover employs the Zero-Inflated Negative Binomial (ZINB) model for dropout probability estimation of each gene and accumulation curves for prediction of dropout number in each cell. By combination with scImpute, SAVER and MAGIC, it not only detects dropout and real zeros at higher accuracy, but also improve the downstream clustering and visualization results.

2. Citation

If you use scRecover in published research, please cite:

3. Installation

To install scRecover from Bioconductor:

if(!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install("scRecover")

To install the developmental version from Bioconductor:

if(!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install("scRecover", version = "devel")

Or install the developmental version from GitHub:

if(!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
BiocManager::install("miaozhun/scRecover")

To load scRecover and other required packages for the vignettes in R:

library(scRecover)
library(BiocParallel)
suppressMessages(library(SingleCellExperiment))

4. Input

scRecover takes two inputs: counts and one of Kcluster or labels.

The input counts is a scRNA-seq read counts matrix or a SingleCellExperiment object which contains the read counts matrix. The rows of the matrix are genes and columns are cells.

Kcluster is an integer specifying the number of cell subpopulations. This parameter can be determined based on prior knowledge or clustering of raw data. Kcluster is used to determine the candidate neighbors of each cell and need not to be very accurate.

labels is a character/integer vector specifying the cell type of each column in the raw count matrix. Only needed when Kcluster = NULL. Each cell type should have at least two cells for imputation.

5. Test data

Users can load the test data in scRecover by

data(scRecoverTest)

The test data counts in scRecoverTest is a scRNA-seq read counts matrix which has 200 genes (rows) and 150 cells (columns).

dim(counts)
#> [1] 200 150
counts[1:6, 1:6]
#>        E3.46.3383 E3.51.3425 E3.46.3388 E3.51.3423 E3.46.3382 E3.49.3407
#> BTG4           22          0         12         26          0          0
#> GABRB1          0          0          0          0          0          0
#> IL9             0          0          0          0          0          0
#> TAPBPL          2          0          5          1          0          2
#> KANK4           0          0          0          0          0          0
#> CPSF2          12          0         95          0          5        115

The object labels in scRecoverTest is a vector of integer specifying the cell types in the read counts matrix, corresponding to the columns of counts.

length(labels)
#> [1] 150
table(labels)
#> labels
#>   1   2 
#>  50 100

The object oneCell in scRecoverTest is a vector of a cell’s raw read counts for each gene.

head(oneCell)
#>         Vsx2_1a_F2
#> Itm2a            0
#> Gm26258          0
#> Sergef         149
#> Gm24106          0
#> Fam109a          0
#> Ssu72           60
length(oneCell)
#> [1] 24538

6. Usage

6.1 Imputation using scRecover

6.1.1 With read counts matrix input

Here is an example to run scRecover with read counts matrix input:

# Load test data for scRecover
data(scRecoverTest)

# Run scRecover with Kcluster specified
scRecover(counts = counts, Kcluster = 2, outputDir = "./outDir_scRecover/", verbose = FALSE)
#> [1] "========================= Running scImpute ========================="
#> [1] "reading in raw count matrix ..."
#> [1] "number of genes in raw count matrix 200"
#> [1] "number of cells in raw count matrix 150"
#> [1] "reading finished!"
#> [1] "imputation starts ..."
#> [1] "searching candidate neighbors ... "
#> [1] "inferring cell similarities ..."
#> [1] "dimension reduction ..."
#> [1] "calculating cell distances ..."
#> [1] "cluster sizes:"
#> [1] 92 54
#> [1] "estimating dropout probability for type 1 ..."
#> [1] "searching for valid genes ..."
#> [1] "imputing dropout values for type 1 ..."
#> [1] "estimating dropout probability for type 2 ..."
#> [1] "searching for valid genes ..."
#> [1] "imputing dropout values for type 2 ..."
#> [1] "writing imputed count matrix ..."
#> [1] "========================= scImpute finished ========================"
#> 
#> Processing 1 of 2 cell clusters
#> 
#> Processing 2 of 2 cell clusters
#> 
#> [1] "======================== scRecover finished! ========================"
#> [1] "The output files of scRecover are in directory:"
#> [1] "./outDir_scRecover/"
#> [1] "============================== Cheers! =============================="

# Or run scRecover with labels specified
# scRecover(counts = counts, labels = labels, outputDir = "./outDir_scRecover/")

6.1.2 With SingleCellExperiment input

The SingleCellExperiment class is a widely used S4 class for storing single-cell genomics data. scRecover also could take the SingleCellExperiment data representation as input.

Here is an example to run scRecover with SingleCellExperiment input:

# Load test data for scRecover
data(scRecoverTest)

# Convert the test data in scRecover to SingleCellExperiment data representation
sce <- SingleCellExperiment(assays = list(counts = as.matrix(counts)))

# Run scRecover with SingleCellExperiment input sce (Kcluster specified)
scRecover(counts = sce, Kcluster = 2, outputDir = "./outDir_scRecover/", verbose = FALSE)
#> [1] "========================= Running scImpute ========================="
#> [1] "reading in raw count matrix ..."
#> [1] "number of genes in raw count matrix 200"
#> [1] "number of cells in raw count matrix 150"
#> [1] "reading finished!"
#> [1] "imputation starts ..."
#> [1] "searching candidate neighbors ... "
#> [1] "inferring cell similarities ..."
#> [1] "dimension reduction ..."
#> [1] "calculating cell distances ..."
#> [1] "cluster sizes:"
#> [1] 92 54
#> [1] "estimating dropout probability for type 1 ..."
#> [1] "searching for valid genes ..."
#> [1] "imputing dropout values for type 1 ..."
#> [1] "estimating dropout probability for type 2 ..."
#> [1] "searching for valid genes ..."
#> [1] "imputing dropout values for type 2 ..."
#> [1] "writing imputed count matrix ..."
#> [1] "========================= scImpute finished ========================"
#> 
#> Processing 1 of 2 cell clusters
#> 
#> Processing 2 of 2 cell clusters
#> 
#> [1] "======================== scRecover finished! ========================"
#> [1] "The output files of scRecover are in directory:"
#> [1] "./outDir_scRecover/"
#> [1] "============================== Cheers! =============================="

# Or run scRecover with SingleCellExperiment input sce (labels specified)
# scRecover(counts = sce, labels = labels, outputDir = "./outDir_scRecover/")

6.2 Estimate dropout gene number in a cell

Function estDropoutNum in the package could estimate the dropout gene number or all expressed gene number (namely observed gene number plus dropout gene number) in a cell:

# Load test data
data(scRecoverTest)

# Downsample 10% read counts in oneCell
set.seed(999)
oneCell.down <- countsSampling(counts = oneCell, fraction = 0.1)

# Count the groundtruth dropout gene number in the downsampled cell
sum(oneCell.down == 0 & oneCell != 0)
#> [1] 2095

# Estimate the dropout gene number in the downsampled cell by estDropoutNum
estDropoutNum(sample = oneCell.down, depth = 10, return = "dropoutNum")
#> [1] 1994

Blow shows the expressed gene number predicted by estDropoutNum with 10% downsampled reads and the groundtruth expressed gene number derived by downsampling when the reads depth varying from 0% to 100% of the total reads.

7. Output

Imputed expression matrices of scRecover will be saved in the output directory specified by or a folder named with prefix ‘outDir_scRecover_’ under the current working directory when is unspecified.

8. Parallelization

scRecover integrates parallel computing function with BiocParallel package. Users could just set parallel = TRUE (default) in function scRecover to enable parallelization and leave the BPPARAM parameter alone.

# Run scRecover with Kcluster specified
scRecover(counts = counts, Kcluster = 2, parallel = TRUE)

# Run scRecover with labels specified
scRecover(counts = counts, labels = labels, parallel = TRUE)

Advanced users could use a BiocParallelParam object from package BiocParallel to fill in the BPPARAM parameter to specify the parallel back-end to be used and its configuration parameters.

8.1 For Unix and Mac users

The best choice for Unix and Mac users is to use MulticoreParam to configure a multicore parallel back-end:

# Set the parameters and register the back-end to be used
param <- MulticoreParam(workers = 18, progressbar = TRUE)
register(param)

# Run scRecover with 18 cores (Kcluster specified)
scRecover(counts = counts, Kcluster = 2, parallel = TRUE, BPPARAM = param)

# Run scRecover with 18 cores (labels specified)
scRecover(counts = counts, labels = labels, parallel = TRUE, BPPARAM = param)

8.2 For Windows users

For Windows users, use SnowParam to configure a Snow back-end is a good choice:

# Set the parameters and register the back-end to be used
param <- SnowParam(workers = 8, type = "SOCK", progressbar = TRUE)
register(param)

# Run scRecover with 8 cores (Kcluster specified)
scRecover(counts = counts, Kcluster = 2, parallel = TRUE, BPPARAM = param)

# Run scRecover with 8 cores (labels specified)
scRecover(counts = counts, labels = labels, parallel = TRUE, BPPARAM = param)

See the Reference Manual of BiocParallel package for more details of the BiocParallelParam class.

9. Evaluation of scRecover

9.1 On downsampling data

We evaluated SAVER, scImpute, MAGIC and their combined with scRecover version SAVER+scRecover, scImpute+scRecover, MAGIC+scRecover on a downsampling scRNA-seq dataset generated by random sampling of reads from a SMART-seq2 scRNA-seq dataset (Petropoulos S, et al. Cell, 2016, https://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-3929/).

9.1.1 Accuracy of dropout prediction

We found after combined with scRecover, scImpute+scRecover, SAVER+scRecover and MAGIC+scRecover will have higher accuracy than scImpute, SAVER and MAGIC respectively.