Exome-Based Rare-Variant Analyses in CKD

Significance Statement

Studies have identified common CKD-associated gene variants, but the contribution of rare variants has not been systematically examined. The authors use exome sequencing and rare-variant collapsing analyses to compare rare genetic variants in 3150 cases (representing broad clinical CKD subtypes) with 9563 controls. For five known CKD-associated genes, they detected a significant enrichment of rare variants in PKD1, PKD2, COL4A5, and found suggestive evidence for rare COL4A3 and COL4A4 variants. They also found evidence for four other genes not previously implicated in CKD. By demonstrating that rare-variant collapsing analyses can validate known genes and identify candidate genes and modifiers for nephropathy, these findings provide a rationale for larger-scale investigation of the rare variants’ contribution to disease risk across major clinical CKD categories.

Visual Abstract

Abstract

Background

Studies have identified many common genetic associations that influence renal function and all-cause CKD, but these explain only a small fraction of variance in these traits. The contribution of rare variants has not been systematically examined.

Methods

We performed exome sequencing of 3150 individuals, who collectively encompassed diverse CKD subtypes, and 9563 controls. To detect causal genes and evaluate the contribution of rare variants we used collapsing analysis, in which we compared the proportion of cases and controls carrying rare variants per gene.

Results

The analyses captured five established monogenic causes of CKD: variants in PKD1, PKD2, and COL4A5 achieved study-wide significance, and we observed suggestive case enrichment for COL4A4 and COL4A3. Beyond known disease-associated genes, collapsing analyses incorporating regional variant intolerance identified suggestive dominant signals in CPT2 and several other candidate genes. Biallelic mutations in CPT2 cause carnitine palmitoyltransferase II deficiency, sometimes associated with rhabdomyolysis and acute renal injury. Genetic modifier analysis among cases with APOL1 risk genotypes identified a suggestive signal in AHDC1, implicated in Xia–Gibbs syndrome, which involves intellectual disability and other features. On the basis of the observed distribution of rare variants, we estimate that a two- to three-fold larger cohort would provide 80% power to implicate new genes for all-cause CKD.

Conclusions

This study demonstrates that rare-variant collapsing analyses can validate known genes and identify candidate genes and modifiers for kidney disease. In so doing, these findings provide a motivation for larger-scale investigation of rare-variant risk contributions across major clinical CKD categories.

CKD affects approximately one in ten adults worldwide.^1–3 Although Mendelian disorders are estimated to be responsible for 10% of ESRD cases,⁴ 15%–30% of patients, including those with presumptively acquired forms of disease such as diabetic or hypertensive nephropathy, report a family history of renal disease, indicating a contribution of hereditary factors across clinical categories.^5,6 Moreover, affected individuals within the same family frequently carry different clinical diagnoses, suggesting that there may be shared genetic contributions across CKD subtypes.^5,6 Recent studies support this hypothesis.⁷ For example, two variants in APOL1 (Online Mendelian Inheritance in Man reference number [OMIM]: 603743) significantly increase the risk for multiple forms of CKD, including FSGS, hypertension-attributed nephropathy, and HIV-1–associated nephropathy.^8–11 Similarly, mutations in the type 4 collagen genes (COL4A3 [OMIM: 120070], COL4A4 [OMIM: 120131], and COL4A5 [OMIM: 303630]) traditionally associated with Alport syndrome also frequently manifest as FSGS.^12–15 Collectively, these findings highlight the potential of genetic analyses to identify shared biology between disease causes traditionally considered to be distinct, enabling more precise classification of CKD on the basis of underlying molecular pathophysiology.

Although genome-wide association studies (GWAS) have identified multiple common genetic variants influencing all-cause CKD and kidney function, these explain a small fraction of the variance in these traits.^16–19 To date, no studies have systematically examined the contribution of rare genetic variation (<1% minor allele frequency [MAF]) with a presumed larger effect on the risk of CKD. Detection of rare independent variants clustered in a single gene can provide significant insight into disease biology and influence clinical therapy in several ways, even if pathogenic variants in a particular gene only explain a small proportion of all cases. First, in a population with both acquired and inherited multifactorial disease, accurate estimates of the proportion of cases caused by known genes can inform the use of existing therapies and diagnostic tests. Second, rare mutations can lead to the identification of widely applicable therapeutic drug targets, such as the discovery of PCSK9 mutations leading to the development of treatment for general forms of hypercholesterolemia.^20,21 Third, it is increasingly recognized that the validation of such drug targets in genetic studies of human populations improves the probability of the success of drug development in clinical trials.²²

In our rare-variant collapsing analyses, we first define a class of variation with properties that are presumed to enrich for biologic impact. We refer to variants that satisfy the imposed criteria as qualifying variants (QVs). Because single-variant association tests for rare variants lack power, the collapsing method aggregates QVs within each protein-coding gene and searches for a statistically significant difference in the rate of QV carriers between cases and controls (individuals with more than one QV in a gene are only counted once). These analyses have allowed gene discovery even when applied to heterogeneous, complex diseases with high allelic and locus heterogeneity; moreover, as long as the effect sizes are high and the causal variants are rare in the wider population, the approach can potentially require relatively small sample sizes.^23–25 For example, this approach has successfully detected new and known genes in general epilepsy, kidney malformations, polycystic liver disease, amyotrophic lateral sclerosis, and idiopathic pulmonary fibrosis.^26–32 Here, we apply a collapsing analysis to evaluate the contribution of rare variants to CKD risk in a large, multiethnic population drawn from two independent patient cohorts.

Methods

Study Cohort

We studied samples drawn from an initial combined cohort of 3315 individuals, reflecting two independent CKD cohorts (Supplemental Table 1).³³ After quality filtering (see Sequencing and Bioinformatic Protocols) and genetic matching to 9563 controls, the final case cohort comprised 3150 individuals. The Columbia University Medical Center (CUMC) cohort comprised 2041 patients with renal disease who were evaluated in CUMC nephrology clinics and who consented for genetics research. The second cohort comprised 1109 individuals from the AstraZeneca A Study to Evaluate the Use of Rosuvastatin in Subjects On Regular Hemodialysis: An Assessment of Survival and Cardiovascular Events trial (AURORA). AURORA recruited 2776 individuals aged between 50 and 80 years who were undergoing hemodialysis,³⁴ and 1109 participants were included in this study, comprising those who had consented to genetic research and for whom DNA was available and successfully sequenced. The recruitment and exclusion criteria for the AURORA cohort have been previously described.³⁴ Reasons for which DNA was not available for all patients in the AURORA study include lack of consent (collection was optional for patients), consent withdrawal, or country regulations not allowing for DNA collection. The clinical diagnosis for samples from both cohorts was classified into one of the following categories: (1) Mendelian CKD, including congenital or cystic renal disease; (2) glomerulopathy; (3) diabetic nephropathy; (4) hypertensive nephropathy; (5) tubulointerstitial disease; and (6) other/unknown, in cases where the primary disease did not belong to one of the above categories (“other”), or was altogether uncertain (“unknown”). We compared the genetic variation in these 3150 cases with that found among 9563 unrelated, healthy individuals and individuals ascertained for studies unrelated to kidney or urological disease in the Columbia Institute for Genomic Medicine (IGM) database (Supplemental Table 2). CUMC patients were more diverse in age and race/ancestry than AURORA patients (Supplemental Table 1). Reflecting different patient ascertainment, the AURORA and CUMC cohorts differed in the representation of clinical diagnostic subgroups, but collectively the two cohorts captured the major clinical categories of CKD (Table 1).³

Table 1.

Clinical diagnosis distributions across populations

Clinical Diagnosis	AURORA		CUMC		Total
	N	%	N	%	N	%
Diabetic nephropathy	178	16.1	176	8.6	354	11.2
Glomerulopathy	228	20.6	1042	51.1	1270	40.3
Hypertensive nephropathy	189	17.0	118	5.8	307	9.7
Mendelian CKD or congenital genitourinary anomalies	156	14.1	406	19.9	562	17.8
Other/unknown	148	13.3	274	13.4	422	13.4
Tubulointerstitial disease	210	18.9	25	1.2	235	7.5
Total	1109		2041		3150

Ethics

Patients recruited for the CUMC cohort provided informed consent for the use of DNA in genetic research, and the study was approved by the Columbia University Institutional Review Board. The AURORA study was performed in accordance with the provisions of the Declaration of Helsinki as defined by the International Conference on Harmonization, Good Clinical Practice, and applicable regulatory requirements, as well as the policy on bioethics and human biologic samples of the trial sponsor, AstraZeneca. For the sequenced AURORA patients, written, informed consent for the use of DNA for genetic research was obtained from patients during the original AURORA study, before initiation of this study. Deidentified data were provided to CUMC investigators for clinical sequence interpretation and collapsing analysis.

Sequencing and Bioinformatic Protocols

Samples were sequenced at the IGM at Columbia University. Exome enrichment was performed using Roche and IDT xGen Exome Research Panel V1.0 capture kits. Sequencing was performed on the Illumina HiSeq2500 and NovaSeq platforms. FASTQ data were processed and aligned to hg19/GRCh37 alignment using BWA-MEM,³⁵ and variants were called using the Genome Analysis Toolkit (GATK) v3.6 best practices. All samples were processed using the same bioinformatic pipeline in the IGM. Analyses were performed using the in-house Analysis Tool for Annotated Variants³⁶ for retrieving, annotating, and filtering variants in large cohorts.

All case (n=3150) and control (n=9563) samples included in the collapsing analysis achieved >10-fold read depth in >80% of the consensus coding sequence (CCDS release 20) and achieved average autosomal read depth of >30-fold. All samples were also confirmed to have concordance between genetic and self-declared sex, had contamination levels <4% according to VerifyBamID,³⁷ and were unrelated up to the second degree according to KING³⁸ (Supplemental Methods). Genetic ancestry was also defined for the test cohort using a panel of 12,840 common exome markers; for all analyses, genetic outliers were removed (Supplemental Methods). For the recessive analysis, which used a higher MAF threshold, test samples were restricted to those whose genetic variation placed them with high confidence among individuals of European genetic ancestry (probability >0.95).

For clinical sequence interpretation, we prioritized rare (population frequency ≤1%) nonsynonymous variants and canonical splice site variants in renal Mendelian disorders, identified using the OMIM database. We classified variants using the American College of Medical Genetics (ACMG) guidelines for clinical sequence interpretation.³⁹

Variant Quality Criteria

Collapsing analyses were applied using similar protocols to those described previously (Supplemental Methods).^28–30,40 To qualify for inclusion in the analysis, a variant was required to have a minimum read-depth of ten and be located within the CCDS or the 2 bp canonical splice sites within the introns. All QVs achieved a “pass” or “intermediate” status according to GATK’s Variant Quality Score Recalibration and passed several GATK quality score thresholds (Supplemental Methods). To reduce bias caused by differential coverage between case and control samples at QV sites, a binomial exact test was performed to exclude variant sites where the proportion of cases and proportion of controls with at least ten-fold read depth significantly differed (binomial exact test P<0.05). In addition to these quality metrics, as defined in our previous studies, QV refers to variants that also fulfill a predefined set of filtering criteria (Supplemental Table 3). A Fisher exact test was then applied to determine if the rate of case carriers was significantly different than the rate of control carriers under identical QV criteria.

Collapsing Analysis Overview

For discovering new CKD genes, we considered 12 distinct, nonsynonymous models for selecting QVs, comprising 11 dominant models and one recessive model incorporating gene-based and regional intolerance–based collapsing procedures (Supplemental Table 3). The recessive model screened for both homozygous genotypes and proxy compound heterozygous genotypes, whereby two QV observations in the same gene in the same patient were considered as a putative compound heterozygous genotype. In addition to running collapsing analyses on the combined case cohort of 3150 cases, we investigated the CUMC (n=2041) and AURORA (n=1109) case collections individually. These analyses were followed by six additional analyses focusing on the major subphenotype groups among the 3150 cases, for a total of nine clinical groupings. All 11 dominant models were run for the nine clinical groups, whereas the single recessive model was only run on the combined case cohort (Supplemental Table 3). Accounting for the 11 dominant runs across the nine different clinical subgroups, plus a single, recessive, nonsynonymous run on the combined case cohort, a total of 100 nonsynonymous models were analyzed.

Regionally Informed Collapsing Analysis

Recent studies have shown that accounting for regional mutation intolerance in coding regions can increase the power of rare-variant association studies.^41–43 Our analyses also included six models that included missense variants only if they fell within subgenic regions that are intolerant to missense variants,^41–45 according to certain thresholds determined by two measures of missense intolerance (MTR50 and localized intolerance model using Bayesian regression [LIMBR], Supplemental Method 3). We apply these filters to all missense-inclusive, nonsynonymous models and provide supporting data to show that these filters optimize the signal detection among known CKD genes. In addition, to determine whether the regional missense intolerance filters improve our missense signal detection, we use our top-ranked CKD gene signals from the primary analyses and compare the signal strength in the missense-only QV model before and after application of the intolerance filters (Supplemental Table 4).

Modifier Analysis

We also performed three secondary analyses, investigating genes that could potentially modify the risk of CKD in individuals harboring (1) the APOL1 risk genotypes, (2) diagnostic mutations in polycystin 1 (PKD1) and polycystin 2 (PKD2), and (3) diagnostic mutations in COL4A3, COL4A4, or COL4A5.

The APOL1 G1 risk allele is represented by both rs73885319 and rs60910145 missense variants, and the G2 risk allele is a 6 bp deletion (rs71785313, formerly rs143830837).⁸ APOL1-associated CKD is thought to follow a recessive model, with disease risk significantly higher among patients who carry two copies of the risk alleles (i.e., G1/G1, G1/G2, or G2/G2).^8,46 We identified 137 cases (4.3%) and 206 controls (2.2%) carrying two APOL1 risk alleles across the full case-control cohort. Note that despite the high frequency of cases carrying these known risk alleles, the G1 or G2 alleles were not expected to drive a signal for APOL1 in the general collapsing analyses because their global MAF (1%–2%) is above the maximum MAF thresholds for all collapsing analysis models used, meaning they will not be classed as QVs in any category used in this study. Genetic ancestry predictions using principal component analysis clustered these high-risk carriers with in-house samples of black or Afro-Caribbean ancestry. We therefore used the top six ancestry principal components to identify 2209 ancestry-matched individuals from the full control group for case-control comparisons (Supplemental Figure 1). We also adopted this ancestry approach to match the dual-risk APOL1 case samples with the remaining cases that did not harbor two risk variants as a negative control group. The cases for the PKD1/PKD2 and COL4A3/COL4A4/COL4A5 modifier analysis were selected on the basis of the presence of a diagnostic mutation in these genes, as classified by the ACMG criteria.^33,39 The modifier analysis for the PKD1/PKD2 and COL4A3/COL4A4/COL4A5 groups was conducted by comparing cases with diagnostic mutations in these genes with the broader set of 9563 multiethnic controls, because the cases with diagnostic mutations were also multiethnic and we have previously shown that ultrarare variants are not expected to be individually strongly associated with ancestry structure.²³

Statistical Analyses

For the collapsing analysis and the power calculation, a two-sided Fisher's exact test was used for the P values and odds ratios (ORs), with Bonferroni post hoc analysis. Calculations were performed in R software (version 3.4.0, 2017–04–21).⁴⁷ We defined a study-wide Bonferroni multiplicity-adjusted significance threshold as P<2.7×10⁻⁸, determined by (0.05/[18650 CCDS genes×100 nonsynonymous models]). This is conservative because of the correlation among the various models being considered. We also performed two additional synonymous QV models (one each for dominant and recessive) to act as negative experimental controls. We also refer to suggestive genes in this article as genes found to be highly ranked in the collapsing analyses, but with an understanding that highly ranked genes that do not achieve study-wide significance are a mixture of true (underpowered) biologic signals and observations from the tail of the null distribution. Thus, all suggestive genes depend on additional replication studies or orthogonal evidence to support clinical relevance. For our power calculation, we varied the samples sizes in our all-cause CKD cohort (n=3150), while maintaining the observed case and control carrier frequencies. We then investigated what case and control cohort sizes provided 80% power to achieve study-wide significance for a dominantly acting gene in which first, the collapsed QVs confer a minimum combined OR of 5.5, and second, QVs are present in at least 0.1% of cases.

We also performed a look-up of a recent analysis of UK Biobank data. We selected all kidney-associated traits from the UK Biobank data and downloaded their GWAS P values for genotyped and imputed variants from GeneATLAS (http://geneatlas.roslin.ed.ac.uk).⁴⁸ Quality control for imputed variants included ensuring an imputation score ≥0.9, a MAF of ≥0.001, and a Hardy–Weinberg equilibrium P value of ≥10⁻¹⁰. We next asked whether the top suggestive genes in the collapsing analysis were enriched for common variants for CKD in GeneATLAS, using a threshold P value of <10⁻⁴ for any of the kidney traits, and we included variants 50 kb up- and downstream of each gene.

Results

Gene-Based Collapsing Analyses

We analyzed exome sequencing data from 3150 individuals collectively representing the major clinical subcategories of CKD (Table 1). Three genes reached study-wide significance: PKD1 and PKD2, associated with autosomal dominant polycystic kidney disease (ADPKD); and collagen type 4 α5 chain (COL4A5), associated with X-linked Alport syndrome (Table 2). The signals for these genes were detectable among the all-cause CKD cohort and across multiple other models (Figure 1, Supplemental Table 5). As each of these genes is associated with a Mendelian form of nephropathy, we surmised PKD1, PKD2, or COL4A5 signals were driven by the presence of patients with ADPKD kidney disease and X-linked Alport syndrome, rather than rare variants in these genes contributing to all-cause CKD. Consistent with this expectation, the case enrichment was much stronger among the 562 cases comprising the “Mendelian or congenital CKD” subgroup (Figure 1, Table 2) and explained the signal in the overall cohort. In addition to these three genes, COL4A3 and COL4A4, which are also causal for Mendelian forms of CKD, were highly ranked across multiple analysis models (Figure 1, Supplemental Table 5, Table 2). We highlight these two genes because each of their association signals was distributed across the overall CKD cohort, indicating a contribution of pathogenic COL4A3 and COL4A4 variants across multiple clinical diagnostic categories (Supplemental Table 6). Synonymous models showed no genes at significant or near-significant enrichment (Supplemental Figure 2).

Table 2.

Best-ranked models

Gene Name	OMIM Phenotype No.	Known Mendelian Disease Inheritance	Best Model (Lowest FET P)
			Clinical Group/Inheritance Model/Collapsing Model	P Value (Fisher Exact Test)	OR [95% CI]
Known Mendelian nephropathy genes
PKD1	173900	Dominant/multiple	Mendelian or congenital CKD/dominant/ultrarare, deleterious predicted	1.6×10−55	29.45 [19.5 to 44.8]
PKD2	613095	Dominant/multiple	Mendelian or congenital CKD, dominant/rare, protein-truncating	5.6×10−26	>352 [56 to >13164]
COL4A5	301050	X-linked dominant/multiple	Mendelian or congenital CKD/dominant/rare protein-truncating	4.7×10−12	>155 [21 to >6512]
COL4A4	203780/141200	Autosomal dominant and recessive	All cases/dominant/ultrarare non-benign/LIMBR50	1.0×10−5	8.53 [3.1 to 23.7]
COL4A3	104200/203780/141200	Autosomal dominant and recessive	All cases/dominant/ultrarare deleterious predicted	3.1×10−5	7.92 [2.8 to 22.2]
No prior associations with Mendelian nephropathy
SLC17A1	NA	Not previously reported	AURORA only/dominant/rare missense-only	3.15×10−6	4.8 [2.5 to 9.0]
CPT2	255110/600649/608836	Autosomal recessive	All cases/dominant/rare missense-only/MTR50	4.0×10−7	4.50 [2.4 to 8.5]
SCLT1	NA	Not previously reported	CUMC only/dominant/rare, nonbenign/MTR50	2.0×10−6	8.25 [3.5 to 19.7]

To calculate OR and 95% CI for undefined events due to zero-cell counts we force a single count to zero cell and specify greater (>) than resulting OR and undefined 95% CI. Gene names are italicized. FET, Fisher Exact Test.

Figure 1.

Quantile-quantile plots show top signals for different models and phenotypes tested. The significant P values for known genes and suggestive P values for some novel genes are shown. The models tested are indicated above each plot. Plots (G–H) represent the modifier analyses for PKD1/PKD2 (G), COL4A3/4/5 (H), and APOL1 (I). The dotted line in plots (A–C) indicates the threshold for study-wide significance (P<2.7×10⁻⁸). No genes achieved study-wide significance threshold for plots (D–I). PTV, indicates the model using exclusively predicted protein truncating variants.

Diagnostic exome analysis using the ACMG guidelines for clinical sequence interpretation³³ detected Mendelian causes of CKD in 293 of the 3150 (9.3%) individuals. After exclusion of these 293 cases, there was no residual signal in COL4A5, COL4A4, or COL4A3, suggesting that the diagnostic analysis identified all or most of the pathogenic variants. However, PKD1 was still significantly enriched in the Mendelian group (ultrarare model, P=7.1×10⁻¹¹; OR, 7.8; 95% confidence interval [95% CI], 4.4 to 13.3) and PKD2 was suggestively enriched in the AURORA group (ultrarare model, P=8.5×10⁻⁴; OR, 20.2; 95% CI, 2.9 to 223.4), suggesting there are residual pathogenic variants in these genes that do not meet ACMG diagnostic criteria. Because most of these variants were in AURORA patients, for whom limited clinical information is available, it is possible that some of these residual variants would be classified as diagnostic using these criteria with more clinical information.

Because the signal from known disease genes such as COL4A3 and COL4A4 did not reach significance, it is possible that novel genes contributing to CKD in our cohort also did not reach our threshold of significance (Table 3). Beyond the known genes, the top-ranked signal belonged to SLC17A1 (OMIM: 182308; Table 2), a gene encoding a urate transporter that has been previously associated with gout and serum uric acid levels by GWAS.^49,50 This signal originated primarily from the AURORA cases and was distributed across all clinical categories (rare, missense-only model; 1.53% AURORA cases versus 0.32% IGM controls; P=3.2×10⁻⁶; OR, 4.8; 95% CI, 2.5 to 9.0). In addition to SLC17A1, we noted a suggestive signal in the sodium channel and clathrin linker 1 gene (SCLT1; OMIM: 611399), which was enriched for QVs among CUMC cases across different clinical categories (rare, nonbenign model; P=5.2×10⁻⁶; OR, 5.0; 95% CI, 2.4 to 10.3). SCLT1 encodes a ciliary protein, and inactivation of its murine ortholog causes cystic kidney disease.⁵¹

Table 3.

Top 25 ranked genes across all groups and models

Gene Name	Clinical Group	Best Model (Lowest FET P)
		Inheritance Model/Collapsing Model	P Value (Fisher Exact Test)	OR [95% CI]
PKD1	Mendelian or congenital CKD	Dominant/ultrarare, deleterious-predicted	1.6×10−55	29.5 [19.5 to 44.8]
PKD2	Mendelian or congenital CKD	Dominant/rare, LOF	5.6×10−26	352.3 [56.1 to 13163]
COL4A5	Mendelian or congenital CKD	Dominant/rare, LOF	4.7×10−12	155.4 [21.5 to 6512]
CPT2	All cases	Dominant/rare, missense/MTR50	4.0×10−7	4.5 [2.4 to 8.5]
AHDC1	APOL1 cases	Dominant/rare	5.91×10−7	14 [5 to 37.9]
SCLT1	CUMC only	Dominant/rare/MTR50	2.0×10−6	8.2 [3.2 to 22.7]
SLC17A1	AURORA only	Dominant/rare, missense	3.15×10−6	4.8 [2.5 to 9]
TNFAIP6	Glomerular neph.	Dominant/rare	4.23×10−6	10.8 [3.7 to 33.6]
SLC9A2	Glomerular neph.	Dominant/rare, missense/LIMBR50	5.04×10−6	4.4 [2.3 to 8.2]
ALDH1L2	Glomerular neph.	Dominant/rare/LIMBR50	5.09×10−6	6.6 [2.9 to 14.9]
COL4A4	All cases	Dominant/ultrarare/LIMBR50	1.01×10−5	8.5 [2.9 to 30.3]
PAX7	Unknown or other	Dominant/rare, missense/LIMBR50	1.04×10−5	5.5 [2.6 to 10.5]
EFEMP2	CUMC only	Dominant/rare, missense/MTR50	1.52×10−5	6.8 [2.7 to 18.1]
BAG3	Hypertensive neph.	Dominant/rare/MTR50	1.73×10−5	8.5 [3.3 to 19.2]
RPS6KL1	Glomerular neph.	Dominant/rare, missense/MTR50	2.06×10−5	6.1 [2.6 to 13.9]
KRTAP1 to 5	COL4A3/4/5 cases	Dominant/rare, missense	2.15×10−5	30.6 [7.2 to 98.8]
PDIA4	All cases	Dominant/rare/MTR50	2.43×10−5	4.4 [2.1 to 9.3]
JPH1	AURORA only	Dominant/rare, missense	2.67×10−5	4.1 [2.1 to 7.7]
GNPNAT1	All cases	Dominant/rare, missense	2.71×10−5	27.4 [3.8 to 1193.7]
CREBRF	Unknown or other	Dominant/rare/LIMBR50	2.78×10−5	22.9 [5.2 to 99.7]
ANXA2	CUMC only	Dominant/ultrarare, deleterious-predicted	2.91×10−5	18.8 [3.7 to 181]
TUBA1C	Hypertensive neph.	Dominant/ultrarare, deleterious-predicted	2.98×10−5	94.4 [7.5 to 4770]
GEMIN2	PKD1/2 cases	Dominant/rare/LIMBR50	3.04×10−5	78.4 [11.3 to 469]
COL4A3	All cases	Dominant/ultrarare, deleterious-predicted	3.05×10−5	7.9 [2.6 to 28.4]
TMEFF1	Tubulo-interstitial	Dominant/rare	3.72×10−5	17.3 [4.7 to 53.2]

To calculate OR and 95% CI for undefined events due to zero-cell counts we force a single count to zero cell and specify greater (>) than resulting OR and undefined 95% CI. Three genes that had high numbers of likely artifactual variants have been excluded. FET, Fisher Exact Test; LOF, loss of function; neph, nephropathy.

Regional Missense Intolerance Informed Collapsing Analyses

Accounting for regional mutation intolerance in coding regions can increase the power of rare-variant association studies. As some coding segments are more tolerant to missense variation, they are more likely to contain nondisease causal missense variants, which can add noise using gene-based collapsing analyses alone. Hence, we included six models that only included rare missense variants that fell within subgenic regions that are intolerant to missense variants, as determined by two measures of missense intolerance (Supplemental Method 3).^42,44,45 We first evaluated the performance of LIMBR⁴⁵ and missense tolerance ratio (MTR)⁴⁴ filters by analyzing known gene signals. Both methods increased enrichment of QVs in cases over controls in PKD1, with the MTR50 filter increasing significance by almost two orders of magnitude, demonstrating the utility of this approach (Table 4). The full comparison of the LIMBR- or MTR-filtered missense-only models across the five known genes (PKD1/2 and COL4A3/4/5) is shown in Supplemental Table 4.

Table 4.

Increase in PKD1 case enrichment with MTR and LIMBR intolerance filters

Gene Name	Subgroup	Model Name	Exome-Wide Rank	Qualified Case Frequency	Qualified Control Frequency	Enriched Direction	OR	P Value	95% CI Lower	95% CI Upper
PKD1	AURORA only	(Dominant) rare missense	6	0.0965	0.0668	Case	1.49	4.73×10−04	1.20	1.85
PKD1	AURORA only	(Dominant) rare missense LIMBR50	1	0.0902	0.0601	Case	1.55	1.95×10−04	1.24	1.93
PKD1	AURORA only	(Dominant) rare missense MTR50	2	0.0631	0.032	Case	2.04	1.06×10−06	1.56	2.66
PKD1	All samples	(Dominant) rare missense	309	0.0797	0.0669	Case	1.21	0.0158	1.04	1.41
PKD1	All samples	(Dominant) rare missense LIMBR50	102	0.0737	0.0602	Case	1.24	0.009	1.06	1.45
PKD1	All samples	(Dominant) rare missense MTR50	14	0.0454	0.0318	Case	1.45	4.35×10−04	1.18	1.77
PKD1	CUMC only	(Dominant) rare missense	9321	0.0701	0.0669	Case	1.05	0.5932	0.87	1.27
PKD1	CUMC only	(Dominant) rare missense LIMBR50	5600	0.0642	0.0602	Case	1.07	0.5075	0.88	1.30
PKD1	CUMC only	(Dominant) rare missense MTR50	5997	0.0358	0.0323	Case	1.11	0.4129	0.86	1.44

Next, we applied the regional intolerance filters on a genome-wide basis (Supplemental Method 3), using predefined models that included missense variants only or both protein-truncating and missense variants (Supplemental Table 3). Although no additional gene achieved study-wide significance, there was now suggestive evidence for one gene, carnitine palmitoyltransferase 2 (CPT2; OMIM: 600650), after adopting the regional missense intolerance filters (Supplemental Figure 3). Compared with the standard missense-only model, the MTR50% filter augmented the significance of the signal by over three orders of magnitude in the combined CUMC-AURORA cohort (rare missense-only model, P=4.7×10⁻⁴; OR, 2.1; 95% CI, 1.4 to 3.3, versus rare missense-only MTR50 model; P=4.0×10⁻⁷; OR, 4.5; 95% CI, 2.4 to 8.5). Because mutations in CPT2 cause a recessive syndrome which can include AKI and cardiomyopathy (carnitine palmitoyltransferase II [CPT II] deficiency, OMIM phenotypes 255110, 600649, and 608836), all CPT2 variant carriers were further investigated for a second, potentially recessive allele. None of the QV carrier cases in our cohort carried a second CPT2 allele of any protein-altering class with a gnomAD MAF <1%, nor did qualifying individuals carry the Ser113Leu SNP (rs74315294), the most commonly reported variant linked to familial CPT II deficiency.^52,53 The CPT2 signal also persisted after removing the cases with diagnostic mutations in known nephropathy genes. These findings indicate that the observed enrichment is not attributable to clinically undiagnosed cases of carnitine palmitoyltransferase II deficiency.

Genetic Modifiers Analysis

We next investigated genes that could potentially modify the risk of nephropathy in individuals with pathogenic variants in established disease-associated loci. We did not detect any noteworthy signals in subgroups with diagnostic mutations in PKD1/PKD2 (n=94) or in COL4A3/COL4A4/COL4A5 (n=87). Ancestry structure within each modifier group is shown in Supplemental Table 7. However, comparing the 137 CKD cases with the APOL1 risk genotypes to 2209 ancestry-matched controls, we identified a suggestive signal in the AT-Hook DNA Binding Motif Containing 1 gene (AHDC1; OMIM:615790) in the rare, nonbenign model (P=5.9×10⁻⁷; OR, 14.0; 95% CI, 5.0 to 37.9; Figure 1). This association was driven by nonsynonymous and indel variants distributed across the gene (nine QVs in 137 cases versus 11 QVs in 2209 controls) and was not detected among cases who carried only one or no APOL1 risk alleles (Supplemental Table 8). Moreover, of 206 controls with the APOL1 high-risk genotypes, only two carried a QV in AHDC1, further supporting the association (P=0.008).

Look-Up of Top Genes in the UK Biobank Data

We next asked whether common variants in our top genes were associated with CKD traits in a recent analysis of the UK Biobank data.⁴⁸ Overall, 45% of the 18,852 genes in this UK Biobank analysis were reported to contain or have close proximity (within 50 kb upstream/downstream of transcript boundaries) to at least one SNP that is nominally associated (P<0.001) with a kidney trait. In comparison, all nine genes of interest from collapsing analysis that either achieved P<3.0×10⁻⁰⁶ or lower in our collapsing analyses (PKD1, PKD2, CPT2, AHDC1, SCLT1, COL4A5, and SLC17A1, specific phenotypes and lowest P values shown in Supplemental Table 9) or were known genes with high diagnostic yield (COL4A3 and COL4A4) had at least one nominal association (P<0.001) with a nephropathy. The probability of detecting nine out of nine genetic associations by chance is low (probability, 7.6×10⁻⁴) and is consistent with an enrichment for true signals in the collapsing analysis. In addition, even after excluding known genes for kidney disorders (PKD1, PKD2, COLA43, COL4A4, and COL4A5), the probability remains low that all four of our novel suggestive genes would have nominal associations with renal phenotypes (P=0.04), when compared with other genes that do not lead to primary renal disorders.

Implications for Future Sequencing Studies of CKD

The observed distribution of case and control carrier frequencies provides an opportunity to estimate the sample sizes that could implicate novel CKD genes at study-wide significance with 80% power (Supplemental Figure 4). Using the dominant ultrarare deleterious predicted model (Supplemental Table 10), we estimated achievable P values and accompanying confidence intervals for increased sample sizes under the assumption that our current carrier frequency estimates reflect the true frequencies and that any future sampling will reflect an identical case composition as the current case sample. Under these stringent assumptions, the set data indicate that over a dozen susceptibility genes for all-cause CKD could be implicated within the range of effect sizes detected in this study, and a three-fold increase in case and control sample sizes would provide 80% power to detect them with collapsing analyses (Supplemental Table 10, Table 2).

Discussion

We present the largest whole-exome association analysis of CKD to date, comprising 3150 cases representing all major clinical phenotypes. We observe three study-wide significant genes, PKD1, PKD2, and COL4A5, and enrichment of rare variants in two others, COL4A3 and COL4A4, all of which are established causes of Mendelian nephropathy. The highly significant signals in known genes confirmed the strength of our analytic approach in detecting cases with hereditary nephropathy from a cohort composed of different subtypes of CKD. Some significant genes such as PKD1 or PKD2 were essentially all detected in the Mendelian CKD category, indicating it is unlikely that variants in these gene contribute to all-cause CKD. On the other hand, it is noteworthy that the signals in COL4A5, COL4A3, and COL4A4 emerged from the combined cohort, spanning across major clinical disease categories. Mutations in these genes cause X-linked and autosomal forms of Alport syndrome, but recent studies report pathogenic variants in patients diagnosed with other forms of CKD, such as FSGS.^12–14 Hence, the association of these genes with CKD across combined clinical categories provides additional and more formal quantitative evidence that rare variation in these genes contributes to many clinical forms of nephropathy.

The AURORA and the CUMC cohorts differed in ascertainment and recruitment strategies, likely explaining variation in signal strength for some genes between these two cohorts. The AURORA patients were composed of older persons (50–80 years old) on maintenance dialysis and the known relative survival advantage of patients with ADPKD on hemodialysis may explain the stronger PKD1/PKD2 representation in this cohort. CUMC’s enrichment for glomerular diseases may explain the stronger COL4A3/COL4A4/COL4A5 representation in that cohort. It should also be noted that we expect to observe some sampling variability especially in the context of sparse observation rates in relatively small individual cohorts.

We detected three interesting suggestive signals (SLC17A1, CPT2, and SCLT1). Although a recent analysis of UK Biobank data provides supportive evidence, replication in larger cohorts will be required. SLC17A1 encodes a renal urate transporter, and common variants in this gene have been associated with uric acid levels in large GWAS analyses.^49,50 A number of genetic and epidemiologic studies have indicated a link between serum uric acid levels and CKD, hypertension, and cardiovascular disease. Elevated uric acid levels have been independently associated with hypertension in most studies but its association with cardiovascular disease is not consistent. Hence, if the SLC17A1 signal is confirmed, it may clarify the causal relationship between uric acid levels, CKD, and cardiovascular disease.

The second gene, CPT2, encodes the CPT II protein, which is located in the mitochondrial membrane. CPT II deficiency (OMIM: 255110, 600649. and 608836) is a recessive disease of fatty acid metabolism that can cause severe cardiac and hepatic failure in its neonatal form, whereas its later-onset, myopathic form can produce acute tubular necrosis from rhabdomyolysis.⁵³ Although heterozygous carriers of CPT2 missense variants may experience mild metabolic defects,^54,55 to our knowledge, renal or cardiac dysfunction has not been reported except in individuals with biallelic, pathogenic mutations.

The third suggestive signal was in SCLT1, encoding a ciliary protein that has been nominated as a candidate gene for human nephropathy⁵⁶ because inactivation of its ortholog in mice causes cystic kidney disease.⁵¹ Although mitochondrial dysfunction and ciliary disorders usually cause pediatric forms of CKD,^56–59 the SCLT1 and CPT2 signals in this adult cohort suggest that dysfunction in these pathways may also contribute to later-onset CKD.

The APOL1 G1 and G2 alleles have risen to high MAF among populations of African ancestry because they confer protection against trypanosomal infection⁸; however, individuals who harbor two risk variants have a greater risk of several types of nephropathy.^10,60,61 In addition, APOL1 risk genotypes may independently predispose to cardiovascular disease.⁶² Because only a minority develop nephropathy, genetic or environmental modifiers have been suspected and a recent study has implicated common variants in UBD as genetic modifiers.⁶³ We did not detect any association with rare variants in UBD, but found suggestive evidence that rare, nonsynonymous variants in AHDC1 increase the risk of CKD in patients with the APOL1 dual-risk G1/G2 alleles. AHDC1 encodes a nuclear protein with a predicted DNA binding domain that is ubiquitously expressed. De novo mutations cause Xia–Gibbs syndrome, an intellectual disability syndrome that also features sleep apnea and various dysmorphologies.⁶⁴ The function of AHDC1 is not completely understood, but this suggestive signal encourages further investigation to confirm this association, and, if replicated, to examine the potential links to APOL1 biology.

A major challenge in rare-variant association studies is distinguishing deleterious variations, particularly missense substitutions, from neutral variation. Algorithms on the basis of evolutionary sequence conservation such as PolyPhen-2⁶⁵ have proven helpful, but they do not leverage the large amount of population genetic data now available,⁶⁶ which enables detection of coding regions that are intolerant to missense variation. We incorporated two filtering methods using regional missense intolerance, MTR and LIMBR, and both showed utility for optimizing missense variant signals in known genes (e.g., PKD1) and also detected a suggestive signal in the CPT2 gene. Systematic examination of MTR/LIMBR filters in exome-wide association studies, and extension of these methods, will better delineate opportunities for application of this approach.

Understanding the genetic basis of nephropathy can clarify the relationship between CKD and its comorbidities. For example, CKD is an established risk factor for cardiovascular mortality but some genetic nephropathies are associated with specific cardiovascular complications such as cerebral aneurysms (e.g., polycystic kidney disease), cardiac structural defects (e.g., coloboma, heart defects, atresia choanae, growth retardation, genital abnormalities, and ear abnormalities [CHARGE], Noonan, Bardet Biedl, Alagille syndromes), cardiomyopathy (e.g., Fabry disease or hereditary amyloidosis), or premature atherosclerosis (e.g., maturity onset diabetes of the young). Hence, exome sequencing can help differentiate cardiovascular complications that are linked to CKD as a manifestation of the same genetic defect, and those that arise as a consequence of renal impairment per se. Such genetic information can also improve risk stratification and adjudication of outcomes in interventional studies such as AURORA, which were undertaken to modify the risk of cardiovascular disease in renal failure.

In summary, this study explored the potential of exome sequencing and collapsing analysis to detect causal genes for nephropathy in a large, ancestrally diverse cohort representing broad clinical subcategories of CKD. We detected study-wide signals for known nephropathy genes and suggestive signals for several candidate genes among the all-cause CKD cohort. These data suggest that genetic risk factors may contribute to all-cause CKD, but their detection would require a larger sample size. For example, we estimate that a doubling of the case size would have yielded a study-wide signal for COL4A4 and other suggestive signals of similar magnitude. With the declining cost of whole-exome sequencing, such studies are increasingly feasible and may enable more precise patient classification on the basis of molecular subtypes. Our findings thus encourage more systematic investigation of the contribution of rare variants to disease risk in larger cohorts and within each CKD disease category.

Disclosures

This study was sponsored by the Innovative Medicines and Early Development (IMED) Biotech Unit, AstraZeneca, AstraZeneca. Authors Cameron-Christie, Fleckner, March, Haefliger, Platt, and Petrovski are employees of AstraZeneca, and Goldstein is an adviser for AstraZeneca. Participants recruited for “A Study to Evaluate the Use of Rosuvastatin in Subjects on Regular Hemodialysis: An Assessment of Survival and Cardiovascular Events” (AURORA), National Institutes of Health code NCT00240331. Sample collection for the Columbia University Irving Medical Center (CUIMC) cohort was partially supported by National Institutes of Health National Institute of Diabetes and Digestive and Kidney Diseases grant R01 DK0800899. The sequencing of the CUIMC cohort was supported by the CUIMC Precision Medicine Initiative.

Supplemental Material

This article contains the following supplemental material online at http://jasn.asnjournals.org/lookup/suppl/doi:10.1681/ASN.2018090909/-/DCSupplemental.

Supplemental Method 1. IGM bioinformatics processing.

Supplemental Method 2. Quality metrics.

Supplemental Method 3. Missense intolerance scores.

Supplemental Method 3. UK Biobank lookup.

Supplemental Figure 1. APOL1 PCA plots.

Supplemental Figure 2. QQ plots of synonymous variants.

Supplemental Figure 3. CPT2 enrichment, MTR filter.

Supplemental Figure 4. Power curve.

Supplemental Table 1. Cohort demographics.

Supplemental Table 2. Control cohort phenotypes.

Supplemental Table 3. Qualifying variant parameters and subgroups.

Supplemental Table 4. LIMBR and MTR comparison.

Supplemental Table 5. All genes across all models where P<0.05.

Supplemental Table 6. Clinical phenotypes in known genes.

Supplemental Table 7. Ancestry prediction in modifier analysis.

Supplemental Table 8. AHDC1 rare QV rates in APOL1.

Supplemental Table 9. Look-up of UK Biobank data.

Supplemental Table 10. Predicting P value signals.