Meta-analysis of genome-wide association studies confirms a susceptibility locus for knee osteoarthritis on chromosome 7q22

Abstract

Osteoarthritis (OA) is the most prevalent form of arthritis and accounts for substantial morbidity and disability, particularly in the elderly. It is characterized by changes in joint structure including degeneration of the articular cartilage and its etiology is multifactorial with a strong postulated genetic component. We performed a meta-analysis of four genome-wide association (GWA) studies of 2,371 knee OA cases and 35,909 controls in Caucasian populations. Replication of the top hits was attempted with data from additional ten replication datasets. With a cumulative sample size of 6,709 cases and 44,439 controls, we identified one genome-wide significant locus on chromosome 7q22 for knee OA (rs4730250, p-value=9.2×10⁻⁹), thereby confirming its role as a susceptibility locus for OA. The associated signal is located within a large (500kb) linkage disequilibrium (LD) block that contains six genes; PRKAR2B (protein kinase, cAMP-dependent, regulatory, type II, beta), HPB1 (HMG-box transcription factor 1), COG5 (component of oligomeric golgi complex 5), GPR22 (G protein-coupled receptor 22), DUS4L (dihydrouridine synthase 4-like), and BCAP29 (the B-cell receptor-associated protein 29). Gene expression analyses of the (six) genes in primary cells derived from different joint tissues confirmed expression of all the genes in the joint environment.

Introduction

Osteoarthritis (OA) is the most prevalent form of chronic joint disease and accounts for substantial morbidity and disability, particularly among the elderly. It is characterized by loss of joint homeostasis. The articular cartilage cannot maintain its integrity and is progressively damaged, the subchondral bone envelope is thickened changing loads in the bone-cartilage biomechanical unit, the synovium shows signs of inflammation and bony spurs (osteophytes) appear at the edges of the bone. Its etiology is multifactorial with a significant genetic component as shown by twin and family studies[1, 2].

Many genetic variants have been considered as potential risk factors for OA but most of the reported associations are inconclusive or not replicated. Recently, a large-scale meta-analyses found evidence that the GDF5 locus on chromosome 20 was associated with the increased risk of knee OA in Caucasians[3-6]. Other genome-wide data have reported an association with the DVWA gene in Asians but not Caucasians [7] and a PTGS2 variant that replicated but did not reach genome-wide significance (GWS)[8]. Recently, a genome wide association (GWA) study identified a locus on chromosome 7q22 that shows an association with combined knee OA and/or hand OA phenotype[9].

In this study we have synthesized available data from four GWA studies under the auspices of the TreatOA (Translational Research in Europe Applied Technologies for Osteoarthritis [www.treatoa.eu]) consortium. A total of 2,371 knee OA cases were available for this first stage of the analysis. The most significant signals were further investigated in additional samples of European descent and SNPs that reached genome-wide significance (GWS) were further evaluated in Asian samples. .

Methods

Study design

A detailed description of all samples used in this study is provided in the Supplementary Material. We used a three-stage design for the identification of any potential associations between sequence variants and knee OA in populations of European ancestry. We first synthesized the available data from 4 GWA studies (deCODE, Rotterdam Study, Framingham, Twins UK) using inverse variance fixed effects models. The variants that reached the 2×10⁻⁵ level of significance were selected for further replication. These SNPs were followed-up in 8 additional European cohorts (arcOGEN, Greek, Spanish, Finnish, Nottingham, Chingford study, GARP, Estonian and Swedish). The SNPs that replicated in the follow-up samples were genotyped in additional 2 European samples (deCODE (Icelandic) and Swedish). One cohort provided in silico replication from an ongoing GWA study (arcOGEN, 12 SNPs were directly genotyped, and 6 were imputed), whereas de novo replication was performed in the other cohorts. Furthermore, the top hits were followed-up in Asian populations (Chinese and Japanese samples). The effect sizes from the meta-analysis of the GWA studies and the effect sizes from the replication effort were all combined to provide an overall estimate. We also synthesized the effect estimates of the European and Asian samples to provide a global summary effect estimate

Phenotype definitions

Study subjects with a radiographic Kellgren and Lawrence (K/L) grade≥ 2 [10] or total knee replacement (TKR) were included as cases in the analysis. When clinical criteria were considered (Greek, Spanish and GARP study groups) the American College of Rheumatology (ACR) classification criteria were used [11]. As controls, we considered subjects who had no known affected joints among those assessed. For example in a cohort that assesses knee, hip and hand OA, controls were participants with no affected hip or hand joints for the knee OA analysis. Population-based controls were used for the arcOGEN study.

Genotyping and imputation

Samples from the GWA studies were genotyped using the Infinium HumanHap300 (Illumina) for deCODE and Twins UK samples, HumanHap550v3 Genotyping BeadCHip (Illumina) for the Rotterdam Study and the Affymetrix GeneChip® Human Mapping 500K for the Framingham cohort. The number of SNPs genotyped ranged from 314,075 to 500,510. Imputations were performed to increase the coverage. All the top SNPs studied had acceptable imputation quality. Finally, the genotyped and imputed SNPs that successfully passed the quality control criteria (n=2,335,627) were considered for the analyses. Detailed information on genotyping platform, quality control and imputation methods for each cohort are presented in Supplementary Table 3.

The replication samples for the Greek, Spanish, Finnish, Chingford and GARP studies were genotyped using the MassArray iPlex Gold from Sequenom. Replication genotyping was carried out by a genotyping contractor (Kbiosciences Ltd) using a competitive allele-specific PCR SNP genotyping system for the Nottingham and the Estonian cohort. The additional 622 Icelandic cases and the samples from the Swedish cohort were genotyped by deCODE genetics using the Centaurus (Nanogen) platform [12]. Detailed information on genotyping is provided in Supplement.

Statistical analysis

Association analysis

Each team performed an association testing per gender for knee OA under a per-allele model. The lambda inflation factor was calculated per gender-specific effect size using the genomic control method [13] and the standard errors were corrected by the square root of the lambda inflation factor (${\mathit{SE}}_{\text{corrected}}={\mathit{SE}}_{\text{observed}}x\sqrt{\lambda }$). Robust standard errors were estimated to adjust for the family relationships (Framingham study and GARP).

Meta-analysis

The effect size for each SNP (odds ratio per copy of minor allele as per HapMap) was calculated using inverse-variance fixed effects models [14], synthesizing all the sex-specific effect sizes and the corrected standard errors. We also performed analyses combining men and women. In family studies results from men and women combined were used in order to account for relatedness between females and males within families. Meta-analyses of the GWA studies were performed using the METAL (www.sph.umich.edu/csq/abecasis/metal) software. Between-study heterogeneity was tested using the Cochran’s Q statistic, which is considered significant at p<0.1. The extent of inconsistency across studies was quantified using the I² metric which ranges from 0 to 100% [15]. Heterogeneity is considered low, moderate, high and very high for 0-24%, 25-49%, 50-74% and >75% respectively [16]. We also computed the 95% CI for the I² [17]. Furthermore, we repeated the calculation with random effects models for all SNPs that were further evaluated in replication datasets. Results are shown in Supplementary Table 4. Meta-analyses of the 18 top-hits were performed using Stata version 10.1.

Assessment of credibility

In order to assess the credibility of the top hit we calculated the Bayes factor under a spike and smear prior using as an alternative an average genetic effect corresponding to an OR of 1.2 and a conservative agnostic prior of 0.0001% [18].

Functional analysis

Two methodological approaches were used to investigate the functional role of genes identified by GWA studies. (A) By assessing their expression in primary human joint cells (synovial fibroblasts, chondrocytes and meniscal cells) and its change in response to the proinflammatory cytokines TNFα and IL1β as well as comparing their gene expression profiles during chondrocyte de-differentiation (3D pellet cultures of vs monolayer culture, see Supplement) and (B) by assessing their expression dynamics by wholemount in situ hybridization using 6h (shield), 10h (bud), 13h (5-9 somites) and 1, 2, 3 and 4 days old zebrafish (Danio rerio) embryos, in order to explore their role during embryogenesis (see Supplement).

Results

Meta-analysis of GWA studies and replication of top findings

The descriptive characteristics of the GWA studies used for the meta-analyses are from Iceland (deCODE), the Netherlands (Rotterdam study), USA (Framingham) and the UK (Twins UK). The characteristics of these studies are presented in Table 1 and Supplement. The 4 GWA datasets included a total of 2,371 cases and 35,909 controls. A quantile-quantile (QQ) plot, comparing the meta-analysis association results of the four studies to those expected by chance, showed an excess of SNP associations indicating a likely true association signal (Figure 1). The data analysis revealed the strongest association on chromosome 7q22 with a p-value of 5.06 ×10⁻⁸ for rs4730250 localized in dihydrouridine synthase 4-like gene (DUS4L) (Figure 2). Other associated signals in 7q22 gene cluster are in high linkage disequilibrium (LD) (r²>0.8) with the top signal (Figure 2).

Table 1.

Characteristics of the studies included in the analysis

Team	Knee OA Cases/ Controls	Platform used	Age mean (range)	BMI mean (range)	Females (%)	Knee OA definition	Control definition
GWA studies
deCODE*	1033/32482	Infinium HapMap 300	69(19-99)	26(14-60)	58%	TKR	Health care records
Framingham	419/1674	Affymetrix GeneChip®	64(29-93)	26(14-54)	56%	Radiographic	Radiographic
Rotterdam	868/1464	Illumina HapMap550v 3	67(55-94)	26(16-56)	59%	Radiographic	Radiographic
TwinsUK	51/289	Infinium HapMap 300	54(37-76)	25(15-51)	100%	Radiographic	Radiographic
Replication cohorts stage 1
arcOGEN	1643/4894	Illumina 610 Quad	NA	NA	71%	Radiographic /clinical	General population
Chingford(a)	64/236	NP	63 (54-77)	26 (17-43)	100%	Radiographic	Radiographic
Finnish	112/210	NP	67 (51-74)	29 (20-42)	75%	TKR	Population-based
Greek	368/606	NP	61(20-90)	26(17-34)	72%	Clinical	Clinical
GARP	161/758	NP	60(30-79)	27(19-47)	63%	Radiographic /clinical	Radiographic/clinical
Spanish	262/294	NP	66(32-94)	31(18-53)		TKR/clinical	Clinical
Nottingham(b)	647/237	NP	66 (40-97)	27 (15-51)	53%	TKR	Radiographic and clinical
Estonian	69/456	NP	47 (32-60)	28(15-47)	69%	Radiographic	Radiographic
Replication cohorts- Stage 2
deCODE	622/32482(c)	Illumina and Centaurus (Nanogen)	77 (40-99)	29 (19-49)	63%	TKR	Population-based
Swedish	390/839	NP	62 (46-73)	29 (18-51)	63%	TKR+conco mitant clinical & radiographic diagnosis of OA	General population without TKR

NP: Not pertinent; TKR: Total knee replacement; THR: total hip replacement

^(a)Numbers excluding the samples already included in the arcOGEN study.

^(b)Numbers excluding the samples already included in the arcOGEN study.

^(c)same controls as for discovery cohort.

Figure 1.

Q-Q plot of the expected vs observed distribution of p-values.

Figure 2.

Regional association plot of rs4730250. Statistical significance of the associated SNPs are illustrated on –log₁₀ scale. The p-value of the rs4730250 and the other 10 selected SNPs are based on the meta-analysis of all datasets (both GWA studies and replication studies). P-values for the rest of the SNPs are based on the meta-analysis of the GWA studies. The sentinel SNP is shown in blue. The correlation of the sentinel SNP is shown on a scale from minimal (gray) to maximal (red). SNPs in red have r²≥0.8 with the sentinel SNP and SNPs in orange have r²≥ 0.5. Chromosome positions are based on HapMap release 22 build 36.

We selected for follow-up in replication samples all SNPs with a p-value <2×10⁻⁵ in the meta-analysis association results. A total of 18 SNPs from 10 chromosomal loci satisfied this criterion (Supplementary Table 1). However, as some of those SNPs are fully equivalent in the HapMap-CEU dataset a total of 11 non-identical SNPs were tested for replication. We analyzed these 11 SNPs for replication in 3,326 cases and 7691 controls from 8 European studies (Table 1 and Supplement). Two SNPs, rs4730250 and rs10953541 both located at 7q22, replicated nominally (p<0.05) in the combined analysis of the follow-up samples, with p-values of 6.3×10⁻⁴ and 8.3×10⁻³ respectively. The two SNPs, rs4730250 and rs10953541, were then further genotyped in two additional replication sets.

Both SNPs reached GWS in a meta-analysis of all European sample sets (the GWA datasets and the replication cohorts) (Table 2). We analyzed a total of 6,709 knee OA cases and 44,439 controls. SNP rs4730250 was genome wide significant GWS with a per-allele summary OR of 1.17 (95% CI: 1.11-1.24) and a p-value of 9.2 ×10⁻⁹. The minor allele frequency was 0.17 in the combined dataset. Low heterogeneity was observed (I²=15%, 95% CI: 0-48%), which was not statistically significant (p=0.26 for Cochran’s Q statistic) (Figure 3). No gender specific effects were seen. The summary estimates did not differ significantly in men and women (p-value=0.74, test of homogeneity) (Figure 3). Analysis where both sexes were analyzed together in all cohorts did not alter the results (OR=1.17 [95% CI: 1.07-1.27], p-value=4.1×10⁻⁸). The summary effect sizes of all loci under study are presented in Table 2. The two significant SNPs at 7q22, rs4730250 and rs10953541, are highly correlated (D′=1, r²=0.63 in HapMap-CEU) and are likely to represent the same underlying association signal as shown by conditional association analysis (Supplementary Table 2).

Table 2.

Summary odds ratios and 95% confidence intervals of SNPs in the analysis including all European descent data.

SNP rs number	Minor (risk)allele	Chr	Position	Gene	MAF	OR (95% CI) Fixed effects	p-value	I2 (95% CI)	Cochran’s Q
rs4730250	G	7	106994931	DUS4L	0.17	1.17 (1.11-1.24)	9.17×10−9	15(0-49)	0.26
rs10953541	T	7	107031781	BCAP29	0.24	1.17 (1.10-1.23)	3.90.×10−8	19 (0-54)	0.23
rs3749132	A	2	68907001	ARHGAP25	0.07	1.17 (1.05-1.30)	4.08×10−3	47 (0-74)	0.04
rs886827	C	7	42285581	GLI3	0.27	1.07 (0.99-1.16)	0.089	65 (43-80)	0.001
rs1886695	G	20	33643949	CPNE1	0.16	0.89 (0.84-0.95)	1.76×10−4	42 (2-66)	0.02
rs10071956	T	5	173093290	Intergenic	0.38	1.12 (1.06-1.19)	5.05×10−5	15 (0-53)	0.29
rs6816070	G	4	16089455	LDB2	0.42	0.91 (0.86-0.95)	1.34×10−4	0 (0-54)	0.46
rs661924	T	10	21353562	NEBL	0.39	1.11 (1.05-1.17)	1.82×10−4	30 (0-67)	0.18
rs436354	G	5	783271	ZDHC11	0.17	1.19 (1.01-1.30)	1.79×10−2	41(2-63)	0.06
rs1994104	T	12	83040643	intergenic	0.13	0.88 (0.80-0.96)	3.13×10−3	46 (2-70)	0.02
rs9857056	G	3	181698548	intergenic	0.12	1.11 (1.02-1.20)	1.65×10−2	72 (43-87)	0.001

MAF: minor allele frequency; Minor allele is the OR allele

Figure 3.

a) Forest plot of study-specific estimates (black boxes) and summary OR estimates and 95% confidence intervals (95% CIs) (diamond) for the association between the rs4730250 SNP and knee osteoarthritis.

Age and BMI are considered to be significant risk factors for the development of knee OA [19-25]. We performed an analysis where the top hit was adjusted for these risk factors in deCODE samples and the Rotterdam Study. The association of the top hit remained largely unchanged in analyses adjusted for BMI and age.

In order to assess the credibility of the associations of the two SNPs, we calculated the Bayes factor[18] under a spike and smear prior using an average genetic effect corresponding to an OR of 1.2 and a conservative agnostic prior (assuming no prior knowledge of the association) of 0.0001%. The posterior credibility of these associations was 98% and remained similarly high even with a small alternative effect size of 1.1.

We also tested if the observed signal at the 7q22 region was replicated in East Asian samples (Japanese and Chinese cohorts). The total number of knee OA cases and controls assessed was 1183 and 1245 respectively. The rs12535761 was used as a proxy for the rs4730250. The two SNPs are in strong LD (r²=1, D′=1, in HapMap Asian samples). The finding was not replicated in the Asian samples with a summary effect size of 1.03 (95% CI: 0.85-1.25). A meta-analysis including both European and Asian samples with 7,892 cases and 45,684 controls yielded a global summary effect of 1.15 (95% CI: 1.10-1.22) with p-value=5.7×10⁻⁸ for rs4730250 with low heterogeneity (I²=19%).

Functional analysis of genes in 7q22 cluster

The associated signal at 7q22 is located within a large (500kb) linkage disequilibrium (LD) block that contains six genes; PRKAR2B (protein kinase, cAMP-dependent, regulatory, type II, beta), HPB1 (HMG-box transcription factor 1), COG5 (component of oligomeric golgi complex 5), GPR22 (G protein-coupled receptor 22), DUS4L (dihydrouridine synthase 4-like), and BCAP29 (the B-cell receptor-associated protein 29).

We performed additional experiments to get more information about the genes in the cluster and their potential role in joint biology and pathology. Analysis of the mRNA expression data in chondrocyte pellet indicates that BCAP29, COG5, DUS4L and HBP1 expression levels were higher than in monolayer cultures suggesting that they are expressed in an environment that more accurately recapitulates articular cartilage. In contrast no difference were seen for GPR22 and PRKAR2B mRNA expression. In a zebrafish model the expression of all genes was detectable from the shield stage onwards. Results are described in detail in Supplement

Discussion

This study provides further evidence for a knee OA signal localizing to the 7q22 cluster region and associated with knee OA. The statistical credibility and confidence of this evidence is very high based on the calculations of the Bayes factor. The same locus has been identified and proposed as an OA susceptibility locus from the Rotterdam Study for the prevalence and progression of OA [9]. Our study and the earlier Rotterdam Study do include overlapping populations. However, our study was specifically targeting the knee OA phenotype. Furthermore, an additional 3 European cohorts and two Asian populations were used for further replication. Our study utilizes the largest sample size in the genetics of knee OA research to date with almost 8000 knee OA cases analyzed.

The most significant hits identified by our study are located within a large (500kb) linkage disequilibrium (LD) block that contains six genes; PRKAR2B, HPB1, COG5, GPR22, DUS4L and BCAP29. The top hit rs4730250 is annotated in intron 3 of the DUS4L gene. Any of the genes at the 7q22 region may confer risk for knee OA, as the LD pattern across the region is high.

The functional analyses support the epidemiological findings but do not exclude any of the 6 candidate genes. Specifically, the zebrafish experiments show that both COG5 and DUS4L are expressed in developing cartilage supporting the notion that either of these genes could have a biological function during chondrogenesis. The studies in the de-differentiation model of human chondrocytes (3D vs 2D culture) show that BCAP29, COG5, DUS4L and HBP1 all have different expression patterns in 3D culture (chondro-like cells) than in 2D culture (de-differentiated cells) suggesting that these four genes may play a role in cartilage metabolism.

A major issue in the field of osteoarthritis is the definition of the disease phenotypes[4, 26]. Different criteria may introduce bias and dilute the effect. The cases in our study were defined either clinically by the presence of a knee replacement or radiographically using the Kellgren/Lawrence (K/L) system. The K/L system is however far from perfect and can be affected by differences in the position of the knee in which the radiographs were obtained, observer biases, interpretation of grading criteria and random error [27, 28]. Similarly there are no standard criteria for replacing knee joints. This may introduce heterogeneity and move the observed effects towards the unity, and so under-estimate the true strength of an association. In our study we synthesized data with a standardized definition of the phenotype, however small individual locus effects with ORs in the range of 1.1-1.2 as for other chronic diseases may well be plausible for knee OA, explaining the paucity of other significant hits despite the reasonable large-scale effort. These findings highlight that even larger collaborative studies and improved standardization of the phenotypes are needed to better understand and identify further genetic variants of OA.

Moreover, even though we were able to accumulate a large sample size, the power of the study to detect very small effect sizes in the range of 1.05-1.15 is inadequate. For example, identification of a GWS signal with an effect size of 1.15 and minor allele frequency of 20%, with 80% power would require almost 7000 additional knee OA cases.

Our results confirm that the 7q22 chromosomal region confers risk for knee OA, which along with our functional work implicates 6 possible genes. Further in depth genetic analysis of the locus, including deep-sequencing of the region and functional work including in vitro assays and animal models will be required to deepen our understanding of the underlying molecular pathways associated with the disease.