Epigenome-Wide Association Study of Fasting Blood Lipids in the Genetics of Lipid Lowering Drugs and Diet Network Study

Abstract

Background

Genetic research regarding blood lipids has largely focused on DNA sequence variation; few studies have explored epigenetic effects. Genome-wide surveys of DNA methylation may uncover epigenetic factors influencing lipid metabolism.

Methods and Results

To identify whether differential methylation of cytosine-guanine dinucleotides (CpGs) correlated with lipid phenotypes, we isolated DNA from CD4+ T-cells and quantified proportion of sample methylation at over 450,000 CpGs using the Illumina Infinium HumanMethylation450 Beadchip in 991 participants of the Genetics of Lipid Lowering Drugs and Diet Network. We modeled percent methylation at individual CpGs as a function of fasting very low density lipoprotein cholesterol (VLDL-C) and triglycerides (TG) using mixed linear regression adjusted for age, gender, study site, cell purity, and family structure. Four CpGs (cg00574958, cg17058475, cg01082498, cg09737197) in intron 1 of carnitine palmitoyltransferase 1A (CPT1A) were strongly associated with VLDL-C (P=1.8*10^-21 to 1.6*10^-8) and TG (P=1.6*10^-26 to 1.5*10^-9). Array findings were validated by bisulfite sequencing. We performed qPCR experiments demonstrating that methylation of the top CpG (cg00574958) was correlated with CPT1A expression. The association of cg00574958 with TG and CPT1A expression were replicated in the Framingham Heart Study (P=4.1*10^-14 and 3.1*10^-13, respectively). DNA methylation at CPT1A cg00574958 explained 11.6% and 5.5% of the variation in TG in the discovery and replication cohorts, respectively.

Conclusions

This genome-wide epigenomic study identified CPT1A methylation as strongly and robustly associated with fasting VLDL-C and TG. Identifying novel epigenetic contributions to lipid traits may inform future efforts to identify new treatment targets and/or biomarkers of disease risk.

Introduction

Genomic studies of lipids and other cardiovascular disease (CVD)- related traits have traditionally focused on heritable allelic variation, namely, genetic polymorphisms in the nucleotide sequence of DNA in populations. Much progress has been made through research focused on this classical genetic paradigm. For example, multiple large-scale meta-analyses of genome-wide association studies (GWAS) have identified numerous loci associated with fasting blood lipids, many of them constituting novel findings.¹ Despite promising discoveries, these loci explain a modest fraction of the observed variance, approximately 2–12%.¹

Epigenetic changes are functional biochemical alterations in DNA that do not alter the underlying DNA sequence.² DNA methylation is an epigenetic process implicated in human disease that involves methylation of cytosine, usually at cytosine-(phosphate)-guanine (CpG) dinucleotides in the promoter region or within genes.² This molecular phenotype plays a pivotal role in gene expression by affecting chromatin structure and altering the availability of coding regions to transcription mechanisms.³ In contrast to DNA sequence variation, epigenetic variation is sensitive to both inherited and environmental inputs.

Genome-wide epigenetic investigation of blood lipids has been largely unexplored. Therefore, we conducted an epigenome-wide association study (EWAS) for fasting very low density lipoprotein cholesterol (VLDL-C) and triglycerides (TG) in 991 participants belonging to the Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) Study.⁴ DNA was isolated from CD4+ T-cells harvested from stored buffy coats, and methylation was quantified using the Illumina (San Diego, CA) Infinium Human Methylation450 Beadchip.⁵ Our top finding was validated through bisulfite resequencing and gene expression assays. Subsequently we sought replication of our top finding for TG in an external study population (the Framingham Heart Study, FHS) with comparable epigenetic, gene expression, and lipid data.

Methods

Discovery Study Population

The GOLDN study was designed to identify genetic determinants of lipid response to two interventions (a high-fat meal challenge and fenofibrate treatment for 3 weeks).⁴ Briefly, the study ascertained and recruited families from the Family Heart Study at two centers, Minneapolis, MN and Salt Lake City, UT, who self-reported to be white. Only families with at least two siblings were recruited for a total of 1327 individuals. Volunteers were required to withhold lipid-lowering agents (pharmaceuticals or nutraceuticals) for at least 4 weeks prior to the initial visit to be eligible. A total of 1053 met all eligibility requirements. The study protocol was approved by Institutional Review Boards at the University of Minnesota, University of Utah, and Tufts University/New England Medical Center and the subjects gave informed consent. For the current study we evaluated fasting TG and VLDL-C among 991 participants for whom epigenetic data were available.

Data Collection

TG and VLDL-C were measured before the diet and drug intervention. Participants were asked to fast for ≥12 hours and abstain from alcohol intake for ≥24 hours. TGs were measured by a glycerol-blanked enzymatic method (Trig/GB, Roche Diagnostics Corporation, Indianapolis, IN). Nuclear Magnetic Resonance (NMR) spectroscopy measured VLDL-C (Liposcience, Raleigh, NC).⁶ Data on medical history, physical activity and other lifestyle factors such as alcohol intake, smoking status, and diet were collected using an interviewer-administered questionnaire.

Methylation Assays

DNA was extracted from CD4+ T-cells harvested from stored buffy coats using antibody-linked Invitrogen Dynabeads.⁷ CD4+ T-cells were selected for three reasons: First, DNA methylation patterns are often tissue specific. For instance, studies of whole blood samples reflect methylation variations within each blood cell type that may act to confound epigenomic association results.^{8, 9} Second, many key genes involved in lipid metabolism are expressed in lymphocytes and other immune cells (eg, PPARs).^10-12 In one study, peripheral blood mononuclear cell gene expression profiles were demonstrated to reflect nutrition-related metabolic changes. Responsive genes were enriched for fatty acid metabolizing enzymes including CPT1, ACAA2, and SCL25A20.¹³ Therefore, we hypothesized this cell type should reflect underlying epigenetic variation influencing blood lipids while minimizing potential confounding. Third, blood collection is the most viable tissue collection method among healthy individuals.

Stored buffy coats were collected at the same time lipid concentrations were measured. We lysed cells captured on the beads and extracted DNA using DNeasy kits (Qiagen, Venlo, Netherlands). We used the Illumina Infinium Human Methylation450 Beadchip (Illumina Inc, San Diego, CA) to interrogate ∼470,000 autosomal CpG sites across the genome.⁵ A description of the array as well as CpG site nomenclature conventions can be found at http://www.illumina.com/products/methylation_450_beadchip_kits.ilmn. For each assay, 500ng of DNA was treated with sodium bisulfite (EZ DNA, Zymo Research, Irvine, CA) prior to standard Illumina amplification, hybridization, and imaging steps. The resulting intensity files were analyzed with Illumina's GenomeStudio which generated beta scores (ie, the proportion of total signal from the methylation specific probe or color channel) and “detection P-values” (defined as the 1-p-value computed from the background model characterizing the chance that the target sequence signal was distinguishable from the negative controls).¹⁴ Beta scores with an associated detection P-value greater than 0.01 were removed and samples with more than 1.5% missing data points were eliminated from further analysis. Furthermore, any CpG probes where more than 10% of samples failed to yield adequate intensity were removed. A total of 58 samples were removed. The quality control filters applied are comparable to other published reports.^15-19 The filtered beta scores were then subjected to batch normalization with the ComBat package for R software in non-parametric mode (http://www.bu.edu/jlab/wp-assets/ComBat/Abstract.html). We performed the normalization in parallel on random subsets of 20,000 CpGs per run where each array of 12 samples was used as a “batch.” Our methods have been extensively described in Absher et al. and the utility of ComBat to correct for batch effects in comparison to other programs is reported.^{19, 20} To correct for differing probe chemistry on the Illumina Infinium Human Methylation450 Beadchip we separately normalized probes from the Infinium I and II chemistries and subsequently adjusted the β scores for Infinium II probes using the equation derived from fitting a second order polynomial to the observed methylation values across all pairs of probes located <50bp apart (within-chemistry correlations >0.99), where one probe was Infinium I and one was Infinium II.¹⁹ Finally, we eliminated any CpGs where the probe sequence mapped either to a location that did not match the annotation file or to more than one locus. We identified such markers by re-aligning all probes (with unconverted Cs) to the human reference genome. After these quality control procedures, there were methylation data from 461,281 CpGs. Principal components (PCs) based on the beta scores of all autosomal CpGs passing QC were generated using the prcomp function in R (V 2.12.1) and used to adjust for cell purity in association analysis similarly to Hidalgo et al.¹⁸ Deconvolution estimated CD4+ T-cell percentages were generated using cell-type specific methylation data from external reference samples by adapting the method of Abbas et al.²¹ Predicted CD4+ T-cell percent purity was highly correlated with PC1 (r²=0.85, P=4*10^-293) but not other PCs, thus supporting the usefulness of methylation PCs in adjusting for cell purity in our analysis. PCs have the added benefit that they can control for unknown confounding and we chose 4 PCs based on the scree plot of the eigenvalues (see Supplemental Figure 1).

Expression Data

qPCR was conducted with Life Technologies (Grand Island, NY) TaqMan probes for CPT1A (at two different sites (exon 11-12 and exon 2-3 boundaries)) and for 5 control genes in buffy coats from 87 GOLDN individuals. These assays yielded nearly identical estimates of CPT1A expression and did not distinguish between the two known splice variants of CPT1A. RNA was extracted with Trizol, followed by QIAGEN (Venlo, Netherlands) RNAeasy purification. 2 μg of input RNA was used in a 40μl reaction using an iScript cDNA synthesis kit (Bio-Rad, Hercules, CA). Template cDNAs were diluted 1:4 and relative mRNA expression levels were quantified in triplicate 10μl reactions by TaqMan gene expression assays with a 7900HT Real Time PCR system according to standard PCR cycling conditions (Life Technologies). All gene expression assays were purchased from Life Technologies, including two assays for CPT1A (Hs00912671_m1 and Hs00912681_m1), plus 5 internal control genes for baseline normalization (Hs00168719_m1 (PPIB), Hs00154728_m1 (DECR1), Hs02786624_g1 (GAPDH), Hs00984230_m1 (B2M), Hs00951083_m1 (TFRC)). Since methylation was assayed in CD4+ T-cells, but CPT1A expression was measured in buffy coats, the control genes were selected for 1) T-cell expression, 2) neutrophil expression, or 3) red blood cell expression (HBZ), the later two of which are likely to be our greatest source of confounding, so these controls served as both loading controls and as estimators of buffy coat composition. Reaction cycle threshold (Ct) values for the two CPT1A probes yielded nearly identical results and were averaged. Individual Ct values for each endogenous control were median centered. Endogenous control Ct values were averaged per sample and subtracted from median centered CPT1A target Ct values to generate ddCt input for relative quantitative method (2^-ddCt). These normalized CPT1A expression values were compared to the methylation level at cg00574958 corrected for methylation PCs but not other covariates (since the samples were from the same individuals) using linear regression.

Bisulfite Resequencing

To validate the array results we used CATCH-Seq (Ubiquity Genomics, Huntsville, AL) target enrichment to perform bisulfite sequencing of ∼200kb around CPT1A in 154 participants chosen at the extremes of the observed cg00574958 beta value. We used the same T-cell DNA samples assayed on the Methylation450 array. The 154 DNA samples were prepared for Illumina sequencing using custom methylated adapters to uniquely barcode each sample during library creation. The library creation followed standard Illumina protocols of shearing, end-repair, and adapter ligation. Prior to bisulfite treatment (QIAGEN Epitect) and PCR amplification, pools of 12 samples were captured with biotinylated probes that were generated from human BAC clones mapping to ∼200kb of genomic sequence containing the CPT1A locus (Life Technologies; RP11 1109D20, RP11 800E23, RP11 154D10, and CTC 337L). The captured libraries were sequenced on an Illumina HiSeq2500 with 2×100bp reads. One-hundred and thirty-two samples achieved a mean coverage of >340X, among which 121 covered cg00574958 at greater than 100X. These 121 individuals were included in the comparison with the array data. Methylation levels at each CpG across the target region were estimated using Bowtie2 for alignment and Bismark for the calculation of the methylation proportions.²²

Statistical Analysis

In the discovery stage, we modeled beta score at each CpG site as a function of the log of fasting TG or VLDL-C using mixed linear regression models adjusted for age, gender, study site, and 4 methylation PCs as fixed effects and family structure as a random effect. After removing an additional 4 observations due to missing phenotype or covariate data (ie, 1053-58-4), a total of 991 participants were considered in the association analysis. A Bonferroni correction for multiple testing was applied to the discovery models adjusted for 4 lipid traits where α=0.05/(461, 281*4)=4.5*10^-8 (note results for HDL-C and LDL-C are not presented as no CpG was significant after correction for multiple testing). Because environmental factors can directly confound EWAS by affecting both the epigenotype and phenotype, sensitivity analysis adjusting for potential major sources of confounding including current alcohol use and current smoking use was conducted. A second sensitivity analysis considered additional sources of confounding including physical activity (hours of moderate to heavy physical activity a week), diet (% energy from carbohydrate and % energy from total fat), and education (as a proxy for socioeconomic status). We have previously described this population to be very genetically homogeneous and have not adjusted for substructure.²³ However, principle components based on SNP data from GWAS (SNP PCs) calculated in Eigenstrat (as described by Aslibekyan et al)²³ were available for 72% of our study population and the first 10 SNP PCs were included in the second sensitivity analysis described. All models were implemented in the R kinship package (lmekin function).²⁴ The lmekin function fits a linear mixed effects model which uses the kinship coefficient to define the correlation of random effects while the fixed effects are used to test for associations and adjust for potential confounders. We selected the maximum number of unrelated individuals from our data (N=278) and report the difference in the R² estimate from a linear model with and without the top CpG term (adjusted for the same covariates described for the primary model) as the variance explained by that term. A similar nested models approach was used to determine the variance explained by the combination of the top 4 CpGs reported.

GOLDN GWAS, MethylQTLs, and Other Annotations

As described above, the majority of the 991 participants in this study had been previously genotyped at 906,600 loci using the Affymetrix Genome-Wide Human SNP Array 6.0 (Santa Clara, CA). ²³ MACH software (Version 1.0.16) was used to impute nongenotyped SNPs using HapMap Phase II (release 22, Human Genome build 36, hg18) as a reference, resulting in genotypes for 2,529,001 SNPs. To investigate an underlying genetic component to observed epigenetic associations we performed cis-methylQTL analyses of SNPs within 1 Mb of top CpG findings similarly to that described by Zhi et al.²⁵ In brief, we used linear mixed models, fit using the lmekin function, to regress the methylation level (beta score) of a CpG site on the genetic variant, adjusting for covariates (age, sex, and study site) as fixed effects and family structure as a random effect. Additionally, we used publically available regional ENCODE annotations accessed Dec 13, 2013 to evaluate transcription factor binding sites, chromatin modifications, histone acetylation, and consistency of tissue specific methylation patterns in the region of our association results. We downloaded the most recent Global Lipids Consortium GWAS data (N=188,577) and used LocusZoom (http://csg.sph.umich.edu/locuszoom) to plot regional SNP results from meta-analysis on Chromosome 11 in the CPT1A region.¹ Finally, we examined expression QTL (eQTL) SNPs in the region of CPT1A available at the blood eQTL browser (http://genenetwork.nl/bloodeqtlbrowser/).

Replication Population

The Framingham Heart Study (FHS) Offspring cohort, initiated in 1971, includes 5124 children and spouses of children of the Original cohort as previously described.²⁶ Genome-wide DNA methylation and gene expression were assessed from peripheral blood samples (N=2846) collected during examination cycle 8 (2005-2008). Genomic DNA was extracted from buffy coat preparations using the Gentra Puregene DNA extraction kit (Qiagen). Bisulfite conversion was conducted on genomic DNA using the EZ DNA Methylation kit and bisulfite-converted DNA was then hybridized to the Illumina Infinium Human Methylation450 Beadchip in two laboratory batches (N=576 and 2270). RNA was extracted from whole blood using the PAXgene Blood RNA System Kit (Qiagen) with mRNA expression profiling from the Affymetrix Human Exon 1.0 ST GeneChip platform. Rigorous quality control (QC) measures were conducted. For the DNA methylation arrays, 71 samples were excluded due to poor SNP matching of control positions; 45 were removed due to missing rate >1%, and 73 were removed as outliers using multidimensional scaling (MDS), for a total of 135 samples excluded (ie, considering overlap among QC measures), leaving 2711 samples. We excluded 833 CpG probes due to missing rate >20%. Methylation data were normalized within laboratory batches using a DASEN method from the wateRmelon package.²⁷ Robust multichip average (RMA) method was used to normalize the gene expression values with QC measures as previously reported.²⁸ Data from 2280 Third Generation FHS participants with both measured cell counts and gene expression obtained during the second examination cycle (2008-2011) was used to predict cell count proportions in the current sample using partial least squares regression.²⁹ Internal validation using training and testing datasets achieved an r²>0.8 in the majority of cell lines (except basophils). Then, the prediction was extended to the Offspring cohort using the resulting coefficients and Offspring cohort gene expression data. We excluded Offspring cohort participants on lipid-lowering therapy and restricted our analyses to participants with both DNA methylation and gene expression, leaving 1261 participants for analyses. We tested associations between levels of log-transformed triglycerides and a) CpG-specific DNA methylation at cg00574958 and b) gene-level CPT1A mRNA expression (transcript cluster ID#3379644), using linear mixed regression models, adjusted for age, sex, and estimated cell count as fixed effects, technical covariates (ie. chip) as random effects, and kinship among family members as a random correlation structure, using the pedigreemm package of R. DNA methylation and gene expression were specified as dependent variables in the regression models. Pooled DNA methylation data from the two laboratory batches, as opposed to meta-analyzed data, was used to avoid exaggeration of the standard error in the meta-analyzed data due to a large population imbalance (10:1) between the two groups. Effect sizes of the pooled results match those of meta-analyzed results. The association between the DNA methylation and gene expression was performed on the residuals after the removal of the fixed and random covariates, along with the kinship correlation structure using simple linear models, primarily to avoid potential confounding by blood count. Mediation analyses were conducted using nonparametric estimation methods³⁰ to model the indirect association between CPT1A DNA methylation at cg00574958 and log-transformed TG levels through changes in CPT1A gene expression. In the mediation analyses, a pathway is specified a priori in which a hypothesized causal factor (cg0574938 DNA methylation) influences a mediator (CPT1A gene expression), which in turn affects the outcome of interest (plasma triglyceride levels). The model assumes no unmeasured confounding or effect modification between the included elements. The proportion mediated describes the average magnitude of indirect association between CPT1A DNA methylation at cg00574958 and log-transformed triglycerides attributed through changes in CPT1A gene expression relative to the average total association. Asymptotic 95% confidence intervals were obtained from nonparametric bootstrapping with 20,000 iterations implemented in the mediation package in R.³¹ Finally, the results for GOLDN and FHS were subsequently meta-analyzed using Fisher's method, obtaining the Chernoff bound of the chi-squared cumulative distribution function for the P-value.

Results

The GOLDN study population (N=991) was on average 48.8±16 years of age. The sample was 52% female and 49% were recruited from the MN field center as opposed to the UT site. Mean fasting VLDL-C (±SD) was 105.3±93 mg/dL and mean fasting TG (±SD) was 137.0±95 mg/dL. Mean TG levels were within the normal range according to ATP III Guidelines.³² The FHS population (N=1261) was on average older (mean age was 64.8±9 years), more likely to be female (60%), and had an average fasting TG level of 112.3±66 mg/dL.

Differential methylation at four CpGs in intron 1 of carnitine palmitoyltransferase 1A (CPT1A) (cg00574958, cg17058475, cg01082498, and cg09737197) were very strongly associated with both TG and VLDL-C (Figures 1 and 2). Increased methylation at each of the four CpGs was inversely associated with each lipid trait, with P-values ranging 1.6*10^-8 to 1.6*10^-26 (Table 1). The association result for the top CPT1A DNA methylation site (cg00574958) with TG was replicated in FHS with comparable effect size and direction ((se)=-0.007 (9.4*10^-4), P=4.1*10^-14) and the meta-analysis P-value was 9.6*10^-36. Sensitivity analysis or further statistical adjustment for potential confounders in GOLDN did not appreciably alter the results (see Supplemental Table 1). Also, we used a subset of unrelated GOLDN individuals to estimate the variance in fasting TG explained by cg00574958, and found that ∼11.6% of TG variance could be explained by methylation at that site. The combination of all four CpGs identified in GOLDN explained 14.7% of TG variance. In FHS 5.5% of TG variance could be explained by methylation at cg00574958. These four CpG sites and their position along with ENCODE annotations are displayed in Figure 3.

Figure 1.

Epigenome-wide association Manhattan plot for TG in the discovery dataset (N=995).

Figure 2.

Epigenome-wide association Manhattan plot for VLDL-C in the discovery dataset (N=991).

Table 1.

Significant EWAS signals observed for very low density lipoprotein cholesterol (VLDL-C) and triglycerides (TG) in the GOLDN discovery dataset (N=991).

Marker*	Chr	Location	Beta (SE)§	P
VLDL-C
cg00574958	11	68607622	-0.022 (2.28*10-3)	1.79*10-21
cg17058475	11	68607737	-0.025 (3.15*10-3)	3.32*10-15
cg01082498	11	68607675	-0.008 (1.24*10-3)	1.06*10-10
cg09737197	11	68608225	-0.019 (3.31*10-3)	1.57*10-8
TG
cg00574958	11	68607622	-0.032 (2.99*10-3)	1.56*10-26
cg17058475	11	68607737	-0.035 (4.14*10-3)	8.90*10-17
cg01082498	11	68608225	-0.011 (1.63*10-3)	8.79*10-11
cg09737197	11	68607675	-0.027 (4.36*10-3)	1.53*10-9

^*results were adjusted for age, sex, center, pedigree, and 4 methylation principal components

^§The regression parameter estimate (beta) represents the change in methylation beta score for each unit change in log (TG) or log (VLDL-C).

Figure 3.

ENCODE annotation of the promoter region and intron 1 of CPT1A. Top CpGs for TG are positioned within the gene along with CpG islands, cell line chromatin state (ChromHMM), cell line methylation at CpG sites on the Methyl450 Beadchip according to Hudson Alpha Institute for Biotechnology (HAIB; note blue, purple and orange highlights correspond to low, medium and high methylation state, respectively) and HMR conserved transcription factor binding sites.

After qPCR in 87 GOLDN participants, % methylation (corrected for methylation PCs) at cg00574958 was significantly negatively correlated (r=-0.378, P=0.00031) with relative CPT1A expression (Supplemental Figure 2). CPT1A expression in GOLDN was positively correlated with TG (r=0.19, P=0.04, N=87). Bisulfite resequencing among 121 participants of the region surrounding CPT1A (coverage ≥100X) demonstrated strong correlation with % methylation at cg00574958 observed by the methyl450 array with r=0.83 and P=5.4*10^-31 (see Supplemental Figure 3), validating variable methylation at that site. In FHS, increased methylation at cg00574958 was also negatively associated with relative CPT1A expression ((se)=-3.812 (0.52), P=3.1*10^-13), while CPT1A expression was positively associated with TG after adjustment for age, gender and family structure ((se)=0.113 (0.02), P=5.6*10^-12). In FHS, the mediation analysis found that 13.5% (95% CI 7.4 – 22.2%) of the association of CPT1A DNA methylation at cg00574958 with log TG can be attributed to changes in gene expression of CPT1A.

Using GOLDN data, we also evaluated cis-meQTLs in the region of the four top CpGs in CPT1A. Results are displayed in Supplemental Figure 4. We show there is little genetic association between SNPs and the CpGs of interest within a 1 Mb window surrounding the CpGs. Additionally, Supplemental Figure 5 shows a regional plot of GWAS results from the Global Lipids Consortium in the region of the CPT1A locus. Data show the most significant SNP P-value for the TG outcome was nearer to a neighboring gene and not statistically significant (MRPL21, P=4.0*10^-4). Finally, there are several strong eQTLs from peripheral blood samples near the 3′ region of CPT1A (Supplemental Figure 6).

Discussion

The current study reports the top results from an epigenome-wide analysis of fasting triglycerides and very low density lipoprotein cholesterol using the Illumina Infinium Human Methylation450 Beadchip. DNA was isolated from CD4+ T-cells harvested from stored lymphocytes. Over 450,000 CpGs were evaluated in 991 participants from the Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) Study. Four CpGs in CPT1A were significantly associated with both lipid traits. Adjustment for potential confounders did not materially change the findings. The top CpG marker in CPT1A was validated by bisulfite resequencing and was associated with CPT1A gene expression in a subset of GOLDN participants. The top CpG association and expression results were robustly replicated in the well characterized Framingham Heart Study population. This study demonstrates the potential importance of epigenetic variation in CPT1A in inter-individual differences in fasting TG and VLDL-C levels.

Carnitine plays an essential role in the transfer of long-chain fatty acids across the mitochondrial membrane. CPT1 is a key enzyme in the carnitine-dependent transport of long-chain fatty acids into the mitochondria, and its deficiency results in a decreased rate of fatty acid beta-oxidation.³³ Three tissue-specific CPT1 isoforms exist, including the liver (CPT1A), muscle (CPT1B), and brain (CPT1C) forms.³⁴ Publically available gene expression data show CPT1A is also expressed in CD4+ T-cells.³⁵ CPT1A deficiency is a very rare autosomal recessive disorder of fatty acid beta-oxidation caused by functional mutations in the gene that have been both directly and indirectly linked to alterations in the active enzyme site.³⁶ Carriers of functional mutations may be at risk for lipid and other metabolic disorders.^37-41 However, to the best of our knowledge, no large SNP GWAS or epigenetic studies have highlighted the gene.

Our study reports an inverse relationship between methylation at multiple loci in intron 1 and TG and VLDL-C. We replicated the top marker's association with TG, including the direction, in 1261 participants from FHS. Gene expression studies in a subset of our GOLDN data show increased methylation at our top CpG site (cg00574958) is associated with decreased relative gene expression (Supplemental Figure 2). These gene expression findings in relation to methylation at cg00574958 were validated in FHS. In both GOLDN and FHS, increased CPT1A expression was associated with increased TG. A previous study in an animal model suggests decreased CPT1A activity correlates with increased lipid levels in myocytes.⁴² Differences in the animal model and our results could be due to our cross-sectional design or the cell types examined. Importantly, all associations including the direction of effect replicated across our studies. Mediation analysis in FHS suggests that ∼14% of the association between the top CpG and TG can be attributed to changes in gene expression. Further evaluation of these trends in more than one tissue type in humans and animals are warranted to help make causal inferences about the relationships between methylation, gene expression and lipid levels.

Following replication of our results we evaluated potential functional genomic mechanisms underlying the epigenetic associations. GOLDN meQTL data demonstrate that nearby SNPs are not strongly associated with the CpGs of interest. Additionally, a regional examination of CPT1A and nearby loci from the Global Lipids Consortium GWAS data found no SNPs (or lipid QTLs) that reach genome-wide significance for the TG trait. Taken together these results suggest that common SNP variation represented by GWAS does not underlie the observed associations. We also investigated peripheral blood eQTLs in the CPT1A region using published data. Several strong eQTLs for CPT1A were identified near the 3′ region of the gene. However, the eQTL SNPs are distant (∼100kb), and most likely exert an effect on gene expression independent from the CpGs highlighted by our study. Results from ENCODE suggest the region of interest is an active regulatory site. There are several transcription factor binding sites near our highlighted CpGs (red dotted rectangle in Figure 3). More than one of the annotated transcription factors is involved in lipid metabolism including sterol regulatory element-binding proteins (SREBPs), peroxisome proliferator-activated receptor gamma (PPARγ), and upstream transcription factor 1 (USF1).^43-45 The 4 CpGs are just upstream of an active promoter region according to Chromatin State Segmentation by HMM in the HepG2 cell line (bright red region on Figure 3) and in between two CpG islands. Additionally, evidence of increased acetylation of lysine 27 of the H3 histone protein (H3K27Ac) also indicates enhanced transcription in the region. Given this functional evidence and observed associations between methylation and gene expression in GOLDN and FHS, we speculate CpG methylation state in intron 1 may facilitate an epigenetic program including open chromatin and histone related enhanced binding of transcription factors, although future studies are needed to evaluate this function.

Few EWAS of lipid traits have been reported and, to the best of our knowledge, have not highlighted CPT1A.^{46, 47} However, a recent metabolomic EWAS provides further in silico validation of our results. Specifically 649 blood metabolomic traits from 1814 participants of the Kooperative Gesundheitsforschung in der Region Augsburg (KORA) study were assessed for association with methylation at 457,004 CpG sites determined on the Infinium HumanMethylation450 BeadChip platform.⁴⁸ Our top finding in CPT1A (cg00574958) was among the top findings for VLDL-A (P=9.23*10^-14) and adjustment of the epigenetic association result for nearby SNP variation measured in GWAS did not alter the result. The authors concluded the association did not exhibit an underlying genetic signal. Validation studies, external replication and a discussion of the biological relevance of the finding were not provided by that report. Also relevant to our findings is a 2006 publication by Shen et al. which showed hypermethylation of the promoter region of CPT1A affects gene expression during differentiation of human embryonic stem cells (hESCs) into neural progenitor/stem cells (NPCs) and suggested a link to lipid metabolism.⁴⁹

The study findings implicate a role for CPT1A methylation in inter-individual variation in blood lipid levels beyond DNA sequence variants. As methylation alters gene expression, findings suggest future drug development for lipid lowering could be centered around therapeutically altering expression or action of CPT1A. However, our results are several steps upstream of therapeutic implications. First, both GOLDN and FHS have a cross-sectional design limiting causal inference. Specifically, we cannot determine if the observed associations are due to methylation effects on lipids or vice-versa. The functional mechanism of methylation variation on gene expression also needs further evaluation. Finally, since both GOLDN and FHS represent healthy Caucasians further replication can expand generalizability of our results to other ethnic and clinical populations. Despite these limitations our study is supported by several strengths including genome-wide epigenetic testing with a dense panel of CpG markers on a large sample of healthy adults with replication complimented by expression and validation studies.

Multiple large studies have evaluated the effect of common DNA sequence variants on fasting lipid traits with many impactful findings, yet known predictors explain only a limited portion of the observed variability. Methylation at CpG sites is an important genomic regulatory mechanism that few studies have evaluated on a genome-wide level in the context of CVD related traits. The current study quantified methylation at over 450,000 CpG sites in participants from the GOLDN study. EWAS identified four CpGs in CPT1A significantly associated with TG and VLDL-C. Results were validated by bisulfite resequencing and the top result was replicated in an independent study population. The association does not appear to be influenced by surrounding genetic variants and ENCODE suggests the region is an active epigenetic regulatory site for the gene. In conclusion, this study emphasizes the importance of expanding genomic studies of lipid related traits beyond sequence variants and to identify additional loci that could become useful in the prevention of CVD.

Clinical implications for Irvin et al's “Epigenome-wide association study of fasting blood lipids in the Genetics of Lipid Lowering Drugs and Diet Network Study” (2014/009158 vers4).

The findings implicate a role for CPT1A methylation in inter-individual variation in triglycerides and VLDL-C beyond DNA sequence variants. Specifically, we report strong association between methylation in the promoter region of CPT1A in CD4+ T-Cells and TG and VLDL-C in over 900 healthy Caucasian adults from the Genetics of Lipid Lowering Drugs and Diet Network Study (GOLDN). Common genetic variation did not underlie the epigenetic signal in GODLN and resequencing studies validated the array findings for the top CpG. Results were robustly replicated in over 1200 adults from the Framingham Heart Study (FHS) using DNA from buffy coat preparations. The top CpG site was associated with gene expression in both populations. As we found that methylation at this site alters gene expression, our findings suggest future lipid lowering treatment could alter expression or action of CPT1A. However, our findings are based on cross-sectional data in healthy Caucasian populations and further research is needed before CPT1A can be targeted for drug development. Importantly, further replication can expand generalizability of our results to other ethnic and clinical populations. Longitudinal studies are needed to determine if the observed associations are due to methylation effects on lipids or vice-versa. Finally, the functional mechanism of methylation variation on gene expression also needs further evaluation. In summary, the current study represents a promising finding in genomics for the potential future treatment of hypertriglyceridemia, but much additional work is needed before drug development can ensue.