Hello,

I revisited recently an EWAS analysis done with minfi v. 1.16. With minfi v. 1.20, the distribution of beta values after preprocessQuantile normalization was very strange, and quite far from the clean bimodal distribution that I was expecting (and that I got originally with v. 1.16).

I saw in the "NEWS" section that a bug with preprocessQuantile was corrected in v.1.19. Was this bug already present in 1.16? Is preprocessQuantile still a good choice for normalization in v. 1.20, or should it be systematically replaced by a more elaborate normalization like preprocessFunnorm?

This is not a reproducible example without the data, but for information, here are the steps that lead to the aberrant distribution of beta values:

targets <- read.metharray.sheet(".")
RGset   <- read.metharray.exp(targets=targets)
GRset   <- preprocessQuantile(RGset, sex=rep("F",12), quantileNormalize=TRUE)
densityPlot(getBeta(GRset))

Thank you very much for your help,

Frederic

The bug should not have that effect.  For example, we re-checked some of
the plots in the minfi paper and concluded they were unchanged.

But it sounds worrisome.  Could you perhaps share some of the data; perhaps
there is a smaller subset which shows this behaviour?  It is unfortunately
basically impossible to even speculate what is happening with the little
information you give.

But to re-visit the plot: could you do a density separately for probes of
type I and type II and do it with minfi 1.16 and 1.20 and compare?

Best,
Kasper

Hello Kasper,

Thank you for your very quick response!

I plotted the density for all probes, then by probe types, with minfi 1.16 and minfi 1.20. Indeed, by probe type, the normalization looks rather OK.

Here is what it looks like:

http://adcloud.genome.ulaval.ca/index.php/s/PdeyTmu8gqNAt2o   (minfi 1.16)

http://adcloud.genome.ulaval.ca/index.php/s/zD4SMrfEMJTRpw0    (minfi 1.20)

The data are still unpublished. I could ask permission to give you access to the dataset, but it would have to be communicated outside of bioconductor's mailing list.

Thanks for your help!

Frederic 

Ok, I have looked into this, and I am not done yet.  I can replicate this problem with the data in `minfiData`, and it is indeed related to the bugfix introduced in preprocessQuantile().  I have (for now) added a `method` argument to preprocessQuantile in minfi 1.21.3 (right now only on Github); with `method = "v1"` (default) you get the old behavior, with `method = "v2"` you get the new (weird) behavior. Now that the issue has been replicated and I have a old and new codebase for testing, I will look further into it - ie. revisit my decision that this was a bug and not a feature.

I have now fully analyzed the issue.  The problem is indeed caused by the bugfix introduced in minfi 1.20.0.  The version of preprocesQuantile in minfi 1.20.0 and 1.21.0-1.21.2 is wrong and should not be used.

I have reverted the change and committed this to minfi 1.21.3 in the devel branch.  Following a successful build (a few days down the road) I will patch minfi 1.20.0 to 1.20.1.

Best,
Kasper