henhaobio
Sign in

bioRxivBioinformaticsPreprint

dmrff: identifying differentially methylated regions efficiently with power and control

Matthew J SudermaniD, James R StaleyiD, Robert French, Ryan ArathimosiD, Andrew J. SimpkiniD, Kate M. TillingiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

A differentially methylated region (DMR) is a genomic region in which DNA methylation is consistently positively or negatively associated with a phenotype or exposure. We demonstrate that existing algorithms for identifying DMRs either fail to consistently control false positive rates (comb-p and DMRcate), suffer from low power (bumphunter) or lack modeling flexibility (seqlm). We introduce a new method, dmrff, that overcomes these shortcomings and can additionally be used to meta-analyze multiple datasets. When applied to investigate associations of age in multiple datasets, dmrff identifies novel DMRs near genes previously linked to age. An R implementation is available on Github (http://github.com/perishky/dmrff).

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (31 Dec 2018) · CC BY-ND 4.0

Combining nearby CpG summary statistics by inverse-variance weighted meta-analysis holds false positives at the nominal level where comb-p and DMRcate inflate, and supports multi-cohort meta-analysis.

Region-level tests may gain power over site-by-site epigenome-wide association studies (EWAS), but comb-p and DMRcate lose error control where CpG correlation structure varies, bumphunter is slow and underpowered, and seqlm is fixed to one model with no obvious meta-analysis route. dmrff reuses EWAS effect sizes and standard errors with the measured CpG correlation matrix, so it controls the error rate, runs in minutes and can be applied across cohorts.

Key results

  1. ShownIn null simulations on real blood methylation profiles, comb-p and DMRcate lost false positive control, most severely when the data contained a strongly associated CpG site, while dmrff, seqlm, EWAS and the bumphunter 'area' statistic stayed at or below the nominal level.

    Simulated random phenotypes: comb-p exceeded 20% false positives on the chromosome 1 subset and bumphunter's 'value' statistic reached 8% on the high-correlation subset; on the augmented high-correlation subset comb-p rose to 91.7% and DMRcate to 51.6%, against a 5% expectation (DMRs overlapping the added sites were excluded). · False positive rates; Simulations; Null simulations

  2. ShownIn power simulations, dmrff ranked with seqlm and above EWAS and both bumphunter variants; comb-p and DMRcate scored higher but their power is not interpretable given their error rates.

    Phenotypes were built to correlate at 24 levels from 0 to 0.4 with the mean of 10 central CpG sites in a cluster, 100 variables per level, power being the fraction with an overlapping DMR call. Reported order: comb-p, then DMRcate, then seqlm and dmrff, then EWAS, then bumphunter; only seqlm, dmrff and EWAS are plotted, and no power values appear in the extracted text. · Power; Power simulations

  3. ShownAcross 14 public blood methylation datasets, dmrff found more age DMRs than seqlm or merged EWAS hits, including many regions containing no individually significant CpG site, and its calls replicated across the other datasets about as well as single CpG sites.

    dmrff and seqlm each found hundreds of DMRs (Bonferroni p < 0.05) with no EWAS-associated CpG site at a relaxed p < 0.2; in over 70% of datasets dmrff found thousands of DMRs absent from seqlm's relaxed list, while seqlm found at most 326 the other way. Replication by leave-one-dataset-out meta-analysis exceeded 96% in 12 of 14 datasets, within 0.13% of the CpG site rate. · Analysis of age; Replication of DMRs; DNA methylation datasets for analyses of age

  4. ShownFor meta-analysis, taking CpG correlations from each contributing dataset (two-step) gave larger and more distinctive regions than borrowing them from one reference dataset, so a reference should only be a fallback.

    Two-step found 70,464 age DMRs versus 83,549 for the reference approach, but 22,973 of them spanned more than one CpG site versus 12,156; among regions covering at least 8 CpG sites, 347 of 9,095 two-step regions were novel versus 9 of 204 reference regions, and the two-step approach gave at least 10x more novel regions at most size thresholds. · Meta-analysis; DNA methylation datasets for analyses of age

  5. SuggestedAge DMRs that contain no individually significant CpG site are not obviously artefacts: their genes were enriched for age-associated expression in blood, the tissue the methylation came from, but not in eight other tissues.

    Of 70,464 two-step DMRs, 482 had no EWAS association even at relaxed p < 0.2, leaving 123 linked genes after removing genes near any associated CpG site. Age-associated expression was reported for 51 (40%) of these in blood versus at most 21 (17%) elsewhere (adipose 14, artery 21, heart 6, lung 4, muscle 5, nerve 0, skin 0, thyroid 0); Wilcoxon p < 0.0012 in blood, p > 0.26 elsewhere. · Novel DMRs are functionally relevant

Methods, data and reproducibility

dmrff takes EWAS effect estimates and standard errors (here from limma) for CpG sites at most 500bp apart that share an effect sign and have nominal p < 0.05, and combines them by inverse-variance weighted meta-analysis adjusted for the CpG correlation matrix. Statistics are computed for every sub-region, candidates picked greedily, and p-values Bonferroni-adjusted over all tests.

  • Blood methylation profiles (Illumina 450k) used for null and power simulations, reduced to a chromosome 1 subset (2577 sites), a high-correlation subset (8906 sites) and an augmented version · human · n = 464 profiles · GSE50660
  • Public peripheral blood 450k datasets from GEO for the analysis of age: 20 eligible, 6 dropped for low cross-dataset agreement (mean correlation < 0.25) · human · n = 14 datasets meta-analysed · GSE40279, GSE42861, GSE55763, GSE87571; all twenty accessions are listed in the paper
  • Dataset chosen as the correlation reference for reference meta-analysis (highest mean agreement in pairwise CpG correlations) · human · n = mean R = 0.75 · GSE72680
  • Published associations of gene expression with age across nine tissues (adipose, artery, heart, lung, muscle, nerve, skin, thyroid, blood), used to test enrichment of DMR-linked genes · human · n = nearly two hundred individuals; 41298 genes
Methods & tools
dmrff, comb-p, DMRcate, bumphunter, seqlm, limma, sva, metafor, Illumina Infinium HumanMethylation450 BeadChip
Reagents & environment
An R implementation is said to be on GitHub; no version, container or dependency list is given.
Experimental conditions
comb-p, DMRcate, seqlm and dmrff each finish a genome-wide dataset in a few minutes, comparable to an EWAS; bumphunter takes about 10x longer with 100 bootstraps.
Reproducibility gaps
No GitHub URL for the R package appears in the extracted text, and figures and tables are missing, so false positive rates, power curves and per-dataset DMR counts are only partly quantified in the prose; the input data are public GEO accessions, listed without links.
Limitations (5)
  • Optimising region boundaries by scoring every sub-region adds tests and so costs some power. · stated by the authors
  • Meta-analysis needs CpG correlations from each contributing dataset, requiring two rounds of exchange with analysts or a precomputed correlation file; where original data are gone, a reference dataset works but is less powerful. · stated by the authors
  • The greedy sub-region search is not suited to whole-genome bisulfite sequencing, where dense CpG coverage would generate far too many tests. · stated by the authors
  • Simulations use small subsets of one 450k blood dataset, with competing tools at mostly default settings and bumphunter run on 100 rather than 1000 simulations, so absolute error and power estimates may not transfer to other designs. · noted in this summary
  • The age analysis has no independent ground truth: more DMRs is treated as better, and functional support rests on an enrichment test in blood for 123 genes, with two datasets (GSE55763, GSE87571) behaving unlike the rest at 87.4% and 61.2% CpG replication. · noted in this summary

Read this if you run EWAS on Illumina methylation arrays and need region-level calls, or a multi-cohort DMR meta-analysis, with evidence about which existing tools hold their error rate.

Citations
120

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGomez '19Daito '26Zhang '24Srivastava '19Korthauer '18Martorell‐M '19
66% semantic matchBioinformatics
Lissette Gomez, Gabriel Odom, Juan I. Young+7 more·Nucleic Acids Research·Jul 8, 2019· 37 citations