dna-methylation-analysis

Overview

dna-methylation-analysis covers measuring and analyzing DNA methylation (5mC): Illumina arrays (450K/EPIC) and bisulfite sequencing (WGBS/RRBS), processing pipelines (minfi, Bismark), differential methylation (DMPs/DMRs), cell-type deconvolution, and epigenetic clocks. Methylation is tissue- and cell-type-specific — every analysis must reckon with cellular heterogeneity.

When to use

• Platform choice: EPIC arrays vs WGBS vs RRBS vs targeted bisulfite-seq.
• Array processing: minfi pipeline — QC, normalization (functional/BMIQ), probe filtering.
• Bisulfite-seq: Bismark alignment, methylation calling, coverage requirements.
• DMP/DMR analysis: limma-based, bumphunter, DMRcate.
• Cell-type heterogeneity: reference-based deconvolution, ReFACTor.
• Epigenetic clocks: Horvath, Hannum, PhenoAge, GrimAge — computing and interpreting.
• EWAS design: confounders, batch effects, multiple testing at 850K CpGs.

Core concepts

• Platform trade-offs. EPIC arrays (~935K CpGs): cheap, standardized, great for EWAS — but biased to promoters/CpG islands and miss most of the methylome. WGBS: whole-genome, unbiased, expensive (30x needed), heavy compute. RRBS: enriches CpG-dense regions at moderate cost. Targeted: deep coverage of candidate regions. Match platform to question and budget.
• Beta vs M values. Beta (0-1, interpretable as methylation proportion) for display; M-value (logit, homoscedastic) for statistics. Run tests on M-values, present betas — testing on betas violates linear-model assumptions at the extremes.
• Array QC (minfi). Detection p-values (failed probes/samples), control probes (bisulfite conversion, staining, hybridization), sex prediction vs reported sex (mislabeled samples), SNP probes for sample identity tracking. Remove failed samples before normalization, not after analysis.
• Normalization. Functional normalization (uses control probes — good default), BMIQ (Type I/II probe bias correction), Noob (background + dye-bias). Type I vs Type II probe chemistry differences must be corrected — uncorrected, they dominate the signal. Check density plots before/after: bimodal betas (0/1 peaks) should be clean.
• Probe filtering. Cross-reactive probes (Chen et al. list), SNP-overlapping probes (genetic variation masquerading as methylation), sex chromosomes (unless studying them). Filter before analysis — these probes generate false DMPs reliably.
• Cell-type heterogeneity. Blood methylation reflects cell composition; a "disease DMP" may just be shifted neutrophil/lymphocyte ratios. Solutions: reference-based deconvolution (Houseman/IDOL libraries) with cell proportions as covariates; ReFACTor (reference-free) when no reference exists; or sort cells upfront (best, expensive). Always address this — reviewers will ask.
• DMPs and DMRs. DMPs: limma on M-values with FDR (850K tests — genome-wide significance ~1e-7-ish by Bonferroni, FDR standard). DMRs: bumphunter/DMRcate find spatially correlated regions — biologically more interpretable than isolated CpGs (single-CpG hits are often technical). Report effect sizes (delta-beta): statistically significant 1% methylation differences are rarely biologically meaningful.
• Batch effects. Arrays are batch-sensitive (plate, chip position): randomize samples across chips, include batch in models, use ComBat/sva when needed. Never confound batch with phenotype.
• Epigenetic clocks. Horvath (multi-tissue), Hannum (blood), PhenoAge/GrimAge (mortality-trained) predict biological age; age acceleration (residual vs chronological age) associates with disease/mortality. Compute with published coefficients; interpret as biomarker, not as "true aging rate" — clocks are correlative and tissue-dependent.

Practical workflow

1. Design. Platform choice; randomization across batches; collect cell-count or sorting data; power for 850K tests (n in hundreds for EWAS).
2. QC. Detection p-values, control probes, sex check, identity SNPs; drop failures.
3. Normalize. Functional normalization + BMIQ; probe filtering (cross-reactive, SNP, sex); density-plot checks.
4. Deconvolve. Cell proportions estimated; include as covariates (or analyze sorted cells).
5. Test. limma on M-values (DMPs, FDR); bumphunter/DMRcate (DMRs); delta-beta effect sizes; batch in the model.
6. Interpret. Annotate to genes/regulatory features; pathway enrichment on DMR genes; validate top loci by targeted bisulfite-seq/pyrosequencing.
7. Clocks (if relevant). Compute age acceleration; associate with phenotype; report clock version.
8. Report. Platform, normalization, filtering, cell-type handling, full summary statistics, GEO deposition.

Example (R sketch):

library(minfi)
rgSet <- read.metharray.exp("idats/")
qc <- getQC(preprocessRaw(rgSet)); plotQC(qc)
mSet <- preprocessFunnorm(rgSet)          # functional normalization
beta <- getBeta(mSet); M <- getM(mSet)
fit <- eBayes(lmFit(M, design))           # DMPs on M-values

Common pitfalls

• Testing on beta values instead of M-values.
• Cell-type confounding presented as disease signal.
• Cross-reactive/SNP probes left in (false DMPs).
• Batch confounded with phenotype.
• Single-CpG hits over-interpreted (prefer DMRs).
• Tiny delta-betas (1-2%) treated as biologically meaningful.
• Array findings extrapolated to the whole methylome.
• Epigenetic age acceleration equated with proven aging biology.