Study population
The GOLDN study determined genetic predictors of participant lipid and inflammatory response to diet and fenofibrate treatment interventions [13,14,15]. Participants were identified through 3-generation families previously screened in the National Heart, Lung and Blood Institute Family Heart Study (FHS) Minnesota and Utah centers. Individuals who used lipid-lowering drugs at the time of screening were required to discontinue them for 4 weeks prior to study participation. Participants fasted for at least 8 h and abstained from alcohol intake for ≥ 24 h. The postprandial lipemia intervention followed the protocol of Patsch et al. [16]. The whipping cream (83% fat) meal had 700 calories/m2 body surface area; 3% of calories were derived from protein, and 14% from carbohydrate. Blood samples were drawn immediately before the meal with the last draw being 6 h later. During the 6-h study period, participants consumed only water and abstained from physical activity. Inflammatory markers and adiponectin were assayed only in the fasting state. The data used in this analysis were collected at the time of the postprandial lipemia challenge and, thus, except for the 6-h lipid measures, was cross-sectional.
Lipids, inflammatory markers, and adiponectin
Triglycerides (TG), high-density lipoprotein (HDL), low-density lipoprotein (LDL), and total cholesterol were measured by nuclear magnetic resonance (NMR) spectroscopy at fasting and during the postprandial intervention (Liposcience, Raleigh, NC) [17]. High-sensitivity C-reactive protein (hsCRP) was measured on the Hitachi 911 using a latex particle-enhanced immunoturbidimetric assay (Kamiya Biomedical Company, Seattle, WA). Interleukin-6 (IL6), interleukin-2 soluble receptor alpha (IL2sRα), tumor necrosis factor (TNF)-α and monocyte chemoattractant protein-1 (MCP1) were measured using quantitative sandwich enzyme immunoassays (ELISA kit assays, R&D Systems Inc., Minneapolis, MN) as described in previous publications [18, 19]. Adiponectin was measured using an ELISA kit from R&D Systems (Minneapolis, MN) [20].
DNA methylation measurements and epigenetic age calculation
DNA from a total 993 GOLDN participants was assayed using Illumina’s Infinium HumanMethylation450 Beadchip as previously described by Irvin et al. [21]. Briefly, CD4+ T cells were harvested from stored buffy coats (from the fasting sample) using antibody-linked Invitrogen Dynabeads [22]. Cells were lysed and DNA was extracted using DNeasy kits (Qiagen, Venlo, Netherlands). After following standard Illumina protocols, the resulting intensity files were analyzed with Illumina’s GenomeStudio to generate beta scores representing the percent of methylated sites at that location across all the cells in a sample. After previously reported QC, methylation data were available for 461,281 5′-cytosine-phosphate-guanine-3′ genomic site (CpGs) [21]. The non-normalized data were uploaded into the DNA methylation age calculator [4] using the advanced analysis in blood option. We removed 140 samples for which the predicted tissue was not blood. To err on the side of caution, we removed 23 individuals from the analysis because they were severe outliers according to the Horvath measure of epigenetic age acceleration (> 4 standard deviations away from the mean), leaving a total sample size of n = 830 individuals for the analysis.
In addition to the Horvath measure of epigenetic age acceleration, we considered another independent measure of age acceleration derived from the Hannum methylation age estimate [12]. Both methylation age acceleration estimates have been associated with outcomes in a side-by-side manner in previous publications [5, 23, 24]. The Horvath methylation age estimate (DNAm age) is based on 353 CpGs and has been robustly correlated with chronological age in multiple studies. DNAm age sometimes systematically underestimates chronological age [25]. To remove this systematic offset from the analysis, this study leverages epigenetic age acceleration defined by regressing DNAm age on chronological age and forming residuals. The second estimate is based on Hannum’s methylation age estimate calculated from 71 CpGs. Figures 1 and 2 are the scatter plots of calculated epigenetic age versus chronological age for the Horvath and Hannum calculators, respectively, in GOLDN. The Horvath and the Hannum measures of DNAm age are related, but they capture slightly different aspects of biology (only 6 CpGs overlap) [5]. The Hannum age estimator is correlated with estimated blood cell counts, which reflects that it was constructed on the basis of whole blood DNA methylation data [5, 26]. In the current analysis, we focused on a modification of the Hannum estimate described (in previous publications) as extrinsic epigenetic age acceleration (EEAA) since it has been found to outperform the original Hannum measure when it comes to predicting lifespan [5]. EEAA is a residual formed from regressing the Hannum estimate weighted by the proportion of naive CD8+ T cells, exhausted CD8+ T cells, and plasmablasts on chronological age which makes the estimate more strongly reflective of immune system aging [5, 23, 27]. Chen et al. reported the Hannum estimator tracks aspects of immunosenescence exhibiting significant (albeit weak) correlations with several markers of immunosenescence, e.g., the abundance of senescent (or conversely naïve) T cells and telomere length [26, 27]. The Horvath DNAm age estimator exhibits substantially weaker correlations with estimated blood cell counts, which reflects the fact that it was constructed across a broad spectrum of tissues and cell types. Overall, the Horvath DNAm age estimator does not relate to measures of immunosenescence (including telomere length) [27]. Rather, it mostly captures cell intrinsic DNAm age changes which might reflect the actions of an innate aging process.
For both methylation age acceleration estimates, a positive value indicates that the epigenetic age is higher than expected based on chronological age (i.e., accelerated epigenetic aging). We note EEAA in GOLDN is limited by the use of DNA methylation data derived from CD4+ T cells as opposed to peripheral blood mononuclear cells (PBMCs) like in previous studies. For a more direct comparison to the Horvath estimate we also report the results for the Hannum residual defined by regressing the (unweighted) Hannum methylation age estimate on chronological age in a secondary analysis.
Statistical analysis
Participant characteristics were compared by the median of each age acceleration measure using chi square tests and t-tests, as appropriate. TG, hsCRP, IL6, IL2sRα, TNFα, MCP1, and adiponectin were ln-transformed to achieve normality. Pearson correlation coefficients between lipids, inflammatory markers, adiponectin, and each measure of age acceleration were also estimated. We used linear mixed models to examine the association of the Horvath epigenetic age acceleration residual with fasting and postprandial total cholesterol, LDL, HDL, and TG as well as inflammatory markers and adiponectin (as outcomes). The model was adjusted for age, study site, sex, corresponding fasting lipid level (if applicable), deconvolution estimated T cell type (naïve, regulatory, memory) percentages (estimated using a method developed by the GOLDN group [28]), age acceleration estimate squared (to account for a potential non-linear effect of this variable), current smoking status, current drinking status and a random effect of family relationship. Linear mixed models were implemented in the R kinship package (lmekin function) [29]. For our analysis of the EEAA estimate we used a similar model, except the model was not adjusted for estimated T cell type percentages since blood cell composition is a component of the age acceleration estimate. For both the Horvath and Hannum estimates we considered a sensitivity model adjusted for history of coronary heart disease (CHD) and diabetes status. Further, we explored a gender by age acceleration interaction term. In total we tested the association of 14 outcomes with each age acceleration estimate making the Bonferroni correction for multiple testing p = 0.003 (0.05/14). In secondary analysis we examined the association of each outcome with the Hannum residual, considering the model described above for EEAA both with and without adjustment for T cell percentages.