Hi Erwin:

I have not met this situation before, but I have an idea about this situation.

Firstly I suspect the quality of the data is not very good, even before normalization it should not be like that.

Secondly, BMIQ method (I assume you have used that) is designed for beta distribution data, so if the original data is even not beta distributed, maybe it would not working (even fail).

Thirdly, BMIQ is done by separating beta distribution into 3 parts, lower-value(for example, 0-0.4), middle-value (0.4-0.7), higher value(0.7-1)... And fit each part for normalization. So in your data, I would suspect BMIQ encountered some problem in the middle and high-value area, for example from beta 0.4 to 1. Thus the lower part looks right, but the rest 2 looks kind of wired. In other words, possibly one type of probe is not very good in the middle-high area. I think maybe you can try check it using QC.GUI() function, on origin matrix, like below:

QC.GUI(myLoad$beta,myLoad$pd$Sample_Group,arraytype="EPIC")

In the second panel, you may see the distirbution of two difference types.

So I think maybe you can try other normalization methods, see if the plot is still like this. I think some other methods, like PBC maybe not be that picky for data distribution.

Again, I have to say I never encounter this situation, but I am just saying some possibilities I think might cause this.

Best Tian

Cross-posted: https://www.biostars.org/p/419178/