BayesAge: A Maximum Likelihood Algorithm To Predict Epigenetic Age

Abstract

DNA methylation is a reaction that results in the formation of 5-methylcytosine when a methyl group is added to the cytosines C5 position. As organisms age, DNA methylation patterns change in a reproducible fashion. This phenomenon has established DNA methylation as a valuable biomarker in aging studies. Epigenetic clocks based on weighted combinations of methylation sites have been developed to accurately predict the age of an individual from their methylome. However, many epigenetic clocks, particularly those that utilize penalized regression, model the changes in methylation linearly with age. Moreover, these models, which use methylation levels as features, are not robust to missing data and do not account the count-based nature of bisulfite sequence data. Additionally, the models are generally not interpretable. To overcome these challenges, we present BayesAge, an extension of the previously developed scAge approach that was developed for the analysis of single cell DNA methylation datasets. BayesAge utilizes maximum likelihood estimation (MLE) to infer ages, models count data using binomial distributions, and uses LOWESS smoothing to capture the non-linear dynamics between methylation and age. Our approach is designed for use with bulk bisulfite sequencing datasets. BayesAge outperforms scAge in several respects. Specifically, BayesAges age residuals are not age associated, thus providing a less biased representation of epigenetic age variation across populations. Moreover, BayesAge enables the estimation of error bounds on age inference and, when run on down-sampled data, its coefficient of determination between predicted and actual ages surpasses both scAge and penalized regression.