Multi-Omics Epigenome-Wide Association Study

A standard EWAS identifies CpG sites statistically associated with a phenotype using methylation data alone. Multi-omics EWAS adds at least one additional omics layer — typically gene expression, genetic variants, or proteomics — and formally integrates them to test whether the methylation signal is accompanied by concordant molecular changes, enabling mechanistic interpretation and stronger causal inference.

Power for the methylation scan typically requires at least 200–500 samples with the phenotype contrast of interest; well-powered EWAS in common diseases often use thousands of samples. Each additional omics layer must be measured in the same individuals, so the practical constraint is usually the cost and availability of matched multi-omics data. Below ~200 matched samples, false-discovery risk is high even with stringent thresholds.

The choice depends on the research question. meQTL / mQTL mapping asks which CpGs are regulated by nearby genetic variants and whether those CpGs in turn regulate expression — relevant for causal inference via Mendelian randomisation. Mediation analysis tests whether methylation statistically mediates a known exposure-outcome path. MOFA and similar latent-factor methods are hypothesis-generating: they find shared axes of variation across all layers without assuming a prior biological direction. Most comprehensive studies use more than one strategy.

Yes. TCGA provides matched methylation and expression data for many cancer types. The UK Biobank and Generation Scotland provide population-level matched methylation and genetic data. GTEx links methylation to expression across tissues. Combining these resources requires careful harmonisation of sample identifiers and batch correction across independently processed datasets, and any results should be replicated in a held-out or independent cohort.

Blood is a mixture of cell types (granulocytes, lymphocytes, monocytes, etc.) whose proportions vary between individuals and confound methylation levels. The standard approach is to estimate cell-type proportions using reference-based deconvolution (Houseman algorithm, EpiDISH, or similar) and include the estimated proportions as covariates in the regression model. If the phenotype of primary interest is a non-blood tissue, you should also consider whether blood methylation is a valid proxy.