Systematic analysis of relationships between plasma branched-chain amino acid concentrations and cardiometabolic parameters: an association and Mendelian randomization study

In 300OB, fat distribution was assessed using MRI, including volumes of visceral adipose tissue and subcutaneous adipose tissue, divided into deep and superficial subcutaneous adipose tissue, respectively. Hepatic fat content was quantified using localized proton magnetic resonance spectroscopy. Non-invasive measures of atherosclerosis of the carotid arteries included measurement of carotid IMT, carotid plaque presence and maximum plaque thickness, and plaque presence and maximum thickness in the common carotid, internal carotid, external carotid artery, and carotid bulbus. The presence of plaque was defined as focal thickening of the wall of at least 1.5 × mean IMT or an IMT > 1.5 mm, according to the Mannheim IMT consensus [51]. All measurements were performed at Radboud University Medical Center. Further details are described in a previous paper [49].

Blood samples of participants from both cohorts were collected separately and frozen at − 80 °C prior to any measurement. For both cohorts, circulating proteins from EDTA plasma samples were measured using commercially available OLINK Proteomics (Uppsala Sweden), the AB Inflammation Panel for 300OB (162 CVD II panel and inflammation panel), and OLINK Proseek Multiplex CVD III panel for LLD (92 CVD proteins). OLINK provides a multiplex immunoassay for high-throughput detection of protein biomarkers in liquid samples. Proteins are recognized by antibody pairs coupled to cDNA strands that bind in close proximity and extend by a polymerase reaction. In 300OB, proteins were excluded from analysis when the detection level of 75% was not met. Quality control (QC) was performed by OLINK Proteomics, with two samples that did not pass QC subsequently excluded from the analysis. OLINK panel data was available for 1294 participants from LLD and 294 samples from 300OB. More detailed information can be found at the OLINK site: http://​www.​olink.​com/​products.

For both cohorts, we profiled a wide range of plasma metabolites using NMR and the Nightingale Biomarker Analysis Platform [52]. This platform provides a wide range of plasma metabolites including lipid concentrations; relative compositions of 14 lipoprotein subclasses; lipoprotein particle sizes and concentrations of apolipoproteins; cholesterol; triglycerides, and phospholipids; and several glycolysis components, fatty acids, inflammation markers, ketone bodies, and amino acids. This platform provides measures of 231 plasma metabolome traits, which included the three BCAA, in both cohorts. For the other traits, we included 228 metabolites in LLD but only 223 metabolites in 300OB due to a large percentage of missing values for five measurements. Our group had previously validated platform precision by comparing several traits and observed a high degree of consistency [49]. All the NMR measurements in LLD samples were performed in one batch, whereas 300OB samples were randomized and measured in two batches. Therefore, for association analyses in the 300OB cohort data, batch number was included as an additional covariate.

We used ultra-high-performance liquid chromatography in combination with isotope dilution tandem mass spectrometry to analyze TMAO, choline, L-carnitine, and betaine for both cohorts and γ-butyrobetaine in plasma specifically in the 300OB cohort, as described elsewhere [49].

Blood cell counts were measured for both cohorts, and we used six cell types related to immunity for further analysis: basophils, eosinophils, granulocytes, lymphocytes, monocytes, and thrombocytes. The cytokines for the LLD cohorts included IL (interleukin)-1β, IL-6, IL-8, IL-10, IL-12p70, and tumor necrosis factor-α (TNF-α) measured by ProcartaPlex multiplex immunoassay (eBioscience, San Diego, CA). Other inflammation markers, including leptin, adiponectin, IL-18, IL-18BP, and resistin, were measured using commercially available sandwich ELISA kits (R&D Systems, Minneapolis, MN). For the 300OB cohort, measured cytokines included IL-1β, IL-6, IL-18 (including IL-18BP), and inflammation markers included adiponectin, leptin, and resistin, all measured according to protocols described earlier [49].

First, we identified age, sex, and BMI as important covariates in assessing the association between BCAAs and CMD parameters. Next, we explored whether other lifestyle and dietary factors can be potential confounders. If a factor is an important confounder, we expected to see a significant association between this factor and the exposure, i.e., BCAAs. We therefore made use of the LLD cohort to assess the associations between BCAAs and all potential lifestyle and dietary covariates using a univariable regression model. The lifestyle and dietary factors assessed included current smoking status, intake of alcohol, meat, fruit, vegetables, and coffee and total energy intake in kcal and macronutrients (carbohydrates, total protein, animal protein, plant protein, and fat) (Additional file 1: Table S2). These lifestyle factors were obtained from questionnaires filled in by each participant. Dietary intake was assessed using a semi-quantitative Food Frequency Questionnaire (FFQ) [48]. While we did observe strong associations with age, sex, and BMI, we did not observe any significant associations with smoking or dietary factors (Additional file 1: Table S2). This led us to conclude that the potential confounding effects of these factors were negligible, if present at all. We therefore only considered age, sex, and BMI as important covariates and assessed pair-wise associations between individual BCAAs and different CMD traits using linear regression models where phenotype ~ BCAA + age + sex + BMI. The independent and continuous dependent variables were scaled using a z-scale transformation (to obtain values ranging from − 1 to 1) to be able to compare the effect size of different categories and in both cohorts. Logistic regression was used for binary phenotypes. Each BCAA was analyzed separately without mutually adjusting for the others.

Multiple regression requires the data to fulfill several assumptions: the response variable is assumed to be a linear function of the model parameters (1) and model errors are assumed to be independent (2), to have a constant variance (3), and to be normally distributed (4). In addition, the predictors are assumed to be measured without error (5), and there are assumed to be no unmeasured confounders that affect both the response and predictor variables (6). To ensure that major assumptions 1–4 were fulfilled, we manually inspected diagnostic plots of the models for each BCAA–phenotype combination. To conform with the normality assumption, we log-transformed 27 phenotypes (see Additional file 1: Table S1). Assumptions 5 and 6 could not be directly assessed, which is one of the limitations of our study.

We note that we included BMI as a covariate because its causal role in BCAAs and cardiometabolic traits has been widely studied. In this study, we wanted to correct for the confounding effects of BMI so we could investigate the associations of BCAAs with various CMD-related parameters, including fat distribution, independent from overall obesity as represented by BMI. NMR measurements in the 300OB cohort were also adjusted for batch number. Where applicable, we included additional covariates, including protein and meat intake and HOMA-IR for NMR analysis in LLD, as mentioned in the “Results” section. In addition, BCAAs have been implicated in kidney function, which is also linked to cardiometabolic health. We therefore also checked the LLD cohort for associations between BCAAs and kidney function, which was assessed by estimated glomerular filtration rate (eGFR). The latter was calculated using the original Modification of Diet in Renal Disease formula. All three BCAAs were significantly associated with eGFR (Additional file 1: Table S2). As eGFR data was not available for the 300OB cohort, we did not include eGFR as a covariate but did compare results with or without eGFR as a covariate in the LLD cohort. To correct for multiple testing, P values were further adjusted for 537 traits × 3 BCAA for each cohort using the Benjamini–Hochberg method, with significance set at a false discovery rate (FDR) < 0.05. These are reported as the FDR-value in both the tables and in the Results section where appropriate. All analyses were performed using R version 3.61.

To determine causality between BCAAs and associated factors, two-sample bi-directional MR was performed using the R package TwoSampleMR v.0.5.6 with default settings [53]. The genetic variants used as instrumental variables for MR analyses and their effect sizes were obtained from publicly available summary statistics for genome-wide association studies (GWASs) on BCAAs and cardiometabolic traits. BCAA-associated SNPs were taken from the largest GWAS on metabolites to date performed in the UK Biobank and available in the OpenGWAS database [54] under accession IDs met-d-Val, met-d-Ile, and met-d-Leu. Summary statistics for all of the cardiometabolic traits were obtained from the OpenGWAS database, where we aimed recover data from the GWAS with the highest sample size, preferably based on European samples, for each CMD parameter tested in the association analysis. Genetic variants were clumped using r² < 0.001 in 1000G EUR samples. Proxies were added automatically by the TwoSampleMR R package. To estimate the causal effect of BCAAs on CMD parameters, we used SNPs associated with the sum of all three BCAAs (accession id met-d-Total_BCAA) to overcome the potential confounding effect of the other two BCAAs when using separate BCAAs as the exposure.

To calculate univariable MR (UVMR) estimates, we used Wald ratios meta-analyzed by the inverse variance weighted (IVW) method [55]. Instrumental variables (IVs) used in MR need to fulfill three major assumptions: (1) the IVs should be associated with the exposure, (2) the IVs should not share a common cause with the outcome, and (3) the IVs should affect the outcome only through the exposure. To reduce the chances of violating these assumptions, extensive QC and sensitivity analyses were performed on the candidate MR results. In detail, results were only considered when they met the following criteria: (1) MR results were based on three or more SNPs, as this allowed us to perform the sensitivity analyses listed below; (2) MR results showed a Benjamini-Hochberg-corrected p value < 0.05 using IVW and nominally significant results (p value < 0.05) using two other MR approaches (weighted median and MR PRESSO test [56, 57]); (3) MR results did not show indications of horizontal pleiotropy or heterogeneity, as estimated using MR Egger [57] (intercept p value > 0.05) and the MR PRESSO [56] outlier-adjusted test (p value < 0.05) that estimates the pleiotropy and tries to correct for it by removing outliers; (4) MR results were not driven by single SNPs, as tested using leave-one-out analyses (no SNP after exclusion resulting in IVW MR p value > 0.05); or (5) genetic instruments were strong as estimated using F-statistics (F > 10). We also estimated heterogeneity using Cochran’s Q-test but did not filter out the results based on this measure.

BMI has previously been shown to affect both BCAA levels and CMD parameters. Moreover, BMI, BCAAs and CMD parameters have shared associated genetic variants. To overcome the potential violation of the 3rd MR assumption in the UVMR analyses, we removed BMI-associated SNPs published by the GIANT consortium [58] from the list of genetic variants. In addition, we corrected for the effect of BMI using multivariable MR analyses (MVMR). In the MVMR analysis estimating the effect of CMD parameters on BCAAs, we used two exposures: the current CMD parameter and BMI (using summary statistics published by the GIANT consortium [58]). In the MVMR in the direction from BCAAs to CMD parameters, we used four exposures: valine, isoleucine, leucine, and BMI. Sensitivity analyses of the MVMR results were performed by the MVMR v. 0.3 R package [59], which estimates the strength of the genetic instruments (F-statistics, which we required to be > 10) and heterogeneity using Cochran's Q-test.

We performed the analyses in both directions and corrected for multiple testing using the Benjamini–Hochberg method separately for each group of phenotypes (general cardiometabolic traits and NMR lipoproteins) and for each direction.

The basic characteristics of the LLD (n = 1400) and 300OB (n = 294) cohorts and their respective fasting plasma BCAA concentrations are summarized in Additional file 1: Table S1. The average plasma levels of BCAAs in the population-based LLD cohort were 44.9 μM for isoleucine, 57.9 μM for leucine, and 149 μM for valine. In the obesity cohort (300OB), the isoleucine concentration was 6.6 μM higher (p = 5.68 × 10⁻⁸), while the average concentrations of leucine and valine were 3.1 μM (p = 5.16 × 10⁻⁵) and 7 μM (p = 0.0017) lower, respectively, compared to the LLD cohort (Additional file 1: Table S1). In addition, BCAA concentrations were generally higher in males than in females and correlated positively with BMI (Additional file 1: Table S2). After correcting for age, sex, and BMI, we performed an association analysis of BCAAs with 537 CMD parameters available in either cohort, including 5 glycemic traits, 6 routine lipid measures, 4 blood pressure measures, 15 atherosclerosis-related parameters, 7 detailed fat distribution parameters, 19 cell counts and cytokines, plasma concentrations of 248 circulating proteins, 5 parameters of TMAO and its precursors, and 228 plasma metabolites (predominantly lipidomic traits) (Fig. 1, Additional file 1: Table S1).

Overall, we detected 838 significant associations for 409 unique CMD parameters in at least one of the cohorts at FDR < 0.05 (Additional file 1: Table S3). In addition, we looked at the confounding effect of kidney function, assessed by eGFR, on the associations between BCAAs and CVD in the LLD cohort. Despite being significantly associated with all three BCAA levels (Additional file 1: Table S2), correcting for eGFR did not influence the association results (Additional file 1: Table S4, Additional file 2: Figure S1).

Different BCAAs showed different numbers of associations with CMD parameters, and the estimated association strength also differed. Isoleucine had the highest number of associations (373 associations), followed by leucine (263) and valine (202) (Fig. 2). We found 153 CMD parameters that were associated with all three BCAAs. However, some associations were significant only for specific BCAAs: 108 parameters were only associated with isoleucine, 1 with leucine, and 24 with valine. In addition, on average, association effect sizes were largest for isoleucine (mean beta_LLD = 0.22), followed by leucine (mean beta_LLD = 0.09) and valine (mean beta_LLD = 0.07) (one-way ANOVA P = 1.71 × 10⁻¹⁶). This observation implies that different BCAAs may have a divergent impact on CMD. While zooming in on different categories of CMD parameters, we found shared associations of BCAAs for all five glycemic traits, including fasting plasma levels of glucose, insulin, HbA1c, HOMA-IR index, and diabetes status. However, BCAA-specific associations were found for lipidomic traits and proteomics (Fig. 2). The estimated association strength with NMR-based lipidomic traits was particularly strong for isoleucine when compared to the estimations with valine (Fig. 3). For example, the estimated effect size of the association between isoleucine and plasma levels of saturated fatty acids in the LLD cohort was 0.47 (p = 1.07 × 10⁻⁵⁵), while the estimated effect size was 0.10 (p = 8.16 × 10⁻⁴) for leucine and -0.07 (p = 7.94 × 10⁻³) for valine (Additional file 1: Table S3). In line with this, isoleucine also showed more or stronger associations with fat distribution in the 300OB cohort. For example, plasma levels of isoleucine were significantly associated with the ratio of visceral adipose tissue to subcutaneous adipose tissue (VAT_SAT_ratio) (β_ile = 0.31), while the association was much weaker for leucine (β_leu = 0.20) and not significant for valine (β_val = 0.11) (Additional file 2: Figure S2). Moreover, only isoleucine was significantly associated with the number of carotid artery plaques (β_ile = 0.16, p = 0.01) in the 300OB cohort (Additional file 1: Table S3). For circulating proteins, proteins from the inflammatory panel measured in the 300OB cohort had more isoleucine-associated proteins, while CVD-related proteins measured in the LLD cohort showed more associations with valine (Additional file 1: Table S2). We did not detect many associations with cytokines (except for adiponectin, leptin, and IL-18BP) and cell counts (except for lymphocytes and leucocytes) (Additional file 2: Figure S3). In recent years, the gut microbiome has been implicated in CMD, especially through gut microbial metabolites such as TMAO, and we found both TMAO and its precursors to be associated with BCAAs (Additional file 2: Figure S4).

Out of 756 pair-wise associations for 252 CMD parameters that were tested in both the LLD and 300OB cohorts, 202 associations showed significant heterogeneity between the two cohorts (I² > 0.75 and heterogeneity FDR < 0.05, Additional file 1: Table S5, Fig. 4), even though we had corrected for age, sex, and BMI as covariates. This observation was not due to the differences in the number of diabetes patients and glycemic traits, as the estimated effect sizes were comparable after removing T2D patients in 300OB (Pearson r = 0.97) (Additional file 1: Table S6, Additional file 2: Figure S4) or correcting for HOMA-IR index in the LLD cohort (Pearson r = 0.98) (Additional file 1: Table S7, Additional file 2: Figure S5). Interestingly, the most heterogeneous associations were related to isoleucine and lipidomic traits, and they seemed to show consistently larger effect sizes in the 300OB cohort than in the LLD cohort (Fig. 4A). Moreover, many heterogeneous associations observed for valine showed opposing directions in LLD and 300OB (Fig. 4A). This observation inspired us to hypothesize that an individual’s obesity status may explain these heterogeneous associations, which cannot be simply corrected for using BMI as a covariate. To examine this, we selected a subset of 185 individuals from the LLD cohort with matched age and BMI to the 300OB cohort. The heterogeneity in effect size in this LLD subset was not much different for leucine and isoleucine (Fig. 4B, C). Strikingly, the association directions of valine were altered in the individuals with obesity from LLD, which became consistent with the estimations in the 300OB cohort (Fig. 4C, Additional file 1: Table S8).

Although we identified numerous associations between BCAAs and cardiometabolic traits, and specifically with plasma lipids, using both clinical measures and detailed NMR plasma lipid levels measurements, the associations alone are insufficient to provide insights into the underlying causality. To investigate the causality of the identified associations, we performed bi-directional two-sample MR analysis using genetic variants as IVs (Fig. 1B, Additional file 3: STROBE-MR checklist) [60]. Genetic variants affecting BCAA concentrations were taken from the largest publicly available GWAS database of metabolites from the UK Biobank cohort [53, 54]. For each CMD parameter, we searched the MRC IEU Open GWAS database for summary statistics, preferably from population-based European GWAS studies to ensure similarity of SNP associations with both exposure and outcome. This resulted in data for 28 phenotypes and 217 lipoproteins (characteristics of the included GWASs can be found in Additional file 1: Table S9). As some of the available studies were based on UK Biobank data (see Additional file 1: Table S9), there was a sample overlap between exposure and outcome data, and some of the MR analyses may not be a truly two-sample MR.

Obesity has previously been reported to causally contribute to both CMD parameters and BCAAs [3, 4, 41]. Therefore, we performed two types of MR analyses to deal with the confounding effect of BMI and removed causal estimates that failed sensitivity analyses (see the “Methods” section, Additional file 1: Tables S10-S11). In the MVMR analysis, including BMI as an additional exposure covariate revealed eight potential causal links from three phenotypes to BCAAs: increases in triglyceride levels and frequency of T2D and a decrease in HDL cholesterol levels were all associated with an increase in BCAA levels (Additional file 1: Table S10); however, the heterogeneity of these associations was high. A UVMR analysis, performed after removing BMI-associated SNPs, identified seven significant results linking four traits to BCAA levels. For example, for each 1 SD increase in genetically predicted leucine, we saw a 0.52 SD increase in fasting insulin, a 0.29 SD increase in fasting glucose and a 0.05 SD decrease in total cholesterol levels (Additional file 1: Table S11, Fig. 5). As BCAA levels are highly correlated, we used the combined levels of all three BCAAs as the exposure when estimating the causal effect of BCAAs on CMD parameters. However, neither method could detect any significant causal relationships from BCAAs to phenotypes in our dataset as these MR results failed QC sensitivity analyses (Additional file 1: Table S11). Altogether, our results suggest altered plasma BCAA levels are more likely the outcome of metabolic syndromes. This was further confirmed by our observation of a large number of causal effects of lipidomic traits on BCAAs: 504 significant estimates for 185 unique lipidomic traits using MVMR (Additional file 1: Table S10) and 252 significant estimates for 96 traits using UVMR (Additional file 1: Table S10). All reported UVMR effects passed the pleiotropy and heterogeneity checks performed using the MR Egger and MR PRESSO methods. However, a high degree of heterogeneity for most of the MR effects was observed based on Cochran's Q-test. This implies complexity in causality and suggests some caution in data interpretation.