Proteins as Mediators of the Association Between Diet Quality and Incident Cardiovascular Disease and All-Cause Mortality: The Framingham Heart Study.

Background Biological mechanisms underlying the association of a healthy diet with chronic diseases remain unclear. Targeted proteomics may facilitate the understanding of mechanisms linking diet to chronic diseases. Methods and Results We examined 6360 participants (mean age 50 years; 54% women) in the Framingham Heart Study. The associations between diet and 71 cardiovascular disease (CVD)-related proteins were examined using 3 diet quality scores: the Alternate Healthy Eating Index, the modified Mediterranean-style Diet Score, and the modified Dietary Approaches to Stop Hypertension diet score. A mediation analysis was conducted to examine which proteins mediated the associations of diet with incident CVD and all-cause mortality. Thirty of the 71 proteins were associated with at least 1 diet quality score (P<0.0007) after adjustment for multiple covariates in all study participants and confirmed by an internal validation analysis. Gene ontology analysis identified inflammation-related pathways such as regulation of cell killing and neuroinflammatory response (Bonferroni corrected P<0.05). During a median follow-up of 13 years, we documented 512 deaths and 488 incident CVD events. Higher diet quality scores were associated with lower risk of CVD (P≤0.03) and mortality (P≤0.004). After adjusting for multiple potential confounders, 4 proteins (B2M [beta-2-microglobulin], GDF15 [growth differentiation factor 15], sICAM1 [soluble intercellular adhesion molecule 1], and UCMGP [uncarboxylated matrix Gla-protein]) mediated the association between at least 1 diet quality score and all-cause mortality (median proportion of mediation ranged from 8.6% to 25.9%). We also observed that GDF15 mediated the association of the Alternate Healthy Eating Index with CVD (median proportion of mediation: 8.6%). Conclusions Diet quality is associated with new-onset CVD and mortality and with circulating CVD-related proteins. Several proteins appear to mediate the association of diet with these outcomes.

adrenomedullin

Alternate Health Eating Index

beta‐2‐microglobulin

Dietary Approaches to Stop Hypertension

Food Frequency Questionnaire

growth differentiation factor 15

gene ontology

granulin

hemopexin

low‐density lipoprotein receptor

Mediterranean‐style Diet Score

matrix metallopeptidase

myeloperoxidase

plasminogen activator inhibitor 1

protein quantitative trait loci

soluble intercellular adhesion molecule 1

uncarboxylated matrix Gla‐protein

Clinical Perspective

What Is New?

A healthy diet is associated with circulating cardiovascular disease‐related protein biomarkers, largely representing regulators of inflammatory pathways in a group of middle‐aged and older participants in the Framingham Heart Study.

Four proteins—B2M (beta‐2‐microglobulin), GDF15 (growth differentiation factor 15), sICAM1 (soluble intercellular adhesion molecule 1), and UCMGP (uncarboxylated matrix Gla‐protein) may mediate the association of diet with health outcomes.

What Are the Clinical Implications?

Our findings provide novel evidence to better understand the mechanisms linking healthy diet with new‐onset cardiovascular disease and all‐cause mortality.

A healthy diet is recommended as an important lifestyle factor to reduce the risk of developing cardiovascular disease (CVD) and other chronic diseases. Substantial epidemiological evidence has shown that a healthy diet, as assessed by diet quality score, is associated with reduced risk of a broad range of clinical outcomes including CVD. For example, higher diet quality scores, estimated using the Alternate Healthy Eating Index (AHEI), the Mediterranean‐style Diet Score (MDS), or the Dietary Approaches to Stop Hypertension (DASH) diet score, were associated with lower incident coronary heart disease , , , and lower all‐cause mortality. Although the association between these diet quality scores and clinical outcomes has been well established, the underlying mechanisms remain elusive.

A proteomics approach may play a critical role in the elucidation of underlying mechanisms of the well‐observed diet‐disease associations. A previous FHS (Framingham Heart Study) investigation examined the association of 71 CVD‐related proteins with CVD and all‐cause mortality. Diet may play a pivotal role in regulating these proteins; however, current research examining the association of habitual diet quality with protein biomarkers is limited. To address this knowledge gap, we used data derived from the FHS, a large community‐based observational study, to test our hypothesis that the association between diet quality and new‐onset CVD and all‐cause mortality is mediated, at least partly, by CVD‐related proteins. We assessed the cross‐sectional association between diet quality and 71 CVD‐related proteins. Gene ontology (GO) is a widely used bioinformatic tool to systematically examine protein function and test the potential protein‐protein interaction. We, therefore, conducted GO functional enrichment analysis to better understand the potential biological mechanisms of diet‐associated proteins. We further tested the potential mediation effects of these proteins on the prospective association of diet quality with incident CVD and mortality.

Methods

Anonymized data and materials used for this analysis have been made publicly available through the database of Genotypes and Phenotypes repository and can be accessed at the following accession number: phs000007.v29.p10.

Study Participants

The present study included FHS participants who attended the seventh examination of the Offspring cohort (1998–2001; n=3530) or the first examination of the Third Generation cohort (2002‒2005; n=4095). , Among these participants (n=7350), 71 CVD‐related protein biomarkers were measured by the SABRe CVD (Systems Approach to Biomarker Research in Cardiovascular Disease) Initiative that was established by the National Heart, Lung, and Blood Institute. After exclusion of participants without Food frequency questionnaire (FFQ) data (n=750) and missing covariates (n=240) at baseline, we analyzed data collected from up to 6360 participants. Participant selection is shown in Figure 1. All FHS protocols and procedures were approved by the Institutional Review Board for Human Research at Boston University Medical Center and all participants provided written informed consent. The current analyses were approved by the Tufts University Institutional Review Board.

Figure 1

The flow diagram of participant selection and study overview.

The number of participants in each model was varied according to the presence of each protein data. CVD indicates cardiovascular disease.

Diet Quality Scores

The FHS used a previously validated 126‐item FFQ to assess habitual dietary intake for the year preceding each examination. The FFQ used in the present study was administered at the seventh examination of the Offspring cohort and the first examination of the Third Generation cohort. Dietary data were excluded if the reported energy intake was <2.5 MJ/day (600 kcal/day) for both men and women, ≥16.7 MJ/day (4000 kcal/day) for women, ≥17.5 MJ/day (4200 kcal/day) for men, or if more than ≥13 food items were left blank. The FFQ was used to calculate 3 commonly used diet quality scores: the AHEI, a modified DASH score, , and a modified MDS. , Although the 3 scores use different scoring strategies, they have largely similar components (Table S1), and higher scores represent a healthier diet.

The AHEI is composed of 11 components including vegetables, fruits, nuts and legumes, whole grains, red and processed meat, sugar‐sweetened beverages and fruit juice, eicosapentaenoic and docosahexaenoic acids, other polyunsaturated fatty acids (without eicosapentaenoic acid and docosahexaenoic acid), trans fatty acids, sodium, and alcohol. Each component score ranges from 0 (unhealthy) to 10 (healthy), with a higher score assigned to moderate alcohol drinking and higher intakes of vegetables, fruits, whole grains, nuts and legumes, eicosapentaenoic and docosahexaenoic acids, and polyunsaturated fatty acids. Lower scores are assigned to higher intakes of sugar‐sweetened beverages and fruit juice, red and processed meat, trans fatty acids, and sodium. The component scores are summed to generate the overall score ranging from possible value of 0 to 110, where a higher score reflects better concordance with the current Dietary Guidelines for Americans. ,

The modified DASH score is calculated on the basis of energy‐adjusted intakes of 8 dietary components including vegetables, fruits, nuts and legumes, whole grains, low‐fat dairy, red and processed meat, sugar‐sweetened beverages, and sodium. , For each component, with the exception of for red and processed meat, sugar‐sweetened beverages, and sodium, we assigned a value of 1 (lowest quintile of intake) to 5 (highest quintile of intake). The order of the scores was reversed for red and processed meat, sugar‐sweetened beverages, and sodium, whereby the highest quintile was assigned a value of 1 and the lowest quintile was assigned a value of 5. The scores of all 8 components were then summed to create an overall DASH score ranging from 8 to 40, with higher scores indicating better adherence to the DASH dietary pattern.

The modified MDS was calculated as a measure of adherence to the Mediterranean diet. , The MDS is composed of 9 components including vegetables, fruits, nuts, legumes, whole grains, fish, red and processed meat, ratio of monounsaturated to saturated fatty acids, and alcohol. Instead of using median intake as a threshold to dichotomize each component as described in a previous study, we categorized consumption of each component into sex‐ and cohort‐specific quartiles. A score of 0 to 3 was assigned to each component based on quartile rank, with the exception of red and processed meat and alcohol. For red and processed meat, the order of the scores was reversed (ie, the highest quartile was assigned a score of 0). For alcohol, we assigned a value of 1 to intake ≥10 and ≤25 g/day for men or ≥5 and ≤15 g/day for women and a value of 0 for all other intakes. All component scores were summed to generate an overall MDS, ranging from 0 (lowest diet quality) to 25 (highest diet quality).

Target Proteins

The selection of target proteins has been previously described. Briefly, the SABRe CVD Initiative measured 85 CVD‐related proteins. These proteins were quantified using a modified Sandwich ELISA method, multiplexed on a Luminex xMAP platform (Sigma‐Aldrich, St. Louis, MO), using frozen fasting plasma samples collected at the same time that the FFQ was administered. , Because >80% samples had values below the lower detection limits, 14 proteins were excluded from the present analysis. A 7‐point calibration curve (in triplicate) was used to calibrate protein quantification. Quality control (QC) used both the “High” and “Low” spike control (QC1 and QC2, respectively). The method provided acceptable reproducibility of assay performance (Table S2), with mean intra‐assay coefficient of variation of 7.8 (range: 2.3–28.9; interquartile range: 6.6) for QC1 and 6.8 (range: 1.2–15; interquartile range: 5.2) for QC2 and inter‐assay coefficient of variation of 8.9 (range: 2.5–21.9; interquartile range: 8.2) for QC1 and 7.9 (range: 2.3–24.5; interquartile range: 5) for QC2. Complete list and selection criteria of the 71 CVD‐related proteins analyzed in the present study are presented in Table S2.

Covariates and Clinical Outcome Ascertainment

Data on smoking status (never, past, or current), alcohol intake (servings/week), body mass index (BMI; kg/m2), and blood pressure (mm Hg) were obtained using questionnaires and physical examinations following standard protocols. Physical activity score was calculated based on the intensity and time spent for each type of activity assessed by the physical activity questionnaire. Participants are under continuous surveillance for CVD events and mortality. A panel of 3 physicians was formed to review all pertinent information including medical and hospital records, death certificates, communication with personal physicians, and next‐of‐kin interviews. Primary outcomes were all‐cause mortality and incident CVD including non‐fatal CVD (coronary heart disease, myocardial infarction, angina, coronary insufficiency, cerebrovascular accident, atherothrombotic infarction of the brain, transient ischemic attack, cerebrovascular disease, and intermittent claudication) and CVD death (fatal coronary heart disease and death due to stroke, peripheral arterial disease, heart failure, or other cardiovascular causes).

Statistical Analysis

Our 3 main analyses are as follows (Figure 1): (1) examine the cross‐sectional association between diet quality and target proteins, (2) visualize the interrelation between diet‐related proteins and biological pathways (ie, the functional network), and (3) perform a mediation analysis to test if diet‐associated proteins mediate the longitudinal association between diet quality and incident CVD and all‐cause mortality.

Diet‐Protein Association

We primarily analyzed the 3 diet quality scores (AHEI, DASH, and MDS) on a continuous scale. To facilitate comparisons of the associations between the 3 diet quality scores and proteins, we standardized each score by dividing its SD. Log‐transformed CVD‐related proteins (n=71) were regressed on age, sex, and cohort index to obtain residuals and the residuals were then inverse normal transformed to a mean of 0 and SD of 1 for subsequent statistical analysis. Linear mixed effect models (implemented using the R nlme package) were used to account for family structure in our study sample with adjustment for sex, age, energy intake, smoking status, physical activity score, alcohol intake, and BMI.

We used a 2‐step strategy to identify diet‐associated proteins. In the first‐step analysis, we examined diet‐protein associations in all study participants. We applied Bonferroni correction with adjustment for the number of proteins to account for multiple testing, that is, to determine if a protein was statistically significant, we required that this protein had 2‐tailed P<0.0007 (0.05/71). In the second‐step analysis, we conducted the internal validation tests to examine robustness of the significant proteins identified in the first‐step analysis. Based on pedigrees, we randomly divided our study participants into 2 independent subcohorts with the allocation ratio of 1:1. We considered a protein significant in the first‐step analysis to be a diet‐associated protein if that protein was associated with diet quality score at P<0.05 and had regression coefficients in the same direction in both subcohorts. In addition, we tested heterogeneity between the 2 subcohorts by calculating the Cochran's Q statistic using the R meta package and required that diet‐associated proteins should not have the heterogeneity P<0.05/the number of significant proteins in the first‐step analysis.

Construction of Functional Network

Diet‐related proteins were analyzed to identify closely related biological processes based on GO terms and interrelations of functional groups in biological network by using the Cytoscape plug‐in ClueGO, which enables the visualization of clustered networks and pathways. We used default selection criteria for relevant pathways, that is, a minimum of 3 proteins from the selected diet‐related proteins, which accounted for at least 4% from the total number of proteins in the GO terms. The statistical test was based on the 2‐sided hypergeometric option with a Bonferroni step‐down correction. The ClueGO network is created with Cohen's Kappa coefficient ≥0.5 and reflects the relationships between the terms based on the similarity of their associated proteins.

Mediation Analysis

In a previous FHS analysis, 46 of the 71 proteins were associated with CVD or all‐cause mortality. We therefore conducted a mediation analysis using a modified approach proposed by Huang and Yang to investigate whether a significant association between diet quality and clinical outcomes is mediated by these proteins. A linear mixed effect model was used to estimate the association between diet quality and CVD‐related proteins, and a mixed effect Cox proportional hazard model (implemented using R coxme package) was adopted to estimate the natural direct and indirect effects of diet quality on clinical outcomes. Family structure was accounted for by using a random intercept. We used the R code provided by the Huang and Yang to calculate the 95% CIs and P values for the nature direct effect, indirect effect, and total effect based on a resampling method taking random draws (repeated for 1E+6 times) from multivariate normal distribution of estimates for model parameters. We considered natural indirect effect (ie, mediation) statistically significant if P<0.05/the number of the proteins significantly associated with diet quality. The proportion of mediation by a target protein was calculated as the ratio of indirect effect to the sum of both direct and indirect effect.

In mediation analysis for incident CVD, participants with history of CVD at baseline were excluded. We adjusted 2 sets of covariates. Covariates in model 1 included sex, age, energy intake, smoking status, physical activity score, alcohol intake, BMI, systolic blood pressure, use of antihypertension medications, high‐density lipoprotein cholesterol, total cholesterol, and type 2 diabetes. In model 2, we additionally adjusted for education (with and without college education), smoking pack‐years, estimated glomerular filtration rate (mL/min per 1.732), family history of CVD, aspirin use, and, for women only, menopausal status, oral contraceptive use, and hormone replacement therapy. Similarly, 2 sets of covariates were adjusted for in the mediation analysis for all‐cause mortality. In model 1, we adjusted for the covariates included in the model 1 mediation analysis for incident CVD and history of CVD and cancer. In model 2, we included the same additional covariates adjusted for in the model 2 mediation analysis for incident CVD. We conducted sensitivity analysis with exclusion of the incident cases or death events occurring in the first 2 or 5 years after baseline. To test robustness of the proteins significant in the mediation analysis, we further adjusted for other diet‐associated proteins that were correlated with the significant proteins. In this multi‐marker analysis, we selected proteins with absolute pairwise Spearman correlation coefficients ≥0.3 and added these proteins as covariates.

We identified 71 independent cis‐pQTL variants (protein quantitative trait loci; linkage disequilibrium R 2 <0.1 and minor allele frequency >0.01) for 18 of the diet‐associated proteins based on our previous study. These cis‐pQTLs reside within 500kb from the transcription start site of the protein coding genes, suggesting that they may directly affect expression of the protein coding genes. , Therefore, analysis using cis‐pQTLs provides evidence to support the causal roles of diet‐associated proteins to clinical outcomes. Information on the 71 cis‐pQTL variants and the 18 diet‐associated proteins are presented in Table S3. We used mixed effect Cox models to examine the association of the 71 cis‐pQTL variants with incident CVD and all‐cause mortality in our study participants (n=7060). Covariates included in models were sex, age, and the first 2 genetic principal components. All statistical analyses were conducted using R statistical analysis software (version 3.5.0; R Foundation for Statistical Computing, available: http://www.R‐project.org).

Results

Participants' characteristics (n=6360; mean age 50 years; 54% women) at baseline are presented in Table 1. Women tended to have higher AHEI and DASH scores. Because MDS was constructed using sex‐specific quartiles, MDS scores were similar in both men and women. Higher diet quality scores tended to be associated with a lower proportion of current smokers, more college education, less alcohol intake, and lower BMI. The 3 diet quality scores were correlated, with Pearson r ranging from 0.67 to 0.75 (P<0.0001 for all; Figure S1). We found no significant difference in participants' characteristics and distribution of CVD‐related proteins in the 2 internal validation subcohorts after Bonferroni correction (Table S4).

Table 1

Baseline Characteristics of Participants According to Tertiles of Diet Quality Score (n=6360)

	AHEI				DASH				MDS
	T1	T2	T3	P‐trend*	T1	T2	T3	P‐trend	T1	T2	T3	P‐trend
Age, y	50±14	51±14	48±13	<0.0001	49±14	50±14	50±14	0.002	49±13	50±14	50±14	<0.0001
Women, n (%)	862 (41)	1156 (54)	1421 (67)	<0.0001	1503 (59)	736 (43)	682 (32)	<0.0001	1222 (54)	1133 (55)	1084 (53)	0.99
College educated, n (%)	751 (36)	918 (44)	1152 (55)	<0.0001	939 (37)	768 (45)	1114 (53)	<0.0001	780 (35)	923 (45)	1118 (56)	<0.0001
Smoking status, n (%)				0.195				0.003				<0.0001
Never	1104 (53)	1225 (58)	1271 (61)		1284 (51)	1003 (59)	1313 (63)		1124 (50)	1172 (58)	1304 (65)
Past	636 (30)	630 (30)	664 (32)		768 (31)	525 (31)	637 (31)		742 (33)	626 (31)	562 (28)
Current	358 (17)	242 (12)	163 (8)		451 (18)	176 (10)	136 (7)		384 (17)	235 (12)	144 (7)
Smoking pack‐years	66±156	31±102	17±78	<0.0001	65±154	28±98	14±69	<0.0001	60±147	34±113	17±78	<0.0001
Alcohol, servings/wk	7±11	5±7	5±5	<0.0001	7±10	5±7	4±6	<0.0001	6±10	5±7	5±6	0.0002
Physical activity score	38±8	37±7	37±7	0.02	38±8	37±7	38±7	0.64	37±8	38±7	38±7	0.03
Body mass index, kg/m2	28±5	28±6	26±5	<0.0001	28±6	28±5	26±5	<0.0001	28±6	28±5	27±5	<0.0001
Systolic blood pressure, mm Hg	123±17	122±17	119±17	<0.0001	122±16	122±18	119±17	<0.0001	121±17	122±18	120±17	0.04
Estimated glomerular filtration rate, mL/min per 1.732	93±19	92±19	93±18	0.902	94±19	92±19	92±18	<0.0001	95±19	92±19	92±18	<0.0001
Hypertension meds, n (%)	447 (21)	423 (20)	354 (17)	0.0002	503 (20)	350 (21)	371 (18)	0.06	406 (18)	444 (22)	374 (19)	0.57
High‐density lipoprotein, mmol/L	0.7±0.7	0.8±0.7	1.0±0.8	<0.0001	0.8±0.7	0.8±0.7	0.9±0.8	<0.0001	0.8±0.7	0.8±0.7	0.9±0.8	0.0001
Aspirin use, n (%)	450 (21)	446 (21)	399 (19)	0.053	511 (20)	359 (21)	425 (20)	0.981	412 (18)	399 (20)	484 (24)	<0.0001
Hormone replacement therapy, n (%)	161 (8)	176 (8)	230 (11)	0.0002	167 (7)	151 (9)	249 (12)	<0.0001	188 (8)	185 (9)	194 (10)	0.139
Postmenopausal status, n (%)	464 (22)	531 (25)	591 (28)	<0.0001	491 (20)	455 (27)	641 (31)	<0.0001	535 (24)	523 (26)	529 (26)	0.054
Oral contraceptive use, n (%)	154 (7)	271 (13)	369 (18)	<0.0001	219 (9)	241 (14)	334 (16)	<0.0001	287 (13)	258 (13)	249 (12)	0.721
Type 2 diabetes, n (%)	129 (6)	159 (8)	111 (5)	0.24	155 (6)	129 (8)	115 (6)	0.41	127 (16)	153 (8)	119 (6)	0.65
Cancer, n (%)	77 (4)	81 (4)	63 (3)	0.23	86 (3)	62 (4)	73 (3)	0.89	65 (3)	75 (4)	81 (4)	0.04
CVD, n (%)	137 (7)	131 (6)	102 (5)	0.02	148 (6)	112 (7)	110 (5)	0.40	118 (5)	132 (6)	120 (6)	0.29
Family history of CVD, n (%)	1564 (75)	1574 (75)	1506 (72)	0.039	1875 (75)	1247 (73)	1522 (73)	0.140	1696 (75)	1480 (73)	1468 (73)	0.080

Data were expressed as means±SDs or absolute numbers (percentage). AHEI, Alternate Healthy Eating Index; CVD, cardiovascular disease; DASH, Dietary Approaches to Stop Hypertension; MDS, Mediterranean‐style Diet Score; and T, tertile.

Test of linear trend across tertile categories of diet quality scores was performed by entering the median value of each tertile category into the model as a continuous variable Unadjusted P‐trends were analyzed by Cochran–Armitage trend tests for categorical variables and linear mixed effects models for continuous variables.

Cross‐Sectional Association of Diet Quality and CVD‐Related Proteins

After adjustment for sex, age, energy intake, smoking status, physical activity score, alcohol intake, and BMI, we found that 34 of the 71 proteins were significantly associated with at least 1 diet quality score at P<0.0007 (Bonferroni corrected P<0.05/71; Table S5). Of these proteins, 31 were associated with AHEI, 25 with DASH, and 14 with MDS. As expected, the observed diet‐protein association patterns were similar across the 3 diet quality scores, with pairwise Pearson correlations for t‐statistics of 0.92, 0.96, and 0.94, respectively (P<0.0001 for all; Figure S2). Correlation coefficients between the 34 proteins ranged from −0.34 to 0.68 with a mean correlation coefficient of 0.12 (Figure S3). Similar associations between diet quality scores and proteins were obtained after excluding the patients with prevalent cases of CVD and type 2 diabetes (n=673; Figure S4).

The internal validation analysis confirmed that 30 of the 34 proteins met our criteria (Figure 2 and Table S6), that is, associated with at least 1 diet quality score with same direction and P<0.05 in both subcohorts. Twenty‐eight proteins were associated with AHEI, 21 proteins were associated with DASH, and 12 proteins were associated with MDS. No significant heterogeneity was detected for the 30 proteins based on the Cochran's Q statistic P<0.001 threshold (0.05/34; Table S6).

Figure 2

Adjusted regression coefficients and corresponding 95% CI for the associations between standardized diet quality scores and CVD‐related proteins in all study participants.

Linear mixed effects model was adjusted for sex, age, energy intake, smoking status, physical activity score, alcohol intake, and body mass index. Regression coefficients are depicted with ● for AHEI, ▲ for DASH, and ■ for MDS. The horizontal lines represent 95% CIs. The complete name of the abbreviated proteins can be found in Table S2. ADM indicates adrenomedullin; AGP1, arabinogalactan protein 1; AHEI, Alternate Healthy Eating Index; ANGPTL3, angiopoietin‐like 3; APOB, apolipoprotein B; B2M, beta‐2‐microglobulin; CD14, cluster of differentiation 14; CNTN1, contactin 1; CRP, C‐reactive protein; CVD, cardiovascular disease; CXCL16, chemokine ligand 16; DASH, Dietary Approaches to Stop Hypertension; GDF15, growth differentiation factor 15; GMP140, granule membrane protein 140; GRN, granulin; HPX, hemopexin; IGF1, insulin‐like growth factor 1; IGFBP1, insulin‐like growth factor binding protein 1; LDLR, low‐density lipoprotein receptor; MCP1, monocyte chemoattractant protein 1; MDS, Mediterranean‐style Diet Score; MMP, matrix metallopeptidase; MPO, myeloperoxidase; PAI1, plasminogen activator inhibitor 1; sICAM1, soluble intercellular adhesion molecule 1; TIMP1, tissue inhibitor of metalloproteinases 1; and UCMGP, uncarboxylated matrix Gla‐protein.

Functional Network for Diet‐Related Proteins

GO functional enrichment analysis for the 30 diet‐associated proteins showed significant enrichment for 5 biological processes (Table S7). This analysis highlighted 3 functional groups (Figure 3). The most significant GO biological processes in these 3 functional groups were regulation of neuroinflammatory response (P=5.6×10−9 for GO:0150076), endothelial cell apoptotic process (P=3.5×10−6 for GO:2000351), and interleukin‐8 production (P=8.7×10−6 for GO:0032677). Seven proteins (out of 30) were highly enriched in these networks including GRN (granulin; GRN), sICAM1 (soluble intercellular adhesion molecule‐1; sICAM1), LDLR (low‐density lipoprotein receptor; LDLR), MMP8 (matrix metallopeptidase 8; MMP8), MMP9 (MMP9), IGF1 (insulin‐like growth factor 1; IGF1), and PAI1 (plasminogen activator inhibitor 1; SERPINE1).

Figure 3

Functional network of diet‐related proteins (n=30).

Three significant function groups including 5 enriched biological processes were identified, followed by a Bonferroni step‐down correction for multiple testing. Nodes indicate enriched GO terms and the same color of nodes means that they are in the same pathway function group. The most significant term is highlighted by a large name label for each group. All terms are compared with each other, and pathway function groups are defined using the Cohen's Kappa coefficients, a measure taking into account how many genes are shared between 2 terms. Each dot represents diet‐related target protein. Edges between nodes and dots represent interactions between protein and terms. The width of edges indicated the value of Cohen's Kappa coefficients. The network was generated by using ClueGO, a plug‐in of Cytoscape. The complete name of the abbreviated proteins can be found in Table S2. ANGPT1 indicates angiopoietin 1; CCL2, C‐C motif chemokine ligand 2; CD14, cluster of differentiation 14; CNTN1, contactin 1; CRP, C‐reactive protein; GRN, granulin; ICAM1, intercellular adhesion molecule 1; IGF1, insulin‐like growth factor 1; GO, gene ontology; LDLR, low‐density lipoprotein receptor; MMP, matrix metallopeptidase; and SERPINE1, serpin family E member 1.

Longitudinal Association of Diet Quality and Incident CVD and Mortality

After exclusion of participants with prevalent CVD at baseline, 5585 participants were included in the longitudinal analysis for incident CVD. We observed 413 nonfatal and 75 fatal CVD events, during a median follow‐up time of 12 years. The analysis for all‐cause mortality included 5890 participants and 512 death events that occurred during a median follow‐up time of 14 years.

Higher diet quality scores were significantly associated with lower risk of incident CVD (fatal and nonfatal) and all‐cause mortality (Table S8). Hazard ratios (HRs; 95% CI; P‐trend) for incident CVD were 0.84 (0.76–0.93; P<0.0001), 0.82 (0.73–0.91; P=0.0002), and 0.89 (0.80–0.98; P=0.025) per SD increase of AHEI (12 points), DASH (5 points), and MDS (4 points), respectively. The HRs (95% CI; P‐trend) for all‐cause mortality were 0.86 (0.78–0.95; P=0.004), 0.85 (0.77–0.94; P≤0.002), and 0.86 (0.77–0.95; P=0.002) per SD increase of AHEI, DASH, and MDS, respectively. In sensitivity analyses, exclusion of events occurring within 2 or 5 years after baseline resulted in similar associations of diet quality scores with incident CVD and all‐cause mortality (Table S8).

We examined the 3‐way associations between diet quality scores, proteins, and CVD and all‐cause mortality. Because of the inverse association between diet quality and incidence CVD and all‐cause mortality, we expected that a protein inversely associated with diet quality scores was likely to be positively associated with the clinical outcomes or vice versa. Overall, the observed cross‐sectional association between diet quality scores and proteins are consistent with the findings in longitudinal association or Mendelian randomization analysis conducted in previous FHS analyses (Table S9). For example, we observed that better diet quality was associated with lower levels of cystatin C. In the prior FHS studies, Mendelian randomization analysis showed that higher levels of cystatin C were associated with increased risk of coronary heart disease, and longitudinal analysis showed higher levels of cystatin C were associated with increased risk of incident CVD and all‐cause mortality.

In addition, we analyzed the association of 71 independent cis‐pQTL variants for 18 diet‐associated proteins with incident CVD and all‐cause mortality. We found that participants who carried A allele of rs10908589, a cis‐pQTL variant of CD5L, had increased CD5L levels (P=7.9×10−26) and increased all‐cause mortality (HR, 1.28; 95% CI, 1.13–1.44; P=6.4×10−5; Table S10). This observation was consistent with the expected 3‐way associations, that is, higher diet quality scores were associated with lower CD5L levels and lower all‐cause mortality. Two other independent cis‐pQTLs associated with CD5L and 2 cis‐pQTLs associated with GMP140 (granule membrane protein 140) and CD14 (cluster of differentiation 14) with nominal significance in the association analyses are also presented in Table S10.

Mediation Analysis of Target Proteins in Relation to Diet Quality and Clinical Outcomes

Among the 30 diet‐associated proteins, 17 proteins were selected on the basis of significant associations with diet quality scores and with CVD outcomes and mortality. With adjustment for the mediation analysis model 1 covariates, 6 proteins (GDF15 [growth differentiation factor 15], UCMGP [uncarboxylated matrix Gla protein], sICAM1, ADM [adrenomedullin], CRP [C‐reactive protein], and B2M [beta‐2‐microglobulin]) significantly mediated the association between at least 1 diet quality score and all‐cause mortality at P for indirect effect <0.003 (0.05/17 proteins; Tables S11 through S13). Among them, UCMGP significantly mediated the association of all‐cause mortality with AHEI (P=1.6×10−5), DASH (P=1.2×10−4), and MDS (P=9.0×10−5). The median proportion of mediation (Table 2 and Table S14) by UCMGP was 20.7% (95% CI, 11.2–43.3%), 21.0% (95% CI, 9.4–82.9%), and 17.4% (95% CI, 8.8–47.4%) for AHEI, DASH, and MDS, respectively. Additional adjustment for education, smoking pack‐years, estimated glomerular filtration rate, family history of CVD, aspirin use, and, for women only, menopausal status, oral contraceptive use, and hormone replacement therapy reduced the strength of the mediation effect (Table 2). Nonetheless, mediation P values for all‐cause mortality remained significant or nominally significant for GDF15, UCMGP, B2M, and sICAM1.

Table 2

Significant Mediation Effect of Diet‐Associated Proteins on Longitudinal Associations of Diet With All‐Cause Mortality and Incident CVD

Diet quality	Mediator	Hazard ratio (95% CI)		Model 1		Model 2
				P value	Mediated proportion, %	P value	Mediated proportion, %
All‐cause mortality
AHEI	GDF15	0.957	(0.941–0.972)	1.2E‐16	30.1	0.003	21.8
	UCMGP	0.966	(0.953–0.979)	1.6E‐05	20.7	0.002	24.0
	Adrenomedullin	0.985	(0.977–0.992)	0.002	10.4	0.11
	CRP	0.985	(0.976–0.993)	0.002	9.9	0.06
	Beta‐2‐microglobulin	0.984	(0.975–0.992)	0.001	9.9	0.02	10.3
	Soluble intercellular adhesion molecule	0.987	(0.979–0.993)	0.002	9.3	0.03	8.6
Dietary Approaches to Stop Hypertension	UCMGP	0.979	(0.969–0.987)	1.2E‐04	21.0	0.003	25.9
	CRP	0.986	(0.977–0.993)	0.002	15.2	0.06
Mediterranean‐style Diet Score	UCMGP	0.978	(0.968–0.987)	9.0E‐05	17.4	0.003	19.1
Incident CVD
AHEI	GDF15	0.982	(0.971–0.991)	0.002	11.3	0.02	8.6

Linear mixed effect and mixed effect Cox proportional hazard models were adopted to estimate the indirect (mediation) effect. Hazard ratios per 1 increase of SD of standardized diet quality score and P values were derived from mixed effect Cox proportional hazard models. Model 1 was adjusted for sex, age, energy intake, smoking status, physical activity score, alcohol intake, body mass index, systolic blood pressure, use of antihypertension medications, high‐density lipoprotein and total cholesterol, type 2 diabetes, and history of CVD and cancer. Model 2 was additionally adjusted for estimated glomerular filtration rate, smoking pack‐years, aspirin use, education, family history of CVD, use of hormone replacement therapy, postmenopausal status, and oral contraceptive use. The median proportion of mediation was calculated as the ratio of indirect effect to the sum of both direct and indirect effect. Complete mediation analysis results are in the Tables S11 through S17. AHEI indicates Alternate Healthy Eating Index; CRP, C‐reactive protein; CVD, cardiovascular disease; growth differentiation factor 15; and UCMGP, uncarboxylated matrix gamma‐carboxyglutamic acid protein.

Mediation analysis results for incident CVD are presented in Tables S15 through S17. GDF15 was significant in the mediation analysis for AHEI and incident CVD with P of 0.002 after adjusting for model 1 covariates (Table S15). The median proportion of mediation by GDF15 was 11.3% (95% CI, 4.9–31.6%). Similarly, additional adjustment for model 2 covariates reduced the strength of the mediation effect (Table 2), mediation P=0.02, and proportion of mediation by GDF15 was 8.6% (95% CI, 2.8–31.2%).

In sensitivity analyses, exclusion of events occurring within 2 or 5 years after baseline did not substantially change the strength of the mediation analysis results (Tables S18 and S19). After adjusting for correlated diet‐associated proteins, mediation P values for the 4 proteins (B2M, sICAM1, GDF15, and UCMGP) remained similar in the mediation analysis for all‐cause mortality with adjustment for model 2 covariates (Table S20). Also, mediation P value for GDF15 was similar tos that in the model 2 mediation analysis for incident CVD (Table S21).

Discussion

We showed that diet quality, represented by 3 diet quality scores, was associated with 30 CVD‐related protein biomarkers in a large group of middle‐aged and older participants in the FHS. The majority of diet‐associated proteins were involved in biological pathways related to inflammatory response. Our mediation analysis demonstrated that 6 proteins (ADM, B2M, GDF15, UCMGP, sICAM1, and CRP) may mediate the association between diet quality scores and all‐cause mortality. In addition, GDF15 significantly mediated the longitudinal association between diet quality scores and new‐onset CVD. Our study provides novel evidence that targeted proteomic analysis may be useful to highlight molecular pathways underlying the beneficial effects of healthy diet for disease prevention.

Our findings are consistent with the literature examining the beneficial effects of healthy diet on CVD and mortality. Although the association between diet and protein biomarkers has not been well studied, we observed associations similar to those reported in previous studies. , The cross‐sectional Toronto Nutrigenomics and Health study demonstrated that a Western‐style dietary pattern was associated with 25 proteins involved in coagulation and lipid metabolism among 54 putative CVD biomarker proteins. A Swedish study of 2 population‐based cohorts analyzed 184 CVD‐related circulating proteins and demonstrated that dietary patterns were associated with 21 proteins. These dietary pattern‐associated proteins are involved in multiple pathways such as inflammation and lipid metabolism. In the present study, we observed diet‐protein associations consistent with those reported in 3 prior studies. , , For example, the unhealthy dietary patterns characterized in these studies were associated with higher concentrations of PAI1 (plasminogen activator inhibitor 1), MPO (myeloperoxidase), APOB (apolipoprotein B), GDF15, and HPX (hemopexin). We found that higher diet quality scores, reflecting healthier diet, were associated with lower levels of these proteins.

The biological mechanisms underlying the relationship between healthy diet and CVD and other chronic diseases are not fully understood. It is postulated that adherence to healthy diet has the potential to reduce vascular damage by alleviating inflammatory responses. A strong inverse association between overall diet quality and inflammation biomarkers has been demonstrated by few studies. , A study of the Women's Health Initiative cohort showed that higher MDS was favorably associated with a series of inflammation markers, including CRP and sICAM‐1, which explained ≈30% of the observed MDS‐CVD association. Another study, also conducted in the Women's Health Initiative showed that GDF15 may partly explain dietary effects on all‐cause mortality. CRP and sICAM1 are well‐known inflammation markers, whereas ADM, B2M, and GDF15 have been linked with inflammatory response. , , These data are consistent with our observations in the mediation analysis and, therefore, provide evidence to support inflammation as an important mechanism underlying the nexus between diet quality and human health.

The present study showed that UCMGP may be a strong mediator with respect to the association of diet quality with all‐cause mortality. MGP is primarily secreted by vascular smooth muscle cells in the arterial wall. The inactive form of MGP, UCMGP, undergoes posttranslational modifications depending upon availability of vitamin K. Dark green leafy vegetables, which are important constituents of healthy diet, are rich in vitamin K. Consistent with our observations, UCMGP is a risk factor for arterial calcification and has been associated with an increased risk of mortality. An observational study showed findings consistent with our results, highlighting that proteins such as UCMGP may partly explain dietary effects on health outcomes.

The strengths of the present study include the use of comprehensive dietary, lifestyle, and clinical data; long‐term follow‐up of clinical outcomes; and well‐quantified circulating targeted protein biomarkers in a large group of FHS participants. There are several limitations that warrant discussion. We examined the cross‐sectional association between diet quality scores and target proteins, which limits our ability to infer causality between diet and protein biomarkers. Nevertheless, by examining the 3‐way associations between diet quality score, proteins, and clinical outcomes, the cross‐sectional analysis showed consistency with longitudinal analysis of diet‐disease and protein‐disease associations. In addition, the cross‐sectional analysis was in line with associations generated from analyses using genetic variants. Our study population was predominately middle‐aged and older White adults, which may limit the generalizability of the present findings to other populations. Dietary intake was assessed using semiquantitative FFQ, which may lead to misclassification. In addition, dietary constituents that were not selected for constructing the diet quality scores may play important roles in the regulation of protein biomarkers and affect the risk of CVD and mortality. Dietary quality may change over time; therefore, further analysis considering the change in diet quality is warranted. Proteins may mediate the relationship between diet quality and non‐CVD mortality, which also need to be examined in future studies. Despite the fact that multiple potential confounders were adjusted for in the present analysis, the possibility of residual confounding could not be ruled out.

CONCLUSIONS

In conclusion, we demonstrated that diet quality, represented by 3 diet quality scores, was associated with 30 CVD‐related, circulating protein biomarkers. We further showed that several proteins significantly mediated the long‐term association of diet quality with incident CVD and all‐cause mortality. Our findings provide novel evidence to better understand the mechanisms underlying the observed association of diet with CVD and all‐cause mortality.

Sources of Funding

This research was supported by grants from the National Heart, Lung and Blood Institute (NHLBI)'s Framingham Heart Study (Contract No. N01‐HC‐25195) and by the NHLBI Career Transition Award (1K22HL135075‐01). Dr Ho was supported by National Institutes of Health grants R01‐HL134893 and R01‐HL140224. The funding sources had no role in study design, collection, analysis, or interpretation of data; writing of the report; or the decision to submit the article for publication.

Disclosures

None.

Tables S1–S21

Figures S1–S4

Click here for additional data file.