How Bimodal Is A Methylation Distribution?

Question

[Pablo]

Analysing methylation from a sequencing experiment, we seem to be having some bimodal distributed values.

The question I was posed (which I'm still not so sure whether makes sense or not) is "Are you sure this is bimodal?". Needless to say, the plot doesn't show a clear "yes" or "no" answer. I.e. it's neither two well defined peaks nor one single peak, but something in between. So the questions are:

1- Is there a way to assign a p-value (or any metric) on how bimodal a distribution is?

2- Does it even make sense to ask this question? (Why or why not?).

Apologies in advanced if the question is too off topic.

Answer 1

[Chris Miller]

1) The advice given in this CrossValidated thread seems solid. Really, if you've got a hairy stats question, those are the guys to ask.

2) Bimodal distribution of methylation scores is is common, from what I understand. Here's a figure showing a distribution of methylation scores in CpG islands, taken from the supplement of this paper:

Harris, et al. Comparison of sequencing-based methods to profile DNA methylation and identification of monoallelic epigenetic modifications

Nature Link: http://www.nature.com/nbt/journal/v28/n10/full/nbt.1682.html

PMC Link: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2955169/?tool=pubmed

Answer 2

[Alastair Kerr]

A google search on testing for modality picks up this bit of advice

From work I am involved in on CpG islands I guess that your trends should be bimodal, it can be seen clearly in CpG island methylation in human and in full transcript methylation in Ciona.

There will be a Deaton et al paper coming out in Genome Research later in the year that will also show that the 'CpG shores' hypothesis is less likely.

[Pablo]

The reference you gave is about someone who comments on a method based on a package that he never used or even looked at.

Sorry, but it does seem a little bit too vague (plus it doesn't answer my question).

Also, minus one for referencing Google.

Answer 3

[Genotepes]

Very preliminary comment before heading home,

looks like a mixture distributions problem. If you data looks normal, then an EM algorithm and likelihood ratio test could solve this.

I am sure this is not quite so easy.

Christian

Answer 4

[Chiahsin Liu]

Actually, I get similar distribution from HumanMethylation450 data.