Bayesian Epigenome-Wide Association Study (Bayesian EWAS)

A standard EWAS fits a linear (or logistic) regression at each CpG site and reports a p-value; significance is declared if p falls below a genome-wide threshold (typically 3.6×10⁻⁸ for the 450K array). A Bayesian EWAS instead computes the posterior probability that the association at each site is non-zero, given the data and a prior. This allows external biological knowledge to be incorporated, provides direct uncertainty quantification via credible intervals, and replaces the binary significance decision with a continuous ranking by posterior evidence.

In the absence of strong external information, a weakly informative prior — such as a normal prior on regression coefficients centered at zero with a variance calibrated to the expected effect-size range in epigenomic studies — is a reasonable default. Empirical Bayes, which estimates hyperparameters from the data itself, is a practical alternative that avoids fully subjective prior choices. The key requirement is to document your prior and run sensitivity analyses varying the hyperparameters to confirm that conclusions are robust.

In large biobanks (n > 5,000), classical EWAS is already well-powered and computationally straightforward. Full Bayesian MCMC across 450K–900K sites at this scale is typically impractical without distributed computing. Variational Bayes approximations or empirical Bayes shrinkage estimators offer a practical middle ground. For many large-cohort questions, the added complexity of a fully Bayesian approach may not yield enough additional insight to justify the computational cost.

Yes. The statistical framework is platform-agnostic; what changes is the preprocessing pipeline and the number of sites tested (the EPIC v2 array covers approximately 935,000 CpG sites). The increased dimensionality raises the computational burden further and makes regularisation through Bayesian priors even more valuable relative to naive frequentist approaches.

Nearby CpG sites in the same genomic region are correlated (linkage disequilibrium of methylation, or 'methylation LD'). Site-by-site Bayesian regression ignores this structure. More sophisticated approaches model spatially correlated effects using Gaussian process priors or region-level summarisation (e.g., bumphunter-style region detection followed by Bayesian region-level testing). Ignoring spatial correlation does not necessarily bias point estimates but can inflate the effective number of independent discoveries.