Proteomic Assessment of the Risk of Secondary Cardiovascular Events among Individuals with CKD

Visual Abstract

Abstract

Key Points

• Machine learning and large-scale proteomics led to a 16-protein secondary cardiovascular risk model in patients with CKD.

• Hepatic fibrosis and liver X receptor activation represented biologic pathways that link kidney disease and risk of secondary cardiovascular events.

• An understanding of the circulating proteins associated with secondary cardiovascular events may help to identify novel therapeutic targets.

Background

Cardiovascular risk models have been developed primarily for incident events. Well-performing models are lacking to predict secondary cardiovascular events among people with a history of coronary heart disease, stroke, or heart failure who also have CKD. We sought to develop a proteomic risk score for cardiovascular events in individuals with CKD and a history of cardiovascular disease.

Methods

We measured 4638 plasma proteins among 1067 participants from the Chronic Renal Insufficiency Cohort (CRIC) and 536 individuals from the Atherosclerosis Risk in Communities (ARIC) Cohort. All had non–dialysis-dependent CKD and coronary heart disease, heart failure, or stroke at study baseline. A proteomic risk model for secondary cardiovascular events was derived by elastic net regression in CRIC, validated in ARIC, and compared with clinical models. Biologic mechanisms of secondary events were characterized through proteomic pathway analysis.

Results

A 16-protein risk model was superior to the Framingham Risk Score for secondary events, including a modified score that included eGFR. In CRIC, the annualized area under the receiver operating characteristic curve (area under the curve) within 1–5 years ranged between 0.77 and 0.80 for the protein model and 0.57 and 0.72 for the clinical models. These findings were replicated in the ARIC validation cohort. Biologic pathway analysis identified pathways and proteins for cardiac remodeling and fibrosis, vascular disease, and thrombosis.

Conclusions

The proteomic risk model for secondary cardiovascular events outperformed clinical models on the basis of traditional risk factors and eGFR.

Introduction

Patients with CKD are known to be at a heightened risk of cardiovascular events,^1,2 and their management has improved by the results of several recent clinical trials.^3–8 Studies evaluating sodium-glucose cotransporter-2 inhibitors, glucagon-like peptide 1 receptor agonists, nonsteroidal mineralocorticoid receptor agonists, and anti-inflammatory drugs such as canakinumab have enrolled patients with CKD and demonstrated reductions in the risk of both incident and secondary cardiovascular events. Despite these advances, cardiovascular risk stratification in the CKD population remains limited. Current approaches are unable to provide precise estimates of risk for developing secondary cardiovascular events. In addition to personalized care, accurate cardiovascular risk stratification could also make the conduct of clinical trials more efficient and less costly by providing a tool to enroll individuals at high risk of such events.

Cardiovascular risk stratification models have been developed primarily to predict incident disease in individuals free of cardiovascular disease, using algorithms such as the Pooled Cohort Equation⁹ and Multiethnic Study of Atherosclerosis Risk Score.¹⁰ Limited models, however, are available that classify the risk of secondary cardiovascular events among people with a history of coronary heart disease, stroke, or heart failure, particularly among individuals with CKD. Notably, the Framingham Risk Score for secondary events was derived in individuals who have known atherosclerotic cardiovascular disease,¹¹ but this algorithm was not developed for individuals with CKD.

Prior work has demonstrated that GFR has a widespread effect on the circulating functional proteome, potentially leading to the heightened cardiovascular risk observed in patients with CKD.¹² In this study, we thus sought to develop a protein-based risk score for cardiovascular events in individuals with CKD and a history of cardiovascular disease. Our objectives were to (1) identify new protein biomarkers of secondary cardiovascular risk in the CKD population with prevalent cardiovascular disease, (2) derive and validate a multiprotein cardiovascular risk score and compare its predictive performance with secondary event risk scores composed of traditional clinical risk factors alone or hybrid models composed of both proteins and traditional clinical risk factors, (3) infer biological mechanisms associated with the risk of secondary cardiovascular events in CKD using the tools of pathway analysis of proteins, and (4) compare and contrast the underlying biological mechanisms for secondary cardiovascular events and previously published incident cardiovascular events.¹³

Methods

We derived risk models by evaluating individuals from the Chronic Renal Insufficiency Cohort (CRIC), a prospective observational study that enrolled adults with CKD, defined as an eGFR <60 ml/min per 1.73 m², between 2003 and 2008.¹⁴ This analysis includes 3416 individuals who completed the study visit at year 1, which was considered this study's baseline, and had cryopreserved plasma samples available for proteomic analysis. Participants who were on dialysis or had a history of kidney transplantation were excluded (n=50). Because of the interference of lupus antibodies with aptamers (communication from SomaLogic), we also excluded 12 participants with SLE. Furthermore, 105 samples were excluded because they did not pass SomaLogic's quality control standards. We excluded an additional 2182 participants who did not have a history of cardiovascular disease and were included in a proteomic analysis of incident events.¹³ As a result, the final CRIC sample for this analysis included 1067 participants with a history of myocardial infarction, heart failure, or stroke at our study baseline.

We developed a proteomic risk score for secondary cardiovascular events consisting of nonfatal myocardial infarction, nonfatal stroke, heart failure hospitalization, or cardiovascular death. We first validated the model in an internal testing set, which comprised CRIC participants not included in model development. We then validated the model in an external cohort of 536 participants from the Atherosclerosis Risk in Communities (ARIC) Cohort,¹⁵ all of whom had CKD (eGFR <60 ml/min per 1.73 m²) not requiring maintenance dialysis and a history of either myocardial infarction, stroke, or heart failure at the time their plasma was sampled for proteins (ARIC visit 5, 2011–2013). We evaluated the associations of individual proteins with secondary cardiovascular events. Pathway analysis of proteins associated with the study outcome was performed to characterize the biological mechanisms of secondary cardiovascular events. We then compared the proteins and biological pathways identified for secondary cardiovascular events with those identified previously for incident cardiovascular events. All analyses were approved by the appropriate institutional review boards, and all participants provided written informed consent.

Proteomic Assessment with Modified Aptamers

Cryopreserved plasma specimens from both the CRIC and ARIC studies were sent to SomaLogic and evaluated using SomaScan version 4.0, which comprises 5284 aptamer-binding reagents. The process for protein quantification, including normalization steps, has been described previously.^13,16 On the basis of exclusion of aptamers developed to nonhuman proteins and on our quality control analysis within the CRIC study,¹⁷ our study reports on 4830 aptamers. They had median intra-assay and interassay coefficients of variation <5% and represented 4638 unique proteins (some proteins are measured by two or more aptamers).

Primary Outcome: Secondary Cardiovascular Events

The primary study outcome was defined as an occurrence of a nonfatal myocardial infarction, nonfatal stroke, heart failure hospitalization, or cardiovascular death within a 5-year time horizon from blood sample collection for proteomic analyses.^18–21 This time horizon matched that for the secondary Framingham Risk Score.¹¹ In further analyses, we extended model assessment beyond 5 years to additional years of follow-up on the basis of data availability in the CRIC and ARIC Cohort.

Clinical Variables

For clinical risk modeling for individuals with prevalent cardiovascular disease, we evaluated prediction using the secondary Framingham Risk Score,¹¹ which included the following variables: age, sex, total cholesterol, HDL cholesterol, systolic BP, diabetes, and smoking status. To optimize the performance of the clinical risk score in the CKD population, we also developed a modified secondary Framingham risk equation that added eGFR to the other variables in the secondary Framingham Risk Score. Because we modified the equation by adding eGFR and to optimize its performance in the CKD population, we refit the coefficients for all variables (and eGFR) to the CRIC study population and for the same 5-year time horizon as the protein model.

Statistical Analyses

The primary model consisted of proteins only. To evaluate its predictive performance, we randomly split the CRIC sample of 1067 participants into a training (80%) set and a testing (20%) set. The training set was used to build prediction models and select the optimal parameters. The testing set was used for internal validation of the model's performance.

We developed risk prediction models using elastic net, a machine learning technique that accommodates a large number of features (proteins) and time-to-event outcomes while providing, via combined ridge (L2) and least absolute shrinkage and selection operator (L1) penalties, parsimonious and interpretable models. Cox regression elastic net models were fitted using the R package glmnet.^22,23 The relative contribution of the two penalties was 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 ten-fold cross-validation of the partial likelihood deviance and associated 1-SEM rule. Having so obtained a final protein set, we refit (relaxed) these selected proteins using a conventional Cox regression model to reduce bias in the elastic net regression coefficients.²⁴ For conclusions not to be overly influenced by the initial (random) training–testing set partitioning, we performed stability analyses on the basis of five alternative random partitions.

We evaluated model performance in both the CRIC testing set and ARIC external validation cohort. Model discrimination was assessed by calculation of each model's area under the receiver operating characteristic curve (AUC) annually using time-dependent methodology.²⁵ 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 model was compared with the clinical or hybrid (comprising proteomic and clinical variables) model using the harmonic mean P value approach²⁶ to generate a combined P value that accommodates dependent tests. The AUC values for the secondary Framingham Risk Score model were the weighted sum of two different equations representing women and men.

We also developed two different hybrid models in the CRIC training set that used the variables from the secondary Framingham Risk Score and all 4638 proteins. In the first hybrid model, which was termed “noncompeting, hybrid model,” all clinical variables from the secondary Framingham Risk Score were forced into the model and were thus not subject to selection. The protein variables were then selected by elastic net. Coefficients for the mandated clinical variables and any selected protein variables were determined by Cox regression. In the second hybrid model, termed “competing, hybrid model,” we allowed all the proteins and clinical terms to compete for inclusion in an elastic net model. The proportional hazards assumption held for all proteins in the selected model. Model performance was assessed in the CRIC testing set.

Model calibration incorporated all the data and depicted bar plots that indicated agreement between average predicted risk and observed risk in each of the quintiles defined by the predicted risk score. The calibration of the CRIC-derived multiprotein model was evaluated in the ARIC Cohort 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.²⁷

Mechanistic Discovery/Biology/Pathway Analysis

To gain biological insights into the mechanisms of secondary cardiovascular events in CKD, we characterized the individual proteins and corresponding biological pathways implicated in secondary cardiovascular events. Cox proportional hazards models were used to assess the association between the plasma levels of individual proteins and secondary cardiovascular events in all 1067 CRIC participants. The log₂ hazards ratios were evaluated, and the Benjamini–Hochberg method was used to control the false discovery rate at 5%. Volcano plots represent the (1) unadjusted associations and multivariable models that sequentially adjusted for (2) eGFR and (3) age, sex, race, body mass index, systolic BP, diastolic BP, hypertension, diabetes, smoking, total cholesterol, HDL cholesterol, eGFR, and proteinuria. We also compared the risk estimates between each protein and secondary cardiovascular events with risk estimates previously reported for incident cardiovascular events in CRIC using the same version of SomaScan and analytical strategy.¹³ To surmise whether there are biological differences between incident compared with secondary cardiovascular events in CKD, we assessed the correlation (Spearman rho) for the two hazard ratios representing associations between each protein and cardiovascular outcome after adjustment for eGFR.

We entered all proteins, their corresponding risk estimates after adjustment for eGFR, and q values in pathway analyses using ingenuity pathway analysis (IPA) software (Ingenuity Systems Inc., Redwood City, CA; www.ingenuity.com).^28,29 A total of 4731 proteins were recognized by IPA and served as the background. Fisher's right-tailed exact test was used to calculate P value to determine the probability that the association of the differentially expressed proteins in the measured dataset and known canonical pathways was explained by chance alone.

Statistical analyses were performed using R, version 4.0.3 (RStudio, Inc., Boston, MA; http://www.rstudio.com/), with the packages 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).

Results

We evaluated 1067 participants from CRIC and 536 individuals from the ARIC Cohort; their characteristics at the time of proteomic measurements are listed in Table 1. Mean (SD) eGFR was 39 (15) ml/min per 1.73 m² in CRIC and 44 (12) ml/min per 1.73 m² in the ARIC Cohort. Compared with those in the ARIC Cohort, CRIC participants were younger, more likely to be men, and more likely to be Black. The CRIC participants were more likely to have a history of hypertension and diabetes and less likely to be active smokers.

Table 1.

Demographic and clinical characteristics at the time of proteomic measurements

Variable	Derivation Cohort (CRIC), N=1067	Validation Cohort (ARIC), N=536
Demographics
Age, yr, mean (SD)	63 (8)	79 (5)
Female, n (%)	424 (40)	241 (45)
Self-reported Black race, n (%)	508 (48)	112 (21)
Self-reported White race, n (%)	452 (42)	424 (79)
Prior cardiovascular disease, n (%)
Coronary heart disease	750 (70)	355 (66)
Heart failure	331 (31)	242 (45)
Stroke	354 (33)	93 (17)
Cardiovascular disease risk factors, n (%)
Hypertension	1021 (96)	448 (86)
Diabetes	665 (62)	260 (51)
Current tobacco use	148 (14)	155 (31)
Measures
Systolic BP, mm Hg, mean (SD)	131 (23)	131 (21)
Diastolic BP, mm Hg, mean (SD)	68 (13)	63 (12)
Body mass index, kg/m2, mean (SD)	33.1 (7.6)	29.9 (6.1)
Total cholesterol, mg/dl, mean (SD)	171 (44)	166 (51)
HDL cholesterol, mg/dl, mean (SD)	46 (14)	46 (13)
LDL cholesterol, mg/dl, mean (SD)	92 (34)	92 (37)
eGFR, ml/min per 1.73 m2, mean (SD)	39 (15)	44 (12)
Urine protein-creatinine ratio, g/g, median (IQR)	0.17 (0.06–0.84)	—
Urine albumin-creatinine ratio, mg/g, median (IQR)	—	21 (9–79)
Medications, n (%)
Taking antihypertensive medication	1054 (99)	442 (85)
Statins	803 (76)	374 (70)

ARIC, Atherosclerosis Risk in Communities; CRIC, Chronic Renal Insufficiency Cohort; IQR, interquartile range.

Risk of Secondary Cardiovascular Events

Over the 5-year follow-up, there were 388 cardiovascular events in CRIC (309 events in the training set and 79 events in the testing set) that occurred after a mean of 3.6±1.7 years. There were 249 cardiovascular events in the ARIC Cohort that occurred after a mean of 3.2±1.8 years. From the CRIC training set, of a total 4638 measured proteins, elastic net regression selected 16 proteins for the proteomics risk model (Table 2). The proteins that had the largest coefficients on cardiovascular risk prediction included N-terminal pro-B-type natriuretic peptide (NT-proBNP), anthrax toxin receptor 2, CUG-binding protein Elav-like family member 2, tenascin, and cartilage intermediate layer protein 2. In addition, several proteins, which are known biomarkers of cardiac stress and tissue injury, were selected. Troponin T,³⁰ troponin I, β-2 microglobulin,^31,32 kidney injury molecule-1 (KIM-1),^33–35 parathyroid hormone levels (parathyroid hormone), ^36,37 and NT-proBNP³⁸ are known, independent markers of risk of cardiovascular disease in individuals with CKD.

Table 2.

Proteins and corresponding coefficients in the risk prediction model

UniProt	Protein Name	Coefficient in the Protein-Only Model	Protein Present in the Hybrid, Noncompeting Model	Protein Present in the Hybrid, Competing Model
P16860	NT-proBNP	0.19	✓	✓
P58335	Anthrax toxin receptor 2	−0.16		✓
O95319	CUGBP Elav-like family member 2	0.11	✓	✓
P24821	Tenascin	0.090	✓	✓
Q8IUL8	Cartilage intermediate layer protein 2	−0.080		✓
P45379	Troponin T, cardiac muscle	0.078	✓	✓
P14555	Phospholipase A2, membrane associated	0.078	✓
P61626	Lysozyme C	0.073
P19429	Troponin I, cardiac muscle	0.067	✓	✓
P14649	Myosin light chain 6B	0.059
Q96D42	KIM-1	0.058		✓
P61769	β-2 microglobulin	0.056	✓	✓
P01270	PTH	0.055		✓
O43251	RNA binding protein fox-1 homolog 2	0.020	✓	✓
Q13790	Apolipoprotein F	0.0050	✓	✓
Q4LDE5	Sushi von Willebrand factor type A, EGF, and pentraxin domain containing protein 1	0.0049	✓	✓

CUGBP, CUG-binding protein; KIM-1, kidney injury molecule-1; NT-proBNP, N-terminal pro-B-type natriuretic peptide; PTH, parathyroid hormone.

The proteomics risk model was consistently associated with a higher risk of secondary cardiovascular events and surpassed the prognostic utility of clinical models. In both the CRIC and ARIC Cohort, higher proteomic quintiles demonstrated similar associations with a higher risk of events across follow-up (Figure 1). In quintile 5, nearly all the participants in both cohorts had a secondary cardiovascular event by year 5, highlighting the utility of the risk model to identify patients at extremely high secondary cardiovascular risk in this population. In addition, the discrimination of the 16-protein model, depicted in Figure 2 as a time-dependent AUC, was superior to both the secondary Framingham Risk Score (P = 0.001) and the modified secondary Framingham risk equation (P = 0.003), which included eGFR and had all the Framingham model's coefficients refit to the CRIC training set. Internal validation among the CRIC testing set demonstrated AUC values across the 10-year follow-up that ranged between 0.75 and 0.81 for the protein model, 0.57 and 0.64 for the secondary Framingham Risk Score, and 0.66 and 0.74 for the modified and refit secondary Framingham equation that included eGFR (Figure 2). Similar findings and gains in AUC for the proteomics model were observed in the external ARIC validation set. AUC values for the 7.5-year follow-up available in the ARIC Cohort ranged between 0.66 and 0.74 for the 16-protein model and were significantly higher than the AUC values for the secondary Framingham Risk Score (AUC, 0.51–0.57; P < 0.001) and modified secondary Framingham risk equation (AUC, 0.59–0.66; P = 0.004). Model performance remained stable with repeated random training–testing set partitioning (data not shown).

Figure 1.

Kaplan–Meier survival curves of secondary cardiovascular events according to proteomic risk quintiles. The risk of secondary cardiovascular events is depicted for each quintile of the proteomic risk score in the (A) internal (CRIC) and (B) external (ARIC Cohort) validation cohorts. Most of the participants in the highest quintile of risk had a secondary cardiovascular event by year 5 of follow-up. ARIC, Atherosclerosis Risk in Communities; CRIC, Chronic Renal Insufficiency Cohort.

Figure 2.

Time-dependent AUC values. The AUC values are depicted for each model in the CRIC testing set (internal validation) and ARIC Cohort (external validation). The follow-up periods were 10 years in CRIC and 7.5 years in the ARIC Cohort. The protein model was derived in the CRIC training set using elastic net regression and consisted of 16 proteins. AUC values for the secondary Framingham Risk Score were the weighted sum of two different equations representing women and men. The modified clinical risk score included eGFR and the risk markers from the secondary Framingham Risk Score. AUC, area under the curve; FRS, Framingham Risk Score.

The 16-protein model was well calibrated in both the CRIC and ARIC Cohort (P value of Greenwood–Nam–D'Agostino test was NS for either calibration plot) and demonstrated a broad dynamic range of risk stratification across quintiles of predicted risk (Figure 3). In CRIC, this gradient translated to 10% of participants in quintile 1 and >70% of participants in quintile 5 having a secondary cardiovascular disease event during the 5-year follow-up. Similarly, in the ARIC Cohort, the gradient of risk spanned from 30% of participants in quintile 1 having a secondary cardiovascular event to >70% in quintile 5 (Figure 3). This risk gradient was attenuated in both cohorts for the modified secondary Framingham risk model that included eGFR with a risk gradient between quintiles 1 and 5 ranging from approximately 25% to 50% in CRIC and slightly more in the ARIC Cohort (Supplemental Figure 1). No gradient in cardiovascular risk was observed when assessing the secondary Framingham Risk Score alone (Supplemental Figure 2).

Figure 3.

Calibration of proteomic risk models. (A) The CRIC comprised 1067 participants with a self-reported history of coronary heart disease, myocardial infarction, stroke, or heart failure. Over the 5-year follow-up period, there were 388 incident cardiovascular events. (B) Similarly, the ARIC study population comprised 536 participants with CKD and a history of cardiovascular disease at the time of proteomic measurements. Over a 5-year follow-up, there were 249 secondary cardiovascular events.

Hybrid Risk Models

Hybrid risk models that comprised both clinical risk factors and proteins performed no better than the protein-only model. When the demographics and clinical risk factors from the secondary Framingham Risk Score were forced into this hybrid model (noncompeting, hybrid model), elastic net selected 11 proteins (Supplemental Table 1), and the predictive performance was similar to that of the primary 16-protein model (Supplemental Figure 3). When the clinical factors were allowed to compete with proteins (hybrid, competing model), elastic net selected 14 proteins and none of the clinical variables (Supplemental Table 2). The predictive performances were similar to that of the primary protein model (Supplemental Figure 3). With exception of one protein in each hybrid model, the remaining proteins selected were contained within the 16-protein model. Furthermore, the proteins with the highest coefficients in both hybrid models were NT-proBNP, CUG-binding protein Elav-like family member 2, and tenascin.

Biologic Insights and Pathway Analysis

We have reported the risk estimates between each of the 4830 aptamers representing 4638 unique proteins and secondary cardiovascular events in Supplemental Table 3 (unadjusted analysis), Supplemental Table 4 (adjusted for eGFR), and Supplemental Table 5 (adjusted for age, sex, race, body mass index, systolic BP, diastolic BP, hypertension, diabetes, smoking, total cholesterol, HDL cholesterol, eGFR, and proteinuria). We found 1197 individual proteins associated with the risk of secondary cardiovascular events in unadjusted analysis (q value <0.05, Figure 4A). After adjustment for eGFR to eliminate protein filtration markers, 213 proteins retained statistical significance (Figure 4B). After complete multivariable adjustment, 149 proteins remained significant (q value <0.05) (Figure 4C). Of the proteins that were significant in each of these models, we noted that the vast majority were the same when using a competing-risk analysis (data not shown).

Figure 4.

Volcano plots: proteomics of secondary cardiovascular disease. The volcano plots depict the hazard ratios per log₂-transformed protein value. Analyses for associations between proteins and secondary cardiovascular disease were evaluated in the full CRIC with a baseline history of cardiovascular disease. There were a total of 1067 participants who contributed 388 secondary cardiovascular events over the 5-year time frame. (A) Unadjusted associations between individual proteins and secondary cardiovascular disease; (B) adjustment for eGFR; and (C) adjustment for age, sex, race, body mass index, systolic BP, diastolic BP, hypertension, diabetes, smoking, eGFR, and albuminuria.

Effect Sizes of Protein Predictors of Incident versus Secondary Cardiovascular Events

It is well-accepted practice in epidemiology to distinguish between incident and secondary cardiovascular events. To clarify whether this distinction is based on biological differences between the two types of events, we compared the protein hazard ratios for secondary cardiovascular events with incident cardiovascular events that we have previously reported using the same methodology¹³ in CRIC. This analysis displayed in a scatterplot demonstrates generally similar effect sizes (hazard ratios) for the two types of events (eGFR model, rho correlation 0.76, P value < 2.2×10⁻¹⁶, Figure 5A; fully adjusted model, rho correlation 0.65, P value <2.2×10⁻¹⁶, Figure 5B).

Figure 5.

Hazard ratios when comparing single-protein associations with incident and secondary cardiovascular events. In (A) and (B), the y axis indicates the hazard ratio for a given protein and secondary cardiovascular events. The x axis indicates the hazard ratio for a given protein and incident cardiovascular events. (A) Comparison of the hazards ratios for all proteins that were significantly associated with incident cardiovascular disease and/or secondary cardiovascular events after adjustment for eGFR. The rho correlation for these hazards ratios was 0.76, P value < 2.2×10⁻¹⁶. (B) A similar comparison of the hazards ratios for all proteins that were significantly associated with incident cardiovascular disease and/or secondary cardiovascular events after adjustment for age, sex, race, body mass index, systolic BP, diastolic BP, hypertension, diabetes, smoking, eGFR, and albuminuria. The rho correlation for these hazards ratios was 0.65, P value < 2.2×10⁻¹⁶. Proteins that had an identical hazard ratio for both cardiovascular end points would have been depicted on the dashed diagonal line. CVD, cardiovascular disease.

Pathway Analysis

The IPA tool recognized a total of 4371 proteins. The analysis, which was based on the significance values of the risk estimates between proteins and secondary cardiovascular events after adjusting for glomerular filtration (to eliminate a large number of filtration markers), demonstrated significant enrichment for proteins involved in hepatic and cardiac fibrosis, liver X receptor (LXR)/retinoid X receptor activation, and atherosclerosis (Table 3). These same pathways were also the ones most significantly linked to incident cardiovascular events in the CRIC study.¹³

Table 3.

Canonical pathway analysis of individual proteins associated with secondary cardiovascular events

IPA Canonical Pathway	P Valuea	Ratiob
Hepatic fibrosis/hepatic stellate cell activation	1.36×10−6	0.35 (40/114)
LXR/RXR activation	1.52×10−6	0.39 (32/83)
Atherosclerosis signaling	3.06×10−4	0.33 (27/83)
Extrinsic prothrombin activation pathway	0.003	0.54 (7/13)
Pyruvate fermentation to lactate	0.005	1 (3/3)
Coagulation system	0.005	0.38 (11/29)
Acute-phase response signaling	0.008	0.25 (33/130)
Osteoarthritis pathway	0.01	0.26 (29/113)
White adipose tissue browning pathway	0.02	0.31 (13/42)
FXR/RXR activation	0.02	0.27 (20/74)

The canonical pathway analysis in ingenuity pathway analysis considered all the individual protein associations with secondary cardiovascular events (after adjustment for eGFR). FXR, farnesoid X receptor; IPA, ingenuity pathway analysis; LXR, liver X receptor; RXR, retinoid X receptor.

^aThe P value refers to the P value of overlap and is calculated using the right-tailed Fisher's exact test. It represents the probability of association of the selected proteins from the current dataset with the canonical pathways on the basis of random chance alone.

^bThe ratio is the quotient of the numbers in parentheses, which includes the number of significant proteins that map to the canonical pathway (numerator) divided by the total number of proteins measured in our study that map to the same pathway (denominator).

Discussion

In 1603 individuals from two cohorts with CKD and prevalent cardiovascular disease, we derived and validated a proteomics model for predicting the risk of secondary cardiovascular events. In the internal and external validation sets, the proteomics model was superior to the Framingham Risk Score for secondary cardiovascular events, even after the performance of the Framingham model was optimized by inclusion of eGFR and by refitting its coefficients to a population with CKD (CRIC). Notably, the proteomics model displayed a broader gradient in the predicted and observed 5-year secondary cardiovascular event risk than any of the clinical models. Biologic assessment of the pathways underlying secondary events identified pathways and proteins that have important roles in cardiac remodeling and fibrosis, vascular disease, and thrombosis.

Our findings provide a potentially important step toward personalized risk stratification of individuals with CKD and history of cardiovascular disease, using proteomics. Approximately 20% were at a risk generally viewed as moderate, with 10%–30% event rates within 5 years. Notably, in the highest 20% of the predicted risk, over 70% of participants were affected by a cardiovascular event within 5 years, in both the CRIC and ARIC Cohort. This exceptionally high-risk subgroup warrants appropriate counseling to ensure an optimal shared decision-making approach when evaluating additional therapies, interventions, participation in clinical trials of new treatments, or palliative care. In addition, the proteomics platform, which has received US Food and Drug Administration support for purposes of monitoring drug effects and stratifying risk among trial participants, could facilitate risk stratification on a time scale consistent with the duration of clinical trials. As such, proteomic modeling has the potential to enrich clinical trials of CKD for participants at moderate to high risk of cardiovascular events.

Clinical guidelines, drug regulatory agencies, and well-accepted epidemiologic practices have traditionally divided cardiovascular outcomes into incident events and recurrent, secondary events. Predictive cardiovascular models on the basis of traditional risk factors have performed less well in patients with prevalent disease because these individuals are often treated aggressively for their risk factors, thereby weakening or even paradoxically reversing the associations of those risk factors with cardiovascular outcomes.³⁹ We identified similar proteins and pathways underlying both incident¹³ and recurrent cardiovascular disease risks. The associations of individual proteins with incident events and secondary events had similar effect sizes (Figure 5) and represented similar biological pathways. In particular, pathways and markers involved in regulating systemic fibrosis and cardiac fibroblast activity were important in linking CKD and cardiovascular disease risk regardless of whether the population included those with prevalent cardiovascular disease, such as in our analysis, or individuals without a history of cardiovascular disease.¹³ In addition, the LXR pathway was another leading canonical pathway identified in both the incident and secondary cardiovascular disease analyses. LXRs regulate fibroblast proliferation and myocardial fibrosis, and these receptors are expressed 10- to 15-fold higher in nonmyocytes, including cardiac fibroblasts and endothelial cells, compared with myocytes.⁴⁰ These findings suggest that the time-honored distinction between incident and secondary cardiovascular events may be less important from the viewpoint of underlying biological mechanisms.

Several limitations should be noted when interpreting our findings. There were more proteins in blood than the approximately 5000 we analyzed. Studies that evaluate an even larger number of proteins than in this study may report additional insights. In addition, the etiology of CKD in our study was predominantly related to its two major population risk factors—type 2 diabetes and hypertension. We do not yet know how our multiprotein cardiovascular risk model might have performed in the setting of less common etiologies of CKD. In addition, the overall generalizability of our findings is limited because we included only those CRIC participants who survived to and completed the year 1 visit. Those who died within the first year of enrollment into the cohort may have represented a particularly high-risk subgroup. Furthermore, although the proteomic risk score provides a more precise estimate of cardiovascular event risk in a secondary prevention population, its current, clinical applications are limited. The lowest quintile remained at substantial 5-year risk of events that it would have justified use of current therapies such as sodium-glucose cotransporter-2 inhibitors, glucagon-like peptide 1 receptor agonists, and nonsteroidal mineralocorticoid receptor antagonists. Furthermore, there is a need to evaluate proteomics for risk stratification in more contemporary populations receiving recently introduced cardioprotective therapies.

In summary, we have successfully conducted the largest proteomic study of secondary cardiovascular events risk in the CKD population to date. The proteomic risk prediction model provided a broad dynamic range of risk stratification and was superior to accepted, secondary prevention models that comprise clinical markers, including eGFR. We reported hundreds of protein biomarkers for secondary cardiovascular events, many not previously reported, and their biological pathways and mechanisms. These results represent a step toward implementation of personalized risk assessment.

Supplementary Material

jasn-36-231-s001.pdf ^{(1.4MB, pdf)}

jasn-36-231-s002.pdf ^{(302.8KB, pdf)}

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/E867.

Funding

This work was supported by National Institute of Diabetes and Digestive and Kidney Diseases (U01DK108809) and National Heart, Lung, and Blood Institute (HL159081 and HL161303).

Author Contributions

Conceptualization: Rajat Deo, Ruth F. Dubin, Harold I. Feldman, Peter Ganz, Ramachandran S. Vasan.

Data curation: Rajat Deo, Peter Ganz, Yue Ren.

Formal analysis: Rajat Deo, Peter Ganz, Hongzhe Li, Yue Ren, Mark Segal, Aditya L. Surapaneni, Jianqiao Wang.

Funding acquisition: Rajat Deo, Ruth F. Dubin, Harold I. Feldman, Peter Ganz.

Investigation: Rajat Deo, Ruth F. Dubin, Peter Ganz, Hongzhe Li.

Methodology: Josef Coresh, Rajat Deo, Ruth F. Dubin, Harold I. Feldman, Peter Ganz, Morgan E. Grams, Hongzhe Li, Yue Ren, Mark Segal, Ramachandran S. Vasan.

Project administration: Rajat Deo, Harold I. Feldman, Peter Ganz, Ramachandran S. Vasan.

Resources: Rajat Deo, Peter Ganz.

Software: Hongzhe Li, Yue Ren.

Supervision: Rajat Deo, Peter Ganz, Hongzhe Li.

Validation: Josef Coresh, Rajat Deo, Peter Ganz, Morgan E. Grams, Yue Ren, Aditya L. Surapaneni, Jianqiao Wang.

Visualization: Rajat Deo.

Writing – original draft: Rajat Deo, Peter Ganz.

Writing – review & editing: Jordana B. Cohen, Josef Coresh, Mirela Dobre, Ruth F. Dubin, Harold I. Feldman, Alan S. Go, Morgan E. Grams, Jiang He, Mayank Kansal, Tanika N. Kelly, Paul L. Kimmel, Hongzhe Li, Mahboob Rahman, Yue Ren, Mark Segal, Haochang Shou, Ramachandran S. Vasan, Jianqiao Wang.

Data Sharing Statement

Data related to transcriptomic, proteomic, or metabolomic data. Original data created for the study are or will be available in a persistent repository upon publication. Analyzable Data. Aggregated Data. CRIC. National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) Repository. The proteomics data will be submitted to the CRIC and NIDDK repository and linked with all the clinical data. This process has started and will be completed in the near future. CRIC repository will have all the proteomic data in the future. For now, we are submitting an Excel file (Supplemental Material) that lists single protein associations (for all approximately 5000 proteins and the risk of a secondary cardiovascular end point).

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/JSN/E866, http://links.lww.com/JSN/E935.

Supplemental Table 1. Clinical markers and proteins and corresponding coefficients in the noncompeting clinical and protein hybrid model.

Supplemental Table 2. Proteins and corresponding coefficients in the competing clinical and protein model.

Supplemental Table 3. Associations between individual proteins and secondary cardiovascular disease events (unadjusted).

Supplemental Table 4. Associations between individual proteins and secondary cardiovascular disease events after adjustment for kidney function.

Supplemental Table 5. Associations between individual proteins and secondary cardiovascular disease events after complete multivariable adjustment.

Supplemental Figure 1. Calibration of modified clinical equation in the CRIC and ARIC cohorts.

Supplemental Figure 2. Calibration of the secondary Framingham Risk Score in the CRIC and ARIC Cohort.

Supplemental Figure 3. Time-dependent AUC values for hybrid models across 10 years of follow-up.