Deep Learning allows accurate age prediction from blood DNA methylation
2Department of Developmental Biology and Cancer Research, The Hebrew University of Jerusalem, Israel
3Department of Endocrinology and Metabolism, Hadassah Medical Center and Faculty of Medicine, The Hebrew University of Jerusalem, Israel
DNA methylation is known to be associated with gene regulation, health and aging. Age prediction using methylation thus paves the way for a simple blood test which indicates poor health and lifespan predictions. Previous studies developed epigenetic clocks based on linear regression of hundreds of DNA methylation sites, measured using methylation arrays. Yet, the high cost and long turnaround time of arrays hamper the widespread applicability of such clocks. Conversely, multiplex targeted PCR allows fast and simple sequencing of a set of predefined age-associated loci.
Here, we present deep learning models that use fully connected networks to predict chronological age, based on the binary methylation patterns across multiple neighboring CpG sites. This model predicts age with high accuracy (median absolute error ≤ 2 years), outperforming state-of-the-art predictors that use hundreds of CpG sites. Overall, our approach offers a compelling alternative to array-based methylation clocks, with various future applications including forensic profiling, monitoring epigenetic processes in cancer and in transplantation medicine, and a useful accurate tool for studying aging and rejuvenation.