strange plotDispEsts result from Deseq

[sadiexiaoyu]

Hi, all,

I am using DESeq to do a paired design RNA-seq experiment. The aim of the analysis is to detect genes differentially expressed between two tissues from the same individual. I have 3 individuals as biological replicates.
First of all, I do the filtering using
> cdsFull=newCountDataSet(x,design)
>rs=rowSums(counts(cdsFull))
>theta=0.1
>use=(rs>quantile(rs,probs=theta))
>table(use)
>cdsFilt=cdsFull[use,]
> cdsFilt=estimateSizeFactors(cdsFilt)
> sizeFactors(cdsFilt)
X1b X2b X3b X1n X2n X3n
1.1493410 1.0751491 1.3231961 0.9078587 0.7904943 0.8936754
Then I normalized the data,
> head(counts(cdsFilt,normalized=TRUE))
X1b X2b X3b X1n X2n X3n
gene1 1359.03967 1283.54286 1158.558474 904.32576 1010.75997 917.55913
gene2 64.38472 73.47818 62.726910 56.17614 72.10678 50.35385
gene3 867.45362 778.49665 929.567464 771.04511 648.96103 788.87706
gene4 72.21530 54.87611 61.215418 53.97316 89.81722 74.97130
gene5 449.82299 424.12721 386.941904 331.54940 407.34006 343.52519
gene6 0.00000 0.00000 1.511492 0.00000 0.00000 0.00000
And using pooled-CR method to estimate dispersion,
> cdsFilt=estimateDispersions(cdsFilt,method=c("pooled-CR"))
> plotDispEsts(cdsFilt)
However, the plot result seems weird... at the left down corner, there is a package of dots....Could anyone tell me what is that? (I have attached the plot result).

Thank you very much!

Best,

Sadiexiaoyu

[Simon Anders]

Don't worry about them. They are just an artifact of the fitting procedure: In RNA-Seq, the real variance is always at least as large as the expected value. However, the sample variance can be smaller than the sample mean, and then, the dispersion is apparently zero. This is why we fit a mean-dispersion value: All these low points are "pulled up" to the red line.

[sadiexiaoyu]

Don't worry about them. They are just an artifact of the fitting procedure: In RNA-Seq, the real variance is always at least as large as the expected value. However, the sample variance can be smaller than the sample mean, and then, the dispersion is apparently zero. This is why we fit a mean-dispersion value: All these low points are "pulled up" to the red line.

Dear Simon,

So happy to know it is Ok! And thank you very much for giving detailed answer. But I am sorry that I am still confused why "In RNA-Seq, the real variance is always at least as large as the expected value. However, the sample variance can be smaller than the sample mean"?

Best,

Sadiexiaoyu