Single-Ancestry versus Multi-Ancestry Polygenic Risk Scores for CKD in Black American Populations

Visual Abstract

Abstract

Key Points

• The predictive performance of an African ancestry–specific polygenic risk score (PRS) was comparable to a European ancestry–derived PRS for kidney traits.

• However, multi-ancestry PRSs outperform single-ancestry PRSs in Black American populations.

• Predictive accuracy of PRSs for CKD was improved with the use of race-free eGFR.

Background

CKD is a risk factor of cardiovascular disease and early death. Recently, polygenic risk scores (PRSs) have been developed to quantify risk for CKD. However, African ancestry populations are underrepresented in both CKD genetic studies and PRS development overall. Moreover, European ancestry–derived PRSs demonstrate diminished predictive performance in African ancestry populations.

Methods

This study aimed to develop a PRS for CKD in Black American populations. We obtained score weights from a meta-analysis of genome-wide association studies for eGFR in the Million Veteran Program and Reasons for Geographic and Racial Differences in Stroke Study to develop an eGFR PRS. We optimized the PRS risk model in a cohort of participants from the Hypertension Genetic Epidemiology Network. Validation was performed in subsets of Black participants of the Trans-Omics in Precision Medicine Consortium and Genetics of Hypertension Associated Treatment Study.

Results

The prevalence of CKD—defined as stage 3 or higher—was associated with the PRS as a continuous predictor (odds ratio [95% confidence interval]: 1.35 [1.08 to 1.68]) and in a threshold-dependent manner. Furthermore, including APOL1 risk status—a putative variant for CKD with higher prevalence among those of sub-Saharan African descent—improved the score's accuracy. PRS associations were robust to sensitivity analyses accounting for traditional CKD risk factors, as well as CKD classification based on prior eGFR equations. Compared with previously published PRS, the predictive performance of our PRS was comparable with a European ancestry–derived PRS for kidney traits. However, single-ancestry PRSs were less predictive than multi-ancestry–derived PRSs.

Conclusions

In this study, we developed a PRS that was significantly associated with CKD with improved predictive accuracy when including APOL1 risk status. However, PRS generated from multi-ancestry populations outperformed single-ancestry PRS in our study.

Introduction

CKD is an independent risk factor of cardiovascular disease and affects approximately one in seven adults in the United States.¹ Given that CKD often progresses asymptomatically until late-stage disease, it is important to identify at-risk individuals for earlier intervention. However, current risk algorithms—for example, Kidney Failure Risk Equation—primarily focus on progression to kidney failure among those who already have CKD.^2–5 In addition, these clinical algorithms do not consider how one's genetic background may contribute to disease risk.

Heritability estimates of CKD range from 30% to 75%.^6–8 Genome-wide association studies (GWASs) demonstrate that CKD is a polygenic disease.^9–22 Thus, polygenic risk scores (PRSs)—a method to quantify genome-wide risk for disease—may improve the utility of genomic information to guide targeted interventions for high-risk (HR) individuals. Recently, PRSs have been shown to capture risk equivalent to monogenic mutations but at higher population prevalence.²³ As such, application of PRS developed for kidney traits may aid in prevention of not only kidney failure but also severe CKD.

One of the current limitations of the field, however, is that PRSs—often derived in European populations—have diminished predictive accuracy in African and admixed ancestry populations, primarily due to differences in linkage disequilibrium patterns and allele frequencies.^24,25 While multi-ancestry approaches to PRS development are increasing, African ancestry representation is lagging, comprising <3% of training datasets.²⁵ Furthermore, most PRSs for CKD to date have been developed using race-based estimates of kidney function.^{12,14,26–35} Previous eGFR equations have included coefficients that increased eGFR by 8%–16% for Black race and have been demonstrated to underestimate kidney disease severity in these individuals.^36–39

The disproportionate burden of CKD, lower accuracy of current PRS in African ancestry populations, and misclassification bias invoked by race-based equations are likely to exacerbate existing disparities in CKD prevention and management.^40–42 More studies are needed in African ancestry populations to prevent bias in clinical implementation of PRSs. Such work can potentially close a gap in genomic risk prediction and improve predictive accuracy using race-free estimators of kidney function. The objective of this study was to leverage genetic data from cohorts of Black American patients to develop a PRS for CKD, defined by the updated eGFR equations, and compare its performance with previously published scores.^{12,14,26,28–30,33}

Methods

Study Populations and CKD Phenotyping

The populations used for this study are summarized in Figure 1. Complete descriptions of enrollment, genotyping, and quality control procedures for each cohort are available in the Supplemental Methods. Briefly, selection of these cohorts was chiefly based on availability of African ancestry genomic data and kidney phenotypic data. Prevalent CKD status was defined as eGFR <60 ml/min per 1.73 m², corresponding to stage 3 or higher disease, according to the National Kidney Foundation's Kidney Disease Improving Global Outcomes (KDIGO) guidelines.⁴³ Controls were defined as individuals without a history of kidney transplant or dialysis and eGFR >60 ml/min per 1.73 m². eGFR was calculated from serum creatinine and/or cystatin C values measured at a baseline visit for the observational cohort studies or the closest date to study enrollment available in the electronic health record (her). When possible, we used the 2021 equations fit without race coefficients, and eGFR equation (creatinine only or combined creatinine-cystatin C) is specified for each cohort below.³⁸

Figure 1.

Study overview. CHS, Cardiovascular Health Study; GWAS, genome-wide association study; HyperGEN, Hypertension Genetic Epidemiology Network; MESA, Multi-Ethnic Study of Atherosclerosis; MVP, Million Veteran Program; PRS, polygenic risk score; REGARDS, Reasons for Geographic and Racial Differences in Stroke; TOPMed, Trans-Omics in Precision Medicine; WGS, whole genome sequence.

Summary Statistics

We obtained GWAS summary statistics for eGFR from 57,336 Black participants of the Million Veteran Program (MVP), which is one of the largest summary statistics on kidney function in an African ancestry population to date.¹⁷ eGFR was computed using the 2009 Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) creatinine equation (2009 eGFRcr), which includes a race coefficient. To increase summary statistics sample size, we also conducted GWAS for eGFR in a subset of 8916 participants in the Reasons for Geographic and Racial Differences in Stroke (REGARDS) study.⁴⁴ eGFR was computed using the 2021 CKD-EPI creatinine-cystatin C equation (2021 eGFRcr-cys) without a race coefficient. We meta-analyzed 13.8 million overlapping autosomal single-nucleotide polymorphisms (SNPs) for a total of 65,957 participants in METASOFT under the random-effects model to account for between-study heterogeneity.⁴⁵ Notably, this method prevents overly conservative P value estimation under the original random-effects model.

PRS Development

PRSs were developed using the PRS-CS software, which applies a Bayesian regression approach and continuous shrinkage priors on SNP effect sizes.⁴⁶ Precomputed linkage disequilibrium blocks are informed by a reference panel (1000 Genomes African, phase 3). Variant selection for PRS construction is exclusive to nonambiguous HapMap3 variants, which allows for computationally efficient calculation of posterior SNP effect sizes within these blocks. In our pipeline, we applied the meta-analyzed eGFR summary statistics, which resulted in a final PRS of 1,142,436 SNPs.

Risk Model Optimization

We then assessed predictive performance of the PRS at each shrinkage prior (φ)—1, 1e-02, 1e-04, and 1e-06—and optimized model parameters in a subset of 1862 participants in the Hypertension Genetic Epidemiology Network (HyperGEN).⁴⁷ CKD status was determined from the 2021 CKD-EPI creatinine equation without a race coefficient (2021 eGFRcr). We calculated PRS for each φ set of SNP weights in PLINK (version 1.9—“--score” function).⁴⁸ Because SNP weights were based on eGFR summary statistics, we multiplied PRS by −1 to reflect a higher PRS in association with lower eGFR. For each φ, we assessed the distribution of the PRS between cases and controls and fit logistic regression models in R (version 4.1.2, glm function) for CKD (outcome) and the PRS (predictor). PRS was considered as a standardized continuous predictor and as a categorical predictor at the top 3%, 5%, 10%, 20%, and 50% thresholds. Models were adjusted for age, sex, diabetes status, and the first ten ancestry principal components; these variables are defined in the Supplemental Methods.

We selected the optimal φ based on the liability R²—the explained variance of the outcome by the predictor, accounting for differences between the cohort-level and population-level prevalence of disease—as well as the area under the curve (AUC).⁴⁹ AUC was calculated for the fully adjusted, covariate-only, and PRS-only models (pROC R package).⁵⁰ Additional metrics to assess predictive performance of PRSs included sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Both PPV and NPV were adjusted for population prevalence of CKD (aPPV, aNPV). Population prevalence of CKD was obtained from the United States Renal Data System.⁴⁰

We also evaluated potential misclassification of CKD by excluding participants from the control group with mildly decreased kidney function (60 ≥eGFR <90, hereafter referred to as G2s in accordance with KDIGO staging).⁴³ In another sensitivity analysis, we adjusted for prevalent hypertension, obesity, and smoking, which, in addition to diabetes, comprise the primary risk factors of CKD. We then compared the distribution of the PRS, stratified by each risk factor, using the independent t test, as well as assessed multiplicative interaction for each risk factor.

Finally, we evaluated effect modification of APOL1 risk genotypes, which have previously been shown to have an additive effect to PRS.²⁸ APOL1 SNPs (rs73885319, rs60910145, and rs143830837) were excluded from PRS construction, and participants who were homozygous for G1 or G2 or had one G1 and one G2 recessive allele were categorized as HR. Carriers of one or neither allele were defined as low risk (LR). We fit logistic regression models for standardized PRS, APOL1 genotype, and a multiplicative interaction term between PRS and APOL1 risk status.

PRS Evaluation

On selection of the optimal φ, model covariates, and APOL1 weight, we evaluated PRS performance in a subset of 6032 participants from the Cardiovascular Health Study, Jackson Heart Study, Multi-Ethnic Study of Atherosclerosis (MESA), and Women's Health Initiative, accessed through the Trans-Omics in Precision Medicine (TOPMed) Consortium.^51–55 We computed PRS in PLINK and calculated eGFR (2021 eGFRcr) to ascertain CKD case status. Logistic regression models were adjusted for age, sex, diabetes, study, and the first ten ancestry principal components. In the final model, we excluded G2s and estimated odds ratios (ORs) for prevalent CKD (outcome) and the APOL1-weighted PRS (predictor). We assessed the PRS as a continuous predictor and at the percentile thresholds.

To compare performance of our optimized PRS with previous studies, we selected PRS for kidney traits (eGFR, CKD, kidney failure) from the polygenic score catalog and applied them to prevalent CKD as an outcome.⁵⁶ Because multiple PRSs had been derived from pipelines that included TOPMed cohorts, we sought to avoid model overfitting in the setting of data overlap. Thus, we leveraged data from 6641 Black participants in the Genetics of Hypertension Associated Treatment (GenHAT) study.⁵⁷ We computed PRS in PLINK and fit logistic regression models, adjusted for the same covariates as above, in addition to trial randomization group. In sensitivity analyses, we compared performance of the risk model when CKD was defined by the 2009 eGFRcr equation and separately among APOL1 risk strata.

Results

Baseline characteristics of all cohorts are summarized in Table 1. The mean age of each cohort ranged from 47 to 72 years (SD 5–13 years). Approximately 85% of participants in the meta-analysis were male, with the primary contribution from MVP (92% male). The prevalence of diabetes ranged from 21% to 60%, and the prevalence of CKD varied between 5% and 17%. Manhattan and quantile–quantile plots for the random-effects meta-analysis are presented in Figure 2. We did not observe any novel SNP associations in the meta-analysis results.

Table 1.

Baseline characteristics of study cohorts

Trait Mean (SD)/N (%)	GWAS Summary Statistics		Development	Evaluation
				TOPMed Cohorts
	MVP (18) (57,336)	REGARDS (8621)	HyperGEN (1862)	CHS (682)	JHS (2919)	MESA (1097)	WHI (1334)	GenHAT (6641)
Age, yr	58 (12)	64 (9)	47 (12)	73 (5)	54 (13)	61 (10)	64 (7)	66 (8)
Male, n (%)	52,670 (92)	3395 (39)	680 (37)	251 (37)	1106 (38)	515 (47)	0 (0)	2984 (45)
Diabetes, n (%)	20,967 (37)	2496 (29)	395 (21)	167 (24)	630 (22)	338 (31)	274 (21)	3963 (60)
CKD, n (%)	—	1143 (13)	103 (6)	104 (15)	149 (5)	62 (6)	34 (3)	1134 (17)
eGFR, ml/min per 1.73 m2	86 (23)	92 (31)	103 (30)	88 (29)	107 (33)	96 (26)	120 (31)	85 (27)
APOL1 HR, n (%)	—	1050 (12)	294 (16)	86 (13)	417 (14)	129 (12)	155 (12)	871 (13)

eGFR equation by cohort: 2009 eGFRcr: Million Veteran Program; 2021 eGFRcr-cys: Reasons for Geographic and Racial Differences in Stroke; 2021 eGFRcr: Hypertension Genetic Epidemiology Network, Cardiovascular Health Study, Jackson Heart Study, Multi-Ethnic Study of Atherosclerosis, Women's Health Initiative, Genetics of Hypertension Associated Treatment. “—”, not reported; CHS, Cardiovascular Health Study; GenHAT, Genetics of Hypertension Associated Treatment; GWAS, genome-wide association study; HR, high-risk; HyperGEN, Hypertension Genetic Epidemiology Network; JHS, Jackson Heart Study; MESA, Multi-Ethnic Study of Atherosclerosis; MVP, Million Veteran Program; REGARDS, Reasons for Geographic and Racial Differences in Stroke; TOPMed, Trans-Omics in Precision Medicine; WHI, Women's Health Initiative.

Figure 2.

Manhattan and quantile–quantile plots for eGFR meta-analysis.

In HyperGEN, the prevalence of CKD was 6%. ORs for CKD for each φ, with and without G2s, are presented in Supplemental Table 1. The optimal shrinkage parameter was φ=1, with a liability R² of 7% and AUC_PRS of 58%. In the unadjusted model, a 1 SD higher PRS was associated with a 36% greater odds of CKD (95% confidence interval [CI], 1.12 to 1.66); adjusted model estimates were similar (OR [95% CI]: 1.35 [1.08 to 1.68]). Furthermore, exclusion of G2s slightly improved prediction (AUC_PRS=59%).

The independent effect of APOL1 was approximately 1.6 in a model that included the main and interaction effects of the PRS and APOL1 (Supplemental Table 2). There was no evidence of multiplicative interaction (p_int=0.84). Thus, we applied weighting of PRS+1.6 for HR individuals: additive weighting of APOL1 improved effect estimates (OR [95% CI]: 1.41 [1.14 to 1.76]) and PRS accuracy (R²=10%, AUC_PRS=60%), summarized in Table 2.

Table 2.

Comparison of unweighted and APOL1-weighted polygenic risk scores in Hypertension Genetic Epidemiology Network

HyperGEN	Cases/Controls	Raw PRS			APOL1-Weighted PRS
		OR (95% CI)	Liability R2	AUC (Crude)	OR (95% CI)	Liability R2	AUC (Crude)
Full cohort	103/1759	1.35 (1.08 to 1.68)	0.07	0.83 (0.58)	1.35 (1.12 to 1.63)	0.10	0.84 (0.59)
No G2s	103/1218	1.28 (1.00 to 1.65)	0.04	0.89 (0.59)	1.41 (1.14 to 1.76)a	0.10a	0.90 (0.60)a

Odds ratios and 95% confidence intervals for CKD are presented for a 1 SD higher polygenic risk score. Logistic regression models were adjusted for age, sex, diabetes, and PC1–PC10. G2s (60 ≤eGFR <90) were excluded. AUC, area under the curve; CI, confidence interval; HyperGEN, Hypertension Genetic Epidemiology Network; OR, odds ratio; PRS, polygenic risk score.

^aIndicates the optimized risk model.

Moreover, associations were robust in sensitivity models adjusting for CKD risk factors (Supplemental Table 3). There was a modestly significant difference in the PRS by hypertension status (P = 0.04) in an independent t test. There was also evidence of multiplicative interaction between body mass index and PRS (standardized p_int=0.02, APOL1-weighted p_int=0.04). However, these effects were attenuated on exclusion of G2s.

Thus, we applied the optimized, APOL1-weighted PRS in the TOPMed validation cohort, excluding G2s. The prevalence of CKD was 8%, and 13% of participants had APOL1 HR genotypes. After quality control filters, 97% of PRS SNPs were available in TOPMed. Not only was the mean PRS higher among CKD cases compared with controls but we also observed a PRS quintile-dependent higher odds of CKD compared with those in the lowest quintile of PRS (Figure 3, A–C). Furthermore, when considered as both a continuous predictor and at top thresholds, higher PRS was consistently associated with greater odds of CKD, as presented in Table 3. These associations, however, were not statistically significant in smaller strata (3%, 5%, and 10%). PRS demonstrated high specificity but low sensitivity, as well as modest predictive accuracy in the unadjusted model (liability R²=3% and AUC_PRS=56%), as shown in Figure 3D. Complete metrics of PRS accuracy are presented in Supplemental Table 4.

Figure 3.

Metrics of PRS performance in TOPMed validation cohort. (A) Histograms showing the distribution of APOL1-weighted PRS, stratified by CKD status. Red: CKD, blue: no CKD. (B) Adjusted OR for APOL1-weighted PRS associations with CKD, stratified by quintile of PRS (Q1 as the reference). (C) Boxplots of PRS distribution between CKD cases and controls. Red: CKD, blue: no CKD. (D) AUC for PRS models. Red: PRS only, blue: covariate only, Black: PRS and covariates. AUC, area under the curve; CI, confidence interval; OR, odds ratio.

Table 3.

Overall and threshold-based performance of polygenic risk score in Trans-Omics in Precision Medicine

PRS Threshold	Cases/Controls	OR (95% CI)	Liability R2	AUC (Crude)
1 SD	349/4047	1.21 (1.08 to 1.36)	0.03	0.89 (0.56)
Top 3%	15/117	—	—	—
Top 5%	26/194	6.88 (0.35 to 44.88)	0.006	0.89 (0.50)
Top 10%	52/388	2.27 (0.53 to 7.33)	0.004	0.89 (0.50)
Top 20%	88/791	1.78 (1.11 to 2.80)	0.02	0.89 (0.52)
Top 50%	211/1987	1.61 (1.22 to 2.13)	0.03	0.89 (0.55)

Odds ratios and 95% confidence intervals for CKD are presented for either a 1 SD higher polygenic risk score or top percentile compared with lower percentile of polygenic risk score (e.g., top 5% with bottom 95% as reference). Logistic regression models were adjusted for age, sex, diabetes, PC1–PC10, and study. G2s (60 ≤eGFR <90) were excluded. “—” indicates that model did not converge. AUC, area under the curve for the adjusted model; CI, confidence interval; Crude, area under the curve for the unadjusted model; OR, odds ratio; PRS, polygenic risk score.

The performance of previously published PRS in GenHAT is summarized in Supplemental Table 5. PRSs were developed using various software (e.g., PRS-CS, LDPred, snpnet) and pruning and thresholding methods; GWAS sample sizes used for summary statistics exceeded 250,000. In general, the greatest predictive accuracy arose from multi-ancestry PRS with higher OR, liability R², and AUC. In threshold models, our PRS was comparable with the European PRS by Ma et al.; however, the Ma PRS demonstrated stronger effect sizes and smaller P values in European PRS. European-derived PRS was less precise than our PRS as evidenced by the larger SEM and wider CIs. In continuous models, multi-ancestry PRS performed similarly; however, the Khan PRS demonstrated better risk stratification among the top percentiles (Table 4).²⁸ For example, individuals in the top 3% of the Khan PRS had 3.30 (95% CI, 2.01 to 5.47) greater odds of CKD, whereas those in the top 3% of our PRS had 1.98 (95% CI, 1.20 to 3.28) greater odds of CKD (Table 4). Interestingly, at top thresholds, the Ma PRS (European) outperformed some multi-ancestry PRS.^28,30 Although associations for each PRS were consistent across eGFR equations, modeling CKD defined by the race-free equation yielded more precise estimates and improved AUC_PRS (Supplemental Table 6). In addition, multi-ancestry PRS emerged as the best predictors in both APOL1 HR and APOL1 LR strata as well (Supplemental Table 7).

Table 4.

Predictive performance of selected kidney function polygenic risk scores in Genetics of Hypertension Associated Treatment, among individuals in the top 3% of polygenic risk score

Lead Author	Trait	Method	GWAS, N	GWAS Ancestry	Development Ancestry	nSNPs (% Overlap)	Cases/Controls	OR (95% CI)	R2	AUC (Crude)
Jones	CKD (Prev)	PRS-CS	65,957	AA		1,142,436 (98)	57/52	1.98 (1.20 to 3.28)	0.03	0.87 (0.51)
Khan		P+T	765,348	Multi (1.8%)		471,316 (94)	61/48	3.30 (2.01 to 5.47)	0.04	0.87 (0.52)
Ma	eGFR	PRS-CS	567,460	European		1,117,334 (97)	52/59	2.98 (1.80 to 4.90)	0.009	0.87 (0.51)
Steinbrenner		LDPred	406,350	Multi (1.8%)		1,496,254 (95)	56/55	2.22 (1.35 to 3.65)	0.005	0.87 (0.51)
Wuttke		GWAS	765,348	Multi (1.8%)	European	147 (97)	30/81	1.35 (0.76 to 2.33)	0.001	0.87 (0.50)
Yu		LDPred	1,159,871	Multi (1.5%)		1,477,661 (96)	56/55	2.13 (1.29 to 3.53)	0.004	0.87 (0.51)
Sinnott-Armstrong	ESKD (Prev)	snpnet	376,520	Multi (1.7%)	European	183,272 (92)	41/70	1.51 (0.90 to 2.51)	0.001	0.87 (0.50)
Tanigawa	ESKD (Inc)	snpnet	269,704	European		158 (84)	36/75	2.14 (1.24 to 3.65)	0.004	0.87 (0.50)

In ancestry columns, percentages indicate proportion of African ancestry individuals included in genome-wide association study cohorts. Cases and controls are presented for those in the top 5% of their respective polygenic risk score. Odds ratios and 95% confidence intervals for CKD are presented for individuals in the top 3% of polygenic risk score compared with those in the bottom 97% of polygenic risk score (reference). Logistic regression models adjusted for age, sex, diabetes, PC1–PC10, and randomization group. G2s (60 ≤eGFR <90) were excluded. AA, African American; AUC, area under the curve for the adjusted model; CI, confidence interval; Crude, area under the curve for the unadjusted model; GWAS, genome-wide association study; Inc, incident; nSNPs, number of single-nucleotide polymorphisms; OR, odds ratio; Prev, prevalent; PRS, polygenic risk score; P+T, pruning and thresholding.

Discussion

In this study, we derived a PRS for CKD exclusively using African ancestry data. Although PRS derivation varied widely in methodology (e.g., PRS-CS, LDPred, snpnet), which can make comparison difficult, a multi-ancestry approach outperformed single-ancestry PRS. We leveraged data from multiple cohort studies that recruited participants throughout the United States, across a wide age span (18–99 years), and we optimized PRS according to race-free eGFR equations. We observed good transferability of PRS across eGFR equations; however, modeling CKD using the race-free equations yielded more precise estimates. It is noteworthy that the low sensitivity of these tests limits their clinical utility. However, it is possible that, with further refinement, PRS can be used to guide CKD prevention, similar to BRCA mutations as indications for prophylactic mastectomies or 10-year risk of cardiovascular disease as indication for initiation of statin therapy.⁵⁸ Given the high prevalence of CKD risk factors in the United States, prevention for high–genetic risk individuals may include stricter BP control and glycemic management. Furthermore, among individuals without clinical risk factors, a high PRS may reinforce pursuit of a healthy lifestyle and/or inform family members to evaluate their own risk.

Carriers of two APOL1 recessive alleles—compared with those with one or neither allele—have a higher risk of progressive kidney disease and organ failure.^59–62 Mutations in this locus arose in response to trypanosomiasis, a parasitic infection endemic to sub-Saharan Africa. Thus, individuals descended from this region (many of whom are racialized as Black in the United States) are more likely to carry these risk alleles. Similar to Khan et al., we demonstrated that weighting PRS by APOL1 status improved predictive accuracy of PRS or APOL1 alone.^28,63 The prevalence of APOL1 HR genotypes in our study was representative of the United States (13% to 14%), but additive effects to PRS may vary when prevalence in the development cohorts deviate from the population prevalence.⁶⁴ Although the Khan PRS observed that HR status conferred additional risk equivalent to 1 SD higher PRS among a cohort with a much lower prevalence of APOL1 HR individuals, we observed a 1.6 SD additive effect to our PRS. Inclusion of constants for monogenic mutations, such as APOL1, should be PRS-specific rather than fixed.

Interestingly, a recent study by Vy et al. suggests that polygenicity of CKD may vary by APOL1 risk, in which an African ancestry–derived PRS outperformed a multi-ancestry PRS among APOL1 HR individuals.⁶⁵ In our study, multi-ancestry scores had the best predictive accuracy in both low and high APOL1 risk status. The reason for this difference in our findings is unclear. Still, continued study of the potential for variation in polygenic risk by monogenic risk is needed to better understand the penetrance of APOL1, especially in the context of kidney transplantation. Studies such as All of Us are likely to address these gaps in genomic data with larger, increasingly diverse sample sizes.⁶⁶

Although we demonstrated that multi-ancestry PRS for CKD outperform single-ancestry approaches, even when summary statistics sample sizes are comparable, current representation of African ancestry in these cohorts are still lagging at <2%.^24,25 In addition, there is a lack of continental diversity even within African ancestry populations that contribute to current analyses. Owing to the trans-Atlantic slave trade and forced migration between the 16th and 19th centuries, many individuals of African ancestry in current PRS development are descended from sub-Saharan West Africa,⁶⁷ and the genetic architecture of CKD among individuals descended from other regions of Africa is likely to be even less well characterized.⁶⁸ Given the vast genetic diversity of Africa, greater representation of individuals with ancestry across the continent is still needed. It is possible that, with increasing the representation of these historically excluded groups in CKD genomic studies, predictive accuracy of African ancestry PRS may improve.

Based on our findings, the current literature,^26–30 and available genomic resources, we conclude that data from multiple populations should be used to develop PRS for CKD in Black American populations. In the absence of diverse consortia or biobanks, meta-analyzing GWASs from several studies may help achieve this goal and improve PRS transferability. We note that our single-ancestry PRS was developed in much smaller summary statistics data in comparison with previously published scores (<66,000 versus >250,000), and performance improves by increasing the training data sample size.⁶⁹ More research is needed to determine whether single-ancestry PRS developed in African ancestry datasets of comparable size would outperform available multi-ancestry PRS.

There were multiple limitations in our study that we attempted to redress. Although MVP summary statistics were based on the 2009 eGFR equation, by restricting the analysis to Black American participants, the effect of the race coefficient was essentially null in linear modeling of GWASs. However, the differences in eGFR equations—2009 eGFRcr for MVP and 2021 eGFRcr-cys for REGARDS—for summary statistics may have obfuscated the PRS training. Still, the PRS was optimized (HyperGEN) and validated (TOPMed, GenHAT) using one equation (2021 eGFRcr). Moreover, summary statistics representation was disproportionately male, mostly owing to MVP. We attempted to balance these effects by meta-analyzing with REGARDS (40% male) and optimizing the PRS in HyperGEN (37% male). In sensitivity analyses, we did not observe any effect modification of PRS by sex. As MVP remains one of the largest GWASs of kidney traits in Black American populations, future studies should weigh these nuances in selection of populations for PRS development.

At the PRS evaluation stage, we used harmonized TOPMed data to increase statistical power and generalizability. Although TOPMed data had been incorporated in prior PRS summary statistics—notably those derived from a multi-ancestry GWAS of eGFR by Wuttke et al.—we assessed the performance of these PRS in GenHAT to avoid model overfitting because of data overlap.¹² Although GenHAT is not necessarily as representative of the general population as our TOPMed cohort, associations in GenHAT and TOPMed were comparable (data not shown). Moreover, although we could not distinguish diabetic kidney disease from nondiabetic kidney disease in our data, multiple sensitivity analyses suggest that diabetes was not a significant effect modifier of this PRS.

While self-reported race was used as proxy for ancestry, we assessed the proportion of African ancestry when possible and observed a distribution concordant with previous literature.⁶⁷ As ESKD was an exclusion criterion for multiple cohorts included in this analysis, the reported associations are likely underestimated. In addition, we excluded stages 1 and 2 CKD from the case definition because of inconsistent availability of proteinuria data in our validation cohorts. In the absence of evidence of kidney damage (e.g., proteinuria), neither KDIGO stage G1 (eGFR ≥90) nor G2 fulfills the criteria for CKD based on eGFR alone.⁴³ Still, this exclusion may have reduced generalizability of these findings.

Our study has several strengths. As previously stated, this study is among the first to develop PRS using the updated race-free estimators of kidney function. In addition, we specifically focused on the population that was adversely affected by prior equations. Although some PRSs that were developed during the transition of equations compared performance of 2009 and 2021 equations, our study was optimized and validated without these race coefficients. Moreover, we demonstrated that ancestry-specific approaches are not an adequate solution to current limitations of PRS transferability. Rather, development of multi-ancestry PRS with proportionate ancestry representation is likely to redress these disparities. Once that is accomplished, PRSs may have broad clinical utility for surveillance and prevention of kidney disease.

In conclusion, we developed and validated a PRS for CKD in cohorts of Black American populations using the updated equations for eGFR. Predictive accuracy was improved by inclusion of monogenic risk (i.e., APOL1). Associations were robust to multiple sensitivity analyses, including accounting for traditional risk factors, suggesting that PRSs have additional utility beyond current clinical algorithms, which is limited for CKD. Multi-ancestry approaches outperformed single-ancestry PRS, and application of race-free estimators of kidney function improved PRS accuracy and precision.

Future studies should seek to increase African ancestry in CKD genomics and ensure proportionate representation across the African continent. GWASs and PRSs for kidney traits should also be reevaluated in the context of the new eGFR equations. Finally, PRSs should be evaluated both in comparison with and in conjunction with current clinical risk algorithms, such as the Kidney Failure Risk Equation. A combination of these measures may aid in the prevention of kidney failure.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/E785.

Funding

This research was funded by the National Institutes of Health (NIH) National Heart, Lung, and Blood Institute (NHLBI) grants R35HL155466 (M.R. Irvin) and R01HL136666 (M.R. Irvin, L.A. Lange). A.C. Jones was also supported by the National Institute of Diabetes and Digestive and Kidney Diseases (F31DK128990 and T32DK116672). Additional support included U01HG011167 funded through the National Human Genome Research Institute (NHGRI) and K24HL133373 and R01HL092173 funded through NHLBI (N.A. Limdi).

Author Contributions

Conceptualization: Nicole D. Armstrong, Bertha A. Hidalgo, Marguerite R. Irvin, Alana C. Jones, Nita A. Limdi, Amit Patki, Vinodh Srinivasasainagendra, Hemant K. Tiwari.

Data curation: Nicole D. Armstrong, Donna K. Arnett, Nisha Bansal, Joshua C. Bis, Jennifer A. Brody, Ninad S. Chaudhary, Yii-Der I. Chen, James J. Cimino, Brittney Davis, Ian H. De Boer, Clarissa J. Diamantidis, Nora Franceschini, Marguerite R. Irvin, Holly J. Kramer, Ethan M. Lange, Leslie A. Lange, Nita A. Limdi, Josyf C. Mychaleckyj, Amit Patki, Bruce M. Psaty, Stephen S. Rich, Jerome I. Rotter, Vinodh Srinivasasainagendra, Sylvia Wassertheil-Smoller, Bessie A. Young.

Formal analysis: Nicole D. Armstrong, Ninad S. Chaudhary, Alana C. Jones, Amit Patki, Vinodh Srinivasasainagendra.

Funding acquisition: James J. Cimino, Marguerite R. Irvin, Alana C. Jones.

Methodology: Nicole D. Armstrong, Brittney Davis, Nora Franceschini, Bertha A. Hidalgo, Marguerite R. Irvin, Alana C. Jones, Atlas Khan, Krzysztof Kiryluk, Holly J. Kramer, Nita A. Limdi, Amit Patki, Vinodh Srinivasasainagendra, Hemant K. Tiwari.

Supervision: Bertha A. Hidalgo, Marguerite R. Irvin, Nita A. Limdi, Hemant K. Tiwari.

Writing – original draft: Marguerite R. Irvin, Alana C. Jones.

Writing – review & editing: Nicole D. Armstrong, Donna K. Arnett, Nisha Bansal, Joshua C. Bis, Jennifer A. Brody, Yii-Der I. Chen, Ian H. De Boer, Clarissa J. Diamantidis, Nora Franceschini, Bertha A. Hidalgo, Marguerite R. Irvin, Alana C. Jones, Atlas Khan, Krzysztof Kiryluk, Holly J. Kramer, Ethan M. Lange, Leslie A. Lange, Nita A. Limdi, Josyf C. Mychaleckyj, Amit Patki, Bruce M. Psaty, Stephen S. Rich, Jerome I. Rotter, Vinodh Srinivasasainagendra, Hemant K. Tiwari, Sylvia Wassertheil-Smoller, Bessie A. Young.

Data Sharing Statement

Previously published data were used for this study. TOPMed: doi: 10.1038/s41586-021-03205-y, REGARDS: doi: 10.3389/fgene.2021.781451, MVP: doi: 10.1038/s41467-019-11704-w, GenHAT: doi: 10.3389/fgene.2021.781451. GWAS summary statistics for MVP can be obtained via the National Center for Biotechnology Information (NCBI) Database for Genotypes and Phenotypes (dbGaP) under accession phs001672.v9.p1. The raw genotypic and phenotypic data in REGARDS are also deposited in dbGaP under accession phs002719.v1.p1. Harmonized WGS and phenotypic data for HyperGEN, Cardiovascular Health Study, Jackson Heart Study, Multi-Ethnic Study of Atherosclerosis, and Women's Health Initiative can be accessed upon request from the TOPMed Consortium.

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/JSN/E783, http://links.lww.com/JSN/E784.

Supplemental Methods.

Supplemental Table 1. Effect estimates for each phi set of PRS weights in HyperGEN.

Supplemental Table 2. Assessment of APOL1 effect modification on PRS in HyperGEN.

Supplemental Table 3. Evaluation of interaction of CKD risk factors with PRS in HyperGEN.

Supplemental Table 4. Complete metrics of African American–specific PRS performance in TOPMed.

Supplemental Table 5. Complete metrics of selected kidney function PRS in GenHAT.

Supplemental Table 6. Selected PRS performance by eGFR equation in GenHAT.

Supplemental Table 7. Selected PRS performance by APOL1 risk status in GenHAT.