Genome-wide association study meta-analysis identifies seven new rheumatoid arthritis risk loci

Abstract

To identify novel genetic risk factors for rheumatoid arthritis (RA), we conducted a genome-wide association study (GWAS) meta-analysis of 5,539 autoantibody positive RA cases and 20,169 controls of European descent, followed by replication in an independent set of 6,768 RA cases and 8,806 controls. Of 34 SNPs selected for replication, 7 novel RA risk alleles were identified at genome-wide significance (P<5×10⁻⁸) in analysis of all 41,282 samples. The associated SNPs are near genes of known immune function, including IL6ST, SPRED2, RBPJ, CCR6, IRF5, and PXK. We also refined the risk alleles at two established RA risk loci (IL2RA and CCL21) and confirmed the association at AFF3. These new associations bring the total number of confirmed RA risk loci to 31 among individuals of European ancestry. An additional 11 SNPs replicated at P<0.05, many of which are validated autoimmune risk alleles, suggesting that most represent bona fide RA risk alleles.

RA is a common autoimmune disease that affects up to 1% of the general adult population worldwide¹. Approximately two-thirds of patients are seropositive for rheumatoid factor (RF) or anti-cyclic citrullinated peptide (anti-CCP) autoantibodies². Genetic studies in autoantibody positive RA among subjects of European ancestry have identified multiple major histocompatibility complex (MHC) region alleles and 25 confirmed RA risk alleles in 23 non-MHC loci^3–15. These alleles explain about 23% of the genetic burden of RA¹¹, indicating that additional alleles remain to be discovered.

To identify novel RA risk alleles, we conducted a GWAS meta-analysis of 5,539 autoantibody positive RA cases and 20,169 controls of European ancestry (Table 1). This study expands upon our previous GWAS meta-analysis of 3,393 cases and 12,462 controls¹¹ by (a) adding a new GWAS dataset of 483 RA cases recruited from the Boston area (Brigham Rheumatoid Arthritis Sequential Study, BRASS) and 1,449 shared controls, (b) adding 513 cases and 431 controls recruited from Sweden (Epidemiological Investigation of Rheumatoid Arthritis, EIRA), (c) incorporating a recently published GWAS of 2,418 cases and 4,504 controls recruited from North America (Canada and North American Rheumatoid Arthritis Consortium-III, NARAC-III)¹³, and (d) adding additional shared controls to the NARAC-III dataset. For each of six GWAS case-control collections, we removed SNPs and individuals that failed quality control, matched case-control samples using principal components analysis (PCA) to minimize bias due to stratification, and imputed¹⁶ genome-wide to infer genotypes at additional central European (CEU) HapMap SNPs. We used logistic regression to conduct a GWAS for 2.56 million SNPs in each collection, corrected for residual inflation using genomic control¹⁷, and combined results across collections by inverse variance-weighted meta-analysis (P_GWAS)¹⁸. We found little evidence of systematic bias, as indicated by the genomic control inflation factor λ_GC (1.04)¹⁷, quantile-quantile (Q–Q) plots, results for markers highly differentiated across Europe¹⁹, and comparisons with an alternative analysis using PCA eigenvectors as covariates (Supplementary Figures 1–4; Supplementary Table 1). See Methods and Supplementary Note for full details of the analysis.

Table 1.

Patient collections.

	Case Collection	Control Collection	Geographical Origin	Case Antibody Status	Cases	Controls	Genotyping Platform	Case-Control Stratification Correction
	Brigham Rheumatoid Arthritis Sequential Study (BRASS)	Shared controls	Boston, USA	100% CCP+	483	1449	Affymetrix 6.0	GWAS data PC matching
Meta-analysis	CANADA	CANADA and Shared controls	Toronto, Canada	100% CCP+	589	1472	Illumina 370K	GWAS data PC matching
	Epidemiological Investigation of Rheumatoid Arthritis (EIRA)	EIRA	Sweden	100% CCP+	1173	1089	Illumina 317K	Epidemiologically matched & GWAS data PC matching
5,539 Cases; 20,169 Controls	North American Rheumatoid Arthritis Consortium (NARAC) I	Shared controls	North America	100% CCP+	867	1041	Illumina 550K	GWAS data PC matching
	NARAC III	Shared controls	North America	100% CCP+	902	4510	Illumina 317K	GWAS data PC matching
	Wellcome Trust Case Control Consortium (WTCCC)	Shared controls from WTCCC	United Kingdom	100% RF+ or CCP+	1525	10608	Affymetrix 500K	Geographically matched
	CANADA II	CANADA II	Toronto & Halifax, Canada	100% CCP+	1076	1269	Sequenom iPlex	Geographically matched
Replication	Dutch	Dutch	The Netherlands	100% RF+	718	697	Sequenom iPlex	Geographically matched
	Genetics Network Rheumatology Amsterdam (GENRA)	GENRA	Amsterdam, The Netherlands	100% CCP+	519	1155	Sequenom iPlex	Geographically matched
6,768 Cases; 8,806 Controls	Genomics Collaborative Initiative (GCI)	GCI	North America	100% RF+	461	460	Kinetic PCR	Epidemiologically matched
	Leiden University Medical Center (LUMC)	LUMC	Leiden, The Netherlands	100% RF+ or CCP+	310	544	Kinetic PCR	Geographically matched
	NARAC II	Shared controls	North America	100% CCP+	462	693	Sequenom iPlex	Ancestry informative marker data matching
	United Kingdom Rheumatoid Arthritis Genetics (UKRAG)	UKRAG	United Kingdom	100% RF+ or CCP+	2906	3494	Sequenom iPlex	Geographically matched
	Nurses Health Study (NHS)	NHS	North America	100% RF+ or CCP+	316	494	Biotrove OpenArray	Epidemiologically matched

Meta-analysis of GWAS results for six collections (top panel) was used to identify SNPs for replication in eight collections (bottom panel). For each collection, we list the source of controls, geographic origin, autoantibody status of cases, numbers of cases and controls, genotyping platform (GWAS collection microarray platform, and replication and validation/prediction collection direct-genotyping technology), and the strategy used to correct for casecontrol population stratification. See Methods and Supplementary Note for details.

Our GWAS meta-analysis found support for RA risk loci previously confirmed among individuals of European ancestry, consistent with power for our study design (Table 2, Supplementary Figure 5). Four of the 25 confirmed non-MHC risk alleles obtained P_GWAS<5×10⁻⁸ (PTPN22, CTLA4, TNFAIP3, and CD40 loci); the remaining 21 confirmed alleles obtained P_GWAS≤0.002. We found modest evidence for association at PADI4 (rs2240340, P_GWAS=0.01, odds ratio (OR)=1.06) and FCRL3 (rs3761959, P_GWAS=0.001, OR=1.08), but not CD244 (rs3753389, P_GWAS=0.26, OR=1.03), loci implicated in RA patients of Asian ancestry^3,20,21 (Supplementary Table 1 online).

Table 2.

Previously known RA risk associated SNPs in Europeans.

SNP			GWAS Meta-Analysis				Power
Locus	ID	Gene(s)	Minor allele	MAF	OR (95% CI)	PGWAS	α = 10−6	α = 5×10−8
1p36	rs3890745*	TNFRSF14	C	0.32	0.89 (0.85,0.94)	3.6×10−6	0.45	0.24
1p13	rs2476601	PTPN22	A	0.10	1.94 (1.81,2.08)	9.1×10−74	1	1
1p13	rs11586238	CD2,CD58	G	0.24	1.13 (1.07,1.19)	1.0×10−5	0.46	0.26
1q23	rs12746613*	FCGR2A	T	0.12	1.13 (1.06,1.21)	0.0004	0.11	0.04
1q31	rs10919563*	PTPRC	A	0.13	0.88 (0.82,0.94)	0.0002	0.10	0.03
2p16	rs13031237	REL	T	0.37	1.13 (1.07,1.18)	7.9×10−7	0.67	0.45
2q11	rs10865035*	AFF3	A	0.47	1.12 (1.07,1.17)	2.0×10−6	0.55	0.33
2q32	rs7574865	STAT4	T	0.22	1.16 (1.10,1.23)	2.9×10−7	0.77	0.58
2q33	rs1980422	CD28	C	0.24	1.12 (1.06,1.18)	5.2×10−5	0.32	0.15
2q33	rs3087243	CTLA4	A	0.44	0.87 (0.83,0.91)	1.2×10−8	0.89	0.74
4q27	rs6822844	IL2,IL21	T	0.18	0.90 (0.84,0.95)	0.0007	0.08	0.02
6p21	rs6910071	HLA-DRB1 (*0401 tag)	G	0.22	2.88 (2.73,3.03)	<10−299	1	1
6q21	rs548234	PRDM1	C	0.33	1.10 (1.05,1.16)	9.7×10−5	0.21	0.08
6q23	rs10499194	TNFAIP3	T	0.27	0.91 (0.87,0.96)	0.0007	0.11	0.03
6q23	rs6920220	TNFAIP3	A	0.22	1.22 (1.16,1.29)	8.9×10−13	1	0.99
6q23	rs5029937	TNFAIP3	T	0.04	1.40 (1.24,1.58)	7.5×10−8	0.95	0.86
6q25	rs394581*	TAGAP	C	0.30	0.91 (0.87,0.96)	0.0006	0.13	0.05
8p23	rs2736340	BLK	T	0.25	1.12 (1.07,1.18)	1.5×10−5	0.34	0.17
9p13	rs2812378*	CCL21	G	0.34	1.10 (1.05,1.16)	0.0001	0.21	0.09
9q33	rs3761847	TRAF1,C5	G	0.43	1.13 (1.08,1.18)	2.1×10−7	0.70	0.49
10p15	rs2104286	IL2RA	C	0.27	0.92 (0.87,0.97)	0.002	0.05	0.02
10p15	rs4750316	PRKCQ	C	0.19	0.87 (0.82,0.92)	2.0×10−6	0.41	0.21
11p12	rs540386*	TRAF6	T	0.14	0.88 (0.83,0.94)	0.0003	0.13	0.05
12q13	rs1678542*	KIF5A,PIP4K2C	G	0.38	0.91 (0.87,0.96)	0.0002	0.20	0.08
20q13	rs4810485	CD40	T	0.25	0.85 (0.80,0.90)	2.8×10−9	0.88	0.73
22q12	rs3218253*	IL2RB	A	0.26	1.09 (1.03,1.15)	0.002	0.07	0.02

GWAS meta-analysis results for previously known SNPs associated with RA risk among European populations. Listed are the chromosome, SNP ID, and candidate gene(s) in the region. The minor allele and frequency (positive strand in HapMap release 22, frequency in controls subjects), odds ratio (95% confidence interval), and association P-value are derived from our GWAS meta-analysis. Most SNPs have achieved P<5×10⁻⁸ in combined analysis from previous studies; those with an asterisk (*) have been validated by replication in independent samples, but may not have attained P<5×10⁻⁸ in any single study. SNPs with strong evidence of association in Asians (including PADI4, rs2240340, P_GWAS = 0.01; FCRL3, rs3761959, P_GWAS = 0.001; and CD244, rs3753389, P_GWAS = 0.26) or suggestive evidence of association in Europeans are shown in Supplementary Table 1 online. Power was calculated based on the odds ratios from our meta-analysis at α = 10⁻⁶ (a threshold for selecting SNPs for replication in the current study) and 5×10⁻⁸ (genome-wide signicance). Details of the power calculations can be found in the Supplementary Materials online.

We used three criteria to select 34 independent SNPs for replication after removing confirmed RA risk alleles. First, we selected all 11 SNPs with P_GWAS<10⁻⁶. Second, we selected 9 SNPs (of 116 tested) with P_GWAS<0.0001 and identified as functionally related to known RA risk loci by GRAIL²², a bioinformatic analysis that identifies connections among genes in published abstracts (Supplementary Figure 6 online). Third, we selected 14 SNPs (of 104 tested) with P_GWAS<0.01 and previously known to be associated with other autoimmune diseases^23–31, as we found evidence for enrichment of autoimmune SNPs in our RA GWAS (Supplementary Figure 7 online). See Supplementary Note online for additional details about our selection of SNPs for replication.

These 34 SNPs were genotyped in an independent set of 6,768 autoantibody positive RA cases and 8,806 matched controls of European ancestry (Table 1). As in our GWAS, all cases were seropositive for disease-specific autoantibodies (anti-CCP or RF). For each SNP, we tested for replication of the GWAS meta-analysis association by calculating a one-tailed P-value (P_replication) for the same allele conferring risk in both analyses. Additionally, we conducted overall association tests across all 41,282 samples (GWAS meta-analysis plus replication, P_overall), and considered P_overall<5×10⁻⁸ to be a reproducible level of significance for RA risk association.

Of the 34 SNPs tested, 12 obtained a Bonferroni-corrected P_replication<0.0015 (=0.05/34), and an additional 9 obtained P_replication<0.05. Ten of these achieved genome-wide significance in the combined analysis, P_overall<5×10⁻⁸, indicating that these are validated RA risk alleles (Table 3). The other 11 SNPs had P_replication<0.05, but P_overall>5×10⁻⁸, indicating highly suggestive but not genome-wide significant association (Table 4). Results for all 34 SNPs in replication and combined analyses are shown in Supplementary Table 2 online.

Table 3.

Newly validated RA risk alleles.

SNP						GWAS Meta-Analysis				Replication				Combined
ID	Chr	Position	Gene(s)	Allele		PGWAS	OR	MAF		Preplication	OR (95% CI)	MAF		Poverall	Cochran Q
				Major	Minor			Case	Control			Case	Control		P
Novel RA and novel autoimmune risk loci
rs934734	2p14	65,507,237	SPRED2	A	G	3.2E-07	1.13	0.52	0.49	0.0002	1.13 (1.06,1.21)	0.53	0.51	5.3E-10	0.95
rs6859219	5q11	55,474,337	ANKRD55, IL6ST	C	A	2.5E-09	0.78	0.18	0.21	0.0002	0.85 (0.78,0.93)	0.19	0.22	9.6E-12	0.19
rs26232	5q21	102,624,619	C5orf13	C	T	4.3E-07	0.88	0.29	0.32	0.004	0.93 (0.88,0.98)	0.30	0.32	4.1E-08	0.82
Novel RA, previously implicated autoimmune risk loci†
rs13315591	3p14	58,531,881	PXK	T	C	3.7E-07	1.29	0.10	0.09	0.002	1.13 (1.04,1.23)	0.09	0.08	4.6E-08	0.12
rs874040	4p15	25,784,466	RBPJ	G	C	1.9E-07	1.14	0.33	0.30	3.0E-11	1.18 (1.12,1.24)	0.34	0.30	1.0E-16	0.57
rs3093023	6q27	167,504,701	CCR6	G	A	3.3E-07	1.13	0.47	0.43	4.5E-06	1.11 (1.06,1.16)	0.46	0.43	1.5E-11	0.42
rs10488631	7q32	128,188,134	IRF5	T	C	2.8E-06	1.19	0.13	0.11	1.2E-06	1.25 (1.14,1.37)	0.13	0.10	4.2E-11	0.60
Associations at known RA risk loci‡
rs11676922	2q11	100,265,458	AFF3	A	T	6.9E-07	1.12	0.49	0.46	1.1E-09	1.15 (1.10,1.20)	0.48	0.45	1.0E-14	0.50
rs951005	9p13	34,733,681	CCL21	A	G	5.4E-07	0.84	0.14	0.16	6.7E-05	0.87 (0.81,0.93)	0.13	0.15	3.9E-10	0.61
rs706778	10p15	6,138,955	IL2RA	C	T	7.9E-08	1.14	0.44	0.40	1.5E-05	1.11 (1.06,1.17)	0.43	0.40	1.4E-11	0.36

GWAS, replication and combined meta-analysis results for SNPs that achieve genome-wide significance for association with RA risk. Listed are the rs ID for each SNP, the chromosomal/cytological band, position in human genome build 36, candidate genes in the region (selected by GRAIL analysis, or manually based on immunological function; see text and Supplementary Note), and major and minor alleles (positive strand in HapMap release 22, major/minor based on frequency in GWAS controls). The association P-value, odds ratio (OR) with respect to minor allele (95% confidence interval for replication analysis) and minor allele frequencies (MAF) in cases and controls are listed for our GWAS and replication analyses. For the combined analysis, the overall association P-value and the P-value for Cochran’s Q test for heterogeneity are listed.

^†: Implicated in other autoimmune diseases: PXK is associated with SLE (rs6445975, r² = 0.15 with rs13315591); RBPJ is associated with T1D (rs10517086, r² = 1 with rs874040); CCR6 is associated with Crohn’s (rs2301436, r² = 0.48 with rs3093023); and IRF5 SNP rs10488631 is associated with SLE.

^‡: Previously implicated in rheumatoid arthritis: AFF3 SNP rs11676922 has recently been reported (Barton et al 2009); CCL21 SNP rs2812378 (r² = 0.06 with rs951005); and IL2RA SNP rs2104286 (r² = 0.25 with rs706778).

Table 4.

Suggestive RA risk alleles.

SNP						GWAS Meta-Analysis				Replication				Combined
ID	Chr	Pos (HG18)	Gene(s)a	Allele		PGWAS	OR	MAF		Prep.	OR (95% CI)	MAF		Poverall	Cochran Q
				Major	Minor			Case	Control			Case	Control		P
rs7543174a	1q21	151,340,745	IL6R	T	C	7.9E-05	1.13	0.18	0.16	0.01	1.07 (1.01,1.13)	0.19	0.18	1.2E-05	0.06
rs840016a	1q24	164,140,328	CD247	C	T	3.6E-05	0.90	0.39	0.42	0.006	0.92 (0.86,0.98)	0.38	0.40	1.6E-06	0.62
rs13119723b	4q27	123,575,918	IL2, IL21†	A	G	0.001	0.89	0.13	0.15	6.7E-05	0.87 (0.81,0.93)	0.15	0.17	6.8E-07	0.46
rs11594656b	10p15	6,162,015	IL2RA†	T	A	0.0002	0.90	0.23	0.25	0.04	0.95 (0.90,1.00)	0.24	0.25	0.0001	0.86
rs2793108b	10p11	31,419,111	ZEB1	T	C	0.002	0.93	0.40	0.43	0.001	0.93 (0.89,0.98)	0.41	0.43	1.4E-05	0.73
rs3184504b	12q24	110,347,328	SH2B3	T	C	0.004	0.93	0.49	0.49	0.0002	0.92 (0.88,0.96)	0.48	0.49	6.0E-06	0.08
rs7155603a	14q24	75,030,289	BATF†	A	G	1.0E-05	1.16	0.21	0.19	0.001	1.12 (1.04,1.20)	0.23	0.21	1.1E-07	0.74
rs8045689a	16p11	28,895,770	CD19, NFATC2IP†	T	C	5.3E-05	1.14	0.32	0.30	0.01	1.06 (1.01,1.12)	0.32	0.30	2.4E-05	0.35
rs2872507b	17q12	35,294,289	IKZF3†	G	A	4.7E-05	1.10	0.49	0.47	0.002	1.08 (1.02,1.14)	0.48	0.46	9.4E-07	0.69
rs11203203b	21q22	42,709,255	UBASH3A†	G	A	2.5E-05	1.11	0.39	0.37	0.02	1.07 (1.00,1.14)	0.38	0.37	3.8E-06	0.49
rs5754217b	22q11	20,264,229	UBE2L3	G	T	0.0007	1.10	0.22	0.19	0.01	1.07 (1.01,1.13)	0.22	0.21	4.8E-05	0.79

GWAS, replication and combined meta-analysis results for SNPs with highly suggestive associations with RA risk (defined as P<0.05 in our replication samples). As in Table 2, listed for each SNP are the rs ID, chromosomal location and position, candidate gene(s), and major and minor alleles, GWAS and replication analysis P-value, odds ratio and case/control minor allele frequencies, and the combined analysis P-values for association and for Cochran’s Q test for heterogeneity.

^a: Selected for replication based on GRAIL P_text score

^b: Selected for replication based on P_GWAS < 0.01 and validated autoimmune disease association: rs13119723, IL2/IL21, RA and Celiac; rs11594656, IL2RA, T1D and MS; rs2793108, ZEB1, T1D; rs3184504, SH2B3, Celiac and T1D; rs2872507, IKZF3, Crohn’s; rs11203203, UBASH3A, T1D; rs5754217, UBE2L3, SLE; the ZEB1 SNP was from the May 2009 release of the online T1D database (www.T1Dbase.org).

^†Autoimmune disease associations for SNPs other than those tested for replication in the present study (see text, Fig. 2 and Supplementary Note): CD247, Crohn’s; IL2/IL21, Celiac, RA and T1D; IL2RA, MS, RA and T1D; BATF, T1D; CD19/NFATC2IP, T1D; IKZF3, T1D; UBASH3A, Crohn’s and T1.

Three of the 10 newly validated RA risk loci have not been implicated in any genetic studies of RA or other autoimmune disease. The SNPs are located at chromosome 2p14 (rs934734, P_overall=5.3×10⁻¹⁰), 5q11 (rs6859219, P_overall=9.6×10⁻¹²) and 5q21 (rs26232, P_overall=4.1×10⁻⁸). Replication sample odds ratios (for minor alleles) were 1.13, 0.85 and 0.93, respectively (Table 3).

Although no single gene can be declared causal, we labeled RA risk loci with the names of the most compelling candidate gene(s) from each region of linkage disequilibrium (LD) based upon GRAIL analysis and/or knowledge of RA pathogenesis^2,22. At 2p14, the most significant SNP (rs934734) is located within intron 1 of the sprouty-related, EVH1 domain containing 2 (SPRED2) gene (Figure 1A), which has been shown to regulate CD45+ hematopoietic cells via the Ras/MAP kinase pathway³². At 5q11, the most significant SNP (rs6859219) is located in ANKRD55, an ankyrin repeat domain-containing gene of unknown function (Figure 1B). A more compelling immunological candidate, interleukin 6 signal transducer (IL6ST), lies ~150 kilobases (kb) proximal but outside the region of LD with associated SNPs. The IL6ST protein product, gp130, functions as a part of the receptor complex for the inflammatory cytokine IL6³³. At 5q21, there is no obvious biological candidate gene; SNP rs26232 lies within the intron of predicted gene C5orf30 (Figure 1C).

Figure 1. Association with RA risk across 4 loci.

Regional association plots show strength of association (-Log(P-value)) versus chromosomal position (kilobases, kb) for all SNPs across 1 Megabase (Mb) regions centered on the newly validated SNPs (labeled). P_GWAS values are plotted with red/white diamonds for all SNPs, shaded by the degree of LD (r², see legend) with the validated SNP (larger red diamond). P_overall in combined analysis of GWAS and replication collections is plotted with the blue diamond. Local recombination rates estimated from HapMap CEU (centi-Morgans per Mb, blue line) are plotted against the secondary y-axis, showing recombination hotspots across the region. Labeled green arrows below the plots indicate genes and their orientations. (a) 2p14, SPRED2 locus. (b) 5q11, IL6ST-ANKRD55 locus. (c) 5q21, C5orf13 locus. (d) 10p15, IL2RA locus.

Our study provides the first convincing evidence that 4 loci implicated in other autoimmune diseases are also associated with risk of RA. Of these, three of the 4 SNPs were selected for replication based on obtaining P_GWAS<10⁻⁶, regardless of their previously reported associations with autoimmune disease. These 3 SNPs are located at chromosome 3p14 (rs13315591, near PXK, P_overall=4.6×10⁻⁸), 4p15 (rs874040, near RBPJ, P_overall=1.0×10⁻¹⁶) and 6q27 (rs3093023, in CCR6, P_overall=1.5×10⁻¹¹). The 3p14 SNP lies 187 kb proximal to a SNP associated with systemic lupus erythematosus (SLE)²⁵; the SLE SNP, intronic in the PXK gene, is only weakly associated with RA risk in our study (rs6445975, P_GWAS = 0.03, r² = 0.15 and D’ = 0.75 with rs13315591). The 4p15/RBPJ SNP is in complete LD (r² = 1) with a SNP associated with risk of type 1 diabetes (T1D, rs10517086)²³, and the same allele confers risk in both diseases. The 6q27/CCR6 SNP rs3093023 is in LD with a SNP associated with Crohn’s disease²⁴, which is only weakly associated with RA risk in our study (Crohn’s SNP rs2301436, P_GWAS = 0.045, r² = 0.48 and D’ = 0.80).

The IRF5 SNP rs10488631 (P_GWAS =2.8×10⁻⁶), chosen because of its association in SLE^34,35, was convincingly associated with autoantibody positive RA in our study (P_replication=1.2×10⁻⁶, P_overall=4.2×10⁻¹¹). Graham et al.³⁶ proposed that the complex SLE risk associations in IRF5 are explained by three independent groups of SNPs, each consisting of SNPs in tight LD with each other. In our dataset, these groups are represented by rs10488631 (group 1), rs729302 (group 2) and rs4728142 (group 3). In addition to the group 1 SNP, rs10488631, we found evidence for association with group 3 SNP rs4728142 (P_GWAS=7.2×10⁻⁶), which is in LD with a variant that alters IRF5 polyadenylation and expression³⁶. However, we found no evidence for association for group 2 (rs729302, P_GWAS=0.15). Conditional analyses indicated that the group 1 and group 3 effects are independent (Supplementary Table 3 online). We note that group 3 SNP rs3807306 has been suggested to be associated with autoantibody negative RA³⁷, and our results suggest that this SNP is associated with autoantibody positive RA (P_GWAS=4.5×10⁻⁶).

Our GWAS meta-analysis identified new alleles exhibiting independent effects from previously implicated RA risk alleles at two other loci, CCL21¹¹ (rs951005, P_overall=3.9×10⁻¹⁰) and IL2RA¹⁵ (rs706778, P_overall=3.3×10⁻¹¹). Our new CCL21 SNP has little LD with the previous RA risk-associated SNP rs2812378¹¹ (P_GWAS=1.6×10⁻⁷; r²=0.06 and D’=0.83). Conditional analyses indicated that the associations are partly independent (Supplementary Table 3 online); the new SNP rs951005 remains significant (P = 1.7×10⁻⁵) conditional on the previously known SNP rs2812378; conditional on rs951005, rs2812378 remains nominally significant (P = 0.01).

SNPs at IL2RA are known to be associated with multiple autoimmune diseases including RA¹⁵, T1D²³ and multiple sclerosis (MS)^38,39. The IL2RA SNP identified in our GWAS (rs706778, P_overall=3.3×10⁻¹¹) is in partial LD with a SNP previously implicated in RA, T1D and MS (rs2104286, r²=0.25 and D’=1)¹⁵ and another SNP implicated in T1D and MS (rs11594656; r²=0.24 and D’=1)²³. In fact, T1D risk is conferred by a haplotype tagged by rs2104286(T)-rs11594656(T)³⁸. In our GWAS dataset, rs706778 correlated strongly with the T1D risk haplotype (r²=0.83), and conditional and haplotype analyses demonstrated that risk of RA is better explained by this haplotype [rs2104286(T)-rs11594656(T)] than by rs706778 (Supplementary Tables 3 and 4 online).

A SNP at the AFF3 locus has been previously implicated in RA¹⁴, with equivocal evidence for association with T1D^23,40. Our study provides definitive evidence that this locus is associated with risk of autoantibody positive RA (rs11676922, P_overall=1.0×10⁻¹⁴, OR=1.14).

Eighteen of the 34 SNPs that we tested in replication (based on the criteria above) are in LD with validated autoimmune risk alleles (Figure 2). Of these, 5 demonstrated genome-wide (P_overall<5×10⁻⁸) and 7 showed suggestive (P_replication<0.05) evidence of association in our study; only 6 showed no evidence of association in replication. While not meant as a complete comparison of all known RA and other autoimmune disease risk alleles, these results underscore the emerging overlap in genetic bases of RA and other autoimmune diseases^41,42. Although additional replication in large sample collections is required to confirm the suggestive associations, many likely represent true RA risk alleles.

Figure 2. Previously validated autoimmune SNPs tested in our replication study.

Eighteen SNPs tested in our replication samples were in LD (defined as r² > 0.3) with a validated autoimmune risk allele. Of these, 5 were validated as RA risk alleles in our study (P_overall < 5×10⁻⁸, inner most circle), 6 were suggestive associations (P_replication <0.05 but P_overall > 5×10⁻⁸), and 6 demonstrated no evidence of association in our replication samples (P_replication ≥0.05). For the 12 SNPs with suggestive or no evidence of association, each SNP is plotted by the strength of association with RA risk in the replication samples; those closer to the inner circle have more significant P_replication. All of the RA risk alleles confer risk in the same direction as the validated autoimmune risk alleles (when the same allele or a near perfect proxy was tested). We include the following as “autoimmune” diseases in our study, listed on the outside of the circle, although these reflect diseases along the autoimmuneinflammatory spectrum: systemic lupus erythematosus (SLE), celiac disease, Crohn’s disease, multiple sclerosis (MS), psoriasis, and type 1 diabetes (T1D); other autoimmune diseases are not included (e.g., autoimmune thyroiditis). The haplotype tagged by the IL2RA SNP, rs706778, is associated with T1D and MS^38,39; the SH2B3 SNP is associated with both T1D and Celiac disease^23,30. The CCR6 SNP rs3093023 is in partial LD with a SNP associated with Crohn’s disease (rs2301436, r² = 0.48)²⁴. The AFF3 SNP has an equivocal association with T1D (where the associated SNP is rs9653442⁴⁰). The IL2-IL21 SNP tested in our study, rs13119723, is in LD (r²=0.67) with a SNP previously implicated in both Celiac disease and RA (rs6822844)^10,29, but only in partial LD (r²=0.09) with a T1D SNP (rs4505848)²³. The CD19-NFATC2IP SNP tested in our study, rs8045689, was selected because of GRAIL; it is in partial LD (r²=0.38) with a SNP associated with T1D (rs4788084)²³. The TNIP1 SNP in our study, rs6889239, is in strong LD with an SLE SNP (rs7708392, r²=0.91)⁴⁹, but not in LD with another TNIP1 SNP associated with psoriasis (rs17728338, r²<0.01)²⁶. The ZEB1 SNP (rs2793108) was from the May 2009 release of T1D base, although this SNP did not appear in a subsequent publication²³. The PXK SLE SNP was tested in this study but not shown, as it is in weak LD with the RA risk SNP (r² = 0.15 between rs6445975 and rs13315591); the RA SNP was selected because of P_GWAS<10⁻⁶, not because of its association with another autoimmune disease. Note that there are SNPs associated with RA and other autoimmune diseases not shown; we only include those SNPs tested as part of the current study.

Several of the SNP associations further implicate T-cells in RA pathogenesis. RBPJ (recombination site binding protein J; also known as CSL) is a transcription factor within the Notch signaling pathway. RBPJ-deficient mice have no T-cell development, while early B-cell development is maintained⁴³. CCR6 is a cell surface protein that distinguishes Th17 cells from other CD4+ helper T-cells⁴⁴. Synoviocytes from arthritic joints of mice and RA patients produce CCL20, a CCR6 ligand. Intriguingly, anti-CCR6 monoclonal antibodies substantially inhibit mouse arthritis, suggesting that CCR6 could be a therapeutic target in RA⁴⁴. CD247 (P_replication=0.006, P_overall=1.6×10⁻⁶; Table 3) encodes the T-cell receptor zeta chain, a subunit of the T-cell receptor-CD3 complex that when mutated causes an inflammatory arthritis in mouse^45,46. The zeta chain plays an important role in coupling antigen recognition to several intracellular signal-transduction pathways. These results build upon knowledge of T-cell activation-differentiation in RA pathogenesis evidenced by genetic associations at HLA-DRB1 (presents antigens to T-cells), PTPN22 (alters T-cell thresholds), CTLA4 (co-stimulator for T-cell activation), IL2RA/CD25 (mediates IL2-dependent T-cell responses), and STAT4 (transcription factor in Th1 cell differentiation), among others.

In conclusion, we find convincing evidence for association with risk of autoantibody positive RA at 10 loci, seven of which represent novel risk loci for RA. We estimate that the now >30 validated non-MHC RA risk alleles explain 3.9% of the total disease variance, with 0.67% due to the new RA risk alleles reported here. It is clear that additional risk alleles remain to be identified, as genetic discoveries explain only 16% of disease variance (including an estimated 12% for the MHC⁴⁷), whereas more than half the risk of autoantibody positive RA is thought to be genetic^47,48. The number of SNPs with suggestive evidence of association in our study (Table 4) further indicates that many more common risk alleles with modest effect sizes remain to be discovered. In addition to common variants, the roles of rare variants, copy number variants and epigenetic modifications will need to be explored with newer genomic technologies.

METHODS

Sample collections

Case-control collections are listed in Table 1 and described in detail in our previous studies^6,7,11,13,19. Collections were composed entirely of individuals of selfdescribed European ancestry, and all cases either met the 1987 American College of Rheumatology (ACR) criteria for diagnosis of RA⁵⁰ or were diagnosed by board-certified rheumatologists. All RA cases were further limited to anti-CCP+ patients, or RF+ patients if anti-CCP status data were missing. The BRASS RA samples have been used in our previous 100K GWAS⁶, but here are presented for the first time genotyped with the Affymetrix 6.0 array. In the current study, controls were matched to BRASS RA cases using principal components analysis from GWAS data from three separate studies: controls from a multiple sclerosis GWAS³¹, controls from an age-related macular degeneration GWAS⁵¹, and controls from a myocardial infarction GWAS⁵². WTCCC collection controls included the 1958 birth and National Blood Service cohorts as well as non-autoimmune disease cases (bipolar disorder, cardiovascular disease, hypertension and type 2 diabetes)¹⁵. All GWAS collections except WTCCC were restricted to control subjects matching cases using principal components analysis of GWAS data. All replication sample collections were composed of epidemiologically and/or geographically matched cases and controls, except NARAC II collection, which was case-control matched based on genotypes at a set of ancestry-informative markers as previously described¹¹. The eight replication samples included: (1) CCP positive cases and controls from Halifax and Toronto (CANADA-II)¹³; (2) RF positive Dutch cases from Groningen and Nijmegen, together with geographically matched controls⁵³; (3) CCP positive Dutch cases and controls collected from the greater Amsterdam region (GENRA)⁵⁴; (4) North American RF positive cases and controls matched on gender, age, and grandparental country of origin from the Genomics Collaborative Initiative (GCI)⁴; (5) CCP or RF positive Dutch cases and controls from Leiden University Medical Center (LUMC)⁵; (6) CCP positive cases drawn from North American clinics and controls from the New York Cancer Project (together this collection is called NARAC-II)⁷; (7) CCP or RF positive cases recruited at multiple sites in the United Kingdom by the United Kingdom Rheumatoid Arthritis Genetics (UKRAG) collaboration⁹; and (8) CCP or RF positive cases identified by chart review from the Nurses Health Study (NHS) and matched controls based on age, gender, menopausal status, and hormone use⁵⁵. We used available SNP data from this and previous studies to identify genetically identical samples from the same country¹³; we assumed these represented duplicated individuals and removed them. Institutional review boards at each collection site approved the study, and all individuals gave their informed consent.

Genotyping

The BRASS GWAS collection was genotyped on the Affymetrix Genechip 6.0 platform at the Broad Institute (Boston, USA). All other GWAS collections were genotyped as previously described^7,13,19. Genotype data for GWAS samples from RA and other disease studies were obtained with permission from the investigators and/or disease consortia. Additional shared control GWAS genotype data were obtained from the NIMH (USA) through a formal application and approval process (part of the BRASS collection), and from the Illumina iControls database (NARAC III). For each GWAS collection, quality control (QC) was implemented in cases and in each control cohort separately, and then again in the merged collection data. QC steps included filtering SNPs and individuals with >5% missing data, followed by filtering SNPs with minor allele frequency (MAF) <1% and a Chi-squared test of Hardy Weinberg equilibrium P_HWE<10⁻⁶. For the WTCCC collection, genotyped on the older Affymetrix 500K platform, we implemented more stringent QC (>1% missing data, MAF<1%, P_HWE<10⁻⁵). We then used individual-pairwise identity-by-state estimates to remove occasional related and potentially contaminated samples. Data processing and QC were performed in PLINK⁵⁶. Additional details are described in the Supplementary Note online.

The 34 SNPs chosen for replication, as well as proxy SNPs, were directly genotyped in each of eight collections (Table 1 and Supplementary Note online). Canada samples were genotyped on the Sequenom iPlex platform at University of Toronto, Mount Sinai Hospital and University Health Network (Toronto, Canada); Dutch and GENRA samples were genotyped on the Sequenom iPlex platform at the Broad Institute; UKRAG samples were genotyped on the Sequenom iPlex platform at The University of Manchester, UK; GCI and LUMC samples were genotyped by kinetic PCR at Celera (Alameda California, USA); NARAC II samples were genotyped on the Sequenom iPlex platform at the NIH (USA); and NHS samples were genotyped on the Biotrove OpenArray platform at the Channing Laboratory, Harvard Medical School (Boston, USA). QC filter criteria for SNPs in each replication or validation collection were 10% missing data, 1% MAF and P_HWE<10⁻³. If a given SNP failed in genotyping or QC in a collection, a proxy SNP (r²>0.8) with the least missing data (if available) was used. See Supplementary Note online for details.

Genome-wide association analyses

To address population stratification and remove outliers in our GWAS for BRASS, Canada, EIRA, NARAC I, and NARAC III, we performed principal component (PC) analysis using EIGENSTRAT⁵⁷. For BRASS, Canada, NARAC I, and NARAC III we further removed poorly matched controls based on case-control Euclidean distances calculated from five PCs¹¹. Once matched, imputation was conducted on GWAS genotype data for each GWAS collection separately, using the IMPUTE software¹⁶ and haplotype-phased HapMap Phase 2 European CEU founders as a reference panel. Imputation yielded posterior genotype probabilities as well as imputation quality scores at SNPs not genotyped with a minor allele frequency >=1% in HapMap CEU.

We conducted logistic regression analysis for each SNP in each GWAS collection, to estimate regression coefficients β and z-scores for allele counts (additive model), which were then genomic-control corrected¹⁷. λ_GC values for genotyped SNPs only versus all SNPs (Supplementary Table 5 online) verified that logistic regression controlled for any deflation in the distribution of association test results¹⁸. We then conducted a meta-analysis to combine results across datasets for 2,554,714 SNPs with high quality genotype data in one or more collections (see Supplementary Note online) by summing inverse variance-weighted β and z-scores¹⁸, and again genomic-control corrected our results. We also conducted Cochran’s Q test for heterogeneity across collections, using the β coefficients for each collection for which results were available for a given SNP. Detailed description of all analyses and results is provided in Supplementary Note online.

Replication analysis

Replication and combined analyses followed the GWAS meta-analysis: after matching in the NARAC II collection (Table 1)¹¹, and removing spurious duplicate samples¹², logistic regression was used to test for association and inverse variance-weighted z-scores were summed across collections. Replication association tests were one-tailed, for the same allele being risk or protective as in the GWAS meta-analysis.