Exploring a deep-learning epigenetic clock based on an interpretable convolutional neural network to unravel the tick-tack of cellular aging

In the field of aging research, DNA methylation patterns have emerged as valuable epigenetic biomarkers for modeling the passage of time at the molecular level. Through so-called first-generation epigenetic clocks, it is possible to estimate an individual’s chronological age with remarkable accuracy using the β-values of multiple CpG sites. Traditionally, these clocks have been built using machine learning models based on regularized linear regression (ElasticNet) for both feature selection and prediction. However, simple linear regression presents certain limitations, as such models are unable to capture nonlinear interactions between CpG sites or to model local dependencies among them. To overcome these constraints, recent approaches have explored deep learning methods capable of addressing these nonlinear and spatial relationships, although further research is still needed in this area. In this study, we compiled a large catalog of DNA methylation data from various tissues of healthy individuals differing in age, sex, and geographic origin. Using this dataset, we propose an approach based on a biologically interpretable convolutional neural network (CNN), which has been trained with images derived from methylation maps, in which CpG sites have been spatially organised according to their genomic position. Our model aims to reduce systematic errors in chronological age estimation and to help identify new genomic regions involved in the epigenetic changes associated with aging.