Major Histocompatibility Complex Associations of Ankylosing Spondylitis are Complex and Involve Further Epistasis with ERAP1

Abstract

Ankylosing spondylitis (AS) is a common, highly heritable, inflammatory arthritis for which HLA-B*27 is the major genetic risk factor, although its role in the aetiology of AS remains elusive. To better understand the genetic basis of the MHC susceptibility loci, we genotyped 7,264 MHC SNPs in 22,647 AS cases and controls of European descent. We impute SNPs, classical HLA alleles and amino acid residues within HLA proteins, and tested these for association to AS status. Here we show that in addition to effects due to HLA-B*27 alleles, several other HLA-B alleles also affect susceptibility. After controlling for the associated haplotypes in HLA-B we observe independent associations with variants in the HLA-A, HLA-DPB1 and HLA-DRB1 loci. We also demonstrate that the ERAP1 SNP rs30187 association is not restricted only to carriers of HLA-B*27 but also found in HLA-B*40:01 carriers independently of HLA-B*27 genotype.

Introduction

Ankylosing spondylitis (AS) is a common, highly heritable ¹, inflammatory arthritis for which HLA-B*27 is the major genetic risk factor. To better understand the genetic basis of the MHC susceptibility loci, we genotyped 7,264 MHC SNPs in 9,069 AS cases and 13,578 population controls of European descent using the Illumina Immunochip microarray. In addition to extremely strong effects due to HLA-B*27:02 and B*27:05, several other HLA-B alleles (B*07:02, B*13:02, B*40:01, B*40:02, B*47:01, B*51:01, and B*57:01) also affect susceptibility to AS. HLA-B independent associations were demonstrated with variants in the HLA-A, HLA-DPB1 and HLA-DRB1 loci. We also demonstrate that the ERAP1 SNP rs30187 association is not restricted only to carriers of HLA-B*27 but also found in HLA-B*40:01 carriers independently of the HLA-B*27 genotype. The presence of associations in both HLA Class I and II loci might reflect effects on antigen presentation to both CD4⁺ and CD8⁺ T lymphocytes in the pathogenesis of AS.

Whilst the classical HLA-B*27 allele is found in over 85% of AS patients ^2-4 it is clearly not sufficient alone to cause disease, with only 1-5% of HLA-B*27 carriers developing the disease. From epidemiological data it is evident that susceptibility to AS is affected by other genes within and outside the MHC ¹. Indeed, 26 risk loci outside the MHC have now been identified by genome-wide association studies^5-8.

The biological mechanism(s) by which HLA-B27 confers risk of disease remains elusive. The main hypotheses regarding this mechanism can be divided into canonical mechanisms based on the known function of HLA-B27 within the adaptive immune system, and non-canonical mechanisms related to unusual properties of HLA-B27, notably its propensity to dimerise or misfold. Suggested canonical mechanisms propose either that HLA-B27 is uniquely capable of presenting particular peptide(s) found at sites of inflammation in AS to cytotoxic T-lymphocytes (the arthritogenic peptide hypothesis) ⁹, or that HLA-B27 is associated with reduced gut mucosal immunity, leading to migration of enteric bacteria across the intestinal mucosa, driving the production of the pro-inflammatory cytokine interleukin (IL)-23 and development of AS (the mucosal immunodeficiency hypothesis) ^10,11. Both these theories place antigenic peptide presentation and handling as critical steps in the pathogenesis of AS. One of the first non-MHC susceptibility loci to be identified in AS was endoplasmic reticulum aminopeptidase 1 (ERAP1) ⁵, the main function of which is to trim peptides in the endoplasmic reticulum (ER) to optimal length for binding to MHC Class I molecules on antigen-presenting cells for subsequent interaction with CD8⁺ T-cells ^12,13. Moreover, this association is so far uniquely found in HLA-B*27 positive disease ⁷.

HLA-B27 has an unusual property of forming homodimers through disulphide bonding of the unpaired cysteine residue at position 67 ¹⁴. It has been proposed that these homodimers may cause AS through abnormal presentation of peptides or by facilitating “abnormal” interaction with natural killer cells ¹⁵. Apart from HLA-B*27, the subtypes of the alleles HLA-B*14, -B*15, -B*38, -B*39 and -B*75 encode a cysteine residue at position 67 but of these there is only evidence that HLA-B*14 may be AS associated ^16,17. It is also unclear if these other non-HLA-B27 Cys67 variants can form homodimers. Additionally, Cys67 is found on all HLA-B27 subtypes, including the subtypes HLA-B*27:06 and –B*27:09 which are not AS-associated ^18,19. A further hypothesis suggests that abnormal folding of the HLA-B27 molecule during assembly results in ER stress and activation of the unfolded protein response ^20,21. ER stress is evident in the HLA-B*27 transgenic rat model of AS and correlates with production of IL-23 ²¹, but has not been demonstrated in HLA-B*27 positive patients^22-24.

Whilst non-B27 HLA associations have been reported, notably with HLA-B40 ^25-27and HLA–A*02⁸, most have not been definitive or replicated in independent studies. In this study we analyze the associations of AS across the MHC aiming to identify functional and potentially causal variants using a large, previously reported, panel of cases and controls of European ancestry ⁸. Here we extend on our primary analysis of this cohort by fine-mapping the MHC region with imputation of SNPs, MHC Class I and II classical alleles, and amino acid residues within the classical HLA proteins ²⁸. In addition to HLA-B27, we identify further HLA-B and other HLA Class I and Class II alleles associated with AS, and demonstrate that HLA-B40 in addition to HLA-B27 interacts with ERAP1 to cause disease. This implicates both CD4 and CD8 lymphocytes in AS-pathogenesis and suggests that HLA-B40 and HLA-B27 operate by similar mechanisms to induce the disease.

Results

HLA-B susceptibility alleles

At the HLA-B locus, 38 classical alleles at four-digit resolution were imputed. All SNP, HLA and amino acid association P-values were determined by logistic regression. As expected, the two common HLA-B*27 alleles in the European population, B*27:02 (odds ratio (OR)=43, p=1.07×10⁻¹²²) and B*27:05 (OR=62, p<10⁻³²¹), were the most significantly associated with disease risk (Figure 1a-b and Tables 1 and 2). Controlling for the effect of the two B*27 alleles, we identified the protective alleles HLA-B*07:02 (OR=0.82; p=5.04×10⁻⁶) and HLA-B*57:01 (OR=0.75; p=5.13×10⁻⁴) (Table 2). Moderate association was also observed, sequentially, with the risk alleles HLA-B*51:01 (OR=1.33; p=2.14×10⁻³), HLA-B*47:01 (OR=2.35; p=2.25×10⁻³), HLA-B*40:02 (OR=1.59; p=4.65×10⁻³), HLA-B*13:02 (OR=1.43; p=4.29×10⁻³), and HLA-B*40:01 (OR=1.22; p=4.93×10⁻³). No evidence of further susceptibility alleles was observed after controlling for the risk and protective alleles identified above (p > 0.05; Figure 1d). The HLA-B associations were similar in both HLA-B*27 positive and negative restricted analyses (Supplementary Tables 1-6).

Figure 1.

Association results with ankylosing spondylitis susceptibility in the MHC. Omnibus SNP and amino acid association tests are shown in panels a), c), e), g) and i), and classical allele association tests with 2 and 4 digit resolution in panels b), d), f), h) and j). The strongest association was found with positions in the polymorphic nucleotide rs41558317 and in the polymorphic amino acid 97 of HLA-B a), and with the HLA-B*27 allele b). Controlling for the effect of HLA-B susceptibility alleles an independent association was observed with SNPs and amino acid position in the HLA-A locus c) corresponding to the HLA-A*02:01 allele d). Further conditioning on HLA-A and -B loci an independent association with SNPs and amino acid positions in the HLA-DPB1 locus was evident e); no HLA-DPB1 classical allele was significant at the same magnitude as the SNPs and amino acids positions f). Further controlling for the effect of variation in the HLA-DPB1 locus association was observed with SNPs in the HLA-DRB1 locus g) and h). SNP association tests are shown in blue circles, colour coded by linkage-disequilibrium from the SNP with the strongest association. Amino acid position tests are shown as red triangles. Classical allele tests are shown as bars for 2 and 4 digit imputation resolution.

Table 1.

Most significant polymorphic positions (omnibus test) and imputed classical alleles associated with ankylosing spondylitis susceptibility (P-value < 1 × 10⁻²⁰⁰⁰).

Position	rs	AA position	Classical allele	χ 2	DF	P-value
31,432,180	-	97	-	14,857	5	< 10−3221
31,432,180	rs1071652	-	-	14,841	3	< 10−3221
31,430,829	rs41558317	-	-	14,823	1	< 10−3221
31,432,179	rs1140412	-	-	14,823	2	< 10−3219
-	-	-	HLA-B*27	14,820	1	< 10−3221
31,432,506	-	70	-	14,812	3	< 10−3215
31,432,129	-	114	-	14,402	2	< 10−3128
31,432,130	rs709055	-	-	14,401	2	< 10−3128
31,432,131	rs1050628	-	-	14,389	1	< 10−3127
-	-	-	HLA-B*27:05	14,220	1	< 10−3090
31,430,834	rs3819282	-	-	13,798	1	< 10−2999
31,430,345	rs3819299	-	-	13,757	1	< 10−2990
31,430,346	rs3819299	-	-	13,757	1	< 10−2990
31,451,646	rs4463302	-	-	12,898	1	< 10−2803
31,432,485	-	77	-	12,871	2	< 10−2795
31,432,486	rs1131217	-	-	12,849	1	< 10−2793
31,377,108	rs2394967	-	-	11,613	1	< 10−2524
31,381,125	rs6905036	-	-	11,552	1	< 10−2511
31,432,208	rs41556113	-	-	10,929	1	< 10−2376
31,432,843	rs41553720	-	-	10,299	2	< 10−2237
31,432,515	-	67	-	9,741	4	< 10−2112
31,432,515	rs1071816	-	-	9,725	3	< 10−2110
31,518,387	rs2844510	-	-	9,525	1	< 10−2071

Table 2.

Evidence for association of HLA-B alleles with susceptibility to ankylosing spondylitis. Effect sizes and levels of significance were estimated in stepwise conditional procedure, where for rounds 2 and onwards the test conditioned on the previous alleles.

Round	HLA-B Allele	OR (95% CI)	P-value
1	27:05	62.41 (56.90-68.45)	< 10−321
2	27:02	43.41 (29.80-63.23)	1.07 × 10−122
3	07:02	0.82 (0.74-0.91)	5.04 × 10−6
4	57:01	0.75 (0.61-0.92)	5.13 × 10−4
5	51:01	1.33 (1.14- 1.56)	2.14 × 10−3
6	47:01	2.35 (1.43-3.86)	2.25 × 10−3
7	40:02	1.59 (1.19-2.14)	4.65 × 10−3
8	13:02	1.43 (1.14-1.80)	4.29 × 10−3
9	40:01	1.22 (1.06-1.40)	4.93 × 10−3
All other alleles			> 0.05

Non HLA-B susceptibility loci in the MHC

To assess whether other MHC loci affect disease susceptibility independently from the HLA-B locus, we performed additional conditional analyses. Adjusting for the HLA-B susceptibility alleles identified we observed an association signal with SNPs in the HLA-A locus (rs2975033, OR=1.22, p=6.16×10⁻¹⁰) and with the classical allele HLA-A*02:01 (OR=1.22, p=1.41×10⁻⁹) (Figure 1c-d). The risk allele “A” of rs2975033 was in near-perfect linkage disequilibrium (LD) with the risk allele HLA-A*02:01 (r²=0.97).

Further controlling for the effect of the susceptibility SNP rs2975033 in HLA-A revealed an independent signal with SNPs (rs1126513, OR=1.21, p=2.46×10⁻⁷) in the Class II locus HLA-DPB1 (Figure 1e); no association of similar strength to those seen with SNPs (p > 10⁻⁵) were observed with classical HLA-DPB1 alleles (Figure 1f). After controlling for the effect of the SNP rs1126513 in HLADPB1 we observed an association with the SNP rs17885388 (OR=1.16, p=1.27×10⁻⁵) in the HLA-DRB1 locus, and a similar level of significance was also observed with the Class II allele HLA-DRB1*01:03 (p=3.78×10⁻⁵). No further associations were observed after controlling for all identified effects (p > 5 × 10⁻⁵) (Figure 1i-j).

Association signals and amino acid positions in HLA proteins

We observed disease-associated alleles at MHC Class I and II loci. Classical alleles at these loci determine the amino acid sequence of the respective HLA proteins, which could in turn influence the specificity of the peptides presented to CD8⁺ and CD4⁺ T lymphocytes. We therefore analyzed the polymorphic amino acid residues at these proteins to assess their effect in disease susceptibility. In this analysis, the strongest association was observed for amino acid position 97 in HLA-B (omnibus p<10⁻³²²¹; Table 1 and Figure 1a-b). In addition, through conditional analysis, we found that the association at the HLA-B*27 allele, and other HLA-B*27-associated polymorphisms, was explained by position 97 while the reverse was not true (Supplementary Table 7). This polymorphic position carries as many as six different amino acid residues in the population (Figure 2), each conferring a different degree of risk (or protection) to disease, consistent with the analysis of HLA-B alleles mentioned above (Table 2). Position 97 lies in the floor of the HLA-B peptide-binding groove (Figure 3), located in the C/F pocket, also referred as the C-terminal pocket, which anchors the side chain of the C-terminal peptide residue ²⁹. Asparagine at position 97 is uniquely observed in HLA-B*27 alleles. Threonine at position 97 (predominantly found in HLA-B*51 alleles) was also found to increase disease risk (OR=1.12, p=4.50×10⁻³); serine (found in HLA-B*07 and *08 alleles) decreased risk of disease (OR=0.86, p=5.2×10⁻⁸); and valine (found in HLA-B*57 alleles) was also protective (OR=0.68, p=1.4×10⁻⁸) (Table 3).

Figure 2.

Amino acid residue frequencies in 13,578 controls and 9,069 cases within associated amino acid positions within HLA proteins.

Figure 3.

Three-dimensional models for the HLA-B, HLA-A and HLA-DPβ1 proteins. These structures are based on Protein Data Bank entries 3LV3, 3UTQ and 3LQZ, respectively, with a direct view to the peptide-binding groove.

Table 3.

Haplotype analysis of SNPs encoding the amino acid 97 of HLA-B

		HLA-B codon 97 position
		1	2	3
SNP	rs41558317	rs1140412	rs1071652	rs41556417	Amino acid residue
Position (HG18)	31,430,829	31,432,179	31,432,180	31,432,181
Reference allele or amino acid in HG18	A	G	C	T	Serine (S)
Alternate allele(s)	G	C/A	G/A/T	A/C
Allele frequency in controls (ref/alt(s))	0.95/0.05	0.29 0.67/0.05	0.82 0.10/0.04/0.05	0.92 0.04/0.04
Single locus univariate P-value	<10−3221	<10−3219	<10−3221	2.10 × 10−65		Multivariate OR (95% CI)	Unadjusted haplotype frequency		P	HLA-B Allele
Risk allele univariate OR (95% CI)	60.36 (55.47-65.74)	59.99 (55.13-65.33)	59.99 (55.14-65.34)	2.03 (1.86-2.21)			Controls	Cases
Haplotype	G/A	A	T	T	Asparagine (N)	16.51 (15.43-17.69)	0.045	0.449	< 10−300	*27:02 *27:04 *27:05
	G/A	C	G	T	Threonine (T)	1.12 (1.03-1.21)	0.097	0.065	4.50 × 10−3	*13:02 *39:06 *40:06 *51:01 *51:08 *52:01 *55:01 *56:01
	G/A	C	C	T	Arginine (R)	1.00 (Reference)	0.493	0.297	1	*15:01 *15:03 *15:10 *15:16 *15:17 *15:18 *18:01 *35:01 *35:02 *35:03 *35:08 *35:12 *37:01 *38:01 *38:02 *39:01 *39:10 *40:01 *41:01 *44:02 *44:03 *44:04 *44:05 *45:01 *47:01 *49:01 *50:01 *53:01 *58:01
	G/A	C	C	A	Tryptophan (W)	1.00 (0.89-1.12)	0.042	0.025	0.95	*14:01 *14:02
	G/A	G	C	T	Serine (S)	0.86 (0.81-0.91)	0.286	0.148	4.81 × 10−8	*07:02 *07:05 *08:01 *15:07 *27:07 *40:02 *41:02 *48:01
	A	C	A	C	Valine (V)	0.68 (0.59-0.78)	0.038	0.016	1.41 × 10−8	*57:01 *57:03

Strong associations were also observed with the amino acid positions 70, 114, 77 and 67 of HLA-B, but these signals were strongly attenuated after conditioning on amino acid position 97. In contrast, none of these positions could explain the association at position 97. In particular, there was little evidence of association at position 67 (i.e. the position where disulphide bonding of unlinked cysteine residues might occur) after conditioning on position 97 (P-value=0.04) (Supplementary Table 8).

The most strongly associated amino acid of the HLA-A molecule, after conditioning on associated HLA-B alleles, was amino acid valine at position 95 (p=3.70×10⁻⁹). The association with this amino acid was statistically equivalent with that observed with the SNP rs2975033 and with the classical allele HLA-A*02:01. This amino acid is positioned within the binding-site of HLA-A (Figure 3).

Independent associations were observed at the two Class II loci HLA-DPB1 and HLA-DRB1, and these were highly correlated with polymorphic amino acids in the peptide-binding site of these molecules (Figure 3). At the HLA-DPB1 locus, rs1126513 showed the strongest association and the risk allele for rs1126513 was perfectly correlated with the presence of leucine at position 11 of the HLA-DPB1 molecule (position 11, OR=1.21, p-value=2.46×10⁻⁷). At the HLA-DRB1 locus, the strongest association with an amino acid was observed with aspartic acid at position 70 (OR=1.16, p-value=3.44×10⁻⁵); due to LD this association was statistical equivalent to the one observed with the SNP rs17885388.

Gene-gene interactions and susceptibility loci

We have previously observed that the association with the variant rs30187 in the ERAP1 locus is restricted to HLA-B*27 positive subjects, consistent with an epistatic interaction between these two loci ⁷. Here we investigated the possibility of interaction between the other HLA-B susceptibility alleles and the variant rs30187. When testing for interaction with the HLA-B*40 alleles we found that rs30187-T increased the risk of disease in the strata where HLA-B*27 was present, as previously shown, or when HLA-B*40:01 was present in the absence of HLA-B*27 (OR=1.41; p=5.81×10⁻³); rs30187 had no effect on disease susceptibility when both HLA-B*27 and HLA-B*40 alleles were absent or in the non-HLA-B*27/HLA-B*40:02 stratum (Figure 4). No evidence of interaction was observed between rs30187 and the other HLA-B susceptibility alleles. This suggests that the rs30187 variant interacts with the HLA-B*40:01 allele; although no evidence to support an interaction was observed with HLA-B*40:02, the study had low power to detect such an effect. There was no evidence of interaction between either of the HLA-B*40 alleles and any of the other independently associated susceptibility SNPs in the loci encoding the aminopeptidases ERAP1, ERAP2 and NPEPPS (p > 0.1).

Figure 4.

Interaction between ERAP1 and HLA-B susceptibility alleles. For each stratified group, effect size for the ERAP1 variant rs30187 is given. Error bars represent 95% confidence intervals. Number of samples in each group (controls/cases) is given below the HLA-B*40 genotype.

We then examined if our data supported a model where the HLA-B*27 and -B*40 alleles increased disease susceptibility beyond their inferred independent effects, as previously reported ³⁰. No support for an interaction between these alleles was observed in this dataset (Supplementary Table 1).

Discussion

Independently of the expected HLA-B associations, this study demonstrates that both HLA-B*40:01 and -B*40:02 are disease associated alleles, and identified three further HLA-B risk alleles, HLA-B*51:01, B*47:01 and B*13:02. The allele HLA-B*51:01 is also the major genetic risk factor for Behçet’s disease ³¹, a seronegative disease complicated by sacroiliitis resembling AS in up to 10% of cases ³². In addition to the seven HLA-B risk alleles we identified two protective alleles at this locus, HLA-B*07:02 and B*57:01. Interestingly, in the HLA-B*27 transgenic rat model of AS, the HLA-B*27-negative control carries the HLA-B*07 allele, and does not develop disease, consistent with the protective effect of this allele in humans ³³. It has recently been shown in HLA-B7/B27 co-expressing mice that there is partial negative selection of HLA-B27+ T-cells in the course of defining the immunodominant response to influenza infection ³⁴. Further, in Erap deficient, influenza-infected HLA-B27-positive mice, there was a marked reduction in presentation of the HLA-B27 immunodominant epitope, and T cell immunity to that epitope, presumed to be because the HLA-B27-related immunodominant flu epitope requires cleavage by Erap to be presented by HLA-B27. In contrast, in HLA-B7-transgenic mice, Erap deficiency had no effect on presentation of the HLA-B7 immunodominant epitope or the corresponding T cell response to it, suggesting that it does not require Erap cleavage for presentation ³⁵. This provides a potential mechanism to explain the genetic effects observed in humans with AS, with ERAP1 loss of function protecting against HLA-B27 associated AS, but having no effect in HLA-B7 carriers, where an HLA-B7 protective association is observed.

Outside the HLA-B locus, we identified three independent significant signals associated with AS; one was in the HLA Class I locus HLA-A, and one each in the HLA Class II loci HLA-DPB1 and HLA-DRB1. The association in the HLA-A locus corresponded to the classical allele HLA-A*02:01, which has also been implicated in multiple sclerosis ³⁶; however, whilst this allele is protective in multiple sclerosis, it increases the risk of AS. Previous studies have hinted at HLA-DPB1 associations with AS, which we have confirmed here. HLA-DPB1, in conjunction with HLA-DPA1, forms the HLA-DP heterodimer which typically plays a role in the presentation of exogenously derived peptides, such as microbial peptides, to CD4⁺ T lymphocytes. The strongest association was found with an amino acid position located in the base of the peptide-binding groove of HLA-DP, suggesting that this polymorphism might impact on the peptide repertoire presented by HLA-DP.

Previous findings that ERAP1 variants influence risk of disease in HLA-B*27 positive, but not negative individuals, strongly supports the notion that both these molecules act in the same biological pathway to affect disease susceptibility ⁷. We have now shown that HLA-B*40:01 interacts with ERAP1 variants in the same manner. Similar genetic interactions involving ERAP1 have been observed in two other immune-mediated disorders - psoriasis with HLA-Cw6 ³⁷ and Behçet’s syndrome with HLA-B*51 ³⁸, two disorders that are already known to share genetic susceptibility factors with AS. It is likely that similar molecular mechanisms are involved in these disorders, and that these include the pathways of MHC Class I antigen presentation. To our knowledge, there is no evidence that HLA-B40, HLA-B51 or HLA-Cw6 have non-canonical disease-related properties such as those by which HLA-B27 is proposed to function in the pathogenesis of AS.

Analysis of polymorphic amino acid positions in these AS-associated HLA molecules showed that the SNPs with the strongest evidence of association at each of these three loci were highly associated with amino acid positions located in the peptide-binding groove of these proteins. From these results we infer that antigen presentation to both CD4⁺ and CD8⁺ T lymphocytes is likely to be important in the pathogenesis of AS and/or its tissue specificity, although other mechanisms underlying the associations cannot formally be excluded.

MHC Class I molecules contain six specificity pockets in the peptide-binding groove, alphabetically named A to F, which serve to anchor particular side-chains of the bound peptide ³⁹. Position 97 of HLA-B is located in the C/F pocket, also referred to as the C-terminal pocket, which anchors the side chain of the C-terminal peptide residue ²⁹. Experimental evidence suggests this position is important for protein function and shaping the peptide repertoire presented by HLA-B. Mutagenesis experiments have shown that Asn97 is important for HLA-B*27:04 surface expression; mutating this residue from Asn97 to Asp97 results in reduced surface expression and increased accumulation of unfolded protein in the ER, as well as reduced homodimers formation ⁴⁰; thus Asn97 relative to Asp97 reduces ER stress and B27-homodimer formation, yet is associated with AS risk. Moreover, work in the mouse homologue has shown that changing residue 97 (W97R) results in altered peptide specificity and affinity for β₂-microglobulin ⁴¹, and previous crystallographic studies of viral peptides bound to HLA-B27 have shown that this position influences the location of the peptide in the binding groove of the molecule ⁴². Lastly, this position was also found to be associated with HIV-1 viral control, where Val97 was found to provide the strongest protective effect to progression to AIDS, hypothesized to be through a mechanism of peptide presentation ⁴³. Asp97 is not shared by the AS-associated subtype HLA-B*27:07, where it is substituted by serine. Serine is also a polar amino-acid and the substitution would be expected to have only minor effects on the protein structure. Whilst AS is known to occur in individuals carrying HLA-B*27:07 its relative strength of association compared to other AS-associated HLA-B*27 subtypes is unknown.

In summary, with high-density genotyping of the MHC we have demonstrated independent association signals located in HLA Class I and II loci. Imputation of amino acid residues in the classical HLA Class I and II proteins resolved the peak of association at each of these loci to an amino acid residue located in the peptide-binding groove of these proteins. Refining this analysis by imputation of classical HLA alleles showed that there are multiple risk and protective haplotypes in the HLA-B locus. Further, epistatic interaction was demonstrated between ERAP1 and the HLA Class I alleles HLA-B*27 and HLA-B*40:01.

METHODS

Sample collection

All cases met the modified New York classification criteria for AS ⁴⁴. 9,069 cases and 13,578 controls were recruited through a multi-center study coordinated by the International Genetics of Ankylosing Spondylitis Consortium ⁸, and all samples were unrelated and met European ancestry criteria as detailed therein. All subjects provided written informed consent and the study was approved by the Princess Alexandra Hospital Research Ethics Committee (reference HREC/05/QPAH/221) and University of Queensland Research Ethics Committee (Project Clearance No: 2006000102). All samples were genotyped with the Illumina (San Diego, CA, USA) Infinium platform Immunochip ⁴⁵, and the current study was restricted to 7,264 markers in the MHC (chromosome 6, bps 29,602,876-33,268,403, NCBI Build 36 human genome coordinates).

Imputation

We imputed SNPs across the MHC, and classical HLA Class I and II alleles (HLA-A, -C, -B, -DRBI, -DQA1, -DQB1, -DPA1, -DPB1) and their corresponding amino acids determinants with SNP2HLA ²⁸. Samples with cumulative dosage above 2.5, across all 4-digit alleles for any one of the HLA loci, were removed from the analysis. SNPs, alleles, or amino acid residues were excluded from the analysis if the r² imputation quality score was below 0.2.

Classical allele imputation at the HLA-B locus resulted in high quality data, with a median sensitivity and specificity for imputed HLA-B alleles of 0.958 and 0.998, respectively (Supplementary Figure 1). With our imputation strategy, similar imputation performance has previously been shown for the other HLA Class I and II loci (HLA-A, -C, -DRB1, -DQA1, -DQB1, -DPA1 and -DPB1), suggesting that imputation performance for these loci was also accurate in our study ^28,43,46,47.

Statistical framework for association analysis

Associations of SNPs, HLA protein amino acid positions and non-HLA-B alleles across the MHC locus were assessed with logistic regression, assuming an additive risk effect on the log-odds scale. To account for population stratification we included as covariates 10 principal components for each individual, computed with 16,145 unlinked autosomal, non-MHC, SNPs with the tool shellfish (http://www.stats.ox.ac.uk/~davison/software/shellfish/shellfish.php). The omnibus association test compares, via likelihood ratio test, the null model H₀, where there is no risk effect at the position tested, against the alternative model H₁, where the risk effect at the position is included in the model as a fixed effect:

$$H_0:\mathit{logit}\left(y_i\right)=\theta +\sum _{k=1}^{10}{\pi }_k{p}_{i,k}$$ (1)

$$H_1:\mathit{logit}\left(y_i\right)=\theta +\sum _{a=1}^{m-1}{\beta }_a{g}_{a,i}+\sum _{k=1}^{10}{\pi }_k{p}_{i,k}$$ (2)

where y_i denotes the binary phenotype code for individual i (0=control, 1=case). The π_k parameter is the effect associated with each of the principal components, and p_i,k is the value of the k^th principal component for individual i. The θ parameter represents the sampling fraction (i.e. the logistic regression intercept). In the alternative model, a indicates the specific allele being tested and g_a,i is the dosage (imputed or genotyped) of allele a in individual i. The β_a parameter represents the effect on the log odds of disease per allele. For testing a multi-allelic locus, nucleic or amino acid positions, with m possible alleles we included m-1 β parameters, one for each allele, where the most common allele was selected as the reference allele. The likelihood ratio test that compares model H₀ with H₁ results in a test statistic that is χ² distributed with m-1 degrees of freedom.

When testing for association with imputed classical HLA alleles we defined a series of binary markers coding the presence or absence of the allele being tested, and each different allele was tested as a biallelic position as described above.

To identify independent effects we performed conditional logistic regression by including the most strongly associated position/polymorphism as a fixed effect in both the null model H₀ and the alternative model H₁. We then analyzed all positions as described above. Conditional analysis was repeated in an iterative fashion by sequentially adding the most significant positions as fixed effects until no significant position or polymorphism was observed. Allelic associations were deemed significant with p<10⁻⁵, this statistical significance threshold accounted for 5,000 independent tests using Bonferroni correction. Two tests were considered independent if the two SNPs had a pairwise correlation (r²)<0.90, which resulted in 3,252 SNPs independent tests. For the special case of HLA-B alleles where we had a higher prior probability of association we defined significance as p<10⁻³ as only 38 alleles were tested.