Introduction

Several recent genome-wide association studies (GWASs) in people of white race have successfully localized common DNA sequence variants that are cross-sectionally associated with blood lipoprotein levels.^1–14 Most recently, Teslovich et al¹² published a meta-analysis of 46 GWASs of >100 000 individuals of European ancestry. In total, 95 loci were significantly associated with lipid phenotypes (ie, total cholesterol, high-density lipoprotein cholesterol [HDL-C], low-density lipoprotein cholesterol [LDL-C], and triglycerides). Given that individual loci only have modest effects, in several GWASs, authors have computed summary lipid gene scores to assess whether the cumulative allelic dose of risk alleles at these loci contribute to quantitative variation in lipoprotein levels in their study populations.^1–3,12

Clinical Perspective on p 80

Virtually all prior work evaluating the relation of lipid gene scores to lipid levels has been cross-sectional. However, lipid levels often change with age or over time. Etiologically, it is of interest to know whether the genotypic effects are stable or whether they vary with time, across the life span. Although largely unexplored, some evidence suggests that the effects of some single-nucleotide polymorphisms (SNPs) may vary across the life course.^3,15–18 It has been hypothesized that the differential effects may be due to age-dependent gene expression, changes in the penetrance of underlying genes, and gene-environment interactions related to the effect of cumulative exposure to certain environmental factors.¹⁶ Further, it is unknown whether sex interactions exist in the effect of lipid genotype scores on circulating lipid levels. Lipid levels are known to vary by sex,¹⁹ and sex-differences in the effects of some alleles have been noted.^2,3,20,21

To better understand the effects of GWAS-identified SNPs on lipid levels, we created phenotype-specific lipid gene scores to test the hypothesis that over 9 years of follow-up, there would be interactions between lipid gene scores and time such that people with less-favorable lipid gene scores would experience more adverse changes in lipid levels and a higher incidence of abnormal lipid levels than people with more-favorable lipid gene scores. We also evaluated whether sex interactions were present in the cross-sectional relations of lipid gene scores to lipid levels.

Methods

Study Design, Population, and Data Collection

The Atherosclerosis Risk in Communities (ARIC) study is a multicenter, population-based prospective cohort study designed to investigate the etiology and natural history of atherosclerosis in middle-aged adults.²² Participants were recruited from 4 US communities: Forsyth County, NC; Jackson, MS; suburbs of Minneapolis, MN; and Washington County, MD. The study cohort included 15 792 white and black men and women aged 45 to 64 years at baseline in 1987 to 1989 (visit 1). Cohort reexamination took place at 3-year intervals; response rates were 93%, 86%, and 81% at visit 2 (1990–1992), 3 (1993–1995), and 4 (1996–1998), respectively.

At each visit, ARIC participants underwent interviews, venipuncture, and measurement of blood pressure and anthropometrics. Trained interviewers ascertained basic demographic data, medical history, information about personal habits, and medication use. Study participants were asked to bring all medications, vitamins, and supplements taken in the 2 weeks before the examination. All medication names were transcribed and coded.

This analysis includes only ARIC white participants who gave consent for genetic analyses and had successful GWAS genotyping. Further, for LDL-C and triglyceride analyses, data points were excluded in instances when the participant did not fast for 8 hours before venipuncture. Local institutional review boards approved the ARIC protocol, and all participants gave informed consent.

Lipid Measurement

Fasting (minimum 8 hours) blood samples were drawn from an antecubital vein into tubes containing EDTA and were sent to the ARIC Central Lipid Laboratory (Baylor College of Medicine; Houston, TX) for processing. Total plasma cholesterol²³ and triglycerides²⁴ were determined by enzymatic methods. HDL-C was measured after dextran-magnesium precipitation,²⁵ and the Friedewald equation²⁶ was used to calculate LDL-C in participants with triglyceride levels <400 mg/dL.

Lipid Medication Use

Because many participants took antihyperlipidemic medications, it was necessary to account for these medications when analyzing lipid levels. As recommended by Tobin et al,²⁷ to estimate among medication users what values might have been had the participants not been taking medication, medication use was taken into account by adding a constant. This approach has been shown to be preferable to other approaches, such as including medication use as an indicator variable in multivariate models or excluding antihyperlipidemic medication users, which may introduce bias, and similar to more complex methods, such as censored normal regression models.²⁷

The constant used depended on the specific type of medications used. For statins and fibric acid derivatives, we used the estimates recommended by Wu²⁸ (statins: LDL-C, +49.9 mg/dL; HDL-C, −2.3; triglycerides, +18.4; total cholesterol, +52.1; fibrates: LDL-C, +40.1; HDL-C, −5.9; triglycerides, +57.1; total cholesterol, +46.1). For bile acid sequestrants and niacin, we used the midpoint of the range provided in the Adult Treatment Panel III Final Report²⁹ applied to ARIC mean values (bile acid sequestrants: LDL-C, +40.5; HDL-C, −1.9; triglycerides, 0.0; niacin: LDL-C, +24.7; HDL-C, −9.9; triglycerides, +89.4). The Adult Treatment Panel III did not report estimates for the effects of bile acid sequestrants and niacin on total cholesterol, and we were unable to find stable estimates elsewhere. As an approximation based on the effects of statins and fibrates on total cholesterol, we estimated the effects of bile acid sequestrants and niacin on total cholesterol by adding the absolute values for change in LDL-C and HDL-C. Thus, for bile acid sequestrant users, we added 42.4 to total cholesterol values, and for niacin users, we added 34.6. Sensitivity analyses were explored to evaluate the robustness of our approach and assumptions.

Genotyping

In the ARIC study, genotyping was performed using the Affymetrix Genome-Wide Human SNP Array 6.0. Genotyping and quality control methods are described in detail elsewhere.³⁰ Briefly, excluded were subjects who disallowed DNA use, those with unintentional duplicates with higher missing genotype rates, suspected mixed/contaminated samples, scans from 1 problem plate, samples with a mismatch between called and phenotypic sex, samples with genotype mismatch with 39 previously genotyped SNPs, suspected first-degree relative of an included individual, and genetic outliers based on average identity by state statistics and principal components analysis using EIGENSTRAT. SNPs were excluded because of no chromosome location; being monomorphic; and having a call rate <95%, minor allele frequency of <1%, or Hardy-Weinberg equilibrium P<10⁻⁵. After the filtering, 669 450 SNPs were used in the imputation to 2 543 887 autosomal SNPs from HapMap Phase II CEU samples using MACH version 1.0.16.³⁰

Gene Score Creation

We used SNPs identified in the Teslovich GWAS¹² to create the lipid gene scores. The Teslovich GWAS is much larger than any previous lipid GWAS and made a conglomerate of the majority of populations included in prior GWASs.

Separate effect size-based gene scores were created for each lipid phenotype. Gene score creation involved several steps, as follows:

1. We first determined the number of unfavorable alleles for each individual. For genotyped SNPs, each unfavorable allele was counted as +1. For imputed SNPs, we used the predicted number of unfavorable alleles as estimated by the MACH imputation.³⁰ MACH provides an approximation of the expected number of copies of a selected allele, ranging from 0 (no chance the person had the selected allele) to 2 (both alleles were the selected allele).

2. To account for the differing effect sizes of the SNPs, the number of unfavorable alleles from each SNP was multiplied by the absolute value of the additive effect size published in the Teslovich GWAS.¹²

3. For each person, we summed the estimates (number of unfavorable alleles×effect size) across all individual SNPs included in the gene score for each phenotype. This sum was then rescaled by dividing by the phenotype-specific average effect size (the average effect of all SNPs included in the gene score).

4. Lastly, we divided each phenotype-specific gene score by 1 SD.

Note that the dual rescaling of steps 3 and 4 are mathematically unnecessary because a single rescaling would yield identical results. We chose to rescale twice to maximize the transparency and interpretability of the gene scores. The rescaling in step 3 permits reporting of the mean number of unfavorable alleles for each score (as shown later in Table 2). The rescaling by 1 SD in step 4 enhances interpretability and comparisons across phenotypes.

On the basis of Teslovich et al,¹² 47 SNPs were included in the HDL-C gene score (ARIC, 13 directly genotyped; 34 imputed), 37 in the LDL-C gene score (ARIC, 12 directly genotyped; 25 imputed), 32 in the triglyceride gene score (ARIC, 6 directly genotyped; 26 imputed), and 52 in the total cholesterol gene score (ARIC, 14 directly genotyped; 38 imputed). SNPs included in the lipid gene score, along with their imputation status, are shown in online-only Data Supplement Table I.

Notably, ARIC was a primary discovery cohort in Teslovich et al,¹² contributing 7.8% (7841/100 184) of individuals of European descent. As such, the initial SNP identification in Teslovich et al and subsequent testing of relations of the gene scores to baseline lipid levels in the present study are not entirely independent.

Statistical Analysis

Mean±SD and frequency (%) were used to describe characteristics of participants at each of the 4 ARIC visits. General linear regression (SAS PROC GLM; SAS Institute Inc, 2007) was used to evaluate the main effects of phenotype-specific lipid gene scores on baseline lipid levels, adjusted for age, sex, and center. The percent variance of lipid levels explained by the phenotype-specific gene scores, after adjustment for age, sex, and center, was determined using the “effectsize” command in the SAS PROC GLM model statement, which is the sum of squared deviations around the gene score mean (type III sum of squares for the gene score effect) divided by the sum of squared deviations around the grand mean plus the sum of squared deviations around the gene score mean (total sum of squares+gene score sum of squares). Interactions between sex and lipid gene scores on lipid levels were explored by using cross-product terms in the models. Sex-stratified results were also reported regardless of whether an interaction was present, given inherent interest.

Associations between lipid gene scores and changes in lipid levels over time were assessed using linear mixed models implemented in PROC MIXED. These models allow for estimation of both a fixed effect (population average) of the gene score over time and the individual-level variability in the intercept and slope. Participants who provided outcome data on at least 1 time point were retained in the model. Given the balanced longitudinal design, the 4 time points of the study were used as an ordinal variable, with time 1 coded as 0. Age, sex, and study center were included in all models as covariates for estimation of the intercept and slope parameters. An unstructured variance structure was specified. Product terms of time and each covariate were used to account for covariate effects on the estimated slope parameter. Before examining the effect of the gene score on each outcome, we first modeled the outcome with a random intercept and with both a random intercept and a random slope. We compared the fit of these models using standard indices (ie, Akaike information criterion, deviance). We then incorporated the gene score and its product term with time to examine the association with the baseline outcome value and the rate of change over time. Examination of the distribution of the random effects and the model residuals were consistent with the assumptions of linear mixed-effects regression. As a final consideration, we incorporated an interaction between age and gene score on the intercept and slope parameters. In the case of a significant interaction, we explored the effect of the interaction by estimating separate models within categories of age.

Note that we did not model gene score-by-age interactions; instead, as has been done in similar analyses,¹⁷ we analyzed the associations of longitudinal change with gene scores as well as gene score-by-time interactions. This approach was selected to maximize statistical power. Participants entered into the study at different ages and were not followed-up annually; therefore, each participant only contributed data to certain age groups, and the sample size for each individual age group was much more limited than that for each follow-up time point.¹⁷ From the standpoint of interpreting the results, however, the most uniform and evident change in participant characteristics during follow-up was age; therefore, the most plausible interpretation is that longitudinal effects would mostly reflect age effects.¹⁷ In secondary analyses, we conducted the linear mixed-models regression analyses stratified by 10-year age category (ie, baseline age of 45–54 or 55–64 years).

Cox proportional hazards regression was used to evaluate the relation of gene scores with incidence of abnormal lipid levels. Incidence was defined by medication use or levels beyond Adult Treatment Panel III²⁹ cut points (mg/dL) as follows: HDL-C, ≤40; LDL-C, ≥160; triglycerides, ≥200; total cholesterol, ≥240. Participants who had abnormal lipid levels or were taking lipid-lowering medications at baseline were excluded from this analysis. Person-years, which were calculated separately for each phenotype, accrued from the participant's baseline examination until incidence of abnormal lipid levels, loss to follow-up, death, disenrollment, or the participant's fourth ARIC visit. Models were adjusted for age, sex, and center.

Results

Cohort Description

The cohort of 9328 white ARIC participants had a mean age of 54 years at baseline, and 53% were women (Table 1). At baseline in 1987 to 1989, mean levels of HDL-C, LDL-C, triglycerides, and total cholesterol were 50.3±16.8, 139±39, 139±94, and 216±42 mg/dL, respectively; 3.4% were taking pharmacological agents to improve their lipid levels.

Over 9 years of follow-up, overall HDL-C levels remained fairly stable, whereas LDL-C levels decreased and triglyceride levels increased (Table 1). The prevalence of hypercholesterolemia and use of medications to lower lipid levels increased greatly over the course of follow-up.

Cross-Sectional Relation of Lipid Gene Scores to Lipid Levels

In cross-sectional analyses, all 4 lipid gene scores were significantly related to their corresponding phenotypes (P<5×10⁻³⁴ for all) (Table 2). After adjusting for age, sex, and center, the HDL-C gene score explained 1.6% of the variance in baseline HDL-C levels, and each SD increase in the HDL-C gene score was associated with a 1.89-mg/dL lower HDL-C level. Sex modified the relation of the HDL-C gene score to HDL-C levels (P=0.05). In stratified analyses, the association per 1-SD higher HDL-C gene score was stronger among women (−2.16 mg/dL) than men (−1.56 mg/dL). For LDL, the LDL-C gene score explained 6.0% of the variation in LDL-C levels, and each 1-SD higher LDL-C gene score was associated with a 9.5-mg/dL higher LDL-C level. There was no evidence of sex interactions in the relation of the LDL-C gene score to LDL-C levels. Similarly, the triglyceride gene score explained 6.0% of the variation in triglyceride levels, and levels were 22.8 mg/dL higher for each 1-SD higher triglyceride gene score. There was a sex interaction (P<0.0001), with each 1-SD higher triglyceride gene score having a greater association among men (26.5 mg/dL) than among women (19.2 mg/dL). For total cholesterol, the gene score explained 6.8% of the variation, and each 1-SD higher total cholesterol gene score was associated with a 10.7-mg/dL higher total cholesterol level. No significant sex interactions were present in the relation of the total cholesterol gene scores to levels of total cholesterol.

Longitudinal Relation of Lipid Gene Scores to Lipid Levels

Using linear mixed-models regression (Table 3 and online-only Data Supplement Tables II–V), each 1-SD higher triglyceride gene score was associated with an increase in triglyceride levels of 0.3 mg/dL (P=0.003) for each 3-year time period. This result remained significant even after adjusting for body mass index (P=0.002). There was a 3-way interaction (age×gene score×time; P=0.01) in that the greater increase in triglyceride levels at higher triglyceride gene scores was stronger among younger participants (age, 45–54 years; time×gene score interaction, β=0.52; P=0.0003) than older participants (age, 55–64 years; β=0.07; P=0.62). The HDL-C gene score, LDL-C gene score, and total cholesterol gene score were not significantly related to change in levels of their respective phenotypes.

In sensitivity analyses, we explored omitting participants based on starting use of lipid-lowering medications, and including only participants with lipid levels for all 4 visits. Results of these sensitivity analyses were similar to those of the primary analyses (data not shown).

Incidence of Abnormal Lipid Levels

Lipid gene scores were positively related to incidence of abnormal lipid levels (Table 4). The hazard ratio (95% CI) for developing abnormal HDL-C levels per each 1-SD higher HDL-C gene score was 1.15 (1.10–1.20). For LDL-C the comparable hazard ratio was 1.41 (1.32–1.50), whereas for triglycerides, it was 1.49 (1.41–1.56) and for total cholesterol, 1.36 (1.29–1.45). Lipid gene scores for LDL-C, triglycerides, and total cholesterol also predicted incidence of lipid-lowering medication use, regardless of actual lipid levels. There was no evidence to suggest that sex or age significantly modified the relations of lipid gene scores to an incidence of abnormal lipid levels or use of lipid-lowering medications.

Notably, for all traits, participants with higher gene scores had mean baseline lipid values closer to abnormal thresholds used to define incidence than did participants with lower lipid gene scores. For example, the cut point for HDL-C was 40 mg/dL; the mean HDL-C for those in the highest quartile of the HDL-C gene score was 48.3±16.0 versus 53.4±17.4 mg/dL for those in the lowest quintile. Similarly, those in the top quartile of LDL-C gene scores had mean LDL-C values closer to the abnormal cut point (160 mg/dL) than those in the lowest quartile (149.3±38.0 versus 125.7±36.8). The same held true for triglycerides (cut point, 200 mg/dL; mean among upper gene score quartile, 166.7±117.0 mg/dL; mean among lower gene score quartile, 112.0±63.8 mg/dL) and total cholesterol (cut point, 240 mg/dL; mean among upper gene score quartile, 227.7±40.8 mg/dL; mean among lower gene score quartile, 201.8±39.1 mg/dL).

Discussion

Using SNPs identified in a recently published lipid GWAS that included >100 000 white participants,¹² we created phenotype-specific effect size-based lipid gene scores and then related these scores to lipid levels in nearly 10 000 white participants of the ARIC cohort. For all 4 phenotypes, the lipid gene scores were cross-sectionally related to lipid levels and longitudinally related to incidence of abnormal lipid levels and use of antihyperlipidemic medications. Further, across time, participants with higher triglyceride gene scores experienced a greater rise in their triglyceride levels than participants with lower triglyceride gene scores. HDL-C, LDL-C, and total cholesterol gene scores were not related to change in levels of their respective phenotypes.

Cross-Sectional Findings

As expected, lipid gene scores were cross-sectionally related to lipid levels. For all phenotypes, however, the gene scores explained relatively little variation (<7%) in lipid levels. Teslovich et al¹² reported, as applied to Framingham Heart Study data, that lipid gene scores explained more variation (9.6%–12.4%) in lipid phenotype levels. Previous studies that created lipid GWAS gene scores found the scores to explain 3% to 6% of lipid level variation.^2,3

We also evaluated whether the association of the lipid gene scores with cross-sectional lipid levels varied by sex. Modest sex interactions were observed, with the HDL-C gene score being more strongly related to HDL-C levels among women and the triglyceride gene score being more strongly related to triglyceride levels among men. Lipid levels are known to vary by sex,¹⁹ and sex differences in the associations for some alleles have been noted.^2,3,20,21 Gonadal hormones are believed to modify both SNP penetrance and expressivity.²¹

Prospective Associations

The primary aim of the present study was to evaluate whether phenotype-specific lipid gene scores were associated with change in levels of corresponding lipid phenotypes. Different lipid phenotypes change at dissimilar rates at different points in the life course. By middle age in the United States, HDL-C levels have reached fairly stable values, and LDL-C levels are stable or, given recent secular trends, decreasing, whereas triglyceride levels continue to rise.^31–33 This may explain, in part, why we observed a positive association between the triglyceride gene score and change in triglyceride levels in these initially 45- to 64-year-olds, whereas for other phenotypes, the results were null. Notably, the modest triglyceride findings persisted even after adjusting for body mass index, which is positively related to triglyceride levels. The HDL-C, LDL-C, and total cholesterol gene scores were not related to changes in their respective phenotypes in this middle-aged population. In a previous study of a younger population (mean age, 40.8 years) higher levels of a total cholesterol gene score were associated with greater increases in total cholesterol across time.¹⁸

For all lipid phenotypes, lipid gene scores predicted incidence of abnormal levels or medication use. Persons with greater gene scores were closer to lipid cut points at baseline and, consequently, were more likely to exceed abnormal cut points over follow-up. The cross-sectional relations of lipid gene scores to lipid levels, along with the incidence of abnormal lipid level findings suggest that as additional SNPs or rare variants are identified and a greater proportion of variance explained, it may be possible that lipid genotype scores could assist in primordial prevention by helping, along with other factors, to identify individuals who would be at greatest risk of developing abnormal lipid levels across their life course.

The current results must be interpreted within the context of typical changes in lipid levels as a consequence of both aging and secular trends. In the ARIC sample, between baseline (1987–1989) and visit 4 (1996–1998), HDL-C levels fluctuated but remained fairly stable, whereas LDL-C levels decreased, and triglyceride levels increased. This mirrors national trends according to National Health and Nutrition Examination Survey data, which show that during a similar time frame (between 1988–1994 and 1999–2002), LDL-C levels decreased, there was no change in HDL-C levels, and there was a nonsignificant increase in triglyceride levels.³² Regardless of secular trends in lipid levels, individual-level linear mixed-models analysis of lipid levels among ARIC participants enabled exploration of contrasts among individuals with different lipid gene scores.

Strengths and Limitations

The strengths of the current study are the large population-based sample and the multiple lipid measurements across time. Limitations include some SNPs being imputed, potential lipid measurement error, and the fact that we had to estimate the effect of medication use on lipid levels. In sensitivity analyses, however, results were similar when medication users were excluded, suggesting that improperly accounting for medication use did not substantially bias our results. Further, because our sample consisted of middle-aged white participants, the results may not be generalizable to populations of different age ranges and to nonwhites. Given that LDL-C and HDL-C levels are fairly stable by middle age,^31–33 it would be interesting to evaluate the relation of HDL-C and LDL-C gene scores to change in HDL-C and LDL-C levels in a younger population. Finally, although there are many advantages to using gene scores, use of gene scores prohibits identification of the individual SNPs most strongly related to lipid levels or change in lipid levels.

Conclusions

In this sample of nearly 10 000 white participants, lipid gene scores were related to cross-sectional lipid levels and prospectively associated with incidence of abnormal lipid levels and use of antihyperlipidemic medications. Additionally, the triglyceride gene score was associated with increases in triglyceride levels over time. As additional relevant genes are identified and a greater proportion of the variance of lipid levels is explained, it is possible that gene scores might be better reflective of lifetime exposure to serum lipids than a measurement taken at a single point in time. Understanding the relation between gene scores and the evolution of blood lipids over time has potential to enhance information used in cardiovascular disease risk prediction.

Sources of Funding

ARIC is carried out as a collaborative study supported by National Heart, Lung, and Blood Institute contracts N01-HC-55015, N01-HC-55016, N01-HC-55018, N01-HC-55019, N01-HC-55020, N01-HC-55021, N01-HC-55022, R01HL087641, R01HL59367, and R01HL086694; National Human Genome Research Institute contract U01HG004402; and National Institutes of Health contract HHSN268200625226C. The infrastructure was partly supported by grant number UL1RR025005, a component of the National Institutes of Health and NIH Roadmap for Medical Research.

Disclosures

None.

Footnotes