DrImpute : imputing dropout events in single-cell RNA-sequencing data
In DrImpute: Imputing Dropout Events in Single-Cell RNA-Sequencing Data

library(knitr)
opts_chunk$set(fig.width = 12)

This vignette illustrates the use of DrImpute software in single cell RNA sequencing data analysis.

Data preparation

Example data is taken from Usoskin et al. (2015), GSE59739. We randomly selected 150 cells from original 799 cells.

library(DrImpute)

Firstly, genes that are expressed less than 2 cells are removed.

data(exdata)
exdata <- preprocessSC(exdata)

Normalization is performed using total read count for simplicity, and then log transformation is applied.

sf <- apply(exdata, 2, mean)
npX <- t(t(exdata) / sf ) 
lnpX <- log(npX+1)

Data analysis

Dropout Imputation can be simply done using DrImpute function.

lnpX_imp <- DrImpute(lnpX)

zero_p <- sum(lnpX == 0)/(dim(lnpX)[1]*dim(lnpX)[2])
zero_p_imp <- sum(lnpX_imp == 0)/(dim(lnpX)[1]*dim(lnpX)[2])

The ratio of zero is r round(zero_p,2), and r round(zero_p - zero_p_imp,2)*100 percent of zero's are imputed by DrImpute.

We visualized single cell RNA sequencing data using PCA with and without imputation by DrImpute.

par(mfrow = c(1,2))

lXc <- scale(t(lnpX), center= TRUE, scale = FALSE)
lXc_imp <- scale(t(lnpX_imp), center= TRUE, scale = FALSE)

library(irlba)
#svd.lXc <- svd(t(lXc))
svd.lXc <- irlba(lXc, nv = 2)
svd.lXc.imp <- irlba(lXc_imp, nv = 2)

PC <- svd.lXc$u %*% diag(svd.lXc$d)
PC.imp <- svd.lXc.imp$u %*% diag(svd.lXc.imp$d)

plot(PC, bg= c("red", "blue","black", "purple")[factor(colnames(exdata))], type = "p", pch = 21, col = "black", main = "Without imputation", xlab = "PC1", ylab="PC2")
plot(PC.imp, bg= c("red", "blue","black", "purple")[factor(colnames(exdata))], type = "p", pch=21, col="black", main = "With DrImpute", xlab = "PC1", ylab="PC2")

add_legend <- function(...) {
  opar <- par(fig=c(0, 1, 0, 1), oma=c(0, 0, 0, 0),
    mar=c(0, 0, 0, 0), new=TRUE)
  on.exit(par(opar))
  plot(0, 0, type='n', bty='n', xaxt='n', yaxt='n')
  legend(...)
}

add_legend("bottomright", legend=levels(factor(colnames(exdata))), pch=19, col = c("red","blue", "black", "purple"), cex=1, horiz = TRUE)

Prior to the use of DrImpute, the NP, TH, and PEP groups are visually indistinguishable in the 2D space. However, after using DrImpute, NP, TH, and PEP have better separation.