Protein Biomarkers of Cardiovascular Disease and Mortality in the Community.

BACKGROUND: The discovery of novel and highly predictive biomarkers of cardiovascular disease (CVD) has the potential to improve risk-stratification methods and may be informative regarding biological pathways contributing to disease. METHODS AND
RESULTS: We used a discovery proteomic platform that targeted high-value proteins for CVD to ascertain 85 circulating protein biomarkers in 3523 Framingham Heart Study participants (mean age, 62 years; 53% women). Using multivariable-adjusted Cox models to account for clinical variables, we found 8 biomarkers associated with incident atherosclerotic CVD, 18 with incident heart failure, 38 with all-cause mortality, and 35 with CVD death (false discovery rate, q<0.05 for all; P-value ranges, 9.8×10-34 to 3.6×10-2). Notably, a number of regulators of metabolic and adipocyte homeostasis were associated with cardiovascular events, including insulin-like growth factor 1 (IGF1), insulin-like growth factor binding protein 1 (IGFBP1), insulin-like growth factor binding protein 2 (IGFBP2), leptin, and adipsin. In a multimarker approach that accounted for clinical factors, growth differentiation factor 15 (GDF15) was associated with all outcomes. In addition, N-terminal pro-b-type natriuretic peptide, C-reactive protein, and leptin were associated with incident heart failure, and C-type lectin domain family 3 member B (CLEC3B; tetranectin), N-terminal pro-b-type natriuretic peptide, arabinogalactan protein 1 (AGP1), soluble receptor for advanced glycation end products (sRAGE), peripheral myelin protein 2 (PMP2), uncarboxylated matrix Gla protein (UCMGP), kallikrein B1 (KLKB1), IGFBP2, IGF1, leptin receptor, and cystatin-C were associated with all-cause mortality in a multimarker model.
CONCLUSIONS: We identified numerous protein biomarkers that predicted cardiovascular outcomes and all-cause mortality, including biomarkers representing regulators of metabolic homeostasis and inflammatory pathways. Further studies are needed to validate our findings and define clinical utility, with the ultimate goal of improving strategies for CVD prevention.

Clinical Perspective

What Is New?

Numerous protein biomarkers predict cardiovascular outcomes and all‐cause mortality, including biomarkers representing regulators of metabolic homeostasis and inflammatory pathways.

What Are the Clinical Implications?

Our findings highlight the power of targeted proteomics in refining current understanding of cardiovascular risk.

Further studies are needed to validate our findings and define clinical utility, with the ultimate goal of improving strategies for cardiovascular disease prevention.

Introduction

Cardiovascular disease (CVD) is the leading cause of death in the United States and is also emerging as the leading cause of death in developing countries in light of rapid epidemiological transitions over the past 2 decades.1 This has brought primary prevention of CVD to the forefront. One of the major challenges in developing preventive strategies is that current CVD risk assessment algorithms have limited predictive value and are subject to miscalibration when applied to different populations.2 The discovery of novel and highly predictive biomarkers of CVD has the potential to improve risk stratification and enable targeted prevention strategies in the preclinical phase of CVD, when intervention is most likely to be effective. In addition, circulating biomarkers may be informative regarding causal biological pathways contributing to disease, and hold the promise of future pathway‐specific therapies and personalized approaches to treatment.

To that end, the Systems Approach to Biomarker Research in Cardiovascular Disease (SABRe CVD) Initiative was established by the National Heart, Lung, and Blood Institute, to identify biomarker signatures of atherosclerotic CVD and its risk factors.3 The SABRe CVD initiative included both discovery and targeted proteomics. We hypothesized that a multimarker protein panel can improve CVD risk prediction above and beyond the established clinical risk factors. To that end, we measured 85 high‐value candidate protein biomarkers for CVD in participants in the FHS (Framingham Heart Study). Biomarkers were formally nominated and selected based on evidence of association with atherosclerotic CVD (ASCVD) from comprehensive literature review, gene expression profiling (FHS and others), published genome‐wide association studies of myocardial infarction or coronary heart disease, and discovery proteomics (FHS and others).

Methods

Study Sample

Analytical methods and study materials will not be made available to other researchers for purposes of reproducing the results or replicating the procedure. Anonymized data have been made publicly available at dbGaP (study accession phs00007.v29.p10). The FHS is a prospective, longitudinal, community‐based observational cohort study. The Offspring cohort was recruited beginning in 1971 and the Third Generation cohort started in 2002. We included participants in the Offspring cohort attending the seventh examination cycle (1998–2001, n=3539) and Third Generation participants attending their first examination (2002–2005, n=4095).4, 5 Because deaths and CVD events were exceedingly rare in individuals aged <50 years, we excluded individuals below this age (n=3833). In addition, we excluded individuals who had missing clinical covariates (n=238), no available biomarker measures (n=21), or missing follow‐up (n=19), leaving 3523 individuals for analyses examining all‐cause mortality. For each analysis, prevalent disease was excluded for the corresponding end point (n=412 for atherosclerotic CVD [ASCVD], n=37 for heart failure [HF]). The study was approved by the institutional review board of Boston University Medical Center, and all participants provided written informed consent.

Clinical Assessment

Participants underwent a comprehensive medical history, anthropometry, and physical examination. Resting, seated blood pressure was obtained by a physician and reported as the average of 2 measurements. Diabetes mellitus was defined as a fasting glucose ≥126 mg/dL or the use of hypoglycemic medications for treating hyperglycemia. Current smoking was defined as smoking >1 cigarette daily on average during the past year.

Biomarker Measures

Blood plasma samples were obtained at the baseline clinical visit, and immediately processed and kept at −80°C until assayed. Candidate biomarkers were selected based on the following criteria: (1) association with ASCVD from a review of published studies; (2) targeting proteins coded for by genes associated with ASCVD in genome‐wide association studies; (3) targeting genes associated with ASCVD or its major risk factors in gene expression analyses; and (4) discovery proteomics in the FHS or elsewhere (Table S1). A total of 85 biomarkers were then assayed using a modified ELISA sandwich approach, multiplexed on a Luminex xMAP platform (Sigma‐Aldrich, St. Louis, MO; Details in Methods S1). A total of 17 different multiplex panels were generated over 4 years based on several factors, including dilution rate, cross‐reactivity, and when the target was added to the assay list. Standard Luminex assays with detailed general methods published previously were used.6, 7 Methods for antibody conjugation and multiplex assay development were conducted by contract lab (Sigma‐Aldrich) and used protocols recommended and developed by Luminex. These include critical steps for measuring sensitivity, specificity, precision, accuracy, and linearity. All targets were initially developed as a singleton assay before compatible targets were combined to “multiplex panels” and further tested for cross‐reactivity and recovery. Commercial reagents were used for assay development when available, but small‐scale protein production procedures were used to generate reference proteins when needed. Measurements were calibrated with a 7‐point calibration curve (in triplicate) and tested for recovery at both ends of the quantitation scale. Both the “Hi” and “Lo” spike control (QC1 and QC2, respectively) were used to QC each marker and provide inter‐ and intra‐assay coefficients of variation of assay performance from the production runs. For further details, please see Methods S1. For low‐abundance biomarkers, depletion of high‐abundance proteins was performed using ProteoPrep 20 (Sigma‐Aldrich), an antibody‐based resin designed to deplete 95% of total protein from plasma. Assay performance characteristics including medians, interquartile ranges, and coefficients of variation are summarized in Table S1. As an internal control, we examined traditional biomarkers (N‐terminal pro‐b‐type natriuretic peptide [NT‐proBNP] and C‐reactive protein [CRP]) in prevalent CVD cases versus noncases and found highest levels of both biomarkers among those with HF, followed by CVD, with much lower levels among noncases (Table S2). In addition, we found strong correlations of CRP (r=0.95, n=3282 Offspring participants, P<0.0001; r=0.86, n=4048 Third Generation, P<0.0001) and NT‐proBNP with previous standard assays (r=0.81; n=4006; P<0.0001).

Clinical Outcome Ascertainment

Participants were followed with annual health history updates, and all medical records relevant to CVD outcomes were reviewed. Cardiovascular events were adjudicated by a 3‐physician panel. Primary outcomes were (1) ASCVD, a composite end point of nonfatal myocardial infarction, revascularization (percutaneous coronary intervention or bypass surgery), atherothrombotic stroke, and coronary heart disease death, (2) HF,8 and (3) all‐cause mortality, with a secondary end point of CVD‐related death.

Statistical Analysis

We recoded 14 biomarkers into binary variables (below/above detection limit) because of low numbers above detection limit, and rank normalized the remaining 71 biomarkers. To relate each biomarker to each outcome, we used multivariable Cox proportional hazards regression models. We ran 2 models accounting for baseline covariates: (1) adjusted for age and sex and (2) also adjusted for systolic blood pressure, antihypertensive drug treatment, diabetes mellitus, body mass index, smoking status, total and high‐density lipoprotein cholesterol, and prevalent atrial fibrillation. In HF and mortality analyses, we also adjusted for prevalent myocardial infarction. To account for multiple testing in single biomarker models, we set a false discovery rate (FDR; q‐value) <0.05.9 Next, we constructed multimarker models using a step‐wise approach: We considered for entry markers with single‐marker P<0.10 and used <0.01 for a marker to enter and stay in the step‐wise model; we forced in the clinical covariates. For each outcome, we compared the multimarker model versus the base model (clinical covariates only) using a likelihood ratio test. In each of 1000 bootstrap samples, we chose markers using step‐wise selection and we calculated frequency of selection for each marker as evidence of its importance as a predictor.

To assess the incremental benefit for predicting outcome events, we compared C‐statistics between covariates‐only and multimarker models.10 We also calculated integrated discrimination improvement and net reclassification improvement metrics.11 For net reclassification improvement, we used 2 categories with the threshold set at observed event rate. We conducted all analyses using SAS software (v9.4; SAS Institute Inc, Cary, NC).

Results

Baseline characteristics of the 3523 participants (mean age, 62 years; 53% women) are displayed in Table 1. During follow‐up (median, 14.3 years; Q1=11.4 and Q3=15.2 years), 755 participants died, including 167 from CVD death. Over the follow‐up period, there were 392 incident ASCVD events, and 226 participants developed incident HF. Absolute ASCVD event rates were 4% at 5 years and 8% at 10 years. Baseline blood samples were analyzed for a panel of 85 distinct biomarkers, of which 71 had detectable levels for >95% of participant samples, with mean interassay coefficient of variation of 8.9±5.0% and mean intra‐assay coefficient of variation 7.8±4.9% (Table S1). The majority of biomarkers were weakly correlated with one another, with correlation coefficients ranging from −0.2 to 0.2 (Table S2). Two percent of pair‐wise correlations were modest or strong with r>0.40. The strongest pair‐wise correlations were observed for cadherin 13 (CDH13) and ADAM metallopeptidase domain‐containing protein 15 (ADAM15; r=0.90).

Table 1

Baseline Characteristics of Participants

Clinical Characteristic	N=3523
Age, y	62 (8)
Women, n (%)	1879 (53)
Systolic blood pressure, mm Hg	128 (18)
Diastolic blood pressure, mm Hg	75 (10)
Body mass index, kg/m2	28.1 (5.3)
Diabetes mellitus, n (%)	402 (11)
Hypertension treatment, n (%)	1239 (35)
Current smoking, n (%)	434 (12)
Total cholesterol, mg/dL	201 (37)
HDL cholesterol, mg/dL	54 (17)
Prevalent myocardial infarction, n (%)	141 (4)
Prevalent heart failure, n (%)	37 (1)
Prevalent atrial fibrillation, n (%)	132 (4)

Data shown are means and SDs, unless otherwise noted. HDL indicates high‐density lipoprotein.

Multiple Biomarkers Predict Incident Cardiovascular Events

In multivariable‐adjusted single‐marker analyses, 8 biomarkers were positively associated with incident ASCVD (FDR‐q <0.05 for all; P‐value range from 3.6×10−6 to 2.6×10−3), including growth differentiation factor 15 (GDF15), tissue inhibitor of metalloproteinase‐1 (TIMP1), and beta‐2‐microglobulin (B2M; Table 2; full results presented in Table S3). Similarly, there were 18 biomarkers associated with incident HF (FDR‐q <0.05 for all; P‐value range from 2.4×10−15 to 6.3×10−3), including NT‐proBNP, GDF15, and adrenomedullin (ADM; Table 2). Of these protein biomarkers, higher concentrations of C‐type lectin domain family 3 member B (CLEC3B; tetranectin) were associated with a lower risk of HF (hazard ratio [HR], 0.82 per 1‐SD increase in CLEC3B; 95% confidence interval [CI], 0.71–0.95; P=6.6×10−3), whereas all other 17 markers were associated with a higher risk of HF. There were 35 biomarkers associated with CVD death (FDR‐q <0.05 for all; Pvalues from 1.6×10−18 to 3.0×10−2), and 38 biomarkers predicted all‐cause mortality (FDR‐q <0.05 for all; P values from 9.8×10−34 to 3.6×10−2; Table 2). Of the 38 biomarkers associated with mortality, CLEC3B, insulin‐like growth factor 1 (IGF1), butyrylcholinesterase (BCHE), paraoxonase 1 (PON1), insulin‐like growth factor binding protein 1 (IGFBP3), CNTN1, kallikrein B1 (KLKB1), and peripheral myelin protein 2 (PMP2) were associated with lower risk of mortality, whereas all others predicted a higher risk of mortality. Age‐ and sex‐adjusted analyses are presented in Table S4.

Table 2

Single Biomarker Multivariable‐Adjusted Associations With Cardiovascular Outcomes

	Biomarker	HR	95% LCL	95% UCL	P Value
Atheroslerotic CVD (N=392)	GDF15	1.38	1.20	1.58	3.6E‐06
	TIMP1	1.32	1.17	1.48	5.5E‐06
	B2M	1.24	1.10	1.40	3.6E‐04
	REG1A	1.19	1.07	1.32	8.5E‐04
	Cystatin‐C	1.21	1.08	1.36	1.1E‐03
	d_Troponin	1.82	1.26	2.63	1.5E‐03
	AGP1	1.19	1.06	1.33	2.2E‐03
	sICAM1	1.17	1.06	1.30	2.6E‐03
Heart failure (N=226)	NT‐proBNP	1.98	1.67	2.34	2.4E‐15
	GDF15	2.08	1.72	2.53	9.1E‐14
	ADM	1.47	1.25	1.73	2.5E‐06
	B2M	1.47	1.24	1.73	4.5E‐06
	Cystatin‐C	1.43	1.22	1.67	6.5E‐06
	CRP	1.38	1.19	1.60	2.5E‐05
	TIMP1	1.39	1.18	1.63	6.0E‐05
	IGFBP1	1.37	1.17	1.61	1.2E‐04
	CD14	1.31	1.14	1.51	1.5E‐04
	MPO	1.30	1.13	1.48	1.4E‐04
	UCMGP	1.32	1.14	1.52	1.8E‐04
	EFEMP1	1.34	1.14	1.57	3.2E‐04
	GRN	1.29	1.12	1.48	4.2E‐04
	Adipsin	1.31	1.12	1.52	5.8E‐04
	IGFBP2	1.30	1.11	1.51	8.3E‐04
	Resistin	1.25	1.09	1.43	1.1E‐03
	A1M	1.21	1.06	1.39	4.5E‐03
	CLEC3B	0.82	0.71	0.95	6.3E‐03
All‐cause mortality (N=755)	GDF15	1.96	1.76	2.17	1.2E‐35
	NT‐proBNP	1.43	1.31	1.57	2.1E‐15
	B2M	1.39	1.27	1.52	1.0E‐12
	TIMP1	1.36	1.24	1.48	1.0E‐11
	IGFBP2	1.34	1.23	1.46	9.5E‐12
	UCMGP	1.32	1.22	1.43	1.8E‐11
	ADM	1.33	1.22	1.45	2.5E‐10
	EFEMP1	1.28	1.17	1.40	2.6E‐08
	CD14	1.24	1.15	1.34	4.2E‐08
	CRP	1.24	1.15	1.34	5.9E‐08
	CLEC3B	0.82	0.76	0.88	2.9E‐07
	sICAM1	1.20	1.12	1.29	1.2E‐06
	Cystatin‐C	1.23	1.13	1.34	1.9E‐06
	d_IL6	1.92	1.45	2.55	4.5E‐06
	GRN	1.19	1.10	1.28	5.1E‐06
	REG1A	1.19	1.10	1.28	7.9E‐06
	FGF23	1.19	1.10	1.28	8.0E‐06
	AGP1	1.17	1.09	1.27	4.2E‐05
	MMP8	1.17	1.08	1.26	4.1E‐05
	IGF1	0.87	0.81	0.94	1.9E‐04
	IGFBP1	1.18	1.08	1.29	2.0E‐04
	CD5L	1.15	1.07	1.24	3.0E‐04
	Adipsin	1.16	1.07	1.26	4.1E‐04
	SAA1	1.14	1.06	1.23	4.6E‐04
	MMP9	1.15	1.06	1.25	4.7E‐04
	BCHE	0.87	0.81	0.94	5.6E‐04
	PON1	0.88	0.81	0.95	9.4E‐04
	MCP1	1.13	1.05	1.22	1.0E‐03
	MPO	1.12	1.04	1.21	1.7E‐03
	sGP130	1.13	1.05	1.21	1.7E‐03
	Ceruloplasmin	1.15	1.05	1.25	2.0E‐03
	Resistin	1.12	1.04	1.20	2.8E‐03
	A1M	1.12	1.04	1.20	3.2E‐03
	COL18A1	1.11	1.03	1.20	4.6E‐03
	IGFBP3	0.90	0.84	0.97	5.1E‐03
	CNTN1	0.90	0.83	0.97	7.9E‐03
	KLKB1	0.90	0.83	0.98	1.3E‐02
	PMP2	0.91	0.85	0.98	1.6E‐02
CVD death (N=167)	NT‐proBNP	2.43	1.99	2.97	1.6E‐18
	ADM	1.75	1.45	2.11	4.1E‐09
	GDF15	1.96	1.56	2.46	6.9E‐09
	B2M	1.72	1.42	2.09	3.0E‐08
	EFEMP1	1.60	1.33	1.94	9.5E‐07
	Cystatin‐C	1.55	1.29	1.87	4.1E‐06
	UCMGP	1.43	1.21	1.70	3.6E‐05
	Adipsin	1.45	1.21	1.73	4.7E‐05
	IGFBP2	1.42	1.19	1.70	1.2E‐04
	AGP1	1.38	1.17	1.62	1.1E‐04
	REG1A	1.37	1.16	1.61	1.6E‐04
	CRP	1.38	1.17	1.63	1.8E‐04
	PMP2	0.75	0.64	0.88	4.4E‐04
	TIMP1	1.40	1.16	1.69	5.0E‐04
	SAA1	1.33	1.13	1.56	6.5E‐04
	Ceruloplasmin	1.40	1.15	1.70	7.3E‐04
	BCHE	0.76	0.64	0.89	8.3E‐04
	NRCAM	0.76	0.65	0.89	8.1E‐04
	CD14	1.32	1.12	1.56	9.4E‐04
	d_IL6	2.46	1.44	4.19	1.0E‐03
	CDH13	0.77	0.66	0.91	2.0E‐03
	Resistin	1.28	1.09	1.49	2.2E‐03
	FGF23	1.28	1.09	1.50	2.6E‐03
	sICAM1	1.27	1.09	1.48	2.6E‐03
	LDLR	0.78	0.66	0.92	2.6E‐03
	MMP9	1.30	1.09	1.54	2.8E‐03
	COL18A1	1.27	1.08	1.49	3.2E‐03
	ADAM15	0.79	0.68	0.92	3.1E‐03
	A1M	1.26	1.08	1.47	3.5E‐03
	sGP130	1.25	1.07	1.47	5.3E‐03
	d_FLT3	0.64	0.47	0.88	5.4E‐03
	IGFBP1	1.31	1.08	1.58	5.6E‐03
	FBN	1.25	1.06	1.48	9.4E‐03
	d_CSF2RB	0.67	0.49	0.91	1.2E‐02
	GRN	1.23	1.05	1.44	1.2E‐02

Multivariable‐adjusted model, adjusted for age, sex, systolic blood pressure, hypertension treatment, diabetes mellitus, body mass index, smoking, total and HDL cholesterol, and history of atrial fibrillation. In addition, heart failure and mortality analyses were adjusted for prevalent myocardial infarction. CVD indicates cardiovascular disease; d_, dichotomous biomarker; HDL, high‐density lipoprotein; HR, hazards ratio per 1‐SD change in rank normalized data; LCL, lower 95% confidence interval; UCL, upper 95% confidence interval.

Biomarkers Associated Across Different Fatal and Nonfatal Cardiovascular Outcomes

We noted prominent overlap in biomarkers that predicted more than 1 outcome (Figure). In total, 28 biomarkers predicted both CVD death and all‐cause mortality, including GDF15, ADM, CRP, cystatin‐C, and NT‐proBNP, among others. There was overlap in prediction of both fatal and nonfatal cardiovascular outcomes, with GDF15, TIMP1, B2M, and cystatin‐C predicting higher risk of atherosclerotic CVD, HF, mortality, and CVD death. Furthermore, all 18 biomarkers that predicted incident HF also predicted all‐cause and/or CVD death.

Figure 1

Heatmap of biomarkers associated with all‐cause mortality (FDR‐q <0.05) are also associated with other concomitant fatal and non‐fatal cardiovascular outcomes, including atherosclerotic CVD, heart failure, and CVD death. ASCVD indicates atherosclerotic cardiovascular disease; CVD, cardiovascular disease; FDR, false discovery rate; HF, heart failure; HR, hazard ratio.

Multimarker Models Predict Incident Cardiovascular Events and Death

Joint associations of multiple biomarkers with incident events were assessed using step‐wise regression, after accounting for clinical covariates (Table 3). When added to a clinical model, only GDF15 was associated with incident ASCVD: (HR, 1.43 per 1‐SD increase in GDF15; 95% CI, 1.24–1.65; P=1.3×10−6). Four biomarkers jointly predicted incident HF: NT‐proBNP (HR, 1.79; 95% CI, 1.51–2.12; P=1.9×10−11), GDF15 (HR, 1.78; 95% CI, 1.44–2.19; P=7.5×10−8), CRP (HR, 1.29; 95% CI, 1.10–1.50; P=0.001), and leptin (HR, 0.73; 95% CI, 0.60–0.89; P=0.002).

Table 3

Multimarker Models Associated With Cardiovascular Outcomes

Outcome	Biomarker	HR	95% LCL	95% UCL	P Value	Bootstrap (n)a
Atherosclerotic CVD	GDF15	1.43	1.24	1.65	1.3E‐06	574
Heart failure	NT‐proBNP	1.79	1.51	2.12	1.9E‐11	1000
	GDF15	1.78	1.44	2.19	7.5E‐08	957
	CRP	1.29	1.10	1.50	1.4E‐03	462
	Leptin	0.73	0.60	0.89	1.5E‐03	577
Mortality	GDF15	1.71	1.50	1.94	3.7E‐16	1000
	CLEC3B	0.80	0.73	0.87	2.4E‐07	941
	NT‐proBNP	1.30	1.18	1.44	2.7E‐07	985
	AGP1	1.24	1.14	1.36	6.8E‐07	821
	sRAGE	0.83	0.77	0.91	3.0E‐05	884
	PMP2	0.86	0.79	0.93	1.1E‐04	517
	UCMGP	1.18	1.08	1.29	2.2E‐04	817
	KLKB1	0.84	0.77	0.92	3.0E‐04	609
	IGFBP2	1.19	1.08	1.32	4.0E‐04	715
	IGF1	0.88	0.81	0.95	9.7E‐04	556
	Leptin R	1.14	1.05	1.24	1.9E‐03	504
	Cystatin‐C	0.87	0.79	0.97	8.6E‐03	412
CVD death	NT‐proBNP	2.09	1.69	2.58	6.0E‐12	1000
	PMP2	0.68	0.57	0.82	2.4E‐05	392
	AGP1	1.44	1.20	1.71	6.6E‐05	462
	MMP9	1.41	1.16	1.72	4.9E‐04	855
	GDF15	1.45	1.12	1.87	5.0E‐03	369
	CLEC3B	0.77	0.64	0.92	5.1E‐03	366

Multivariable‐adjusted model, adjusted for age, sex, systolic blood pressure, hypertension treatment, diabetes mellitus, body mass index, smoking, total and HDL cholesterol, and history of atrial fibrillation. In addition, heart failure and mortality analyses were adjusted for prevalent myocardial infarction. CVD, cardiovascular disease; HDL, high‐density lipoprotein; HR, hazards ratio per 1‐SD change in rank normalized data; LCL, lower 95% confidence interval; UCL, upper 95% confidence interval.

Bootstrap analysis: number of times biomarker entered into the model out of 1000 samples.

Twelve biomarkers were significantly associated with all‐cause mortality in a multimarker model: GDF15, CLEC3B, NT‐proBNP, arabinogalactan protein 1 (AGP1), soluble receptor for advanced glycation end products (sRAGE), PMP2, uncarboxylated matrix Gla protein (UCMGP), KLKB1, insulin‐like growth factor binding protein 2 (IGFBP2), IGF1, leptin receptor, and cystatin‐C. Six biomarkers were associated with CVD death: NT‐proBNP, PMP2, AGP1, matrix metalloproteinase 9 (MMP9), GDF15, and CLEC3B.

We used bootstrap resampling and fitted step‐wise multimarker models in 1000 samples to assess frequencies with which biomarkers remained in the multimarker model (Table 3). GDF15 entered the model for all‐cause mortality 100% of the time and for HF 96% of the time. Similarly, NT‐proBNP entered the model for both HF and CVD death 100% of the time and for all‐cause mortality 99% of the time. Other markers were noted to be less robust, including CRP for HF, cystatin‐C for mortality, and PMP2 AGP1, GDF15, and CLEC3B for CVD death, each of which entered the specific models <50% of the time.

Performance Metrics of Multimarker Models

Model discrimination was assessed by improvement in the C‐statistic with the addition of multimarkers to a clinical model. Specifically, the C‐statistic increased from 0.753 to 0.758 for ASCVD, from 0.843 to 0.873 for HF, from 0.783 to 0.818 for all‐cause mortality, and from 0.851 to 0.880 for CVD death (P<0.0001 for all, comparing the likelihood ratio of the model with and without biomarkers; Table 4). The integrated discrimination improvement was highest for CVD death and HF (0.188 and 0.124, respectively). We examined the 2‐category net reclassification improvement, using event rates to define high‐ versus low‐risk categories. The net reclassification improvement for the addition of multiple biomarkers to the base clinical model was <1% for ASCVD, 3.5% for HF, 7.7% for all‐cause mortality, and 9.5% for CVD death.

Table 4

Performance Metrics of Multimarker Models

Performance Metric	Atherosclerotic CVD	Heart Failure	Mortality	CVD Death
C‐statistic (95% CI)
Clinical model	0.753	0.843	0.783	0.851
Clinical+multimarkera	0.758	0.873	0.818	0.880
Change in C‐statistics	0.004	0.030	0.036	0.030
P value (LR)	0.00008	<0.0001	<0.0001	<0.0001
IDI (95% CI)	0.012 (0.004, 0.023)	0.124 (0.084, 0.173)	0.083 (0.068, 0.11)	0.188 (0.126, 0.259)
2‐category NRI (95% CI)	0.007 (−0.018, 0.065)	0.035 (−0.015, 0.109)	0.077 (0.045, 0.122)	0.094 (0.021, 0.189)

2‐category NRI indicates clinical vs clinical+multimarker models, using outcome‐specific event rate to define high‐ vs low‐risk categories; CI, confidence interval; CVD, cardiovascular disease; IDI, integrated discrimination improvement; LR, likelihood ratio comparing clinical vs clinical+multimarker models.

Multimarker models are outcome specific and include markers included in step‐wise models described in Table 3.

Discussion

Using a proteomic platform that targeted 85 high‐value proteins for CVD, we identified 8 individual biomarkers associated with incident ASCVD, 18 with incident HF, 38 with all‐cause mortality, and 35 with CVD death after accounting for potential clinical confounders. In a multimarker approach that accounted for clinical factors, GDF15 was associated with all outcomes. In addition, NT‐proBNP, CRP, and leptin were associated with incident HF, and CLEC3B (tetranectin), NT‐proBNP, AGP1, sRAGE, PMP2, UCMGP, KLKB1, IGFBP2, IGF1, leptin receptor, and cystatin‐C were associated with all‐cause mortality in a multimarker model. We demonstrate multiple new associations of protein biomarkers that regulate metabolic and inflammatory pathways with clinical outcomes using a multiplexed high‐throughput platform. Previous proteomics approaches among large cohorts have been predominantly focused on CVD case‐control designs.3, 12 We now extend previous promising results and test the association of targeted proteomic biomarkers with incident CVD events and death in the community. These findings may have future potential implications with respect to clinical risk prediction.

Whereas GDF15, NT‐proBNP, ADM, cystatin‐C, fibroblast growth factor 23 (FGF23), and CRP are known to predict incident CVD and other outcomes,13, 14, 15, 16 there are a number of novel associations in our study to highlight. Specifically, a number of regulators of metabolic and adipocyte homeostasis appeared to predict adverse events in our study: IGF1 is involved in adipocyte proliferation and differentiation, and whereas some studies support a protective role with respect to atherosclerosis, others have not shown any association with coronary artery disease.17, 18 In the FHS, lower IGF1 levels were previously reported to be associated with adverse metabolic risk,19 and we now show an association with mortality in both single‐ and multimarker models. IGF1, in turn, is regulated by insulin‐like growth factor binding protein 1 (IGFBP1), which plays a role in glucose counter‐regulation. Higher IGFBP1 levels were previously found to be associated with lower metabolic risk, but greater mortality,20 and IGFBP1 was reported to predict incident HF among older adults.3, 21 We now extend these observations and show that IGFBP1 predicts incident HF, all‐cause mortality, and CVD mortality in our cohort, along with 3 other regulators of metabolic disease and adipocyte function: IGFBP2, adipsin, and leptin. IGFBP2 is the predominant insulin‐like growth factor binding protein (IGFBP) produced by white adipose tissue and is thought to protect against obesity and insulin resistance in mice.8, 22 Previous studies, however, did not demonstrate any relation to CVD outcomes.10, 21 Adipsin has been shown to regulate inflammation in adipose tissue and may improve beta‐cell function in diabetes mellitus.11, 23, 24 Among women, adipsin is correlated with metabolic parameters, including body mass index, insulin resistance, and carotid atherosclerosis,3, 12, 25 but it has not been linked to adverse CVD outcomes previously. Last, resistin is an adipokine that impairs insulin action and is thought to represent a mechanistic link of obesity with diabetes mellitus.17, 18, 26 In previous FHS analyses, resistin was reported to be associated with incident HF, and it also predicted adverse outcomes among patients with stable coronary disease in the Heart and Soul study.19, 27, 28

Our results for multiple inflammatory cytokines and regulators are also notable. These include B2M, cluster of differentiation 14 (CD14), soluble intracellular adhesion molecule 1 (sICAM1), granulin (GRN), AGP1, CD5 molecule‐like (CD5L), serum amyloid A1 (SAA1), monocyte chemoattractant protein 1 (MCP1), soluble glycoprotein 130 (sGP130), interleukin 6 (IL6), and myeloperoxidase (MPO), which predicted all‐cause mortality in our study. B2M is a glomerular filtration marker and acute‐phase reactant that has been associated with CVD events and mortality among patients with chronic kidney disease,20, 29 and was also associated with all‐cause mortality in the National Health and Nutrition Examination Survey.30 In addition to predicting all‐cause mortality, B2M is also a predictor of incident ASCVD, HF, and CVD mortality in our study. Soluble CD14 is also an acute‐phase protein and serves as the receptor for lipopolysaccharide.31 In clinical studies, soluble CD14 predicted all‐cause mortality among older adults.32 We now extend these findings to a wider age range, and show that soluble CD14 predicts incident HF and all‐cause and CVD mortality. sICAM1 is a cell adhesion molecule that previously has been associated with CVD, including myocardial infarction, in case‐control studies.33, 34 We now show that sICAM1 predicts not only ASCVD, but also all‐cause mortality and CVD mortality. GRN has previously been linked to atherosclerosis through inflammatory activation in experimental studies35 and predicts incident HF, all‐cause, and CVD mortality in our study. AGP1 is another acute‐phase reactant, and its protein, orosomucoid, predicted mortality among elderly individuals.36 In our analyses, AGP1 predicted ASCVD, as well as all‐cause and CVD mortality in single‐ and multimarker models. CD5L inhibits macrophage apoptosis and is thought to regulate infarct size in animal models.37 SAA1 is an acute‐phase reactant that previously has been associated with future coronary disease in various populations.38 We report an association of SAA1 with all‐cause and CVD mortality. MCP1 is a chemokine that regulates monocyte trafficking and previously has been associated with myocardial infarction in FHS and adverse outcomes in patients with acute coronary syndrome.39, 40 sGP130 is a natural inhibitor of IL6 responses and has been associated with adverse outcomes among ischemic HF patients.41 We now show an association of GP130 with all‐cause and CVD mortality. Last, MPO is secreted by activated leukocytes and is thought to mediate oxidative stress and atherogenesis.42 It has been associated with future coronary disease and HF in case‐control studies.43, 44 In our study, MPO predicted incident HF, in addition to all‐cause mortality.

Last, our study sought to validate previous genetic associations, with selection of a number of targeted biomarkers from genome‐wide association studies. For example, the precursor to GRN is progranulin, and serves as the ligand for sortilin 1 (SORT1),45 which, in turn, is thought to regulate lipoprotein metabolism and has been linked to risk of myocardial infarction in previous genome‐wide association studies studies.46 We now demonstrate an association of GRN with both CVD and all‐cause mortality, supporting the association of the GRN/SORT1 pathway with CVD events.

Numerous studies have examined the utility of multiple biomarkers in CVD risk prediction, with the ascertainment of markers representing different pathophysiological pathways that might add complementary information with respect to risk prediction. The incremental information provided by multiple protein biomarkers to existing clinical risk factors, however, has remained modest.14, 47, 48 When examined in a multimarker approach, we similarly found modest improvements in model discrimination metrics. Interestingly, 12 biomarkers independently predicted all‐cause mortality on top of clinical variables. In contrast, in attempting to create a protein multimarker for ASCVD, only GDF15 entered the model. Although biomarkers were targeted on the basis of existing evidence for associations with atherosclerosis, our findings need further validation in external cohorts. However, 45 of 85 biomarkers were selected on the basis of known associations with ASCVD from the literature and 41 on the basis of previous discovery proteomics evidence, highlighting the fact that the majority of biomarkers selected were motivated by existing clinical, genetic, gene expression profiling, or discovery proteomics data. There are other limitations that deserve mention. Ours are observational data, and causal or mechanistic inference cannot be drawn. Furthermore, limitations of the multiple reaction monitoring mass spectrometry platform include bias toward detection of more‐abundant proteins.

In conclusion, we identified 8 individual biomarkers that predicted incident ASCVD, 18 that predicted incident HF, and 38 that predicted all‐cause mortality. Many biomarkers represent regulators of metabolic and adipocyte homeostasis; others are involved in inflammatory pathways. Using a targeted discovery proteomic platform, we were able to extend previous findings, demonstrate novel biomarker associations with incident CVD events, and validate previous genetic associations. Large‐scale mass spectrometry or newer aptamer‐based technology49, 50 may allow for profiling of many more proteins in large study samples that may refine current multimarker models. Further studies are needed to validate our findings and define clinical utility, with the ultimate goal of improving targeted strategies for CVD prevention.

Sources of Funding

This work was partially supported by the National Heart, Lung, and Blood Institute's Framingham Heart Study (Contracts N01‐HC‐25195 and HHSN268201500001I) and by the Division of Intramural Research (Courchesne, Chen, Liu, Hwang, and Levy) of the National Heart, Lung and Blood Institute. Dr Ho is supported by NIH grants K23‐HL116780 and R01 HL140224 and a Massachusetts General Hospital Hassenfeld Scholar Award. The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the US Department of Health and Human Services.

Disclosures

None.

Table S1. Biomarker Assay Performance Characteristics

Table S2. Correlation Between Biomarkers

Table S3. Multivariable‐Adjusted Associations of Single Biomarkers and Cardiovascular Outcomes

Table S4. Age‐ and Sex‐Adjusted Associations of Biomarkers and Cardiovascular Outcomes

Click here for additional data file.

Data S1.

Click here for additional data file.