Search
Question: PCA from deseq and r function differ
0
2.5 years ago by
tonja.r40
United Kingdom
tonja.r40 wrote:

I stabilize my data with rlog and then plot PCA with the DESeq proposed method.

dds = DESeqDataSetFromMatrix(countData=histone_m, colData=Design, design=~condition) cds=estimateSizeFactors(dds)  met = rlog(dds)      data <- plotPCA(met, intgroup=c("condition"), returnData=TRUE) percentVar <- round(100 * attr(data, "percentVar")) myPlot = ggplot(data, aes(PC1, PC2, color=name)) +        geom_point(size=3) +         xlab(paste0("PC1: ",percentVar[1],"% variance")) +         ylab(paste0("PC2: ",percentVar[2],"% variance"))

Then I decided to do a normal pca from r:

met =assay(met) pca<- prcomp(t(met)) screeplot_percent(pca) col=c("red","pink","black","blue") idx <- seq_len(3) print(splom(pca$x[,idx], col=col,pch=19)) Function for screeplot: screeplot_percent <- function(x, npcs = min(10, length(x$sdev)), ...) {   idx <- seq_len(npcs)   sum_var <- sum(x$sdev ^ 2) vars <- 100 * (x$sdev[idx] ^ 2 / sum_var)   cumvar <- cumsum(vars)      barplot(vars, width = 0.9, space = 0.1, names.arg = idx, ylim = c(0, 100),           xlab = "Principal Component", ylab = "Percent Variance",           xaxp = c(1, npcs, npcs - 1), las = 1)   lines(x = idx - 0.5, y = cumvar, type = "b", lty = 2)   legend("bottomright", legend = c("Proportion", "Cumulative"), lty = c(1, 2),          pch = c(19, 1)) }

red -> WEN1
pink -> WEN3
black -> WNN1
blue -> WNN3

In fact PCA plots differ a bit and screeplot differ a lot. However, I do not understand why.
As you see WEN3 and WNN3 are quite apart on the second plot in comparison to the first plot. Additionally, the scale on the y-axis and x-axis is different. What is the reason for this?

Also my screeplot tells me that PC1 explains app. 75% of variance whereas ggplot claims 87%.

modified 2.5 years ago by Michael Love17k • written 2.5 years ago by tonja.r40
2
2.5 years ago by
Michael Love17k
United States
Michael Love17k wrote:
Whenever you have a question about a function in Bioconductor, a good to start is with the help page for that function. For ?plotPCA in DESeq2, you'll see there is an extra step of filtering to use the top high variance genes.