DMRcate reporting DMRs when all CpG's have very high FDR

[johnz...@gmail.com]

I am using DMRcate to call DMR's on RnBeads processed data. After cleaning the data and running SVA, I am using the code below to identify DMR's:

grset <- makeGenomicRatioSetFromMatrix(Mval.clean, what = "M",
                                       array = "IlluminaHumanMethylationEPIC",
                                       annotation = "ilm10b2.hg19", mergeManifest = TRUE)

dmr <- cpg.annotate("array", grset, arraytype="EPIC", what="M",
                    analysis.type="differential", design=design, coef=2)

dmrcoutput <- dmrcate(dmr, lambda=1000, C=2)
results.ranges <- extractRanges(dmrcoutput, genome = "hg19")


cpg.annotate says there's only 1 CpG with FDR < 0.99.

Then I increased the FDR to 1 just to be able to run dmrcate(). When running dmrcate(), a number of regions is reported as significant (~2,000) at FDR < 0.05, with low minfdr.

Why are significant DMR's reported? Are they truly significant even though individual CpG's are not?

You can download my R session here: https://www.dropbox.com/sh/sm43si4h4spioiu/AACNCqJ0yPYg1bte-mArekkHa?dl=0

Thank you!

[Tim Peters]

Hi John,

Significant DMRs are reported because you increased the FDR threshold in cpg.annotate() to 1. DMR significance is indexed at the rate of individual CpG sites - this is what this parameter controls. You've done the right thing by increasing this value to see if you start getting "significant" CpG sites, but unfortunately, with only 1 CpG with FDR < 0.99, the genome-wide effect size is negligible. So when you make fdr=1, this means every single CpG site in your fit will form a "DMR". For a reliable metric of overall significance I would use the Stouffer combined p-value - but in your case it's very likely that they are all equal to 1. So I think calling DMRs is a superfluous exercise for this particular hypothesis.

Best,

Tim