DeepStrataAge: an interpretable deep-learning clock that reveals stage- and sex-divergent DNA methylation aging...

Area of Science:

• Epigenetics and computational biology
• Genomics and aging research
• Systems biology and multi-omics

Background:

• Aging is a primary risk factor for major chronic diseases.
• Current DNA methylation (DNAm) clocks often use linear models, missing complex biological dynamics.
• There is a need for more accurate and mechanistically interpretable biological age measures.

Purpose of the Study:

• To develop a deep neural network (DNN)-based DNAm clock for precise biological age estimation.
• To identify non-linear dynamics and sex-specific patterns in epigenetic aging.
• To uncover mechanistic insights into the aging process using interpretable AI.

Main Methods:

• Trained a DNN model on 29,167 samples using Illumina EPIC v1.0 and v2.0 arrays.
• Selected 12,234 CpGs based on sex- and age-stratified correlations.
• Employed Shapley Additive Explanations (SHAP) for model interpretability and uncovering epigenetic dynamics.

Main Results:

• Achieved high accuracy in biological age prediction (1.89 years), outperforming existing clocks.
• Discovered phase-structured, wave-like dynamics in age-influential CpGs with distinct sex-specific timings.
• Identified sex-specific epigenetic aging phases related to developmental, cytoskeletal, immune, and transcriptional pathways.

Conclusions:

• The DNN-based DNAm clock offers superior accuracy and mechanistic interpretability for biological age.
• Epigenetic aging exhibits non-linear, sex-specific wave-like patterns that align with multi-omic aging signatures.
• This framework reveals critical sex-specific windows of rapid molecular aging, advancing our understanding of age-related disease.