Variational autoencoders to predict DNA-methylation age and provide biological insights in age-related health and disease

Motivation Epigenetic ageing clocks based on age-associated changes in DNA-methylation (DNAm) profiles have shown great potential as predictors to quantify age-related health and disease. Most of these clocks are built using penalized linear regression models with the aim to predict chronological and/or biological age. However, the precise underlying biological mechanisms influencing these clocks remain unclear. In this work, we explored if a Variational Autoencoder (VAE) can be trained to model DNAm age, and whether this VAE captures biologically relevant features of ageing in the latent space. Results Our results indicate that VAEs present a promising framework to construct embeddings that capture complex interactions and that can be used to extract biological meaningful features upon predicting DNAm age. By using deep learning interpretation methods, we show it is possible to determine which genomic loci and pathways are important in making the predictions, both on a population and individual level, paving the way to unravel what makes the DNAm clock tick.