Proteomic cardiovascular risk assessment in chronic kidney disease

Abstract

Aims

Chronic kidney disease (CKD) is widely prevalent and independently increases cardiovascular risk. Cardiovascular risk prediction tools derived in the general population perform poorly in CKD. Through large-scale proteomics discovery, this study aimed to create more accurate cardiovascular risk models.

Methods and results

Elastic net regression was used to derive a proteomic risk model for incident cardiovascular risk in 2182 participants from the Chronic Renal Insufficiency Cohort. The model was then validated in 485 participants from the Atherosclerosis Risk in Communities cohort. All participants had CKD and no history of cardiovascular disease at study baseline when ∼5000 proteins were measured. The proteomic risk model, which consisted of 32 proteins, was superior to both the 2013 ACC/AHA Pooled Cohort Equation and a modified Pooled Cohort Equation that included estimated glomerular filtrate rate. The Chronic Renal Insufficiency Cohort internal validation set demonstrated annualized receiver operating characteristic area under the curve values from 1 to 10 years ranging between 0.84 and 0.89 for the protein and 0.70 and 0.73 for the clinical models. Similar findings were observed in the Atherosclerosis Risk in Communities validation cohort. For nearly half of the individual proteins independently associated with cardiovascular risk, Mendelian randomization suggested a causal link to cardiovascular events or risk factors. Pathway analyses revealed enrichment of proteins involved in immunologic function, vascular and neuronal development, and hepatic fibrosis.

Conclusion

In two sizeable populations with CKD, a proteomic risk model for incident cardiovascular disease surpassed clinical risk models recommended in clinical practice, even after including estimated glomerular filtration rate. New biological insights may prioritize the development of therapeutic strategies for cardiovascular risk reduction in the CKD population.

Structured Graphical Abstract

Structured Graphical Abstract.

ARIC, Atherosclerosis Risk in Communities; CRIC, Chronic Renal Insufficiency Cohort; CKD, chronic kidney disease; CV, cardiovascular; CVE, cardiovascular event; ROC AUC, receiver operating characteristic area under the curve.

See the editorial comment for this article ‘Revolutionizing cardiovascular risk prediction in patients with chronic kidney disease: machine learning and large-scale proteomic risk prediction model lead the way’, by R. Avram, https://doi.org/10.1093/eurheartj/ehad127.

Introduction

Chronic kidney disease (CKD) affects more than 10% of the population worldwide and increases the risk for developing cardiovascular disease.^1–3 Despite widespread recognition of CKD as a strong cardiovascular risk factor, risk stratification tools that can accurately inform cardiovascular prognosis in this population are lacking. The 2013 American College of Cardiology/American Heart Association (ACC/AHA) Pooled Cohort Equations (PCE) for cardiovascular risk assessment, which are relied upon by the cholesterol and blood pressure management guidelines,^4,5 were developed from traditional risk factors to identify the 10-year risk of incident atherosclerotic cardiovascular disease in the general population.⁶ Yet, clinical models that utilize such traditional risk factors have performed poorly in individuals with CKD even after accounting for CKD by adding an estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m² to the risk equations.^7,8 Perhaps as a result, the European Society of Cardiology has broadly categorized individuals with CKD as either high or very high risk for developing a cardiovascular event;⁹ however, prognostic equations that are capable of providing personalized, accurate, and modifiable estimates of cardiovascular risk across a broad dynamic range to be clinically useful have remained an area of great need in the CKD population.

We surveyed the plasma proteome using an aptamer-based proteomic platform in the years preceding cardiovascular disease in individuals with CKD. As such, the current study takes a relatively unbiased approach for discovery of circulating protein biomarkers of incident cardiovascular risk in CKD from which risk equations could be constructed and biological mechanisms responsible for increased cardiovascular risk inferred.^10,11 In the present study conducted in the Chronic Renal Insufficiency Cohort (CRIC) with validation of key findings in the Atherosclerosis Risk in Communities (ARIC) cohort, we aimed (1) to identify numerous new protein biomarkers of incident cardiovascular risk in individuals with CKD; (2) to derive and validate a multi-protein cardiovascular risk score in CKD and compare its predictive performance to risk scores composed of traditional risk factors; (3) to evaluate potential causal relationships between protein biomarkers identified in this study and their cardiovascular disease phenotypes through Mendelian randomization analysis; and (4) to infer biological mechanisms associated with the risk of incident cardiovascular disease in CKD through the tools of pathway analysis.

Methods

Study design

We developed and validated proteomic risk models in two independent cohorts. The derivation cohort included participants from CRIC, a prospective, observational study that enrolled individuals aged 21 to 74 years with CKD, defined as an eGFR < 60 mL/min/1.73 m² .¹² Those with end-stage renal disease (ESRD) and on dialysis were excluded. The current analysis consists of individuals who completed the Year 1 visit in CRIC, considered our study’s ‘baseline’, and had cryopreserved plasma samples available for proteomic analysis (n = 3261). We excluded 1067 individuals, who at study baseline had self-reported or documented history of coronary heart disease, myocardial infarction (MI), stroke, or heart failure. Thus, the final analytical cohort consisted of 2182 participants free of cardiovascular disease.

In our primary analyses, we identified associations of individual proteins with a composite cardiovascular outcome (described below) and constructed a multi-protein prognostic model. We validated the multi-protein model in 485 participants from the ARIC cohort,¹³ all of whom had CKD (eGFR < 60 mL/min/1.73 m²) and were free of cardiovascular disease at the time their plasma was sampled (ARIC Visit 3) for proteins (Figure 1). In addition, we conducted a series of analyses including associations between individual proteins and cardiovascular disease events, Mendelian randomization to investigate causality of the protein-outcome associations, and pathway analyses aimed at biologic discovery. All analyses were approved by the appropriate institutional review boards, and all participants provided written informed consent.

Figure 1.

Methods overview.

Proteomic profiling by modified aptamers

Stored samples from CRIC and ARIC were thawed, aliquoted, and shipped as frozen specimens on dry ice to SomaLogic. Proteomic analysis for both cohorts was performed with the same version 4.0 of SomaScan. The method of quantification of proteins, including normalization steps, has been described previously^10,11,14 and is detailed further in the Supplementary data online, Methods. Of the 5284 aptamers available on the SomaScan v.4.0 proteomic platform, we removed 305 for non-human targets and an additional 130 incompletely characterized investigational aptamers (see Supplementary data online, Figure S1). We also removed 19 aptamers with a coefficient of variation >50%, based on our quality control analysis of 129 split duplicate plasma samples. The 4830 aptamers, which passed quality control, had median intra-assay and inter-assay coefficients of variation <5% and represented 4638 unique proteins.

Primary outcome

The primary study outcome in both the CRIC and ARIC cohorts was defined as the first event among MI, stroke, heart failure hospitalization, or cardiovascular death within a 10-year time horizon from blood sample collection for proteomic analyses. The 10-year time horizon was selected to harmonize with the time horizon for PCE.⁶ The outcome definitions and corresponding adjudication protocols in each cohort are specified in the Supplementary data online, Methods.

Clinical variables

As part of clinical cardiovascular risk modelling, we evaluated the prediction by the 2013 Pooled Cohort Risk Equations,⁶ which includes age, sex, race/ethnicity, systolic blood pressure, use of anti-hypertensive therapies, total cholesterol, high-density lipoprotein (HDL) cholesterol, history of diabetes, and current smoking in both CRIC and ARIC using standard protocols.^15–17 Additionally, we evaluated a history of hypertension, diastolic blood pressure, eGFR, and proteinuria. In ARIC, we relied on the Chronic Kidney Disease Epidemiology Collaboration equation using creatinine and cystatin C to calculate eGFR.¹⁸ We assessed all variables including log-transformed values and interaction terms when relevant.

Statistical analysis

For proteomic risk modelling, our objective was to determine the predictive performance of models that consisted of proteins only or hybrid models comprised of both proteins and clinical risk factors. We randomly split the CRIC data into a training (80%) and a testing (20%) set. We used the training set (n = 1746 participants) to build prediction models and select the optimal parameters. Of note, four participants in the training set were excluded due to limited exposure time leaving 1742 individuals for elastic net modelling of proteins. The CRIC testing set (n = 436 participants) was withheld from the model development and was used solely to evaluate the models’ performance. To ensure that the testing set included an identical set of participants when evaluating the proteomics and clinical models, we removed 46 participants who were missing at least one clinical variable. As such, the final testing set included 390 participants.

Our frontline technique for developing risk prediction models was elastic net Cox regression which combines ridge (L2) and lasso (L1) penalties. Model fitting was conducted using the R package glmnet.^19,20 The relative contribution of the two penalties is controlled by a mixing parameter α which we set to 0.5. The shrinkage (regularization) parameter λ, which controls model complexity (number of included proteins), was determined by 10-fold cross validation of the partial likelihood deviance and the attendant ‘1 standard error rule’. After arriving at the final selection of proteins, to reduce bias in estimated regression coefficients, we refit the selected features for the elastic net model in a Cox regression model. For the preliminary multi-variable models, stability assessment was based on five repeats of the training–testing set partitioning. Alternate methods that can accommodate time-to-event endpoints such as random forests and lasso modelling were also evaluated; however, no predictive advantages were detected.

We developed two sets of hybrid models in the CRIC training set using elastic net penalized Cox regression with 15 clinical terms from the PCE and all 4638 proteins. In the first hybrid model termed ‘Clinical and Protein Hybrid Model’, the 15 standard clinical variables were not subject to selection and were forced into the model that selected proteins by elastic net. As a result, this approach evaluated whether proteins could enhance cardiovascular risk stratification beyond a clinical model. The coefficients for the mandated clinical variables, along with any selected protein variables, were estimated by Cox regression. In the second hybrid model termed ‘Competing Clinical and Protein Hybrid Model’, we allowed all proteins and all clinical terms to compete for inclusion in an elastic net model. We used the cox.zph function from the R survival package to check for proportionality using Schoenfeld residuals and attendant formal testing.²¹ The proportional hazards assumption held for all but one of the 32 proteins in the selected model. In addition, we assessed covariate functional form using plots of smoothed Martingale residuals and attendant confidence envelopes. No departures from linearity, as evidenced by excursions outside these envelopes, were detected.

Model performance was assessed by discrimination and calibration using both the CRIC testing set and the ARIC external validation cohort. In particular, we calculated the prediction models’ receiver operating characteristic (ROC) area under the curve (AUC) annually between 1 and 10 years using time-dependent AUC methodology.²² We calculated standard errors for time-dependent AUCs via bootstrapping.²³ Specifically, we performed 1000 bootstrap samples of the testing data set and calculated the time-dependent AUC for each model at a given follow-up year. To construct the bootstrap null distribution, we subtracted the original AUCs from these 1000 AUCs. At each follow-up year, the one-sided P-value was obtained by comparing the observed difference with its null distribution. The overall performance of the proteomics models was compared to the standard 2013 PCE⁶ and the modified PCE using the harmonic mean P-value approach²⁴ to generate a combined P-value that accommodates dependent tests. The AUC values for the PCE model were the weighted sum of four different equations representing Black women, White women, Black men, and White men.

Because the standard PCE did not include eGFR, was not developed for a CKD population, and did not include heart failure in the composite endpoint, we trained and validated a modified PCE model. In addition to including eGFR, we refit the same PCE cardiovascular risk factors (age, sex, race/ethnicity, systolic blood pressure, use of anti-hypertensive therapies, total cholesterol, HDL cholesterol, history of diabetes, and current smoking) in the same study populations and for the same time horizon as the protein model. Our inclusion of heart failure alongside atherosclerotic events in the modified PCE model resembles the ‘global’²⁵ or ‘general’²⁶ cardiovascular outcomes used in other studies.²⁷ Combining atherosclerotic and heart failure events in a model is facilitated by their clinical predictors being virtually identical.^27,28

Model calibration utilized all the data with bar plots indicating agreement between average predicted risk and the observed risk in each of the quintiles defined by the predicted risk score. The calibration of the CRIC-derived multi-protein model was evaluated in ARIC after adjusting for differences in baseline hazard while retaining the original coefficients for each predictor variable. Calibration performance was assessed using the Greenwood-Nam-D’Agostino test statistic.²⁹

For discovery of cardiovascular risk biomarkers, Cox proportional hazards models were used to assess the association between plasma levels of individual proteins and risk of the primary outcome in all 2182 participants without a history of cardiovascular disease from the CRIC study. The hazard ratios (HRs) were standardized to median absolute deviation (MAD), which is better suited for skewed data than standard deviation. The Benjamini-Hochberg (BH) method was used to control the false discovery rate (FDR) at 5%. The estimated HRs and 95% confidence intervals were reported for the significant proteins. Visualization of findings made recourse to volcano plots that represented (a) unadjusted associations, and multi-variable models that sequentially adjusted for (b) eGFR and (c) age, sex, race, body mass index, systolic blood pressure, diastolic blood pressure, hypertension, diabetes, smoking, total cholesterol, HDL cholesterol, eGFR, and proteinuria.

Biologic mechanisms and Mendelian randomization analyses

To evaluate potential causal relationships between individual protein biomarkers and the cardiovascular outcome of the study, we searched a proteome–phenome (PheWAS) database of Mendelian randomizations (https://www.epigraphdb.org/pqtl).³⁰ Prior work by Zheng and colleagues reported on 989 proteins and 225 phenotypic traits including cardiovascular diseases and risk factors in cis-genetic variant analyses.³⁰ Thus, for all protein biomarkers that remained significantly associated with the incident cardiovascular disease outcome after full multi-variable adjustment, we searched the PheWAS database and assessed phenotypes representing either cardiovascular disease endpoints or risk factors. We report the proteins for which these cardiovascular phenotypes were significantly associated with plasma protein levels (P < 0.05) in Mendelian randomization analyses.

Pathway analysis

Based on the analyses in CRIC, individual proteins, risk estimates for associations of individual proteins with the primary study outcome, and corresponding q-values after adjustment for eGFR were used in the pathway analyses by Ingenuity Pathway Analysis (IPA) software (Ingenuity Systems Inc., Redwood City, CA, USA; www.ingenuity.com).^31,32 Proteins were identified according to their UniProt identification annotation. The totality of proteins included in the SomaScan assay and recognized by IPA (n = 4371) served as background proteins. Fisher’s right tailed exact test was used to calculate a P-value to determine the probability that the association of the differentially expressed proteins in the measured data set and known canonical pathways is explained by chance alone.

Statistical analyses were performed using R, version 4.0.3 (RStudio, Inc., Boston, MA. URL http://www.rstudio.com/), with the packages of glmnet (version 4.0-2), survival (version 3.2-7 pec (version 2019.11), compareC (version 1.3.1), forestplot (version 1.10), and ime4 (version 1.1-26). Finally, with the exception of the P-values created for the calibration plot and for comparing the time-dependent AUCs using bootstrap, all other P and q-values are two-sided.

Results

At our study’s baseline, the CRIC (n = 2182) and ARIC (n = 485) participants were (by design) free of known coronary heart disease, heart failure, or stroke and had a broad spectrum of CKD severity (mean ± SD eGFR 45 ± 17 mL/min/1.73 m² in CRIC and 50 ± 11 mL/min/1.73 m² in ARIC). Compared to those in ARIC, CRIC participants were somewhat younger, more likely to be men, and more likely to be Black (Table 1). In addition, CRIC participants were more likely to have a history of hypertension, diabetes, and less likely to be active smokers. Physiologic and laboratory measures such as systolic blood pressure, diastolic blood pressure, and body mass index were similar among the two cohorts, while total cholesterol levels were higher and HDL levels were lower in ARIC compared to CRIC participants.

Table 1.

Baseline characteristics across derivation and validation cohorts

	Derivation cohort (CRIC)	Validation cohort (ARIC)
	n = 2182	n = 485
Demographics
Age, years	58 (11)	64 (5)
Female	1023 (47%)	283 (58%)
White	1116 (51%)	373 (77%)
Black	836 (38%)	112 (23%)
CVD risk factors
Hypertension	1860 (85%)	290 (60%)
Diabetes	926 (42%)	114 (24%)
Current tobacco use	248 (11%)	189 (39%)
Measures
Systolic blood pressure, mmHg	125 (21)	133 (24)
Diastolic blood pressure, mmHg	71 (12)	72 (12)
Body mass index, kg/m2	31.7 (7.7)	30.1 (5.9)
Total cholesterol, mg/dL	187 (43)	211 (42)
HDL cholesterol (mg/dL)	50 (16)	46 (16)
eGFR, mL/min/1.73 m2	45 (17)	50 (11)
Urinary protein to creatinine ratio (g/g), median [IQR]	0.11 [0.05, 0.57]
Medications
Taking anti-hypertensive medication	1147 (53%)	273 (57%)
Incident cardiovascular events
Annualized rate	2.1%	4.6%

Continuous variables are reported as mean (standard deviation) for continuous variables or N (%) for categorical variables. CRIC indicates Chronic Renal Insufficiency Cohort; ARIC indicates Atherosclerosis Risk in Communities study; HDL, high-density lipoprotein; and IQR, interquartile range.

Development and validation of a protein model and its comparison to clinical models

Over the 10-year follow-up period, there were 459 incident cardiovascular events in CRIC (373 events in the training set and 86 events in the testing set) that occurred after a mean 7.8 ± 3.1 years (annualized event rate of 2.1%); similarly, there were 173 incident cardiovascular events in ARIC that occurred after a mean 7.7 ± 3.2 years (annualized event rate of 4.6%). From the CRIC training set, elastic net regression selected 32 proteins (out of a total 4638 measured proteins) that comprise the primary proteomics risk model. The risk prediction equation, proportionate contribution of each protein, and biological features of each of the 32 proteins in the model are presented in Table 2. Only half of the proteins selected for the model have been previously implicated in atherogenesis, cardiac hypertrophy or dilation, conduction disease, or vascular disease, whereas the remaining 16 proteins have previously not been linked to cardiovascular disease. The discrimination of the 32-protein model, depicted in Figure 2 as a time-dependent AUC over the entire 10-year time horizon, was superior to both the original PCE (P = 0.0099) and the modified PCE (P = 0.0049), which includes eGFR and whose coefficients for all the variables were refit to the CRIC training set. Specifically, internal validation among the CRIC testing set demonstrated annualized AUC values ranging between 0.84 to 0.89 for the protein model and 0.70 to 0.73 for the two clinical equations. Similar findings and gains in AUC were observed in the external ARIC validation set. The AUC values for the 32-protein model ranged between 0.73 and 0.78 and were higher than those for the two clinical models (protein model vs. PCE, P = 0.0026; protein model vs. modified PCE, P = 0.0023). In both validation sets, the protein model demonstrated excellent short-term risk prediction with its highest discrimination over the initial 3-year period. Further, both clinical models (without and with eGFR) had similar prediction indicating that the inclusion of eGFR in PCE did not appreciably enhance its prediction of incident cardiovascular events. Modelling performance remained stable with repeated training–testing set partitioning (data not shown).

Table 2.

Proteins and corresponding coefficients in risk prediction model^a

UniProt	Protein name	Coefficient in prediction modela	Features of protein	Association with cardiovascular disease
Q14508	WAP four-disulfide core domain protein 2 (human epididymis protein – 4)	0.019	• Plays a major role in sperm maturation • Protease inhibitor that likely confers protection against microbial virulence factors	• No clear evidence reported for the association of this protein and cardiovascular disease.
O43251	RNA binding protein fox-1 homolog 2	0.011	• Key regulator of alternative exon splicing, especially in the nervous system • Interacts with the oestrogen receptor 1 (ER1) transcription factor and regulates ER1 activation	• Impaired function associated with cardiac conduction defects in myotonic dystrophy.59
Q9UBX5	Fibulin-5	0.11	• Essential for elastic fibre formation • Stabilizes and organizes elastic fibres in skin, lung, and vasculature • Regulator of alternative exon splicing, especially in the nervous system	• Diminution of elastic fibres has been associated with vascular anomalies.60
P16860	N-terminal pro-BNP	0.074	• BNP is a natriuretic peptide hormone secreted from cardiac muscle cells. • BNP functions as a paracrine anti-fibrotic factor in the heart. • BNP plays a critical role in cardiovascular homeostasis and inhibits renin-aldosterone secretion.	• Plasma natriuretic peptide levels may predict the risk of CV events and death.61 • Circulating BNP or NT-proBNP aids in heart failure management. • BNP is a prognostic marker for patients with acute myocardial infarction.62
Q9H5V8	CUB domain–containing protein 1	0.073	• Involved in cell adhesion and cell matrix association63 • May be a novel marker for immature hematopoietic stem cell subsets64	• No clear evidence reported for the association of this protein and cardiovascular disease.
P35754	Glutaredoxin-1	0.052	• Cytoplasmic enzyme that catalyzes the reversible reduction of glutathione-protein mixed disulfides • Enzyme contributes to the antioxidant defense system.65	• Important role in maintaining cardiovascular homeostasis and preventing ischaemia, peripheral arterial disease, and cardiac hypertrophy66
Q8IZP7	Heparan-sulfate 6-O-sulfotransferase 3	−0.072	• Involved in heparin metabolism • Associated with diabetic retinopathy in diabetes • Polymorphisms in this gene associated with obesity and elevated triglyceride levels.67	• No clear evidence reported for the association of this protein and cardiovascular disease.
P00533	Epidermal growth factor receptor (EGFR)	0.024	• Tyrosine kinase receptor that binds ligands of the EGF family and activates several signalling cascades to convert extracellular cues into appropriate cellular responses including NF-KB. • Involved in the pathogenesis of non-small cell lung cancer68	• Activation of EGFR has been implicated in endothelial dysfunction, neointimal hyperplasia, atherogenesis, and cardiac remodelling.69 • Increased circulating EGF-like ligands mediate accelerated vascular disease.69
P41271	Neuroblastoma suppressor of tumorigenicity 1	0.046	• Considered a tumour suppressor gene of neuroblastoma and plays an important role in preventing cells from entering the final stage (G1/S) of the transformation process. • Inhibits bone morphogenic proteins. Dysregulation can also lead to kidney disease and pulmonary arterial hypertension.70	• Elevated levels are associated with the prevalence of CAD in patients with obstructive sleep apnea.71
Q03167	Transforming growth factor beta (TGFß) receptor type 3	−0.17	• Binds to TGFß • Involved in regulating cellular homeostasis, embryonic development, immune surveillance, angiogenesis, and apoptosis • Leads to either suppression or promotion of various cancer types	• Dysregulated in atherosclerotic vascular diseases72
P23280	Carbonic anhydrase 6	−0.11	• Localized in the salivary glands and secreted into the saliva73 • May precipitate regulation of salivary pH	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q99988	Growth differentiation factor 15 (GDF-15)	−0.037	• Regulates food intake, energy expenditure, and body weight in response to metabolic stresses74 • Triggers the activation of the central amygdala, which is part of the emergency circuit that shapes feeding responses to stress74 • Higher levels associated with CKD and cancer	• Cardiovascular disease is a major driver of GDF-15 production.75 • In population-based cohorts, GDF-15 levels are associated with an increased risk of developing cardiovascular disease. • In patients with CKD, higher levels are associated with death and heart failure.
P39900	Macrophage metalloelastase	0.030	• Part of the matrix metalloproteinase (MMP) family • This protein is involved in macrophage migration, tissue injury and remodelling.	• Elevated plasma levels are associated with cardiovascular disease.76 • MMP-12 may be prognostic for cardiovascular events in patients with stable coronary heart disease.10 • MMP-12 inhibition may worsen post-MI cardiac dysfunction.
P45379	Troponin T, cardiac muscle	0.020	• Tropomyosin-binding subunit of troponin, the thin filament regulatory complex which confers calcium-sensitivity to striated muscle actomyosin ATPase activity	• Mutations can lead to hypertrophic,77 restrictive, or dilated cardiomyopathy.78 • Elevated plasma levels are associated with cardiovascular disease.
P05452	Tetranectin	−0.079	• Binds to kringle 4 of circulating plasminogen, resulting in enhanced activation of plasminogen into plasmin	• Increased levels are associated with the presence and severity of coronary artery disease.79
Q96EU7	C1GALT1-specific chaperone 1	−0.058	• X-borne gene encoding cosmc, chaperone required for functioning of T-synthase80 • T-synthase is essential for biosynthesis of O-glycans, which is important for cellular metabolism and interactions including leukocyte trafficking. • Dysregulation implicated in breast and colon cancer and autoimmune disorders including Tn polyagglutination syndrome.81	• No clear evidence reported for the association of this protein and cardiovascular disease.
P01270	Parathyroid hormone (PTH)	0.044	• PTH elevates calcium levels by dissolving the salts in bone and preventing renal excretion. • It also stimulates glycogen synthesis in osteoblastic cells.	• Excess PTH is associated with fatal and non-fatal cardiovascular events.82 • Higher concentrations are associated with CHD, increased left ventricular mass, and heart failure.83
Q96EE4	Coiled-coil domain–containing protein 126	−0.043	• Part of domain containing proteins • Functions in various physiologic processes including regulation of signal transduction, gene expression, cell division, and motility • Diverse roles in tumour progression	• No clear evidence reported for the association of this protein and cardiovascular disease.
O75078	Disintegrin and metalloproteinase domain–containing protein 11	0.11	• Protein implicated in cell-cell or cell-extracellular matrix interactions • Coding region is known to be disrupted by tumour suppressor genes implicated in breast and ovarian cancers.84	• No clear evidence reported for the association of this protein and cardiovascular disease.
P07998	Ribonuclease pancreatic	0.014	• Endonuclease that catalyzes the cleavage of RNA on the 3’ side of pyrimidine nucleotides • In addition to the pancreas, it is also released by endothelial cells suggesting a role in vascular homeostasis.85	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q9H665	Insulin growth factor-like family receptor 1 (IGFLR1)	−0.032	• IGFLR1 is widely expressed in lymph nodes, spleen, and kidney. • Increased expression has been associated with poor survival in renal cell cancer.86 • Similar structure to tumour necrosis factor receptor family, which are widely expressed in the immune system and have a regulatory effect on both congenital and adaptive immunity87	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q9NY97	N-acetyllactosaminide beta-1,3-N-acetylglucosaminyltransferase 2	−0.057	• These golgi-resident glycosyltransferases are involved in the biosynthesis of poly-N-acetyl-lactosamine chains, which are known to suppress excessive immune responses.88	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q14956	Transmembrane glycoprotein NMB	0.038	• Augments bone mineral deposition by stimulating osteoblast differentiation • Also has anti-inflammatory and reparative functions • Demonstrated to be neuroprotective in animal models of neurologic disease • It is also relevant to bone marrow function including cell differentiation.89	• No clear evidence reported for the association of this protein and cardiovascular disease.
O00300	Tumour necrosis factor receptor superfamily member 11B	0.024	• Involved in maintaining bone homeostasis by regulating osteoclast activity90 • Mutations can result in Paget’s disease of bone, characterized by osteoclastic overactivity.	• Controversy exists regarding the role of osteoprotegerin in cardiac disease. Serum levels could either be increased due to vascular insults or could be atherogenic91
Q2I0M5	R-spondin-4	0.059	• Activator of canonical Wnt signalling pathway • R-spondin family of proteins include two furin-type cysteine-rich domains. These regions are required for activation and stabilization of ß-catenin.92	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q12805	EGF-containing fibulin-like extracellular matrix protein 1	0.041	• Binds to the epidermal growth factor (EGF) receptor and promotes cell adhesion and migration in human adenocarcinoma93 • Also up-regulated in gliomas and undetected in normal brain or cultured astrocytes	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q8IUL8	Cartilage intermediate layer protein 2 (CILP2)	−0.11	• CILP2 is localized to the intermediate to deep zone of articular cartilage. • CILP2 participates in cartilage scaffolding.	• CILP1 is a novel marker for cardiac fibrosis.94
P51858	Hepatoma-derived growth factor	0.073	• Family of proteins that have growth stimulating, angiogenesis-inducing, and anti-apoptotic roles95 • Ubiquitously expressed in non-cancerous tissue • Contributes to the development of digestive malignancies and promotes proliferation of hepatocellular carcinoma cells	• No clear evidence reported for the association of this protein and cardiovascular disease.
Q96D42	Hepatitis A virus cellular receptor 1	0.037	• Member of the T cell transmembrane, immunoglobulin and mucin family • Regulates immune cell activity especially regarding the host response to viral infection	• Variants in this gene have been associated with coronary heart disease and ischaemic stroke.96
Q6ZUB0	Putative spermatogenesis-associated protein 31D4	−0.076	• Plays a role in cell differentiation and spermatogenesis	• No clear evidence reported for the association of this protein and cardiovascular disease.
P15559	NAD(P)H dehydrogenase [quinone] 1	0.091	• Cytoplasmic 2-electron reductase and reduces quinones to hydroquinones (encoded by NQO1) • The ubiquitin-independent p53 degradation pathway is regulated by NQO1. • Individuals with decreased NQO1 expression have reduced p53 stability, which may lead to resistance to chemotherapeutics. • Enzyme also involved in biosynthetic processes such as the vitamin K-dependent gamma-carboxylation of glutamate residues in prothrombin synthesis	• The enzyme has anti-inflammatory effects and a protective role in protecting against cardiovascular injury including atherogenesis, dyslipidaemia, and insulin resistance.97

The cells with orange fill designate proteins for which a clear association with cardiovascular disease has not been reported.

Predicted risk: 1-[0.82exp(0.019*HE4 + 0.011*RBM9 + 0.11*Fibulin 5 + 0.049*BNP + 0.074*CDCP1 + 0.052*GLRX1 − 0.072*H6ST3 + 0.024*ERBB1 + 0.046*DAN − 0.17*TGFBRIII − 0.11*Carbonic anhydrase 6 − 0.037*MIC-1 + 0.030*MMP-12 + 0.020*Tropinin T − 0.079*Tetranectin − 0.058*C1GLC + 0.044*PTH − 0.043*CC126 + 0.11*ADA11 + 0.014*RNase 1 − 0.032*TM149 + 0.071*N-terminal pro-BNP − 0.057*B3GN2 + 0.038*GPNMB + 0.024*OPG + 0.059*RSPO4 + 0.040*FBLN3 − 0.11*CILP2 + 0.073*HDGF + 0.037*TIM-1 − 0.076*S31D4 + 0.091*NAD(P)H dehydrogenase − 0.28)].

Figure 2.

Time-dependent area under the curve values across 10 years of follow-up. The area under the curve values are depicted for each model in the Chronic Renal Insufficiency Cohort testing set (internal validation) and the Atherosclerosis Risk in Communities cohort (external validation). Area under the curve values for the Pooled Cohort Equation were the weighted sum of four different equations representing Black women, White women, Black men, and White men. The modified Pooled Cohort Equation included estimated glomerular filtration rate and the Pooled Cohort Equation risk factors (age, sex, race, systolic blood pressure, use of anti-hypertensive therapies, total cholesterol, high-density lipoprotein cholesterol, history of diabetes, and current smoking), all of which were refit to the Chronic Renal Insufficiency Cohort training set. The protein model was derived in the Chronic Renal Insufficiency Cohort training set using elastic net regression and consisted of 32 proteins.

Model dynamic range and calibration

The primary 32-protein model was well calibrated in both the CRIC and ARIC cohorts (Figure 3) and demonstrated a broad dynamic range of risk stratification across quintiles of predicted risk (P-value of Greenwood-Nam D’Agostino test for the calibration plot was not significant for either calibration plot). Further, within a 10-year horizon, the absolute predicted as well as observed incident cardiovascular event rates were ∼60% in the highest predicted risk quintile in both cohorts. The PCE and modified PCE both demonstrated relatively poor calibration with discordance between the predicted and observed risks in CRIC and ARIC.

Figure 3.

Calibration of proteomic risk model in the Chronic Renal Insufficiency Cohort and Atherosclerosis Risk in Communities cohort. (A) The Chronic Renal Insufficiency Cohort training set was comprised of 1742 participants without a self-reported history of coronary heart disease, myocardial infarction, stroke, or heart failure. Over the 10-year follow-up period, there were 373 incident cardiovascular events. (B) Similarly, the Atherosclerosis Risk in Communities study population was comprised of 485 participants with chronic kidney disease and without a history of cardiovascular disease at the time of proteomic measurements. Over a 10-year follow-up, there were 173 incident cardiovascular events.

Hybrid risk models

Hybrid risk models that were comprised of both clinical risk factors and proteins (Clinical and Protein Hybrid Model) performed similarly to the proteins only model. When the 15 clinical terms from the PCE were forced into the hybrid model (see Supplementary data online, Table S1), elastic net selected 11 proteins and the predictive performance was similar to the primary 32-protein model described earlier (Figure 4). When the clinical factors were allowed to compete with proteins (Competing Clinical and Protein Hybrid Model), elastic net regression selected only protein measures and none of the clinical variables (see Supplementary data online, Table S2). All 27 proteins selected in this manner were contained within the 32-protein model described above, and the predictive performance was similar.

Figure 4.

Time-dependent area under the curve values for hybrid models across 10 years of follow-up. The area under the curve values are depicted for each hybrid model in the Chronic Renal Insufficiency Cohort testing set (internal validation). Both hybrid models were developed in the Chronic Renal Insufficiency Cohort training set. In the Clinical and Protein Hybrid Model, estimated glomerular filtration rate and all variables from the Pooled Cohort Equation were forced into a final hybrid model that selected protein variables by elastic net regression. The final model consisted of 15 clinical terms and 11 proteins. In the Competing Clinical and Protein Hybrid Model, all clinical terms and proteins were allowed to compete for inclusion in a hybrid model derived by elastic net regression. No clinical variables were selected, and the final model consisted of 27 proteins (all of which are also contained in the 32-protein model shown in Figure 2).

Large-scale discovery of individual biomarkers of cardiovascular risk

We evaluated the associations of each of the 4830 aptamers representing 4638 unique proteins with the risk of incident cardiovascular disease. We have listed the HRs and their confidence intervals, nominal P-values, and FDR q-values for each of the 4830 aptamers in Supplementary data online, Table S3 (unadjusted analysis), Supplementary data online, Table S4 (adjusted for eGFR), and Supplementary data online, Table S5 (adjusted for age, sex, race, body mass index, systolic blood pressure, diastolic blood pressure, hypertension, diabetes, smoking, total cholesterol, HDL cholesterol, eGFR, and proteinuria). We found 2258 individual proteins associated with the risk of incident cardiovascular disease in unadjusted analysis (q-value < 0.05) (Figure 5A). After adjustment for eGFR to eliminate filtration markers, 396 proteins retained statistical significance (Figure 5B). After complete multi-variable adjustment, 74 proteins met statistical significance (q-value < 0.05) (Figure 5C).

Figure 5.

Volcano plots: proteomics of incident cardiovascular disease. The volcano plots depict the hazard ratios per log 2–transformed protein value. Analyses for associations between proteins and incident cardiovascular disease were evaluated in the full Chronic Renal Insufficiency Cohort without a baseline history of cardiovascular disease. There were a total of 2182 participants who contributed 459 incident cardiovascular events over the 10-year timeframe. (A) represents the unadjusted associations between individual proteins and incident cardiovascular disease in univariate analysis, (B) represents adjustment for estimated glomerular filtration rate; and (C) represents adjustment for age, sex, race, body mass index, systolic blood pressure, diastolic blood pressure, hypertension, diabetes, smoking, total cholesterol, HDL cholesterol, estimated glomerular filtration rate, and proteinuria.

Inferences of causality through Mendelian randomization

In the PheWAS database of phenome-wide association studies using cis-only genetic variant analyses,³⁰ protein quantitative trait loci (pQTLs) were identified for 36 of the 74 proteins that were independently associated with incident cardiovascular disease after complete multi-variable adjustment. The proteins, corresponding cardiovascular diseases and/or traits, causal effect size, and P-value assessed through Mendelian randomization analysis are reported in Supplementary data online, Table S6. There were 18 proteins that had causal associations with a cardiovascular endpoint including stroke, coronary heart disease, MI, or heart failure. In addition, 16 proteins had causal associations with cardiovascular risk factor traits such as systolic or diastolic blood pressure, body mass index, body fat, weight, diabetes, hypertension, total cholesterol, or HDL cholesterol.

A detailed review of the 18 proteins that were statistically significant in our study’s final multi-variable model and identified in the PheWAS browser as potentially causal of a clinical cardiovascular disease outcome suggests their fundamental role in either immunologic functions, vascular development, metabolic functions, or nervous system development/functions. Sushi von Willebrand factor type A, epidermal growth factor, and pentraxin domain–containing protein 1 (SVEP1) is a large extracellular protein expressed in vascular smooth muscle cells and implicated in atherosclerosis, coronary heart disease, hypertension, and diabetes.^33,34 SVEP1 is also important in the development of the lymphatic system³⁵ and contributes to the risk of dementia.³⁶ Sialic acid-binding Ig-like lectin 7 (SIGLEC7) prevents cytotoxicity by inhibiting natural killer cells activity. Scavenger receptor cysteine-rich type 1 protein M130 (CD163) is an acute phase receptor involved in the clearance and endocytosis of haemoglobin complexes that have been oxidized and are at risk of tissue damage. Other immune-related proteins that are potentially causative of incident cardiovascular disease include macrophage metalloelastase (MMP12), macrophage scavenger receptor type 1 and 2 (MSR1), and β2-microglobulin (B2M), which is part of the Class I major histocompatibility complex.

In addition, several of the proteins significant in both multi-variable modelling and previous Mendelian randomization studies have been implicated in both vascular development and neuronal and axonal development. Tenascin (TNC) is an extracellular matrix protein that guides migrating neurons and axons during development.³⁷ Pleiotrophin (PTN) is a growth factor that regulates neuronal cell proliferation, survival, differentiation, and migration.^38–40 In addition to promoting growth and glucose uptake, insulin-like growth factor 1 (IGF1) leads to synapse maturation.⁴¹ Two additional proteins also identified in the PheWAS browser and related to IGF1 included insulin-like growth factor binding protein 3 (IGFBP3) and SPARC-like protein 1 (SPARCL1). These two proteins are involved in the regulation of insulin-like growth factor. SPARCL1 also promotes hepatic fibrosis.^42,43 All these proteins are multi-functional and have also been associated with vascular remodelling including connective tissue and vascular integrity,^44,45 fibrosis,⁴⁴ and atherosclerosis.^46,47

Biologic pathway analysis

We based the IPA on 396 proteins whose HRs were associated with the primary outcome after eGFR adjustment (i.e. after eliminating kidney filtration biomarkers) at FDR of 5% and had a total of 4371 background proteins measured and recognized by IPA. This analysis yielded significant enrichment for proteins involved in immunologic and vascular function, cholesterol biosynthesis, and hepatic fibrosis (Table 3). Most of these canonical pathways were involved in immunologic or vascular signalling (acute phase response signalling, atherosclerosis signalling, complement system, coagulation system, extrinsic prothrombin activation, and granulocyte adhesion and diapedesis). In addition, the pathway for hepatic fibrosis had one of the most significant associations in this analysis.

Table 3.

Canonical pathways associated with incident CVD

IPA canonical pathway	P-value	Ratio
LXR/RXR heterodimer activation	4.04 x 10−8	0.434 (36/83)
Hepatic fibrosis/hepatic stellate cell activation	7.63 x 10−7	0.368 (42/114)
Acute phase response signalling	1.77 x 10−4	0.308 (40/130)
Atherosclerosis signalling	3.13 x 10−4	0.337 (28/83)
Complement system	4.35 x 10−4	0.464 (13/28)
FXR/RXR activation	6.32 x 10−4	0.338 (25/74)
Osteoarthritis pathway	0.00174	0.292 (33/113)
Coagulation system	0.00247	0.414 (12/29)
Extrinsic prothrombin activation pathway	0.00348	0.538 (7/13)
Granulocyte adhesion and diapedesis	0.00362	0.29 (29/100)

Ratio indicates the number of significant proteins that map to the canonical pathway divided by the total number of proteins measured in our study that map to the same pathway. For the IPA analysis, all protein associations with the primary cardiovascular outcome were adjusted for eGFR.

IPA, Ingenuity Pathway Analysis.

Discussion

In a population of 2667 individuals from two cohorts with a broad range of severity of non-dialysis-dependent CKD and free of overt cardiovascular disease, we applied a large-scale proteomic analysis using machine learning (elastic net) to generate a 32-protein model that identified those at extremely high or relatively low risk for developing MI, heart failure, stroke, or cardiovascular death (Structured Graphical Abstract). Those in the highest quintile of predicted risk had an observed incident cardiovascular event rate of 60% over 10 years. The reliable identification of such high-risk individuals with no prior history of cardiovascular disease will undoubtedly be of great interest to patients, their providers, and healthcare systems. Compared to their respective lowest quintiles, the risk in the top quintile was over 10-fold higher in CRIC and 4.3-fold higher in ARIC indicating a broad dynamic range of risk stratification, meaningfully separating patients at various levels of risk. The proteomics model consistently outperformed the currently recommended clinical risk model, the 2013 PCE, and its modified version that included eGFR for predicting incident cardiovascular events in both the CRIC testing set and ARIC validation cohort with gains in AUC of ∼0.20 at 3 years and ∼0.10 at 10 years in both cohorts.

Biological plausibility for proteomic models is high. Proteins orchestrate biological processes that under adverse conditions lead to diseases and their clinical manifestations.⁴⁸ Proteins integrate the effects of genes with the environment, age, comorbidities, lifestyle, and drugs. Proteins are thus more proximate to the pathogenesis of diseases and their outcomes than upstream traditional risk factors and can thus report an individual’s ‘risk of the risk factors’. In addition, unlike traditional risk factors many of which are not modifiable (e.g. age, sex, history of diabetes, and history of hypertension), proteins are well-suited for the task of risk modelling as they are imminently mutable as conditions change.²⁷ In addition, proteins are the targets of 95% of all known drugs,⁴⁹ and multi-protein models can accurately capture the effects of beneficial or harmful intervention on cardiovascular risk, regardless of the biological mechanisms involved.²⁷ As part of our large-scale analyses, we have reported the associations between each of the 4830 aptamers representing 4638 unique proteins and the risk of incident cardiovascular events before and after controlling for various markers of risk. These invaluable data sets inform hundreds of new biomarkers and potential mediators of cardiovascular risk and have the potential to guide future investigations aimed at developing cardioprotective therapies in CKD patients.

We have also demonstrated that the Clinical and Protein Hybrid Model, which consists of both clinical risk factors and proteins, does not enhance risk prediction over the 32-protein model alone. In fact, when all clinical and protein variables were allowed to compete for inclusion in the risk model (Competing Clinical and Protein Hybrid Model), only proteins and none of the clinical variables were selected by elastic net. Likely, the proteins selected into the risk model encode information conferred by demographics and traditional clinical risk factors, while providing additional independent information of prognostic value.⁴⁸

While the mean eGFR was similar between the two study cohorts, CRIC participants were younger and had a higher proportion of men and Blacks, and a higher prevalence of hypertension and diabetes than those in the ARIC study. Given the differences between these cohorts and diversity within each cohort recruited broadly, our findings suggest that the protein risk score may be generalizable and meaningfully superior to the traditional clinical risk scores in a broad population with CKD.

The clinical utility for accurate cardiovascular risk prediction in patients with CKD is exemplified by recent randomized clinical trials in CKD populations that have identified novel therapies to reduce the risk of cardiovascular events.^50,51 For example, in the Dapagliflozin and Prevention of Adverse Outcomes in CKD (DAPA-CKD) trial, CKD participants assigned to treatment with the sodium-glucose cotransporter 2 inhibitor dapagliflozin had a reduced risk of adverse cardiovascular events.⁵⁰ This benefit was observed independently of underlying diabetes and suggests that a broad population may potentially benefit from this class of medications. However, given the relatively high expense associated with this treatment, assessment of absolute risk by a proteomics risk score might identify individuals at the highest risk in whom this treatment (and other expensive therapies) might be most cost-effective.

Exploration of the underlying biology linking CKD to the development of incident cardiovascular disease consistently pointed to various immunologic/inflammatory canonical pathways.^52–54 Three of the 10 canonical pathways identified are classified under innate immune activation—acute phase response signalling, complement system, and granulocyte adhesion and diapedesis. Similarly, Mendelian randomization identified at least 6 of the 18 proteins (SVEP1, SIGLEC7, CD163, MMP12, MSR1, and B2M) that were statistically significant in our final modelling, potentially causative for an incident cardiovascular endpoint, and important in immunologic functions. Immunomodulatory strategies that target the downstream effectors of innate immunity including interleukin (IL)-1β reduce cardiovascular risk. Compared to placebo, canakinumab, a monoclonal antibody that selectively inhibits IL-1β, reduced the rates of recurrent cardiovascular events in patients with a history of MI and elevated high-sensitivity C-reactive protein levels.⁵⁵ Similar findings were observed in the CKD subgroup of this trial, and the cardiovascular benefits were directly correlated with the magnitude of reduction in the inflammatory response.⁵⁶ Our analyses provide further biologic support to investigating existing and novel therapies targeting some of the key proteins noted above that are within these pathways.

Identification of both the hepatic fibrosis pathway and SPARCL1 from Mendelian randomization studies provides evidence for another important intermediary linking CKD to cardiovascular disease. SPARCL1 is highly up-regulated in adipose tissue and correlates with pathological features including hepatic fibrosis in non-alcoholic steatohepatitis (NASH) patients.^42,43 Although CRIC participants were not specifically monitored for NASH, which has been reported to be associated with liver inflammation, fibrosis, and a high risk of cardiovascular disease, the consistent identification of this pathway and biomarker suggests a potential participation of liver abnormalities in CKD several years before the onset of clinical cardiovascular disease.^57,58

Several limitations should be considered when interpreting our findings. There are potentially more proteins in blood than the ∼5000 we analysed. Studies that evaluate an even larger number of proteins than reported in the present study are underway. In addition, the aetiology of CKD in our study is predominantly related to its two major population risk factors, type 2 diabetes and hypertension. We do not yet know how our multi-protein cardiovascular risk model might perform in the setting of less common aetiologies of CKD. Further, the PCE risk prediction tool was developed for fatal and non-fatal stroke, fatal coronary heart disease, and non-fatal MI, rather than the broader class of cardiovascular disease events that include heart failure and cardiovascular death. However, as the risk variables for estimation of the risk of incident heart failure are similar to those already contained in the PCE, its coefficients were refit to optimize its prediction for the broader endpoint while eGFR was added. Further, novel cardiac risk factors such as N-terminal pro-B-type natriuretic peptide and troponin have targeted assays; however, these were not considered in routine clinical evaluation as they are not part of the established PCE prediction tool. Finally, we did not assess albuminuria in the prediction modelling as it was not present at ARIC, Visit 3, which is the baseline for our external validation. We still adjusted for proteinuria in our multi-variable models from the CRIC data set when identifying proteins that are independently associated with incident cardiovascular disease in CRIC participants.

In summary, to the best of our knowledge, we have successfully conducted the largest proteomic study of incident cardiovascular risk in the CKD population to date. The proteomic risk prediction model provides a broad dynamic range of risk stratification and is meaningfully superior to models used in clinical practice that are comprised of clinical markers including eGFR. These results represent an important step toward implementation of personalized risk assessment. Further, genetic and pathway analysis, as well as examination of the biology of individual proteins associated with cardiovascular risk in this population, may suggest new targets of treatment.

Author contributions

Drs. Deo, Dubin and Ganz had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Deo, Dubin, Ganz; Acquisition, analysis, or interpretation of data: All authors; Drafting of the manuscript: Deo and Ganz; Critical revision of the manuscript for important intellectual content: Deo, Dubin, Ren, Wang, Coresh, Grams, Surapaneni, Segal, Li, Ganz; Statistical Analysis: Ren, Wang, Haotian Zheng, Zihe Zheng, Surapaneni, Segal, Li; Obtained funding: Deo, Dubin, Ganz; Study supervision: Deo, Dubin, Ganz

Supplementary material

Supplementary material is available at European Heart Journal online.

Data availability

The data underlying this article are available in the manuscript and in its online supplementary material.

Funding

Funding for this work was obtained from the National Institutes of Health U01DK108809 and R01HL159081. In addition, funding for the CRIC Study was obtained under a cooperative agreement from the National Institute of Diabetes and Digestive and Kidney Diseases (U01DK060990, U01DK060984, U01DK061022, U01DK061021, U01DK061028, U01DK060980, U01DK060963, U01DK060902, and U24DK060990). In addition, this work was supported in part by the following: the Perelman School of Medicine at the University of Pennsylvania Clinical and Translational Science Award NIH/NCATS UL1TR000003, Johns Hopkins University UL1 TR-000424, University of Maryland GCRC M01 RR-16500, Clinical and Translational Science Collaborative of Cleveland, UL1TR000439 from the National Center for Advancing Translational Sciences (NCATS) component of the National Institutes of Health and NIH roadmap for Medical Research, Michigan Institute for Clinical and Health Research (MICHR) UL1TR000433, University of Illinois at Chicago CTSA UL1RR029879, Tulane COBRE for Clinical and Translational Research in Cardiometabolic Diseases P20 GM109036, Kaiser Permanente NIH/NCRR UCSF-CTSI UL1 RR-024131, and Department of Internal Medicine, University of New Mexico School of Medicine Albuquerque, NM R01DK119199. Partial support for this work was also provided by the Winkelman Family Fund in Cardiovascular Innovation. The opinions expressed in this paper do not necessarily reflect those of the National Institute of Diabetes Digestive and Kidney Disease, the National Institutes of Health, and the Department of Health and Human Services or the Government of the United States of America.