Abstract
DNA methylation (DNAm) patterns, or epigenetic scores (EpiScores), can serve as stable biomarkers for complex traits, potentially offering a more consistent assessment of chronic exposures than single-point protein measurements. This study presents the first large-scale epigenetic analysis of the high-abundance serum proteome, measured via mass spectrometry in 14,671 participants from the Generation Scotland cohort. We identified thousands of significant CpG-protein associations and developed 112 validated protein EpiScores. Notably, sixteen of these EpiScores were significantly associated with incident cardiovascular disease (CVD), outperforming directly measured proteins. These findings suggest that protein EpiScores are powerful clinical biomarkers for disease risk prediction.
Introduction
Circulating proteins are vital indicators of physiological state, yet most proteomic research focuses on low-abundance signaling molecules. The "high-abundance" proteome—comprising 99% of total protein mass, including albumin and transport proteins—remains understudied despite its roles in immunity and coagulation. Furthermore, protein levels can be transiently affected by factors like diet or acute infection. In contrast, DNA methylation (DNAm) is relatively stable and may capture cumulative biological history. This research explores whether DNAm patterns can act as reliable proxies (EpiScores) for high-abundance proteins and improve the prediction of chronic conditions like cardiovascular disease (CVD).
Methods and Study Design
The study utilized data from 14,671 participants in the Generation Scotland cohort. Researchers employed two primary statistical frameworks to map the relationship between CpG sites and 133 proteins:
- OSCA (Marginal Linear Regression): Tested individual CpG sites against proteins, adjusting for age, sex, batch, and other covariates.
- GMRMomi (Bayesian Penalized Regression): A joint modeling approach used to fine-map influential CpG sites and generate parsimonious models for EpiScore creation.
The cohort was split into a training set (N=6,816) and a holdout test set (N=3,463) to validate the EpiScores. Finally, Cox proportional hazards models assessed the predictive power of these scores for incident CVD over a 17-year follow-up.
Key Results
The marginal models identified 15,855 significant CpG–protein associations across 125 proteins. The more stringent Bayesian approach narrowed this to 697 high-confidence associations. Validation in the test set confirmed that 112 of the 133 protein EpiScores significantly correlated with their measured protein counterparts.
Pleiotropy and Regulatory Insights
The study highlighted significant pleiotropy, where specific CpG sites were associated with multiple proteins. For example, one site (cg06072257) was linked to 22 different proteins, suggesting a broad regulatory role in metabolic processes.
Clinical Predictive Power
The most significant finding was the superior performance of EpiScores in predicting incident CVD. While only one directly measured protein showed a significant association with CVD, sixteen protein EpiScores were significant predictors. For instance, the EpiScore for Gelsolin (GSN) showed a clear protective association (Hazard Ratio: 0.61), which was not as evident in the raw protein measurements.
Discussion
The "chronic advantage" of EpiScores likely stems from DNAm's ability to reflect long-term environmental and biological impacts, whereas protein measures provide only a temporal snapshot. The discovery of strong associations between blood-based DNAm and liver-synthesized proteins suggests that white blood cell methylation captures systemic states like inflammation that influence overall proteomic output.
While limited by a primarily European ancestry cohort and a cross-sectional design for the initial mapping, this work demonstrates that the high-abundance proteome is a rich source of clinical biomarkers. Future diagnostic and prognostic tools may benefit from integrating both direct protein measurements and their corresponding epigenetic proxies to better monitor and predict disease risk.
Summary updated