Plasma Proteomic Signatures of Adiposity Are Associated With Cardiovascular Risk Factors and Type 2 Diabetes Risk in a Multiethnic Asian Population

Obesity is a metabolic disorder resulting from a complex interplay between genetic, psychosocial, and environmental factors (1). Although obesity is a major risk factor for metabolic diseases, such as type 2 diabetes and cardiovascular disease (CVD) (2), the biologic mechanisms connecting obesity and these metabolic diseases are not fully understood.

Recent studies in European populations have identified proteins associated with BMI (3–5), highlighting pathways involved in lipid metabolism and inflammation that may contribute to obesity-related diseases (4). However, these studies were based on earlier versions of the SomaScan platforms, which covered a limited number of proteins (∼1,100–3,600 proteins), used BMI as the only measure of adiposity, and included participants of predominantly European ancestry. The latter may be particularly important, given the well-recognized differences in the association of BMI with type 2 diabetes in populations of Asian compared with European ancestry (6). Finally, previous studies did not examine whether obesity-related proteins are associated with cardiometabolic health.

We evaluated the associations between ∼5,000 plasma proteins and BMI and waist circumference (WC) in a multiethnic Asian population comprising adults of Chinese, Malay, and Indian ethnicity. We also developed protein signatures of BMI and WC and evaluated their associations with total and visceral body fat, cardiometabolic abnormalities and metabolically unhealthy obesity, and incidence of obesity, metabolic syndrome, and type 2 diabetes.

An overview of this study is shown in Fig. 1. Participants in this study were sampled from the Singapore Multi-Ethnic Cohort Phase 1 (MEC1) (7). MEC1 is a population-based cohort comprising 14,465 ethnic Chinese, Malay, and Indian adults recruited between 2004 and 2010. Between 2011 to 2016, 6,112 participants completed a follow-up study. Participants completed an interviewer-administered questionnaire and were invited to a physical examination visit for both the baseline and follow-up studies. Height was measured without shoes on a portable stadiometer, and weight was measured on a digital scale. WC was measured by trained research staff using stretch-resistant tape at the midpoint between the last rib and iliac crest. Blood was drawn for biomarker measurements and biobanking at −80°C.

At follow-up, a subset of Chinese participants underwent DXA and computed tomography scan for measurements of total body fat mass, trunk fat mass, android fat mass, lean body mass, visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) (8) (Supplementary Material).

Selected MEC1 participants were profiled on the SomaScan proteomic assay. First, a random sample of 720 participants with equal numbers in each sex and ethnic group were included as the population set. The second set included 759 incident type 2 diabetes patient cases and 1,484 matched controls as the type 2 diabetes case-control (T2DCC) set. After accounting for shipment issues and exclusion criteria, the final number of participants for analyses was 631 from the population set and 616 patient cases and 1,200 controls from the T2DCC set (Supplementary Material and Supplementary Table 1).

We externally replicated our findings in the Atherosclerosis Risk in Communities (ARIC) study, a population-based cohort study of White and African American U.S. adults. Details of the ARIC study have been reported elsewhere (9). We included a total of 8,428 ARIC participants with proteomic data measured on the SomaScan V4 platform using samples collected during visit 2 (1990–1992) in our analyses (Supplementary Material).

Relative protein abundances for 4,978 SOMAmers targeting 4,775 human proteins were measured in plasma samples using aptamer-based technology (SomaScan V4 assay) by SomaLogic, Inc. (Boulder, CO) (10,11). Standardization and quality control analyses are described in the Supplementary Material.

All data analyses were based on log₂-transformed values and winsorized at ±5 SDs to reduce the impact of extreme outliers. All P values reported are from two-sided tests. R software (version 4.2.0) was used.

Using the population set at baseline, we evaluated ethnicity-specific associations between protein levels and BMI and WC adjusted for age and sex. The ethnicity-specific results were pooled using fixed-effects inverse-variance meta-analysis.

We defined associated proteins as those that were 1) significantly associated with BMI or WC at the Bonferroni threshold (P < 1.00 × 10⁻⁵), 2) had the same direction of association across ethnicity, and 3) had no evidence of statistical heterogeneity using the Cochran Q test (P_het >1.00 × 10⁻⁵). For proteins targeted by multiple SOMAmers, the SOMAmer with the smallest P value was selected. Using the associated proteins, we performed 100 replicates of 10-fold cross-validation elastic net regression analysis (12). Elastic net regression is a regularized regression method that enables automatic variable selection and variance reduction (12). For each participant, the signature scores were calculated as the weighted sum of the abundances of the selected proteins multiplied by the β-coefficients for these proteins from the elastic net regression model.

Using the T2DCC set as validation, we examined cross-sectional correlations between the protein signatures and BMI and WC at baseline and follow-up and the correlations between changes in the protein signatures and changes in BMI and WC over a period of ∼6 years. We also examined the correlations with DXA body fat measurements and VAT and SAT measurements.

We evaluated associations between the protein signatures and systolic blood pressure, HDL cholesterol, LDL cholesterol, triglycerides, fasting glucose, and A1C among participants in the T2DCC set. We used linear regression models adjusted for age, sex, and ethnicity and evaluated the difference in standardized coefficients using the z test. Participants receiving antihypertensive or lipid-lowering medications were excluded from analyses involving blood pressure and blood lipids, respectively.

We evaluated the ability of our protein signatures to differentiate between metabolically healthy and unhealthy obesity (13) among obese individuals in the T2DCC set at baseline. We defined obesity as BMI ≥25.0 kg/m² according to the World Health Organization recommendations for Asian populations (14). We defined metabolically unhealthy as having two or more of four metabolic abnormalities: 1) systolic blood pressure ≥130 mmHg or diastolic blood pressure ≥85 mmHg, 2) fasting plasma glucose ≥5.6 mmol/L, 3) triglycerides ≥1.7 mmol/L, and 4) HDL cholesterol <1.0 mmol/L for men and <1.3 mmol/L for women (15). We used logistic regression models adjusted for age, sex, ethnicity, and BMI.

We applied the protein signatures in the T2DCC set to examine the association with incidence of obesity (BMI ≥25.0 kg/m²), metabolic syndrome (having two of four metabolic abnormalities as described earlier), and type 2 diabetes using logistic regression models adjusted for age, sex, and ethnicity. We also evaluated the predictive utility of the protein signatures to predict type 2 diabetes using the area under the receiver operating characteristic curve (AUROC) (Supplementary Material).

Using the ARIC study, we evaluated the population-specific associations between anthropometric measurements and each protein in the BMI and WC protein signatures, adjusted for age, sex, and study center. Subsequently, we pooled the estimates using a fixed-effects inverse-variance meta-analysis. We considered an adiposity-protein association in MEC1 replicated if it was statistically significant at the Bonferroni threshold (0.05/number of proteins in the signature) and directionally consistent with the ARIC study. Next, we computed the BMI and WC protein signatures for each ARIC participant using the coefficients identified in MEC1 and assessed their correlations with BMI and WC. Finally, we evaluated associations between protein signatures and type 2 diabetes incidence using the Cox proportional hazards model, adjusted for age, sex, population, and study center.

We performed protein subcellular location annotations using the UniProt Knowledgebase following an approach similar to that described previously (16,17) (Supplementary Material). We further included information on the specificity of the SomaScan platform based on previous studies (18,19) (Supplementary Material).

We queried the Gene Ontology (20), Kyoto Encyclopedia of Genes and Genomes (21), and Reactome (22) databases using the gprofiler2 R package (version 0.2.1) (23) to identify annotations and pathways significantly overexpressed in the protein signatures using the hypergeometric test at a false discovery rate (FDR) of 0.05 (24), with all measured proteins as the background. Results were visualized using the enrichplot R package (version 1.24.0).

Researchers can request data from the MEC1 study for scientific purposes through an application process on the listed website (https://blog.nus.edu.sg/sphs/data-and-samples-request/). Data will be shared through an institutional data-sharing agreement.

The baseline characteristics of study participants are listed in Supplementary Table 2. Participants were 44.0% male, and the mean (±SD) age was 47.6 (±12.0) years. The ethnic distribution was 34.8% Chinese, 32.6% Malay, and 32.7% Indian_.

Using the population set as discovery, we evaluated the ethnicity-specific associations between plasma protein levels with BMI and WC adjusted for age and sex (Supplementary Table 3). In the meta-analysis across ethnic groups, 410 proteins were associated with BMI, and 385 proteins were associated with WC, with 359 proteins in common that were directionally consistent (Fig. 2). Across the three ethnic groups, there was no evidence of statistical heterogeneity (P_het >1.00 × 10⁻⁵), and all associations were in the same direction except one protein (complement factor D) for WC, which was excluded from additional analyses. The top four most strongly associated proteins were the same for BMI and WC, namely, leptin, heart-type fatty acid binding protein, insulin-like growth factor–binding protein 1 (IGFBP1), and growth hormone receptor. We performed a lookup of the 410 BMI-associated proteins in three recent proteomic association studies in European adults (3–5). Of the 410 BMI-associated proteins observed in our study, 90 proteins were not measured in these three studies, and 263 (82%) of the remaining 320 BMI-associated proteins were directionally consistent and significant at the Bonferroni threshold in at least one study (Supplementary Table 3). Examples of novel proteins that were not previously reported to be associated with BMI include serine protease high-temperature requirement A1 (HTRA1), syndecan-3 (SDC3), and complexin-2 (CPLX2).

We performed feature selection using elastic net regression analysis based on the BMI- and WC-associated proteins. We identified 124 proteins for the BMI protein signature and 125 proteins for the WC protein signature, with 60 overlapping proteins in both signatures (Fig. 3 and Supplementary Table 4).

We validated the protein signatures identified in the population set using the T2DCC set. The BMI and WC protein signatures were strongly correlated with BMI and WC at baseline and follow-up (r range 0.778–0.842) (Table 1 and Supplementary Fig. 1). Among 814 participants with both baseline and follow-up data, changes in BMI (r = 0.632) and WC (r = 0.438) were directly correlated with changes in the respective protein signatures over the same period (Supplementary Fig. 2). All correlation coefficients reported here were significant (P < 0.001).

In a subset of Chinese participants with DXA and computed tomography measurements, the BMI protein signature was directly and strongly correlated with total body fat mass (r = 0.782) and moderately correlated with SAT (r = 0.596) and VAT (r = 0.676) (Table 1). In comparison with the BMI protein signature, the WC protein signature was less strongly correlated with total body fat mass (r = 0.619) and SAT (r = 0.433) but more strongly correlated with VAT (r = 0.735). The scatterplots suggested a direct linear relationship between total body fat mass and the BMI protein signature (β = 4.98 kg per SD; P < 0.001) and between VAT area and the WC protein signature (β = 48.1 cm² per SD; P < 0.001) (Fig. 4).

In the T2DCC set, all adiposity-related predictors (BMI, BMI protein signature, WC, and WC protein signature) were significantly associated with higher systolic blood pressure, LDL cholesterol, triglycerides, fasting glucose, and A1C and lower HDL cholesterol after adjustment for age, sex, and ethnicity (Supplementary Table 5). Protein signatures were significantly more strongly associated with HDL cholesterol, triglycerides, and A1C than anthropometric measurements. Moreover, with the exception of the BMI protein signature and LDL cholesterol, the associations between the BMI and WC protein signatures and CVD risk factors remained statistically significant after further adjustment for BMI or WC.

We evaluated the ability of our protein signatures to differentiate between metabolically healthy and unhealthy obesity. Among 900 obese (BMI ≥25 kg/m²) participants in the T2DCC set, 542 (60.8%) were metabolically unhealthy. After adjusting for age, sex, ethnicity, and BMI, having a higher BMI protein signature (per SD increment; odds ratio [OR] 2.00; 95% CI 1.62–2.47) or WC protein signature (OR 2.29; 95% CI 1.87–2.82) was associated with being metabolically unhealthy.

Among participants in the T2DCC set who were nonobese at baseline and had available data at follow-up (n = 414), both the BMI (OR 2.23; 95% CI 1.65–3.00) and WC protein signatures (OR 2.07; 95% CI 1.52–2.83) were associated with a higher incidence of obesity at follow-up. Results were similar when we used the international BMI cutoff of 30.0 kg/m² to define obesity (Supplementary Material).

Similarly, in participants without metabolic syndrome at baseline (n = 408), both the BMI (OR 2.98; 95% CI 2.23–4.00) and WC protein signatures (OR 3.29; 95% CI 2.44–4.45) were associated with a higher incidence of metabolic syndrome. Furthermore, the associations between the BMI and WC protein signatures and metabolic syndrome remained significant after further adjustment for BMI and WC, respectively

In the T2DCC set, the BMI (OR 2.44; 95% CI 2.15–2.77) and WC protein signatures (OR 2.84; 95% CI 2.47–3.25) were significantly associated with type 2 diabetes incidence, with larger effect sizes (P_diff < 0.05) than BMI (OR 1.91; 95% CI 1.70–2.14) and WC (OR 2.09; 95% CI 1.86–2.36) (Table 2). Moreover, the associations between the BMI and WC protein signatures and type 2 diabetes remained significant after further adjustment for BMI and WC, respectively.

Because the effects of adiposity on insulin resistance may be modified by ethnicity (25,26), we conducted additional analyses stratified by ethnicity (Supplementary Table 6). We found a stronger association between the BMI protein signature and type 2 diabetes in Malay (OR 2.96; 95% CI 2.33–3.77) and Chinese participants (OR 2.37; 95% CI 1.94–2.88) than in Indian participants (OR 1.91; 95% CI 1.55–2.35; P_interaction = 0.019). Similarly, the WC protein signature was more strongly associated with type 2 diabetes incidence in Malay (OR 3.59; 95% CI 2.77–4.67) and Chinese participants (OR 2.72; 95% CI 2.20–3.37) than in Indian participants (OR 2.21; 95% CI 1.76–2.77; P_interaction = 0.028).

We compared the use of the protein signatures versus BMI or WC in type 2 diabetes prediction models (Supplementary Material and Supplementary Fig. 3). The findings suggest that the protein signatures had significantly higher AUROC values compared with BMI or WC in unadjusted prediction models but did not significantly improve the AUROC compared with BMI or WC in prediction models that included other clinical predictors, such as fasting plasma glucose.

Participants in the ARIC study had a mean age of 56.7 (±5.7) years, were mostly women (57%), and were White (81%) or African American (19%) adults (Supplementary Table 7). In the meta-analysis across White and African American participants, 121 (97.6%) of 124 proteins in the BMI protein signature and 124 (99.2%) of 125 proteins in the WC protein signature were directionally consistent and significantly associated with BMI and WC, respectively, after Bonferroni adjustment (P < 0.05/124 for BMI and 0.05/125 for WC) (Supplementary Table 4 and Supplementary Fig. 4). Concordant with findings in MEC1, the BMI protein signature was highly correlated with BMI (r = 0.824), and the WC protein signature was highly correlated with WC (r = 0.774). Higher BMI protein signature (hazard ratio [HR] per SD increment 1.81; 95% CI 1.72–1.89) and WC protein signature (HR 1.98; 95% CI 1.88–2.08) at baseline were strongly associated with type 2 diabetes incidence over a median follow-up of 19.4 years. The association between the BMI protein signature and type 2 diabetes was stronger in White (HR 1.92; 95% CI 1.82–2.03) compared with Black (HR 1.56; 95% CI 1.42–1.72) participants (P_interaction <0.001). No significant interaction with population was observed for the association between WC protein signature and diabetes risk.

Based on the UniProt subcellular location annotation, 42.9% of the 189 proteins in the BMI and/or WC protein signatures were secreted proteins, 12.7% were membrane and secreted proteins, 30.7% were membrane proteins, and 10.7% were intracellular proteins.

We identified 36 significant annotations from the BMI protein signature and 28 significant annotations from the WC protein signature (Supplementary Table 8 and Supplementary Figs. 5 and 6). Pathways related to posttranslational protein phosphorylation, regulation of IGF transport and uptake by IGFBPs, and complement and coagulation cascades were significantly enriched (P_FDR <0.05) in both BMI and WC protein signatures. In addition, pathways involved in adenosine monophosphate–activated protein kinase (AMPK) signaling, extracellular matrix (ECM)–receptor interaction, and cell adhesion molecules (CAMs) were significantly enriched (P_FDR <0.05) in the WC protein signature.

We present the cis protein quantitative trait loci (pQTLs), confidence tiers, and confirmation by mass spectrometry from previous studies for the proteins included in our protein signatures (18,19) (Supplementary Table 9). In total, 83.8% of the proteins in the BMI protein signature and 83.2% of the proteins in the WC protein signature had cis pQTLs on the SomaScan platform. Furthermore, 62.1% of the proteins in the BMI protein signature and 53.6% of the proteins in the WC protein signature were classified as tier 1 measurements or were confirmed by mass spectrometry in terms of target specificity.

We identified 410 proteins associated with BMI and 385 proteins associated with WC that were consistent across all Chinese, Malay, and Indian ethnic groups in our study population. We derived protein signatures of BMI and WC, which we validated in an independent data set across two time points and externally in a cohort of U.S. White and African American participants. Compared with anthropometric measurements, our protein signatures were more strongly associated with cardiometabolic risk factors and were able to distinguish between metabolically healthy and unhealthy obesity. In prospective analyses, both protein signatures were strongly associated with the risk of type 2 diabetes in MEC1 and the ARIC study. Biologic pathways related to posttranslational protein phosphorylation, regulation of IGFs and IGFBPs, coagulation cascades, AMPK signaling, ECM-receptor interaction, and cell adhesion were overrepresented in the BMI and WC protein signatures.

Of the 410 proteins associated with BMI in our Asian cohort, >80% of the proteins measured in common were replicated in previous studies in European populations (3–5), highlighting the robustness of these findings to ethnic and geographic variations. We replicated several candidate proteins that were recently identified to be obesity-associated proteins in European populations. These proteins include kallistatin (3,4,16,27), E-selectin (3,4,16,28–30), seizure-6-like protein (31), reticulon-4 receptor (4,16,32), and neuronal growth regulator 1 (16,33,34). Previous studies suggested that kallistatin, E-selectin, and seizure-6-like protein may be involved in the regulation of inflammation and the immune system (29,35–37), and reticulon-4 receptor and neuronal growth regulator 1 may be involved in the regulation of energy metabolism through AMPK activation (32,38). In addition, we identified novel proteins that were not previously reported, which include HTRA1, SDC3, and CPLX2. HTRA1 has been shown to regulate the availability of IGFs (39,40). In addition, HTRA1 may inhibit adipogenesis by regulating the formation of adipocytes and could be an indicator of adipose tissue dysfunction (41). Variants of the SDC3 gene have been associated with obesity in Korean (42) and Taiwanese adults (43). SDC3 is involved in regulating appetite through the melanocortin system (44) and may therefore be a potential therapeutic target for treating obesity. CPLX2 may affect the translocation of glucose transporter 4, which plays a critical role in glucose uptake regulation (45). However, it should be noted that glucose transporter 4 translocation is an intracellular process, and whether plasma levels of CPLX2 are related to this process is unclear. In addition, CPLX2 may also have a role in the regulation of the immune system (46). These novel adiposity-related proteins, if replicated, can inform research on the pathophysiology of obesity and related cardiometabolic conditions.

A large proportion (85%) of the proteins significantly associated with WC and BMI in MEC1 overlapped. However, greater differences existed in proteins selected by machine learning for the BMI and WC signatures, with slightly fewer than half of the proteins overlapping. We further showed that the BMI protein signature was highly correlated with total body fat mass, whereas the WC protein signature was more strongly correlated with VAT. These findings suggest that the two protein signatures were able to capture the differences in plasma proteins associated with adiposity in different anatomic compartments.

Compared with anthropometric measures, the BMI and WC protein signatures were more strongly associated with HDL cholesterol, triglycerides, hemoglobin A_1c levels, and incidence of type 2 diabetes. In addition, the associations of these protein signatures with CVD risk factors and type 2 diabetes remained significant after further adjustment for BMI or WC. These findings suggest that the protein signatures are likely to capture physiologically relevant effects of body fat better than traditional measures of adiposity.

The identified plasma protein signatures differentiated between metabolically healthy and metabolically unhealthy obesity and were associated with type 2 diabetes risk. Thus, the biologic pathways represented by the adiposity-related proteins may provide mechanistic insights into the physiologic changes associated with metabolically unhealthy obesity (47). Pathways related to complement and coagulation cascades, IGF and IGFBP regulation, and posttranslational protein phosphorylation were significantly enriched in both the BMI and WC protein signatures. In addition, pathways involving AMPK signaling, ECM-receptor interaction, and CAMs were enriched in the WC protein signature. The dysregulation of the complement system resulting from overexpression of cytokines and adipose tissue–derived factors may lead to chronic inflammation and the development of metabolic disorders, such as insulin resistance and type 2 diabetes (48). Similarly, imbalances in levels of IGFs and IGFBPs have been linked to the dysregulation of fat metabolism and insulin resistance (49–51). AMPK is a key enzyme involved in regulating energy and nutrient metabolism, and the dysregulation of AMPK has been implicated in the development of obesity, insulin resistance, and type 2 diabetes (52). The overactivation of ECM-receptor signaling pathways may lead to inflammation and insulin resistance (53). Elevation of CAMs has been previously associated with obesity and type 2 diabetes (54,55) and may affect inflammation and atherogenesis (56). Taken together, the pathways enriched in the adiposity-related protein signatures depict a state of dysregulated cell signaling, systemic inflammation, and impaired glucose and fat metabolism. In addition, several of the pathways (e.g., complement pathways, ECM-receptor interaction, and IGFBP regulation) enriched in our adiposity-related protein signatures were also enriched in type 2 diabetes–associated proteins in European populations (57), underscoring the central role of obesity-related pathways in the pathogenesis of type 2 diabetes.

Previous studies have reported ethnic differences in the association between obesity and insulin resistance (25,58). Khoo et al. (25) demonstrated that the association of body fat percentage with insulin resistance was weaker in those of Indian ethnicity than in those of Chinese or Malay ethnicity residing in Singapore. Furthermore, Retnakaran et al. (58) found that prepregnancy BMI was more strongly associated with insulin resistance in East Asian women than in South Asian women in Canada. Here, although associations between plasma proteins and adiposity measures were consistent across ethnic groups, we observed stronger associations between the BMI and WC protein signatures and type 2 diabetes risk in Chinese and Malay participants than in Indian participants. These findings suggest that ethnic differences in the associations between obesity and type 2 diabetes may be explained by differential effects of obesity-related proteins on type 2 diabetes.

We acknowledge the limitations of our study. First, our study population included only ethnic Chinese, Malay, and Indian participants, and future studies in other Asian populations are needed. Second, ∼40% of the proteins in our protein signature are not secreted and may therefore be leakage proteins. Furthermore, a subset of the plasma proteins may be contained within extracellular vesicles, such as exosomes. Although the plasma proteome, including both leakage and exosomal proteins, has been shown to be informative about current and future states of health (18,30), the abundance of leakage and exosomal proteins in the plasma may not be a direct reflection of biologic activity in the cell or exosome. Third, we were not able to infer causality, because the associations between plasma proteins and anthropometric measurements were cross-sectional. Fourth, it is possible that some protein measurements on the SomaScan platform may have low specificity (59). Results from previous studies indicate that >80% of the proteins in the protein signatures identified in our study had cis pQTLs on the SomaScan platform, and more than half were measured with high confidence or validated by mass spectrometry (18,19). Still, a subset of protein measurements requires confirmation in future studies.

In conclusion, we provide a proof of principle that plasma protein markers can capture metabolically unhealthy adiposity better than anthropometric measures, which may have future implications for clinical practice. For example, protein biomarkers in a single blood sample, if confirmed, could be alternatives for anthropometric measures that reflect the detrimental metabolic effects of adiposity. This could be a valuable approach, because direct measurement of specific fat depots with scans is often not feasible in clinical practice (60). Our findings of consistent associations across diverse populations are highly relevant in this context, because the relationship between anthropometric measures and body fat depots differs greatly by race/ethnicity (61).

Acknowledgments. The authors thank all participants, the study team, and the investigators for their research contributions.

Funding. The MEC1 study is supported by individual research and clinical scientist award schemes from the Singapore National Medical Research Council (including MOH-000271-00) and the Singapore Biomedical Research Council and infrastructure funding from the Singapore Ministry of Health (Population Health Metrics and Analytics), National University of Singapore, and National University Health System, Singapore. The ARIC study has been funded in whole or in part with federal funds from the National Heart, Lung, and Blood Institute, National Institutes of Health, Department of Health and Human Services (75N92022D00001, 75N92022D00002, 75N92022D00003, 75N92022D00004, 75N92022D00005).

The funders had no role in the design, implementation, analysis, or interpretation of the data. The National University of Singapore (NUS) and John Hopkins University (JHU) have each signed separate collaboration agreements with SomaLogic to conduct SomaScan of MEC (NUS) and ARIC (JHU) stored samples at no charge in exchange for the rights to analyze linked MEC and ARIC phenotype data.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. C.G.Y.L., X.S., and R.M.v.D. contributed to the drafting of the manuscript. B.O., M.R.R., C.E.N., E.S.T., and J.Co. contributed to the critical revision of the manuscript. X.S. and R.M.v.D. designed the study. All authors contributed to the acquisition, analysis, or interpretation of data. X.S. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.