Abstract
Our study identified the relationship between epigenetic biomarkers and the risk of diabetes and cancer, and used machine learning techniques to analyze the contributions of various epigenetic biomarkers to disease risk.
Introduction
Chronological age is often a "noisy" metric for assessing health; two individuals of the same age can exhibit vastly different physiological profiles based on genetics, environment, and lifestyle. Epigenetic biomarkers, specifically DNA methylation (DNAm) patterns, have emerged as a more precise "biological clock" that reflects the cumulative wear and tear on the human body. While previous research has linked specific epigenetic clocks to chronic diseases, most studies have focused on a narrow set of markers. This paper addresses the need for a comprehensive, multi-model comparison to identify which specific epigenetic signals—including clocks, immune markers, and mortality-related proxies—best predict the risk of diabetes and cancer.
Methodology
Data Source and Participants
The study utilized data from the National Health and Nutrition Examination Survey (NHANES) database spanning 2011–2018. The cancer model included 1,291 participants (177 cases), while the diabetes model included 2,205 participants (394 cases). Epigenetic data were derived from blood samples using the Illumina MethylationEPIC BeadChip.
Feature Selection
The researchers analyzed 30 distinct epigenetic biomarkers alongside clinical and sociodemographic data, including sex, age, race, education, marital status, BMI, and the family income-to-poverty ratio. These biomarkers encompass a wide range of biological signals, from estimated chronological age to predicted immune cell proportions.
Machine Learning Pipeline
To identify the most effective predictive models, the team compared nine different algorithms: AdaBoost, GBM, KNN, LightGBM, MLP, Random Forest (RF), SVM, XGBoost, and Logistic Regression. To address the class imbalance inherent in disease datasets, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. Model interpretability was achieved through SHapley Additive exPlanations (SHAP) values, and clinical utility was evaluated using Decision Curve Analysis (DCA).
Key Results and Insights
Model Performance
LightGBM consistently emerged as the top-performing algorithm. The diabetes model achieved an impressive Area Under the Curve (AUC) of 0.96, while the cancer model reached an AUC of 0.98. DCA curves indicated that LightGBM and Gradient Boosting provided the highest net benefit for clinical screening scenarios.
Drivers of Cancer Risk
For cancer, epigenetic age acceleration—the discrepancy between biological and chronological age—was a primary driver. As shown in the comparison between tumor and non-tumor patients, nearly all epigenetic clocks (e.g., HorvathAge, HannumAge, PhenoAge) showed statistically significant differences (p < 0.001), with tumor patients exhibiting higher biological ages.
Beyond age clocks, gender, ethnicity (specifically non-Hispanic White), and PACKYRSMort (an epigenetic proxy for smoking history) were top contributors to the cancer model. Interestingly, cancer risk showed a non-linear relationship with smoking, increasing significantly between 10 and 30 pack-years before showing a declining trend, potentially due to survival bias.
Drivers of Diabetes Risk
In contrast to cancer, diabetes risk was more heavily influenced by metabolic proxies and socioeconomic factors. The biomarker logA1Mort—a DNAm-based proxy for glycated hemoglobin (HbA1c)—emerged as the most critical feature. This suggests that epigenetic signatures of long-term blood sugar levels are more predictive for diabetes than general biological clocks. Additionally, the family income-to-poverty ratio was a top-three feature, with a sharp decline in risk observed once the ratio exceeded 4, highlighting how economic stressors are biologically encoded.
Discussion and Limitations
This study demonstrates that machine learning models incorporating a broad array of epigenetic biomarkers can accurately predict the risk of chronic diseases. However, several limitations remain. The cross-sectional design prevents the establishment of definitive causality. Furthermore, the reliance on North American NHANES data may limit generalizability to global populations. Future research should focus on longitudinal data, multi-omics integration (combining epigenetics with proteomics and transcriptomics), and greater granularity regarding specific types of cancer and diabetes.
Summary updated