Precision medicine relies on the identification of robust disease and risk factor signatures from omics data. However, current knowledge-driven approaches may overlook novel or unexpected phenomena due to the inherent biases in biological knowledge. In this study, we present a data-driven signature discovery workflow for DNA methylation analysis utilizing network-coherent autoencoders (NCAEs) with biologically relevant latent embeddings. First, we explored the architecture space of autoencoders trained on a large-scale pan-tissue compendium (n = 75 272) of human epigenome-wide association studies. We observed the emergence of co-localized patterns in the deep autoencoder latent space representations that corresponded to biological network modules. We determined the NCAE configuration with the strongest co-localization and centrality signals in the human protein interactome. Leveraging the NCAE embeddings, we then trained interpretable deep neural networks for risk factor (aging, smoking) and disease (systemic lupus erythematosus) prediction and classification tasks. Remarkably, our NCAE embedding-based models outperformed existing predictors, revealing novel DNA methylation signatures enriched in gene sets and pathways associated with the studied condition in each case. Our data-driven biomarker discovery workflow provides a generally applicable pipeline to capture relevant risk factor and disease information. By surpassing the limitations of knowledge-driven methods, our approach enhances the understanding of complex epigenetic processes, facilitating the development of more effective diagnostic and therapeutic strategies.

Introduction DNA methylation (DNAm) age clocks are powerful tools for measuring biological age, providing insights into aging risks and outcomes beyond chronological age. While traditional models are effective, their interpretability is limited by their dependence on small and potentially stochastic sets of CpG sites. Here, we propose that the reliability of DNAm age clocks should stem from their capacity to detect comprehensive and targeted aging signatures.Methods We compiled publicly available DNAm whole-blood samples (n = 17,726) comprising the entire human lifespan (0-112 years). We used a pre-trained network-coherent autoencoder (NCAE) to compress DNAm data into embeddings, with which we trained interpretable neural network epigenetic clocks. We then retrieved their age-specific epigenetic signatures of aging and examined their functional enrichments in age-associated biological processes.Results We introduce NCAE-CombClock, a novel highly precise (R2 = 0.978, mean absolute error = 1.96 years) deep neural network age clock integrating data-driven DNAm embeddings and established CpG age markers. Additionally, we developed a suite of interpretable NCAE-Age neural network classifiers tailored for adolescence and young adulthood. These clocks can accurately classify individuals at critical developmental ages in youth (AUROC = 0.953, 0.972, and 0.927, for 15, 18, and 21 years) and capture fine-grained, single-year DNAm signatures of aging that are enriched in biological processes associated with anatomic and neuronal development, immunoregulation, and metabolism. We showcased the practical applicability of this approach by identifying candidate mechanisms underlying the altered pace of aging observed in pediatric Crohn's disease.Discussion In this study, we present a deep neural network epigenetic clock, named NCAE-CombClock, that improves age prediction accuracy in large datasets, and a suite of explainable neural network clocks for robust age classification across youth. Our models offer broad applications in personalized medicine and aging research, providing a valuable resource for interpreting aging trajectories in health and disease.