Abstract Background The study aimed to show the relationship between a large number of circulating metabolites and subsequent cardiovascular disease (CVD) and subclinical markers of CVD in the general population. Methods and Results In 2278 individuals free from CVD in the EpiHealth study (aged 45–75 years, mean age 61 years, 50% women), 790 annotated nonxenobiotic metabolites were measured by mass spectroscopy (Metabolon). The same metabolites were measured in the PIVUS (Prospective Investigation of Vasculature in Uppsala Seniors) study (n=603, all aged 80 years, 50% women), in which cardiac and carotid artery pathologies were evaluated by ultrasound. During a median follow‐up of 8.6 years, 107 individuals experienced a CVD (fatal or nonfatal myocardial infarction, stroke, or heart failure) in EpiHealth. Using a false discovery rate of 0.05 for age‐ and sex‐adjusted analyses and P<0.05 for adjustment for traditional CVD risk factors, 37 metabolites were significantly related to incident CVD. These metabolites belonged to multiple biochemical classes, such as amino acids, lipids, and nucleotides. Top findings were dimethylglycine and N‐acetylmethionine. A lasso selection of 5 metabolites improved discrimination when added on top of traditional CVD risk factors (+4.0%, P=0.0054). Thirty‐five of the 37 metabolites were related to subclinical markers of CVD evaluated in the PIVUS study. The metabolite 1‐carboxyethyltyrosine was associated with left atrial diameter as well as inversely related to both ejection fraction and the echogenicity of the carotid artery. Conclusions Several metabolites were discovered to be associated with future CVD, as well as with subclinical markers of CVD. A selection of metabolites improved discrimination when added on top of CVD risk factors. Keywords: amino acids, cardiovascular disease, epidemiology, mass spectroscopy, metabolomics Subject Categories: Cardiovascular Disease, Epidemiology, Risk Factors __________________________________________________________________ Nonstandard Abbreviations and Acronyms MS mass spectroscopy PIVUS Prospective Investigation of Vasculature in Uppsala Seniors Clinical Perspective. What Is New? * In a population‐based sample, we found 37 metabolites to be related to incident cardiovascular disease. * These metabolites belonged to multiple biochemical classes, such as amino acids, lipids, and nucleotides. * Most metabolites were also associated with subclinical markers of cardiovascular disease, such as enlarged dimeter of the left atrium and a reduced ejection fraction. What Are the Clinical Implications? * A selection of 5 metabolites improved discrimination of cardiovascular disease by 4% (C statistic) when added to traditional cardiovascular disease risk factors. * If reproduced by others, metabolites might be included in risk prediction scores in the future. Omics technologies could be used in epidemiology either to search for novel insights into the pathogenesis of diseases or to identify clinically relevant risk markers that could add to the predictive power of established risk factors. One such example is the use of circulating proteomics to predict cardiovascular disease (CVD), where studies have identified both already well‐known biomarkers, such as NT‐proBNP (N‐terminal pro‐B‐type natriuretic peptide), but also some new proteins linked to incident CVD.[32] ^1 , [33]^2 Another omics modality is metabolomics, the study of small molecules, measured either by mass spectroscopy (MS) or by nuclear magnetic resonance techniques. In a systematic review from 2017,[34] ^3 the authors found 12 studies in which metabolomic markers were associated with incident CVD. Metabolites from several metabolite classes, such as carnitines, amino acids, and lipids, were related to future CVD. However, the authors concluded that a quantitative synthesizing of the previous findings was difficult, given the use of different analytical platforms and statistical methods. Since 2017, a few additional large studies have been published, using MS or nuclear magnetic resonance.[35] ^4 , [36]^5 , [37]^6 , [38]^7 , [39]^8 Also, in this case, the results are not easily comparable, because the number and nature of annotated metabolites varied greatly between studies (see overview in Table [40]S1). Biomarker discovery could be used for several purposes, including to identify new pathophysiological pathways that potentially could be used as targets for drug development, identify new biomarkers that could be used as diagnostics at the individual level, and identify new biomarkers that could be used for improved risk stratification. The primary aim of the present study was to evaluate if we could identify new pathophysiological pathways for CVD by large‐scale metabolomics. If so, we could, as a secondary aim, evaluate if those metabolites could improve risk stratification in relation to traditional CVD risk factors. For the primary aim, we investigated the metabolomic profile associated with incident CVD (myocardial infarction, stroke, or heart failure) using MS‐based data on almost 800 nonxenobiotic annotated metabolites in the EpiHealth study. This is the study with the greatest number of metabolites evaluated on incident CVD performed to date. To support our findings in the EpiHealth study, we investigated if associations between the metabolites of interest were also related to indices of subclinical CVD in an independent sample, the PIVUS (Prospective Investigation of Vasculature in Uppsala Seniors) study. The 5 indices of subclinical CVD evaluated (left atrial diameter, left ventricular ejection fraction, left ventricular mass, intima‐media thickness, and the echogenicity of the carotid artery wall) have all previously been shown to be associated with incident CVD.[41] ^9 , [42]^10 , [43]^11 , [44]^12 , [45]^13 , [46]^14 METHODS Samples EpiHealth is a population‐based study conducted with the same protocol in 2 Swedish cities, Uppsala and Malmö. Approximately 25 000 individuals participated from 2011 to 2018. The age range was 45 to 75 years, and 50% were women. Traditional CVD risk factors were measured. In the first 2342 subjects included in Uppsala, metabolomics was analyzed on frozen plasma samples. The cohort has been followed for almost 10 years on incident CVD. Details on the examination have been provided previously.[47] ^15 The PIVUS study is a population‐based study conducted in Uppsala, Sweden. At age 70 years, 1016 subjects were investigated from 2001 to 2004 (50% were women). Ten years later, all subjects still alive were invited to a new examination, in which 604 participated. At that time point, traditional CVD risk factors were measured, and an echocardiogram and a carotid ultrasound were obtained. Metabolomics was analyzed on frozen plasma samples in all but 1 individual (because of a lost vial). Thus, the present study used only data from those aged 80 years. Details on the examinations have been given previously.[48] ^16 Eleven percent had a history of myocardial infarction, 10% had suffered from a stroke, and 9% reported a heart failure diagnosis at the time of the investigation. Both studies have been approved by the ethics committee at Uppsala University, and all study participants have given their informed consent to participate. The data that support the finding of this study are available from the corresponding author upon reasonable request. CVD Risk Factors Blood pressure was measured in the supine position in the PIVUS study and in the sitting position in EpiHealth (M10‐IT; Omron, Kyoto, Japan). Blood was drawn after an overnight fast in the PIVUS study and after a minimum of 6 hours in EpiHealth. Glucose and low‐density lipoprotein and high‐density lipoprotein cholesterol were measured by standard techniques at the clinical chemistry laboratory at the university hospital. Plasma was frozen at −80 °C for later metabolomics analysis. Smoking and medications were assessed by a questionnaire. Body mass index (BMI) was calculated by measured height and weight. Diabetes was defined as glucose ≥7.0 mmol/L or use of antidiabetic medication. Metabolomics In both samples, the same nontargeted metabolomics (Metabolon) was performed on plasma samples. Samples were prepared using the automated MicroLab STAR system (Hamilton). Several internal standards were added before the first step in the extraction process for quality control purposes. To remove protein, we dissociated small molecules bound to protein or trapped in the precipitated protein matrix, and to recover chemically diverse metabolites, proteins were precipitated with methanol under vigorous shaking for 2 minutes (Glen Mills GenoGrinder 2000) followed by centrifugation. The resulting extract was divided into 5 fractions: 2 for analysis by 2 separate reverse phases/ultraperformance liquid chromatography–MS/MS methods with positive ion mode electrospray ionization, 1 for analysis by reverse phases/ultraperformance liquid chromatography–MS/MS with negative ion mode electrospray ionization, 1 for analysis by hydrophilic interaction/ultraperformance liquid chromatography–MS/MS with negative ion mode electrospray ionization, and 1 sample was reserved for backup. Only annotated, nonxenobiotic metabolites with a detection rate >75% in all samples were used in the analyses (n=790). The values were normalized and given in arbitrary units. CVD End Point Data from the Swedish registries of mortality and in‐hospital care were used to follow up incident cases of CVD in the EpiHealth study. In this study we used a combined CVD end point consisting of myocardial infarction (International Classification of Diseases, Tenth Revision [ICD‐10] code I21), ischemic stroke (ICD‐10 code I63), and heart failure (ICD‐10 codes I50 and I11.0). Both fatal and nonfatal events were included in the combined end point. Ultrasound In the PIVUS study, a 2‐dimensional echocardiography examination was performed with an Acuson XP124 cardiac ultrasound unit (Acuson). A 2.5‐MHz transducer was used for the majority of the examinations. Left ventricular dimensions were measured with M‐mode online from the parasternal projection, using a leading‐edge‐to‐leading‐edge convention. Measurements included left atrial diameter, interventricular septal thickness, posterior wall thickness, and left ventricular diameter in end‐diastole and end‐systole. Left ventricular mass was determined from the Penn conversion. Left ventricular mass was then indexed for height (g/m^2.7) to obtain left ventricular mass index. Left ventricular volumes were calculated according to the Teichholz formula (7*D3/[2.4+D]), and from that the ejection fraction was calculated. Carotid artery ultrasound was performed using a 10‐MHz probe investigating both arteries. Intima‐media thickness was measured over a 10‐mm distance proximal to the bifurcation in the far wall by semiautomated software. In the same segment, the gray scale of the intima‐media complex was determined (echogenicity of the intima‐media complex). The mean value of the 2 arteries was used for both indices. Details on the carotid artery ultrasound measurements have been reported previously.[49] ^17 Statistical Analysis The metabolomic variables were subjected to inverse rank normalization to obtain normally distributed variables on an SD scale, in both EpiHealth and the PIVUS study. In EpiHealth, we investigated the association between metabolites and incident CVD using separate Cox proportional hazard analysis for each metabolite. Two levels of adjustment were performed. First, adjustment for age and sex, and second, additional adjustment for traditional CVD risk factors (systolic blood pressure, diabetes, low‐density lipoprotein and high‐density lipoprotein cholesterol, BMI, and current smoking). For these analyses, a false discovery rate <0.05 for the age‐ and sex‐adjusted analysis and P<0.05 for adjustment for traditional CVD risk factors were regarded as significant. A lasso logistic regression analysis using the 37 significant metabolites from the previous step as independent variables together with age and sex was conducted to sort out the most influential independent metabolites based on their β coefficients. The sample was split in 2 parts, and we used lasso with 10 cross‐validations. This procedure selects λ* to be the λ that gives the minimum of the cardiovascular function. Because we knew that we had 107 incident CVD cases during the follow‐up period and we had 6 independent significant traditional CVD risk factors (including age and sex), we decided a priori that we should only add 5 metabolites, at most, to the model, including the risk factors, so as not to induce any overfitting problems (based on a rule of thumb that there should be around 10 cases per covariate in the model). First, the traditional CVD risk factors (including age and sex) entered a logistic regression model, and the C index was calculated (after excluding BMI and diabetes, which were far from significant in the model and did not influence the C index). Second, the top 5 metabolites in the lasso regression model (all showing P<0.05) entered a logistic regression model together with age and sex to calculate the C statistics. In the third step, we entered the 5 metabolites together with age, sex, and traditional risk factors and calculated the C index. In a fourth step, the model with the traditional risk factors only were compared with the model including both traditional risk factors and the 5 metabolites. In the PIVUS study, the 37 metabolites identified to be of interest in EpiHealth were one‐by‐one related to each of the 5 markers of subclinical CVD by linear regression analysis. Similar to the previous analyses in EpiHealth, 2 levels of adjustment were performed: (1) adjustment for sex (the age was the same in all subjects) and (2) additional adjustment for traditional CVD risk factors. In this exploratory validation, P<0.05 was considered statistically significant. Stata 16.1 was used for the calculations (StataCorp, College Station, TX). RESULTS Basic characteristics of the 2 cohorts are provided in Table [50]1. After excluding 64 participants with prevalent CVD at baseline, 2278 persons were at risk during a median follow‐up of 8.6 years (maximum 9.6 years, 18 852 person‐years at risk). During that period, 107 individuals suffered from CVD. Table 1. Basic Characteristics of the 2 Samples EpiHealth, n=2278 PIVUS, n=603 Age, y 61.1 (8.4) 80.1 (0.2) Female sex, % 50.2 50.4 Systolic blood pressure, mm Hg 134.1 (16.8) 146.4 (19.2) LDL cholesterol, mmol/L 3.9 (0.9) 3.4 (0.8) HDL cholesterol, mmol/L 1.5 (0.3) 1.4 (0.4) Diabetes (%) 8.0 11.2 Smoking 6.7 (8.9) y smoked 3.1% current smokers Body mass index, kg/m^2 26.4 (3.8) 26.8 (4.3) Left atrial diameter, mm NA 42.2 (6.6) Ejection fraction, % NA 64.8 (10) Left ventricular mass index, g/m^2.7 NA 45.1 (12) Intima‐media thickness, mm NA 0.95 (0.16) Echogenicity of the carotid artery wall NA 59.8 (15.4) [51]Open in a new tab Mean and SD or proportions are given. HDL indicates high‐density lipoprotein; LDL, low‐density lipoprotein; NA, not assessed; and PIVUS, Prospective Investigation of Vasculature in Uppsala Seniors. Metabolites Versus Incident CVD Of the 790 evaluated metabolites, 37 were associated with incident CVD (overview in Figure [52]1 and details in Table [53]2 and Table [54]S2). These metabolites belonged to multiple biochemical classes, such as amino acids, lipids, and nucleotides. Also, within the lipid group, different functional classes were represented, such as steroids and phosphatidylethanolamines. The top findings were 2 amino acid derivates, dimethylglycine and N‐acetylmethionine. As seen in Table [55]2, the relationships were generally positive, but a few of the steroid hormones were (inversely) associated with incident CVD. Figure 1. Overview of the 37 significant metabolites related to incident cardiovascular disease in the EpiHealth sample. Figure 1 [56]Open in a new tab The intensity of the red color is proportional to the estimate (hazard ratio) so that the pale dots correspond to hazard ratios <1.0. Subpathway and metabolite names for these 37 metabolites, with the start from 1‐stearoyl‐2‐oleoyl‐GPE (18:0/18:1) at the top of the figure and then in clockwise order, are given in Table [57]S2. GPE indicates glycero‐3‐phosphoethanolamine; and SAM, S‐adenosylmethionine. Table 2. Relationships Between Metabolites and Incident Cardiovascular Disease Super pathway Subpathway Chemical name Age and sex adjusted Multiple adjusted HR Low 95% CI High 95% CI P value HR Low 95% CI High 95% CI P value Amino acid Alanine and aspartate metabolism N‐acetylalanine 1.52 1.25 1.86 0.000036 1.51 1.22 1.88 0.00017 Amino acid Glycine, serine, and threonine metabolism N‐acetylserine 1.54 1.26 1.88 0.000030 1.54 1.25 1.9 0.000064 Amino acid Glycine, serine, and threonine metabolism Dimethylglycine 1.58 1.3 1.95 0.000010 1.63 1.32 2.01 5.9 e‐06 Amino acid Histidine metabolism Imidazole propionate 1.38 1.13 1.7 0.0019 1.36 1.11 1.68 0.0045 Amino acid Leucine, isoleucine, and valine metabolism N‐acetylvaline 1.43 1.17 1.75 0.00031 1.4 1.15 1.73 0.00095 Amino acid Methionine, cysteine, SAM, and taurine metabolism N‐formylmethionine 1.42 1.17 1.73 0.00040 1.4 1.15 1.73 0.00093 Amino acid Methionine, cysteine, SAM, and taurine metabolism 2,3‐dihydroxy‐5‐methylthio‐4‐pentenoate 1.43 1.16 1.79 0.00074 1.42 1.13 1.77 0.0028 Amino acid Methionine, cysteine, SAM, and taurine metabolism N‐acetylmethionine 1.58 1.3 1.93 6.8 e‐06 1.55 1.27 1.92 0.000027 Amino acid Phenylalanine metabolism Phenylalanine 1.36 1.13 1.67 0.0016 1.35 1.11 1.65 0.0038 Amino acid Polyamine metabolism 5‐methylthioadenosine 1.38 1.14 1.68 0.0010 1.36 1.11 1.68 0.0037 Amino acid Tryptophan metabolism C‐glycosyltryptophan 1.4 1.13 1.73 0.0021 1.38 1.09 1.72 0.0053 Amino acid Tyrosine metabolism 1‐carboxyethyltyrosine 1.45 1.17 1.79 0.00047 1.43 1.15 1.8 0.0017 Carbohydrate Aminosugar metabolism N‐acetylglucosamine/N‐acetylgalactosamine 1.38 1.13 1.68 0.0021 1.32 1.06 1.65 0.012 Cofactors and vitamins Ascorbate and aldarate metabolism Gulonate 1.43 1.17 1.75 0.00048 1.4 1.14 1.72 0.0011 Cofactors and vitamins Nicotinate and nicotinamide metabolism Quinolinate 1.38 1.13 1.68 0.0021 1.35 1.07 1.68 0.0091 Lipid Androgenic steroids Dehydroepiandrosterone sulfate 0.68 0.54 0.84 0.00035 0.67 0.54 0.84 0.00029 Lipid Ceramides N‐palmitoyl‐sphingosine (d18:1/16:0) 1.36 1.12 1.65 0.0023 1.43 1.13 1.84 0.0037 Lipid Dihydroceramides N‐palmitoyl‐sphinganine (d18:0/16:0) 1.4 1.15 1.7 0.00063 1.36 1.09 1.7 0.0063 Lipid Fatty acid, dihydroxy 3,4‐dihydroxybutyrate 1.39 1.14 1.7 0.0012 1.35 1.09 1.67 0.0052 Lipid Phosphatidylethanolamine 1‐stearoyl‐2‐oleoyl‐GPE (18:0/18:1) 1.36 1.12 1.65 0.0019 1.27 1.03 1.57 0.022 Lipid Phosphatidylethanolamine 1‐stearoyl‐2‐arachidonoyl‐GPE (18:0/20:4) 1.35 1.12 1.63 0.0017 1.25 1.02 1.52 0.029 Lipid Phosphatidylethanolamine 1‐palmitoyl‐2‐arachidonoyl‐GPE (16:0/20:4) 1.39 1.15 1.7 0.00059 1.31 1.08 1.6 0.0067 Lipid Phosphatidylethanolamine 1‐stearoyl‐2‐linoleoyl‐GPE (18:0/18:2) 1.36 1.13 1.65 0.0016 1.28 1.05 1.57 0.015 Lipid Phosphatidylethanolamine 1‐palmitoyl‐2‐oleoyl‐GPE (16:0/18:1) 1.45 1.19 1.75 0.00021 1.34 1.09 1.65 0.0051 Lipid Phosphatidylethanolamine 1‐palmitoyl‐2‐linoleoyl‐GPE (16:0/18:2) 1.39 1.15 1.68 0.00080 1.32 1.09 1.62 0.0044 Lipid Pregnenolone steroids Pregnenediol sulfate (C21H34O5S) 0.65 0.51 0.82 0.00026 0.64 0.51 0.81 0.00024 Lipid Pregnenolone steroids 17α‐hydroxypregnenolone 3‐sulfate 0.64 0.49 0.84 0.00090 0.6 0.46 0.79 0.00020 Lipid Pregnenolone steroids Pregnenolone sulfate 0.7 0.56 0.87 0.0011 0.68 0.54 0.84 0.00063 Lipid Pregnenolone steroids Pregnenetriol sulfate 0.66 0.52 0.84 0.00093 0.64 0.5 0.82 0.00036 Lipid Progestin steroids 5α‐pregnan‐3β, 20α‐diol monosulfate 0.7 0.57 0.85 0.00053 0.71 0.58 0.88 0.0015 Nucleotide Purine metabolism, adenine containing N6‐carbamoylthreonyladenosine 1.38 1.13 1.68 0.0021 1.34 1.08 1.65 0.0076 Nucleotide Purine metabolism, guanine containing 7‐methylguanine 1.42 1.16 1.72 0.00053 1.34 1.09 1.63 0.0046 Nucleotide Pyrimidine metabolism, cytidine containing N4‐acetylcytidine 1.36 1.12 1.65 0.0020 1.34 1.08 1.63 0.0054 Nucleotide Pyrimidine metabolism, uracil containing 5,6‐dihydrouridine 1.38 1.13 1.7 0.0019 1.38 1.12 1.7 0.0027 Peptide γ‐glutamyl amino acid γ‐glutamyltryptophan 1.38 1.14 1.68 0.0012 1.35 1.11 1.65 0.0031 Peptide γ‐glutamyl amino acid γ‐glutamylphenylalanine 1.52 1.23 1.86 0.000056 1.51 1.21 1.86 0.00026 Peptide γ‐glutamyl amino acid γ‐glutamyltyrosine 1.42 1.16 1.73 0.00050 1.38 1.12 1.68 0.0030 [58]Open in a new tab Only metabolites with false discovery rate <0.05 for the age‐ and sex‐adjusted analysis and P<0.05 for the multiple‐adjusted analysis are given. GPE indicates glycero‐3‐phosphoethanolamine; HR, hazard ratio; and SAM, S‐adenosylmethionine. Metabolites and Prediction of CVD A model with age, sex, and traditional cardiovascular risk factors (systolic blood pressure, low‐density lipoprotein and high‐density lipoprotein cholesterol, and current smoking) was developed, with a C index of 0.74 (95% CI, 0.70–0.78) for discrimination of subsequent CVD. A lasso regression identified the 5 metabolites of top importance for incident CVD (5‐methylthioadenosine, dimethylglycine, pregnenediol sulfate, imidazole propionate, and 1‐carboxyethyltyrosine). When those were used in a model together with age and sex, the model resulted in a C index of 0.77. When the 5 metabolites were added to the established risk factors, the C index increased by 4.0% (0.78 [95% CI, 0.74–0.81] versus 0.74 [95% CI, 0.70–0.78]) when compared with the model with traditional risk factors only (P=0.005; Figure [59]2). Figure 2. Area under the curve for traditional risk factors vs traditional risk factors plus metabolites on the outcome incident cardiovascular disease in the EpiHealth study. Figure 2 [60]Open in a new tab The ROC curve (old ROC area) for the traditional risk factors of importance (systolic blood pressure, low‐density lipoprotein and high‐density lipoprotein cholesterol, and current smoking) is given in red (denoted old ROC curve), whereas the curve in black (denoted new ROC curve) also includes the 5 metabolites 5‐methylthioadenosine, dimethylglycine, pregnenediol sulfate, imidazole propionate, and 1‐carboxyethyltyrosine. P=0.0054 for the difference between curves. ROC indicates receiver operating characteristic. When we performed a bootstrap analysis with 10 000 repetitions of the difference in C statistics between the traditional risk factors and the model that included 5 metabolites together with the traditional risk factors, the mean difference across the subsamples was close to the calculation in the total sample, with a 95% CI being highly significant (bootstrapped mean C statistic difference between models 0.0391 [95% CI, 0.0134–0.0649]; P=0.003). Metabolites Versus Subclinical Markers of CVD As can be seen in the overview in Figure [61]3, 35 of the 37 metabolites identified in EpiHealth were related to subclinical markers of CVD evaluated in the PIVUS study. Eleven of the 37 metabolites showed P<0.05 in relation to left ventricular mass index in the age‐ and sex‐adjusted analysis, but none of these metabolites were significantly associated with left ventricular mass index following adjustment for established cardiovascular risk factors (see Table [62]S3 for details). Thirty‐four of the metabolites were related to ejection fraction. Most of these relationships were inverse. Twenty‐nine of those associations with ejection fraction were statistically significant following multivariable adjustment. Twenty‐nine of the metabolites were related to left atrial diameter. Only 9 of those relationships versus left atrial diameter still showed P<0.05 following multivariable adjustment. Figure 3. Hierarchically clustered heat map of the relationships (given as regression coefficients) between the 37 metabolites found to be related to incident CVD and markers of subclinical CVD in the PIVUS study following adjustment for age and sex. Figure 3 [63]Open in a new tab Red filling indicates positive relationships, whereas blue filling indicates inverse relationships. A star denotes that the relationships showed P<0.05. The heat map was created in R 4.2 using the package pheatmap. CVD indicates cardiovascular disease; EF, ejection fraction; GPE, glycero‐3‐phosphoethanolamine; IMGSM, echogenicity of the intima‐media complex; IMT, carotid artery intima‐media thickness; LA, left atrial diameter; LVMI, left ventricular mass index; and PIVUS, Prospective Investigation of Vasculature in Uppsala Seniors. Four metabolites were related to intima‐media thickness, and 2 of those showed P<0.05 following multivariable adjustment (both inverse). Nine metabolites were related to echogenicity of the intima‐media complex, and 2 of those showed P<0.05 following multivariable adjustment. Of particular interest was 1‐carboxyethyltyrosine, being related to left atrial diameter, as well as inversely related to both ejection fraction and the echogenicity of the carotid artery (echogenicity of the intima‐media complex) also after adjustment for traditional cardiovascular risk factors. DISCUSSION The present study, which investigated almost 800 different nonxenobiotic metabolites, showed that 37 metabolites were related to an increased risk of subsequent CVD. As expected from previous studies in this field, these metabolites represent several different biochemical classes and biological pathways. The top 5 metabolites improved the discrimination of incident CVD when added to established cardiovascular risk factors. Most of these 37 metabolites were related to different markers of subclinical CVD in an independent sample. Of particular interest was 1‐carboxyethyltyrosine, being related to both cardiac performance and structural changes in the carotid artery wall. Comparison With the Literature As summarized in Table [64]1, a great number of metabolites have previously been linked to incident CVD. The most consistent findings have been for certain amino acids, such as branch‐chained and aromatic amino acids, acetylcarnitines, and certain lipid classes, such as lysophophatidylcholines and sphingomyelins. The present study highlighted the previously mentioned amino acids and phosphatidylethanolamines, but also found steroids, mainly from the pregnenolone pathway, and metabolites, from the nucleotide purine metabolism, to be linked to future CVD. Although this investigation and previous studies have found relationships between CVD and many different metabolites, the biological significance remains to be established. First, associations are not always causal. One way to evaluate causation is by Mendelian randomization using genetic information. However, although information on the genetic associations exists for some metabolites,[65] ^18 the genetic basis for the majority of the 790 metabolites evaluated in the present study does not exist. Another fact that hampers the use of Mendelian randomization in metabolomics research is that a certain genetic locus is often related to several metabolites in the same pathway or chemical class. Such pleiotropy violates the fundamental assumptions of the Mendelian randomization analysis. Thus, at present, it is hard to evaluate if the majority of the reported associations are causally related to CVD or not, but we hope that the identification of these metabolites and metabolic pathways could serve as inspiration to evaluate such associations in experimental models in which interventions could be made to evaluate causality. In a pathway enrichment analysis (using MetaboAnalyst; [66]www.metaboanalyst.cd) of the 37 metabolites of interest, no pathway was significantly enriched and, in most cases, only 1 metabolite was identified per pathway, suggesting that no physiological pathway was of major importance compared with the others. Thus, incident CVD is linked to many pathophysiological pathways, and the present metabolomics analysis could not identify a pathway of superior interest. Steroids from the pregnenolone pathway identified in the present study were generally inversely related to future CVD. This was somewhat surprising, given that pregnenolone metabolites have been reported to be elevated in patients with myocardial infarction.[67] ^19 One reason for increased levels of pregnenolone metabolites following an acute myocardial infarction might be activation of the hypothalamic–pituitary–adrenal axis caused by the stressful event.[68] ^20 However, in this case, the pregnenolone metabolites were evaluated years before the event, and it has been published that low levels of downstream metabolites, such as testosterone[69] ^21 and cortisol,[70] ^22 , [71]^23 are associated with future CVD or a high CVD risk. A recent experimental study showed that testosterone increased endothelial nitric oxide synthesis and activity of superoxide dismutase and reduced catalase activity, with potential vascular protective effects,[72] ^24 suggesting potential pathophysiological events at the molecular level by low levels of this androgen. It is well established that urate levels, being a breakdown product of purine metabolism, is a risk factor for CVD (see meta‐analysis in Li et al[73] ^25 ) A recent Mendelian randomization study has suggested this association to be causal.[74] ^26 There are several possible explanations whereby high urate levels could induce CVD. Urate could induce oxidative stress by upregulating xanthine oxidoreductase.[75] ^27 Urate could also increase proinflammatory cytokines and active platelets.[76] ^28 Whether or not these pathophysiological mechanisms are also valid for the purine metabolites identified in the present study is not known. However, the upstream nucleotide metabolism is not well studied in the context of CVD. Here, 4 of the 37 metabolites of interest belonged to purine or pyramidine metabolism. This is a novel and potentially important finding that merits further studies. Glucose reacts nonenzymatically with other metabolites, such as proteins, to form advanced glycated end products. This is seen with aging, but is also accelerated in diabetes. The metabolite 1‐carboxyethyltyrosine, being related to both cardiac performance and composition of the carotid artery wall, is such an example. Advanced glycated end products have mainly been linked to diabetes and its complications.[77] ^29 As reviewed in Barlovic et al[78] ^30 and Del Turco et al,[79] ^31 advanced glycated end products could induce inflammation and oxidative stress by binding to a receptor, receptor of advanced glycated end‐products (RAGE), and could also affect extracellular matrix composition. Elevated levels of RAGE have been linked to atherosclerosis in humans.[80] ^32 Whether 1‐carboxyethyltyrosine could influence myocardial function and the composition of the arterial wall by an advanced glycated end product/RAGE‐dependent mechanism remains to be established. Other aromatic amino acids, such as phenylananine and tryptophan, have previously been related to CVD,[81] ^33 , [82]^34 as found in the present study, but reports on tyrosine and CVD are scarce. In a nested case–control study, tyrosine levels were not different between subjects with future CVD and controls.[83] ^33 We demanded a metabolite to show P<0.05 also after adjustment for traditional risk factors, including BMI, to qualify as significantly related to incident CVD. The inclusion of BMI as a confounder is of great importance in this setting, because a large number of metabolites are related to obesity, as reviewed in Rangel‐Huerta et al.[84] ^35 However, we were not interested in only identifying a great number of BMI‐associated metabolites, and therefore we adjusted for BMI. This could be the reason why we were not able to replicate some of the findings in previous studies, such as lysophophatidylcholines and sphingomyelins. We are not aware of any other sample with a similarly high number of metabolites measured at the same platform and with sufficient cases to have an appropriate power that could be used for validation of the findings in the EpiHealth study, so our 37 candidate metabolites have to be reproduced by others in the future to be regarded as validated. However, we would expect that metabolites being related to future CVD to be related to some marker of subclinical CVD, because impairments in those markers usually precede an overt CVD event. We could see that almost all of our 37 metabolites of interest were related to at least 1 or 2 of the 5 markers of subclinical CVD in an independent sample. Although not a formal validation of the 37 metabolites, it is reassuring to see that the vast majority of the metabolites of interest are related markers of subclinical CVD in our supportive analysis in the PIVUS study, and therefore most likely are not mere chance findings. In the present study we conducted an analysis to evaluate if some metabolites could improve the discrimination of CVD on top of traditional CVD risk factors and found the top 5 metabolites to increase the C statistics by 3.9%. We are fully aware that this is most likely to be an overestimation of the effect of the metabolites, because the selection of metabolites and the C statistic test were performed in the same sample, but we see this test as an inspiration for future studies to improve risk prediction by metabolites. An alternative to validation in an independent sample is to split the present sample in 2 parts and then generate a list of metabolites of interest in 1 subsample and perform the C statistics test in the other subsample. However, because of the limited number of cases, we do not have the power to do so in a meaningful way. As an alternative, we performed a bootstrap analysis of the C statistic test in the total sample and found that the increase in the C statistics found in the traditional analysis was stable across 10 000 repetitive subsamples of the cohort, which would serve as an internal validation. In this study, we limited the number of metabolites to be used in the discrimination analysis to 5 because of the restricted number of CVD cases, although the lasso analysis disclosed 15 metabolites of interest. Thus, in future samples with a higher number of incident cases, it is likely that >5 metabolites could be evaluated to improve discrimination further. The major strength of the present study is the use of a metabolomic platform that made it possible to evaluate a great number of nonxenobiotic metabolites in the same sample. It is also a strength that we used the same metabolomic platform in another study in which we had measured several markers of subclinical CVD, although that study (PIVUS) did not have sufficient follow‐up time to generate the number of cases needed to evaluate incident CVD in a meaningful way. Some limitations have already been discussed above, and we should also acknowledge that we have been studying almost exclusively individuals with European ancestry, so future studies in other ethnic groups are warranted. The way to calculate ejection fraction from M‐mode in PIVUS is rather old fashioned but has been used since the first investigations in PIVUS back in 2001 to be comparable throughout this longitudinal study. If anything, an old‐fashioned technique would drive associations toward the null hypothesis and is not likely to produce false‐positive results. In the present study, we used a false discovery rate of <0.05 for the association adjusted for age and sex and P<0.05 for the following multiple‐adjusted analysis as the criteria of significance for a metabolite. This is a commonly used approach, but it must be emphasized that a couple of the 37 metabolites identified to be of interest could be false‐positive findings on multiple‐adjusted analysis, so these results should be taken with caution until reproduced by others. The plasma metabolome is the summary of metabolic processes in a great variety of tissues. If a metabolite is associated (either up‐ or downregulated) with a certain disease, it cannot convey whether the altered concentration is caused by an increased production or an increased extraction and which tissues are involved. Thus, findings of an altered plasma metabolome in a certain disease should mainly be taken as an indication that metabolic processes could be altered, but the more detailed knowledge of these processes must be verified in experimental studies or in more sophisticated studies in humans. The PIVUS sample has a mean age 10 years older than the EpiHealth sample. Ideally, the supportive information gained in the PIVUS study should be derived from a sample with the same age distribution, but we still find it reassuring that most metabolites found to be linked to incident CVD in EpiHealth are linked to deviations in subclinical markers of CVD, even if PIVUS is an older sample. In conclusion, several metabolites were discovered to be associated with subsequent CVD as well as with subclinical markers of CVD. A selection of those improved discrimination for the prediction of incident cardiovascular events when added to established cardiovascular risk factors. Sources of Funding The EpiHealth study is funded as a strategic research area by the Swedish government. PIVUS was funded by Uppsala University Hospital. Disclosures None. Supporting information Tables S1–S3 [85]Click here for additional data file.^ (265.9KB, pdf) Supplemental Material is available at [86]https://www.ahajournals.org/doi/suppl/10.1161/JAHA.122.026885 For Sources of Funding and Disclosures, see page 11. REFERENCES