Comprehensive Metabolomic Profiling and Incident Cardiovascular Disease: A Systematic Review
Background Metabolomics is a promising tool of cardiovascular biomarker discovery. We systematically reviewed the literature on comprehensive metabolomic profiling in association with incident cardiovascular disease (CVD).
Methods and Results We searched MEDLINE and EMBASE from inception to January 2016. Studies were eligible if they pertained to adult humans; followed an agnostic and/or comprehensive approach; used serum or plasma (not urine or other biospecimens); conducted metabolite profiling at baseline in the context of examining prospective disease; and included myocardial infarction, stroke, and/or CVD death in the CVD outcome definition. We identified 12 original articles (9 cohort and 3 nested case‐control studies); participant numbers ranged from 67 to 7256. Mass spectrometry was the predominant analytical method. The number and chemical diversity of metabolites were very heterogeneous, ranging from 31 to >10 000 features. Four studies used untargeted profiling. Different types of metabolites were associated with CVD risk: acylcarnitines, dicarboxylacylcarnitines, and several amino acids and lipid classes. Only tiny improvements in CVD prediction beyond traditional risk factors were observed using these metabolites (C index improvement ranged from 0.006 to 0.05).
Conclusions There are a limited number of longitudinal studies assessing associations between comprehensive metabolomic profiles and CVD risk. Quantitatively synthesizing the literature is challenging because of the widely varying analytical tools and the diversity of methodological and statistical approaches. Although some results are promising, more research is needed, notably standardization of metabolomic techniques and statistical approaches. Replication and combinations of novel and holistic methodological approaches would move the field toward the realization of its promise.
What Is New?
Metabolomic profiling may identify metabolites potentially useful as clinical biomarkers for risk stratification and early identification of cardiovascular disease (CVD).
This systematic review showed that only a small number of longitudinal studies have used comprehensive profiling of circulating metabolites in plasma or serum to identify an early “metabolic fingerprint” for CVD.
Currently, metabolite species associated with higher CVD risk include acylcarnitines, dicarboxylacylcarnitines, as well as several amino acids such as phenylalanine and glutamate, and several lipid classes; however, the addition of these metabolites to CVD risk prediction already employing traditional risk factors yields only small improvements in predictions.
What Are the Clinical Implications?
Current data are promising, although metabolomics approaches and results appear to be heterogeneous.
The lack of robust replications is one of the main limitations in the existing literature due to heterogeneity in study designs, definitions of end points, features of the metabolomics platforms, and small sample sizes.
Additional studies are needed to identify clinically useful metabolic fingerprints for early identification of individuals at high CVD risk.
Cardiovascular disease (CVD) continues to be a major global public health challenge. In 2013, coronary heart disease and stroke were globally the first and third leading causes of years of life lost, respectively.1 In the United States, 85 million adults currently have at least 1 type of CVD, and approximately half of them are under 60 years of age.2 Globally, population aging and growth have led to increasing numbers of CVD deaths.3 Moreover, premature cardiovascular mortality is estimated to continue at present rates or even to increase if policies to combat CVD risk factors are not successful.4 This scenario supports a strong need to improve CVD prevention.
A key factor in the fight against CVD is broadening our knowledge of the pathophysiological processes of this complex disease. Among the “omics” sciences, metabolomics has brought a paradigm shift to metabolic research. Metabolomics is the identification and quantification of small molecules that reflect the state of the organism at a particular moment in time. Currently, high‐throughput technologies allow the quantification of hundreds of circulating metabolites across multiple pathways in a single measurement. This approach is advantageous because it is not limited to a single enzymatic reaction or pathway; rather, it captures the complexity of metabolic networks. Metabolomics has considerably increased interest in metabolism across cardiovascular research.5
Large‐scale metabolomic profiling, including “metabolome‐wide” studies, may identify metabolic changes that precede irreversible organ damage and the appearance of disease and thereby may lead to the early identification of individuals at high CVD risk. For this reason the search for metabolites that could be used as clinical biomarkers is probably 1 of the most interesting aspects of metabolomics in CVD research.6 The identification of metabolomic risk profiles has the potential to improve risk stratification and early identification of CVD. In fact, metabolomics and its sister science, lipidomics, are among the newest approaches in the search for novel biomarkers.7 Single biomarkers are no longer sufficient to interpret or characterize complex biological phenomena, and new metabolomic approaches recognize the importance of characterizing the interrelation of metabolites—the metabolic “fingerprint” of disease and preclinical disease states. An inherent interest in using metabolomics in cardiovascular medicine is also driven by the hypothesis that metabolomics findings may lead to a better understanding of the pathophysiology and biological mechanisms involved in the genesis of clinical CVD events. Such an understanding would pave the way to new, evidence‐based approaches in preventing and managing CVD.
Comprehensive metabolomic profiling applied to CVD is still in its relative infancy.8, 9 Currently, there is no single approach that provides comprehensive coverage of the human metabolome, and many approaches have been used alongside a wide range of analytical platforms, each requiring specific sample preparation, approaches (eg, targeted versus untargeted), and post‐data acquisition statistical methods.10
Given the wide variety of metabolomic profiling approaches, in this systematic review, we aimed to assess and summarize existing literature on comprehensive profiling of circulating metabolites, following an agnostic or hypothesis‐free approach, and incident CVD, focusing our review on analytical methods, metabolites assessed and associated with incident CVD risk, and the predictive value of these metabolites.
The review protocol was registered in PROSPERO International Prospective Register of Systematic Reviews (crd.york.ac.uk/prospero/index.asp Identifier: CRD42015015594). This systematic review was performed according to the MOOSE (Meta‐analysis Of Observational Studies in Epidemiology) checklist11 (Table 1).
Data Sources and Search Strategies
We conducted a comprehensive search in MEDLINE (via Ovid and PubMed) and EMBASE from inception through December 2016. Our search strategy included medical subject headings and key terms related to metabolomics and CVD (Table 2). The search in EMBASE was limited to English, Catalan, Czech, French, German, Italian, Portuguese, Slovak, or Spanish, reflecting the competencies of the first authors. No language limits were set in MEDLINE. We also manually searched references in relevant articles that were identified during screening.
Two investigators (M.R.‐C. and A.H.) independently reviewed all titles and abstracts identified by the search using an online tool for title and abstract screening (http://abstrackr.cebm.brown.edu/). Studies with discrepant decisions were full‐text reviewed, and disagreements between reviewers were resolved by consensus.
Studies were eligible if they met the following criteria: studies had to have been conducted in adult, nonpregnant humans; metabolites studied had to be related to more than 1 specific biological pathway or come from different chemical classes (ie, following an agnostic and/or comprehensive approach); serum or plasma was the biospecimen (we excluded metabolomics profiling conducted in urine samples); metabolite profiling had to have been conducted at least at baseline in the context of a prospective study; and myocardial infarction, stroke, and/or CVD death were included as part of the definition of the main CVD outcome(s).
Data, extracted independently by 2 investigators (M.R.‐C. and A.H.), included first author, year of publication and journal, study name and location, design of the study, duration of follow‐up, sample size, analysis technique, biospecimen (serum/plasma), primary outcome, number, type, and identity of metabolites investigated, analysis approach (targeted/untargeted), statistical tests used, covariates included in the fully adjusted model, and main findings.
We identified 629 titles from electronic databases after the removal of duplicates (Figure). Following the screening of titles and abstracts, 202 articles were eligible for full‐text review; 11 of these were selected, and 1 article was added after a hand search, for a total of 12 original articles included in the present systematic review.12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 Most of the 191 articles were excluded because they were cross‐sectional or did not include in their outcome definition at least 1 of our prespecified CVD outcomes (ie, MI, stroke, and/or CVD death).
Characteristics of Included Studies
General characteristics of the 12 selected articles12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 are shown in Table 3. Half of the articles (6/12) were published in 2014, and all except 220, 23 were conducted using European and/or US populations. These 12 articles include 19 separate primary discovery (or “learning”) and replication (or “validation”) analyses of metabolites in relation to CVD risk. Three articles12, 13, 15 included replication analyses conducted in samples derived from the same population, and another 4 articles included replication analyses conducted in 118, 19, 23 or 222 samples from different populations. Publications by Shah and colleagues12 and Zheng and colleagues21 presented the associations between metabolites and CVD risk only as secondary analyses. Several articles12, 13, 17 included cross‐sectional analyses as well.
Most of the articles used a cohort design for 1 or more of their analyses,12, 14, 16, 18, 19, 20, 21, 22, 23 4 articles exclusively or additionally included a case‐cohort design,17, 19, 22, 23 and 3 exclusively or additionally included a case‐control design,12, 13, 15 either as the discovery or replication sample analysis. The average follow‐up was 10 years or less in most analyses, except for 6 analyses in 2 separate articles that each included follow‐up longer than 10 years.21, 22 Participant numbers in a given analysis ranged from 67 participants16 to 7256 participants.22 Sample size or power calculations were not explicitly mentioned in any article, although Rizza and colleagues acknowledged their small sample size (67 participants) as the main limitation of their study and performed survival random forest analysis as a way to strengthen their results.16
Participants in 6 articles were free of CVD at baseline,15, 17, 18, 19, 21, 22 but 1 of them was conducted in individuals initiating hemodialysis.15 In 3 articles12, 13, 14 participants had a previous history of suspected coronary disease at baseline. Another article included older participants, of whom 68% had a prior history of CVD,16 and 2 articles included exclusively individuals with type 2 diabetes mellitus coupled with history of CVD or other CVD risk factors.20, 23
In most studies the main outcome was a composite of several end points in addition to MI, stroke, and/or CVD death, with additional CVD conditions including, for example, angina, revascularization, or heart failure.
All of the studies used variations of mass spectrometry (MS) for analyzing metabolite features, while 2 studies also used nuclear magnetic resonance spectroscopy (NMR) or both NMR and MS techniques.17, 22 Additionally, the consistency between NMR and liquid chromatography (LC)‐MS for biomarker associations with CVD was assessed by Würtz and colleagues.22 Twelve analyses relied on plasma, and 7 used serum samples.
Methodological and Statistical Approaches
Four studies used an untargeted profiling approach to identify both unknown and known compounds, including up to 10 162 metabolite features,13, 17, 19, 21 although Zheng and colleagues21 only analyzed 356 named compounds (Table 4). In the reporting of results of these untargeted analyses, independent associations of each metabolite feature were not presented in either the main text or supplemental material (likely due to limitations of space); thus, all unreported associations are presumed to be null. However, 11 out of 12 articles,12, 13, 15, 16, 17, 18, 19, 20, 21, 22, 23 whether targeted or untargeted in their approach, included univariate‐ or multivariate‐adjusted estimates of statistically significant single metabolite associations with CVD incidence in either the main text or supplemental material. Among the studies including only targeted/known metabolites, there was a minimum of 31 metabolites in the study by Kume and colleagues20 and a maximum of 310 lipid species in the study by Alshehry and colleagues.23 Targeted metabolite features tended to include groups of amino acids and related metabolites, acylcarnitines, and lipids.
Different data reduction approaches were applied, including principal component analysis (PCA), stepwise selection, correlation minimization, and others. PCA was implemented in 3 (primary) analyses,12, 14, 16 and another employed PCA in secondary analyses.23 The derived factors were then used as independent variables potentially associated with CVD risk. A combination of learning/discovery and validation/replication samples were used in 6 studies, in which features that were found to be significant in the learning set were carried into the validation set(s).12, 13, 15, 19, 22, 23 The least absolute shrinkage and selection operator (LASSO) algorithm was applied in 3 articles.17, 18, 23
In 6 of 12 articles,12, 14, 16, 17, 20, 21 a score was developed combining between a minimum of 4 and a maximum of 16 metabolites; scores were subsequently used as an independent variable to predict CVD risk. Two of these articles calculated a score by summing the regression coefficients of metabolites independently associated with CVD, multiplied by the metabolite levels, and then used the score to prospectively assess the association with CVD.17, 20 Another article used the sum of quartile ranks according to the association between metabolites and alcohol and considered 3 specific metabolic pathways.21
All articles except 1 used Cox regression models to estimate the association between metabolites, components, or scores and the risk of incident CVD. There was wide variation in the covariates included in fully adjusted models, although all except 213, 21 included classic cardiovascular risk factors (ie, age, sex, smoking, body mass index, diabetes mellitus, hypertension or blood pressure, and total or HDL cholesterol).
Metabolites Associated With CVD Risk
The metabolite features in the 3 articles12, 14, 16 primarily evaluating PCA‐derived components were carnitines and amino acids. Higher CVD risk was found for participants with higher levels of short‐,12, 14 medium‐,16 and long‐chain14 carnitines. PCA components including the amino acids alanine16 and proline14 were also associated with higher CVD risk.
The 3 articles17, 20, 21 that developed scores based on regression coefficients of individual metabolites observed higher CVD risk with higher scores. Vaarhorst and colleagues17 developed a metabolite score derived from untargeted NMR with signals corresponding to 36 different compounds. In a LASSO algorithm, 16 metabolite signals were included in the score: creatinine, serine, glucose, 1,5‐anhydrosorbitol, trimethylamine N‐oxide (TMAO), ornithine, citrate, glutamate, glycoproteins, an unsaturated lipid structure, valine, and 5 nonannotated signals. Kume and colleagues20 targeted 31 amino acids using high‐performance LC‐electrospray ionization‐MS/MS and calculated the area under the curve for models including all possible combinations of 6 or fewer amino acids. The final model included ethanolamine, hydroxyproline, glutamic acid, 3‐methylhistidine, tyrosine, and tryptophan, and it was defined as the amino acid‐based index. Zheng and colleagues, in an untargeted NMR approach capturing 356 named compounds (mainly lipids and amino acids), developed scores from the sums of quartile ranks of alcohol‐related metabolites belonging to 3 metabolic pathways—the γ‐glutamyl dipeptide pathway (γ‐glutamyl, valine, phenylalanine, leucine, isoleucine, tyrosine, glutamate, and alanine), the lysophosphatidylcholine pathway (1‐palmitoleoyl‐glycerophosphocholine, 1‐stearoyl‐glycerophosphoethanolamine, 1‐pentadecanoyl‐glycerophosphocholine, and 2‐arachidonoyl‐glycerophosphoethanolamine), and the 2‐hydroxybutyrate pathway (2‐aminobutyrate, α‐hydroxyisovalerate, α‐hydroxyisobutyrate, α‐hydroxyisocaproate, 2‐hydroxy‐3‐methylvalerate, and 2‐hydroxybutyrate). Higher lysophosphatidylcholine scores and the 2‐hydroxybutyrate scores were associated with 4% to 7% higher CVD risk, and the γ‐glutamyl dipeptide score was associated with 2% reduced risk.21
Five articles13, 15, 18, 19, 22 reduced the initial number of metabolites (ranging from 6822 to >10 00019 features) using either false discovery rate or Bonferroni approaches, resulting in 415 to 3219 features, which were subsequently further analyzed.
Wang and colleagues13 initially reduced >2000 high‐performance LC‐MS features to 18 using Bonferroni and trend criteria, 17/18 of which were independently associated with CVD (identified in the article by their mass‐to‐charge ratios and retention times only). Subsequent analyses, however, focused on 3 phosphatidylcholine metabolites that were highly correlated with each other. Choline, TMAO, and betaine, metabolites related to gut flora, were associated with higher CVD risk, with higher choline notably associated with 18 times (95%CI 4.9‐66.5) the risk when the highest quartile was compared with the lowest.
Kalim and colleagues15 initially targeted 165 LC‐MS/MS metabolites (amino acids and derivatives, carnitines, urea cycle intermediates, nucleotides, positively charged polar metabolites) and, after Bonferroni correction, focused on 4 acylcarnitines plus TMAO. The 4 acylcarnitines (oleoylcarnitine, lineoylcarnitine, palmitoylcarnitine, and stearoylcarnitine) were highly correlated; thus, only oleoylcarnitine was evaluated against CVD risk in both discovery and replication analyses, where higher levels were associated with 2.7‐ and 1.5‐fold higher odds of CVD, respectively. TMAO was nominally associated with higher CVD risk in the discovery but not replication analyses.
Stegemann and colleagues18 initially targeted 135 lipid features in triple‐quadrupole‐MS, reduced to 28 after false discovery rate, finally focusing on 3 lipid metabolites consistent across LASSO and 2 other selection methods (backward stepwise and best subset). Triacylglycerol 54:2, cholesterol ester 16:1 and phosphatidylethanolamine 36:5 were each associated with between 16% and 24% higher risk of CVD.
In an untargeted ultraperformance LC‐MS analysis of 10 162 features, Ganna and colleagues19 identified 32 unique metabolites after false discovery rate correction, and these were carried forward to a validation sample. In meta‐analyses of the discovery and validation samples, 3 metabolites were significantly associated with lower CVD risk (lysophosphatidylcholine 18:1, 18:2, and sphingomyelin 28:1), and another (monoglyceride 18:2) was associated with higher risk. However, in a Mendelian randomization analysis only a weak positive causal effect was suggested for the association between monoglyceride 18:2 and CVD.
In an analysis of 68 targeted features including lipids, amino acids, and other metabolites, Würtz and colleagues22 identified 19 metabolites in the discovery cohort significant after Bonferroni correction, which were subsequently carried into 2 separate replication cohorts. In meta‐analyses of the 3 cohorts, 5 metabolites were associated with risk of CVD: phenylalanine and monounsaturated fatty acids were associated with 18% and 17% higher risk, respectively, and polyunsaturated fatty acids, ω‐6 fatty acids, and docosahexanoic acid were associated with 12%, 11%, and 10% lower risk of CVD, respectively.
Finally, Alshehry and colleagues23 evaluated 310 LC‐MS–derived lipid species in relation to CVD. After removing highly correlated species in the discovery sample, 27 lipid species were directly associated (including di‐ and trihexosylceramides, alkylphosphatidylcholines, and lysoalkylphosphatidylcholines), and 5 (including several phosphatidylcholines) were inversely associated with CVD events and/or death. Results of analyses for CVD death alone overlapped with those for events and death combined.
Nine articles assessed whether metabolites or scores significantly associated with CVD risk were useful in discriminating and/or improving prediction of cases versus noncases beyond that obtained with traditional risk factors alone and reported Harrell C discrimination, net reclassification improvement, and integrative discrimination improvement indices (Table 5). Tiny improvements in CVD prediction were observed when metabolites were added to predictive models already containing traditional CVD risk factors. Specifically, after adding the metabolites into the prediction model with traditional CVD risk factors, the C index increased between 0.006 points in the study by Shah and colleagues14 and 0.05 points in the study by Rizza and colleagues.16
After a 2‐step metabolite‐ranking procedure and optimal model selection, 7 lipid species were retained in the study by Alshehry and colleagues23: alkylphosphatidylcholine [PC(O‐36:1)], cholesteryl ester [CE(18:0)], alkylphosphatidylethanolamine [PE(O‐36:4)], phosphatidylcholines [PC(28:0) and PC(35:4)], and lysophosphatidylcholines [LPC(20:0) and LPC(18:2)]. Despite the inclusion of these species in optimized CVD prediction models, only 3 were independently significantly associated with CVD risk in weighted Cox models in the discovery sample, and none was significant in the replication sample. As shown in Table 5, the C index improved from 0.680 using only clinical variables to 0.700 after addition of these 7 lipid species.
This systematic review identified 12 articles including 19 analyses that have prospectively assessed the association between a wide circulating metabolomic profile and risk of CVD events.12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 These articles included metabolite features measured at baseline using predominantly MS as the analytical method. The number and chemical diversity of metabolites were very heterogeneous. Several data reduction approaches were followed in order to identify a smaller subset of metabolites associated with CVD risk. Most of these articles also evaluated the incremental discriminative and predictive capability of metabolites beyond the use of only clinical information and traditional risk factors. Our systematic review reveals the diversity and complexity of current metabolomic profiling in human CVD and, moreover, the challenge of currently drawing any summary conclusions regarding specific circulating metabolites as they relate to CVD risk.
Metabolites Associated With CVD
According to this systematic review, the following types of metabolites are, individually or as a group, associated with CVD risk: acylcarnitines and dicarboxylacylcarnitines,12, 14, 15, 16 TMAO,13, 15, 17 several amino acids such as phenylalanine,21, 22 glutamate,17, 20, 21 and several lipid classes.18, 19, 21, 22, 23 In hypothesis‐based analyses with participants from the PREDIMED study, we have also found an association between CVD risk and branched‐chain amino acids,24 acylcarnitines,25 glutamate,26 ceramides,27 and tryptophan.28 Other studies have also found an association between a score of 3 amino acids,29 as well as ceramides30 and the risk of CVD. This consistency reinforces the potential causal relationship between these metabolites and CVD or at least the role of these metabolites as biomarkers of biological dysfunction related to CVD.31
Interestingly, dicarboxylcarnitines and acylcarnitines,32 plasma branched and aromatic amino acids,33, 34, 35 phenylalanine,33, 34, 35 α‐hydroxybutyrate,36, 37 and ceramides38, 39 have been associated with cardiovascular risk factors including obesity, insulin resistance, and diabetes mellitus. Different complex mechanisms, including inflammation and stress oxidation, are underlying processes potentially explaining these associations.39
In order to address specific questions regarding the utility and practice of broad (agnostic) metabolic profiling (“metabolic fingerprint”) in predicting CVD, we restricted our systematic review to prospective studies assessing more than 1 specific biological pathway and/or metabolites from different chemical classes. Therefore, we excluded previous studies using only 1 metabolite or only a small set of targeted compounds. For example, we did not include a meta‐analysis of 22 prospective studies published between 2001 and 2013 that found an association between a single metabolite, asymmetric dimethylarginine, and CVD outcomes.40 Although informative, 1 drawback of a relatively narrow, targeted approach is that it does not render a global metabolic picture of our understanding of the complex biological mechanisms underlying CVD, thus potentially leading us down less‐than‐fruitful paths.41 Numerous parallels may be drawn with genetics studies in this regard, where single variant studies have frequently yielded different results from genome‐wide studies. However, genetics studies now benefit from considerably more uniform techniques, statistical approaches, databases, and reporting methods, all of which are still broadly lacking in metabolomics research.
Eight of the articles in this review used a targeted approach and analyzed metabolites from the same chemical families. Among them, 5 articles targeted acylcarnitines and/or amino acids,12, 14, 15, 16, 20 and 2 studies analyzed a group of lipids.18, 23 Würtz and colleagues22 included 68 metabolites, although some compounds evaluated (eg, ω‐3 fatty acids) cannot be properly considered “small” molecules. Among these targeted approaches, the number of metabolites ranged from 31 amino acids20 to 310 lipid species.23 Targeted metabolomics quantifies the levels of metabolites, and it is preferable when the aim is a specific pathway.42 However, limiting the number or variety of metabolites may be considered a source of bias when the aim is to define a more global metabolic profile or identify the most important circulating biomarkers related to CVD risk.43 In other words, when taken at face value, the results of this systematic review could suggest that acyclarnitines are reliable predictors of incident CVD given that 4 articles showed elevated CVD risk with higher acylcarnitine levels. However, this would belie 2 important points: first, that which is not studied cannot be evaluated, and second, the studies using an agnostic, untargeted approach did not find—to our knowledge—associations of acylcarnitines with CVD risk. However, this latter approach is likely to be less powerful because of issues related to multiple testing in comparison with targeted approaches based on a priori hypotheses. Additionally, it is difficult to compare the results when the number and nature of initial metabolites differ so greatly among studies.
Four articles included in this systematic review followed an untargeted approach.13, 17, 19, 21 This approach is initially considered an unbiased and unsupervised manner of biomarker discovery. However, the initial range of metabolite features ranged from 10017 to more than 10 000.19 In addition to the potential risk of bias in those studies with less comprehensive extraction procedures, an additional problem in an untargeted approach is the identification of metabolites. There is still a relatively low percentage of known metabolites with annotated spectra.44 This results in additional difficulties in comparing results from studies using different untargeted methods.
In this review, articles using untargeted methods also differed from those using targeted methods in the reliability of quantification. An untargeted analysis usually provides a relative quantification, whereas a targeted approach presents absolute quantities, usually expressed in micromolar units. This difference in analytical approaches is also an important factor in terms of comparability between studies. However, improvements in instrumentation will likely allow the merging of untargeted data collection with quantification of metabolites. This would present an optimal situation, combining a hypothesis‐free approach with an approach driven by an a priori hypothesis.
A related problem is with the reporting of findings. Journal space limitations naturally preclude the publication of hundreds of metabolite associations, even in supplemental material. Even if the vast majority of these associations are null, without the availability of such data, there is no reliable way to include effect/association sizes in future quantitative meta‐analyses without risking a biased presentation. This will likely become a more important issue as calls for full data and results sharing continue to gain momentum.
Study Designs, Populations, and Outcomes
All the articles included in this systematic review measured metabolites at baseline and assessed the association of these profiles with incident CVD. Previous studies45, 46, 47 have compared the metabolite phenotype of CVD patients with that of healthy participants. Such an approach might allow the identification of abnormalities in metabolism present in diagnosed CVD, which may drive secondary prevention endeavors.7 In contrast, the studies presented in this systematic review were designed to study circulating metabolites associated with future risk of CVD through the identification of early metabolic changes.5
We found important differences among articles in baseline characteristics of participants and in outcome definitions. Four articles included participants with baseline coronary stenosis12, 13, 14 or previous history of CVD.16 This makes it difficult to compare the findings for associations between metabolite profiles and CVD across all the included articles because the presence of baseline coronary disease is likely already influencing the levels of metabolites. Cardiometabolic risk factors such as anthropometric markers,48 obesity,33, 48, 49 and diabetes mellitus34, 50 have been shown to have a metabolic fingerprint. In previous research we found that an association between baseline branched‐chain amino acids and CVD was no longer statistically significant after adjusting for diabetes mellitus, dyslipidemia, and hypertension, which may be intermediate factors in the causal path.24 Moreover, the metabolite profile may be different according to the stage of disease.51 Thus, the varying selection criteria and control for confounders add more challenges to understanding the complexity of metabolic networks when comparing results among published articles.52
Eleven of 12 studies used MS with some separation technique, with 1 of these additionally including NMR in a targeted approach22; 1 study exclusively used NMR and followed an untargeted approach.17 The study using both NMR and MS techniques concluded that metabolite associations with CVD obtained from NMR were largely consistent with those obtained using a LC‐MS platform.22 Individual metabolites are separated by their mass‐to‐charge ratio in the first case and by their magnetic resonance shift in the second. NMR requires minimal sample preparation and is less expensive that MS.53 These characteristics make it more appealing in the cardiovascular clinical context.54 However, compared with MS, NMR has lower sensitivity and is limited to the analysis of around 100 of the most abundant metabolites in a sample. In fact, 100 signals were initially identified in the study using NMR and an untargeted approach,17 whereas thousands of signals were identified in the untargeted studies using MS.13, 19 Therefore, MS is ostensibly a better approach to discovering new biomarkers because it enables the measurement of low concentrations of metabolites.
Another difference among the studies was the type of sample and sample‐preparation method. Seven studies used plasma samples, 3 used serum, and 2 used both.19, 22 Samples were also not uniformly drawn in the fasted state. Plasma and serum samples are similar but not equivalent, and care should be taken before extrapolating results obtained from plasma to serum or vice‐versa. The sample preparation and extraction protocol are key aspects of metabolomics analyses. Physical aspects such as the extraction solvent, temperature, derivatization reagents, and so on may affect the extraction process. The use of coagulant could also affect results. Thus, although reproducibility between plasma and serum samples is possible, differences among metabolite concentrations can be found between types of blood samples.
All the articles included in this systematic review used methods to identify a combination of metabolites or a reduced number of individual metabolites related to CVD. The approach of combining metabolite signals may lead to data overfitting when using β coefficients obtained from analyses of individual metabolites, in which weights were used to calculate combined scores and thereafter to relate scores to study outcomes. Six articles applied univariate analyses (ie, when 1 metabolite is analyzed at a time) to select the metabolites associated with CVD. Four of them13, 19, 22, 23 initially used a discovery sample, followed by a correction method for multiple testing (false discovery rate or Bonferroni) and testing in a validation sample. Shah and colleagues12 presented as main results metabolite features without correction for multiple comparisons because they considered the analyses to be exploratory and because a Bonferroni correction was too conservative. Besides considering the multiple testing issue, univariate analyses do not take into account the relationship between and among metabolites in similar or different pathways. In addition, many of the analyses including 1 metabolite at a time (univariate analyses) were unadjusted for potential confounders.
Six articles used data‐driven multivariate methods in which all variables are included simultaneously, and some variable‐reduction technique is subsequently applied to deal with the relationships among them. Of these studies, 412, 14, 16, 23 applied PCA, which is frequently used to deal with a high number of interrelated variables in several biological pathways. The other 2 articles17, 18 applied the LASSO algorithm, which is an automated variable selection method, and another used both multivariable models optimized with Akaike information criteria and the LASSO algorithm.23 The study by Stegemann and colleagues also applied 2 alternative selection algorithms (backward stepwise and best subset) and included a network analysis.18 Another data‐driven approach—assessing CVD risk according to metabolites grouped in scores—presents an additional difficulty in synthesizing results among the presently included studies. In the studies reviewed, score calculations such as PCA, were based on approaches specific to the samples studied, thus making it difficult to extrapolate the relevance of these scores from 1 sample population to another.
In short, the different statistical approaches show the need for clearer standards about the statistical analyses that should be applied in metabolomics. Data pretreatment methods, including scaling, centering, and transformations, are another source of heterogeneity between and among studies. This methodological aspect is often overlooked, although it can be an important determinant in the selection process of those metabolites that may become more influential in the results. Issues of sample size and statistical power, addressed in a very limited fashion in the studies included in this review, are also key aspects that should be addressed more thoroughly in future metabolomic analyses.
Limitations and Strengths
Several limitations of this systematic review should be acknowledged. First, we excluded those studies that assessed the association between only 1 or a small set of targeted compounds and CVD risk. The typical rationals for these a priori hypothesis‐based studies are already known associations between specific metabolites and/or pathways and CVD risk. Some of this knowledge was obtained in some cases before the development of metabolomics. In contrast, this review was focused on studies following a wider metabolomic approach and where a statistical method was used for data reduction. Our rationale for this decision was our aim to obtain a robust summary of the best available evidence relating a broad metabolomic fingerprint and a prospective design with hard clinical end points. Second, for this same reason we excluded studies using other types of biospecimens such as urine or saliva. We opted to focus on circulating blood metabolites to reduce the already known variability in metabolomic profiles among different types of biosamples. Third, we excluded cross‐sectional studies that examined the metabolomic profile of CVD patients and controls because our aim was to identify metabolites associated with early metabolic changes to predict the future risk of CVD. Fourth, we were unable to conduct a quantitative analysis, mainly because of the heterogeneity and limitations of the articles already noted above, that is, widely differing metabolite targets and approaches and differential reporting or nonreporting of associations. Nevertheless, our study is the first attempt to systematically review the results and methodological aspects of studies aimed to assess the association between a wide peripheral blood metabolomic profile and risk of future CVD events.
Metabolomics holds considerable promise as an emerging field applied to the discovery of novel biomarkers for the future risk of CVD. There are still a small number of longitudinal studies assessing the association between baseline metabolomic profiles and the risk of CVD. Current data are promising, although approaches and results are heterogeneous. The lack of robust replications is 1 of the main problems in the current literature because of heterogeneity in study designs, end points, metabolomics platforms, and small sample sizes. Toward this end, standardization of platforms, data analysis approaches, and study designs is critical. We also need larger numbers of cases, longer durations of follow‐up, and repeated measures of metabolites, if possible. A pooled analysis of multiple studies would further help to improve statistical power and standardize analytic approaches. Finally, basic science research is needed to achieve better understanding of the biological mechanisms underlying the epidemiologic findings.
Sources of Funding
This work was supported by the National Institutes of Health, NHLBI 1R01HL118264 and NIDDK R01DK 102896. Hruby's contribution to this work was partially supported by the US Department of Agriculture–Agricultural Research Service, under Agreement No. 58‐1950‐4‐003.
- ↵Mozaffarian D, Benjamin EJ, Go AS, Arnett DK, Blaha MJ, Cushman M, Das SR, de Ferranti S, Després J‐P, Fullerton HJ, Howard VJ, Huffman MD, Isasi CR, Jiménez MC, Judd SE, Kissela BM, Lichtman JH, Lisabeth LD, Liu S, Mackey RH, Magid DJ, McGuire DK, Mohler ER, Moy CS, Muntner P, Mussolino ME, Nasir K, Neumar RW, Nichol G, Palaniappan L, Pandey DK, Reeves MJ, Rodriguez CJ, Rosamond W, Sorlie PD, Stein J, Towfighi A, Turan TN, Virani SS, Woo D, Yeh RW, Turner MB. Heart disease and stroke statistics—2016 update: a report from the American Heart Association. Circulation. 2016;133:e38–e60.
- ↵Roth GA, Nguyen G, Forouzanfar MH, Mokdad AH, Naghavi M, Murray CJL. Estimates of global and regional premature cardiovascular mortality in 2025. Circulation. 2015;132:1270–1282.
- ↵Taegtmeyer H, Young ME, Lopaschuk GD, Abel ED, Brunengraber H, Darley‐Usmar V, Des Rosiers C, Gerszten R, Glatz JF, Griffin JL, Gropler RJ, Holzhuetter H‐G, Kizer JR, Lewandowski ED, Malloy CR, Neubauer S, Peterson LR, Portman MA, Recchia FA, Van Eyk JE, Wang TJ. Assessing cardiac metabolism. A scientific statement from the American Heart Association. Circ Res. 2016;118:1659–1701.
- ↵Brown JM, Hazen SL. Seeking a unique lipid signature predicting cardiovascular disease risk. Circulation. 2014;129:1799–1803.
- ↵Ussher JR, Elmariah S, Gerszten RE, Dyck JRB. The emerging role of metabolomics in the diagnosis and prognosis of cardiovascular disease. J Am Coll Cardiol. 2016;68:2850–2870.
- ↵Shah SH, Kraus WE, Newgard CB. Metabolomic profiling for the identification of novel biomarkers and mechanisms related to common cardiovascular diseases form and function. Circulation. 2012;126:1110–1120.
- ↵Stroup DF, Berlin JA, Morton SC, Olkin I, Williamson GD, Rennie D, Moher D, Becker BJ, Sipe TA, Thacker SB. Meta‐analysis of observational studies in epidemiology: a proposal for reporting. Meta‐analysis Of Observational Studies in Epidemiology (MOOSE) group. JAMA. 2000;283:2008–2012.
- ↵Shah SH, Bain JR, Muehlbauer MJ, Stevens RD, Crosslin DR, Haynes C, Dungan J, Newby LK, Hauser ER, Ginsburg GS, Newgard CB, Kraus WE. Association of a peripheral blood metabolic profile with coronary artery disease and risk of subsequent cardiovascular events. Circ Cardiovasc Genet. 2010;3:207–214.
- ↵Wang Z, Klipfell E, Bennett BJ, Koeth R, Levison BS, Dugar B, Feldstein AE, Britt EB, Fu X, Chung Y‐M, Wu Y, Schauer P, Smith JD, Allayee H, Tang WHW, DiDonato JA, Lusis AJ, Hazen SL. Gut flora metabolism of phosphatidylcholine promotes cardiovascular disease. Nature. 2011;472:57–63.
- ↵Shah SH, Sun J‐L, Stevens RD, Bain JR, Muehlbauer MJ, Pieper KS, Haynes C, Hauser ER, Kraus WE, Granger CB, Newgard CB, Califf RM, Newby LK. Baseline metabolomic profiles predict cardiovascular events in patients at risk for coronary artery disease. Am Heart J. 2012;163:844–850.e1.
- ↵Kalim S, Clish CB, Wenger J, Elmariah S, Yeh RW, Deferio JJ, Pierce K, Deik A, Gerszten RE, Thadhani R, Rhee EP. A plasma long‐chain acylcarnitine predicts cardiovascular mortality in incident dialysis patients. J Am Heart Assoc. 2013;2:e000542. DOI: 10.1161/JAHA.113.000542.
- ↵Vaarhorst AAM, Verhoeven A, Weller CM, Böhringer S, Göraler S, Meissner A, Deelder AM, Henneman P, Gorgels APM, van den Brandt PA, Schouten LJ, van Greevenbroek MM, Merry AHH, Verschuren WMM, van den Maagdenberg AMJM, van Dijk KW, Isaacs A, Boomsma D, Oostra BA, van Duijn CM, Jukema JW, Boer JMA, Feskens E, Heijmans BT, Slagboom PE. A metabolomic profile is associated with the risk of incident coronary heart disease. Am Heart J. 2014;168:45–52.e7.
- ↵Stegemann C, Pechlaner R, Willeit P, Langley SR, Mangino M, Mayr U, Menni C, Moayyeri A, Santer P, Rungger G, Spector TD, Willeit J, Kiechl S, Mayr M. Lipidomics profiling and risk of cardiovascular disease in the prospective population‐based Bruneck study. Circulation. 2014;129:1821–1831.
- ↵Ganna A, Salihovic S, Sundström J, Broeckling CD, Hedman ÅK, Magnusson PKE, Pedersen NL, Larsson A, Siegbahn A, Zilmer M, Prenni J, Ärnlöv J, Lind L, Fall T, Ingelsson E. Large‐scale metabolomic profiling identifies novel biomarkers for incident coronary heart disease. PLoS Genet. 2014;10:e1004801.
- ↵Kume S, Araki S, Ono N, Shinhara A, Muramatsu T, Araki H, Isshiki K, Nakamura K, Miyano H, Koya D, Haneda M, Ugi S, Kawai H, Kashiwagi A, Uzu T, Maegawa H. Predictive properties of plasma amino acid profile for cardiovascular disease in patients with type 2 diabetes. PLoS One. 2014;9:e101219.
- ↵Zheng Y, Yu B, Alexander D, Steffen LM, Nettleton JA, Boerwinkle E. Metabolomic patterns and alcohol consumption in African Americans in the Atherosclerosis Risk in Communities Study. Am J Clin Nutr. 2014;99:1470–1478.
- ↵Würtz P, Havulinna AS, Soininen P, Tynkkynen T, Prieto‐Merino D, Tillin T, Ghorbani A, Artati A, Wang Q, Tiainen M, Kangas AJ, Kettunen J, Kaikkonen J, Mikkilä V, Jula A, Kähönen M, Lehtimäki T, Lawlor DA, Gaunt TR, Hughes AD, Sattar N, Illig T, Adamski J, Wang TJ, Perola M, Ripatti S, Vasan RS, Raitakari OT, Gerszten RE, Casas J‐P, Chaturvedi N, Ala‐Korpela M, Salomaa V. Metabolite profiling and cardiovascular event risk: a prospective study of 3 population‐based cohorts. Circulation. 2015;131:774–785.
- ↵Alshehry ZH, Mundra PA, Barlow CK, Mellett NA, Wong G, McConville MJ, Simes J, Tonkin AM, Sullivan DR, Barnes EH, Nestel PJ, Kingwell BA, Marre M, Neal B, Poulter NR, Rodgers A, Williams B, Zoungas S, Hillis GS, Chalmers J, Woodward M, Meikle PJ. Plasma lipidomic profiles improve on traditional risk factors for the prediction of cardiovascular events in type 2 diabetes mellitus. Circulation. 2016;134:1637–1650.
- ↵Ruiz‐Canela M, Toledo E, Clish CB, Hruby A, Liang L, Salas‐Salvadó J, Razquin C, Corella D, Estruch R, Ros E, Fito M, Gómez‐Gracia E, Aros F, Fiol M, Lapetra J, Serra‐Majem L, Martínez‐González M, Hu FB. Plasma branched‐chain amino acids and incident cardiovascular disease in the PREDIMED trial. Clin Chem. 2016;62:582–592.
- ↵Guasch‐Ferré M, Zheng Y, Ruiz‐Canela M, Hruby A, Martínez‐González MA, Clish CB, Corella D, Estruch R, Ros E, Fitó M, Dennis C, Morales‐Gil IM, Arós F, Fiol M, Lapetra J, Serra‐Majem L, Hu FB, Salas‐Salvadó J. Plasma acylcarnitines and risk of cardiovascular disease: effect of Mediterranean diet interventions. Am J Clin Nutr. 2016;103:1408–1416.
- ↵Zheng Y, Hu FB, Ruiz‐Canela M, Clish CB, Dennis C, Salas‐Salvado J, Hruby A, Liang L, Toledo E, Corella D, Ros E, Fitó M, Gómez‐Gracia E, Arós F, Fiol M, Lapetra J, Serra‐Majem L, Estruch R, Martínez‐González MA. Metabolites of glutamate metabolism are associated with incident cardiovascular events in the PREDIMED PREvención con DIeta MEDiterránea (PREDIMED) Trial. J Am Heart Assoc. 2016;5:e003755. DOI: 10.1161/JAHA.116.003755.
- ↵Wang DD, Toledo E, Hruby A, Rosner BA, Willett WC, Sun Q, Razquin C, Zheng Y, Ruiz‐Canela M, Guasch‐Ferré M, Corella D, Gómez‐Gracia E, Fiol M, Estruch R, Ros E, Lapetra J, Fitó M, Aros F, Serra‐Majem L, Lee C‐H, Clish CB, Liang L, Salas‐Salvadó J, Martínez‐González MA, Hu FB. Plasma ceramides, Mediterranean diet, and incident cardiovascular disease in the PREDIMED Trial. Circulation. 2017;135:2028–2040.
- ↵Yu E, Ruiz‐Canela M, Guasch‐Ferré M, Zheng Y, Toledo E, Clish CB, Salas‐Salvadó J, Liang L, Wang DD, Corella D, Fitó M, Gómez‐Gracia E, Lapetra J, Estruch R, Ros E, Cofán M, Arós F, Romaguera D, Serra‐Majem L, Sorlí JV, Hu FB, Martinez‐Gonzalez MA. Increases in plasma tryptophan are inversely associated with incident cardiovascular disease in the Prevención con Dieta Mediterránea (PREDIMED) Study. J Nutr. 2017;147:314–322.
- ↵Havulinna AS, Sysi‐Aho M, Hilvo M, Kauhanen D, Hurme R, Ekroos K, Salomaa V, Laaksonen R. Circulating ceramides predict cardiovascular outcomes in the population‐based FINRISK 2002 cohort highlights. Arterioscler Thromb Vasc Biol. 2016;36:2424–2430.
- ↵Johnson CH, Ivanisevic J, Siuzdak G. Metabolomics: beyond biomarkers and towards mechanisms. Nat Rev Mol Cell Biol. 2016;17:1–9.
- ↵Liepinsh E, Makrecka‐Kuka M, Makarova E, Volska K, Svalbe B, Sevostjanovs E, Grinberga S, Kuka J, Dambrova M. Decreased acylcarnitine content improves insulin sensitivity in experimental mice models of insulin resistance. Pharmacol Res. 2016;113:788–795.
- ↵Newgard CB, An J, Bain JR, Muehlbauer MJ, Stevens RD, Lien LF, Haqq AM, Shah SH, Arlotto M, Slentz CA, Rochon J, Gallup D, Ilkayeva O, Wenner BR, Yancy WS, Eisenson H, Musante G, Surwit RS, Millington DS, Butler MD, Svetkey LP. A branched‐chain amino acid‐related metabolic signature that differentiates obese and lean humans and contributes to insulin resistance. Cell Metab. 2009;9:311–326.
- ↵Guasch‐Ferré M, Hruby A, Toledo E, Clish CB, Martínez‐González MA, Salas‐Salvadó J, Hu FB. Metabolomics in prediabetes and diabetes: a systematic review and meta‐analysis. Diabetes Care. 2016;39:833–846.
- ↵Ferrannini E, Natali A, Camastra S, Nannipieri M, Mari A, Adam K‐P, Milburn MV, Kastenmüller G, Adamski J, Tuomi T, Lyssenko V, Groop L, Gall WE. Early metabolic markers of the development of dysglycemia and type 2 diabetes and their physiological significance. Diabetes. 2013;62:1730–1737.
- ↵Adams JM, Pratipanawatr T, Berria R, Wang E, DeFronzo RA, Sullards MC, Mandarino LJ. Ceramide content is increased in skeletal muscle from obese insulin‐resistant humans. Diabetes. 2004;53:25–31.
- ↵Haus JM, Kashyap SR, Kasumov T, Zhang R, Kelly KR, Defronzo RA, Kirwan JP. Plasma ceramides are elevated in obese subjects with type 2 diabetes and correlate with the severity of insulin resistance. Diabetes. 2009;58:337–343.
- ↵Willeit P, Freitag DF, Laukkanen JA, Chowdhury S, Gobin R, Mayr M, Di Angelantonio E, Chowdhury R. Asymmetric dimethylarginine and cardiovascular risk: systematic review and meta‐analysis of 22 prospective studies. J Am Heart Assoc. 2015;4:e001833. DOI: 10.1161/JAHA.115.001833.
- ↵Rhee EP, Gerszten RE. Metabolomics and cardiovascular biomarker discovery. Clin Chem. 2012;58:139–147.
- ↵Klepacki J, Klawitter J, Klawitter J, Karimpour‐fard A, Thurman J, Ingle G, Patel D, Christians U. Amino acids in a targeted versus a non‐targeted metabolomics LC‐MS/MS assay. Are the results consistent? Clin Biochem. 2016;49:955–961.
- ↵Hoefer IE, Steffens S, Ala‐Korpela M, Bäck M, Badimon L, Bochaton‐Piallat M‐L, Boulanger CM, Caligiuri G, Dimmeler S, Egido J, Evans PC, Guzik T, Kwak BR, Landmesser U, Mayr M, Monaco C, Pasterkamp G, Tuñón J, Weber C. Novel methodologies for biomarker discovery in atherosclerosis. Eur Heart J. 2015;36:2635–2642.
- ↵Vinaixa M, Schymanski EL, Neumann S, Navarro M, Salek RM, Yanes O. Mass spectral databases for LC/MS‐ and GC/MS‐based metabolomics: state of the field and future prospects. Trends Analyt Chem. 2016;78:23–35.
- ↵Jung JY, Lee H‐S, Kang D‐G, Kim NS, Cha MH, Bang O‐S, Ryu DH, Hwang G‐S. 1H‐NMR‐based metabolomics study of cerebral infarction. Stroke. 2011;42:1282–1288.
- ↵Kimberly WT, Wang Y, Pham L, Furie KL, Gerszten RE. Metabolite profiling identifies a branched chain amino acid signature in acute cardioembolic stroke. Stroke. 2013;44:1389–1395.
- ↵Bachlechner U, Floegel A, Steffen A, Prehn C, Adamski J, Pischon T, Boeing H. Associations of anthropometric markers with serum metabolites using a targeted metabolomics approach: results of the EPIC‐Potsdam study. Nutr Diabetes. 2016;6:e215.
- ↵Ho JE, Larson MG, Ghorbani A, Cheng S, Chen M‐H, Keyes M, Rhee EP, Clish CB, Vasan RS, Gerszten RE, Wang TJ. Metabolomic profiles of body mass index in the Framingham Heart Study reveal distinct cardiometabolic phenotypes. PLoS One. 2016;11:e0148361.
- ↵Gilstrap LG, Wang TJ. Biomarkers and cardiovascular risk assessment for primary prevention: an update. Clin Chem. 2012;58:72–82.
- ↵Soininen P, Kangas AJ, Würtz P, Suna T, Ala‐Korpela M. Quantitative serum nuclear magnetic resonance metabolomics in cardiovascular epidemiology and genetics. Circ Cardiovasc Genet. 2015;8:192–206.