dmrcate: DMR identification

In timpeters82/DMRcate-release: Illumina 450K methylation array spatial analysis methods

Description

The main function of this package. Computes a kernel estimate against a null comparison to identify significantly differentially (or variable) methylated regions in hg19.

Usage

dmrcate(object, 
        lambda = 1000,
        C=2,
        p.adjust.method = "BH", 
        pcutoff = "limma", 
        consec = FALSE, 
        conseclambda = 10, 
        betacutoff = NULL
        ) 

Arguments

object	A class of type "annot", created from cpg.annotate.
lambda	Gaussian kernel bandwidth for smoothed-function estimation. Also informs DMR bookend definition; gaps >= lambda between significant probes will be in separate DMRs. Support is truncated at 5*lambda. Default is 1000 nucleotides. See details for further info.
C	Scaling factor for bandwidth. Gaussian kernel is calculated where lambda/C = sigma. Empirical testing shows that when lambda=1000, near-optimal prediction of sequencing-derived DMRs is obtained when C is approximately 2, i.e. 1 standard deviation of Gaussian kernel = 500 base pairs. Cannot be < 0.2.
p.adjust.method	Method for p-value adjustment from the significance test. Default is "BH" (Benjamini-Hochberg).
pcutoff	p-value cutoff to determine DMRs. Default is automatically determined by the number of significant probes returned by limma for that contrast, but can be set manually with a numeric value.
consec	Use DMRcate in consecutive probe mode. Treats CpG sites as equally spaced.
conseclambda	Bandwidth in probes (rather than nucleotides) to use when consec=TRUE. When specified the variable lambda simply becomes the minumum distance separating DMRs.
betacutoff	Optional filter; removes any region from the results that does not have at least one CpG site with a beta fold change exceeding this value.

Details

The values of lambda and C should be chosen with care. We recommend that half a kilobase represent 1 standard deviation of support (lambda=1000 and C=2). If lambda is too small or C too large then the kernel estimator will not have enough support to significantly differentiate the weighted estimate from the null distribution. If lambda is too large then dmrcate will report very long DMRs spanning multiple gene loci, and the large amount of support will likely give Type I errors. If you are concerned about Type I errors we recommend using the default value of pcutoff, although this will return no DMRs if no DM probes are returned by limma either.

Many gene loci have lengths reaching into the hundreds of thousands of base pairs, so it is quite possible that multiple signficant regions will have identical values in results$gene_assoc. This is fine; these regions are distinct in that they are at the very least lambda nucleotides apart, and is preferable to attempting collapse into a super-DMR by increasing lambda.

Value

A list containing 2 data frames (input and results) and a numeric value (cutoff). input contains the contents of the annot object, plus calculated p-values:

• ID: As per annotation object input

• weights: As per annotation object input

• CHR: As per annotation object input

• pos: As per annotation object input

• gene: As per annotation object input

• group: As per annotation object input

• betafc: As per annotation object input

• raw: Raw p-values from the significance test

• fdr: Adjusted p-values from the significance test

results contains an annotated data.frame of significant regions, ranked by minpval:

• gene_assoc: Complete list of gene loci overlapping the region, comma-separated

• group: Complete list of gene annotations (e.g. TSS1500, 5'UTR etc.) overlapping the region, comma-separated

• hg19coords: Coordinates of the significant region in hg19. IGV-friendly.

• no.probes: Number of probes constituting the significant region. Tie-breaker when sorting probes by minpval. A few regions may report no.probes=1, which may seem counter-intuitive, but this is only because the adjacent probes are either just below the significance threshold, or it is a highly DM probe in a sparse region. Unless pcutoff is highly conservative, it is unlikely that these regions will report at the head of the sorted list.

• minpval: Minimum adjusted p-value from the probes constituting the significant region.

• meanpval: Mean adjusted p-value from the probes constituting the significant region.

• maxbetafc: Maximum absolute beta fold change within the region

cutoff is the signficance p-value cutoff provided in the call to dmrcate.

Author(s)

Tim J. Peters <[e-mail]>, Mike J. Buckley <[e-mail]>, Tim Triche Jr. <[e-mail]>

References

Peters T.J., Buckley M.J., Statham, A., Pidsley R., Samaras K., Lord R.V., Clark S.J. and Molloy P.L. De novo identification of differentially methylated regions in the human genome. Epigenetics & Chromatin 2015, 8:6, doi:10.1186/1756-8935-8-6

Wand, M.P. & Jones, M.C. (1995) Kernel Smoothing. Chapman & Hall.

Duong T. (2013) Local significant differences from nonparametric two-sample tests. Journal of Nonparametric Statistics. 2013 25(3), 635-645.

Examples

## Not run: 
data(dmrcatedata)
myMs <- logit2(myBetas)
myMs.noSNPs <- rmSNPandCH(myMs, dist=2, mafcut=0.05)
patient <- factor(sub("-.*", "", colnames(myMs)))
type <- factor(sub(".*-", "", colnames(myMs)))
design <- model.matrix(~patient + type) 
myannotation <- cpg.annotate(myMs.noSNPs, analysis.type="differential",
    design=design, coef=39)
dmrcoutput <- dmrcate(myannotation, lambda=1000)

## End(Not run)

timpeters82/DMRcate-release documentation built on May 31, 2019, 2:29 p.m.