Genome-Wide Association Studies in Nephrology: Using Known Associations for Data Checks

Abstract

Prior to conducting genome-wide association studies (GWAS) of renal traits and diseases, systematic checks to ensure data integrity and analytical workflow should be conducted. Using positive controls (ie, known associations between a single-nucleotide polymorphism [SNP] and a corresponding trait) allows for identifying errors that are not apparent solely from global evaluation of summary statistics. Strong genetic control associations of chronic kidney disease (CKD), as derived from GWAS, are lacking in the non-African CKD population; thus, in this perspective we provide examples of and considerations for using positive controls among patients with CKD. Using data from individuals with CKD who participated in the CRIC (Chronic Renal Insufficiency Cohort) Study or PediGFR (Pediatric Investigation for Genetic Factors Linked to Renal Progression) Consortium, we evaluated 2 kinds of positive control traits: traits unrelated to kidney function (bilirubin, body height) and those related to kidney function (cystatin C, urate). For the former, the proportion of variance in the control trait that is explained by the control SNP is the main determinant of the strength of the observable association, irrespective of adjustment for kidney function. For the latter, adjustment for kidney function can be effective in uncovering known associations among patients with CKD. For instance, in 1,092 participants of the PediGFR Consortium, the p-value for association of cystatin C concentrations and rs911119 in the CST3 gene decreased from 2.7*10^-3 to 2.4*10^-8 upon adjustment for serum creatinine–based estimated glomerular filtration rate. In this perspective, we give recommendations for the appropriate selection of control traits and SNPs that can be used for data checks prior to conducting GWAS among patients with CKD.

Genome-wide association studies (GWAS) investigate millions of genetic markers per person to identify genomic regions in which genetic variation associates with a trait or disease. For each single-nucleotide polymorphism (SNP), summary statistics (eg, a p-value) are calculated for its association with the phenotype. The resulting GWAS file contains millions of lines, which makes visual data and plausibility checks challenging. While there are excellent tools for systematically checking genome-wide summary statistics, including allele frequencies and computed association statistics for their global distribution^1,2, other systematic errors can go unnoticed. For example, incorrect association results may arise from a mismatch between the genotypes and phenotypes of the individuals (ie, inadvertent scrambling of the data). Because most SNPs are not expected to show an association with the phenotype of interest, such errors would escape global checks of summary statistics. Hence, additional checks to reliably assess the integrity of data and analytical workflow are required.

These considerations highlight the importance of using a positive control, one or several genomic markers that are known to reproducibly associate with an available trait or phenotype. A SNP that is a suitable positive control should have an effect strong enough to be detected in as little as a few hundred samples. As outlined in Box 1, a practical approach is to survey the GWAS Catalog of the National Human Genome Research Institute (NHGRI) for phenotypes and diseases available in a given study and to identify SNPs that previously have shown genome-wide significant associations (∼p<1*10^-7 or <1*10^-8, depending on the study). The SNPs should have been replicated successfully and should have shown associations in samples of the same ancestry as the data to be analyzed. We suggest reviewing the cited publications from the GWAS Catalog to select the SNP(s) that explain the largest amount of the trait variance. If this is not reported, summary statistics usually can help guide the selection of SNPs with the largest effect estimate and lowest p-values in a given study. Lastly, it is advantageous to have control SNPs represented on a given genotyping chip so that imputation is not required.

Box 1. Workflow to Select Phenotypes and Genetic Markers to Assess Known Associations for Quality Control Purposes.

Step 1: Select one, or preferably several, available control phenotypes. Preference should be given to phenotypes/traits that are continuous and measured (eg, biomarker concentrations) rather than to those that are self-reported. Ideally, the chosen biomarker is not generated in the kidney and does not exhibit net renal secretion or reabsorption.

Step 2: Look for previous GWAS of the corresponding phenotype/trait in the GWAS Catalog (www.genome.gov/gwastudies) and in PubMed. Ensure that the published association was found among individuals of the same ancestry as your study population, that the GWAS was sufficiently powered (large sample size) and the findings were replicated.

Step 3: Among significantly associated markers (typically P < 5 × 10^-8), select the one that explains the largest amount of phenotype/trait variance. If this is not reported in the original publication, select for large effect size estimates and low p-values instead. If several markers can be considered, prefer those with high minor allele frequencies and ones that have been genotyped (rather than imputed) in your own study.

Step 4: In your study, to the extent possible model the association between control trait and marker in the same way as was done in the original report, including trait transformation and units. Ensure the modeled allele and strand match those in the published report of the association.

Step 5: Compare direction and effect size of your association to the published result. Also assess whether the p-value meets statistical significance in your study, but (especially in smaller studies) do not expect the p-value to be as low as the ones initially published, which often originate from very large meta-analyses.

Step 6: If the blood concentrations of the chosen biomarker might be influenced by kidney function, rerun the association analyses adjusting for eGFR.

Step 7: If the positive control does not show the expected direction of association or the magnitude differs substantially, attempt to evaluate at least a second control trait. A typical mistake that can cause the repeated absence of known associations (and is not identified in any other data checks such as quality control, exploratory data analysis, data cleaning of phenotype and genotype information, and repetition of association analyses using a different statistical program) is a mismatch of the order of individuals in the phenotype and in the genotype file. This mistake results in the random shuffling of genotypes and phenotypes, giving rise to null associations.

Abbreviations: GWAS, genome-wide association study; eGFR, estimated glomerular filtration rate.

Finding a good positive control is challenging for GWAS in the field of kidney disease. In individuals of African descent, variants in the APOL1 gene have been shown to associate strongly with focal segmental glomerulosclerosis (FSGS), hypertension-attributed end-stage renal disease, and CKD from a variety of causes^3,4. Therefore, these markers might serve as positive controls. Since these variants are ancestry specific, data checks in samples that are not of African ancestry require the use of other positive controls.

Using quantitative control phenotypes such as biomarker concentrations generally is recommended due to the superior statistical power to detect associations with continuous phenotypes as compared to binary phenotypes (Box 1). Further, many of them are widely available. However, reduced kidney function influences blood concentrations of many biomarkers by altering their production, metabolism, and/or elimination. As a result, the genetic influence on marker concentrations can become less apparent.

A feasible and straightforward solution would be to use a biomarker with extrarenal production and, at least partially, extrarenal elimination, or a phenotypic trait unaffected by decreased kidney function. As an alternative, when evaluating the positive control association, it may be feasible to adjust for glomerular filtration rate (GFR) to reduce the effect of reduced kidney function on the biomarker blood concentration. In the following paragraphs, and summarized in Table 1, we present several examples and considerations.

Table 1.

Association of known SNPs for different phenotypes with and without adjustments for eGFR_cr.

Row	SNP	Gene	Phenotype, Unit	Adj eGFRcr	Effect /other allele	Effect Size	SE	p	Study	N	Published Effect Size	Reference
A	rs6742078	UGT1A	Bilirubin (total), log units	N	T/G	0.29	0.02	6 × 10-32	CRICa	1527	0.23	5
B	rs6742078	UGT1A	Bilirubin (total), log units	Y	T/G	0.29	0.02	4 × 10-32	CRICa	1527
C	rs6440003	ZBTB38	Body Height, cm	N	A/G	0.42	0.24	0.08	CRICa	1549	0.48*	7
D	rs6440003	ZBTB38	Body Height, cm	Y	A/G	0.41	0.24	0.09	CRICa	1549
E	rs911119	CST3	Cystatin C, mg/dl	N	T/C	0.13	0.04	3 × 10-03	PediGFR	1092	positive **	23
F	rs911119	CST3	Cystatin C, mg/dl	Y	T/C	0.14	0.03	2 × 10-08	PediGFR	1092
G	rs7442295	SLC2A9	Uric acid, mg/dl	N	G/A	-0.11	0.07	0.11	4C	535	-0.34 ***	24
H	rs7442295	SLC2A9	Uric acid, mg/dl	Y	G/A	-0.12	0.07	0.08	4C	535

See Box 2 for detailed methods. All models were adjusted for age, sex, and principal components; additional adjustment for eGFR_cr as indicated. SNP effect allele frequencies: 0.33 (rs6742078), 0.43 (rs6440003), 0.81 (rs911119), 0.24 (rs7442295).

^acomprises patients of European American ancestry

^*1 SD in these studies is equal to 6.82 cm; the published effect size of 0.07 SD is equal to 0.48 cm

^**rs911119 T allele decreases cystatin C–based eGFR and thus increases cystatin C

^***Urate unit conversion: the published effect size of 0.02 mmol/l is equal to 0.34 mg/dl

Abbreviations: adj, adjusted; SNP, single-nucleotide polymorphism; eGFR_cr, serum creatinine–based estimated glomerular filtration rate; CRIC, Chronic Renal Insufficiency Cohort; PediGFR, Pediatric Investigation for Genetic Factors Linked to Renal Progression; 4C, Cardiovascular Comorbidity in Children with Chronic Kidney Disease; SE, standard error; rs, reference SNP identifier; SD, standard deviation.

One suitable marker with extrarenal production and largely extrarenal elimination is bilirubin. Serum concentrations of bilirubin reflect the balance between its production and elimination. Although liver and kidney disease can coexist, CKD has not been described as having a major effect on hepatic function, and serum bilirubin values of patients with CKD are usually in the reference range. Polymorphisms in UGT1A, which encodes the bilirubin UDP (uridine diphosphate)-glucoronosyltransferase, were identified initially as associated with bilirubin concentrations in some cohorts forming the CHARGE (Cohorts for Heart and Aging Research in Genomic Epidemiology) Consortium⁵. In this study, the SNP rs6742078 explained 18% of the variance in bilirubin concentrations, with a p-value of <5*10^-324 and a 0.23-unit higher log(total serum bilirubin concentration in μmol/l) per T allele among 9,464 individuals studied. Effect sizes of this magnitude should be easily detectable at genome-wide significance even in samples of smaller size.

Row A of Table 1 shows the association result for rs6742078 with log(serum bilirubin concentration) in 1,527 participants of European American ancestry of the CRIC (Chronic Renal Insufficiency Cohort) Study⁶. Allele frequency, effect direction, and effect size were consistent with previously published results, with an association p-value of 6*10^-32. Additional adjustment for serum creatinine-based estimated GFR (eGFR_cr) did not change the effect size of the SNP on bilirubin concentrations and left the p-value essentially unchanged (4*10^-32; Table 1, row B). Panels A and B of Figure 1 show a graphical representation of this example, highlighting the feasibility of using biomarkers with extrarenal metabolism as positive controls in association studies of renal phenotypes. This strategy also increases confidence in genotyping quality, correct data handling, and the approach to generate association statistics.

Figure 1.

Effects of eGFRcr adjustment on SNP associations with blood bilirubin and cystatin C concentrations. See Box 2 for detailed methods.

X-axis: genomic position; y-axis: logarithmic association p-value and recombination rate calculated from the HapMap rel 22 data. The control SNP selected from previous publications is displayed in purple; the color-coding indicates the pair-wise correlation (r2 from 1000 Genomes EUR panel, March 2012 release) of each SNP with the control SNP.

Panel A and B: Associations at the UGT1A locus with bilirubin, without (A) and with (B) conditioning on GFRcr

Panel C and D: Associations at the CST3 locus with cystatin C, without (C) and with (D) conditioning on GFRcr

Body height should not be affected by CKD of adult onset. Because this anthropometric measure is available in many studies, it represents another phenotype that may serve as a positive control. Ideally, body height should be obtained by standardized measurement rather than self-report. We evaluated the association between height and rs6440003 (the SNP, in the gene ZBTB38, which showed the strongest association with body height in one of the first GWAS of the trait)⁷ among 1,549 participants of European American ancestry in the CRIC Study (Table 1, row C). Although allele frequency, as well as effect direction and magnitude, were consistent with previous findings, the association was only of borderline significance (P = 0.08). As expected, adjustment for eGFR_cr did not change the evaluated association. The SNP only explained 0.3% of the variance in body height in the original publication⁷; therefore, it is likely that this association would not be detected in studies of small size and thus is not an ideal positive control. This example highlights the importance of finding a variant where the proportion of explained phenotypic variance is large. Furthermore, study-specific characteristics need to be taken into consideration. For example, height is not suitable as a positive control in children with CKD because CKD influences growth, and height in children varies as a function of age.

Cystatin C is a protein encoded by CST3 and ubiquitously found in nuclei. Because of its clearance by the kidney, it is commonly used as a biomarker of kidney function and data on its concentration may be available in studies of patients with kidney disease. Studies in the general population found a sizeable effect of variants in CST3 (eg, rs911119) on serum concentrations of cystatin C⁸. Since a marker that is cleared by the kidney like cystatin C is affected substantially by reduced kidney function, associations between the marker and a corresponding SNP are expected to strengthen when adjusting for eGFR_cr. In the PediGFR Consortium, formed by 3 studies of CKD in children (CKiD [Chronic Kidney Disease in Children] Study⁹, ESCAPE [Effect of Strict Blood Pressure Control and ACE Inhibition on the Progression of CRF in Pediatric Patients] Trial¹⁰ and 4C [Cardiovascular Comorbidity in Children with Chronic Kidney Disease] Study¹¹), the association between serum cystatin C concentrations and rs911119 was nominally significant (P = 0.003; Table 1, row E). As shown in Figure 1 (panels C and D), adjusting for eGFR_cr substantially decreases the p-value to 2*10^-8 (Table 1, row F), accompanied by a reduction in the standard error arising from the variation in cystatin C concentrations caused by reduced clearance from the kidneys. This example illustrates that even a phenotype affected by kidney function can be used as a positive control, as long as data are adjusted for GFR.

Uric acid is another biomarker with extrarenal production that may be available as part of a biomarker panel in studies of kidney function and disease. We evaluated rs7442295 in the urate transporter gene SLC2A9, a SNP that showed the strongest association with serum urate concentrations in a previous population-based study among individuals of central European ancestry¹², explaining 3.5% of the urate variance. The association between the SNP and serum urate in the 4C Study was of the same direction as the published association but not statistically significant. Adjustment for eGFR_cr did not lead to appreciable changes in the association p-value (Table 1, rows G and H).

This observation might be due to the more complex homeostasis of urate compared to cystatin C. As a housekeeping gene, CST3 produces cystatin C at a constant rate. However, urate concentrations are influenced by several factors; for example, urate concentrations are affected by purine metabolism (of which it is the end product), renal transport (which is complex and strongly influenced by genetic variation), and environmental influences like dietary intake. All of these factors may be affected in the setting of CKD. Therefore, statistically adjusting for eGFR_cr alone may not be adequate to account for all determinants of urate variability in a way that allows for the detection of a genetic effect on urate concentrations in subjects with CKD. Conversely, cystatin C is freely filtered at the glomerulus and largely undergoes catabolism after tubular reabsorption; the remainder of cystatin C is eliminated in the urine. Because blood concentrations of this biomarker are mainly affected by its excretion, the association of cystatin C level and CST3 SNP become more apparent upon adjusting for eGFR_cr.

If sufficient biomaterial and financial resources are available, phenotyping a trait that can be used to assess the presence of known associations can add to the value of a study. Cystatin C is an obvious choice because many CKD studies already have measured (or are interested in measuring) this marker, as it is required for using the latest CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) GFR estimating equation¹³. Purely from a quality control perspective, measuring bilirubin may be preferable because of the existence of a polymorphism that explains an exceptionally large amount of the variation in serum bilirubin concentrations. An added advantage bilirubin is that it an economical biomarker to measure.

In conclusion, the conduct of systematic data checks prior to conducting GWAS in CKD populations remains challenging because of the lack of previously validated positive controls in populations of non-African ancestry. This brief perspective highlights alternative approaches for assuring data integrity by using extrarenal and renal biomarkers and traits as positive controls.

Box 2. Methods.

Genotype and phenotype data sets were acquired from the ESCAPE, 4C, CKiD and CRIC studies^6,9-11. Serum creatinine–based eGFR was calculated using the Modification of Diet in Renal Disease [MDRD] Study equation¹⁴ for adults and the enzymatic serum creatinine equation by Schwartz et al¹⁵ for children. After stringent data cleaning according to standard protocols¹⁶, imputation was performed using the public 1000 Genomes data as a reference panel (phase 1, release v3, release date 2010-11-23, ALL subset)^17-19. Association analyses were conducted with the software SNPtest v2.5²⁰. Age, sex, significantly (p<0.05) associated principal components, and optionally eGFR_cr were used as covariates. Metaanalysis of results across the PediGFR studies (ESCAPE, 4C, CKiD) was performed using the software GWAMA²¹. Regional association plots were created using the software LocusZoom²².