How can I correct skewed distribution of FPKM values?

Question

[natalie]

I have an excel output file with alternative transcript counts in the form of FPKM values. I'm interested in whether there are case control differences in a handful of the transcripts, but all of the transcripts have super skewed distribution and so i can't run parametric tests. Usually i would transform to the natural logarithm and then proceed with parametric tests (ANOVA, ANCOVA etc). I have done this and it seems to be okay, although when i graph the log values they still look pretty strange and not normally distributed at all.

Does anyone have any advice on whether this is okay or what other method they would use? Thanks!!

Answer 1

[mforde84]

If the values aren't normally distributed then you'll have to use statistical test which either uses the type of distribution it is, or use a non-parametric (rank order tests etc).

Since it's RNAseq data as well, my assumption is that the raw counts data follows a negative binomial distribution. So in that instance, you could use negative binomial GLM directly on the counts data instead of calculating FPKM / log normalizing / etc.

Another thing you might want to consider trying is quantile normalization of the log2 FPKM value. Plot those and see how they look.