Should one combine normalization methods in RNA-seq?

[BGould]

Does anyone know if RPKM normalization and quantile normalization of RNA-seq read counts can or should be combined?

For normalization of paired-end reads in RNA-seq, is it accurate to normalize the raw read counts using the quantile method, then following that calculate the RPKM value for each transcript using the quantile normalized values?

It seems that RPKM normalization using the total millions of mapped reads is somewhat redundant with an initial quantile normalization of read counts. Are these two methods mutually exclusive or should they be combined? Any recommendations?

thanks!

[Macspider]

It's a bit late for this answer but maybe future users will benefit from that:
When you use FPKM / RPKM as a measure unit you already have a normalized unit, which accounts for the gene length. However, when you have many samples sometimes you have different sequencing depths and thus different expression values that are the same thing but at two different depths. To account for that, using a global scaling or a full quantile normalization will help. In fact, when doing it you will end up with all the different "lanes" having the same statistical properties and thus remove the problem of the sequencing depth.

So, in conclusion, use RPKM/FPKM and then do a full quantile on all the samples, it works and is a good practice.

[dpryan]

A better solution would be to never use RPKM/FPKM, since it tends to kill your statistical power, which is related to the raw counts and gets normalized away by the RPKM transformation.