Uncovering dynamics of age-related epigenetic changes with an interpretable deep-learning framework

Aging is the strongest risk factor for chronic diseases such as cardiovascular disease, Alzheimer’s, and cancer. DNA methylation (DNAm) clocks offer a promising measure of biological age, but most rely on linear models that miss non-linear dynamics and CpG interactions. To address this, we developed a deep neural network (DNN)-based DNAm clock trained on 29,167 samples profiled on Illumina EPIC v1.0 and v2.0 arrays. Using 12,234 CpGs selected through sex-and age-stratified correlations, our model achieved high accuracy (1.89 years) and outperformed published deep learning and elastic net based epigenetic clocks in a separate validation cohort. Using Shapley Additive Explanations (SHAP), we further uncovered phase-structured, wave-like dynamics in age-influential CpGs: an early-life module, a midlife transition, and late-life remodeling, with distinct timings by sex. These epigenetic waves cohere with non-linear, multi-omic “aging waves” reported in proteomics and longitudinal omics. SHAP further enabled interpretable CpG attribution, revealing structured, sex-specific aging phases: early-life male clocks involved developmental pathways, while female clocks emphasized cytoskeletal regulation; late-life divergence included immune activation in males and transcriptional remodeling in females. Our framework thus unites accuracy with mechanistic interpretability, revealing sex-specific windows when molecular aging reconfigures most rapidly.