Networks of gut bacteria relate to cardiovascular disease in a multi-ethnic population: the HELIUS study

Abstract

Aims

Gut microbiota have been linked to blood lipid levels and cardiovascular diseases (CVDs). The composition and abundance of gut microbiota trophic networks differ between ethnicities. We aim to evaluate the relationship between gut microbiotal trophic networks and CVD phenotypes.

Methods and results

We included cross-sectional data from 3860 individuals without CVD history from 6 ethnicities living in the Amsterdam region participating in the prospective Healthy Life in Urban Setting (HELIUS) study. Genetic variants were genotyped, faecal gut microbiota were profiled, and blood and anthropometric parameters were measured. A machine learning approach was used to assess the relationship between CVD risk (Framingham score) and gut microbiota stratified by ethnicity. Potential causal relationships between gut microbiota composition and CVD were inferred by performing two-sample Mendelian randomization with hard CVD events from the Pan-UK Biobank and microbiome genome-wide association studies summary data from a subset of the HELIUS cohort (n = 4117). Microbial taxa identified to be associated with CVD by machine learning and Mendelian randomization were often ethnic-specific, but some concordance across ethnicities was found. The microbes Akkermansia muciniphila and Ruminococcaceae UCG-002 were protective against ischaemic heart disease in African-Surinamese and Moroccans, respectively. We identified a strong inverse association between blood lipids, CVD risk, and the combined abundance of the correlated microbes Christensenellaceae–Methanobrevibacter–Ruminococcaceae (CMR). The CMR cluster was also identified in two independent cohorts and the association with triglycerides was replicated.

Conclusion

Certain gut microbes can have a potentially causal relationship with CVD events, with possible ethnic-specific effects. We identified a trophic network centred around Christensenellaceae, Methanobrevibacter, and various Ruminococcaceae, frequently lacking in South-Asian Surinamese, to be protective against CVD risk and associated with low triglyceride levels.

Graphical Abstract

Graphical Abstract.

Time for primary review: 46 days

1. Introduction

Modifiable risk factors such as hypertension, dyslipidaemia, obesity, and hyperglycaemia have significant impact on cardiovascular disease (CVD) risk.^1,2 Recent evidence has shown that these risk factors are influenced by the gut microbiome³ or gut-derived metabolites, including short-chain fatty acids (SCFA).⁴ However, the gut microbiome and its metabolites may also directly affect CVD risk. For example, dietary ingested choline or L-carnitine is processed by the gut microbiota into trymethylamine, further metabolized into trimethylamine N–oxide, whose concentrations vary according to diet and microbiota composition, and is associated with major adverse cardiovascular events.^5,6

While diet is one of the most important determinants of the gut microbiota composition,⁷ after correction for diet, specific differences in the gut microbiota composition between ethnicities remain, as seen in a multi-ethnic cross-sectional study of the inhabitants of Amsterdam, The Netherlands.⁷ This study minimized geographical, environmental, and sample processing influence on gut microbiota results, and suggests that other factors related to ethnicity contribute to differences in the gut microbiota composition.

Changes in gut microbiota composition related to genetic variants in the host can also affect disease phenotypes, as already described for inflammatory bowel diseases.^8,9 Mendelian randomization (MR) studies have been used to infer causality between gut microbiota and metabolic diseases such as Type 2 diabetes through genetic variants that promote the production capacity of gut microbiota for SCFA.¹⁰ In their pioneering study, Sanna et al.¹⁰ identified several bacteria able to produce the SCFA butyrate to have a causative influence on glucose metabolism. A causal role for gut microbiota in other CVD risk factors such as blood pressure and lipid levels has also been established by an MR approach.¹¹ However, as a single bacterium is part of a complex ecosystem in the gut, recent literature suggests that rather than individual bacteria, groups of co-dependent bacteria may be the driving force behind beneficial pathways via cross-feeding and syntrophic interactions. These co-dependent networks are referred to as bacterial guilds or trophic networks and may influence host metabolism as they alter the availability of substrates for themselves and for their host.

Therefore, we hypothesized that ethnic-specific gut microbiota compositions affect CVD, and interrogated this hypothesis using a multi-layered approach. To identify microbes associated with CVD risk in our multi-ethnic population, we performed a machine learning analysis. Subsequently, we inferred causality between single taxa and hard CVD outcomes by performing two-sample Mendelian randomization in a multi-ethnic cohort with outcome data from the UK Biobank.¹⁰ Finally, we evaluated the microbes identified by machine learning and Mendelian randomization from a trophic network perspective to provide greater biological insights into the association between CVD and the gut microbiome composition.

2. Methods

2.1. Study population

The primary discovery cohort for this study was the Healthy Life in Urban Setting (HELIUS) study in Amsterdam, The Netherlands, as previously described.^12,13 In short, in this prospective cohort study, health differences between the six most common local ethnicities, namely, Dutch, African-Surinamese, South-Asian Surinamese, Turkish, Moroccan, and Ghanaian backgrounds, are evaluated. Inhabitants of Amsterdam between the ages of 18 and 70 years were randomly sampled, stratified by ethnicity, through the municipality registry of Amsterdam. Ethnicity was defined based on the country of birth as well as the parents’ country of birth, although for the African- and South-Asian Surinamese populations self-reported ethnicity was used.¹³ After a positive response to a written invitation, the subjects received a confirmation letter of an appointment for a physical examination and a digital or paper version of the questionnaire (depending on the preference of the subject) to fill out at home. Of 24 789 HELIUS participants, 22 165 participants underwent a physical examination and completed the questionnaire, 10 283 participants were genotyped, and 6048 handed in a faeces sample. A total of 4117 subjects were selected and used in this study based on the availability of the clinical, microbiome, and genetic data. For machine learning analysis, 3860 subjects with faecal samples but without CVD history were selected. The HELIUS study was approved by the institutional review board of the Academic Medical Centre (AMC) Amsterdam, The Netherlands, and was conducted in accordance with the Declaration of Helsinki (6th, 7th revisions). All participants provided written informed consent. Researchers who wish to conduct analyses on this cohort can submit a proposal as outlined at http://www.heliusstudy.nl/en/researchers/collaboration, by email: heliuscoordinator@amsterdamumc.nl. The proposals will be examined by the HELIUS Executive Board. If the proposals comply with the ethical guidelines, general objectives, and informed consent forms of the HELIUS study, access will be granted.

2.2. Blood and faeces processing

Anthropometric data and blood samples were collected during the first study visit along with information on drug use. The participants were asked to collect faeces prior to or during the study visit, as previously described.¹³ Exclusion criteria for faeces sampling included having diarrhoea 1 week before the study visit or the use of antibiotics three months prior (or unknown antibiotics use). If samples were collected more than 6 h before the study visit, the participants were instructed to store the sample in a freezer. At the study visit, the samples were frozen at −20°C for up to 1 day and subsequently at −80° in the permanent freezer. DNA was extracted at the Wallenberg Laboratory (Sahlgrenska Academy at University of Gothenburg, Sweden) with a repeated bead-beating protocol.¹⁴ The microbiota composition of faecal samples was assessed by sequencing the V4 region of the 16S rRNA gene on an Illumina MiSeq (Illumina RTA v1.17.28; MCS v2.5, San Diego, CA, USA) chip using 515F and 806R primers designed for dual-indexing¹⁵ and the V2 Illumina kit (2 × 250 bp paired-end reads). Polymerase chain reaction (PCR) and pre-processing of raw sequencing data were performed, as previously described.²

2.3. Microbial data preparation

The faecal samples and 16S sequencing data processing has been previously described.^2,16 In short, faecal samples were collected and stored at −80°C, before total genomic DNA was extracted and paired-end sequencing of the V4 region of the 16S rRNA gene was carried out with an Illumina MiSeq system (2 × 250 bp). The resulting sequencing reads were merged and quality filtered, before the contigs were dereplicated and the UNOISE3 algorithm was used to denoise unique sequences to generate Amplicon Sequence Variants (ASVs). The assign Taxonomy function from the DADA2 R package (v. 1.12.1) was used to assign taxonomy using the SILVA (v. 132) database. For the genome-wide association studies (GWAS), the gut microbiome sequencing data were summarized to phylum, family, genus, or species level, and individual ASVs. The taxa were then filtered to keep taxa with >20 counts in >5% of the subjects. The findCorrelation function from the R caret package v. 6.0-90 was used to remove highly correlated taxa, to reduce redundancy. The final microbial dataset used for GWAS consisted of 401 taxa (237 ASVs, 60 species, 21 families, 76 genera, 7 phyla). If <5% of subjects had zero abundance, a pseudocount of 1 was added and the data were log₁₀ transformed. Alternatively, if >5% of subjects had zero abundance, the data were encoded into a presence/absence pattern and analysed as a binary trait. Detailed methods have previously been published.^16,17 For the machine learning approach and trophic network analysis, the microbes were kept as ASVs but were filtered by abundance to keep the microbes with the highest average abundance. The sequencing data generated in this study have been deposited in the European Genome-Phenome Archive database under the accession code EGAD00001004106.

2.4. Genotype data

The Illumina GSA 24v1-0 array was used to genotype participants at the Erasmus MC Human Genomic Facility (Rotterdam, The Netherlands). GenomeStudio software was used for genotyping of the array according to in-house protocols from the Human Genomic Facility. The Sanger Imputation Service was used to impute autosomal chromosomes according to instructions (https://imputation.sanger.ac.uk/). For phasing, EAGLE2 was used and the Positional Burrows-Wheeler Transform method with the HAPLOTYPE Reference Consortium (release 1.1) was used, including the 1000 Genomes. After imputation, single nucleotide polymorphisms (SNPs) with an INFO score ≤0.3 were removed. The SNPs were annotated with rs numbers if possible; if this was not known, the chromosomal location was used for annotation based on GRCh37 coordinates. Detailed methods including details on further quality control and genotype-to-microbiome associations have previously been published.^16,18 Relevant summary statistics can be found in Supplementary material online, Table S1.

2.5. Framingham score and CVD outcomes

The Framingham score was used to assess CVD risk in our cohort. This score predicts the 10-year risk to develop CVDs in subjects who are free of CVD and cancer at baseline. This score fitted best with our relatively young population. Other scores such as the European SCORE¹⁹ and Pooled Cohort Studies Equations²⁰ were similar in our population with correlation test R² ranges between 0.79 and 0.98 (see Supplementary material online, Figure S1).

This score uses the following parameters: sex, age, systolic blood pressure, use of antihypertensive medication, smoking (defined by self-report and categorized as currently smoking vs. non-smoking), diabetes mellitus (defined as a fasting glucose ≥7 mmol/L or use of glucose-lowering medication), total cholesterol, and HDL cholesterol levels.^21,22

For the two-sample Mendelian randomizations, summary statistics of hard cardiovascular outcomes were obtained from the multi-ethnic Pan-UK Biobank (Pan-UKBB) cohort (https://pan.ukbb.broadinstitute.org, 2020). Three phenotypes were relevant to our study: ischaemic heart disease (phecode: 411), myocardial infarction (phecode: 411.2), and coronary atherosclerosis (phecode: 411.4). These phenotypes are composites of different International Classification of Diseases and related health problems (ICD) codes into a single disease endpoint (see Supplementary material online, Table S6). This study used participants from European, Central/South Asian, African, and Middle Eastern ancestry from the Pan-UKBB.

2.6. Machine learning

Continuous outcomes such as the Framingham score were estimated with a machine learning approach using a regression predictor, for N = 3860 participants. For this, the gradient boosted trees method XGBoost (version 0.90) was used. The participants with a reported CVD event in their medical history were excluded. Nested cross-validation was performed to avoid overfitting and ensure robustness of results, which was employed over 100 iterations on a completely reshuffled dataset to improve robustness, and the average predictive performance was reported as root mean squared error (RMSE). We have used the stability selection approach as previously described,²³ to ensure reliability of the findings. We also use a specialized regularization strategy that makes our methodology applicable to the high-dimensional regime.

Machine learning was performed in Python (version 3.7.4) using the following packages: pandas (version 0.25.1), numpy (version 1.16.4), and scikit-learn (version 0.21.2); for hyperparameters and more detailed methods, see Supplementary material online, Methods.

2.7. Mendelian randomization

In order to infer potential causal relationships between bacteria as exposure and CVD outcomes, we used the inverse variance weighted (IVW) method;²⁴ however, if only one SNP was available, the Wald ratio was used, as a two-sample Mendelian randomization analysis of the summary association statistics. Mendelian randomization relies on three core assumptions: (i) the genetic instrument variables are related to the risk factor; (ii) the confounders are not related to the exposure, outcome, and genetic instruments; and (iii) there is no relationship between the genetic instrument and outcome other than through the risk factor. A threshold of P < 1 × 10⁻⁸ was initially used to select SNPs associated with the exposure and a second threshold of P < 1 × 10⁻⁵ was used to select variants associated to the exposure to avoid reliance on individual SNPs during Mendelian randomization, as previously described.¹⁰ Only SNPs with an F statistic >10 in regard to the association with the exposure were used. These thresholds strengthen the first assumption of Mendelian randomization analysis. The GWAS on the gut microbiome was performed with appropriate confounder correction [age, sex body mass index (BMI), top 10 genetic principal components, metformin use, proton-pump inhibitor use, educational level, and 16S rRNA sequencing run] to reduce the risk of confounded MR results.¹⁶ Finally, the third assumption cannot be formally tested although it is important to consider in the interpretation of the results. Clumping was performed using PLINK v. 2.0 with default parameters. This linkage disequilibrium-reduced SNP set was used for the MR analysis. Furthermore, MR Egger, leave-one-out, and weighted median were used as sensitivity analyses approaches for significant results. Two-sample MR was performed by using the ‘TwoSampleMR’ R library (https://github.com/MRCIEU/TwoSampleMR) v. 0.5.5. The Pan-UKBB subgroups were matched to the HELIUS groups as follows: the ‘Meta’ group in the Pan-UKBB with the total group in HELIUS, the ‘EUR’ group with the Dutch and Turkish separately in HELIUS, the ‘AFR’ group with the African-Surinamese and Ghanaian groups in HELIUS together as well as separately, ‘CSA’ with the South-Asian Surinamese in HELIUS, and the ‘MID’ with the Moroccan group in HELIUS. This is in line with previous results that showed that African populations from the HELIUS cohort are genetically similar to Africans in the 1000 genome cohort; the same holds true for the Dutch and South-Asian with regard to their respective geographical origins.¹⁸ Similarly, UKBB ethnicities were also matched to the 1000 genome cohorts, which also confirmed the geographical origin.²⁵ Reported P-values are nominal and should be interpreted as explorative results. For genetical instruments, see Supplementary material online, Tables S3–S5. Beta coefficients correspond to the log of the odds ratio (OR) per standard deviation increase.

To confirm that causal relationships were not bidirectional, we performed bidirectional MR using CVD outcomes as exposure and microbiome data as outcome.

2.8. Co-occurring networks

To identify correlated microbial networks with potential causative effect on CVD, we selected the 30 bacteria with the strongest Spearman’s correlation to bacteria identified from the MR analyses (Figure 1). These groups of bacteria were summed and used as one composite signal (SumTn). Subsequently, a correlation analysis was performed for the most important cardiovascular and metabolic parameters and the SumTn. An alternative approach for defining composite signals was to conduct hierarchical clustering (Euclidean distance, complete linkage) of Spearman ρ coefficients of the 250 most abundant ASVs. Specifically, several bacteria causally linked to CVD were found to be part of the Christensenellaceae–Methanobrevibacter–Ruminococcaceae (CMR) cluster. This trophic network was subsequently also summed up together and used as one composite signal CMR and correlated with important cardiovascular and metabolic parameters. The CMR cluster was validated in the METSIM cohort²⁶ (a cohort of 531 Finish men) and in a Korean cohort²⁷ (623 healthy and hypertensive Korean men and women). We used the same methods for clustering in all three cohorts.

Figure 1.

Study design. To study the relation between microbes and CVD risk, we predicted the Framingham score with faecal microbiome data using a machine learning approach. To study causal relations between gut microbes and CVD events, we performed a two-step MR approach. For this, summary statistics from a GWAS in the multi-ethnic HELIUS cohort with microbes and in the multi-ethnic Pan-UKBB cohort were used. Bacteria correlated to bacteria deemed to be causally related with CVD events based on Mendelian randomization were compared with results from the machine learning approach. As a third method, an unbiased hierarchical clustering approach was used to identify co-occurring bacterial trophic networks. GWAS, genome-wide association study. Images from Flaticon.com.

2.9. Statistical analysis

A GWAS of genotype and microbiome data was performed using PLINK v. 2.0, considering the following parameters: -glm and -covar-variance-standardize. The association was adjusted for covariates (top 10 genetic principal components calculated by pca in PLINK v. 2.0), age, sex, BMI, proton-pump inhibitor use, and metformin use, since all these parameters are known to modulate the gut microbiota composition.²⁸ The GWAS was conducted in both the total cohort and separately per ethnic subset. Further details on the microbiome GWAS can be found in the previously published study.¹⁶ Spearman ρ correlations were used to evaluate the associations among microbes and metabolic parameters. Mann–Whitney U testing was used for non-parametric comparison of groups with normal or elevated triglyceride levels. Linear regression was used to assess the dose-dependent effect of CMR abundance and elevated triglyceride prevalence and to model interaction, which was compared with a non-interaction model using an analysis of variance (ANOVA). The false discovery rate (FDR) method was used to correct for multiple correlation comparisons and q < 0.05 was considered statistically significant. FDR correction was applied within each ethnicity. All other statistical analyses were performed in R (v 3.6.1) and ggplot2 was used for data visualization. One author had full access to all the data in the study and takes responsibility for its integrity and the data analysis.

3. Results

3.1. Baseline characteristics

Table 1 displays clinical characteristics of patients stratified by ethnic groups. The median age was 51 years, with the Dutch and South-Asian Surinamese being the oldest (median age 54 years), while the Moroccan and Turkish groups were the youngest (median age 46 years). The Turkish group had the highest BMI, while the Dutch had the lowest. African-Surinamese and Moroccans had the highest and lowest Framingham scores, respectively. Most antihypertensive usage was found in the South-Asian Surinamese group with 20.2% and the least in the Moroccan group with 1.5% (Table 1 and Supplementary material online, Table S2).

Table 1.

Baseline characteristics in the HELIUS cohort, stratified per ethnicity

	Overall	Dutch	South.Asian.Surinamese	African.Surinamese	Ghanaian	Turkish	Moroccan
N	4117	1159	527	1021	400	416	594
Age [median (IQR)]	51.00 [42.00, 58.00]	54.00 [43.00, 62.00]	54.00 [46.00, 60.00]	53.00 [47.00, 59.00]	49.00 [43.00, 55.00]	46.00 [36.00, 51.00]	46.00 [37.00, 54.00]
Sex: female (%)	2175 (52.8)	571 (49.3)	273 (51.8)	617 (60.4)	231 (57.8)	208 (50.0)	275 (46.3)
Weight [median (IQR)]	76.70 [67.45, 86.76]	77.10 [67.65, 87.05]	70.20 [62.20, 79.70]	77.80 [68.64, 88.80]	76.03 [68.38, 85.66]	77.95 [68.61, 88.75]	77.68 [69.50, 88.46]
BMI, kg/m2 [median (IQR)]	26.66 [23.84, 30.15]	24.83 [22.33, 27.80]	26.18 [23.66, 29.07]	27.45 [24.46, 31.54]	27.92 [25.14, 30.96]	28.67 [25.60, 31.74]	27.56 [25.10, 31.00]
Waist circumference, cm [mean (SD)]	94.05 (13.03)	92.09 (13.22)	93.76 (12.47)	94.46 (13.74)	94.07 (11.26)	96.17 (12.99)	95.92 (12.54)
Systolic blood pressure, mmHg [median (IQR)]	127.50 [116.50, 140.00]	125.00 [115.00, 137.00]	130.50 [118.50, 143.00]	131.50 [120.00, 143.50]	135.75 [125.00, 148.00]	122.50 [112.50, 132.00]	122.50 [113.00, 134.88]
Fasting glucose, mmol/L [median (IQR)]	5.30 [5.00, 5.80]	5.30 [5.00, 5.70]	5.60 [5.10, 6.40]	5.30 [4.90, 5.80]	5.20 [4.80, 5.70]	5.30 [5.00, 5.70]	5.40 [5.00, 5.90]
HbA1c, mmol/mol [median (IQR)]	38.00 [35.00, 42.00]	37.00 [34.00, 39.00]	41.00 [37.00, 47.00]	39.00 [36.00, 43.00]	39.00 [34.00, 43.00]	37.00 [34.00, 41.00]	37.00 [34.00, 41.00]
Total cholesterol, mmol/L [median (IQR)]	4.96 [4.31, 5.67]	5.17 [4.54, 5.89]	4.91 [4.21, 5.70]	4.89 [4.25, 5.61]	4.96 [4.30, 5.63]	4.90 [4.26, 5.55]	4.74 [4.12, 5.37]
HDL-c, mmol/L [median (IQR)]	1.39 [1.13, 1.71]	1.49 [1.21, 1.81]	1.26 [1.04, 1.49]	1.45 [1.20, 1.78]	1.59 [1.31, 1.89]	1.21 [1.02, 1.49]	1.26 [1.03, 1.52]
Triglycerides, mmol/L [median (IQR)]	0.87 [0.59, 1.27]	0.92 [0.63, 1.32]	1.07 [0.73, 1.51]	0.76 [0.54, 1.05]	0.63 [0.46, 0.89]	1.02 [0.69, 1.57]	0.92 [0.63, 1.33]
LDL-c, mmol/L [median (IQR)]	3.05 [2.45, 3.67]	3.16 [2.55, 3.80]	3.02 [2.39, 3.75]	3.00 [2.36, 3.61]	2.99 [2.42, 3.67]	3.06 [2.46, 3.59]	2.98 [2.42, 3.52]
Creatinine, µmol/L [median (IQR)]	73.00 [63.00, 84.00]	74.00 [66.00, 84.00]	72.00 [62.00, 86.00]	76.00 [66.00, 88.00]	77.00 [66.75, 91.00]	66.00 [57.00, 78.00]	67.00 [56.00, 77.00]
Framingham score [mean (SD)]	9.77 (5.92)	9.67 (5.95)	11.99 (6.13)	10.70 (5.76)	9.06 (4.79)	8.10 (5.24)	7.72 (6.04)
Antihypertensive use: no (%)	3772 (91.6)	1111 (95.9)	480 (91.1)	873 (85.5)	319 (79.8)	404 (97.1)	585 (98.5)
Statin use: no (%)	3615 (87.8)	1042 (89.9)	384 (72.9)	913 (89.4)	364 (91.0)	369 (88.7)	543 (91.4)

3.2. Association between gut microbiota and CVD risk

As overt CVD outcome data were not available in the HELIUS study, we used the Framingham score as a surrogate marker for CVD to study the relationship between CVD and the gut microbiome. We performed a gradient boosted machine learning analysis for regression predictive modelling on all subjects together (total group) and on ethnic subgroups separately to predict the Framingham score based on the gut microbiota composition. The total group (RMSE: 0.386 ± 0.006, P = 1.0 × 10⁻⁴n = 3860), Dutch (RMSE: 0.501 ± 0.018, P = 1.0 × 10⁻⁴, n = 1196), and African-Surinamese (RMSE: 0.514 ± 0.018, P = 9.2 × 10⁻³n = 998, Figure 2) had the most relevant microbiome profiles for Framingham score estimation. The microbiome composition of the Moroccan (RMSE: 0.458 ± 0.034, P = 0.09, n = 489), Figure 3, and South-Asian Surinamese (RMSE: 0.582 ± 0.059, P < 0.05 n = 460) subgroup had similar relevance value for CVD risk. For the Ghanaian (RMSE: 0.550 ± 0.045, P = 0.21, n = 383) and Turkish (RMSE: 0.522 ± 0.047, P = 0.23 n = 334) groups, the microbiome was not found to be relevant for the Framingham score. The 30 most contributing features differed between ethnicities (see Supplementary material online, Figures S2–S5, for relative feature importance).

Figure 2.

Results for the African-Surinamese subgroup. (A) The 30 most important features to predict Framingham score based on the faecal gut microbiota composition. *The bacteria were present in the SumTn of A. muciniphila, indicating that this bacterium is part of the trophic network that has a protective effect against ischaemic heart disease (n = 998). (B) Scatter plot of instruments with sensitivity analyses. (C) Forestplot with the effect of all SNPs used for the Mendelian randomization for A. muciniphila. IVW, inverse variance weighted.

Figure 3.

Results for the Moroccan subgroup. (A) The 30 most important features to predict Framingham score based on the faecal gut microbiota composition (n = 489). *The bacterium was present in the SumTn of Ruminoccocaceae UCG-002, which was protective against ischaemic heart disease, indicating that this bacterium may be part of the trophic network that has a protective effect against ischaemic heart disease. (B) Scatter plot of instruments with sensitivity analyses. (C) Forestplot with the effect of all SNPs used for the Mendelian randomization for Ruminoccocaceae UCG-002. IVW, inverse variance weighted.

3.3. Relations between the gut microbiome and CVD phenotypes

We performed a two-sample MR using the HELIUS cohort for exposure (microbial taxa abundance) and three CVD outcomes (ICD codes, see Supplementary material online, Table S6) from the Pan-UKBB. Subsequently, results from this analysis were compared with machine learning results and microbiome clustering to better understand the role of bacterial clusters or trophic networks in relation to CVD. In total, 303 MR analyses were performed (see Supplementary material online, Tables S3–S5) of which 21 were nominally statistically significant.

3.3.1. Ischaemic heart disease

In the total group, Dorea formicigenerans was potentially protective against ischaemic heart disease (ASV 72, IVW: OR = 0.92, P = 3.8 × 10⁻²). In the Dutch group, Acidaminococcus (ASV 167, OR = 0.99, P = 2.7 × 10⁻²) was shown to be protective against this phenotype. The OR represent increase or decrease of risk per standard deviation change.

In the Moroccan subgroup, three bacteria showed to be associated with a protective effect against ischaemic heart disease. Namely, the two genera Ruminococcaceae UCG-002 (ASV 176, OR = 0.89, P = 1.2 × 10⁻², Figure 3) and Ruminococcaceae UCG-005 (ASV 105, OR = 0.89, P = 4.9 × 10⁻³) as well as Roseburia (ASV 119, OR = 0.84, P = 1.4 × 10⁻²). Ruminococcaceae UCG-002 and Ruminococcaceae UCG-005 also showed a significant result for the weighted median but not for the MR Egger as a sensitivity test (see Supplementary material online, Table S3 for remaining results and sensitivity analyses). Lachnospira pectinoschiza was also shown to have a potential protective effect (ASV 54, OR = 0.87, P = 2.4 × 10⁻²).

In the South-Asian Surinamese subgroup, Alloprevotella (ASV 66, OR = 0.96, P = 3.9 × 10⁻³) was significantly associated with a protective effect against ischaemic heart disease, whereas Solobacterium (ASV 191, OR = 1.03, P = 5.3 × 10⁻³) contributed to ischaemic heart disease.

In the African-Surinamese, Akkermansia muciniphila showed to be associated with a protective effect against ischaemic heart disease (ASV 29, OR = 0.94, P = 4.2 × 10⁻²), whereas Desulfovibrionaceae (ASV 617, Wald Ratio: OR = 1.89, P = 1.8 × 10⁻²) contributed to the disease. In the Ghanaian group, Alloprevotella (ASV 66, OR 1.04, P = 2.5 × 10⁻²) and Lachnospiraceae (ASV 30, OR = 1,39, P = 4.2 × 10⁻²) contributed to this phenotype. In the Turkish groups, no significant role of the microbiome was observed.

3.3.2. Myocardial infarction

Narrowing our analysis to a more focused clinical phenotype (i.e. myocardial infarction), we found in the African-Surinamese Ruminiclostrididium 6 (ASV 84, IVW: OR = 0.90, P = 3.8 × 10⁻²) and Bacteroidetes (OR = 0.32, P = 4.8 × 10⁻²) were potentially protective against myocardial infarction. In the Ghanaian group, Blautia faecis (ASV 40, OR = 0.54, P = 2.1 × 10⁻²) was potentially protective whereas a relationship between Dorea longicatena (ASV 19, OR = 1.45, P = 4.4 × 10⁻²) and myocardial infarction was found.

However, the relationship with this phenotype disappeared after correction for multiple testing. In the South-Asian Surinamese group, Solobacterium appeared to be contributing to myocardial infarction (ASV 191, OR = 1.04, P = 7.2 × 10⁻³).

3.3.3. Coronary atherosclerosis

The third phenotype, which had the smallest selection of ICD codes and thus the most focused phenotype, was coronary atherosclerosis (see for overlap between phenotypes in ICD codes Supplementary material online, Table S6).

In the Moroccan group, Ruminococcaceae UCG-005 (ASV 105, IVW: OR = 0.84, P = 1.6 × 10⁻⁴), L. pectinoschiza (ASV 54, OR = 0.86, P = 2.7 × 10⁻²), and Collinsella aerofaciens (ASV 15, OR = 0.33, P = 2.6 × 10⁻²) were all associated with a protective effect against this phenotype. In the South-Asian Surinamese group, Alloprevotella (ASV 66, OR = 0.95, P = 2.5 × 10⁻³) was potentially protective against coronary atherosclerosis (see Supplementary material online, Table S5, for all MR results and sensitivity analyses).

In the Dutch, Turkish, African-Surinamese, and Ghanaian groups, no significant associations were found (see Supplementary material online, Table S5).

3.3.4. Bidirectional MR

We performed bidirectional MR to assess if there is a potential bidirectional relationship between microbes and CVD outcomes. This was performed for the total and Dutch groups using gut microbiota as outcome and CVD event as exposure (see Supplementary material online, Table S7). For the African-Surinamese, South-Asian Surinamese, Moroccan, and Turkish groups, no SNPs passed the F statistic threshold for the three phenotypes. Significant taxa identified in the initial MR were not significant in the bidirectional analysis MR in the total group. In the Dutch group, ischaemic heart disease was significantly contributing to the relative abundance of Acidaminococcus (ASV 167), suggesting that this microbe does not have a causal effect on ischaemic heart disease, but that ischaemic heart disease or a mechanism related to ischaemic heart disease may influence the abundance of this bacteria.

3.4. Trophic networks related to cardiovascular health

To determine the involvement of taxa identified through machine learning and MR in a trophic network, we employed two approaches. First, we identified the 30 most strongly positively correlated ASVs for MR bacteria and merged them into a composite signal termed SumTn. Secondly, we utilized hierarchical clustering of Spearman’s ρ coefficients for the 250 most abundant ASVs as an alternative method for defining composite signals. Within this framework, we observed that certain MR bacteria associated with CVD (ASVs 105 and 176) were interconnected within the CMR trophic network cluster (Figure 4 and Supplementary material online, Figures S8–S14). This cluster was then consolidated into a composite signal called CMR, which we correlated with cardiovascular and metabolic parameters. By comparing both approaches for generating composite signals of co-abundant bacteria, we assessed whether these composite methods enhanced the association with the metabolic parameters of interest when compared with each individual MR taxon.

Figure 4.

Heatmap of 16S faecal microbiome data of combined ethnicities. Six clusters were identified with this heatmap. The Prevotella enterotype-associated cluster, bacteria related to metformin, the CMR cluster core and a larger bacterial cluster, the Bacteroides cluster and a small intestinal bacterial cluster. Bacteria related to the heatmap based on n = 4117 individuals of different ethnicities. Hierarchical cluster based on Euclidean distances from a Spearman correlation.

The Moroccan group displayed a protective effect of Ruminococcaceae UCG-002 and Ruminococcaceae UCG-005 (ASV 176 and ASV 105, respectively) against ischaemic heart disease and myocardial infarction, both of which were also present in the CMR cluster. When examining the correlation coefficient (Spearman’s ρ), it was observed that the SumTn of Ruminococcaceae UCG-002 (ASV 176) exhibited amplification in the Moroccan subgroup across 13 out of 14 correlations, with four of them demonstrating significance after correcting for multiple comparisons (refer to Table 2). Among the SumTn correlation results, eight showed higher coefficients within the CMR cluster compared with the SumTn group. Notably, the inverse correlation with the Framingham score became statistically significant by using the SumTn bacteria (Ruminococcaceae UCG-002: ρ = −0.089, P = 0.042, q = 0.158; SumTn: ρ = −0.107, P = 0.015, q = 0.036; CMR cluster: ρ = −0.116, P = 0.008, q = 0.026). A similar effect was observed for triglyceride levels, smoking, HbA1c, and Lp(a) as indicated in Table 2.

Table 2.

Correlation to cardiovascular and metabolic markers is amplified with the sum of the trophic network (SumTn) or sum of the CMR cluster

	Ruminococcaceae UCG-002/ASV 176				SumTn			CMR cluster
	N	ρ	P-value	q value	ρ	P-value	q value	ρ	P-value	q value
Sex	594	−0.005	0.910	0.910	0.038	0.352	0.376	0.037	0.374	0.427
Age	594	0.025	0.539	0.578	0.035	0.393	0.393	0.027	0.507	0.541
Framingham score	518	−0.089	0.042	0.158	−0.107	0.015	0.036	−0.116	0.008	0.026
Triglycerides	594	−0.116	0.004	0.034	−0.147	3.19E-04	0.002	−0.154	1.71E-04	0.001
LDL-c	594	−0.055	0.184	0.276	−0.061	0.139	0.171	−0.080	0.051	0.079
Systolic blood pressure	594	−0.064	0.120	0.276	−0.071	0.083	0.120	−0.080	0.052	0.079
Smoking	580	0.102	0.014	0.070	0.122	0.003	0.014	0.125	0.003	0.010
BMI	594	−0.049	0.232	0.317	−0.076	0.065	0.115	−0.079	0.055	0.079
Waist circumference	594	−0.060	0.141	0.276	−0.092	0.025	0.050	−0.093	0.023	0.047
Fasting glucose	594	−0.025	0.539	0.578	−0.073	0.077	0.120	−0.064	0.119	0.147
Diabetes	594	−0.047	0.256	0.320	−0.042	0.304	0.348	−0.022	0.592	0.592
HbA1c	594	−0.076	0.063	0.188	−0.099	0.016	0.036	−0.094	0.022	0.047
CRP	93	−0.144	0.167	0.276	−0.155	0.137	0.171	−0.170	0.103	0.138
Lpa	93	−0.139	0.183	0.276	−0.255	0.014	0.036	−0.245	0.018	0.047
Ruminococcaceae UCG-002/ASV 29		1.000	0.000	0.000	0.690	4.05E-85	3.24E-84	0.6736108	8.99E-80	7.19E-79
SumTN/Sum CMR					1.000	0.000	0.000	1	0	0

Correlations shown are only performed in the Moroccan subgroup.

Within the African-Surinamese group, A. muciniphila (ASV 29) exhibited a protective effect against ischaemic heart disease, as illustrated in Figure 2. Furthermore, when comparing the correlation coefficient of A. muciniphila with the SumTn group, it was found to be higher in 9 out of 14 parameters concerning metabolic and cardiovascular health, with 7 of them demonstrating statistical significance (Table 3). These findings highlight a strong association between these combined bacteria and metabolic parameters. It is important to note that there is significant overlap between the ASVs associated with the SumTn from ASV 29, ASV 176, and the CMR cluster, as indicated in Supplementary material online, Table S9.

Table 3.

Correlation to cardiovascular and metabolic markers is amplified with the sum of the trophic network (SumTn) or sum of the CMR cluster

	A. muciniphila				SumTn			CMR cluster
	N	ρ	P-value	q value	ρ	P-value	q value	ρ	P-value	q value
Sex	1021	0.148	2.00−06	1.50E−05	0.208	2.01E−11	1.07E−10	0.119	1.37E−04	4.10E−04
Age	1021	−0.003	0.919	0.962	0.001	0.986	0.986	−0.028	0.370	0.512
Framingham score	968	−0.073	0.024	0.072	−0.122	1.49E−04	4.12E−04	−0.104	0.001	0.003
Triglycerides	1021	−0.076	0.015	0.072	−0.174	2.35E−08	9.39E−08	−0.17565	1.61E−08	5.79E−08
LDL-c	1021	0.028	0.363	0.605	0.029	0.355	0.474	0.024	0.440	0.563
Systolic blood pressure	1020	−0.059	0.059	0.111	−0.087	0.005	0.011	−0.073	0.019	0.034
Smoking	1021	0.060	0.056	0.111	0.114	2.65E−04	0.001	0.077	0.013	0.027
BMI	1019	0.006	0.840	0.962	−0.023	0.463	0.569	−0.012	0.707	0.796
Waist circumference	1020	−0.021	0.510	0.695	−0.079	0.012	0.021	−0.059	0.059	0.097
Fasting Glucose	1021	−0.063	0.044	0.111	−0.118	1.54E−04	4.12E−04	−0.085	0.007	0.015
Diabetes	1021	−0.001	0.962	0.962	−0.058	0.064	0.103	−0.046	0.141	0.212
HbA1c	1021	−0.072	0.022	0.072	−0.043	0.166	0.241	−0.023	0.469	0.563
CRP	198	−0.048	0.501	0.695	−0.041	0.566	0.647	0.008	0.910	0.963
Lpa	198	−0.034	0.638	0.798	−0.002	0.978	0.986	−0.003	0.963	0.963
A. muciniphila/		1.000	0.000	0.000	0.495	3.22E−64	2.58E−63	0.378	4.05E−36	1.82E−35
ASV 29
SumTn/Sum CMR					1	0	0	1	0	0

Correlations shown are only performed for the African-Surinamese subgroup.

3.5. CMR cluster replication cohort

The presence of the CMR cluster was also validated using the same clustering approach in the independent METSIM cohort and in a Korean cohort (see Supplementary material online, Figures S15 and S16 and Table S12) showing that the CMR trophic network cluster can reproducibly be identified in other ethnicities (Finnish and Korean). In fact, the exact same representative sequences are found in both cohorts for most ASVs of their respective CMR clusters, with ASVs 105 and 176 (uncultured Ruminococcaceae) from the HELIUS cohort being identical with ASVs 120 and 245 of the METSIM cohort. The main difference with regard to the CMR clusters is the inclusion of various Tenericutes ASVs in the METSIM cohort. However, by extending the top 250 most abundant ASVs by including less abundant ASVs, we also found Tenericutes ASVs to belong to the CMR cluster in the HELIUS cohort, and that Tenericutes ASVs 273, 235, and 525 from HELIUS match the sequence of the METSIM Tenericutes ASVs 91, 188, and 243. From the literature it is known that Tenericutes are indeed strongly associated with what is here called the CMR trophic network cluster.²⁹ In addition to this, the inverse association of the CMR cluster was also replicated in the Korean study, as was the negative correlation with triglycerides (ρ −0.23, P = 5.10 × 10⁻⁶, Supplementary material online, Table S16). See Supplementary material online, Tables S13 and S14, which summarize the overlapping findings from the different applied methods.

3.6. Triglyceride levels and the CMR cluster

The identified CMR cluster was especially strongly inversely correlated with triglyceride levels. To better understand if this can be interpreted as a dose-dependent effect, we calculated the average combined CMR abundance per ethnicity and plotted this against the prevalence of elevated triglyceride levels (triglyceride levels ≥1.7 mmol/L or lipid-lowering drugs) per ethnicity (Figure 5) by Mann–Whitney U testing. In addition, we compared individual CMR abundances per ethnicity between individuals with high triglyceride or low triglyceride levels. In Eurasian (South-Asian Surinamese, Turkish, Dutch, and Moroccan) ethnicities, a dose-dependent effect seems to be observed between the prevalence of CMR per ethnicity and the prevalence of high triglyceride levels (Figure 5).

Figure 5.

High plasma triglyceride level prevalence and average CMR abundance between ethnicities. South-Asian Surinamese had the lowest abundance of CMR but the highest prevalence of elevated triglyceride levels, followed by the Turkish, Dutch, and Moroccan subgroups. Eurasians and Sub-Saharan groups have different CMR-dependent triglyceride associations. Asterisks and P-value indicate results from Mann–Whitney U testing comparing CMR abundance between individuals with normal fasting triglyceride levels or increased triglyceride levels (≥1.7 mmol/L or use of lipid-lowering drugs). Higher CMR abundance is seen in individuals with lower prevalence of elevated triglyceride levels. R² indicates the result of a linear regression between average CMR abundance from Eurasian subgroups (e.g. SA, T, D, and M). SA, South-Asian Surinamese; T, Turkish; D, Dutch; M, Moroccan; AS, African-Surinamese; G, Ghanaian.

In the METSIM and Korean validation cohort, Christensenellaceae and Methanobrevibacter were similarly reported²⁶ or shown (see Supplementary material online, Table S16) to be associated with lower triglyceride levels (ρ = −0.23, P = 5.10 × 10⁻⁵). Moreover, many of the exact same Ruminococcaceae as reported in this study were associated with Christensenellaceae and Methanobrevibacter in the validation cohorts, and in the METSIM cohort they were also reported to be depleted in the pre-diabetic subjects.

3.7. Ethnicity interaction

In order to not only identify ethnicity-specific effects as reported previously for MR and machine learning analyses, we also conducted a linear regression to identify ethnicity interactions in the relationship between microbes and triglycerides. We chose the nominally significant microbial taxa from the MR results and compared a basic linear model (microbial abundance and fasting serum triglyceride concentration) with a linear model including ethnicity, and compared the two models with an ANOVA. Here, we found several bacteria that also had an ethnicity interaction, including L. pectinoschiza, D. Formicigenerans, and Acidaminococcus, in the ischaemic heart disease outcome group. In the myocardial infarction group, B. faecis, D. Longicatena, and the genus Sutterella showed significant interaction. For coronary atherosclerosis, L. pectinoschiza and C. aerofaciens had an interaction with ethnicity (see Supplementary material online, Tables S8–S10).

4. Discussion

To date, this is the first study assessing the role of single and co-abundant bacteria in relation to CVD in a multi-ethnic cohort. In our study, we show that several of the bacteria associated with the Framingham score and microbes that are potentially causally related to CVD in different ethnic groups are correlated and form the CMR cluster. The use of the CMR cluster, containing these causative bacteria, as a combined composite signal amplified the negative association between the Framingham score and elevated triglyceride incidence, which was replicated in two other cohorts. The CMR cluster encompassing the bacteria found to be causatively associated with CVD was not only found to be associated with CVD risk in the African-Surinamese and Moroccan individuals but was also found to be of relevance in all Eurasian ethnicities. Furthermore, a dose-dependent effect was observed between the CMR abundance per ethnicity and the prevalence of elevated triglyceride levels in Eurasian ethnicities (Figure 5), further providing proof of the potential protective importance of this cluster.

The identified SumTn in the African-Surinamese and Moroccan groups, which is similar to the CMR cluster, has already been shown to be strongly associated with parameters of health. Christensenellaceae has previously been shown to be inversely associated with BMI and age,^29,30 and supplementation of Christensenella minuta into human donor faeces introduced into germ-free mice reduced weight gain.²⁹ A lower BMI is associated with improved metabolic parameters and a lower CVD risk.³¹Christensenellaceae are known to co-occur with Methanobrevibacter,^29,30,32 which is based on the cross-feeding relationship between these two groups as C. minuta produces H₂, which is a substrate for Methanobrevibacter smithii.³⁰ Interestingly, various Ruminococcaceae, a family abundantly represented in the CMR cluster, are also known to ferment complex dietary carbohydrates to hydrogen and may thus in this fashion contribute to this trophic network.³³ Furthermore, previous studies also observed a direct protective effect against CVD of bacteria identified herein. In the African-Surinamese group, A. muciniphila had a negative association with the Framingham score and protected causally against the ischaemic heart disease phenotype. A similar effect has been observed in apolipoprotein E-deficient mice, where A. muciniphila reversed atherosclerosis induced by a Western diet through induction of intestinal tight junction expression and subsequent attenuation of endotoxemia-induced inflammation.³⁴

Interestingly, in our study, the A. muciniphila SumTn was inversely associated with triglyceride levels, and in the CMR cluster this association was even stronger. This is of interest, as increased triglyceride levels are a risk factor to develop CVD³⁵ and might point towards a potential additional protective mechanism. Indeed, 4 weeks oral gavage of A. muciniphila lowered body weight, total blood cholesterol, and triglyceride levels in a hyperlipidaemic ApoE3 E3L CETP mice model.³⁶ The results imply a mechanistic role for this species in lipid metabolism.

Using MR, we assume in our study that the genetic constitution of participants can influence gut microbiota, which in turn can have a casual effect on CVD. This concept has previously been introduced by studying the microbiome of monozygotic and dizygotic twins, where it was shown that monozygotic twins have a more similar gut microbiota compared with one another than dizygotic twins. Of additional interest is that the highest heritability based on phylogeny was seen for the families Ruminococcaceae and Lachnospiraceae, whereas Christensenellaceae was the most highly estimated heritable taxon.²⁹ This strongly suggests that the human genetic constitution can influence the gut microbiota composition. As the genotype differs between ethnicities, genetic differences may explain observed ethnic differences in gut microbiota composition.⁷ In addition to this, the genotype influence on gut microbiota composition may in part explain why different bacteria were found to be of importance in different ethnicities in our study. However, age also affects gut microbiota³⁷ and there were differences in the average age of the different ethnicities included in this study, yet this does not account for the associations found with the CMR cluster. This cluster is, for example, slightly positively correlated with age. For instance, the Asian Surinamese have the highest age on average, yet have the lowest abundance of the CMR cluster. Furthermore, despite the various age differences, the CMR cluster appears to be associated with cardiovascular health in all ethnicities except in Ghanaians. Interestingly, some of the bacteria found in our study were already identified as being related to CVDs in previous MR studies. The species Alistipes shahi and unclassified Alistipes have been shown to reduce triglyceride levels in two independent cohorts.³⁸ This is in line with findings in our study, where A. shahi was part of the A. muciniphila SumTn, which enhanced the inverse association with triglyceride levels compared with A. muciniphila only. Furthermore, we found that Lachnospiraceae were contributing to ischaemic heart disease in African-Surinamese subjects. Lachnospiraceae have previously been shown to increase the risk of heart failure, which can be the result of ischaemic heart disease.³⁹ This suggests that previously identified bacteria can be part of a relevant bacterial cluster that could affect CVD events.

A limitation of the study was the use of the Framingham score as a surrogate marker for CVD based on clinical parameters in the HELIUS cohort, as overt CVD cases were not available in this relatively young cohort. In addition, the Framingham score is not validated in all ethnic groups we tested, which might influence found differences. The combination of outcomes from the Pan-UKBB however made it possible to infer potential causality of several bacteria on hard CVD outcomes. The multi-ethnic design of the HELIUS study improves the representation of findings. Furthermore, the exact same methods were used to process all samples, which resulted in reliable data. Matched ethnicities between the UKBBand the HELIUS cohort were relatively similar, although not exactly equal, and may introduce some bias. Another limitation of this study includes possible horizontal pleiotropy, which is indicated by some sensitivity analyses. Horizontal pleiotropy can be a problem in MR studies, especially if the studied population is relatively small compared with other GWAS, if study results rely only on MR, and if there is no biological mechanisms of observed results.⁴⁰ This must be considered, and could be due to low statistical power. In addition to this, testing for interactions between ethnicities provided varying results, which also may be due to the sample size, as a higher sample size is necessary for interaction estimation. However, our analyses are based on the biological hypothesis that the host genetic makeup influences gut microbiota composition and that gut microbiota derived metabolites or interactions with lipid absorption can affect host metabolism. In addition to this, we did not rely on the MR only, but support our hypothesis by a machine learning approach and conventional statistical analyses. Most machine learning models in our study showed low RMSE values and low P-values, indicating an overall good fit of the models. RMSE was the highest in the Ghanaian and Turkish groups and the lowest in the total and Moroccan groups, suggesting that the fit in these models was the best. However, RMSE may also be affected by sample size, which limits generalization of interaction effects by ethnicity. In addition to this, we were able to successfully replicate the CMR cluster in two independent external cohorts and replicate the inverse association of the CMR cluster with triglycerides, supporting our hypothesis that this cluster may have cardio-protective effects. This may be the reason why Christensenellaceae and Methanobrevibacter are so commonly found to be more abundant in centenarians.^41–44 In our study, we used summary statistics of two GWAS that were both corrected for common covariates such as age, sex, and BMI.^16,45 However, recent research has indicated that correction for covariates in GWAS for MR studies can introduce bias.⁴⁶ The results of this study must therefore be considered in light of potential bias due to the use of covariate-adjusted summary statistics. Future studies should avoid correcting for covariates or publish uncorrected summary statistics along with corrected summary statistics to limit the risk of bias.

In this study we demonstrated that ethnic-specific gut microbiota profiles can estimate CVD risk, and certain gut microbes may have a causal relation to CVD events. However, by switching from a single microbe to a trophic network approach, stronger associations with CVD risk in all but the Ghanaian ethnicity were observed. The CMR trophic network appears to represent a relevant estimator of CVD risk in nearly all ethnicities, though the link with human genetics remains equally relevant as microbes in the CMR cluster are strongly associated with human genetics, as determined using twin and GWA studies.^29,32 The dose-dependent relationship of elevated triglyceride prevalence and the CMR cluster in the Eurasian ethnicities highlight the importance of gut microbiome differences between the studied ethnicities. Species encompassed by a trophic network together represent a much larger fraction of the entire gut microbiome than single species, and can hence theoretically more significantly influence host metabolism. Future studies should focus on the metabolic functions in vivo and in vitro of the CMR cluster in relation to host metabolism in order to better understand the interaction between host genetics, gut microbiota, and host CVD risk.

Author contributions

Study conceptualization: M.V.W., U.B., J.A.-W., D.H.v.R., A.K.G., M.N. Statistical analysis, data interpretation: M.V.W., U.B., M.C.de.G., R.E.A., E.L., L.R.F.B. Data curation and handling: H.H., K.C., H.G., B.F., B.J.H.v.d.B., A.K., J.F., A.H.Z.. Manuscript drafting: M.V.W., U.B., D.H.v.R., H.H., A.K.G., M.N., M.R.S., J.F. Funding: A.K.G., M.N.

Supplementary material

Supplementary material is available at Cardiovascular Research online.

Funding

This work was supported by the Netherlands Cardiovascular Research Committee (CVON) IN-CONTROL-II grant (2018-27). M.N. is supported by a ZonMw Vici grant 2020 (09150182010020) and a Le Ducq foundation grant 17CVD01. H.H. is supported by a Senior Fellowship of the Dutch Diabetes Research Foundation (2019.82.004). J.F. is supported by the Dutch Heart Foundation IN-CONTROL (CVON2018-27), the ERC Consolidator grant (grant agreement no. 101001678), NWO-VICI grant VI.C.202.022, the AMMODO Science Award 2023 for Biomedical Sciences from Stichting Ammodo, and the Netherlands Organ-on-Chip Initiative, an NWO Gravitation project (024.003.001) funded by the Ministry of Education, Culture, and Science of the government of The Netherlands. The HELIUS study is conducted by the Amsterdam University Medical Center, location AMC, and the Public Health Service of Amsterdam. Both organizations provided core support for HELIUS. The HELIUS study is also funded by the Dutch Heart Foundation (K. Stronks, 2010T084), the Netherlands Organization for Health Research and Development (ZonMw: K. Stronks, 200500003), the European Union (FP-7, K. Stronks, 278901), and the European Fund for the Integration of non-EU immigrants (EIF, K. Stronks, 2013EIF013). U.B. is supported by European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement no. 813781.