[BioC] MEDIPS: how does MEDIPS define a methylated region / cluster?

Fri Sep 12 19:12:32 CEST 2014

Dear Allen,

please see my comments below.

Best,
Lukas

On Fri, Sep 12, 2014 at 3:00 AM, Chong Kim San Allen <patcksa at nus.edu.sg>
wrote:

> Dear Lukas,
> Firstly, thank you so very much for your kind replies. You have been so
> patient and I am very, very grateful for your help.
>
> I have been away and so I'm trying now to gather my thoughts on this again.
>
> 1) I did read section 7.7 of the MEDIPS manual. My understanding (correct
> me if I am wrong) is that MEDIPS identifies differentially methylated
> regions (DMRs) by a sample-against-sample method rather than a
> group-against-group method. So, if I have 5 controls (A,B,C,D,E) and 6
> (1,2,3,4,5,6) test samples, then EACH individual control sample is compared
> against EACH test sample. So, A vs. 1, A vs. 2, A vs. 3, A vs. 4, A vs. 5,
> A vs. 6, B vs. 1, B vs. 2, ......etc. instead of comparing ALL of the
> control samples against ALL of the test samples. Is my understanding
> correct?
>

No that is not correct. MEDIPS identifies differentially enriched regions
by comparing two groups of samples.

>
> You are therefore saying that because of this all-against-all method that
> MEDIPS uses, correcting for CpG density is actually not necessary. But
> that's for identifying differentially methylated regions in MEDIPS. I don't
> quite understand how this relates to my Principal Component Analysis since
> I believe that I need to export my data out from MEDIPS and maybe use a
> different R package to do the PCA. So I am guessing that I need to export
> the methylation profile files for each sample and then input them into the
> PCA analysis software.
>

The motivation of CpG normalization is to quantify MeDIP signals at genomic
regions with varying CpG density across a genome. The differential
enrichment analysis compares IP-seq derived signals at genomic windows
across samples. Therefore, the signal for a specific window will not be
affected by different CpG densities across samples. The same assumption
applies when you are comparing different samples with each other employing
PCA. Due to a lack of experience I cannot recommend to use CpG normalized
methylation estimates or for example variance stabilized counts as done by
DESeq for RNA-seq data.

> 2) I have decided to try and do a PCA/hierarchical clustering using only
> the methylation levels of a specific genomic feature (eg. all gene
> promoters) to make things simple. I have decided to take your suggestion. I
> believe I can use the MEDIPS.createROIset() to calculate the methylation
> levels at all gene promoters. I can then use the methylation profile for
> the ROI (in this case, promoter) from each sample to do a a hierarchical
> clustering or PCA.
>
> 3) You said: " In case you want to apply a PCA on CpG density methylation
> estimates anyway, you can either use the rms values ......."...however, I
> still don't know if RMS is a linear or logarithmic value. If I have 4
> 100-bp windows where the RMS is 3, 5, 8, 9, can I say that the RMS for the
> this 400-bp region is 25? If RMS is a log value then obviously I cannot
> just add them up.
>
The rms values are not log transformed. However, the signals in
neighbouring windows are not independent. Extended reads spanning more than
one window will be counted several times. Therefore, adding up counts or
rms values is not valid. You can use MEDIPS.createROIset() instead.

>
> 4) Also, I was looking at the methylation profile file and I noticed that
> some RMS values are given as "Inf" or "NA". I am confused as to how to use
> or interpret this. I also don't understand why it is "NA"...why not just
> put a 0 (zero) instead of an "NA"?
>
What are the CF and count values for windows where rms values are NA or
Inf? Can you send examples?

>
> 5) Thanks for the suggestion to use the MEDIPS.meth() function to remove
> windows with low or zero counts. That's really helpful. I hadn't even
> considered that.
>
> Have a great weekend.
>
> Warmest regards,
> Allen
>
>
>
>
> ________________________________________
> From: Lukas Chavez [lukas.chavez.mailings at googlemail.com]
> Sent: Tuesday, September 09, 2014 1:09 AM
> To: Chong Kim San Allen
> Cc: Allen [guest]; bioconductor at r-project.org; MEDIPS Maintainer
> Subject: Re: [BioC] MEDIPS: how does MEDIPS define a methylated region /
> cluster?
>
> Dear Allen,
>
> I was referring to section 7.7 of the MEDIPS Tutorial (or vignette)
> available at:
>
> http://www.bioconductor.org/packages/release/bioc/vignettes/MEDIPS/inst/doc/MEDIPS.pdf
>
> I recommend to apply the a PCA to the counts at the genomic windows
> returned by the MEDIPS.meth() function where you probably want to remove
> windows with zero or low counts first. Here, the assumption is that CpG
> density correction of MeDIP signals might not be necessary as you are
> comparing different samples with each other (please see section 7.7 of the
> MEDIPS tutorial). In case you want to apply a PCA on CpG density
> methylation estimates anyway, you can either use the rms values or
> calculate methylation estimates using  BayMeth, a recent tool that appears
> to be superior for such methylation estimates.
>
> All the best,
> Lukas
>
> On Sun, Sep 7, 2014 at 11:10 AM, Chong Kim San Allen <patcksa at nus.edu.sg
> <mailto:patcksa at nus.edu.sg>> wrote:
> Dear Lukas,
> Again, I am so grateful for your quick reply.
>
> The only manual for MEDIPS that I know of is this one:
> http://www.bioconductor.org/packages/release/bioc/manuals/MEDIPS/man/MEDIPS.pdf
> and I have looked and just can't find the section 7.7 that you mentioned.
>
> The reason I asked about how MEDIPS defines a cluster is because I wanted
> to try and reduce the number of methylated regions that I look at and then
> use these "consolidated" regions to do a principal component analysis
> (PCA). I have 12 cell types and I wanted to see if I could either do a PCA
> or Unsupervised Hierarchical Clustering to separate these 12 cell types
> into subgroups, based on their methylation profile.
>
> You said: "When windows with significant differential enrichment between
> groups have been identified, the MEDIPS.mergeFrames() function can be used
> to merge adjacent significant windows into extended regions."
> But at this point, I don't have any groupings. Normally, one would have a
> control and test group and so you can do what you suggested. But I have
> only, for example, a test group and I want to know if this test group can
> be further subdivided into a few subgroups based on their methylation
> profile. I would think that it would be computationally intensive if I used
> original 100bp windows generated by MEDIPS and entered all the rms values
> for each window of each cell type into the PCA.
>
> You said: "Instead of merging any values across windows, I recommend to
> make use of the MEDIPS.createROIse() function which allows for analyzing
> specific regions of interest."
> I didn't want my PCA analysis to be biased by a predefined "region of
> interest", be it whether the ROI is a promoter, CpG, etc. I wanted to just
> input the raw methylation scores (rms) to see what I could get. So that is
> why I asked if the rms can be added up (in question 4 of my last
> email)...also I was curious if it could be done.
>
> Once again., thanks so very much for taking time from your busy schedule
> to answer my questions.
>
> Best regards,
> Allen
>
> ________________________________________
> From: Lukas Chavez [lukas.chavez.mailings at googlemail.com<mailto:
> lukas.chavez.mailings at googlemail.com>]
> Sent: Saturday, September 06, 2014 11:09 AM
> To: Allen [guest]
> Cc: bioconductor at r-project.org<mailto:bioconductor at r-project.org>; Chong
> Kim San Allen; MEDIPS Maintainer
> Subject: Re: [BioC] MEDIPS: how does MEDIPS define a methylated region /
> cluster?
>
> Dear Allen,
>
> please see my comments below. Moreover, please consider reading section
> 7.7 of the MEDIPS manual.
>
> Best,
> Lukas
>
>
> On Fri, Sep 5, 2014 at 1:14 AM, Allen [guest] <guest at bioconductor.org
> <mailto:guest at bioconductor.org><mailto:guest at bioconductor.org<mailto:
> guest at bioconductor.org>>> wrote:
> Hi,
> I have aligned my sequencing data for one sample (filename: Ca2_MAPQ20.bam
> --- because I filtered out everything with MAPQ20 or less) and I've put it
> through MEDIPS. At the moment, I am not making a comparison between
> samples. Therefore, the result file shows the methylation profile for only
> one sample. In this results file, I noticed that the data is organized as
> such:
> chr     start   stop    CF      Ca2_MAPQ20.bam.counts
>  Ca2_MAPQ20.bam.rpkm     Ca2_MAPQ20.bam.rms      Ca2_MAPQ20.bam.prob
>  MSets1.counts.mean      MSets1.rpkm.mean        MSets1.rms.mean
> MSets1.prob.mean
>
> 1) I tried reading the Down et al. (2008) paper that explains the concept
> of coupling factor and I think it is, simply put, a measure of local CpG
> density. I am not sure if my understanding of 'coupling factor' is correct?
>
> That is correct.
>
> 2) This lead to question how or what MEDIPS defines as a "region or
> cluster"? That is to say, on Chromosome 1, I have reads aligning from
> position "1002501" to "1003300" but then there is a region of 1400 bp (from
> "1003301" to "1004700") where there are no reads aligned (rpkm = 0). Then,
> again, from "1004701" to "1005400", there reads aligning to this region. So
> my question is does MEDIPS consider this to be a case of 2 methylated
> regions, that is, "1002501-1003300" and "1004701-1005400" or does MEDIPS
> try to consolidate these 2 methylated regions into one methylated
> region/cluster (that is, from "1002501-1005400") since they are fairly
> close to one another.
>
> Each window is considered as one region; CpG density, counts, rpkm, rms,
> and differential coverage will be calculated for each window separately.
> When windows with significant differential enrichment between groups have
> been identified, the MEDIPS.mergeFrames() function can be used to merge
> adjacent significant windows into extended regions.
>
> 3) I used "uniq=TRUE" to get rid of stacked/clonal reads and so for each
> 100 bp bin, I am usually getting rpkm=1 within each bin but occasionally,
> it may be as many rpkm=3. This lead me to wonder how the relative
> methylation score ("Ca2_MAPQ20.bam.rms") is calculated? For example, 3 bins
> have a value of 1 rpkm each, but one has an rms value of "1660.409928",
> another an rms of "7122.10254" and the third is only "679.1814523". How is
> that possible that 3 bins that have the same rpkm can have such varying rms
> values?
>
> This is due to the different CpG densities in these windows.
>
> 4)Can the rms be added up for a region so as to represent the cumulative
> methylation level for that region. I am asking because I do not know if the
> rms value is a log value or what? So, in my above question (3), if the 3
> bins are adjacent to one another and I decide to cluster them together and
> consider them as one methylated region, then would the rms value for this
> new 300bp bin that I have created be 1660.409928 + 7122.10254 + 679.1814523
> = 9461.68.
>
> Instead of merging any values across windows, I recommend to make use of
> the MEDIPS.createROIse() function which allows for analyzing specific
> regions of interest.
>
> 5)Finally, what does "Ca2_MAPQ20.bam.prob" represent? I figured that is
> some sort of a probability score but I am not sure of what.
>
> As mentioned in the manual, please disregard the prob values which will be
> removed in a future version.
>
> Thanks in advance for any assistance in answering my questions.
>
> Best regards,
> Allen
>
>  -- output of sessionInfo():
>
> error
>
> --
> Sent via the guest posting facility at bioconductor.org<
> http://bioconductor.org><http://bioconductor.org>.
>
> _______________________________________________
> Bioconductor mailing list
> Bioconductor at r-project.org<mailto:Bioconductor at r-project.org><mailto:
> Bioconductor at r-project.org<mailto:Bioconductor at r-project.org>>
> https://stat.ethz.ch/mailman/listinfo/bioconductor
> Search the archives:
> http://news.gmane.org/gmane.science.biology.informatics.conductor
>
> Important: This email is confidential and may be privileged. If you are
> not the intended recipient, please delete it and notify us immediately; you
> should not copy or use it for any purpose, nor disclose its contents to any
> other person. Thank you.
>
> Important: This email is confidential and may be privileged. If you are
> not the intended recipient, please delete it and notify us immediately; you
> should not copy or use it for any purpose, nor disclose its contents to any
> other person. Thank you.
>

	[[alternative HTML version deleted]]