DNA methylation-based predictors of metabolic traits in Scottish and Singaporean cohorts

Hannah M. Smith, Hong Kiat Ng, Joanna E. Moodie, Danni A. Gadd, Daniel L. McCartney, Elena Bernabeu, Archie Campbell, Paul Redmond, Adele Taylor, Danielle Page, Janie Corley, Sarah E. Harris, Darwin Tay, Ian J. Deary, Kathryn L. Evans, Matthew R. Robinson, John C. Chambers, Marie Loh, Simon R. Cox, Riccardo E. Marioni, Robert F. HillaryThe American Journal of Human Genetics 2024 · DOI

TL;DR

This study leverages DNA methylation (DNAm) data to develop "epigenetic scores" (EpiScores) that predict metabolic health traits like BMI and cholesterol. These EpiScores, trained on a large Scottish cohort, show promise in predicting metabolic traits and even cognitive function in independent Scottish and Singaporean populations, highlighting their broad applicability.

What Problem Does This Paper Address?

Metabolic health is crucial for overall well-being, but traditional measures like BMI can be imprecise. Researchers are exploring molecular markers, like DNA methylation, as more objective and potentially predictive indicators of metabolic health. However, identifying reliable DNAm associations with metabolic traits through epigenome-wide association studies (EWASs) can be challenging due to the complex, correlated nature of the methylome. This paper aims to address this by developing and validating DNAm-based predictors (EpiScores) for key metabolic traits and assessing their generalizability and utility in predicting cognitive function.

Methods

The study employed two primary EWAS approaches:

  • Marginal Linear Regression: This standard approach tests each DNA methylation site (CpG site) individually for association with a metabolic trait. To account for potential confounding factors and the complex structure of DNAm data, they performed analyses both with and without adjusting for the first 20 DNA methylation principal components (PCs). They also applied a bias and inflation correction method (bacon package).
  • Bayesian Penalized Regression (BayesR+): This more sophisticated method estimates the effect of each CpG site while simultaneously considering the effects of all other sites. This approach accounts for the correlation structure across the methylome, aiming to provide more robust and high-confidence associations.

Following the EWAS, epigenetic scores (EpiScores) were trained using penalized regression (elnet) on the Generation Scotland (GS) cohort. These EpiScores are essentially weighted sums of DNAm levels at specific CpG sites, designed to predict metabolic traits. The performance of these EpiScores was then evaluated by projecting them onto independent cohorts (LBC1936 and HELIOS) to assess their predictive power and generalizability.

Datasets

  • Generation Scotland (GS) cohort: A large, family-based Scottish cohort (N > 17,000) used for training the EWAS models and developing the EpiScores. DNAm data was measured using the Illumina EPIC array.
  • Lothian Birth Cohort 1936 (LBC1936): A Scottish longitudinal aging study (N = 861) used as an independent test cohort for evaluating the EpiScores. DNAm data was measured using the Illumina 450k array.
  • Health for Life in Singapore (HELIOS) cohort: A multi-ancestry cohort in Singapore (N = 2,245) used as another independent test cohort. DNAm data was measured using the Illumina EPIC array.

Key Insights and Ideas

  • The importance of statistical methodology: The choice of EWAS method significantly impacts the number of identified associations. Marginal linear regression, while identifying many potential sites, can be prone to false positives due to the correlated nature of DNAm. Bayesian penalized regression offers a more stringent approach, yielding fewer but more reliable "lead" findings.
  • EpiScores as powerful predictors: DNAm-based EpiScores can effectively predict metabolic traits, even in populations different from where they were trained. This suggests that these epigenetic signatures are not entirely population-specific.
  • Broader applicability of EpiScores: The study demonstrates that EpiScores trained in a Scottish cohort can generalize to a Singaporean cohort, indicating potential for cross-population utility.
  • EpiScores and cognitive function: Metabolic EpiScores showed associations with general cognitive function level in the LBC1936 cohort, suggesting a link between metabolic health, as reflected by DNAm, and cognitive abilities.

Key Results

  • EWAS findings: Marginal linear regression identified a large number of significant CpG associations (e.g., 57,307 for BMI). However, after adjusting for DNAm PCs and applying Bayesian methods, the number of high-confidence associations was substantially reduced (e.g., 27 for BMI).
  • EpiScore performance:
    • In the LBC1936 cohort, EpiScores explained: 14.4% of BMI variance, 3.2% of total cholesterol variance, and 18.5% of HDL cholesterol variance.
  • In the HELIOS cohort, EpiScores explained: 20.8% of BMI variance, 7.1% of total cholesterol variance, and 12.1% of HDL cholesterol variance.
  • Notably, the body fat percentage EpiScore showed differential performance across ethnic subgroups in HELIOS, explaining ~9% in Chinese and Malay subgroups but only ~3% in the Indian subgroup.
  • Cognitive associations: In the LBC1936, EpiScores for WHR, glucose, body fat percentage, and BMI were significantly associated with general cognitive function level (standardized β range: -0.08 to -0.12).

Limitations and Future Work

  • Methodological differences: The study acknowledges limitations due to methodological differences in measuring metabolic traits across cohorts (e.g., bioimpedance vs. DEXA for body fat).
  • Population-specific performance: While EpiScores showed generalizability, the performance varied across ethnic subgroups within the HELIOS cohort, suggesting that ancestry might influence the predictive power of these markers. Future work could explore training and testing across more diverse subpopulations.
  • Feature selection: Alternative strategies for pre-selecting CpG sites before training EpiScores could potentially lead to improved predictors.
  • Causality: The study notes that previous Mendelian randomization studies suggest DNAm might be a consequence rather than a cause of metabolic differences. Future work could prioritize lead CpGs to explore causal pathways.
  • Longitudinal cognitive changes: While associations with cognitive function level were found, no significant associations were observed with longitudinal cognitive change. Further longitudinal studies are needed.
  • Imputation potential: The study highlights the potential of using EpiScores to impute measured traits in studies where they are not available, which could be a valuable application in future research.

Summary updated

Abstract

Exploring the molecular correlates of metabolic health measures may identify their shared and unique biological processes and pathways. Molecular proxies of these traits may also provide a more objective approach to their measurement. Here, DNA methylation (DNAm) data were used in epigenome-wide association studies (EWASs) and for training epigenetic scores (EpiScores) of six metabolic traits: body mass index (BMI), body fat percentage, waist-hip ratio, and blood-based measures of glucose, high-density lipoprotein cholesterol, and total cholesterol in >17,000 volunteers from the Generation Scotland (GS) cohort. We observed a maximum of 12,033 significant findings (p < 3.6 × 10<sup>-8</sup>) for BMI in a marginal linear regression EWAS. By contrast, a joint and conditional Bayesian penalized regression approach yielded 27 high-confidence associations with BMI. EpiScores trained in GS performed well in both Scottish and Singaporean test cohorts (Lothian Birth Cohort 1936 [LBC1936] and Health for Life in Singapore [HELIOS]). The EpiScores for BMI and total cholesterol performed best in HELIOS, explaining 20.8% and 7.1% of the variance in the measured traits, respectively. The corresponding results in LBC1936 were 14.4% and 3.2%, respectively. Differences were observed in HELIOS for body fat, where the EpiScore explained ∼9% of the variance in Chinese and Malay -subgroups but ∼3% in the Indian subgroup. The EpiScores also correlated with cognitive function in LBC1936 (standardized β<sub>range</sub>: 0.08-0.12, false discovery rate p [p<sub>FDR</sub>] < 0.05). Accounting for the correlation structure across the methylome can vastly affect the number of lead findings in EWASs. The EpiScores of metabolic traits are broadly applicable across populations and can reflect differences in cognition.
BibTeX
@article{smith2024dna,
  author = {Hannah M. Smith and Hong Kiat Ng and Joanna E. Moodie and Danni A. Gadd and Daniel L. McCartney and Elena Bernabeu and Archie Campbell and Paul Redmond and Adele Taylor and Danielle Page and Janie Corley and Sarah E. Harris and Darwin Tay and Ian J. Deary and Kathryn L. Evans and Matthew R. Robinson and John C. Chambers and Marie Loh and Simon R. Cox and Riccardo E. Marioni and Robert F. Hillary},
  title = {{DNA methylation-based predictors of metabolic traits in Scottish and Singaporean cohorts}},
  journal = {The American Journal of Human Genetics},
  year = {2024},
  doi = {10.1016/j.ajhg.2024.11.012},
}
Figure