From Clinical Risk to Biological Aging - A Modular AI Framework for Traceable CpG Biomarker Prioritization

Abstract: Artificial intelligence is making real inroads in epigenetics, especially when it comes to using DNA methylation to dig into aging and disease. Predictive models do a decent job at flagging CpG sites tied to clinical outcomes, but there’s a problem—they’re often black boxes. You get predictions, but the biological context stays hidden. That disconnect makes it tough to reproduce findings or pull out insights that actually matter for biomedicine. That’s what pushed us to build a modular framework that links CpG selection straight to biological annotation. Instead of predictions floating on their own, you can track them back to the biology of aging.
We worked with a Canadian methylation cohort—92 samples, over 485,000 CpG sites. After normalizing and filtering the data, we ran log-rank tests, Random Survival Forests, Cox proportional hazards models, and ensemble learners to find CpGs tied to simulated survival outcomes. SHAP values helped us spotlight the CpGs that reliably drove predictions. Then we mapped these CpGs to enhancer regions, transcription factor motifs, and downstream pathways. Visualizations made it clear how each marker connects to regulatory elements and broader biological processes.
Our framework delivered stable concordance indices across training, test, and cross-validation sets. Certain CpGs kept showing up as influential and linked back to genes like IL7R and CDKN2A—core players in immune aging and cell- cycle regulation. Because we combined modeling with annotation, it’s easier to trace how methylation changes might actually influence biological pathways.
In the end, this approach gives researchers a practical, interpretable way to prioritize CpG biomarkers. By weaving together statistical modeling and enhancer-aware annotation, our framework nudges methylation research toward transparency and real biological relevance, especially in aging and personalized medicine.

Keywords: Artificial Intelligence, Bioinformatics, DNA methylation, CpG biomarkers, aging biology, artificial intelligence, survival modeling, SHAP, enhancer annotation, epigenetics