Polygenic Risk Scores of Coronary Heart Disease in Chronic Kidney Disease Population: The CRIC Study

Abstract

Background:

Cardiovascular disease is a major cause of death in individuals with chronic kidney disease (CKD), thus improving its prediction is critical for cardiovascular disease risk reduction. Studies have not assessed whether a polygenic risk score (PRS) of coronary heart disease (CHD) developed from genetic studies in the general populations can improve cardiovascular disease risk prediction in individuals with CKD, especially when PRS is added to the American Heart Association Predicting Risk of Cardiovascular Disease EVENTs (PREVENT) equations.

Methods:

Focusing on a 10-year composite CHD events, we applied PRSs of CHD obtained from the polygenic score catalog to the multi-centered Chronic Renal Insufficiency Cohort (CRIC) participants, selected the best-performed PRS and evaluated the added value of the best PRS while accounting for PREVENT linear predictors. Three Cox proportional hazard models were fitted with PRS only, PREVENT predictors only, and PRS+PREVENT risk factors. We evaluated the prediction accuracy, discrimination and calibration of the combined model compared to the PREVENT model.

Results:

A PRS of CHD was significantly associated with 21% hazard of CHD in the PRS+PREVENT risk model (95% CI: [1.06, 1.39], two-sided Wald test p-value = 5.5e-3). The PRS+PREVENT model demonstrated higher area under the curve (AUC 72.4%, 95% CI: [0.66, 0.79] versus 70.8%, 95% CI: [0.64, 0.77]), better discrimination ability and improved calibration compared to the PREVENT model predictors. Individuals were more correctly classified in prediction quartiles using the PRS+PREVENT model for the 10-year CHD event proportions.

Conclusions:

Our results demonstrate the added value of a CHD PRS to the PREVENT risk prediction models in individuals with CKD. The combined PRS+PREVENT model has the potential to improve clinical decisions on disease prevention and treatment strategies.

Introduction

Cardiovascular disease is the leading cause of death worldwide and a major contributor to disability¹. Individuals with chronic kidney disease (CKD) are at high risk of developing cardiovascular disease, as they are exposed to nontraditional, uremia-related cardiovascular disease risk factors, including inflammation, oxidative stress, and abnormal calcium-phosphorus metabolism^2,3. Acute myocardial infarction (MI) has four times higher prevalence in individuals with CKD than those without, while the odds for stroke is twice higher in CKD^4,5. Further, cardiovascular disease is the main cause of death in people with CKD^2,3. Therefore, risk prediction of cardiovascular disease is essential for prevention and timely management in individuals with CKD.

Recently, the American Heart Association Predicting Risk of Cardiovascular Disease EVENTs (PREVENT) equations have been developed for cardiovascular disease risk prediction^6,7. These equations include several cardiovascular, kidney and metabolic risk factors for the prediction of incident coronary heart disease (CHD), fatal and nonfatal atherosclerotic cardiovascular disease (ASCVD) and total cardiovascular disease (composite of ASCVD and heart failure). In addition, studies incorporating genetic profiles, specifically polygenic risk scores (PRSs), have been proposed for risk prediction of CHD^8–18, to incorporate an individual’s genetic predisposition. PRSs combine small effects of genetic risk variants identified in genome-wide association studies (GWASs) and have been reported to provide additional information to traditional cardiovascular disease risk factors^19,20. However, the generalizability of CHD PRS derived from GWAS performed in the general population to individuals with CKD remains unknown, especially given potential differences in genetic determinants of CHD between CKD patients and the general population²¹.

In this study, we sought to assess whether incorporating CHD PRS improves prediction of CHD among individuals with CKD in addition to the PREVENT equation among participants from the Chronic Renal Insufficiency Cohort (CRIC) study²². We calculated CHD PRS from 11 studies available on the polygenic score catalog, evaluated their performance, selected the best PRS and performed downstream analysis to assess the value of CHD PRS in CRIC participants (Figure 1). We found that CHD PRS developed in general population added to the prediction of CHD beyond the PREVENT equation, leading to better risk classifications. Our results indicate that individuals with CKD can also benefit from CHD PRS in cardiovascular disease risk prevention.

Figure 1. Study Overview.

We first calculated 11 PRSs from the polygenic score catalog³⁹ that were not derived specifically for individuals affected with CKD, validated in our CRIC study participants, and selected the best performed PRS for downstream analyses. We aim to ask the question whether incorporating PRS into traditional risk prediction models (the PREVENT equation) will improve prediction of 10-year CHD risk in this population. Our results demonstrated better risk classification using the PRS+PREVENT equation. PRS: polygenic risk score; CHD: coronary heart disease; PREVENT: American Heart Association’s Predicting Risk of Cardiovascular Disease EVENTs.

Methods

Data Availability

CRIC study data are available for request at the NIDDK Central Repository (NIDDK-CR) website, Resources for Research (R4R), https://repository.niddk.nih.gov.

Ethics statement

All CRIC participants provided written informed consent, and the study was approved by the institutional review boards for each of the participating clinical centers.

Study population

CRIC (Chronic Renal Insufficiency Cohort) is an ongoing, prospective cohort study of racially and ethnically diverse individuals with chronic kidney disease (CKD) drawn from seven academic clinical centers across the United States with the aim of identifying novel predictors of CKD progression and the epidemiology of cardiovascular disease in this population. Between 2003 and 2008, 3939 adults attending a study site for renal care were recruited and the study design and participant characteristics have previously been described^22,23. Briefly, eligible individuals were aged 21–74 years with impaired renal function diagnosed as an estimated glomerular filtration rate (eGFR) 20–70 ml/min per 1.73 m². Age-related entry criteria for eGFR were established to limit the proportion of older individuals with age-related declines in eGFR. Individuals with one or more of the following conditions were excluded: polycystic kidney disease, on active immunosuppression (in the previous 6 months) for glomerulonephritis, pregnant women, previous dialysis for 1 month or more, previous transplant, Human Immunodeficiency Virus infection, known cirrhosis, New York Heart Association class III-IV heart failure, multiple myeloma or renal cancer, and institutionalized care. Participants were invited to annual visits, where patient-reported medical histories were obtained and samples for blood and urine were collected alongside other visit-specific measures.

Genotypes, Imputation Methods and Quality Control

CRIC participants were genotyped using the Illumina HumanOmni 1-Quad Array Platform and were called using Illumina’s BeadStudio Genotyping Module (Illumina Inc, San Diego, CA, USA)²⁴. Quality control (QC) at both SNP (single nucleotide polymorphism) level and sample level was performed separately. At variant level, we removed variants with poor clustering or replication failures, low completeness (< 5%) and a minor allele count (MAC) ≤ 1. The final genotyped dataset included 918,638 variants. At sample-level, we removed samples with discordance and duplicate study IDs.

We performed genotype imputation using the Trans-Omics for Precision Medicine (TOPMed) freeze 8 reference panel on the TOPMed imputation server^25,26, with Eagle2²⁷ for phasing and minimac 4²⁸ for imputation, following our previous procedures^29–31.

We used HARE (harmonized ancestry and race/ethnicity)³², a machine learning algorithm that produced a surrogate variable of inferred ancestry by separating multi-ethnic cohorts into mutually exclusive groups based on both self-identified race/ethnicity and genetically inferred ancestry. Individuals for whom HARE inferred ancestry was undefined were excluded (Figure S1). We identified two sample clusters based on the inferred ancestry groups: EUR (European inferred-ancestry) and AFR (non-European/predominantly African inferred-ancestry). We recalculated minor allele frequency (MAF) within the two groups.

As prior studies have demonstrated that imputation quality differs across population groups^30,33, we re-calculated estimated imputation quality (Rsq) for each population group. Following our previous post-imputation QC strategy³⁴, we removed variants with Rsq < 0.3 for common variants (MAF > 0.01), and variants with Rsq < 0.8 for low frequency or rare variants (MAF < 0.01). We also calculated genotype principal components (PCs) jointly across both EUR and AFR individuals with standard procedures³⁵ using plink 1.9^36,37.

Outcome assessment

The primary outcome of interest was the composite of incident myocardial infarction (MI) and cardiovascular death at 10 years. In brief, we use the short name “CHD” referring to this composite event in line with the definition used in the PREVENT equations.

In CRIC, outcome ascertainment occurred through annual clinic visits and telephone interviews every 6 months^22,23,38. Following self- or proxy-report, major cardiovascular events underwent central, blinded physician-adjudication using medical records. Myocardial infarction was defined as a typical rise and fall in cardiac enzymes combined with symptoms or other signs of myocardial ischemia, or ECG compatible with ischemia or infarction. Alternatively, MI was defined as a new fixed perfusion abnormality with associated wall motion or electrocardiographic changes.

Deaths in CRIC were identified through next of kin, retrieval of death certificates, hospital records and periodic searches of the National Death Index for all individuals considered lost-to-follow-up. In-hospital deaths were adjudicated by two physicians and classified as a cardiovascular event (heart failure, myocardial infarction, cerebrovascular event, arrhythmia or other cardiovascular cause)³⁸. Deaths not adjudicated were classified as cardiovascular-related based on causes of death identified on the death certificate using a super-learning algorithm relying on adjudicated death events as the gold standard³⁸.

Follow-up time started at the participant’s study enrollment visit and was defined as time to the composite event, censoring for non-cardiovascular or undefined death, loss to follow-up or end of administrative follow-up.

Polygenic risk score (PRS) calculation and evaluation

We downloaded 11 PRS weights from the Polygenic Score Catalog³⁹ developed by various studies for cardiovascular events^8–18 (Table S1). After calculating these PRSs in CRIC, we fitted Cox proportional hazard models separately for each PRS, with the composite 10-year CHD event as the outcome, separately for each ancestry group. We included age, sex, eGFR calculated with the 2021 CKD-EPI race-free equation⁴⁰, and the first 10 genetic PCs as covariates. We assessed each PRS performance using hazard ratio and its 95% confidence intervals (CIs), model C-index and incremental R² in liability scale (Table S1, Figure S2). We selected the best-performing model based on the highest C-index and the largest incremental R² in liability scale in both EUR and AFR populations, developed by Patel et al.⁹, which was then used for all the downstream analyses.

PREVENT equation

We also assessed the additional value of PRS compared to existing risk prediction model, using the recently developed sex-specific, race-free American Heart Association’s PREVENT equations among US adults aged 30–79^6,7. We estimated the 10-year risk of the composite CHD events for each CRIC individual without stroke, myocardial infarction/coronary revascularization or heart failure at baseline, with all the available risk factor data that met criteria for calculation of PREVENT 10-year risk of coronary heart disease (composite of myocardial infarction and cardiovascular death). Of note, unlike previous cardiovascular risk equations, the PREVENT equations incorporate eGFR alongside traditional risk factors, including smoking status, systolic blood pressure, total cholesterol, use of antihypertensive medication or statins, and diabetes status. Such scores are also sex-specific and race-free. We note that the CRIC study was used for validation but not development of these prediction models.

Statistical analyses

Focusing on 10-year CHD risk defined as a composite of incident MI and cardiovascular death, we calculated 11 PRSs obtained from the polygenic score catalog³⁹ and selected the one from Patel et al.⁹, which showed the highest R² increment, for downstream analyses. Splitting CRIC samples into 70% training and 30% testing, we fitted three Cox proportional hazard models in training samples using (1) only the PRS, (2) base PREVENT linear predictor (combined cardiovascular disease risk factors and eGFR), and (3) the combination of both, and applied such models to the 30% testing CRIC samples to assess if the PRS contributed to prediction when accounting for the PREVENT CHD linear predictor. For models including PRS as predictors (model 1 and 3), we also adjusted the first 10 PCs as covariates. We calculated the area under the receiver operating characteristic curve (AUC), compared distributions of predicted values separately for individuals with and without 10-year CHD events, and assessed the clinical impact of using the combined model by investigating individuals with different risk classification categories.

Sensitivity analyses

We performed sensitivity analyses for ASCVD outcome, combining our defined CHD events with stroke. We switched to the PREVENT linear predictors for ASCVD and followed the same evaluation procedures for our primary analysis. We further performed another sensitivity analysis evaluating the performance of the three prediction models (PRS only, PREVENT only, PRS+PREVENT) for CHD outcomes, separately in EUR and AFR populations, using AUROC.

Results

Cohort Characteristics

We included 2,775 individuals in the analyses, after removing those with missing genotype or phenotype data, withdrawal of consent, and prior MI or coronary artery revascularization (Figure S1). The mean age was 56.4 with 1313 (47.3%) females (Table 1). 1196 (43.1%) participants self-identified as non-Hispanic white, 1162 (41.9%) as non-Hispanic Black, 322 (11.6%) as Hispanic, and 95 (3.4%) as others. 1342 (48.4%) had diabetes and hypertension at baseline. At 10-years of follow-up, 381 (13.7%) participants experienced the primary event of CHD (Table 1).

Table 1.

Cohort characteristics of CRIC participants.

Variable	Number of participants (%)
age
Mean (SD)	56.4 (11.3)
Median (range)	58 (21 – 75)
sex
Female	1313 (47.3)
Male	1462 (52.7)
Self-reported race/ethnicity
Non-Hispanic White	1196 (43.1)
Non-Hispanic Black	1162 (41.9)
Hispanic	322 (11.6)
Other	95 (3.4)
Diabetes at baseline
Yes	1342 (48.4)
No	1433 (51.6)
10-year CHD event
Yes	381 (13.7)
No	2394 (86.3)

SD: standard deviation; CHD: coronary heart disease.

Selection of best PRS derived from general populations

We then calculated 11 PRSs obtained from the polygenic score catalog³⁹ and compared their performance in CRIC in terms of the R² increment in liability scale. We selected a PRS showing the highest R² increment (Figure S2, Table S1), a multi-ancestry and multi-phenotype PRS developed by Patel et al.⁹, for all the downstream analyses. This CHD PRS achieved 4.3% and 1.2% R² increment in liability scale and had hazard ratios of 1.48 (95% CI [1.26, 1.74]) and 1.23 (95% CI [1.04, 1.46]) per each standard deviation increase in PRS for CRIC EUR and AFR participants (Figure S2, Table S1), respectively. In addition, quintiles of PRS were able to stratify individuals for their 10-year risk of CHD (Figure S3), suggesting good discrimination ability of the PRS model.

We next investigated whether PRS could provide additional information for CHD risk prediction in addition to the PREVENT CHD model^6,7. We calculated the 10-year CHD risk using the PREVENT CHD equations for 1834 individuals without stroke or heart failure at baseline, clinical and biochemical parameters in line with the use of the PREVENT equations and no missing risk factor data. We then fitted three Cox proportional hazard models with different predictors: (1) PRS only, (2) PREVENT linear predictors, and (3) PRS and PREVENT linear predictors. We compared the C-index for the three models, which were 59.8% (se = 1.82%), 73.5% (se = 1.55%), and 74.3% (se = 1.53%) (Table S2), respectively. In the PRS+PREVENT model, the PRS remained a significant predictor with hazard ratio of 1.21 (95% CI: [1.06, 1.39], two-sided Wald test p-value = 5.5e-3), (Table S3) and, therefore, an independent predictor beyond the known risk factors.

PRS+PREVENT equation shows advantages in testing set

To ensure an unbiased evaluation, we split the 1834 individuals into 70% training set for model fitting and applied such models to the remaining 30% for prediction (testing set). We first compared the model performance using AUC, with the 10-year CHD event as the outcome. We observed that adding PRS to the PREVENT model improved AUC from 70.8% (95% CI: [0.64, 0.77]) to 72.4% (95% CI: [0.66, 0.79]) (Figure 2a, Table S4). In addition, we compared the distribution of linear predictors from the PREVENT model and the PRS+PREVENT model, separately for individuals with and without 10-year CHD events. We found that linear predictors from the PRS+PREVENT model showed higher mean and median for CHD event group (two-sided paired t-test p-value = 6.6e-3) while similar values for non-event group (two-sided paired t-test p-value = 0.66) (Figure 2b, Table S5), indicating the combined model demonstrated better discrimination ability than the PREVENT model only. We also evaluated the performance of these models separately for EUR and AFR populations, and the conclusion holds in both populations (Figures S4–5).

Figure 2. PRS improves risk prediction in addition to the PREVENT equation in CRIC.

A. Receiver Operating Characteristic Curves for three models in independent testing cohort. We fitted three Cox proportional hazard models using (1) PRS only, (2) PREVENT linear predictors, and (3) PRS and PREVENT linear predictors in randomly selected 70% training participants. The three models were then applied to the rest 30% testing individuals for evaluation. Outcomes are 10-year CHD events. PRS: polygenic risk score; CHD: coronary heart disease; PREVENT: American Heart Association’s Predicting Risk of Cardiovascular Disease EVENTs; AUC: area under the curve. B. Distribution of risk prediction from PRS, PREVENT and PREVENT+PRS models in testing individuals, separately for participants with and without 10-year CHD events. The combined model showed higher mean and median for individuals with 10-year CHD while similar values for individuals without 10-year CHD event. C. Risk quartile transition when switching from PRS to PREVENT and to PRS+PREVENT model. We classified testing individuals into different categories according to the quartiles of predicted risks from PRS, PREVENT and PRS+PREVENT models. Color represents individuals in different quartiles of the PREVENT model. Q4 represents the highest risk, and Q1 represents the lowest risk.

PRS+PREVENT equation better classifies individuals

Beyond the improved discrimination, we also assessed potential impact by investigating whether individuals would be classified into different categories when switching from PREVENT to the PRS+PREVENT model. We divided the 30% testing individuals into different categories according to the quartiles of linear predictors from the PREVENT and the PRS+PREVENT models, and plotted the state transition (Figure 2c), with increased quartiles (i.e. Q1 to Q4) indicating increased risk. We observed that all the transmitted individuals are towards the correct direction (Table 2). For example, 16 individuals were classified in Q3 using PREVENT equation but in Q4 using PRS+PREVENT, and they showed substantially higher 10-year CHD cumulative incidence (53.9%) compared to agreed Q3 (17.8%) and even agreed Q4 (35.5%) groups. Similarly, 19 individuals originally in PREVENT Q3 were re-classified in PRS+PREVENT to Q2, with a similar cumulative incidence (6.3%) compared to agreed Q2 (7.6%) rather than agreed Q3 (17.8%).

Table 2.

Comparison of 10-year CHD event cumulative incidence for individuals with different categories.

PREVENT quartiles	PRS+PREVENT quartiles	N	10-year CHD event cumulative incidence (95% CI) (%)
Q1	Q1	124	4.6 (0.6, 8.4)
Q1	Q2	14	8.3 (0, 22.7)
Q2	Q1	13	8.3 (0, 22.7)
Q2	Q2	104	7.6 (2.0, 12.4)
Q2	Q3	21	11.8 (0, 25.2)
Q3	Q1	1	NA
Q3	Q2	19	6.3 (0, 17.4)
Q3	Q3	101	17.8 (8.0, 24.1)
Q3	Q4	16	53.9 (9.6, 64.3)
Q4	Q2	1	NA
Q4	Q3	15	14.8 (0, 30.8)
Q4	Q4	122	35.5 (20, 38.8)

We focus on individuals with state transition, defined as that the quartiles from PREVENT and PRS+PREVENT models are different. N: sample size; CHD: coronary heart disease. Cumulative incidence was estimated using Kapler-Meier curve accounting for censoring.

Sensitivity analyses of ASCVD outcome

We performed sensitivity analyses for ASCVD outcome, combining our defined CHD events with stroke. First, we evaluated the added value of CHD PRS compared to the PREVENT equation for the 10-year ASCVD outcome on independent testing samples, with the same 7:3 sample split. We observed that adding CHD PRS to the PREVENT model improved AUC from 70.6% (95% CI: [0.64, 0.77]) to 72.4% (95% CI: [0.66, 0.79]) (Figure S6). We then similarly evaluated individual re-classification based on risk quartiles if using the CHD PRS + PREVENT equation for the 10-year ASCVD outcome. We observed similar pattern that adding CHD PRS has the potential to improve clinical decision (Table S6). For example, 23 individuals were classified in Q3 using PREVENT equation but in Q4 using PRS+PREVENT, and they showed substantially higher 10-year ASCVD cumulative incidence (37.5%) compared to agreed Q3 (15.7%) and even agreed Q4 (33.4%) groups. Similarly, 22 individuals originally in PREVENT Q4 were re-classified in PRS+PREVENT to Q3, with a much smaller cumulative incidence (9.3%) compared to agreed Q4 (33.4%), even smaller than agreed Q3 (15.7%).

Discussion

In this study, we applied CHD PRSs developed from general populations to individuals with CKD and validated their performance in this population. Leveraging the CHD PRS with the highest prediction power (from Patel et al.⁹), we compared the prediction accuracy, model discrimination and classification adding PRS to the existing PREVENT equations to assess the 10-year CHD risk. Our results indicated the additional value of PRS in CRIC participants, demonstrating higher AUC, better discrimination and improved clinical risk stratification. We also performed a sensitivity analysis for the ASCVD outcome (including stroke), where adding CHD PRS to the PREVENT predictor similarly showed improved results (Figure S4, Table S6). Note that we have excluded individuals with prevalent stroke at the baseline visit (Figure S1), and the sample size of individuals who had an incident stroke was very small (<10). Thus, the sensitivity analysis results on ASCVD adding stroke outcomes were very similar to the primary analysis using only CHD outcomes. Nevertheless, these findings support the potential utility of PRS for risk prediction in individuals with CKD, which has not been previously evaluated. However, additional research is needed to test PRS developed in CKD patients and estimate the cost-effectiveness of applying PRS to risk prediction before their use can be recommended for clinical care.

Our study used comprehensive genetic and clinical data from the CRIC study that recruited individuals with clinical CKD and collected prospective data on CHD and ASCVD events. CRIC is a unique cohort to study risk factors for ASCVD in people with CKD and test risk prediction models for incident events. We also found that PREVENT equation performs slightly worse in AFR individuals compared to EUR (AUC 64.8% versus 69.2%, Figures S4–5), which reinforces the potential health disparity issues in AFR populations^41–43. Note that our study is limited in sample sizes, thus our primary analysis focused on the combined individuals regardless of ancestry but we also performed a sensitivity analysis to evaluate model performance separately in EUR and AFR populations. Other limitations include that we did not re-fit models jointly using PREVENT risk factors and PRS, and the observed improvement in the AUC when adding PRS to the PREVENT predictor was small. We note that our study serves as a proof of concept for potential applications of PRS for risk prediction in people with CKD, that will need to be further validated in the context of clinical care for risk prediction. Finally, future studies are warranted to assess the potential of constructing new PRSs using CKD-specific genetic profiles of non-traditional risk factors for cardiovascular disease.

In summary, our study demonstrated a potential application of PRS in CHD risk prediction for individuals with CKD. It also highlights the applications of PRSs generated in the general population to CHD risk in people with CKD. As genetic data becomes cheaper to generate and more widely accessible, more research is needed to address the utility of PRSs in different clinical settings including that among individuals with CKD, who have a high risk of CHD and ASCVD. This research may have potential applications for clinical decisions.

Supplementary Material

Tables S1–S6

Figures S1–S6

Nonstandard Abbreviations and Acronyms

CKD

chronic kidney disease

MI

myocardial infarction

PREVENT

Predicting Risk of Cardiovascular Disease EVENTs

CHD

coronary heart disease

ASCVD

atherosclerotic cardiovascular disease

PRS

polygenic risk score

GWAS

genome-wide association study

CRIC

Chronic Renal Insufficiency Cohort

eGFR

estimated glomerular filtration rate

MAC

minor allele count

TOPMed

Trans-Omics for Precision Medicine

MAF

minor allele frequency

EUR

European

AFR

African

PCs

principal components

AUC

area under the receiver operating characteristic curve

Appendix:

CRIC Study Investigators are Amanda H Anderson PhD, MPH; Lawrence J. Appel, MD, MPH; Jing Chen, MD, MMSc, MSc; Debbie L Cohen, MD; Laura M Dember, MD; James P. Lash, MD; Vallabh O Shah, PhD, MS; Mark L. Unruh, MD, MS.