Longitudinal prediction of DNA methylation to forecast epigenetic outcomes

Abstract

Epigenetic changes in early life play an important role in the development of health conditions in children. Longitudinally measuring and forecasting changes in DNA methylation can reveal patterns of ageing and disease progression, but biosamples may not always be available. We introduce a probabilistic machine learning framework based on multi-mean Gaussian processes, accounting for individual and gene correlations across time to forecast the methylation status of an individual into the future. Predicted methylation values were used to compute future epigenetic age and compared to chronological age. We show that this method can simultaneously predict methylation status at multiple genomic sites in children (age 5–7) using methylation data from earlier ages (0–4). Less than 10% difference between observed and predicted methylation values is found in approximately 95% of methylation sites. We show that predicted methylation profiles can be used to estimate other molecular phenotypes, such as epigenetic age, at any timepoint and enable association tests with health outcomes measured at the same timepoint. Limited longitudinal profiling of DNA methylation coupled with machine learning enables forecasting of epigenetic ageing and future health outcomes.