INLA methods identified hypervariable CpG sites faster than MCMC in healthy individuals

DNA methylation is an epigenetic regulator of gene expression and cell identity, which can be shaped by both physiological and pathological factors, including environmental exposure. The identification of sites with high methylation variability can be computationally challenging, especially in large-scale studies. To address this, we propose a framework based on the integrated nested Laplace approximation (INLA) to model methylation with Bayesian generalized linear mixed models (GLMMs), accounting for subject covariates, genomic annotations, and cell composition. To validate the methodology, we sequenced 158 healthy subjects with nanopore and analyzed a panel of 13 genes related to inflammation and stress response. We identified a set of hypervariable CpG sites whose genomic context and methylation levels were consistent with a regulatory role, making them potential candidates for epigenomic association studies. In our comparison, INLA results were concordant with those obtained with MCMC-based methods, with runtimes shorter by orders of magnitude. The computational efficiency of the framework allows for fast exploratory data analysis, model testing, and iterative prototyping, making it viable for large-scale studies that otherwise would be computationally prohibitive.