Identification of multiple risk variants for ankylosing spondylitis through high-density genotyping of immune-related loci

Abstract

Ankylosing spondylitis is a common, highly heritable inflammatory arthritis affecting primarily the spine and pelvis. In addition to HLA-B*27 alleles, 12 loci have previously been identified that are associated with ankylosing spondylitis in populations of European ancestry, and 2 associated loci have been identified in Asians. In this study, we used the Illumina Immunochip microarray to perform a case-control association study involving 10,619 individuals with ankylosing spondylitis (cases) and 15,145 controls. We identified 13 new risk loci and 12 additional ankylosing spondylitis–associated haplotypes at 11 loci. Two ankylosing spondylitis–associated regions have now been identified encoding four aminopeptidases that are involved in peptide processing before major histocompatibility complex (MHC) class I presentation. Protective variants at two of these loci are associated both with reduced aminopeptidase function and with MHC class I cell surface expression.

Inflammatory arthritis in ankylosing spondylitis causes pain and stiffness and progressively leads to new bone formation and ankylosis (fusion) of affected joints. It affects 0.55% of populations of European ancestry (herein termed Europeans)¹ and 0.23% of Chinese², but is uncommon in Africans and Japanese, mostly owing to the low prevalence in these ancestry groups of HLA-B*27, the major genetic variant associated with ankylosing spondylitis. Whereas effective treatments are available that suppress inflammation and improve symptoms, there are not yet any treatments that have been shown to robustly slow the rate of ankylosis or induce disease remission.

Ankylosing spondylitis is highly familial (sibling recurrence risk ratio of >52)³ and heritable (h² > 90%)⁴. It is two to three times more prevalent in men than in women, and men tend to be more severely affected. More than 80% of cases are positive for the HLA-B*27 allele, but only a minority of HLA-B*27 carriers develop ankylosing spondylitis (1–5%). The low proportion of HLA-B*27 carriers who develop ankylosing spondylitis reflects the fact that numerous other non–HLA-B*27 variants are likely to influence disease susceptibility³. In addition to HLA-B*27, 12 loci have previously been confirmed to be associated with ankylosing spondylitis in Europeans (ANTXR2, CARD9, ERAP1, IL12B, IL23R, KIF21B, PTGER4, RUNX3, TBKBP1, TNFRSF1A and chromosomes 2p15 and 21q22)^5–7, and 2 loci have recently been reported in Han Chinese (HAPLN1-EDIL3 and ANO6)⁸.

The Immunochip Consortium has developed a custom microarray SNP genotyping chip (the Immunochip), the design of which has been informed by available genome-wide association study (GWAS) and deep sequencing data from various autoimmune and inflammatory diseases to provide a cost-effective platform for immunogenetic studies^9,10. Genetic data from ankylosing spondylitis, psoriasis, Crohn's disease and ulcerative colitis, along with several classic autoimmune diseases, were used in the chip design, making it a powerful platform for studies of pleiotropic genetic effects in these related diseases. In this study, we aimed to identify new associations with ankylosing spondylitis and to dissect and refine the boundaries of known associated loci by performing a dense SNP genotyping study in 10,619 cases and 15,145 controls of European, East Asian and Latin American ancestry using the Immunochip.

Results

Primary association findings

After all sample quality control filters, the European cohort consisted of 9,069 cases and 13,578 controls, and the East Asian cohort consisted of 1,550 cases and 1,567 controls. The genomic inflation factor (λ) calculated using 1,922 SNPs included on the Immunochip from studies of reading and writing ability, psychosis and schizophrenia was 1.047 (λ₁₀₀₀ for an equivalent study of 1,000 cases and 1,000 controls = 1.0285), indicating minimal evidence of residual population stratification in the overall data set (quantile-quantile plots are presented in Supplementary Fig. 1).

Association at genome-wide significance (P < 5 × 10⁻⁸) was observed for 25 loci, including the MHC (Table 1 and Supplementary Fig. 2; genomic control–corrected results are shown in Supplementary Table 1). Suggestive association (P < 5 × 10⁻⁷) was observed at six additional loci (Supplementary Table 2). As with all GWAS, there is uncertainty as to the genes contributing to association at specific loci. At previously reported loci, association (P < 5 × 10⁻⁸) was seen with the most strongly associated previously reported SNPs at CARD9, ERAP1, IL12B, IL23R, KIF21B, RUNX3, NPEPPS-TBKBP1-TBX21, TNFRSF1A and chromosomes 2p15 and 21q22 (Table 1). At PTGER4, the previously associated SNP (rs10440635) also showed moderate association in the current study (P = 3.0 × 10⁻⁵; imputed). No SNPs at the ANTXR2 locus were included on the Immunochip.

Table 1. Non-MHc associations with ankylosing spondylitis susceptibility.

							Europeans			East Asians
SNP	Chr.	Positiona	Nearby gene(s)	Combined P	Risk/non-risk allele	Combined OR	RAF (case/control)	OR	P	RAF (case/control)	OR	P
Loci previously associated with ankylosing spondylitis at genome-wide significance
rs6600247	1p36	25177701	RUNX3	2.6 × 10−15	C/T	1.15	0.540/0.501	1.16	1.3 × 10−14	0.731/0.708	1.12	0.047
rs11209026	1p31	67478546	IL23R	2.0 × 10−27	G/A	1.62	0.959/0.934	1.65	6.0 × 10−28	1.000/1.000	NA	NA
rs41299637	1q32	199144473	GPR25-KIF21B	1.9 × 10−15	T/G	1.19	0.757/0.715	1.20	7.0 × 10−16	0.999/0.998	1.05	0.42
rs6759298	2p15	62421949	Intergenic	4.9 × 10−47	C/G	1.29	0.447/0.378	1.31	3.6 × 10−41	0.437/0.374	1.28	1.6 × 10−6
rs12186979	5p13	40560617	PTGER4	4.3 × 10−6	G/A	1.08	0.516/0.498	1.09	5.4 × 10−6	0.202/0.191	1.06	0.26
rs30187	5q15	96150086	ERAP1	4.4 × 10−45	T/C	1.29	0.405/0.338	1.32	1.3 × 10−41	0.542/0.486	1.36	2.0 × 10−5
rs6871626	5q33	158759370	IL12B	3.1 × 10−8	A/C	1.10	0.360/0.337	1.12	6.0 × 10−8	0.324/0.310	1.08	0.17
rs1128905	9q34	138373660	CARD9	7.0 × 10−9	C/T	1.10	0.529/0.503	1.12	1.6 × 10−9	0.316/0.315	1.00	0.99
rs1860545	12p13	6317038	LTBR-TNFRSF1A	2.8 × 10−10	C/T	1.13	0.634/0.605	1.13	8.3 × 10−10	0.862/0.851	1.07	0.21
rs9901869	17q21	42930205	NPEPPS-TBKBP1-TBX21	6.0 × 10−15	A/G	1.14	0.548/0.516	1.15	2.3 × 10−12	0.667/0.626	1.18	0.002
rs2836883	21q22	39388614	Intergenic	6.5 × 10−17	G/A	1.18	0.768/0.734	1.19	1.8 × 10−14	0.817/0.774	1.30	3.5 × 10−5
New loci associated with ankylosing spondylitis at genome-wide significance
rs4129267	1q21	152692888	IL6R	3.4 × 10−13	C/T	1.14	0.635/0.592	1.18	2.1 × 10−15	0.619/0.620	1.00	0.99
rs1801274	1q23	159746369	FCGR2A	1.4 × 10−9	T/C	1.11	0.487/0.476	1.12	9.9 × 10−10	0.706/0.698	1.04	0.46
rs12615545	2q31	181756697	UBE2E3	1.0 × 10−9	C/T	1.12	0.451/0.421	1.11	2.3 × 10−7	0.710/0.673	1.20	8.5 × 10−4
rs4676410	2q37	241212412	GPR35	9.9 × 10−9	T/C	1.13	0.232/0.209	1.13	2.1 × 10−7	0.346/0.312	1.15	8.4 × 10−3
rs17765610	6q15	90722494	BACH2	5.3 × 10−8	G/A	1.15	0.131/0.118	1.17	3.3 × 10−8	0.017/0.018	1.00	0.96
rs1250550	10q22	80730323	ZMIZ1	1.5 × 10−9	G/T	1.11	0.678/0.652	1.11	5.8 × 10−7	0.583/0.539	1.20	3.9 × 10−4
rs11190133	10q24	101268715	NKX2-3	4.9 × 10−14	C/T	1.15	0.737/0.707	1.18	1.7 × 10−14	0.629/0.617	1.10	0.30
rs11065898	12q24	110346958	SH2B3	4.7 × 10−8	T/C	1.11	0.237/0.216	1.13	1.7 × 10−7	0.348/0.329	1.10	0.082
rs11624293	14q31	87558574	GPR65	1.5 × 10−10	C/T	1.20	0.106/0.087	1.23	1.8 × 10−10	0.158/0.145	1.11	0.14
imm_16_ 28525386	16p11	28525386	IL27-SULT1A1	2.6 × 10−9	A/G	1.11	0.421/0.393	1.11	1.4 × 10−7	0.258/0.232	1.16	0.012
rs2531875	17q11	23172294	NOS2	1.2 × 10−10	G/T	1.12	0.396/0.367	1.12	1.3 × 10−8	0.296/0.256	1.22	4.6 × 10−4
rs35164067	19p13	10386181	TYK2	3.4 × 10−10	G/A	1.14	0.819/0.796	1.16	6.5 × 10−9	0.574/0.549	1.11	0.039
rs7282490	21q22	44440169	ICOSLG	6.2 × 10−9	G/A	1.11	0.411/0.390	1.10	1.4 × 10−6	0.581/0.543	1.18	1.3 × 10−3

Locus plots for reported associations are shown in supplementary Figure 5. Chr., chromosome; RAF, risk allele frequency; NA, not available.

^aNCBI Build 36 human genome coordinates.

We observed little evidence of association with the two previously reported loci in Han Chinese, either in Europeans, East Asians (Chinese, Taiwanese and Koreans) or in the combined data set (P > 0.05)⁸. To increase power for these variants, we genotyped a total of 2,998 East Asian cases and 5,547 East Asian controls. No association was seen (P > 0.05) for either rs4552569 (chromosome 5q14, between HAPLN1 and EDIL3) or rs17095830 (chromosome 12q12, ANO6) (Supplementary Table 3). rs4552569 showed only nominal significance in Europeans (P = 0.02). rs17095830 was not directly typed on the Immunochip, but, in a previous GWAS⁶, no association was observed with this SNP (P > 0.1).

Genome-wide significance was seen at 13 loci not previously known to be associated with ankylosing spondylitis (Table 1). The strongest association at each locus was with a common variant (minor allele frequency (MAF) > 5%), but several associations were also seen with rare variants (MAF < 1%) at these loci, including in the genes CARD9, IL23R, LNPEP and TYK2. Both rare variants in IL23R were nonsynonymous coding variants, whereas the CARD9, TYK2 and LNPEP variants were located at exon-intron boundaries and were predicted to influence splicing.

In total, 24.4% of the heritability of ankylosing spondylitis is now explained: 4.3% from loci other than HLA-B and 20.1% due to HLA-B*27 itself.

IL-23 pathway genes

Genetic studies provided the first evidence that interleukin (IL)-23 is involved in the pathogenesis of ankylosing spondylitis, and variants in several genes involved in the IL-23 proinflammatory cytokine pathway have been shown to be associated with the disease. This study adds to that list, with loci containing TYK2, IL6R and IL27 achieving genome-wide significance.

IL-6 signaling through IL-6R has diverse proinflammatory effects. rs4129267, the most strongly associated IL6R SNP in this study, is also associated with asthma¹¹ but with the opposite direction of association to ankylosing spondylitis. The allele associated with risk of ankylosing spondylitis at this SNP is strongly associated with lower serum concentrations of the soluble form of the IL-6 receptor (sIL-6R), with each allele associated with a 1.4-fold variation in serum IL-6R conc-netrations¹². We found that sIL-6R concentrations varied strongly by rs4129267 genotype, both overall and separately, in cases and controls (Fig. 1). Overall, homozygous carriers of the T allele at rs4129267 had sIL-6R concentrations 73% higher than homozygous carriers of the C allele (28.9 versus 16.7 ng/ml; P = 7.8 × 10⁻¹⁷). This SNP has previously been associated with serum C-reactive protein (CRP) concentrations at genome-wide significance; in the current study, serum CRP concentrations were 30% higher in cases homozygous for the T allele than in cases homozygous for the C allele (19.2 versus 14.8 mg/l).

Figure 1.

IL6R polymorphism alters IL-6R serum concentrations. IL-6R concentrations were determined in cases and controls who were homozygous for either the T or C allele at the rs4129267 SNP. In both cases and controls, individuals homozygous for the C allele showed significantly lower concentrations of circulating IL-6R. *** P < 0.0001; NS, not significant. Bars represent mean ± s.e.m.

At IL23R, we previously identified two independent disease-associated haplotypes⁶ tagged by rs11209026 and rs11209032. Here, after conditioning on rs1 1209026, the strongest association was with rs12141575 (odds ratio (OR) = 1.15; P = 9.4 × 10⁻¹¹); this SNP is in strong linkage disequilibrium (LD) with rs1 1209032 (r² = 0.993) and also with rs1495965, which has been reported to be associated with Behçet's disease^13,14. Behçet's disease is complicated by sacroiliitis resembling ankylosing spondylitis in up to 10% of cases¹⁵.

Six loci showed suggestive association (5 × 10⁻⁸ < P < 5 × 10⁻⁷), including a region on chromosome 2q11 encoding IL1R2 and IL1R1. Conditional analyses showed that there were two separate signals at this locus (Table 2), one in each gene. Both genes encode receptors for the cytokine IL-1, which, among diverse proinflammatory functions, also promotes T helper 17 (T_H17) lymphocyte differentiation.

Table 2. Secondary signals in Europeans at loci known to be associated with ankylosing spondylitis.

Chr.	SNP	Position	Nearby gene(s)	Risk/non-risk allele	Conditional SNP	P	OR	RAF (case/control)	LD (r2/D′) with conditional SNP
1p31	rs12141575	67520024	IL23R	A/G	rs11209026	9.4 × 10−11	1.15	0.370/0.330	0.034/0.983
1q23	rs2039415	159121069	FCGR2A	C/T	rs1801274	7.4 × 10−5	1.09	0.702/0.682	0.002/0.062
2q12	rs2192752	102135805	IL1R2-IL1R1	C/A	rs4851529	4.1 × 10−6	1.11	0.239/0.222	0.007/0.192
5q15	rs10045403	96173489	ERAP1-ERAP2	A/G	rs30187 and rs2910686	5.8 × 10−14	1.20	0.783/0.730	0.178/0.958 0.091/0.429
5q15	rs2910686	96278345	ERAP1-ERAP2	C/T	rs30187	4.5 × 10−17	1.17	0.450/0.440	0.153/0.617
5q33	rs6556416	158751323	IL12B	C/A	rs6871626	4.4 × 10−6	1.11	0.704/0.675	0.048/0.443
6q15	rs639575	91047852	BACH2	A/T	rs17765610	8.6 × 10−5	1.08	0.624/0.609	0.000/0.042
12p13	rs7954567	6361386	LTBR-TNFRSF1A	A/G	rs1860545	1.2 × 10−7	1.11	0.363/0.341	0.002/0.068
16p11	rs35448675	28236248	IL27-SULT1A1	A/G	imm_16_28525386	2.4 × 10−4	1.24	0.007/0.006	0.003/0.955
17q11	rs2297518	23120724	NOS2	A/G	rs2531875	6.3 × 10−7	1.13	0.212/0.190	0.000/0.048
19p13	rs6511701	10486067	TYK2	A/C	rs35164067	1.4 × 10−4	1.10	0.220/0.218	0.162/0.419

Aminopeptidase genes

We previously identified a strong association between ERAP1 and ankylosing spondylitis⁵ that is restricted to HLA-B*27–positive disease⁶. This observation was replicated here (Table 3 and Supplementary Figs. 3 and 4), with strong interaction observed in the European cases between HLA-B*27 and both independently associated ERAP1 haplotype–tagging SNPs (rs30187: β = 0.390, P = 2.9 × 10⁻¹¹; rs10045403: β = −0.282, P = 9.3 × 10⁻⁶; Supplementary Table 4). We observed no evidence of epistasis between HLA-B*27 and rs30187 (or other SNPs in ERAP1) in the East Asian samples, probably because of inadequate statistical power. In this analysis, we had 35% power to detect an association at a significance level of P < 0.05 with the observed allele frequency of rs30187 in controls and the effect size observed in the overall analysis. The low power to detect a major effect at rs30187 in HLA-B*27–positive East Asian samples suggests that the power to detect an interaction would be quite small. The observed OR (95% confidence interval (CI)) in HLA-B*27–positive subjects at rs30187 was 0.73 (0.53–1.03) compared to 1.20 (0.89–1.61) in HLA-B*27–negative subjects. Although these analyses did not detect significant interaction (P > 0.05), they are consistent with an interaction between rs30187 and HLA-B*27, where rs30187 is only associated in HLA-B*27–positive disease. This finding warrants further exploration with larger sample sizes in East Asian populations.

Table 3. Association analysis of rs30187 and rs2910686 haplotypes in samples positive and negative for HLA-B*27.

Sample	SNP (ERAP1)	Haplotype counts (controls/cases)	SNP (ERAP2)	Haplotype	Haplotype counts (controls/cases)	OR (95% CI)	rs2910686 association P value
	rs30187[T]	790/5,641	rs2910686[T]	TT	662/4,595	1.18 (0.96–1.44)	0.12
			rs2910686[C]	TC	128/1,046
HLA-B*27 positive
	rs30187[C]	1,580/7,979	rs2910686[T]	CT	646/2,923	1.20 (1.07–1.34)	1.5 × 10−3
			rs2910686[C]	CC	934/5,056
	rs30187[T]	8,298/763	rs2910686[T]	TT	6,902/629	1.05 (0.87–1.28)	0.64
			rs2910686[C]	TC	1,396/134
HLA-B*27 negative
	rs30187[C]	16,240/1,515	rs2910686[T]	CT	6,842/548	1.28 (1.15–1.43)	7.7 × 10−6
			rs2910686[C]	CC	9,398/967

Association was assessed by 1-degree-of-freedom χ² test.

LNPEP and ERAP2 are both members of the endoplasmic reticulum (ER) aminopeptidase family and have substantial sequence homology with ERAP1. They are both encoded on chromosome 5q15, immediately centromeric to the ERAP1 locus. A previous study, which did not control for the association of ERAP1 with ankylosing spondylitis, observed no association of an ERAP2 loss-of-function variant with ankylosing spondylitis¹⁶. Other studies have identified associations with ERAP2 but have not dissected them from the known associations of ERAP1 SNPs^5,17. Here, controlling for the association of ERAP1 with ankylosing spondylitis, two functionally important SNPs in ERAP2 were found to be associated with ankylosing spondylitis: rs2549782, which leads to a change in ERAP2 catalytic activity¹⁸, and rs2248374, where the protective G allele causes complete loss of ERAP2 mRNA and absence of ERAP2 protein¹⁹. In the European samples, controlling for the association with ERAP1, we identified SNPs in ERAP2 and LNPEP that were associated with ankylosing spondylitis (lead SNP, rs2910686: OR = 1.2, P = 4.5 × 10⁻¹⁷; Supplementary Fig. 3b). Because the association of ERAP1 variants was restricted to HLA-B*27–positive ankylosing spondylitis, analyzing ERAP2 and LNPEP SNPs in HLA-B*27–negative cases and controls produced similar results to analyses of the combined (HLA-B*27–positive and HLA-B*27–negative) cases and controls when the association with ERAP1 was controlled for. Thus, we observed association with ERAP2 SNPs in HLA-B*27–negative ankylosing spondylitis cases (rs2910686: OR = 1.19, P = 2.13 × 10⁻⁵). To investigate the possibility that this finding might be an artifact caused by the strong LD between the ERAP2 locus and HLA-B*27 in cases, we tested association at this locus in a multivariate analysis, controlling in the one analysis for association of the two ankylosing spondylitis–associated ERAP1 hap-lotypes (tagged by rs30187 and rs10045403), their interaction with HLA-B*27 and HLA-B*27 itself. In this analysis, association with rs2910686 was robust (P = 6.6 × 10⁻⁸), suggesting that ERAP2-LNPEP is independently associated with ankylosing spondylitis. Haplotype counts for rs2910686 and rs30187 in both HLA-B*27–positive and HLA-B*27–negative cases are shown in Table 3.

In addition to these ERAP1 and ERAP2 associations, we observed genome-wide significant association of SNPs on chromosome 17q21 around the gene NPEPPS, which encodes puromycin-sensitive aminopeptidase (rs9901869: OR = 1.14, P = 6.0 × 10⁻¹⁵) (Supplementary Fig. 5). The NPEPPS protein localizes to the cytoplasm and is thought to be involved in processing proteasome-derived peptides before their transport to the endoplasmic reticulum and presentation by human leukocyte antigen (HLA) class I molecules²⁰. Association had previously been reported at this locus and was ascribed to TBKBP1 or TBX21. Here SNPs mapping to NPEPPS and TBX21 were independently associated with ankylosing spondylitis, and conditional analysis suggested that there are at least two independent signals at this locus. When conditioning on rs9901869, SNP rs1 1657479 in the 3′ UTR of TBX21 remained significantly associated with ankylosing spondylitis (OR = 1.09; P = 1.8 × 10⁻³). The data did not allow us to determine whether TBKBP1 or TBX21 was primarily associated with ankylosing spondylitis, but both represent attractive candidates. TBKBP1 is a component of the tumor necrosis factor (TNF) signaling pathway, and TBX21 is a transcription factor that influences the differentiation of T helper 1 (T_H1) and natural killer (NK) cells²¹.

Genes influencing lymphocyte activation and differentiation

It has recently been shown that cell type–specific trimethylation of histone H3 at lysine 4 (H3K4me3) chromatin marks can inform the fine mapping of associated SNPs to identify causal variation²². We therefore tested all ankylosing spondylitis-associated SNPs against H3K4me3 chromatin marks in different cell lines from the Encyclopedia of DNA Elements (ENCODE) Project²³. This analysis showed a strong enrichment of disease-associated SNPs associated with H3K4me3 chromatin marks in cells of immune origin (Supplementary Fig. 6).

Because of this association and taking into account the likely pathogenic role of T lymphocytes in ankylosing spondylitis and the involvement of several genes associated with ankylosing spondylitis in T-lymphocyte differentiation, we tested association of the SNPs for ankylosing spondylitis with CD4⁺ and CD8⁺ T cell counts in a previously published GWAS data set²⁴. Association (P < 0.005) was seen between CD8⁺ lymphocyte counts and ankylosing spondylitis– associated SNPs in the loci harboring the genes IL7R, RUNX3 and ZMIZ1 (Supplementary Table 5). Association was also observed between SNPs in EOMES and CD8⁺ lymphocyte counts, but these were not the same EOMES SNPs that were associated with ankylosing spondylitis. We also showed association of the genes SH2B3 and BACH2 with both ankylosing spondylitis and CD4⁺ lymphocyte counts (Supplementary Table 5). We previously showed that CD8⁺ lymphocyte counts are lower in ankylosing spondylitis cases than in healthy age- and sex-matched controls⁶. In contrast, in this study, we found that ankylosing spondylitis cases not on biological therapy had similar CD4⁺ lymphocyte counts as age-matched controls (Supplementary Fig. 7).

HLA Region

After SNP imputation in the MHC region, rs1 16488202 was found to tag HLA-B*27 more accurately in both Europeans and Asians than our previously reported tagging SNP rs4349859 and also rs13202464, reported to tag HLA-B*27 in Asian populations⁸ (Supplementary Table 6). The expected strong association was observed with HLA-B*27 (OR = 46; P < 1 × 10⁻¹⁰⁰) (Supplementary Fig. 8). Risk of ankylosing spondylitis was further increased in HLA-B*27 homozygotes; HLA-B*27 homozygosity was more prevalent in HLA-B*27-positive cases than in HLA-B*27-positive controls (OR = 2.07; P = 0.0025).

Controlling for association with HLA-B*27, there was residual signal with SNPs near HLA-A and HLA-B (Fig. 2). The residual signal at HLA-B may reflect either imperfect HLA-B*27 tagging by rs1 16488202 or association of other HLA-B alleles with ankylosing spondylitis. No other individual non–HLA-B*27 allele was associated with ankylosing spondylitis, although this may represent imperfect HLA-B imputation using single SNPs.

Figure 2.

Ankylosing spondylitis susceptibility associations in the MHC region conditioning on the HLA-B*27–tagging SNP rs116488202 and further conditioning on the HLA-A*02–tagging SNP rs2394250. The 85-kb gap between positions 32,465 kb and 32,550 kb corresponds to an assembly correction between NCBI Genome Builds 36 and 37 of the human genome.

The most strongly associated SNP near HLA-A, rs2394250, tags the classical allele HLA-A*0201 (Supplementary Table 7). Association of the HLA-A*0201 allele was independent of HLA-B*27 genotype, present in both HLA-B*27–positive (OR = 1.21, P = 6.5 × 10⁻¹²; conditioning on rs1 16488202) and HLA-B*27–negative (OR = 1.36, P = 3.2 × 10⁻¹³) disease (Supplementary Table 7b). No significant correlation was noted between HLA-B*27 and SNPs tagging HLA-A*0201 (r² < 0.01), and, thus, this association is not a manifestation of LD with HLA-B*27.

Overlap with other immune-mediated diseases

Considering the loci associated with ankylosing spondylitis in this study, we found substantial overlap with other immune-mediated diseases (Fig. 3 and Supplementary Table 8), notably, inflammatory bowel disease (either Crohn's disease or ulcerative colitis) and celiac disease. Ankylosing spondylitis–associated loci were associated with the same SNP in the same direction of association at 12 loci shared with Crohn's disease, at 11 loci shared with ulcerative colitis and at 6 of 7 loci shared with celiac disease. Overlap of associated loci with other diseases was not as marked, including for rheumatoid arthritis (one concordant, one discordant), psoriasis (two concordant), multiple sclerosis (three concordant, three discordant) and type 1 diabetes (two concordant, two discordant).

Figure 3.

Ankylosing spondylitis genetic susceptibility loci overlap with those of other autoimmune diseases. Diseases are represented in columns, and ankylosing spondylitis susceptibility loci are represented in rows. Shared susceptibility loci are colored green if effect size is concordant and purple if effect size is discordant. Data are shown in supplementary table 8. SLE, systemic lupus erythematosus; T1D, type 1 diabetes.

Refinement of disease associations and secondary signals

In the design of the Immunochip, eight loci already known to be associated with ankylosing spondylitis were selected for fine mapping. Compared with available ankylosing spondylitis GWAS data, the current study had greater marker density at these loci and larger sample size (from 3,023 cases and 8,779 controls to 10,619 cases and 15,145 controls). Nonetheless, for most loci studied, the disease-associated region was not substantially narrowed; less than 10% narrowing of the region was observed for four of the eight loci that were fine mapped (Supplementary Table 9). This suggests that, for many loci associated with common variants, the extent of LD at the locus will be too great to permit substantial refinement of the locus using sample sizes and marker densities of the magnitude employed here.

Two or more independent signals were identified at 12 of the 25 genome-wide significant loci (P < 5 × 10⁻⁴; Table 2 and Supplementary Fig. 9), including 1 locus (ERAP1 on chromosome 5q15) with 3 associated haplotypes (Supplementary Fig. 3). This is a similar proportion to that found in celiac disease (13 of 36 loci)¹⁰. Taken together, these secondary signals contribute 0.75% of the heritability of ankylosing spondylitis (Supplementary Table 10).

Rare variants

Because of the likelihood of population stratification affecting rare variant associations, rare variant associations (MAF < 1%) were only tested on a reduced sample subset of individuals from the UK (7,447 cases and 11,479 controls). Considering loci with common variant associations achieving genome-wide significance, we identified six that harbored rare SNPs that were disease associated (P < 5 × 10⁻³; Supplementary Table 11). These associations remained significant after conditioning on the common variants, and variants at three loci (IL23R, TYK2 and KEAP1) remained significant after controlling for the number of variants studied per locus. Four loci had exonic rare variant associations in the absence of a common variant association (P < 5 × 10⁻⁵) (KLKB1, RAD50, PRDM1 and DYRK4; Supplementary Table 11b). The rare variant with the largest effect was a predicted splice-site variant in TYK2 (rs280518: OR = 7.7, P = 0.002).

Comparisons across ancestry groups

The power of our East Asian case-control cohort was much lower than for our European cohort; nonetheless, at least nominal association (P < 0.01) was detected in East Asians at 13 of the 23 loci for which we identified associations at genome-wide significance in the overall data set (Table 1).

At some loci, association was seen in both East Asians and Europeans but with different SNPs. At IL23R, the primary associated variant in Europeans, rs11209026, was not polymorphic in East Asians, as we and others have previously reported²⁵. However, association was observed at IL23R with a low-frequency nonsynonymous SNP in East Asians (rs76418789: p.Gly149Arg, OR = 1.5, P = 8.2 × 10⁻⁴). The minor, protective allele was predicted to be deleterious by both SIFT²⁶ and PolyPhen analysis²⁷. The same SNP was also nominally associated with ankylosing spondylitis in Europeans (P = 0.01). The MAF of rs76418789 was ∼10 times greater in East Asians than in Europeans (East Asians, 3.7%; Europeans, 0.34%). Of the other loci associated with ankylosing spondylitis in Europeans but not in East Asians, only at BACH2 was the key associated variant present at a much lower frequency in East Asians (rs17765610; MAF of 11.8% in Europeans and 1.8% in East Asians), suggesting that, at most loci with discordant association between ancestry groups, this was not due to differences in the population frequency of the associated SNP.

At PTGER4, association in Europeans peaked 155 kb 5′ of the gene (peak associated SNP, rs12186979: OR = 1.1, P = 5.4 × 10⁻⁶), whereas the peak of association in East Asians was in intron 2 of the gene (rs13354346: OR = 1.3, P = 1.5 × 10⁻⁵) (Supplementary Fig. 5).

Discussion

This study confirmed the association of 12 of the 13 previously reported loci associated with ankylosing spondylitis in Europeans and identified 13 additional loci at genome-wide significance. We found no independent support for two loci previously reported to be associated with ankylosing spondylitis in Han Chinese, suggesting that the original associations of these loci may have been false positives. Additional associated haplotypes were identified at 13 loci, increasing the total number of distinct ankylosing spondylitis associations to 43.

These findings highlight the role of some major biological pathways in the pathogenesis of ankylosing spondylitis, including the IL-23 pathway, gut immunity, T-lymphocyte differentiation or activation, and peptide processing before HLA class I presentation.

We identified three new ankylosing spondylitis–associated genes (TYK2, IL27 and IL6R) with known effects on the IL-23 pathway. TYK2 is a member of the Janus kinase family of intracellular signaling proteins and is involved in signal transduction from IL-23R, as well as other cytokine receptors, including those in interferon (IFN)-α, IFN-β, IL-6, IL-10 and IL-12 signaling. Common TYK2 variants are also associated with Crohn's disease²⁸ and psoriasis²⁹. A different, rare TYK2 variant, rs34536443, is associated with multiple sclerosis³⁰. IL-6 signaling through IL-6R has diverse proinflammatory effects. In combination with transforming growth factor (TGF)-β, it influences the ratio of T_H17 to regulatory T (T_reg) cells, promoting the differentiation of T_H17 cells from naive T cells and inhibiting TGF-β–induced differentiation into T_reg cells³¹. Previous studies have reported no increase in T_H17 lymphocyte counts in ankylosing spondylitis, suggesting that the IL6R association with ankylosing spondylitis operates through mechanisms other than effects on T_H17 lymphocytes^32,33. Whether IL-6 has a role in the differentiation or activation of non-canonical cellular sources of IL-17 such as γδ T cells, NK cells, neutrophils and mast cells, which have been implicated in ankylosing spondylitis, is unclear. IL-27 potentiates the differentiation of CD4⁺ T_H1 cells, while suppressing the differentiation of T helper 2 (T_H2) and T_H17 cells. IL27 has previously been associated with both Crohn's disease and type 1 diabetes (Supplementary Table 8), with the association in type 1 diabetes being in the opposite direction to that observed in both ankylosing spondylitis and Crohn's disease.

At IL23R, we identified rare variant associations with ankylosing spondylitis in addition to the two known common variant haplotypes at this locus. Although one of these haplotypes has been shown to be due to association with the rs11209026 coding SNP^6,25, it is not clear from these genetic studies whether the second haplotype tagged by the intergenic SNP rs12141575 influences ankylosing spondylitis through effects on IL23R or IL12RB2. IL12RB2 encodes one of the two subunits of IL-12R, the stimulation of which drives CD4⁺ lymphocyte differentiation toward the T_H1 lineage and away from the T_H17 lymphocyte phenotype.

We identified six ankylosing spondylitis–associated genes that were also associated either with variation in CD8⁺ lymphocyte counts (EOMES, IL7R, RUNX3 and ZMIZ1) or CD4⁺ lymphocyte counts (BACH2 and SH2B3). EOMES encodes eomesodermin, a transcription factor involved in CD8⁺ T cell differentiation whose expression is induced by RUNX3 (refs. 34–36). Where eomesodermin is deficient, CD8⁺ T cells have been shown to express IL-17 (ref. 37). IL-7 acts through IL-7R to induce RUNX3 expression in developing T cells, in turn favoring differentiation toward the CD8⁺ T cell lineage³⁸. ZMIZ1 is a transcriptional coactivator of the protein inhibitor of activated STAT (PIAS)-like family and thus may have effects on STAT-mediated cytokine signaling. ZMIZ1 has recently been shown to cooperate with activating NOTCH1 mutations in inducing T cell acute lymphoblastic leukemia, consistent with it having a role in T cell differentiation³⁹. Whether these genes affect risk of ankylosing spondylitis directly through effects on CD8⁺ T cell differentiation is unclear. For example, although the risk haplotype at RUNX3 is associated with lower CD8⁺ T cell counts, at IL7R, the opposite phenotype is observed, suggesting that the mechanisms involved are more complex than a simple effect on CD8⁺ T cell counts. IL-7 treatment has been shown to increase T_H17 lymphocyte counts, and it may be that the association of IL7R with ankylosing spondylitis operates directly through such an effect⁴⁰.

One of the genes associated with both ankylosing spondylitis and CD4⁺ lymphocyte counts, BACH2, encodes a B cell–specific transcription factor with diverse effects on B cell differentiation and function⁴¹; association of ankylosing spondylitis with CD4⁺ lymphocyte counts may thus be an indirect effect mediated by B cells. This is particularly noteworthy given the recent evidence suggesting that rituximab, a B cell–targeted therapy, may have beneficial effects in ankylosing spondylitis⁴². SH2B3 (also known as LNK) encodes an adaptor protein involved in T cell receptor signaling⁴³. CD8⁺ lymphocytes are activated by the interaction of MHC class I peptides with their T cell receptors and may in turn become cytotoxic or memory T cells.

The association of four aminopeptidases involved in peptide trimming before HLA class I presentation is particularly noteworthy. We have shown here and previously that genetic variants associated with reduced function of ERAP1 and loss of expression of ERAP2 are protective for ankylosing spondylitis. Whether LNPEP is also involved is uncertain, and identification of the key associated variants at NPEPPS will require further studies. It is possible that these genes operate in ankylosing spondylitis through a quantitative effect on HLA class I peptide presentation or a qualitative effect on the peptide repertoire presented. Downregulation of ERAP1 (refs. 44,45) and ERAP2 (ref. 19) expression has been shown to reduce cell surface expression of HLA class I molecules. ERAP1 preferentially cleaves hydrophobic amino acids, whereas ERAP2 preferentially cleaves basic residues. ERAP1-ERAP2 heterodimers may thus act in concert, particularly in cleaving longer peptides⁴⁵. It has been suggested that misfolding of nascent HLA-B*27 in the ER, leading to ER stress, may be involved in the pathogenesis of ankylosing spondylitis⁴⁶. It is also possible that, by influencing the quantity of peptide available during HLA-B*27 folding, ERAP1 and ERAP2 variants associated with disease risk slow the rate of this folding, thereby increasing ER stress.

In this study, over one-third of loci with common variant associations were found to harbor more than one disease-associated haplotype. Identifying these additional haplotypes increased the proportion of genetic variance explained in ankylosing spondylitis and, more notably, led to valuable biological insights. For example, association of ankylosing spondylitis with SNPs on chromosome 12p13 has previously been reported, although it was not clear whether the association was primarily with TNFRSF1A or LTBR, both plausible candidate genes^6,7,47. The current study shows that there are two signals at this locus, one in TNFRSF1A and the other in LTBR. The primary associated SNP at TNFRSF1A (rs1860545) is in strong LD with a multiple sclerosis–associated SNP, rs1800696 (r² = 0.96, D′ = 0.98) but with the opposite direction of association. rs1800696 has recently been shown to lead to the splicing out of exon 6 of TNFR1, resulting in loss of the transmembrane domain⁴⁸. The resulting protein acts as a soluble decoy receptor for TNF, akin to the TNF inhibitor drug etanercept. TNF inhibitors are highly effective therapeutic drugs in ankylosing spondylitis, but their use can lead to induction or exacerbation of multiple sclerosis. The association with ankylosing spondylitis suggests the possibility that disease activity and response to TNF inhibitor therapy may be affected by this SNP.

Comparison of the genetic associations of ankylosing spondylitis with other diseases reinforces the considerable overlap with Crohn's disease, ulcerative colitis and ankylosing spondylitis. Ankylosing spondylitis frequently complicates inflammatory bowel disease (both Crohn's disease and ulcerative colitis), and increased cofamiliality with inflammatory bowel disease has been demonstrated^49,50, suggesting shared etiopathogenesis. Overlapping genetic susceptibility between ankylosing spondylitis and inflammatory bowel disease has previously been reported^51,52. The genes involved include many with effects on the IL-23 pathway, supporting the notion of this being a key pathway in the pathogenesis of these conditions, most likely through effects on gut mucosal immunology. However, the major loci for each disease are not shared, with ankylosing spondylitis showing no association with NOD2 or ATG16L1 and neither Crohn's disease nor ulcerative colitis showing association with HLA-B*27. This suggests that these disease-specific loci contribute to the organ and tissue specificity of the diseases with which they are associated, whereas the IL-23 pathway is involved in the core immunological pathway underlying all these conditions.

We also identified associations with three loci encoding G protein–coupled receptors, including GPR35, GPR37 and GPR65, and a fourth (GPR25) is close to KIF21B, an established ankylosing spondylitis locus where the key associated variants are not yet defined. The functions of these genes and their ligands are not well established. GPR35 is reported to act as a receptor for 2-acyl lysophosphatidic acid, and GPR65 is reported to be a receptor for glycosphingolipids and protons; the ligands for GPR25 and GPR37 are not known. GPR65 has previously been associated with Crohn's disease⁵³ and multiple sclerosis⁵⁴. The mouse homolog of GPR65, T cell death–associated gene 8, inhibits proinflammatory cytokine production (including of TNF-α and IL-6) in acidic conditions⁵⁵, suggesting a potential mechanism in diverse autoimmune diseases. However, it also has anti-apoptotic effects and an ability to activate not only cyclic AMP (cAMP) intracellular signaling but also other pathways, including mitogen-activated protein kinase (MAPK) and MEK/ERK signaling and thus is likely to have multiple functions⁵⁶. Further research is needed into the functions of these genes and their roles in autoimmune diseases.

It has long been suspected that associations in the MHC region with ankylosing spondylitis are not completely explained by HLA-B*27. The association of HLA-A*0201 with ankylosing spondylitis at genome-wide significance in both HLA-B*27–positive and HLA-B*27–negative cases confirms that suspicion. HLA-A*02 has previously been reported to be associated with anterior uveitis complicating ankylosing spondylitis⁵⁷ and is a risk factor for vitiligo⁵⁸. HLA-A*0201 has a protective effect in multiple sclerosis⁵⁴ and is a risk allele for type 1 diabetes⁵⁹, but no HLA-A association has previously been reported with ankylosing spondylitis itself.

In conclusion, we have increased the number of ankylosing spondylitis–associated loci to 31, identifying 13 new loci and 12 additional ankylosing spondylitis–associated haplotypes at 11 loci, bringing the total number of genetic signals independently associated with anky-losing spondylitis to 43. These loci reinforce the mounting evidence that aberrant peptide processing before MHC class I presentation and alterations of the IL-23 pathway are key elements in the pathogenesis of ankylosing spondylitis.

URL. Haploxt, http://genome.sph.umich.edu/wiki/Haploxt.

Online Methods

Samples

All cases had definite ankylosing spondylitis according to the modified New York criteria⁶⁰. Written informed consent was obtained from all cases with approval from the relevant research ethics authorities at each participating center. A total of 12,252 DNA samples passed genotyping control filters. The case collection consisted of 10,417 individuals of European ancestry, 1,560 of Asian ancestry and 275 from Latin America (Colombia and Mexico). Of these, 2,425 cases of European ancestry have previously been reported in GWAS.

We obtained 12,338 controls of European ancestry, 1,570 of East Asian ancestry and 445 from Latin America. These included shared controls from the UK 1958 Birth Cohort, the UK Blood Services Common Controls and the United States and from participating centers from France, The Netherlands, Norway, Spain, Mexico, Colombia, China, Taiwan and Korea.

Genotyping

Samples were genotyped using the Immunochip, an Illumina Infinium platform, according to the manufacturer's recommendations. Control samples from the UK were genotyped at the Sanger Centre (Hinxton, Cambridge, UK) and at the University of Virginia (Charlottesville, Virginia, USA), control samples from the United States were genotyped at the Feinstein Institute (New York, New York, USA), and all other controls and all cases were genotyped at the University of Queensland Diamantina Institute (Brisbane, Queensland, Australia). Bead intensity data were processed and normalized for each sample in GenomeStudio; data for successfully genotyped samples were extracted, and genotypes were called within collections using optiCall⁶¹. NCBI Build 36 (hg18) mapping was used (Illumina manifest file Immuno_BeadChip_11419691_B.bpm).

SNPs rs4552569, rs13210693 and rs 17095830 were genotyped in 2,998 East Asian cases and 5,547 East Asian controls using TaqMan probes according to the manufacturer's instructions.

The genotyping accuracy of seven disease-associated SNPs with MAF < 1% was confirmed by custom TaqMan genotyping assays. Taking into consideration the five successfully developed assays, genotypes were completely concordant with Immunochip genotyping. Assays could not be developed for two SNPs (rs280518 in TYK2 and rs75430612 in IL2R2). However, given the clean clustering achieved for these SNPs in microarray genotyping and the fact that they had previously been reported, it is likely that they represent true positives.

Data quality control

We first excluded, for each of the collections separately, SNPs with call rate below 95% or with a Hardy-Weinberg equilibrium P value of <1 × 10⁻⁶ in controls, as well as samples with call rate below 95%. For the overlapping SNPs, we performed pairwise missingness tests between the collections and removed all SNPs with differential missingness (P < 1 × 10⁻⁷). After merging data sets, SNPs with call rate below 98% and samples with call rate below 98% were removed. A total of 128,935 SNPs were analyzed. Cluster plots were visually inspected for all SNPs used to inform conclusions.

The origins of samples of European and Asian ancestry were confirmed by principal-component analysis. Immunochip data were merged with genotype information from seven HapMap 3 populations (CEU, TSI, YRI, MEX, JPT, CHD and CHB), and samples were identified as European or East Asian on the basis of their projection onto the first five principal components of genetic variation (Supplementary Figs. 10–12). Ancestry outliers were removed by assigning samples to an ancestry group using a model-based unsupervised clustering approach⁶². A second round of principal-component analysis was performed on the European and Asian populations independently to better resolve ancestry differences within the cohorts. In both cases, only the first principal component was correlated with the places of origin of the samples, and only in the European collection was the principal component marginally correlated with case-control status (P = 0.016). The projection of the samples analyzed onto the first two principal components is shown in Supplementary Figures 3 and 4.

Duplicated samples (intentional or unintentional) and those showing cryptic relatedness were assessed for the European and Asian cohorts separately by calculating identity by descent using PLINK (v1.07)⁶³. For each pair of related samples (PI_HAT > 0.2), the sample with the lower call rate was removed, and, where the pair involved a case and a control with similar call rates (both above 98%), cases were preferentially selected for inclusion.

Association analysis

Association analysis and population stratification control was performed using linear mixed models as implemented in FaST-LMM (v1.05)⁶⁴. For each chromosome, a relationship matrix was constructed with all SNPs, excluding those in the chromosome being analyzed and in the MHC region. Conditional analysis for secondary signal detection was performed by fitting the primary SNP as a fixed effect as implemented in FaST-LMM and using the same SNP set for the relationship matrix.

The decision to apply linear mixed models rather than principal-component analysis for population stratification was informed by a comparison of logistic regression with principal components as covariates against linear mixed models, including different strategies for computing the kinship (or similarity) matrix. In this analysis, we concluded that linear mixed models, when using a kinship matrix, outperformed logistic regression with principal components as covariates, as assessed by the genomic inflation factor in all SNPs (data not shown). We also noticed that including all SNPs when computing the kinship matrix resulted in further loss of power, particularly matrices calculated including HLA-B*27–tagging SNPs, which are a good proxy for phenotype status in a case-control study such as ours. Thus, including SNPs tagging HLA-B*27 in the kinship matrix would have the effect of controlling for affected status, reducing statistical power in a case-control study.

Transformation of SNP effects from the linear 0–1 scale to the liability scale is described in the Supplementary Note. Association plots were generated using LocusZoom, with recombination rates estimated from the HapMap CEU panel (Utah residents of Northern and Western European ancestry) and pairwise LD r² values estimated from the set of control samples⁶⁵.

Low-frequency SNP association analysis

A collection of 7,447 cases and 11,479 controls were selected for low-frequency SNP association analysis. These samples were of UK European origin according to principal-component analysis (Supplementary Fig. 11) and recruitment center informaiton. Association analysis was performed with Fisher's exact test and with logistic regression conditioning on common variant association within the locus.

Imputation

We imputed genotypes in candidate regions for the European and Asian cohorts using the EUR and ASN reference panels, respectively, from the 1000 Genomes Project (Phase 1, 2010–2011 data freeze)⁶⁶. Genotype data were phased with MACH, and genotypes were imputed with Minimac⁶⁷. SNPs with low imputation quality (r² < 0.5) were excluded. Association analysis in imputed genotypes was assessed with probabilistic genotypes, correcting for population stratification with the first five principal components as covariates, using logistic regression as implemented in mach2dat.

HLA-B*27–tagging SNP

Imputed SNPs in the MHC region were tested for association and for tagging of HLA-B*27. Performance as a tagging SNP was assessed with a cohort of samples previously genotyped at this locus (Supplementary Table 6b), including 754 controls from the 1958 Birth Cohort population⁶⁸, 542 ankylosing spondylitis cases from the UK and Australia and 104 ankylosing spondylitis cases and 5 controls from China. SNP rs116488202 was used to tag HLA-B*27 in all analyses.

Overlap with other autoimmune diseases

For all genome-wide significant and suggestive loci for ankylosing spondylitis susceptibility, we searched in a chromosomal window of 0.5 cM around the lead SNP for associations with other autoimmune diseases in the National Human Genome Research Institute (NHGRI) GWAS catalog (accessed 5 June 2012). We then computed the LD between the ankylosing spondylitis–associated SNP and the reported SNP using phased data from the 1000 Genomes Project. If the ankylosing spondylitis–associated SNP was not found, we searched for the next most significant SNP in the locus. When a pair of SNPs was in LD (either r² > 0.40 or D′ > 0.40), a connection between the two diseases was noted. Positively correlated alleles were then compared for their risk or protective effect on the two diseases to determine whether the directionali-ties of effect were concordant or discordant. LD was computed with Haploxt.

HLA-B*27 experiment-wide interactions

Testing for interaction between HLA-B*27–tagging SNP rs116488202 and all other non-MHC SNPs was performed in samples of European ancestry by logistic regression fitting a dominant term for the HLA-B*27–tagging SNP and an additive term for each test SNP, including a multiplicative interaction term, and five principal components for ancestry correction:

$$y={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+{\beta }_{12}{x}_1\times x_2+\sum _{i=1}^5{\beta }_{{\text{PC}}_i}{\text{PC}}_i$$

where y is the log odds of disease, β_o is the intercept, x₁ is a SNP variable reflecting an HLA-B*27 dominant effect (0 or 1), x₂ is a SNP variable coded to reflect an underlying additive effect (0, 1 or 2), PC_i codes for the projection of samples onto the ith principal component and β terms are regression coefficients estimated from the data.

CD4⁺ lymphocyte counts in ankylosing spondylitis cases and controls

Peripheral blood was obtained from 20 individuals with active ankylosing spondylitis (erythrocyte sedimentation rate of >25 mm/h and CRP concentration of > 10 mg/l) who were naive for TNF inhibitor and 20 age-matched healthy controls. Peripheral blood mononuclear cells (PBMCs) were extracted using standard density gradient centrifugation over Ficoll-Paque Plus (GE Healthcare). Extracted PBMCs were frozen in FBS with 10% DMSO until needed. Frozen PBMCs were thawed into RPMI supplemented with 20% FBS and were washed once in RPMI supplemented with 10% FBS. Cells were rested in RPMI supplemented with 10% FBS for approximately 1 h at 37 °C at 5% CO₂ before further use. Cells were stained with CD3 ECD (UCHT1, Beckman Coulter) and CD4 Pacific Blue (13B8.2, Beckman Coulter). Antibodies were used at final dilutions of 1:100 of antibody stocks. Dead cells were excluded using a Live/Dead Fixable Dead Cell Stain kit (Invitrogen). Cells were acquired on a Gallios flow cytometer (Beckman Coulter), and staining was analyzed using Kaluza software (Beckman Coulter).

IL-6R measurements in serum

Serum was collected from ankylosing spondylitis cases and controls of European ancestry attending the Princess Alexandra Hospital Brisbane Ankylosing Spondylitis Specialist Clinic who were either homozygous for the T or C allele of rs4129267. Serum concentrations of IL-6R were measured using an IL-6R Quantikine ELISA (R&D Systems), and optical density (OD) was determined on a Synergy 2 Microplate reader (BioTek). Serum was diluted 1:100, and ELISAs were performed according to the manufacturer's instructions. All data shown in graphs represent mean ± s.e.m., and differences between groups were analyzed using non-parametric one-way ANOVA; Kruskal-Wallis test with Dunn's multiple-comparison post-hoc test in GraphPad Prism 5 software. Statistical significance was accepted at a significance level of P < 0.05.