Abstract

Epigenetic clocks, predictors of biological age built from DNA methylation (DNAm) data, are widely used as biomarkers of aging and disease risk. Yet they are typically reported as single point predictions, with no measure of individual-level uncertainty. Because these predictors are built from ultra-high-dimensional, highly correlated CpG sites, this uncertainty can be substantial. Ignoring it may lead to over-interpreting differences in biological age or epigenetic age acceleration, and to downstream analyses that disregard the error carried by the prediction itself.
In this talk, I present a Bayesian approach to epigenetic prediction that provides credible intervals for each individual's predicted age, allowing uncertainty to be quantified and propagated into subsequent analyses. To make this feasible with hundreds of thousands of CpG sites, the approach is combined with a pre-selection step that reduces the data to a smaller, informative set before model fitting. I illustrate the methods using data from a cohort of people with multiple sclerosis (MS) at The Ohio State University and show how accounting for prediction uncertainty affects the conclusions drawn about epigenetic aging and clinical outcomes.