Metabolomic and microbiome profiling reveals personalized risk factors for coronary artery disease

Abstract

Complex diseases, such as coronary artery disease (CAD), are often multifactorial, caused by multiple underlying pathological mechanisms. Here, to study the multifactorial nature of CAD, we performed comprehensive clinical and multi-omic profiling, including serum metabolomics and gut microbiome data, for 199 patients with acute coronary syndrome (ACS) recruited from two major Israeli hospitals, and validated these results in a geographically distinct cohort. ACS patients had distinct serum metabolome and gut microbial signatures as compared with control individuals, and were depleted in a previously unknown bacterial species of the Clostridiaceae family. This bacterial species was associated with levels of multiple circulating metabolites in control individuals, several of which have previously been linked to an increased risk of CAD. Metabolic deviations in ACS patients were found to be person specific with respect to their potential genetic or environmental origin, and to correlate with clinical parameters and cardiovascular outcomes. Moreover, metabolic aberrations in ACS patients linked to microbiome and diet were also observed to a lesser extent in control individuals with metabolic impairment, suggesting the involvement of these aberrations in earlier dysmetabolic phases preceding clinically overt CAD. Finally, a metabolomics-based model of body mass index (BMI) trained on the non-ACS cohort predicted higher-than-actual BMI when applied to ACS patients, and the excess BMI predictions independently correlated with both diabetes mellitus (DM) and CAD severity, as defined by the number of vessels involved. These results highlight the utility of the serum metabolome in understanding the basis of risk-factor heterogeneity in CAD.

Extended Data Fig. 1 |. Cohort selection and data acquisition pipeline.

This study includes a total of 199 participants with ACS and 970 non-ACS individuals. Each cell shows the number of individuals who were profiled for the corresponding omic platform indicated on the left. Colored bars connecting cells represent the number of overlapping individuals. For example, there were 170 non-ACS individuals that were profiled both for serum metabolomics using the Metabolon platform and for microbiome composition. The 156 samples of individuals with ACS that were profiled using the Metabolon platform are the first to be enrolled in this study. The 473 samples of non-ACS individuals that were profiled using the Metabolon platform, were profiled as part of our previous study (Bar et al. 2020). All samples of individuals with ACS (n = 191) and of non-ACS individuals (n = 961) for which we had available serum obtained during their recruitment, were profiled using the Nightingale platform. While microbiome data were available for all individuals with and without ACS, we only considered samples for which the collection, DNA extraction and sequencing procedures were identical (n = 199 for ACS; n = 340 for non-ACS). Differential abundance analysis was performed based on subcohorts resulting from 1:1 matching for age, sex, BMI, DM, and smoking status. ACS, Acute Corony Syndrome; BMI Body Mass Index; DM, Diabetes Mellitius.

Extended Data Fig. 2 |. Breakdown of ACS serum metabolomics pattern by the origin of metabolites and biological pathway.

(a) Box plots (y axis: centre, median; box, IQR; whiskers, 1.5*IQR) showing the explained variance (EV) of metabolites by different feature groups (x-axis) separated to metabolites enriched in ACS (N = 175; orange) and enriched in matched non-ACS controls (N = 358; blue). (b) EV of metabolites (y axis: center, median; box, IQR; whiskers, 1.5*IQR box, IQR; whiskers, 1.5*IQR) by their super pathways (x axis) separated to metabolites enriched in ACS (orange) and enriched in matched non-ACS controls (blue). The number of metabolites per group is shown below each box. Trad., Traditional; C&V, cofactors and vitamins.

Extended Data Fig. 3 |. Depletion of ACS-related bacteria SGB 4712 replicates in an independent validation cohort.

(a) Box plots showing the relative abundance of the unknown bacterial species SGB 4712 (y-axis: centre, median; box, IQR; whiskers, 1.5*IQR; log scaled) in our ACS and matched controls (x-axis; n = 80 each). The P-value shown is computed using the Mann-Whitney U test. (b) Relative abundance of the unknown bacterial species SGB 4712 (y - axis: centre, median; box, IQR; whiskers, 1.5*IQR; log scaled) in four groups from the MetaCardis validation cohort (x-axis; HC, healthy controls, n = 275; MMC, metabolically matched controls, n = 218; UMCC, untreated metabolically compromised controls, n = 211; IHD, ischaemic heart disease, n = 319). The P - value shown is computed using the Mann-Whitney U test. r.a., relative abundance.

Extended Data Fig. 4 |. Clinical data correlates with metabolic deviations.

(a) The mean weighted R2 of genetics for ACS-enriched metabolites (y - axis) versus chronological age (x - axis). Dots are colored by sex. Spearman correlation is computed over all samples (Spearman ρ = 0.18; p = 0.032). (b) The mean weighted R2 of traditional risk factors for ACS-depleted metabolites (y axis) versus chronological age (x axis; Spearman ρ = 0.33; p = 7.7 ∙ 10⁻⁵). (c) The mean weighted R2 of genetics for ACS-enriched metabolites (y axis: centre, median; box, IQR; whiskers, 1.5*IQR) in ACS patients who had a combined CVD outcome (defined as either: acute myocardial infarction, acute stroke, unplanned PCI, or cardiovascular-related death; x axis) versus not (two-sided Mann-Whitney U, p = 0.002).

Extended Data Fig. 5 |. Replication of higher predicted BMI in ACS individuals based on NMR metabolomics.

Figure panels refer to results of serum metabolomics-based prediction model of BMI trained in a non-ACS control cohort (n = 763) and evaluated on held-out test sets consisting of both controls (n = 179) and individuals with ACS (n = 179; Methods). (a) Measured (x axis) versus predicted (y-axis) BMI for both controls (blue) and ACS (orange) individuals. Line and shaded coloring represent the fitting of a linear model and the 95% confidence interval. (b) Difference between predicted and measured BMI (y axis: centre, median; box, IQR; whiskers, 1.5*IQR) of individuals, binned into three BMI groups (<25, 25–30, >30; x - axis). The P - values shown are computed using the Mann-Whitney U test. BMI, body mass index.

Extended Data Fig. 6 |. Replication of higher predicted BMI in IHD individuals in the MetaCardis study.

Figure panels refer to results of serum metabolomics-based prediction model of BMI trained in a cohort of individuals without IHD and evaluated on held-out test sets consisting of both 319 IHD and 319 non-IHD individuals (Methods). (a) Measured (x axis) versus predicted (y axis) BMI for healthy controls (HC; blue), metabolically matched controls (MMC; blue), and untreated metabolically compromised controls (UMCC; blue), and individuals with ischaemic heart disease (IHD; orange). Line and shaded coloring represent the fitting of a linear model and the 95% confidence interval. (b) Difference between predicted and measured BMI (y axis: centre, median; box, IQR; whiskers, 1.5*IQR) of individuals, binned into three BMI groups (<25, 25–30, >30; x axis). (c) Same as in (b) only for individuals with IHD, and each bin is separated into normoglycemic versus T2DM patients. Higher predicted BMI is associated with an increased incidence of T2DM (OR = 1.13, 95% CI =1.05–1.22, p = 0.002; a logistic regression model adjusted for BMI and age; Methods). The p values shown are computed using the Mann-Whitney U test. BMI, body mass index; T2DM, type 2 diabetes mellitus; OR, odds ratio; CI, confidence interval.

Fig. 1 |. Microbiome and serum metabolomics signatures of ACS.

a, A circular heatmap showing the top 200 metabolites that differ significantly between ACS and non-ACS control cohorts, matched for age, sex, BMI, smoking status and DM (Methods). Each slice represents a single metabolite, with its name indicated around the outer layer of the chart. The color code is indicated at the top of the panel. The outermost layer indicates the –log₁₀(P value) (a logistic regression model adjusted for age and sex; Methods) for the enrichment of metabolites between the two cohorts, where orange/blue colors correspond to metabolites enriched/depleted in the ACS cohort. The next layer shows the –log₁₀(P value) (a logistic regression model adjusted for age, sex and BMI) for each metabolite in diabetic versus normoglycemic ACS patients. Here, black/red colors correspond to metabolites enriched/depleted in diabetic patients. The next four layers show the EV of each metabolite by feature groups, as previously estimated. The metabolites are first sorted by their categories, as indicated in the inner layer, and then by their directional enrichment between the two cohorts. b, The distribution of average phylum abundance (normalized to sum to 1.0) among non-ACS and ACS participants (unmatched controls, n = 335; matched controls, n = 64; unmatched ACS, n = 199; matched ACS, n = 64). P values refer to comparisons between the matched cohorts (Kruskal–Wallis). c, A circular heatmap showing 15 metabolites that significantly correlate with the relative abundance of SGB 4712 (FDR < 1%, Spearman) in the control cohort. Each slice represents a single metabolite, with its name indicated at the outermost layer of the chart. The color code of each layer is indicated at the top of the panel. The outermost layer indicates the –log₁₀(p value) (two-sided Mann–Whitney U-test) for the enrichment of metabolites between the two cohorts, where orange and blue colors correspond to metabolites enriched and depleted, respectively, in the ACS cohort. The next layer shows the Spearman correlation between each metabolite and the relative abundance of SGB 4712. The metabolites are sorted by their biological pathways, as indicated in the inner layer. Trad., Traditional; C&V, cofactors and vitamins; PCM, partially characterized molecules; P-C-G*, p-cresol-glucuronide*; PAGln, phenylacetylglutamine; DMTPA*, 2,3-dihydroxy-5-methylthio-4-pentenoate (DMTPA)*.

Fig. 2 |. Metabolic deviations explained by potential determinants and correlate with clinical parameters.

a,b, Density plots showing the distribution of ACS participants (y axis) versus the mean weighted R² of potential determinants (microbiome, diet, traditional risk factors or genetics) for metabolites (x axis); metabolites enriched in ACS (a); metabolites depleted in ACS (b). c, Heatmap showing the standardized mean weighted R² of potential determinants for ACS-enriched and ACS-depleted metabolites (x axis) for a subgroup of 17 ACS patients (y axis). d, Radar chart showing the standardized mean weighted R² of potential determinants for the two ACS subjects who are marked with blue and orange outlines in c. e,f, Box plots (y axis: center, median; box, interquartile range (IQR); whiskers, 1.5×IQR) showing the mean weighted R² of potential determinants for metabolites in two groups of ACS patients. e, The mean weighted R² of traditional risk factors for ACS-depleted metabolites compared between ACS patients that had a combined CVD outcome within 12 months (n = 13) versus not (n = 122) (P = 0.005; two-sided Mann–Whitney U-test). f, the mean weighted R² of microbiome for ACS-enriched metabolites compared between ACS patients with T2DM (n = 39) versus normoglycemic ACS patients (n = 96) (P = 0.003; two-sided Mann–Whitney U-test).

Fig. 3 |. Microbiome and diet-related metabolic deviations are present in control participants with metabolic impairment.

a–d, Metabolic deviation scores attributed to diet (a), microbiome (b), traditional risk factors (c) and genetics (d) computed for three subgroups: (1) ACS individuals (n = 135) versus non-ACS controls matched for age, sex and BMI (orange); (2) non-ACS controls with metabolic impairment (defined as either: diagnosed with T2DM, hypertension or dyslipidemia, or BMI > 35; n = 102) versus other non-ACS controls matched for age, sex and BMI (blue); and (3) a random set of non-ACS individuals (n = 132) versus other non-ACS controls matched for age, sex and BMI (gray). Violin plot-line elements: center line, median; thick limits, upper and lower quartiles; whiskers, 1.5×IQR. P values shown are computed using the two-sided Mann–Whitney U-test.

Fig. 4 |. A metabolomics-based model of BMI predicts higher BMI in ACS patients and correlates with disease severity.

a, Measured (x axis) versus predicted (y axis) BMI for controls (n = 156; blue) and ACS (n = 156; orange) individuals. Line and shaded coloring represent the fitting of a linear model and the 95% confidence interval. b, Difference between predicted and measured BMI (y axis) of individuals, binned into three BMI groups (<25, 25–30, >30; x axis). P values shown are computed using the two-sided Man–Whitney U-test. c, Same as in b, only for ACS participants, and each bin is separated into normoglycemic (n = 111) versus T2DM patients (n = 44). Higher predicted BMI was associated with an increased incidence of T2DM (odds ratio (OR) = 1.48; 95% confidence interval (CI) = 1.09–2.01; P = 0.01; a logistic regression model adjusted for BMI and age; Methods). d, Same as in b, only for ACS participants; each bin is separated into patients with one vessel (n = 71) versus three vessels involved (n = 36). Higher predicted BMI was associated with an increased incidence of having three vessels involved (OR = 1.5; 95% CI = 1.03–2.19; P = 0.03; a logistic regression model adjusted for BMI, age and T2DM; Methods). Box-plot elements: center, median; box, IQR; whiskers, 1.5×IQR.