Introduction Standard measures of plasma lipoprotein concentration do not reveal heterogeneity in the size of lipoprotein particles or their content of cholesterol and triglycerides. Yet recognizing this heterogeneity may be essential for understanding qualitative differences in lipid metabolism among individuals. Some reports identify a pattern in the size distribution of lipoprotein sub-fractions as intimately connected with coronary heart disease [1],[2]. Related findings identify a link between lipoprotein profile and metabolic syndrome, and by inference to diabetes [3]. While these observations remain controversial for prognostic use [4], they point to alterations in lipoprotein metabolism in disease. The variation in particle size and lipid content can be quantified accurately by NMR-based methods that determine lipoprotein particle concentration according to lipid class and particle size. Thus, NMR methods can measure concentration of large and small low density lipoprotein (LDL) particles as well as concentration of the related intermediate density lipoprotein (IDL) particles, and similarly concentration of small, medium, and large high density lipoprotein (HDL) or very low density lipoprotein (VLDL) particles. HDL and LDL particle concentration can also be estimated by chemical measures of apolipoprotein A1 (ApoA1) and apolipoprotein B (ApoB) protein concentration, respectively, but neither these assays nor other standard clinical assays provide information about particle size distribution, and consequently the apportionment of cholesterol and triglycerides to different sized particles. The greater precision in characterizing lipoprotein profiles using NMR-based techniques provides an opportunity for correspondingly greater detail in understanding lipid metabolism, for example by genome-wide genetic analysis, as has been done recently for plasma concentration LDL-C, HDL-C, triglycerides, ApoA1, and ApoB [5]–[13]. Results Genome-wide association analysis of 22 NMR-based and conventional lipoprotein fractions Among 17,296 WGHS participants with confirmed European ancestry (Table 1), we performed genome-wide association analysis assuming an additive genetic model for 22 plasma lipoprotein measures determined either by NMR methods or by standard clinical assay. On the basis of genome-wide significance (P 150kb) from known genic regions. Among the standard clinical measures LDL-C, HDL-C, and triglycerides only, novel genome-wide loci were found at KLF14 (7q32.2) and CCDC9/DNAH10/ZNF664 (12q24.31.B), both for triglycerides. The association at the novel locus 8p23.1 (which differentiated the fasting sample from the whole sample on the basis of mean VLDL particle size) is over 1.8 Mb from a recently described association at 8p23.1 between SNP rs7819412 and triglycerides [6]. The remaining 24 unique loci suggested genes recognized for a diversity of roles in lipid metabolism, broadly defined (Figure S1). Thus, SNPs with genome-wide significance, were confirmed in or near PCSK9 (at 1p32.3), APOA2 (1q23.3), APOB (2p24.1), ABCG5/8 (2p21), HMGCR (5q13.3), LPL (8p21.3), APOA1-A5 (11q23.3), ABCA1 (9q31.1), FADS1-3 (11q12.2), LIPC (15q22.1), CETP (16q13), LIPG (18q21.1), LDLR (19p13.2), the APOC-APOE complex (19q13.32), and PLTP (20q13.12). Similarly, association at 9q34.2 implicating the ABO gene recapitulates and extends the known association between blood group antigen and total cholesterol [14],[15]. Less well characterized genic regions, which nonetheless have been validated recently for roles in lipid metabolism, were confirmed for ANGPTL3 (1p31.3), CELSR2/MYBPHL/PSRC1/SORT1 (1p13.3), GCKR (2p23.3), MLXIPL (7q11.23), and TRIB1 (8q24.13), HNF1A (12q24.31.A), and HNF4A (20q13.12). The association at COBLL1/GRB14 (2q24.3) with HDL-C was recently described elsewhere in this same cohort and validated by replication [16]. The previous study found much stronger association in women than men, suggesting a potential interaction with gender. At this locus, the gene GRB14 is thought to inhibit receptors in the insulin receptor class [17],[18]. The current analysis extends associations at this locus to concentrations of LDL, HDL, and VLDL particles according to size (Table S1). Consistent with a high degree of correlation among the lipoprotein measures (Table S2), the rank order by p-value among the highly significant SNPs was similar for each measure with at least one genome-wide significant association (Figure S1). A notable exception was the APOB gene (2p24.1), where the ordering of the p-values, conditional analysis, and patterns of linkage disequlibrium (LD) among the top SNPs (Table S1) revealed three classes of associations. One class included VLDL-related fractions, triglycerides, and mean LDL size for which either rs673548 or rs676210 (LD r2 = 1.0) had the strongest association; a second class included ApoB, large LDL particles, and total LDL particles for which either rs1713222 or rs506585 (LD r2 = 0.5) had the strongest association; and a final class including only LDL-C for which rs137117 was most strongly associated (Figure 1A). Between SNPs in different classes, maximum LD ranged from r2 = 0.04–0.11. Similarly, at APOA5-APOA1 (11q23.3), p-values revealed two classes of associations seemingly segregating between effects nearer the APOA5 gene involving triglycerides and effects nearer the APOA1 gene involving HDL related lipoprotein fractions (Figure 1B). 10.1371/journal.pgen.1000730.g001 Figure 1 Loci with distinct classes of SNP associations among lipoprotein fractions with genome-wide significance. (A) APOB locus (2p24.1), (B) APOA1-A5 locus (11q23.3). Recombination rates are from [41]. Large, well-characterized cohorts with NMR-based measurement of lipoprotein fractions are scant, but sub-samples of about 2700 participants in the Framingham Heart Study Offspring cohort (FHS) [19] and about 2000 total CHD cases and controls from PROCARDIS [20] had both the NMR-based lipoprotein measures and genome-wide genetic data already determined. Among all candidate loci, concordance of direction of effects was observed respectively at 124 out of 146 (84%) [84% in fasting sub-sample] and 125 out of 133 (94%) [99% in fasting sub-sample] of the candidate associations for which there was genotype information in FHS and PROCARDIS (Table S3 [whole WGHS sample candidates], Table S4 [fasting WGHS subsample candidates]). For each of the previously known loci except ABCA1 (9q31), at least one of the candidate associations was nominally significant (P 0.05; Table S3). However, a recent genome-wide meta-analysis of LDL-C, HDL-C, and triglycerides found significant, but not genome-wide significant, associations among these fractions with candidate SNPs from the WGHS at PCCB/STAG1 (3q22.3), BTNL2 (6p21.32), KLF14 (7q32.2), and 8p23.1 [10], although the significant SNP associations at PCCB/STAG1 (3q22.3) and BTNL2 (6p21.32) were not fully concordant between the two studies (Table 3). Independent evidence for functional consequence of the candidate SNP (rs10778213) at 12q23.2 is its genome-wide significant association in a smaller sample from the WGHS with plasma C-reactive protein (CRP), a biomarker of inflammation that is slightly correlated if at all with the two HDL measures associated at this locus (total HDL particle concentration [HDL:T], Spearman r = 0.22; HDL cholesterol estimated by the NMR [HDL:N], Spearman r = −0.04) [21]. With the larger sample of WGHS genotype information in the current study, the association with plasma CRP is more significant (P 0.05). e Genome-wide significant association with plasma C-reactive protein in the WGHS [21]. Magnitudes of genetic effects To assess the contribution of common genetic variation at each of the candidate loci to each of the adjusted lipoprotein fractions, we constructed regression models by stepwise selection of SNPs in the vicinity of the primary genome-wide significant associations. Most of these models explain less than 1% of the variation in the adjusted lipoprotein fractions (Figure 2, Table S5, and Table S6). The top three effects, all at APOC-APOE complex (19q13.32), explain 8.9%, 8.4%, and 7.1% of the variance in ApoB particle concentration, the related total LDL particle concentration, and LDL-C, respectively. Fasting status had an influence on retention of SNPs in the model selection procedure, but only for loci with modest effects (Compare Table S5 and Table S6). There were no genetic contributions remaining from the model selection procedure for any of LDL-C, HDL-C, triglycerides, ApoA1, or ApoB concentration at APOA2 (1q23.3) in the whole sample and at WIPI1 (17q24.2) in the fasting subsample, suggesting that these loci would not have been identified for genome-wide association with the five conventional lipoprotein fractions even in a much larger sample with the genome-wide SNP genotyping panel used in this study. Clustering loci on the basis of the profile of associated lipoprotein fractions suggests sub-groups of loci with related patterns of effects (Figure S2, Figure S3), perhaps suggesting distinct but possibly overlapping biological pathways for lipoprotein metabolism. For example, HNF1A, LDLR, ABCG5/8, PCSK9, and CELSR2/PSRC1/SARS/SORT1 largely share associations with IDL, small VLDL, total VLDL large LDL, LDL-C, total LDL, and ApoB. 10.1371/journal.pgen.1000730.g002 Figure 2 Variance explained in adjusted lipoprotein measures by common variation at the candidate loci by SNPs retained in model selection procedures. See also Figure S2 and Figure S3. The total genetic effects for each lipoprotein determined by summing over the effects at all loci ranged from 2.1% for mean VLDL size to 17.2% for ApoB (Table 4). The effects were not substantially different when the entire model selection procedure was performed in the fasting subsample (Table 4), and only slightly smaller in general among the unadjusted lipoprotein fractions (Table S7). Notably, the common genetic variation in this study at the genome-wide loci had a greater total effect on mean particle size than on standard clinical cholesterol measures for HDL but not for LDL or VLDL (Table 4). 10.1371/journal.pgen.1000730.t004 Table 4 Proportion (%) variance in fully adjusted lipoprotein fractions explained by common variation at candidate loci. lipoprotein fraction whole sample fasting subsample LDL large 12.0 11.4 LDL small 8.9 9.4 LDL mean size 8.5 8.7 IDL total 3.5 3.5 LDL total 15.2 15.0 LDL-C assay 13.7 13.8 ApoB assay 17.2 16.8 HDL total 5.6 5.6 HDL large 13.1 12.5 HDL medium 4.6 4.4 HDL small 6.4 5.7 HDL mean size 12.2 11.7 HDL-C by NMR 10.3 9.9 HDL-C assay 9.9 9.1 ApoA1 assay 8.3 7.8 VLDL total 8.9 8.6 VLDL large 3.8 4.1 VLDL medium 6.0 6.0 VLDL small 7.6 7.4 VLDL mean size 2.1 2.5 TG by NMR 7.9 7.6 TG assay 7.7 8.1 Secondary genome-wide analysis To examine the possibility that other loci might include SNPs with genome-wide significant association conditional on effects at the primary loci, we adjusted the primary lipoprotein fraction measurements (which were already adjusted for clinical covariates) for SNPs retained by the model selection procedure at the candidate loci, and repeated the genome-wide association testing. Quantile-quantile analysis confirmed that all of the excess of extremely small p-values in the original analysis could be explained by the variation at the candidate loci (not shown). Similarly, genotype-based statistical models (as opposed to the allele-based additive models used in the primary analysis) did not reveal other loci with genetic influences at the genome-wide significance level in the whole sample. While we adjusted the lipoprotein measures with a full set of clinical characteristics to reduce variance and enhance power in the primary analysis, it remained possible that relevant SNPs would be overlooked if they acted through effects on the adjustment covariates. Similarly, subtle effects on the association estimates due to non-normality of the (possibly log-transformed) adjusted lipoprotein measures or due sub-European population stratification might confound hypothesis testing. To evaluate whether our discovery procedure was robust, we performed secondary analyses repeating the entire genome-wide discovery procedure for alternative nested subsets of clinical covariates with and without further adjustment for population structure and quantile normalization (Table S8). Comparing the full adjustment procedure to alternatives using either a reduced set of clinical covariates or age only, with or without additional adjustment for potential sub-European population stratification and quantile normalization yielded further genome-wide significant associations at three loci with known lipid metabolic genes, LPA (6q25.3), LCAT (16q22.1), and APOH (17q24.2), and two additional loci, 6p22.3 and 10q21.3. All of the additional loci were present in the age-adjusted analysis. Associations at 6p22.3 and 10q21.3 appear to be novel and implicate, respectively the GMPR or MYLIP genes and the JMJD1C gene. The lead SNPs at each of these loci were significantly associated with at least one of LDL-C, HDL-C or triglycerides in the recently published meta-analysis (Table 5) [10]. Similarly, in internal replication among the additional 4639 WGHS samples with genotype available after the main analysis was complete, associations at the candidate SNPs were all significant and the trends of effects were all consistent with effects in the discovery sample (Table 5). We note that at JMJD1C (10q21.3), the candidate SNPs have minor allele frequency near 0.5, and that available data does not allow us to determine whether the differences in the direction of the minor allele effect on VLDL fractions in the WGHS and triglycerides in the previously published replication study are truly physiological or rather that the frequency of the coded (i.e. minor) allele from the WGHS is greater than 0.5 in the replication cohort resulting in an opposite sign of the effect estimates. 10.1371/journal.pgen.1000730.t005 Table 5 Genome-wide significant associations (p 0.05). # Abbreviation as in Table 2. ∧ P-value (two-sided) for association in additional 4639 samples from the WGHS. All association trends were consistent with discovery sample. Combining these new samples with the original discovery samples leads to p–values for the extended WGHS sample. Since lipoprotein particle size is closely related to triglyceride content, we also performed secondary analysis examining genome-wide significant associations after adjustment of the lipoprotein fractions by the full set of clinical covariates and (log-transformed) triglyceride levels (Table 5 and Table S8). This analysis identified only one new genome-wide significant association. At 11p15.4, rs7938647 in the intron of the SBF2 gene was associated with full-plus-triglyceride adjusted total HDL particle concentration. Again, internal replication provided support for this association although there was no association (P>0.05) with LDL-C, HDL-C, or triglycerides in the recent meta-analysis for replication. Associations distinguishing NMR-based from conventional lipoprotein measures Among its unique characteristics, the NMR-based methodology provides information about IDL and VLDL particle concentration, both aspects of lipoprotein profiles that are difficult to measure by conventional methods. For IDL, genetic associations were observed at many of the candidate loci (Figure 2, Table 2, Table S1) and most strongly at LIPC (15q22.1), where rs1532085 had an estimated 0.11 nmol/l shift in particle concentration for each copy of the minor allele (p = 1.5×10−20). For total VLDL concentration, association with genetic variation was observed at many loci but none more strongly than at the APOC-APOE complex where rs439401, which is in perfect LD with rs7412 (the SNP that distinguishes APOE alleles E2 and E3), had an estimated −2.4nmol/l shift in concentration per copy of the minor allele (p = 2.1×10−12; Table S1). Loci strongly affecting the relative concentration of NMR-based estimates of small, medium, and large particle size could be identified on the basis of genome-wide effects on mean particle size, and these associations were of special interest when there was no accompanying association with the corresponding cholesterol measure retained in the model selection procedures (Table 6, Figure S4). For LDL, mean particle size was associated with genome-wide significance at 12 loci (Table 2), among which the model selection procedures failed to identify any association with LDL-C at MLXIPL (7q11.23), LPL (8p21.3), CCDC92/DNAH10/ZNF664 (12q24.31.B), and LIPG (18q21.1). These loci implicate genes related to glucose or triglyceride metabolism as well as unrecognized biological function at one novel locus (CCDC92/DNAH10/ZNF664 [12q24.31.B]). The associations with mean LDL particle size were a consequence of strong inverse effects on large and small LDL particles (MLXIPL [7q11.23], LPL [8p21.3], LIPG [18q21.1]) or of exclusive effects on small LDL (CCDC92/DNAH10/ZNF664 [12q24.31.B]) [see Figure S4]. In the fasting subsample, the associations with the NMR based measures at LPL (8p21.3) and LIPG (18q21.1) also met genome-wide significance, but the associations at MLXIPL (7q11.23) and CCDC92/DNAH10/ZNF664 (12q24.31.B) did not. For HDL, 9 loci had genome-wide significance for mean particle size (Table 2), among which the clinical measure of HDL-C was not associated with genetic variation only at GCKR (2p23.3), as was also found in the fasting subsample (Figure 2, Table 6). The discordant effects on LDL size and cholesterol content at LPL (8p21.3), CCDC92/DNAH10/ZNF664 (12q24.31.B), and LIPG (18q21.1) but not those of HDL size and cholesterol content were independent of triglyceride level in as much as associations persisted in analysis that further adjusted the lipoprotein fractions for (log-transformed) triglycerides, although only at nominal significance rather than genome-wide significance (Table 6). 10.1371/journal.pgen.1000730.t006 Table 6 Loci with genome-wide significant association (P 1%, successful genotyping in 90% of the subjects, and deviations from Hardy-Weinberg equilibrium not exceeding P = 10−6 in significance. A total of 335,603 unique SNPs, of which 32,521 derive from the custom content, remained in the final data. Although assays for two non-synonymous SNPs at the APOE locus (19q13.32), rs429358 and rs7412, which determine ApoE isotype, failed in the design of the Illumina custom content, genotypes for these two SNPs were determined separately by an allele-specific, PCR based method (Celera, Alameda, CA) [34]. These additional SNPs are in linkage disequilibrium with SNPs in the Illumina panel. The targeted genotypes for APOE were included during the model selection procedures but not during the primary analysis to discover loci with genome-wide significant associations. Analytic methods Primary analysis to discover loci with highly significant associations in the WGHS discovery cohort was performed by linear regression in PLINK [35] assuming an additive relationship between the number of copies of the minor allele of each SNP and the mean values of the adjusted lipoprotein measures. A conservative threshold of P 0.4) among the HapMap CEU, YRI, and JPN+CHB populations [37]. Discrepancy between self reported European ancestry and the clustering pattern was observed only for 68 samples ( 0.3) were used in the analysis. In the PROCARDIS study [20], where genotype data derive from the Illumina (San Diego, CA) Human 1M platform representing a superset of the SNPs in the WGHS data, lipoprotein fractions were adjusted for case/control specific effects of age at baseline (continuous), gender, country of recruitment (Germany, Italy, Sweden, United Kingdom), self-reported hypertension (yes/no), diabetes (yes/no), current smoking status by questionnaire (yes/no), and statin therapy (yes/no). Regression models assumed a linear relationship between the number of copies of the minor allele and adjusted mean lipoprotein measure. Supporting Information Figure S1 Locus p-values for lipoprotein fractions with at least one SNP reaching genomewide significance at each of the candidate loci. All plots correspond to analysis in the whole sample except for locus 8p23.1, for which genomewide association was observed only in the fasting subsample as shown. (0.41 MB PDF) Click here for additional data file. Figure S2 Primary loci clustered hierarchically according to Cartesian distance corresponding to whether ( = 1) or not ( = 0) there were associations with each of the lipoprotein fractions in the model selection procedures (see Materials and Methods). (0.02 MB PDF) Click here for additional data file. Figure S3 Dendorgram showing bierarchical relationships between loci clustered as in Figure S2. (0.01 MB PDF) Click here for additional data file. Figure S4 Normalized SNP effects (beta coefficients) from univariate regression models. All plots correspond to analysis in the whole sample except for locus 8p23.1, for which genome-wide association was detected only in the fasting subsample as shown. Locus SNPs are shown if they were retained in the model selection procedure for at least one lipoprotein fraction. Absence of shading indicates the univariate beta coefficient was not significant (p>0.05). A small black dot for some combinations of SNPs and lipoprotein fractions indicates genomewide significance for the univariate beta coefficient. (0.10 MB PDF) Click here for additional data file. Table S1 Best genome-wide associations with the lipoprotein fractions at each candidate locus. (1.10 MB DOC) Click here for additional data file. Table S2 Correlations between all pairs of lipoprotein fractions. (0.12 MB DOC) Click here for additional data file. Table S3 Replication of WGHS candidate associations from whole sample in PROCARDIS and the Framingham Heart Study. (0.56 MB DOC) Click here for additional data file. Table S4 Replication of WGHS candidate associations from fasting sub-sample in PROCARDIS and the Framingham Heart Study. (0.45 MB DOC) Click here for additional data file. Table S5 Proportion of variance in fully adjusted lipoprotein fractions explained in the whole sample by genetic variation at the candidate loci. (0.15 MB DOC) Click here for additional data file. Table S6 Proportion of variance in fully adjusted lipoprotein fractions explained in the fasting sub-sample by genetic variation at the candidate loci. (0.15 MB DOC) Click here for additional data file. Table S7 Total proportion of variance explained by candidate loci for each of the unadjusted lipoprotein fractions. (0.04 MB DOC) Click here for additional data file. Table S8 Sensitivity analysis for locus discovery procedure. (0.10 MB DOC) Click here for additional data file. Table S9 Lipoprotein associations in the whole sample at loci in previous lipid fraction GWAS. (0.34 MB DOC) Click here for additional data file. Table S10 Lipoprotein associations in the fasting sub-sample at loci in previous lipid fraction GWAS. (0.16 MB DOC) Click here for additional data file.
