Bio Methylation Ewas Design

Version Compatibility

Reference examples tested with: meffil 1.3+, sva 3.50+, bacon 1.30+, limma 3.58+, missMethyl 1.36+, pwrEWAS 1.16+, pandas 2.2+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters
• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Two versions decide everything downstream. The ARRAY (450K vs EPIC v1 ~850K vs EPIC v2 ~935K) sets the CpG universe and therefore the effective number of tests and the genome-wide threshold - confirm against the current manifest before quoting a threshold. The CELL-TYPE REFERENCE panel must match the tissue and age (cord blood, adult blood, and tumor need different references); a mismatched reference silently biases the cell-fraction covariates. BACON slot names, pwrEWAS argument names, and sva method options have shifted across Bioconductor releases - check ?function on the installed build.

EWAS Design

"Run an EWAS for my phenotype" -> Build the covariate model and the chip randomization BEFORE the per-CpG test, because the confounders are larger than the signal - the p-value is the last and least interesting thing.

• R: meffil.ewas(beta, variable, covariates=data.frame(age, sex, cellprops, chip, plate)) then BACON-correct the test statistics.

Scope: the design-and-inference layer of an EWAS - confounding, randomization, batch/SVA/RUV correction, genomic inflation/BACON, genome-wide thresholds, power, meta-analysis, replication, and methylation risk scores. The per-CpG test mechanics (beta vs M-value, limma) -> differential-cpg-testing. Cell-fraction estimation algorithms (Houseman/IDOL/EpiDISH) -> cell-type-deconvolution. Normalization internals (funnorm/noob) -> array-preprocessing. mQTL Mendelian randomization for causal orientation -> causal-genomics/mendelian-randomization. MRS weight-learning (penalized regression, nested CV) -> the machine-learning category.

The Single Most Important Modern Insight -- An EWAS Hit Is a Cell-Composition Difference Until Proven Otherwise

The defining feature of the field is that the confounders are LARGER than the signal: a true effect is typically a sub-2% absolute methylation difference, while cell mix, batch, age, sex, smoking, and ancestry each move methylation 10-50%. An EWAS is therefore won or lost BEFORE the per-CpG test - in the chip randomization, the covariate set, and the replication cohort. Four corollaries, each of which a common misuse violates:

1. Composition before regulation. Whole blood is a leukocyte mixture; a CpG can be 90% methylated in granulocytes and 10% in lymphocytes. Any phenotype that shifts the mixture (age, inflammation, infection, stress, smoking, most diseases) produces methylation differences that are composition artifacts, not within-cell regulatory change. Cell-proportion covariates are mandatory, not optional; the default suspicion for any unadjusted hit is "a blood-count difference."
2. Confounder > signal. Design (randomization, matching) is the primary defense; statistical adjustment is the backstop. No analysis rescues a design that confounded chip/plate with phenotype at the bench.
3. Blood is the lamppost, not the keys. Blood is convenient; the disease tissue (brain, adipose, tumor) is usually inaccessible and weakly correlated (blood-brain mean r ~0.15, Hannon 2015). A blood EWAS for a brain trait tests a confounded surrogate, and a cross-sectional design cannot orient cause vs consequence (reverse causation).
4. Discovery is a hypothesis; replication is the finding. With tiny effects and pervasive confounding, a single-cohort genome-wide-significant CpG is a lead, not a result. The EWAS Catalog/Atlas exist so a "novel" hit can be checked against the generic age/smoking/cell-comp CpGs everyone finds.

Organize the analysis around defending these four, not around listing sva/ComBat/bacon functions.

The Confounding Hierarchy (ordered by damage, not convenience)

Default covariate set for a blood EWAS: age, sex, cell proportions, chip/position (or SVs), and where relevant genetic PCs and a smoking score. The omission of any one is the most common EWAS error.

Design Beats Correction -- The No-Rescue Theorem

If all cases ran on chip A and all controls on chip B, batch and phenotype are inseparable - ComBat will remove the batch AND the signal, or preserve a spurious one. No statistical adjustment recovers a confounded design (general principle: experimental-design/batch-design).

Goal: Make technical batch orthogonal to phenotype before any sample touches the array.

Approach: Randomize (or block) sample-to-chip, sample-to-position, and sample-to-plate assignment so case/control, age, and sex are balanced across chips; reserve no chip for one group. This single uncorrectable decision matters more than every analysis choice that follows.

# Stratified randomization of samples to 8-position EPIC BeadChips (450K: 12), balancing case/control per chip
meta$chip <- NA_integer_
for (grp in split(seq_len(nrow(meta)), meta$case)) {
  meta$chip[sample(grp)] <- rep(seq_len(ceiling(length(grp) / 4)), each = 4, length.out = length(grp))
}
# Then interleave the two case groups across chips so no chip is single-group; verify balance:
with(meta, table(chip, case))   # every chip should hold both groups

Batch / SVA / RUV Correction (and the over-correction trap)

The central tension: cell-composition and technical variation MUST be removed, but aggressive correction removes REAL biology when the unwanted variation overlaps the phenotype. Every tool above can erase a true effect.

The positive-control check (the pragmatic referee). Monitor a known signal through correction. In a blood EWAS containing smokers, smoking->cg05575921 (AHRR, ~18% hypomethylation, Joehanes 2016) MUST survive. If adding surrogate variables makes the QQ plot look clean BUT kills AHRR, the pipeline is over-corrected. A clean QQ with a dead positive control is a broken pipeline, not a good one. Do not keep adding SVs until lambda hits 1.0.

ComBat and limma/EWAS modeling run on M-values (logit of beta), not betas (betas are bounded [0,1] and heteroscedastic); effect sizes are reported back on the beta / delta-beta scale for interpretability. Pass the biological covariate to ComBat's mod= so it is protected. sva: pass the FULL model (mod, including the variable of interest) AND the null model (mod0) so SVs are orthogonal to the phenotype; choose the number with num.sv(method='be').

Genomic Inflation and BACON -- Why GWAS Intuition Fails

Lambda (the genomic inflation factor) = median observed chi-square / expected median. In GWAS, lambda > 1 signals stratification and is corrected by genomic control (divide all statistics by lambda). This reasoning is WRONG for EWAS:

1. EWAS statistics are routinely BOTH inflated AND deflated for non-GWAS reasons - residual cell-composition variation, the strong correlation among CpGs, un-modeled technical variation, and the fact that a strong exposure (smoking) genuinely associates with a large fraction of the genome (real signal that legitimately inflates lambda). A lambda of 1.2 may be real biology in a smoking EWAS and cell-composition leakage in an under-powered case-control study - the SAME number means different things.
2. Genomic control assumes a single multiplicative inflation on a mostly-null genome and ignores BIAS (a systematic mean shift off zero). EWAS violates all three assumptions.

BACON (van Iterson 2017 Genome Biol 18:19) fits a Bayesian Gaussian mixture to the observed test statistics, estimates the empirical-null distribution, and reports BOTH a bias (mean shift) AND an inflation (scale) - then standardizes statistics against that empirical null without assuming the genome is mostly null. Report lambda but correct with BACON; show QQ plots before and after. Caveat: BACON can over-deflate a highly polygenic exposure if the alternative component is large - apply it with the QQ plot, not as a reflex.

Genome-Wide Thresholds and the FWER-vs-FDR Decision

Naive Bonferroni on the probe count is over-conservative because CpGs are correlated (co-methylation, shared mQTLs); raw p is anti-conservative because there are ~850K tests. Use the array-specific FWER threshold for the cross-study-comparable headline claim and the EWAS Catalog; use BH-FDR for hypothesis-generating discovery. BH's independence assumption is imperfect under probe correlation but robust to positive dependence. Region-level (DMR) multiple-testing is owned by dmr-detection; mechanical p.adjust is owned by differential-cpg-testing. EPIC v2 (~935K probes, renamed/dropped vs v1) may shift the threshold - verify against the manifest.

Power and Meta-Analysis

Goal: Size an EWAS for a realistic (tiny) effect using site-specific methylation variance, not a single assumed sd.

Approach: pwrEWAS (Graw 2019 BMC Bioinformatics 20:218) simulates DNAm using empirical per-CpG variance from a reference dataset, so power reflects the real site-specific variance structure (bimodal sites near 0/1 behave differently from intermediate sites). Specify target delta-beta, sample size, and the genome-wide threshold; expect to need hundreds-to-thousands of samples for a 1-2% effect - which is why meta-analysis dominates.

EWAS meta-analysis is the field standard for power: each cohort runs an identical pre-specified pipeline (meffil enables this), BACON-corrects per cohort, uploads summary statistics, and a central site does fixed-effect inverse-variance pooling with heterogeneity testing (Cochran's Q, I^2). meffil's distinguishing feature is distributed normalization - cohorts normalize locally without sharing individual data, reducing meta-analysis heterogeneity - plus automated selection of the number of normalization PCs.

Study Designs

Replication and Look-Up

Replication is the real significance bar. After discovery, triage every hit against both databases - a "novel" CpG that is in fact a top smoking/age/blood-cell CpG is almost certainly residual confounding.

• EWAS Catalog (Battram 2022 Wellcome Open Res 7:41; ewascatalog.org) - published associations at P < 1e-4 (so a Catalog "hit" is a lookup, not a genome-wide claim) plus de-novo EWAS; check whether a CpG was reported for any trait.
• EWAS Atlas / Open Platform (Li 2019 Nucleic Acids Res 47:D983) - curated associations with a trait-ENRICHMENT tool for interpreting a CpG set.