Hello,

I am a bit confused about the rlog() and vsd() functions in DESeq2.

Using log2 transformation, it is quite clear how to proceed for normalisation and then log2 transformation of data. It is however not possible to normalise the data before applying rlog or vst on them. I see from the document that normalisation is done sort of behind the scene in these two cases.

I used the three transformations on my data. Please see the graph in the link below: https://github.com/Homap/datascience/blob/master/docs/log2rlog_vst.pdf

What I see is that the shape of the distribution looks quite different using log2 vs. rlog vs. vsd. In the case of the graph above, it seems if sufficient filtering of raw counts are done, log2 transformation works quite good. I was wondering what is your idea about that and also why rlog values start from below zero?

Thank you!

I am not sure why one would want to do a blind log [base 2] transformation on count data. There are specific functions in DESeq2 for logging data and utilising it in downstream applications.

The general procedure can be regarded as:

1. Derive raw or estimated counts (outside DESeq2)
2. Import counts to DESeq2 (tximport and/or DESeq2)
3. Normalise the counts via an estimation of size factors and gene-wide dispersion (DESeq2)
4. conduct differential expression analysis on the normalised counts (DESeq2)
5. transform the data for downstream applications (e.g., PCA, clustering, 'machine learning', etc.) via variance stabilisation or regularised log (DESeq2)

If, at any point, you wish to obtain 'normalised counts', then you can use:

counts(dds, normalized = TRUE)

Kevin