DESeq2 beta prior

[ysnapus]

I have a dataset where the condition of interest is a factor with 3 levels. DESeq2 issues me a warning to use betaPrior=False otherwise the log2 FoldChange will be affected by the base level. I thought the fold-change is always calculated w.r.t to a base level.

In my case, I have 3 time points (0 - control, 15, and 60). I get results for 15_vs_0 and 60_vs_0, and of course, these are different and depend on the fact that I chose 0 as the base point. I ran the DESeq function with and without betaPrior=F, and I get different results for 0_vs_15 fold change and p-values for the betaPrior=T/F cases.

What is the beta prior for? I tried to look in the DESeq2 paper at bioRXiv (http://biorxiv.org/content/early/2014/02/19/002832) and I don't think there is a beta distribution involved. Or is it talking about the normal prior for the estimated beta parameter in the GLM? i.e., is it the equation (10) of the "Moderated estimation of fold change and dispersion for RNA-Seq data with DESeq2" paper?

[Michael Love]

hi,

Yes you've got it right (it's the zero-centered Normal prior). In the paper, we say "logarithmic fold change prior", and as logarithmic fold changes are written as beta in the GLM, in the software we use the shorter "betaPrior".

Note this is described in the help file for ?DESeq and ?nbinomWaldTest:

betaPrior whether or not to put a zero-mean normal prior on the non-intercept coefficients (Tikhonov/ridge regularization) See nbinomWaldTest for description. Only used for the Wald test.

The warning is discussed on the Bioc list here

http://permalink.gmane.org/gmane.science.biology.informatics.conductor/52331

and here

http://permalink.gmane.org/gmane.science.biology.informatics.conductor/51749

briefly, in version 1.2, the LFC estimates can change if you choose a different base level, i.e. 15_vs_0 was not the same as -1 * 0_vs_15

We have fixed this in version 1.3, so that the LFCs are base-level independent. In this case, you can make contrasts of any level against any other level, using the contrast argument to results(). This is accomplished with the expanded model matrix described in the paper and in ?nbinomWaldTest.

[ysnapus]

Thanks! It's probably better to use version 1.3. How can I install the development version? Bioconductor installed 1.2.10.

[Michael Love]

it will install automatically, if you are running the development version of R.

it's not trivial to install devel R alongside release R, but we are really happy to have devel testers

there is some information on the Bioconductor website under 'developers'

The devel version 1.3 will become release 1.4 on April 14, 2014.