with or w/o pseudocounts in DESeq?

[ekimmike]

Hi,

I'd like to ask about your opinion on using pseudocounts in DE analysis, as some people say they're absolutely necessary some disagree ...
I've read all the vignettes - still no unclear, hence decided to test it on my data

I did comparison between DEseq, edgeR, BaySeq and cuffDiff on RNAseq data (these are tophat genes, HTseq tables, FDR = 10% for all)

please see the file

http://www.filedropper.com/pseudo

In case of condition 1 it seems clear DESeq performs best, but in condition 2 it appears most affected by pseudocounts detecting about 1600 in case 1, 2000 in case2 with only 1000 overlap

any suggestions?

[Gordon Smyth]

I suggest that you use the packages that you mention according to the instructions that come with them. None of the package permit you to do ad hoc transformations of the counts prior to analysis.

BTW, the term 'pseudocount' has a technical meaning for the edgeR package, but it part of the internal computations, not something that is manipulated by a user. It is not a meaningful concept for any of the other packages.

[Simon Anders]

I really wonder where this idea that "pseudocounts are absolutely necessary" comes from.

The DESeq vignette states "The count values must be raw counts of sequencing reads. This is important for DESeq’s statistical model to hold, as only the actual counts allow assessing the measurement precision correctly. Hence, please do do not supply other quantities, such as (rounded) normalized counts, or counts of covered base pairs – this will only lead to nonsensical results."

I guess in the next release I will put the first sentence in red boldface. Every other week somebody comes up with a new idea what to try as input. So, again: If the numbers in your count tables are not counts of reads, they are not suitable. Pseudocounts are not counts of reads, but counts of nothing.

So, kudos for your work of testing all these methods with pseudocounts, but why did you bother?

[ekimmike]

I guess in the next release I will put the first sentence in red boldface.

I know, I know ... I'll move this discussion to bioconductor forum, and try to explain better

So, kudos for your work of testing all these methods with pseudocounts, but why did you bother?

mostly because I switched from nonparametric tools