Different PCA plots using rlog and vsd on the same data set

Question

[lirongrossmann]

Hi All,

I have been using Deseq2 to analyze a dataset I have and ran into a problem I am not sure how to solve.

I have been using the following code to run deseq2 on my dataset:

dds <-DESeqDataSetFromMatrix(countData = ep,colData = cp,design = ~Risk)
dds <- estimateSizeFactors(dds)
rld <- rlog(dds)
plotPCA(rld, intgroup="Risk")
vsd <- varianceStabilizingTransformation(dds)
plotPCA(vsd, intgroup="Risk")

The two PCA plots I got look completely different, so I am not sure which transformation I should rely on for further analysis.

Any help?

[Kevin Blighe]

Just a comment that may be of interest:

DESeq2's PCA function automatically removes a proportion of variables of low variance before printing the plot. As you're using a variance-stabilised transformation versus a regularised log transformation, I would fully expect the PCA plots to look different based on this behaviour of DESeq2.

If you try the code that I posted in your other thread on both the rld and vst counts, then I think that you will see more similar PCA bi-plots, as, with my code, no variables are removed prior to generating the plots.

[lirongrossmann]

Thank you! I have been using variance-stabilised transformation mainly because rlog runs very slowly. In your experience, does rlog give better results?

[Kevin Blighe]

I prefer rlog counts purely because the data distribution looks better, however, as you mentioned, it can be very very slow! There is nothing wrong with using variance-stabilised counts.

In DESeq2, you have to remember, too, that the statistical tests are performed on just the normalised counts, i.e., before any rlog or vst.

[GouthamAtla]

You should post the PCA Plot. You can anonymise the sample labels if you don't wish to show the sample labels.

[lirongrossmann]

I would love to! Is there a way to upload the plots?

[Devon Ryan]

You'll need to upload them elsewhere and then link to them.

[Michael Love]

Please add links when you cross post to multiple forum sites.