DESeq with no biological replicates for identification of Diff. Expressed Genes

[dbaud]

Hi,

I would like to use the R-package DESeq for the identification of differentially expressed genes between 2 conditions. I have got 4 different conditions and 3 technical replicates per conditions (that is 12 samples of count data). I have read the vignette, and have understood that technical replicates should be merged into one column. When running DESeq this way, i get very few genes called as differentially expressed.
I have several questions:

1.
This question is about the merging part. Suppose we have two technical replicates from a given condition. For a certain gene we get 150 counts for replicate 1 and 75 counts for replicate 2. If this is simply do to the fact that the sequencing depth of replicate 1 is twice the one of replicate 2, when we add counts from these replicates together without normalization, aren't we committing a mistake?

2.
When using DESeq to identify differentially expressed genes with technical replicates merged, I get very few hits. Is this what we expect?

3.
Should all of the 12 samples be considered in the initial count data object in order to estimate dispersion and normalize the data, or should there be a new count data object for each pair of compared conditions? (that is only 2 conditions per count data object)

[raphael123]

1. To normalise sample 1 THEN sample 2 is equivalent as normalizing sample 1 PLUS sample 2 .. AS long as you accept to normalize with the DESeq way.

2. Some experiment have no hit .. You can t expect your number of hit to be something. You have to accept that p value is hard to interpret and take your best hit and then look at your data .
here is something to read about p value

3 I would say that you should have a maximum of sample to estimate dispersion ( but I didn t get what is use by DESeq as dispersion) Here is what I use (not being sure of what I do) :

cds=newCountDataSet(roundmiRNA,miRNAdesign$response)
cds=estimateSizeFactors(cds)
cds=estimateDispersions(cds,method='pooled-CR',sharingMode="maximum",fitType="local",modelFormula = count ~ miRNAdesign$subject)
cds2=cds[,is_escitalopram]
... then my analysis

[jwfoley]

It sounds like you have no biological replicates, only technical. Is that correct? If so, you cannot do statistics on these data using DESeq or any other tool. The statistical tests ask Is the difference between conditions greater than the variation within conditions? and you are missing the second part. You can't do statistics with N = 1.