Human Plasma Lipidome Is Pleiotropically Associated with Cardiovascular Risk Factors and Death

Abstract

Background

Cardiovascular disease (CVD) is the most common cause of death in the United States and is associated with a high economic burden. Prevention of CVD focuses on controlling or improving the lipid profile of patients at risk. The human lipidome is made up of thousands of ubiquitous lipid species. By studying biologically simple canonical lipid species, we investigated whether the lipidome is genetically redundant and whether its genetic influences can provide clinically relevant clues of CVD risk.

Methods and Results

We performed a genetic study of the human lipidome in 1,212 individuals from 42 extended Mexican American families. High-throughput mass spectrometry enabled rapid capture of precise lipidomic profiles, providing 319 unique species. Using variance-component based heritability analyses and bivariate trait analyses, we detected significant genetic influences on each lipid assayed. Median heritability of the plasma lipid species was 0.37. Hierarchical clustering based on complex genetic correlation patterns identified 12 genetic clusters that characterized the plasma lipidome. These genetic clusters were differentially but consistently associated with risk factors of CVD, including central obesity, obesity, type 2 diabetes, raised serum triglycerides and metabolic syndrome. Also these clusters consistently predicted occurrence of cardiovascular deaths during follow-up.

Conclusions

The human plasma lipidome is heritable. Shared genetic influences reduce the dimensionality of the human lipidome into clusters that are associated with risk factors of CVD.

Mexican Americans have a high prevalence of cardiovascular morbidity and mortality.^{1, 2} The high risk for cardiovascular disease (CVD) in this ethnic group is partly explained by a high propensity to metabolic syndrome¹ which is a constellation of clinical states that also contribute to cardiovascular death. However, since the prediction models for cardiovascular deaths have limited accuracy³, the search for important biomarkers of CVD continues. There is now a renewed interest in the potential contribution of lipids to CVD.⁴ Revolutionary advances in the methods for measuring the wide spectrum of lipid molecules in different tissues have helped characterize the “lipidome”, the complete universe of fundamental lipid species.^{5, 6} There are more than 1,000 such lipid species comprising the lipidome.^{7, 8} The emerging field of lipidomics allows the simultaneous assay of large numbers of these canonical lipids.^{8, 9} The lipidome is a rich compilation of phenotypes that may represent novel and accurate predictors of CVD risk.^{10, 11} However, since research on lipidome is still in its infancy the putative role of many lipids in health and disease is far from understood. Thus, a better appreciation of the lipidomic landscape, especially on the changing backdrop of cardiovascular health, is required.

The San Antonio Family Heart Study (SAFHS) is a unique study of Mexican American families followed for over two decades for genetic research.¹² We have demonstrated previously that various lipid species belonging to the diacylgylcerol, triacylglycerol, dihydroceramide and phosphatidylcholine classes can provide clinically meaningful information relating to complex diseases like obesity¹³, hypertension¹⁴, type 2 diabetes¹⁵ and metabolic syndrome¹⁶ in this population. These studies indicate the possibility of a redundancy in the plasma lipidome such that strong inter-lipid species correlations might partly explain why only some of the lipid species are associated with the aforementioned clinical states when examined in a multivariate context. Although such inter-lipid species correlations have been conceptually implied, their complete characterization is currently lacking. To understand the genetic influence on CVD, it is imperative to understand i) if the human plasma lipidome is itself genetically controlled and ii) whether the plasma lipidome and CVD share genetic influences. Since the SAFHS participants represent large and complex pedigrees, this study population permits assessment of the possible genetic basis of cardiovascular and other diseases with a focus on the plasma lipidome.

In this study, we set out to characterize the genetic correlations among the lipid species comprising the plasma lipidome. Our central question was whether or not these fundamental lipid components represent unique phenotypes that are closely related to polygenes which may be causally involved in cardiovascular risk. We investigated the potential genetic redundancy in the human plasma lipidome in Mexican Americans and the association of these genetically derived phenotypes with common risk factors of CVD. We also examined the association of human plasma lipidome with prospectively monitored cardiovascular deaths.

Methods

Study participants

The SAFHS began in 1991 and has enrolled large, extended Mexican American families residing in San Antonio. The enrollment procedures, inclusion and exclusion criteria, and phenotypic assessments of the study participants have been described in detail previously.^{17, 18} This is an ongoing longitudinal observational investigation which has had four phases of data collection over a 23 year period. The data and samples used in this study were collected during the first phase of data collection that lasted from 1992 to 1996. Informed consent was obtained from all participants before collection of samples. The Institutional Review Board of the University of Texas Health Science Center at San Antonio approved the study. Plasma lipidomic data was available on 1,212 participants (from 42 families). Other phenotypic data was available for 1,198 participants. The biological relationships observed in the study sample were monozygotic twins (one pair), parents-offspring (922 pairs), siblings (1,111 pairs), grandparents-grandchildren (310 pairs), avuncular relatives (2,064 pairs), half siblings (148 pairs), double first cousins (8 pairs), 3^rd degree relatives (3,321 pairs), 4^th degree relatives (2,876 pairs), , 5^th degree relatives (1,204 pairs), and 6^th degree relatives (316 pairs). The clinical characteristics of the study participants are shown in Table 1.

Table 1.

Characteristics of the Study Participants

Characteristic	Value
Age [mean (SD)] y	39.52 (16.92)
Females [n (%)]	726 (60.60)
Fasting glucose [mean (SD)] mmol/l	5.60 (2.48)
2-hour post-glucose load glucose [mean (SD) mmol/l	7.31 (5.00)
Fasting insulin [mean (SD)] μU/ml	16.19 (20.19)
2-hour post-glucose load insulin [mean (SD)] μU/ml	77.48 (73.56)
Homeostasis Model of Assessment – Insulin Resistance [mean (SD)]	4.38 (7.44)
Waist circumference [mean (SD)] cm	95.42 (17.31)
Body mass index [mean (SD)] Kg/m2	29.31 (6.62)
Waist-hip ratio [mean (SD)]	0.90 (0.10)
Systolic blood pressure [mean (SD)] mmHg	120.64 (18.72)
Diastolic blood pressure [mean (SD)] mmHg	70.70 (10.20)
Total serum cholesterol [mean (SD)] mg/dl	190.55 (38.79)
Serum triglycerides [mean (SD)] mg/dl	149.48 (102.91)
HDL cholesterol [mean (SD)] mg/dl	50.04 (12.77)
LDL cholesterol [mean (SD] mg/dl	112.32 (32.66)
Participants receiving lipid-lowering drugs [n (%)]	22 (1.86)
Participants receiving anti-hypertensive drugs [n (%)]	116 (9.80)
Prevalence of clinical states [n (%)]
Obesity	470 (39.23)
Central obesity	588 (49.08)
Raised serum triglycerides	433 (36.14)
Low HDL cholesterol	495 (41.32)
Hypertension	170 (14.19)
Type 2 diabetes	179 (14.88)
Metabolic syndrome	421 (35.14)
Cardiovascular deaths during follow-up [n (%)]	52 (4.34)

Lipidomic studies

We estimated the concentrations of a total of 319 lipid species (representing 23 lipid classes and subclasses shown in Table 2) in fasting plasma samples by combining high performance liquid chromatography and mass spectroscopy. These assays were conducted in the Metabolomics Laboratory, Baker IDI Heart and Diabetes Institute, Melbourne, Australia. The experimental protocols used have been described elsewhere.¹⁹

Table 2.

Summary of lipid classes included in the plasma lipidome

Lipid class	Acronym	Number of species
Dihydroceramide	dhCer	6
Ceramide	Cer	6
Monohexosylceramide	MHC	6
Dihexosylceramide	DHC	6
Trihexosylceramide	THC	6
GM3 ganglioside	GM	6
Sphingomyelin	SM	19
Phosphatidylcholine	PC	46
Alkylphosphatidylcholine	PC(O)	18
Alkenylphosphatidylcholine	PC(P)	8
Lysophosphatidylcholine	LPC	21
Lysoalkylphosphatidylcholine	LPC(O)	6
Phosphatidylethanolamine	PE	18
Alkylphosphatidylethanolamine	PE(O)	12
Alkenylphosphatidylethanolamine	PE(P)	9
Lysophosphatidylethanolamine	LPE	6
Phosphatidylinositol	PI	17
Phosphatidylserine	PS	7
Phosphatidylglycerol	PG	4
Cholesteryl ester	CE	26
Cholesterol	COH	1
Diacylglycerol	DG	22
Triaclyglycerol	TG	43
Total		319

Analytical approach

Heritability of the plasma lipidome

Using rich phenotypic and kinship data, we first examined the degree to which plasma levels of individual lipid species are heritable within a variance components framework. This analytical framework does not require large-scale genotyping data since kinships can inform about the extent of shared genetic information between individuals. We used polygenic regression models that predict the inverse-normalized plasma concentration of each lipid species after accounting for the kinship structure as follows: Ω = 2 φ σ²_G + I σ²_E where, Ω is the total phenotypic covariance matrix of a trait, φ is the matrix of kinship coefficients, I is the identity matrix, σ²_G is additive genetic variance and σ²_E is residual environmental variance. All the models were adjusted for the following covariates: age, age², sex, age × sex interaction, age² × sex interaction, receipt of lipid-lowering and anti-hypertensive drugs. Heritability was estimated as σ²_G/(σ²_G + σ²_E) and represented the proportion of phenotypic variance explained by the genetic similarity (represented in the kinship matrix). Inverse-normalization of lipid species was achieved by ranking the observations, generating a cumulative density function and then converting this probability function into a standardized deviate. Thus, all transformed lipid species concentrations had a mean of zero and standard deviation of unity.

Genetic redundancy among the plasma lipid species

The variance components approach has been extended to permit bivariate trait analyses^20-22 in which it is possible to further partition the phenotypic variance into genetic and environmental components. Specifically, in the context of bivariate trait analyses, the phenotypic covariance $\left({\rho }_P^2\right)$ is regarded as a function composed of the additive genetic $\left({\rho }_G^2\right)$ and environmental $\left({\rho }_E^2\right)$ covariances between two traits (denoted below as i and j).

$${\rho }_P(i,j)={\rho }_G(i,j)\sqrt{{h_i}^2{h_j}^2}+{\rho }_E(i,j)\sqrt{(1-{h_i}^2)(1-{h_j}^2)}$$

These parameters are estimated using an estimation-maximization algorithm by jointly utilizing all available pedigree information with a multivariate normal model for continuous traits.^23-25 Conceptually, a significant genetic correlation (ρ _G (i,j)) between two phenotypes indicates shared pleiotropic effect of causal genes. We used this approach to estimate the genetic correlation between each pair of lipid species. We generated a genetic correlation matrix of all the 319 lipid species with each other. We then used hierarchical clustering methods to reduce the dimensionality of this correlation matrix into meaningful clusters that are hereinafter referred to as genetic clusters of lipid species.

Clinical associations of genetic clusters

We investigated the usefulness of the genetic clusters through their association with seven prevalent clinical states, and observed cardiovascular deaths in the study participants using polygenic regression models. The clinical states that we studied were: central obesity [waist circumference ≥102 cm in males and ≥88 cm in females.²⁶], obesity [body mass index ≥30 Kg/m²] , raised triglycerides [serum triglycerides ≥150 mg/dl (1.7 mmol/l) or receipt of lipid-lowering drugs.²⁷], low high-density lipoprotein cholesterol (HDL-C) [serum HDL-C <40 mg/dl (1.03 mmol/l) in males, <50 mg/dl (1.29 mmol/l) in females or receipt of lipid-lowering drugs²⁷], hypertension [systolic blood pressure ≥140 or diastolic blood pressure ≥90 mm Hg or receipt of antihypertensive drugs²⁷], type 2 diabetes [American Diabetes Association criteria^{28, 29}] and metabolic syndrome defined as: presence of central obesity combined with any two of the following: raised triglycerides, low HDL-C, high blood pressure and raised fasting plasma glucose (≥5.6 mmol/l), previously diagnosed type 2 diabetes or receiving anti-diabetic medication [International Diabetes Federation²⁷ definition]. In addition, we had information on the deaths of study participants which was derived from death certificates provided by the San Antonio Metro Health Department, and the causes of death in deceased SAFHS participants. Using the International Classification of Diseases coding scheme ICD-10, we found that 73 deaths were reported to have cardiovascular diseases (ICD code category I) as primary or contributory cause of death in participants followed up till October 31, 2009. Deaths with following primary or contributory causes (ICD-10 codes) were defined as cardiovascular deaths: Rheumatic fever with heart involvement (I01), essential hypertension (I10), rheumatic mitral valve disease (I05.9), hypertensive heart disease (I11.9), hypertensive chronic kidney disease (I12.9), ST elevation myocardial infarction (I21.3 and I21.9), atherosclerotic heart disease of native coronary artery (I25.0 and I25.1), ischemic cardiomyopathy (I25.5), chronic ischemic heart disease (I25.9), diseases of pericardium (I31.9), hypertrophic cardiomyopathy (I42.2), cardiomyopathy (I42.9), cardiac arrest (I46.9), paroxysmal tachycardia (I47.2), cardiac arrythmia (I49.9), heart failure (I50.0), heart failure (I50.9), nontraumatic intracerebral hemorrhage (I61.9 and I62.0), sequelae of nontraumatic intracranial hemorrhage (I69.2 and I69.4), atherosclerosis (I70.9), peripheral vascular disease (I73.9), other disorders of arteries and arterioles (I77.6) and hypertension (I95.9). In the subsample on which lipidomic studies were done there were 52 cardiovascular deaths during follow-up.

Statistical Analysis

To compare distribution of continuous variables across two groups, we used the non-parametric Mann-Whitney U test. To test the assumptions of a normal distribution of a variable, we used the skewness/kurtosis test of D'Agostino, Balanger and D'Agostino Jr.³⁰ To test the variability of estimated parameters across subgroups, we used Cochrane's Q statistic. We used the R packages hclust³¹ (for hierarchical clustering), and corrplot³² (for depicting the large genetic correlation matrix). Enrichment of lipid classes within genetic clusters was assessed using Fisher's exact test and applying Bonferroni correction for multiple comparisons.

Association analyses were conducted using the SOLAR software package.³³ These analyses also used the variance components approach in a polygenic regression model as follows:

CT_i = m + Σ^b_k a_ik + g_i + e_i where, CT is the liability of a clinical trait; m is the mean; a is the covariate vector of dimension k with b as the vector of corresponding regression coefficients; g is the polygenic effect and e is the residual error for an individual indexed by i. Since all the seven clinical traits were discrete in nature, we used the liability threshold approach to model the likelihood of these traits. We modeled the term g as a random variable based on the coefficients of relationship in the kinship matrix. All models included adjustments for age, age², sex, age × sex interaction and age² × sex interaction and use of lipid-lowering and anti-hypertensive drugs and 12 cluster scores as covariates. We generated a cluster score by calculating the average of the inverse-normalized plasma concentrations of all lipid species belonging to that genetic cluster. Statistical significance of the association was tested by constraining the regression coefficient to zero and comparing the log-likelihoods of the constrained and unconstrained regression models in a likelihood ratio χ² test. Statistical significance was tested at a global type I error rate (α) of 0.05, however to correct for multiple tests, we used the Benjamini-Hochberg method of controlling false-discovery rates (FDR). We used Stata 12.0 (Stata Corp, College Station, TX) package for the Mann-Whitney U test and multiple test correction.

Results

The SAFHS participants were middle-aged with a majority of females (Table 1). The prevalence of obesity, central obesity, type 2 diabetes and metabolic syndrome was very high in this sample. Approximately 10% of the study participants were already receiving anti-hypertensive treatment and another 54 subjects had clinically detectable hypertension at the time of enrollment. Only a small proportion of the study participants (<2%) were already receiving lipid-lowering medications. Follow-up of 9,314 person-years revealed that there were 52 cardiovascular deaths in this sample with an estimated cardiovascular mortality incidence of 5.58 deaths per 1000 participants per year. The class membership of each lipid species and its relative plasma concentration are given in Supplementary Table 1. The observed relative concentrations indicate a substantial variability of the plasma lipidome across lipid species, classes, subclasses and individuals.

Human plasma lipidome is heritable

We estimated the narrow-sense heritability of each lipid species in the study participants. We found that each of the 319 lipid species had a statistically significant heritability even after correction for multiple testing (Supplementary Table 1). The heritability estimates ranged from a minimum of 0.09 (p = 0.0226 after multiple test correction) for dihydroceramide 16:0 to a maximum of 0.60 (p = 4.2×10⁻³⁴ after multiple test correction) for dihexosylceramide 24:1. The histogram of the estimated heritabilities (Figure 1A) indicated a potential asymmetry around the central value. When we tested the assumption for normal distribution of the heritability estimates using the skewness/kurtosis test, we found that the skewness significantly deviated from normality (p = 0.011) but the kurtosis was normal (p = 0.816). We therefore generated a box plot of this distribution (Figure 1B) which showed that the median heritability of the plasma lipidome was 0.3705 with an inter-quartile range of 0.1255.

Figure 1.

Heritability of the human plasma lipidome. (A) Histogram showing distribution of the estimated heritability of plasma lipid species. (B) Box plot summarizing the heritability distribution. (C) Box plot showing the distribution of heritability within each lipid class. Lipid classes are abbreviated as shown in Table 2.

We then explored whether or not the estimated heritabilities were similar across the lipid classes and subclasses. A box plot (Figure 1C) showed that there was considerable variation in the estimated heritability across lipid classes and subclasses with the alkylphosphatidylethanolamines (PE(O)) showing the least median heritability (0.2220) and the monohexosylceramides (MHC) showing the highest median heritability (0.5002). There was a statistically significant heterogeneity in heritability across lipid classes and subclasses (Q = 69.64, degree of freedom = 22, p = 7.53×10⁻⁷). In general, phospholipids had lower heritabilities than the sphingolipid or glycerolipid classes.

Genetic correlations among lipid species

We next conducted a series of analyses to characterize the complex genetic correlations among the plasma lipid species. First, we estimated pair-wise genetic correlation coefficients for all pairs of the plasma lipidome (total 50,721 pairs) using bivariate trait analyses. Since presence of obesity, type 2 diabetes, metabolic syndrome and hypertension can influence the plasma levels of lipid species, we adjusted the genetic correlations between species by including these clinical states as covariates. We also adjusted for the receipt of lipid-lowering and anti-hypertensive drugs. The resulting genetic correlation matrix is provided fully in Supplementary Table 2 and shown pictorially in Figure 2. We observed that a total of 23,477 (46.3% of lipid species pairs) genetic correlations were statistically significant at a nominal p-value of 0.05 and 3,492 (6.9%) were significant at an FDR-corrected p-value of 0.05. Figure 2 shows that there was generally a high positive genetic correlation between lipid species of the same class (concentration of blue squares along the diagonals).

Figure 2.

Heatmap depicting genetic correlation coefficient between each pair of lipid species. Lipid classes are separated by dark colored boxes. Lipid classes are abbreviated as shown in Table 2.

Second, we used unsupervised hierarchical clustering method to reduce the complex genetic correlation matrix into genetically meaningful clusters. Using eight criteria for clustering validation, we chose a solution that yielded 12 non-overlapping clusters as shown in Figure 3A and Supplementary Table 3. The relative performance of various clustering indices is shown in Supplementary Figure 1. To test whether the clusters were genetically coherent, we compared the average genetic correlation of lipid species pairs that belonged to the same cluster to those that belonged to different clusters. We observed (Figure 3B) that the median absolute genetic correlation between lipid species of the same cluster was twice that of the lipid species belonging to different clusters (0.52 versus 0.26) – a difference that was highly significant. The average genetic correlation within each specific cluster was also high (Figure 3C).

Figure 3.

Hierarchical clustering of lipid species based on the pair-wise genetic correlations. (A) Results of hierarchical clustering. Twelve genetic clusters with deep separation are identified as color-coded branches. The numbers on the right correspond to the reordered lipid species identifier (see Supplementary Table 3). (B and C) Efficiency of the clustering algorithm as examined using average inter- and intra-cluster genetic correlation. Panel B compares the overall average genetic correlation between (box labeled as B) and within (box labeled as W) the genetic clusters. z and p refer to the results from Mann-Whitney U test. Panel C shows the average (± standard deviation) genetic correlation within each genetic cluster. (D) Agreement between chemically defined lipid classes and genetically derived clusters. The columns are color-coded in accordance with the colors shown in Panels A and C. Grey background cells indicate statistically significant over-representation of a lipid class within a cluster. Results of Fisher's exact tests used to assess significant over-representation are shown in Supplementary Table 4.

A cross-tabulation based on lipid classes and genetic clusters (Figure 3D) indicated that none of the 23 lipid classes and subclasses completely corresponded to any single genetic cluster. Thus, we reasoned that the genetic clusters might contain information that is different from that contained in lipid classes and subclasses. Statistical test for enrichment of lipid classes within a cluster showed that of the 66 observed non-zero cells in Figure 3D, 17 were statistically significant using Fisher's exact test for over-representation (detailed results are in Supplementary Table 4). Of interest, clusters 1 and 10 were enriched for some species of the phosphatidylcholine and lysophosphatidylcholine classes; clusters 5 and 7 were enriched for some dihydroceramides and ceramides while clusters 11 and 12 were enriched for diacylglycerols and triacylglycerols, respectively.

Association of genetic clusters with clinical states

We assessed whether or not the pleiotropically related lipid species can provide clinically meaningful information. We found (Figure 4) interesting patterns of association. First, the strongest associations were found with raised serum triglycerides. Second, cluster 10 scores were consistently associated with a significantly reduced likelihood of central obesity, obesity, raised triglycerides, low HDL- C and metabolic syndrome. This cluster, as can be seen from Figure 3D, is mainly comprised of lysophosphatidylcholine species. Third, cluster 7 scores were associated with an increased likelihood of central obesity, raised serum triglycerides and type 2 diabetes. It is noteworthy that certain dihyroceramides, ceramides and cholesteryl esters were the most common lipid classes that defined cluster 7. Further, clusters 6 and 7 were consistently associated with increased likelihood of most of the clinical states. Fourth, cluster 5 scores were significantly associated with an increased likelihood of central obesity, obesity and metabolic syndrome. This cluster contained only five lipid species – 3 dihydroceramide and 2 sphingomyelin species. Fifth, cluster 9 scores (a cluster predominantly defined by phosphatidylcholine species containing linoleic acid) were specifically associated with an increased likelihood of low HDL cholesterol. Sixth, cluster 1 scores (comprised primarily of lipids containing polyunsaturated fatty acids) demonstrated remarkable specificity of association – significantly increased likelihood of obesity but a significantly reduced likelihood of type 2 diabetes.

Figure 4.

Association of genetic clusters with clinical conditions. For each indicated trait, the results are from a single polygenic regression model that was adjusted for age, age², sex, age × sex interaction, age² × sex interaction and receipt of anti-hypertensive and lipid-lowering drugs. Average inverse-normalized scores for lipid species belonging to each genetic cluster were then included as covariates. The plots report the polygenic regression coefficients. The squares and error bars represent the point and 95% confidence intervals. The squares are color-coded as follows: red – significantly increased liability of the trait, gray – statistically not significant, and blue – significantly decreased liability of the trait. The significance values are indicated at the top of each plot.

Finally, when we examined the association between genetic cluster scores and risk of cardiovascular deaths during follow-up, we found that cluster 1 scores were associated with a significantly reduced risk while cluster 6 scores were associated with a significantly increased risk of cardiovascular deaths. This cluster is dominated by sphingomyelins and both the ether-linked and plasmalogen derivatives of phosphatidylethanolamines. This pattern of association was similar to that with type 2 diabetes. Of interest, type 2 diabetes (polygenic regression coefficient = 0.56, p = 0.0002) was a highly significant predictor of cardiovascular deaths in the SAFHS participants (data not shown).

Discussion

We report three novel and important findings. First, all the plasma lipid species were significantly heritable – a finding that strongly places the plasma lipidome in a genetic context. To our knowledge, such a finding has not been reported previously and implies that a significant proportion of the inter-individual variability in plasma levels of bioactive lipids may be determined by polygenes. Considering that CVD and its risk factors are themselves genetically determined³⁴, our findings raise the possibility that there may be a genetic concordance between the plasma lipid species and clinical states predisposing to CVD in Mexican Americans.

Second, our results of hierarchical clustering suggest that there may be a substantial promiscuity in the genetic control of plasma lipid species. Moreover the fact that the genetically derived clusters were conceptually different from the chemically defined lipid classes indicated a significant pleiotropy in the genetic control of the plasma lipidome that may not be fully explained by the chemical structure of the lipid class or subclass. For example, chemically very similar phosphatidylcholine species (PC 38:6a and PC 38:6b) had distinct cluster memberships (to clusters 1 and 11, respectively; Supplementary Table 5). This divergence appears justifiable on the basis that the genetic correlation between these two species was both negative and modest (-0.45, Supplementary Table 2). Also, the fatty acid composition of these two species is different (primary constituent of PC 38:6a is PC 16:0/22:6 while that for PC 38:6b is PC 18:2/20:4). Interestingly, several other members of cluster 1 contain the omega-3 fatty acid, docosahexaenoic acid (DHA, C22:6) although some also contain the omega-6 fatty acid, arachidonic acid (AA, C20:4). This is further supported by the distribution of diacylglycerol and triacylglycerol species among clusters 11 and 12; the species in cluster 12 contain primarily saturated and monounsaturated fatty acids while cluster 11 contains a high proportion of polyunsaturated species particularly reflecting the omega-6 fatty acid, linoleic acid (C18:2) in the triacylglycerols (Supplementary Table 5). These results raise the possibility that some of the genetic control of the plasma lipids may reside at the level of fatty acid metabolism. Future studies are needed to evaluate this possibility.

Third, the genetically derived clusters of plasma lipid species were not only associated with the risk factors of CVD but also with the prospectively measured, hard outcome of cardiovascular deaths confirmed from death certificates. In this regard, it is conceivable that cluster 1 may reflect an association with DHA metabolism which has been implicated in CVD pathogenesis^35-37 while cluster 6 may relate to the known association of sphingomyelins with CVD³⁸. It is noteworthy that recent studies have demonstrated strong associations between the plasma lipidome and cardiovascular mortality^{39, 40} but our findings show the importance of genetics in characterization of the plasma lipidome and the relevance of this genetic information in potential risk-stratification of CVD.

When interpreting the association results, it should be remembered that the clusters were derived based on shared genetic influences and therefore are inherently designed to detect associations with clinical states that have a high likelihood of a strong genetic basis. For example, in a large study from two cohorts we recently demonstrated that type 2 diabetes is significantly associated with dihydroceramides, ceramides and cholesterol ester species.¹⁵ The fact that in this study cluster 7 was significantly associated with type 2 diabetes points to a possible genetic basis for these associations. Concordantly, we had observed that four dihydroceramide species (18:0, 20:0, 22:0, and 24:1) were significantly associated with waist circumference in Mexican Americans¹³ and here we find that clusters 5 and 7 (which contain these four lipid species) showed significant association with central obesity again indicating a possible genetic basis to the previously reported associations. Finally, our observation about the consistent negative association of cluster 10 with several risk factors of CVD indicates a likely genetic explanation for the reported association of reduced levels of plasma lysophsophatidylcholine with obesity and type 2 diabetes.⁴¹

Some limitations of the present study need to be considered. First, our findings should be seen as indicative and need confirmation by replications across cohorts. Demonstration of heritability and genetic correlation is a useful initial step in the quest to uncover genetic underpinnings of CVD. Deciphering the inheritance patterns and underlying molecular composition of risk traits influencing common, yet genetically complex, diseases remains largely unachieved. A logical future research direction is to identify specific sequence variants and other epigenetic control mechanisms regulating the plasma lipidome. Second, with the exception of the prospective component of cardiovascular deaths, all other inferences in this study are based on cross-sectional data. The associations therefore do not automatically imply a causal role of plasma lipid species in the pathogenesis of CVD. Rather, the main goal of the study was to query the existence of potential correlations in the, as yet latent, genetic regulators of CVD. That we obtained a consistent pattern of associations with the prospectively monitored cardiovascular deaths furthers the likelihood that the genetic clusters derived in the study are clinically meaningful. Third, since the study participants are all Mexican Americans, it is not possible to generalize these results to other ethnic groups. Future studies need to investigate the similarities and differences of the genetic clusters on the background of differing ethnicity. Fourth, lipid concentrations were estimated in stored samples and, in theory, storage can influence the estimated concentrations. However, it has been demonstrated⁴² that quantitative lipidomic techniques such as the ones used in this study are unlikely to be affected by storage even if the samples were exposed to multiple freeze/thaw cycles. Of note, the samples used in this study did not undergo any freeze/thaw cycles prior to analysis. Thus, the results presented here are unlikely to have been influenced by storage of samples. Fifth, it is conceivable that the variability of the plasma lipid species across individuals is only partly explained by genetics. Environmental factors such as shared households, dietary profiles and lifestyle factors can all contribute to both lipidomic variability and CVD risk. Our study did not evaluate these aspects but future studies need to dissect out these additional contributors and likely confounders of CVD. Lastly, the choice of number of clusters can influence the strength of association of the clusters with clinical states. Therefore the results shown here should be considered indicative of patterns rather be considered conclusive about a genetic structure. Alternative definitions and strategies for cluster identification can be envisioned and need to be evaluated.

Cardiovascular disease management currently accounts for 17% of the national health expenditure in the United States and the costs associated with CVD have been projected to triple by 2030.⁴³ Therefore prevention of CVD is a feasible and economically viable alternative to treatment programs. To that end, novel insights into biological mechanisms that predispose individuals to CVD hold the promise of potential new therapies and significant reduction of this considerable economic burden. Modern genomic technologies can be exploited to rapidly identify genes involved in disease susceptibility. However, the cost-effectiveness of such exploratory endeavors can be greatly augmented if genetic basis to a phenotype is strongly suspected. That was the motivation for the present study. Identifying novel lipid-related endophenotypes that are genetically correlated with CVD offers the potential to discover biomarkers which will quickly lead us to causal genes. Since the axis of management is now heavily tilted in favor of personalized medicine, identification of genetic predilection to CVD is important. Our study represents a preliminary step in that direction. Specifically, our results urge that the plasma lipidome needs to be carefully examined in future studies as a potential harbinger of CVD risk due to shared genetic influence and concurrence with the risk factors of CVD.