RNAseq analysis using DESeq

[katussa10]

We are using DESeq to find differentially expressed genes for RNAseq experiment with two biological replicates. When we did the analyses considering them as biological replicates, we found that among differential expressed genes, there was a very high number of counts in one biorep compared to the others. Then we took biorep1 from experimental group and compared with biorep1 of the control group and there were only 3 genes differentially expressed at padj 0.05. However, there are lot of genes with thousands reads vs 0 in the two groups. Then we did the same with the second biorep and we found about 100 differentially expressed genes. Does anyone know why this is happening?

[Simon Anders]

I'm not sure I understand your question. Could you give an example, please?

[katussa10]

Hi Simon,
Here is an example (attached txt file) for the genes that showed differential expression, but between the same experimental group variation was very high. Please let me if it is still not clear. I will try to explain again.

[Gangcai]

I'm not sure I understand your question. Could you give an example, please?

Dear Simon,
I have quite similar problem for the significant genes detected by DESeq and edgeR. Both of them output some significant candidates which have quite large variation within groups. Such as:
wt1: 2
wt2: 345
treat1: 3
treat2:1

or
wt1: 0
wt2: 345
treat1: 0
treat2: 0

Is it normal to get low p value for such kind of expression pattern? Thanks

[Simon Anders]

Short answer: Please try again with the 'deve' version of DESeq (version 1.5.19), and this oddity should vanish.

Long answer: In the current release version of DESeq (version 1.4.1), we estimate a variance for each gene, fit a line through the mean-variance plot, and then use the fitted value of the variance, i.e., the value typical for a gene of the same expression strength. The 'nbinomTest' function gives you, besides the p values, two columns with the "variance residuals", i.e., the ratio of the gene's variance estimate over the fitted value. Cases such as your should show up as having a large value there and the vignette advises to disregard such hits in the downstream analysis.

Nobody ever read this sentence in the vignette, and also, the solution was rather unsatisfactory anyway, and so we have now changed this. Now, we do not use anymore always the fitted value, but instead the maximum of the per-gene estimate and the fitted value. This avoids artifacts like the ones you see. Have a look at the help page for 'estimateDispersion' in the new version, and also at the vignette, which we have extensively overhauled.

[Gangcai]

Short answer: Please try again with the 'deve' version of DESeq (version 1.5.19), and this oddity should vanish.

Long answer: In the current release version of DESeq (version 1.4.1), we estimate a variance for each gene, fit a line through the mean-variance plot, and then use the fitted value of the variance, i.e., the value typical for a gene of the same expression strength. The 'nbinomTest' function gives you, besides the p values, two columns with the "variance residuals", i.e., the ratio of the gene's variance estimate over the fitted value. Cases such as your should show up as having a large value there and the vignette advises to disregard such hits in the downstream analysis.

Nobody ever read this sentence in the vignette, and also, the solution was rather unsatisfactory anyway, and so we have now changed this. Now, we do not use anymore always the fitted value, but instead the maximum of the per-gene estimate and the fitted value. This avoids artifacts like the ones you see. Have a look at the help page for 'estimateDispersion' in the new version, and also at the vignette, which we have extensively overhauled.

Hi Simon,
Thanks for your quick reply. One more question about installation of the devel DESeq.
I have downloaded the newest version of Biobase from bioconductor, but DESeq require even advanced version.(http://www.bioconductor.org/packages...l/Biobase.html )
Do you know where to download >=2.13.6 Biobase? Thanks.
"
Error : package 'Biobase' 2.12.2 was found, but >= 2.13.6 is required by 'DESeq'
"

[labunit]

Hi Simon,
Thanks for your quick reply. One more question about installation of the devel DESeq.
I have downloaded the newest version of Biobase from bioconductor, but DESeq require even advanced version.(http://www.bioconductor.org/packages...l/Biobase.html )
Do you know where to download >=2.13.6 Biobase? Thanks.
"
Error : package 'Biobase' 2.12.2 was found, but >= 2.13.6 is required by 'DESeq'
"

You need to download the development version of R (2.14) to be able to install the development branch of Bioconductor packages including DESeq 1.5.19

[Simon Anders]

At http://www.bioconductor.org/packages...l/Biobase.html

However, installing a 'devel' version of Biobase over a 're'ease' installtion of Bioconductor might cause chaos. Better install the devel version of R and then, 'bioclite' will pull 'devel' versions of all Bioc packages.

[Gangcai]

Hi Simon,
I have tried the deve version of DESeq. The number of significant genes drop quite a lot. It does look better comparing with previous result. But still have some genes have high variation within biological repilcates.
wt1 wt2 treat1 treat2 pvalue_adjusted
928 0 0 0 <<0.01
0 135 0 0 <<0.01

[Simon Anders]

Could you please try again with

cds <- estimateDispersions( cds, method="pooled" )

It seems that our improved method removes these oddities reliably only if one uses a pooled dispersion estimation. I guess, we should hence change the default in 'estimateDispersions' to this, but at the moment, it is still method="per-condition" (which is the same as method="normal" in the old version).