Integrative Metabolomics Differentiate Coronary Artery Disease, Peripheral Artery Disease, and Venous Thromboembolism Risks

Abstract

BACKGROUND:

Arterial and venous cardiovascular conditions, such as coronary artery disease (CAD), peripheral artery disease (PAD), and venous thromboembolism (VTE), are genetically correlated. Interrogating underlying mechanisms may shed light on disease mechanisms. In this study, we aimed to identify (1) epidemiological and (2) causal, genetic relationships between metabolites and CAD, PAD, and VTE.

METHODS:

We used metabolomic data from 95 402 individuals in the UK Biobank, excluding individuals with prevalent cardiovascular disease. Cox proportional-hazards models estimated the associations of 249 metabolites with incident disease. Bidirectional 2-sample Mendelian randomization (MR) estimated the causal effects between metabolites and outcomes using genome-wide association summary statistics for metabolites (n=118 466 from the UK Biobank), CAD (n=184 305 from CARDIoGRAMplusC4D 2015), PAD (n=243 060 from the Million Veterans Project), and VTE (n=650 119 from the Million Veterans Project). Multivariable MR was performed in subsequent analyses.

RESULTS:

We found that 196, 115, and 74 metabolites were associated (P<0.001) with CAD, PAD, and VTE, respectively. Further interrogation of these metabolites with MR revealed 94, 34, and 9 metabolites with potentially causal effects on CAD, PAD, and VTE, respectively. There were 21 metabolites common to CAD and PAD and 4 common to PAD and VTE. Many putatively causal metabolites included lipoprotein traits with heterogeneity across different sizes and lipid subfractions. Small VLDL (very-low-density lipoprotein) particles increased the risk for CAD while large VLDL particles decreased the risk for VTE. We identified opposing directions of CAD and PAD effects for cholesterol and triglyceride concentrations within HDLs (high-density lipoproteins). Subsequent sensitivity analyses including multivariable MR revealed several metabolites with robust, potentially causal effects of VLDL particles on CAD.

CONCLUSIONS:

While common vascular conditions are associated with overlapping metabolomic profiles, MR prioritized the role of specific lipoprotein species for potential pharmacological targets to maximize benefits in both arterial and venous beds.

Highlights.

• Across 249 metabolites in 95 402 free from cardiovascular disease followed for a median of 12.1 years, 196, 115, and 74 metabolites were associated with incident CAD, PAD, and VTE, respectively.

• Using Mendelian randomization, we observed that 78 metabolites increased and 16 decreased risk for CAD, 15 increased and 19 decreased risk for PAD, and 2 increased and 7 decreased risk for VTE. While extensive concordant effects were observed for CAD and PAD, such genetic sharing (despite biomarker correlations) was not observed for VTE with either CAD or PAD. These observations indicate distinct metabolomic effects on atherosclerosis versus thrombosis.

• Smaller VLDL particles increased the risk for CAD while larger VLDL particles decreased the risk for VTE, providing evidence that pharmacological modulation of lipoprotein particles requires careful consideration of particle size and vascular location to maximize risk-benefit balance in both arterial and venous beds.

• HDL particles demonstrated complex heterogeneity, in which there were opposing effects between cholesterol subfractions as protective and triglyceride subfractions as harmful.

• Despite extensive correlation of metabolites, the aforementioned observations were robust upon further interrogation via sensitivity analyses for pleiotropy and multiphenotype modeling.

Cardiovascular diseases, including coronary artery disease (CAD), peripheral artery disease (PAD), and venous thromboembolism (VTE), remain the leading causes of morbidity and mortality worldwide.¹ These 3 common vascular diseases often comorbid and share many clinical risk factors. Indeed, upward of 42% of individuals with CAD also have PAD.² In the past decade, genome-wide association studies (GWAS) of CAD, PAD, and VTE have probed the genetic architectures of and between these 3 outcomes.^3–6 In addition to such genetic studies, assessment of the human metabolome is becoming an attractive approach to study mechanistic pathogenesis, as metabolomics capture exposures from several sources, including genetic, dietary, and lifestyle profiles, that all influence an individual’s lifetime risk of various cardiovascular diseases.⁷ Combining both genetic and metabolomic data, leveraging recent GWAS performed on large-scale metabolomics data⁸ may help to probe the role of the metabolome in explaining the genetic commonalities and distinctions between CAD, PAD, and VTE.⁸ Specifically, identifying upstream metabolites that affect an individual’s risk of CAD, PAD, or VTE and vice versa (ie, identifying metabolites that are affected by disease progression of CAD, PAD, or VTE) may uncover novel mechanistic insights.

Mendelian randomization (MR) can estimate causal effects between metabolites and CAD, PAD, and VTE. MR is a causal inference approach that uses genetic instruments as instrumental variables to prioritize potentially causal relationships between exposures and outcomes, with corresponding lowering risks of confounding and reverse causation.⁹ MR is particularly well-suited for estimating the causal effects of metabolites, especially given their strong heritabilities¹⁰ and recently available in-depth metabolomic profiling in the UK Biobank.^11,12 This profiling spans common amino acids, inflammatory markers, and lipoproteins, including specific subfractions of lipids within diverse lipoproteins (eg, VLDL [very-low-density lipoprotein], IDL [intermediate-density lipoprotein], LDL [low-density lipoprotein], and HDL [high-density lipoprotein]), such as the cholesterol and triglyceride content. Given that previous work has largely focused on standard laboratory measurements, such as the cholesterol content of HDL and LDL, interrogation of more comprehensive and specific biomarkers, such as lipoprotein subfractions, is critical.

In this study, we aimed to (1) identify and compare associations of metabolites with CAD, PAD, and VTE and (2) prioritize and compare causal effects of metabolites with CAD, PAD, and VTE. Specifically, we performed association analyses and bidirectional 2-sample MR analyses in the UK Biobank with 95 402 individuals with metabolomics and nonoverlapping genetic data for CAD, PAD, and VTE from large-scale GWAS consortia to infer novel unbiased mechanistic insights.

MATERIALS AND METHODS

Data Availability

All data supporting the findings of this study are available in the article, in the Supplemental Material, or from the corresponding author upon reasonable request.

Study Cohort

The UK Biobank study cohort has been described previously.¹² Briefly, the UK Biobank is a large population-based prospective study that contains genotype and phenotype data from 502 639 individuals living in the United Kingdom and recruited between 2006 and 2010.¹² Individuals who were closely related, withdrew consent, or had prevalent CAD, PAD, or VTE at baseline were excluded. CAD, PAD, and VTE were defined by the International Classification of Diseases system, specifically Tenth Revision codes (Table S1). In CARDIoGRAMplusC4D 2015, CAD cases were similarly defined by an inclusive CAD diagnosis (eg, myocardial infarction, acute coronary syndrome, chronic stable angina, or coronary stenosis >50%).¹³ In the Million Veterans Program, which consisted of individuals of European, African, and Hispanic ancestry, PAD and VTE were similarly defined by International Classification of Diseases, Ninth and Tenth Revision, and Current Procedural Terminology codes (Table S1).^5,6

Metabolomic Profiling

A random subset of nonfasting baseline plasma samples from 118 466 individuals and 1298 repeat visit samples were measured for 249 metabolomic biomarkers using a high-throughput nuclear magnetic resonance metabolomic biomarker profiling platform developed by Nightingale Health (Helsinki, Finland) in the UK Biobank.⁸ The metabolites span multiple metabolic pathways, including lipoprotein lipids in 14 subclasses, fatty acids and fatty acid compositions, as well as various low-molecular-weight metabolites, such as amino acids, ketone bodies, and glycolysis metabolites, all quantified in molar concentration units.⁸ A total of 95 402 individuals with complete, nonmissing metabolomic profiling were included in this study.

Genetic Correlation Between Cardiovascular Phenotypes

After variant filtering (eg, information score, metric of imputation quality, observed variance of imputed genotype/expected variance under perfect imputation >0.9 and minor allele frequency >0.01), linkage disequilibrium score regression was used to calculate the genetic correlation between CAD and PAD, CAD and VTE, and PAD and VTE using HapMap3 single-nucleotide polymorphisms in the 1000 Genomes Project European reference panel.¹⁴ Summary statistics for CAD were obtained from CARDIoGRAMplusC4D 2015 that did not include any individuals from the UK Biobank,¹³ and summary statistics for PAD and VTE were obtained from the Million Veterans Program.^5,6

Association Analysis of Metabolites With Cardiovascular Phenotypes

To estimate epidemiological associations of 249 metabolites with CAD, PAD, and VTE, we constructed Cox proportional-hazards models regressing metabolite measurements on time to event for CAD, PAD, and VTE, with covariates including age, sex, genotyping array, 5 genetic principal components, and self-reported statin use. A correction factor of 0.5 multiplied by the minimum nonzero value of each metabolite was applied before log transformation of metabolite values. Transformed metabolite values were scaled with a mean of 0 and an SD of 1. Rather than using Bonferroni correction for multiple hypothesis testing (which may substantially mitigate power given the high degree of correlation between the 249 metabolites), we adjusted for the number of eigenvectors (n=41) that explain 99% of the variance in the data (eg, α=0.05/41=0.001), which was consistent with prior studies to minimize both type I and type II errors. As such, α was defined as 0.05/41=0.001. We also reported metabolites that remained significant at a Bonferroni-corrected P value threshold of 0.05/249=0.0002.

MR of Metabolites and Cardiovascular Phenotypes

MR is a method that uses genetic variants as instrumental variables to prioritize causal relationships between exposures and outcomes.⁹ MR relies on several assumptions: (1) genetic instruments must be robustly associated with the exposure, (2) genetic instruments must not be associated with confounders, and (3) genetic instruments must not influence the outcome except through the exposure.⁹ Two-sample MR utilizes GWAS summary statistics in nonoverlapping cohorts for exposures and outcomes.¹⁵ We used 2 nonoverlapping cohorts for metabolites and cardiovascular outcomes. Summary statistics for metabolites were obtained from publicly available GWAS summary statistics from the OpenGWAS platform (https://gwas.mrcieu.ac.uk/), which contained 249 metabolites from the UK Biobank measured by Nightingale Health (Table S2).¹⁶ Summary statistics were not adjusted for statin use. Summary statistics for CAD were obtained from CARDIoGRAMplusC4D 2015,¹³ and summary statistics for PAD and VTE were obtained from the Million Veteran Program.^5,6

We performed bidirectional, 2-sample MR using the TwoSampleMR^17,18 R package to estimate the causal effects between metabolites and CAD, PAD, and VTE. All analyses were consistent with current recommendations for MR.¹⁹ We performed linkage disequilibrium clumping with a window >10 000 kb and an R² >0.001. Genetic instruments were defined to be genetic variants that were significant at the genome-wide significance level after linkage disequilibrium clumping (P<5×10⁻⁸). Each instrumental variable constructed for the metabolites consisted of at least 7 single-nucleotide polymorphisms and at most 112 single-nucleotide polymorphisms (SD=19.9; median number of single-nucleotide polymorphisms, 66), and instrumental variables were robust, with all F statistics >50. We utilized the inverse-variance weighted method, which regresses associations between genetic instruments and outcomes upon associations between genetic instruments and exposures.²⁰

In the primary analysis, in which metabolites were exposures and CAD, PAD, and VTE were outcomes, genetic instruments were clumped and selected based on the above criteria (eg, P<5×10⁻⁸ for metabolite association from GWAS). In the secondary analysis, in which metabolites were outcomes and CAD, PAD, and VTE were exposures, genetic instruments were clumped and selected based on the above criteria (eg, P<5×10⁻⁸ for CAD, PAD, or VTE association from GWAS). To estimate the correlations of causal effects between CAD, PAD, and VTE, we calculated the Pearson correlation coefficients of βs from MR as a measure of metabolomic overlap.

MR Sensitivity Analyses Adjusting for Horizontal Pleiotropy and Correlation

In subsequent sensitivity analyses, we utilized the MR Egger, weighted median, simple mode, and weighted mode methods as well, as such methods may be more robust to pleiotropy. We also conducted tests for pleiotropy and heterogeneity. A test for pleiotropy is calculated using the MR Egger method, as the slope coefficient estimates the causal effect that is consistently asymptotically even if individual genetic variants have pleiotropic effects.²¹ A test for heterogeneity assesses the compatibility of instrumental variable estimates based on individual genetic variants and is calculated using the Cochran Q test on the causal estimates from individual genetic variants through the MR-PRESSO (Mendelian Randomization Pleiotrophy Residual Sum and Outlier) method.²¹ We considered a strict set of requirements to prioritize an exposure as a potential causal metabolite: (1) significant epidemiological association via Cox proportional-hazards model (P<0.001), (2) same direction of effect between Cox proportional-hazards model and 2-sample MR using the inverse-variance weighted and MR Egger methods, (3) significant causal effect via 2-sample MR using the inverse-variance weighted method (P<0.001), (4) no evidence of horizontal pleiotropy via 2-sample MR using the MR Egger intercept method (P>0.001), and (5) no evidence of heterogeneity via 2-sample MR using the MR-PRESSO method (outlier-corrected P value <0.001).

Given the high degree of correlation among metabolites, we also performed sensitivity analyses using multivariable MR (MVMR), which uses genetic variants for ≥2 exposures to simultaneously estimate the causal effect of each exposure on the outcome, controlling for the effect of other included exposures. MVMR requires the same assumptions as univariable MR, but the genetic instruments must be associated with the set of exposures rather than the single exposure, but it is not necessary for each genetic instrument to be significantly associated with every exposure.²² In other words, instruments are selected for each exposure, and then all exposures for such instruments are regressed against the outcome together, weighting for the inverse variance of the outcome. MVMR was performed for all metabolites that were identified to have causal effects in the primary analysis, such that each model for an index metabolite for either CAD, PAD, or VTE included all metabolites that were epidemiologically correlated (P>0.70 and P<0.001) with the index metabolite and also identified to have causal effects in the primary analysis.

All analyses were performed using R, version 4.3.0, and all plots were generated using the R package ggplot2. A heatmap was generated using the R package gplots.

RESULTS

We estimated associations of 249 metabolites with CAD, PAD, and VTE in 95 402 individuals in the UK Biobank, which included 3209 CAD cases, 846 PAD cases, and 1474 VTE cases (Figure 1). The study cohort of 95 402 comprised 52 065 (55%) women with a mean age of 56.3 (SD=8.0) years, and 13 204 (14%) individuals prescribed statins (Table). The GWAS summary statistics for CAD included 60 801 cases and 123 504 controls across 48 nonoverlapping studies¹³; for PAD included 31 307 cases and 211 753 controls from the Million Veterans Program⁵; and for VTE included 26 066 cases and 624 053 controls from the Million Veterans Program.⁶ We calculated the genetic correlations across the 3 sets of summary statistics. CAD and PAD shared the highest genetic correlation (genetic correction, 0.6615; SE=0.044; P=4.51×10⁻⁵¹), followed by PAD and VTE (genetic correction, 0.2463; SE=0.053; P=3.28×10⁻⁶) and CAD and VTE (genetic correction, 0.1300; SE=0.0480; P=6.80×10⁻³).

Figure 1.

Study schematic. A, Epidemiological association analysis was performed using individual-level data from the UK Biobank, regressing incident coronary artery disease (CAD)/peripheral artery disease (PAD)/venous thromboembolism (VTE) events on metabolite levels. B, Bidirectional Mendelian randomization analysis was performed using metabolite summary statistics from the UK Biobank as exposures and CAD/PAD/VTE summary statistics from either CARDIoGRAMplusC4D 2015 or Million Veterans Program (MVP) as outcomes and (C) vice versa. D, Taking only the metabolite with significant causal effects from B, multivariable Mendelian randomization analysis was performed using metabolite summary statistics from the UK Biobank as exposures and CAD/PAD/VTE summary statistics from either CARDIoGRAMplusC4D 2015 or MVP as outcomes. GWAS indicates genome-wide association study.

Table.

Baseline Characteristics of Analyzed UK Biobank Participants

Epidemiological Associations Reveal Overlap of Metabolites Between Cardiovascular Outcomes

We found 196 (P<0.001; 184 with P<0.0002 per Bonferroni-corrected P value threshold), 115 (P<0.001; 92 with P<0.0002), and 74 (P<0.001; 54 with P<0.0002) metabolites associated with CAD, PAD, and VTE, respectively (Figure S1; Table S3). We estimated the overlap of metabolomic associations across CAD, PAD, and VTE. There were 52 metabolites that had significant (P<0.001; 38 with P<0.0002 per Bonferroni-corrected P value threshold) associations with all 3 cardiovascular phenotypes, of which 49 metabolites had the same direction of effect and 3 metabolites had the opposite direction of effect of CAD compared with PAD and VTE. HDL-related traits, despite making up a small fraction of the metabolites that were assessed (56/249 metabolites or 22.5%), made up nearly half of the 49 metabolites that had the same direction of effect (22/49 metabolites or 44.9%). We then compared the overlap in metabolites between CAD, PAD, and VTE (Figure S2). CAD and PAD shared 30 concordant metabolites (ie, with the same direction of effect), CAD and VTE shared 16 concordant metabolites, and PAD and VTE shared 3 concordant metabolites (Figure S3). The metabolites shared by CAD and PAD covered all 3 lipoprotein classes, from HDL, LDL, and VLDL, as well as some nonlipoprotein particles like ApoA1. All metabolites shared by CAD and VTE were larger (eg, medium-sized to XXL-sized particles) lipoprotein traits covering HDL and VLDL but not LDL. None of the metabolites shared by PAD and VTE were lipoprotein traits.

Estimation of Genetic Effects of Metabolites on Cardiovascular Outcomes Reveals Overlap Between CAD and PAD

Then, we used bidirectional MR to estimate the potential causal relationships between metabolites and CAD, PAD, and VTE (Table S4). Overall, we found that many of the metabolites that demonstrated strong epidemiological associations did not maintain genetic effects upon interrogation via MR, particularly nonlipid traits, like glucose and amino acids, as well as inflammatory biomarkers, like GlycA. However, we found several metabolites, particularly lipoprotein traits, that were considered to have robust evidence for causal effects passed the strict set of requirements, including (1) significant epidemiological association via Cox proportional-hazards model (P<0.001), (2) same direction of effect between Cox proportional-hazards model and 2-sample MR using the inverse-variance weighted method and MR Egger method, (3) significant causal effect via 2-sample MR using the inverse-variance weighted method (P<0.001), (4) no evidence of horizontal pleiotropy via 2-sample MR using the MR Egger intercept method (P>0.001; Table S5), and (5) no evidence of heterogeneity via 2-sample MR using the MR-PRESSO method (outlier-corrected P<0.001; Table S6). We found 78 metabolites (74 metabolites at Bonferroni-corrected P value threshold of P<0.0002) that increased and 16 (13 at P<0.0002) that decreased the risk for CAD, 15 (8 at P<0.0002) metabolites that increased and 19 (17 at P<0.0002) that decreased the risk for PAD, and 2 (0 at P<0.0002) metabolites that increased and 7 (4 at P<0.0002) that decreased risk for VTE (P<0.001; Figure S4). In the opposite direction, we found 53 metabolites (44 at P<0.0002) that were decreased by CAD itself, no metabolites that were changed by PAD, and 2 (1 at P<0.0002) metabolites that were increased by VTE itself (Figure S5).

We then compared the overlap in metabolites between CAD, PAD, and VTE (Figure S6). The greatest genetic overlap in metabolites was between CAD and PAD (R²=0.72; P=5.80×10⁻⁷⁰), followed by CAD and VTE (R²=0.18; P=4.34×10⁻¹²) and PAD and VTE (R²=0.17; P=1.29×10⁻¹¹), reflecting the genetic correlation from GWAS. MR revealed 8 metabolites, largely LDL traits, that increased the risk for both CAD and PAD and 13 metabolites, largely HDL traits that decreased the risk for both CAD and PAD (Figure 2). Many of the metabolites that changed the risk for VTE had opposite directions of effect for CAD or PAD.

Figure 2.

Venn diagram of metabolites with shared causal effects between coronary artery disease (CAD), peripheral artery disease (PAD), and venous thromboembolism (VTE) estimated using bidirectional Mendelian randomization. Bidirectional, inverse-variance weighted 2-sample Mendelian randomization using summary statistics from the UK Biobank and CARDIoGRAMplusC4D 2015 or Million Veterans Program was performed to estimate causal effects of metabolites on CAD, PAD, and VTE (P<0.001). Metabolites that had causal effects on >1 cardiovascular phenotype but demonstrated an opposite direction of effect were indicated with white asterisks. Metabolites that maintained significant associations after applying a Bonferroni-corrected P value threshold (P<0.0002) are bolded.

Estimation of Genetic Effects of Lipoprotein Particles Reveals Heterogeneity

Given the strong genetic effects demonstrated by lipoprotein particles (Figure 3), we then closely examined the components of LDL, VLDL, and HDL (eg, total cholesterol, lipid subfraction, triglyceride subfraction) for all available sizes of lipoprotein particles. LDL traits, regardless of the component or size of the particle, demonstrated strong genetic effects associated with increased risk for cardiovascular outcomes, particularly CAD (Figure S7). The other lipoprotein particles demonstrated more heterogeneity. VLDL traits, especially smaller particles, demonstrated strong genetic effects associated with increased risk for arterial vascular diseases, particularly CAD (Figure S8). On the contrary, VLDL traits, especially larger particles, demonstrated strong genetic effects associated with decreased risk for VTE, a venous vascular disease. The smaller VLDL particles had null genetic effects with VTE. In other words, we found that while the cholesterol content of larger VLDL particles increased the risk for CAD and PAD (without risk for VTE), the cholesterol content of smaller VLDL particles decreased the risk for VTE (with modest risk for CAD and PAD). HDL traits interestingly demonstrated protective effects for certain cholesterol and lipid traits but atherogenic effects for triglyceride traits for CAD (Figure S9).

Figure 3.

Forest plots of estimated epidemiological associations and causal effects of lipoprotein particles and cardiovascular outcomes. Points are effect estimates per SD change of metabolites, and error bars represent 95% CIs. Transparent bars represent associations that were not significant based on a multiple testing correction of P<0.001. C indicates cholesterol content; CAD, coronary artery disease; HDL, high-density lipoprotein; L, lipid content; LDL, low-density lipoprotein; PAD, peripheral artery disease; TG, triglyceride content; VLDL, very-low-density lipoprotein; and VTE, venous thromboembolism.

Given the high degree of correlation of metabolites among each other, we used MVMR to identify metabolites that had causal effects on CAD, PAD, and VTE independent of other correlated metabolites. After identifying correlations between metabolites (Table S7), we constructed models for the 94, 34, and 9 metabolites with causal effects with CAD, PAD, and VTE, respectively, with their epidemiologically correlated metabolites (Table S8). For CAD, we identified 4 metabolites with robust, independent effects: phospholipid/total lipid ratio in very large VLDL (effect size [β], 0.57; P=7.46×10⁻⁶), cholesteryl ester/total lipid ratio in small VLDL (β=0.60; P=3.46×10⁻⁶), phospholipid/total lipid ratio in small VLDL (β=8.73; P=3.02×10⁻⁵), and triglycerides in very large HDL (β=1.73; P=3.79×10⁻⁵). For PAD, we identified 2 metabolites with robust, independent effects: triglyceride content in large LDL (β=1.71; P=1.01×10⁻³; adjusted for 8 other correlated metabolites) and triglyceride content in very small VLDL (β=4.02, P=1.61×10⁻⁴; adjusted for 8 other correlated metabolites). The only metabolite that maintained a robust, marginally significant P value (P<0.05) on VTE after MVMR was the triglyceride/total lipid ratio in small HDL (β=−1.91; P=4.49×10⁻²).

DISCUSSION

Here, we performed hypothesis-free epidemiological and genetic association analyses to evaluate the metabolomic profiles of CAD, PAD, and VTE, covering 249 metabolites in 95 402 individuals in the UK Biobank. This allowed us to compare metabolomic profiles among these 3 cardiovascular phenotypes to provide insight into potential similarities and differences in their metabolomic pathogenesis. These observations have important implications for the prevention and treatment of common vascular conditions.

First, we found an overlap in the causal metabolic profiles of CAD and PAD, but not VTE, via MR analyses, consistent with their strong, shared genetic correlation. This is consistent with the high co-occurrence of CAD and PAD, as well as the genetic correlation, reported by previous studies.^2–6 We identified several upstream metabolites that affected an individual’s risk of CAD (n=94), PAD (n=34), and VTE (n=9). Specifically, 21 metabolites had common causal effects for CAD and PAD, 4 had common causal effects for PAD and VTE, and 2 had common causal effects for CAD and VTE. Overall, lipid traits may not be as strongly implicated in thrombus formation compared with plaque formation, as VTE had few causal metabolites compared with CAD and PAD. Previous studies have highlighted carnitine species, glucose, phenylalanine, among other metabolites, but not lipid-related metabolites, as possibly related to VTE.²³ We similarly found that several epidemiological associations identified via association analyses, such as several amino acids, did not maintain robust causal effects upon interrogation via MR.

Second, we found that small VLDL particles may be particularly deleterious for CAD, while large VLDL particles may be protective for VTE. Indeed, VLDL pathways have been implicated in both CAD²⁴ and PAD.^25,26 VLDL is a precursor to LDL and is considered to be a marker of atherogenic lipoprotein remnants.²⁷ The complex relationship between CAD and VTE for VLDL highlights important considerations for VLDL therapeutic modulation to reduce the risk for CAD. VLDL size may be influenced by a variety of factors, including diet and metabolism. For example, low-carbohydrate and high-fat diets were associated with smaller VLDL particles whereas high-carbohydrate and low-fat diets were associated with larger VLDL particles.²⁸ In addition, certain apoB mutations resulting in smaller VLDL particles, characteristically with less triglyceride content, were associated with dyslipidemia in the background of hepatic steatosis.²⁹ Smaller VLDL particles may also be more similar in profile to LDL particles. The convergence of VLDL cholesterol concentrations between CAD and PAD metabolomic profiles strengthens a shared disease mechanism. This result supports the prioritization of atherogenic lipoprotein remnants as a high-yield therapeutic target for both CAD and PAD honing on smaller species. There has been some evidence that in the presence of VLDL particles, plasma phospholipid transfer proteins exert antithrombotic effects by inhibiting the generation of prothrombotic factors.³⁰ On the contrary, we observe a potential protective effect of VLDL particles on VTE, suggesting potential complexity in the context of the role of lipid-lowering medications on CAD and PAD versus VTE. A previous case-control study found that lipid-lowering medications decrease VLDL particles and decrease VTE.³¹ However, the mechanism through which these canonically LDL-lowering medications have an effect on VLDL particles is still being explored. A recent study found that PCSK9 (proprotein covertase subtilisin/kexin type 9) inhibitors have been found to result in increased VLDL particle size, suggesting increased clearance of smaller atherogenic VLDL remnant particles.³² This supports the hypothesis that increased larger sized VLDL particles may be reflecting decreased smaller atherogenic VLDL particles that are more similar in profile to LDL particles, potentially contributing to the protective effect of large VLDL particles for VTE described in the present study. Another study found that VLDL lowering with pemafibrate lead to increased risk for VTE,³³ suggesting that the relationship between VLDL particles and VTE, especially in the context of routinely used lipid-lowering medications, may be more complex than currently appreciated.

Third, HDL exhibits a complex relationship with CAD. Previous MR analyses³⁴ and clinical trials have converged null effects of increasing HDL to reduce the risk of CAD.³⁵ However, with more detailed profiling of HDL, we find that HDL cholesterol may be protective, whereas HDL triglycerides may be deleterious. A nested case-control study similarly showed an independent association of increased HDL triglycerides with increased myocardial infarction risk.³⁶ Among patients referred for coronary angiography, HDL apolipoproteins, influenced by cholesterol and triglyceride composition, are associated with atherosclerosis burden independent of HDL concentrations.³⁷ HDL enriched for triglycerides has been shown to have altered herniated structures.³⁸ The present study raises the hypothesis that both influence CAD in opposing manners. Therefore, strategies raising HDL, and therefore raising HDL cholesterol and HDL triglycerides, have demonstrated a net null effect. Indeed, neither niacin nor fibrates were found to favorably influence this balance in a small study.³⁸

The limitations of this study should be considered. First, the power and precision of MR are limited by the heritability and discovery GWASs, which directly influence the strength and availability of instrumental variables. Because our causal inferences were informed by both the observational, epidemiological analysis and the MR analysis, some metabolites that were prioritized in the observational analysis may have been subsequently excluded in the MR analysis due to limited power. Some truly causal metabolites may have low power in this framework due to relatively reduced heritability biasing such analyses to the null. As more powerful GWASs are performed that produce more, high-quality instrumental variables, the MR results will necessarily become stronger and more robust. Second, MR infers causality based on observational data and, therefore, provides causal estimates that increase the body of evidence in favor of causality without definitively establishing causality. As such, our findings necessitate subsequent experimental or in vivo work to confirm the evidence contributed by the causal estimates from the MR analyses. Third, the study of the human metabolome is complex. In the UK Biobank, metabolites are measured from nonfasting plasma samples, with an average of 4 hours since last meal.⁸ The metabolomic panel in the UK Biobank is highly enriched for circulating lipoproteins and related features. CAD, PAD, and VTE may have more complex metabolomic profiles, and interrogation of nonlipoprotein metabolites beyond those measured in the UK Biobank may reveal other metabolomic overlap. Fourth, this study adopted a conservative, rules-based approach with stringent requirements to prioritize robust, causal evidence, including the requirement that both epidemiological and MR analyses have statistically significant, concordant directions of effect. This approach could result in false-negative results for truly causal metabolites that may be involved in complex pathways (eg, in compensatory or feedback mechanisms that are insufficiently characterized by concordance of observational and causal estimations). This consideration, in addition to previously mentioned limitations (eg, power of MR analyses, lipoprotein-predominant profiling), suggests that metabolites with null effects should not be definitely ruled out as noncausal metabolites. Fifth, using genetically driven metabolites in MR analyses is complicated by the fact that many metabolites are heavily correlated and demonstrate pleiotropic effects.³⁹ While we conduct several stringent sensitivity analyses to address this correlation (eg, MVMR) and pleiotropy (eg, MR-PRESSO), additional modeling and instrument selection may benefit the body of evidence contributing to the causality of certain metabolites. In particular, novel extensions to MVMR, such as MR-BMA (Mendelian Randomization Bayesian Model Averaging) based on Bayesian model averaging, may strengthen the confidence of true causal effects for heavily correlated risk factors.^26,40 Lastly, the UK Biobank is predominantly composed of individuals from European ancestry, and metabolomic profiling within the cohort is heavily dominated by lipid traits. As such, the generalizability of our findings to other ancestry groups may be limited.

Taken together, we highlight the complexities of metabolomics and their influences on common vascular disorders. While LDL uniformly influences CAD and PAD, small VLDL most strongly influences CAD and PAD, but large VLDL may be protective for VTE, and new analyses suggest HDL cholesterol may be protective while HDL triglycerides may be deleterious for CAD. In retrospect, these may help explain previously reported clinical trial results and inform optimal strategies to address metabolomics-mediated vascular disease risk.

ARTICLE INFORMATION

Acknowledgments

The authors thank all of the individuals of the UK Biobank for their generous participation.

Sources of Funding

T.C. Gilliland was partially supported by the National Heart, Lung, and Blood Institute T32 grant 5T32HL125232. T. Nakao is supported by the Japan Society for the Promotion of Science Overseas Fellowship. G.M. Peloso is supported by grants from the National Heart, Lung, and Blood Institute (R01HL142711 and R01HL127564). P. Natarajan is supported by grants from the National Heart, Lung, and Blood Institute (R01HL142711, R01HL127564, R01HL148050, R01HL151283, R01HL148565, R01HL135242, and R01HL151152), the National Institute of Diabetes and Digestive and Kidney Diseases (R01DK125782), Fondation Leducq (TNE-18CVD04), and Massachusetts General Hospital (Paul and Phyllis Fireman Endowed Chair in Vascular Medicine).

Disclosures

P. Natarajan reports research grants from Allelica, Amgen, Apple, Boston Scientific, Genentech/Roche, and Novartis; personal fees from Allelica, Apple, AstraZeneca, Blackstone Life Sciences, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly & Co, Foresite Labs, Genentech/Roche, GV, HeartFlow, Magnet Biomedicine, Merck, and Novartis; scientific advisory board membership of Esperion Therapeutics, Preciseli, and TenSixteen Bio; being a scientific cofounder of TenSixteen Bio; equity in MyOme, Preciseli, and TenSixteen Bio; and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. The other authors report no conflicts.

Supplemental Material

Tables S1–S9

Figures S1–S9

Major Resources Table

Supplementary Material

atv-44-2108-s001.pdf ^{(4.3MB, pdf)}

atv-44-2108-s002.xlsx ^{(3.4MB, xlsx)}