DESeq on "small genome" data

[enelkinsan]

Hi guys,
I'm interested in your opinion, whether it is a good idea to use DESeq application for DE in an experiment where number of transcripts present in the sequencing mixture would be extremely small.
I'm dealing with RNA-seq sequencing data for 50 amplicons, that have been tested for differential expression. This is some kind of test for tissue-specific expression of biomarkers.
On one hand, DESeq may be more resistant to variance inequality in compared series than lmFit, on the other hand - is it right to estimate variance-mean dependance using just 50 points???
limma's lmFit-eBayes worked just fine last time with similar data, but this time the variance in some amplicons is just dragging pVals towards 1.
PS. study of 50 amplicons behaviour via RNA-seq was NOT my idea, I'm just dealing with the output.

Thanks,
Elizabeth

[Simon Anders]

You are right, estimating variance-mean relation from just 50 data points is indeed not such a great idea. In our new DESeq2 package, we now offer the new option fitType="mean" for estimateDispersions which just fits a single mean dispersion and shrinks the gene-wise estimates towards it. This might work well. However, if the variance it too high for limma, DESeq2 will also not see more, I suppose.

Do you mean that the amplicons with high variance have high p values? That would be correct. Or do you mean that the high-variance amplicons pull up the variance estimates for the other amplicons? Then you may want to use some independent filtering beforehands.