|Home | About | Journals | Submit | Contact Us | Français|
Whereas epidemiologic studies show that levels of low-density-lipoprotein cholesterol (LDL-C) and high-density-lipoprotein cholesterol (HDL-C) predict incident cardiovascular disease (CVD), there is limited evidence relating lipoprotein subfractions and composite measures of subfractions to risk for CVD in prospective cohort studies.
We tested whether combinations of lipoprotein subfractions independently predict CVD in a prospective cohort of 4,594 initially healthy men and women (the Malmö Diet and Cancer Study, mean follow-up 12.2 years, 377 incident cardiovascular events). Plasma lipoproteins and lipoprotein subfractions were measured at baseline with a novel, high-resolution ion mobility technique. Principal component analysis (PCA) of subfraction concentrations identified three major independent (i.e., zero correlation) components of CVD risk, one representing LDL-associated risk, a second representing HDL-associated protection, and the third representing a pattern of decreased large HDL, increased small/medium LDL, and increased triglycerides. The last corresponds to the previously described “atherogenic lipoprotein phenotype.” Several genes that may underlie this phenotype—CETP, LIPC, GALNT2, MLXIPL, APOA1/A5, LPL—are suggested by SNPs associated with the combination of small/medium LDL and large HDL.
PCA on lipoprotein subfractions yielded three independent components of CVD risk. Genetic analyses suggest these components represent independent mechanistic pathways for development of CVD.
Prospective epidemiologic studies have shown that plasma levels of LDL-C and HDL-C predict incident atherosclerotic cardiovascular disease (CVD).1 However, there is considerable heterogeneity among lipoproteins, e.g., ranging from small, dense, lipid-depleted LDL particles to large, buoyant, cholesterol-enriched LDL particles.2 Various studies have evaluated the associations of small LDL particles,3 total LDL particle number,3 specific HDL subfractions,4,5 and combined measures such as the LDL-C/HDL-C and apoB/apoA-I ratios6-8 with cardiovascular risk. While a number of these studies suggest clinically useful metrics, no study has systematically addressed the interrelationships among the lipoproteins and asked whether there are distinct combinations of lipoprotein subfractions—possibly reflecting specific underlying mechanistic pathways—that independently confer cardiovascular risk.
We set out to identify any such combinations of lipoproteins in a systematic fashion in a prospective cohort study. To directly measure lipoprotein particles with high resolution, we used a novel ion mobility method.9,10
Our study had two objectives. First, we applied principal component analysis to identify interrelated combinations of subfractions and investigated their relationships to incident CVD in a large prospective cohort study of men and women. Second, to better understand the mechanistic pathways underlying these combinations, we tested their associations with common variants in 29 genetic loci that we have recently shown to be associated with LDL-C, HDL-C, or triglyceride levels.11
The Malmö Diet and Cancer Study Cardiovascular Cohort is a prospective, community-based epidemiological cohort of 6,103 residents of Malmö, Sweden. The women in the cohort were born between 1923 and 1950, and the men between 1923 and 1945. At baseline examinations (conducted between 1991 and 1996), all participants underwent a complete history, physical examination, and laboratory assessment of cardiovascular risk factors, including cigarette smoking, family history of myocardial infarction, blood pressure, presence of diabetes mellitus, C-reactive protein, lipid-lowering therapy, and a fasting lipid profile measured according to standard procedures. LDL-C levels were calculated with Friedewald’s formula, except when triglycerides exceeded 400 mg/dl. Apolipoprotein B (apoB) and A-I (apoA-I) levels were not available. Blood samples were obtained and placed in long-term storage at −80°C.
An ion mobility method was used on these blood samples to directly quantify the full spectrum of lipoprotein particles, from the smallest, most dense HDL particles to large, buoyant VLDL particles, in archived baseline plasma samples (Supplementary Table I). This method employs an ion separation/particle detector system that separates ions by size and which can count lipoprotein particles over a wide range of sizes. A full description of this technique is published elsewhere.9,10 Since the original publication of the method, refinements to the technique have been made that address concerns about the method.10,12 These refinements were fully incorporated into the ion mobility measurements used for this study and are described in Supplementary Methods.
We pre-specified a primary composite cardiovascular endpoint of myocardial infarction, stroke, and death from coronary heart disease, as well as a secondary coronary endpoint of myocardial infarction and death from coronary heart disease. Events were ascertained from local or national records on the basis of ICD-9/-10 codes as previously described.13 All subjects still being followed were censored after December 31, 2005. The mean follow-up time was 12.2 years.
We selected 29 single nucleotide polymorphisms (SNPs) in 29 loci that displayed associations with HDL-C, LDL-C, and/or triglycerides at genomewide significance (P < 5 × 10−8) in a recently completed meta-analysis of seven genome-wide association studies (Supplementary Table II).12 Genotyping was performed as previously described.13 All SNPs were in Hardy-Weinberg equilibrium (P > 0.001). The estimated genotyping error rate based on a subset of samples was < 1%.
Non-parametric Spearman rank correlations were computed for all combinations of each of the eleven lipoprotein subfractions and HDL-C, LDL-C, and triglycerides. Principal component analysis was performed on the eleven lipoprotein subfractions with or without lipids, using all individuals (4,594 total, 2,749 women, 1,845 men) with complete data for all subfractions. The principal component (PC) vectors were used to calculate PC values for each individual from the sum of weighted terms for each lipoprotein or lipid, with each term being the product of the vector-derived weighting and the particle concentration minus the subfraction mean, divided by the standard deviation for the subfraction (all values used for the first three principal components are available in Supplementary Table I and Table 2).
For analyses of associations between lipoprotein subfractions, PCs, or lipid measures and incident cardiovascular events or coronary events, we excluded subjects with prior myocardial infarction/stroke or on baseline lipid-lowering therapy. For the remaining 4,368 subjects, we constructed multivariable proportional-hazards models treating the PCs, lipid levels, particle concentrations/ratios, and peak LDL particle diameter as continuous variables and used the time to the first cardiovascular or coronary event as the dependent variable, adjusting for age, gender, systolic blood pressure, use of antihypertensive medications, diabetes, and current smoking status. Proportionality of hazards for each variable was assessed by Wald chi-square testing. The three PC measures were tested for association with the endpoint separately and together in a single model; they were also tested in models additionally adjusting for HDL-C, LDL-C, and triglyceride levels, as well as in analyses stratified by gender.
For the 29 SNPs and each of the subfractions, PCs, or lipid measures, multivariable linear regression analyses were used to test whether each measure differed according to an increasing copy number of the SNP minor allele, adjusted for age, gender, and diabetes status. P < 0.001 was considered statistically significant, accounting for multiple testing.
SPSS (version 16.0) and STATA (version 10) were used for the analyses.
Of the 6,103 individuals in the MDC-CC cohort, 4,822 individuals had available plasma samples that were subjected to ion mobility analysis for quantification of lipoprotein subfractions. Of these, 4,758 samples yielded complete lipoprotein subfraction profiles, of which 4,594 (2,749 women, 1,845 men) had standard lipid data (LDL-C, HDL-C, and triglycerides) available and were used for analyses. Percentiles, means, and standard deviations for these measurements are listed in Supplementary Table I. All subjects were of self-reported European ancestry. Other baseline characteristics of the MDC-CC cohort have been reported previously.13
In the correlation analyses, we found small and medium-sized LDL particle concentrations to be inversely correlated with large HDL particle concentrations (ρ = −0.46, – 0.55; Table 1) and directly correlated with triglycerides (ρ = 0.45, 0.54). Notably, these correlations were far stronger than that between small/medium LDL and small HDL (ρ = 0.04, 0.04), as well as stronger than that between small/medium LDL and LDL-C (ρ = 0.22, 0.29). Large HDL but not small HDL was highly correlated with HDL-C (ρ = 0.82, 0.17).
Principal component analysis is a statistical technique that analyzes the interrelationships between numerous variables and yields a fewer number of components that explain most of the correlation information of the original variables. Each of these components is an independent, linear combination of the original variables. We reasoned that the correlations observed among the lipoprotein subfractions may reflect hitherto unmeasured biological processes that may be more relevant to cardiovascular risk than the standard lipoprotein measures. Thus, we applied principal component analysis to the ion mobility measurements of the MDC-CC cohort.
Three major principal components (PC1, PC2, PC3) were found to account for 76% of the overall variance among the subfractions (Table 2). Similar results were obtained when HDL-C, LDL-C, and triglycerides were included in the model (Table 2). The strongest contributors to PC1 were the LDL and VLDL subfractions, whereas HDL subfractions were the strongest contributors to PC3, with lesser contributions from LDL subfractions. Notably, PC2 had the strongest contributions from small LDL/medium LDL/triglycerides in the positive direction and large HDL in the negative direction, indicating that as the first group of subfractions increase, large HDL decreases.
Of 4,368 individuals, 377 (8.6%) experienced an incident cardiovascular event during the follow-up period, with 206 (4.7%) being coronary events. With all three major PCs incorporated into a single model, PC2 and PC3 were strongly associated with cardiovascular events (P = 0.001 for both), and PC1 was associated at borderline significance (P = 0.07; Table 3). The same levels of association were seen when each of the major PCs was analyzed in a model separately (data not shown). The fourth through eleventh (minor) PCs were not significantly associated with events (data not shown).
Associations of individual subfractions, composite subfraction measures, and traditional lipid traits with cardiovascular events are shown in Table 3. Among individual subfractions, the strongest effects on risk (either increasing or decreasing risk) for cardiovascular events were seen for small LDL (HR 1.14; 95% CI, 1.04-1.25; P = 0.004), medium-sized LDL (HR 1.17; 95% CI, 1.07-1.28; P = 0.001), and large HDL (HR 0.76; 95% CI, 0.66-0.88; P < 0.001) particles. The other subfractions were either nominally or not associated with cardiovascular events. Analyses performed with coronary events yielded similar results (Table 3).
The weak association of PC1 with cardiovascular events was eliminated by adjustment for LDL-C and HDL-C but not for triglycerides (Table 4). For both PC2 and PC3, the association with cardiovascular events was weakened by adjustment for HDL-C but not for either LDL-C or triglycerides. Of the three PCs, only PC2 was associated with cardiovascular events in both men and women in subgroup analyses (Table 5).
We assessed the strength of association between each of 29 SNPs and the three PCs (Supplementary Table II). For those SNPs showing association with HDL-C, LDL-C, or triglycerides, we also tested for association with large HDL, small LDL, and medium LDL (Table 6). In total, eleven loci were highly associated (P < 0.001) with HDL-C, LDL-C, or triglycerides (Table 6). Of these, six loci—CETP, LIPC, GALNT2, MLXIPL, APOA1/A5, LPL—were highly associated with PC2 and were associated to some degree with both small/medium LDL and large HDL (Table 6). None of the six loci was associated with large LDL or small HDL (data not shown). CETP alone was associated with PC3 and was also weakly associated with PC1 (P = 0.02). Of the other five loci, three were strongly associated with both LDL-C and PC1—APOB, APOE, CELSR2/PSRC1/SORT1.
In this prospective cohort of healthy individuals, we used principal component analysis to identify three independent combinations of correlated lipoprotein subfractions and assess the contribution of these combinations to cardiovascular risk. One of the components (PC2) represents a relationship among the lipoprotein particle classes—small/medium LDL and large HDL—that were most highly associated with risk for CVD in our study, consistent with a recently published NMR lipoprotein analysis of the Women’s Health Study.5 PC2 was associated with cardiovascular events in both men and women, and it was strongly associated with six genetic loci.
We believe our data are of mechanistic relevance for several reasons. First, the PCs represent three independent axes of CVD risk that are not weakened by combining them into a single regression model, unlike traditional lipid measures. PC1 mainly comprises the LDL and VLDL subfractions and appears to best capture the component of CVD risk related to LDL-C; hazard ratios and confidence intervals of PC1 and LDL-C for cardiovascular events were very similar (Table 3), and PC1 was the only PC whose association with events was eliminated by adjustment for LDL-C (Table 4).
PC2 is unique because it represents an inverse correlation between small LDL/medium LDL/triglycerides and large HDL particles, i.e., the magnitude of PC2 is largest for individuals with high levels of lipoproteins in the first group and low levels of large HDL particles, or vice versa. This pattern matches the “atherogenic lipoprotein phenotype” first proposed two decades ago.14 This phenotype comprises a triad of decreased HDL-C, increased smaller LDL particles, and increased triglycerides that confers substantially higher risk for CVD15,16 and is characteristic of type 2 diabetes, insulin resistance, and the metabolic syndrome.17,18 Thus, principal component analysis—which by design yields independent components with zero correlation—provides unbiased evidence that the atherogenic lipoprotein phenotype represents an independent mechanism of cardiovascular risk.
PC3, which has the greatest contributions from both large and small HDL particles, was inversely associated with events and appears to reflect a contribution of HDL-C to CVD risk that, in contrast to PC2, has smaller contributions from non-HDL lipoproteins including an inverse correlation with large LDL particles (Table 3).
Second, we found that most of the SNPs that were strongly associated with HDL-C or triglycerides were associated with PC2, with opposite effects on LDL and HDL subfraction particle concentrations. Four of the implicated genes—CETP, LIPC, APOA1/A5, LPL—have been reported to be associated with LDL or HDL particle size/distribution,19-27 while the roles of two genes—GALNT2, MLXIPL—in lipoprotein metabolism are less clear. MLXIPL encodes a transcriptional regulator whose targets include genes involved in triglyceride metabolism.28 The genes implicated by SNPs in these loci appear to define a distinct biological pathway that (when dysregulated) gives rise to the atherogenic lipoprotein phenotype.
Third, we note that the PCA technique used in this study can be viewed as a “redistribution” of the CVD risk associated with the traditional lipid profile—HDL-C, LDL-C, and triglycerides—into three independent risk factors. As observed above, PC1 appears to be a proxy for plasma LDL-C levels, with only PC1’s association with cardiovascular events eliminated by adjustment for LDL-C. In contrast, all three PCs’ associations with events were weakened by adjustment for HDL-C, indicating that the HDL-C-associated CVD risk overlaps with and is “redistributed” among the three PCs.
Thus, we hypothesize that the strong epidemiological association of plasma HDL-C levels with CVD seen in numerous studies emerges as a result of HDL lipoprotein particles contributing to cardiovascular risk through at least three independent mechanisms, embodied by the three PCs. As suggested by the SNP analyses, several genes—LIPC, APOA1/A5, LPL, GALNT2, MLXIPL—appear to act primarily through one of these mechanisms (PC2), whereas CETP is unique in acting through multiple risk factors (PC2 and PC3 and, perhaps more weakly, PC1).
Finally, the present findings help clarify the conflicting conclusions of previously published reports on lipoprotein particles and cardiovascular risk.3 While a number of studies have found that small LDL particle concentrations are associated with cardiovascular disease,4,29-35 a few studies have not found this relationship.36-38 Our study, demonstrating that (1) levels of small/medium LDL particles are associated with and (2) levels of large LDL particles are not significantly related to cardiovascular disease, is consistent with other large prospective cohort studies whose data were obtained by two different lipoprotein measurement techniques: the Québec Cardiovascular Study,4,34 which used gradient gel electrophoresis, and the Multi-Ethnic Study of Atherosclerosis35 and Women’s Health Study,5 which both used NMR.
The disproportionate influence of large LDL on the measured plasma LDL-C level accounts for the latter’s weaker association with cardiovascular risk than small LDL particle concentration or total LDL particle number, which may be more appropriate than LDL-C as targets of therapeutic intervention. Measurement of the PCs may provide even greater specificity in evaluating the effectiveness of lipid-altering treatment. Accordingly, it would be of interest to know how lipid-modifying medications affect each of the three PCs. The application of principal component analysis in trials of lipid-modifying therapies, as well as specific analyses on the utility of the PCs for clinical risk prediction and as therapeutic targets with which to select and titrate medications, will be the focus of future studies.
Funding Sources: K.M. is supported by a T32 grant in Cell and Molecular Training for Cardiovascular Biology from the United States National Institutes of Health (NIH). M.O.-M. and O.M. are supported by the Swedish Medical Research Council, the Swedish Heart and Lung Foundation, the Medical Faculty of Lund University, Malmö University Hospital, the Albert Påhlsson Research Foundation, the Crafoord foundation, the Ernhold Lundströms Research Foundation, the Region Skane, the Hulda and Conrad Mossfelt Foundation, the King Gustaf V and Queen Victoria Foundation and the Lennart Hanssons Memorial Fund. S.K. is supported by a Doris Duke Charitable Foundation Clinical Scientist Development Award, a charitable gift from the Fannie E. Rippel Foundation, the Donovan Family Foundation, and a K23 career development award from the NIH. R.M.K. is supported by U01HL069757 from the NIH and has received research support from Quest Diagnostics, Inc. For this study, ion mobility measurements of study participants were provided free of charge by Quest Diagnostics, which had no role in the analysis and interpretation of the data or in the preparation, review, or approval of the manuscript.
Disclosures: Dr. Musunuru has served as a consultant for Alnylam Pharmaceuticals. Dr. Kathiresan has received research support from Alnylam Pharmaceuticals and serves on an Advisory Board for Merck & Co. Dr. Krauss has served as a consultant to Celera Genomics, has received research support from Quest Diagnostics, and is a co-inventor on patents for the technology underlying the ion mobility method. Drs. Caulfield, Li, Salameh and Reitz are employees of Quest Diagnostics. The other authors declare no conflicts of interest or relevant financial interests. None of the authors has been specifically compensated by any of the listed companies in the writing of this manuscript. Drs. Musunuru, Orho-Melander, and Melander served as independent analysts with no relationship with Quest Diagnostics.