Localization of type 1 diabetes susceptibility to the MHC class I genes HLA-B and HLA-A

Abstract

The major histocompatibility complex (MHC) on chromosome 6 is associated with susceptibility to more common diseases than any other region of the human genome, including almost all disorders classified as autoimmune. In type 1 diabetes the major genetic susceptibility determinants have been mapped to the MHC class II genes HLA-DQB1 and HLA-DRB1 (refs ^1-3), but these genes cannot completely explain the association between type 1 diabetes and the MHC region^4-11. Owing to the region’s extreme gene density, the multiplicity of disease-associated alleles, strong associations between alleles, limited genotyping capability, and inadequate statistical approaches and sample sizes, which, and how many, loci within the MHC determine susceptibility remains unclear. Here, in several large type 1 diabetes data sets, we analyse a combined total of 1,729 polymorphisms, and apply statistical methods—recursive partitioning and regression—to pinpoint disease susceptibility to the MHC class I genes HLA-B and HLA-A (risk ratios>1.5; P_combined=2.01×10^-19 and 2.35×10^-13, respectively) in addition to the established associations of the MHC class II genes. Other loci with smaller and/or rarer effects might also be involved, but to find these, future searches must take into account both the HLA class II and class I genes and use even larger samples. Taken together with previous studies^4-8,10-16, we conclude that MHC-class-I-mediated events, principally involving HLA-B*39, contribute to the aetiology of type 1 diabetes.

The MHC spans 4 megabases (Mb) and contains 149 genes, of which eight (the class II loci HLA-DRB1, HLA-DQB1, HLA-DQA1, HLA-DPB1, HLA-DPA1; the class I loci HLA-A, HLA-B and HLA-C) are the highly polymorphic immune response genes. There are many other candidate genes with common variants—any one of which or a combination thereof—that might also be involved in disease susceptibility. We studied 850 type-1-diabetes-affected sibling-pair (ASP) families from the United Kingdom and the United States and a first set of 2,049 type 1 diabetes patients and 1,912 controls from across Great Britain, in which we genotyped a combined total of 254 polymorphic MHC loci, including HLA-DQB1, HLA-DRB1, HLA-A and HLA-B (Table 1 and Supplementary Table 1). A second independent set of 1,050 type 1 diabetes cases and 1,125 controls was used for validation. After these analyses were completed, 1,475 additional single nucleotide polymorphisms (SNPs) in 1,964 of our type 1 diabetes cases and 2,923 controls became available as part of our collaboration with the Wellcome Trust Case Control Consortium (WTCCC)¹⁷ (Table 1).

Table 1.

Data sets used to test for unconditional single locus and MHC-class-II-independent type 1 diabetes associations in the MHC

Data set	Size of data set	Classical MHC genes genotyped in data set	SNPs and microsatellite markers genotyped in data set
Family set	850 affected sibling-pair (ASP) families from the United Kingdom and the United States	HLA-DQB1, HLA-DRB1, HLA-DQA1, HLA-DPB1, HLA-A, HLA-B and HLA-C*	27 microsatellites; 46 SNPs in candidate genes†
First case–control set	2,049 type 1 diabetes patients and 1,912 controls from Great Britain	HLA-DQB1, HLA-DRB1, HLA-A and HLA-B*	7 SNPs in the candidate genes‡; 169 nsSNPs from the GWA scan§
Second case–control set (for validation of the HLA-A association)	Additional 1,050 type 1 diabetes patients and 1,125 controls from Great Britain	HLA-DQB1, HLA-DRB1 and HLA-A	Not applicable
WTCCC case–control set	1,964 type 1 diabetes patients and 2,923 controls—overlaps with the main case– control set in 1,281 type 1 diabetes cases and 860 controls	HLA-DQB1, HLA-DRB1, HLA-A and HLA-B: available for 1,281 type 1 diabetes patients and 860 controls	1,475 SNPs from WTCCC GWA scan
WTCCC follow-up case–control set	2,484 type 1 diabetes patients and 2,019 controls	HLA-DQB1, HLA-DRB1, HLA-A and HLA-B	The eight most class-II-independently- associated SNPs from the WTCCC scan

^*See Supplementary Table 1.

^†Forty-six newly typed SNPs from fifteen candidate immune genes in or near the MHC (ITPR3, HLA-DPB1, HLA-DMA, HLA-DMB, PPP1R2P1, TAP1, TAP2, HLA-DOB, BTNL2, C6orf25, LY6G6C, NCR3, TNFA, LTA and NFKBIL1).

^‡SNPs rs241447 and rs241448 from TAP2, rs1800750 from TNFA, and rs2296336, rs3131020, rs1233478 and rs389419 in the recently reported type-1-diabetes-associated genes ITPR3 (ref. ²³), UBD and MAS1L (ref. ²⁴).

^§One-hundred-and-sixty-nine nsSNPs were part of a genome-wide association (GWA) scan of over 12,000 nsSNPs³⁰, spanning the entire 10-Mb extended MHC region. These SNPs were analysed in the case–control set and an additional 2,077 type 1 diabetes patients and 2,482 controls from Great Britain that were not typed at the MHC class II loci.

As expected^1-3,18, the strongest type 1 diabetes associations mapped to the MHC class II genes HLA-DQB1 and HLA-DRB1 (P=10^-117 and P=10^-124, respectively, for the genotype model in the families, and similarly, P<10^-300 and P=10^-300, respectively, in the first case–control set; Fig. 1 and Supplementary Table 2, see also http://dil.t1dbase.org/page/poster/mhc_association). The data did not fit the multiplicative model owing to the known epistatic interactions between alleles and dominance effects of HLA-DRB1 and HLA-DQB1 genotypes^2,19.

Figure 1. Association analyses across the MHC.

a, b, -log₁₀(P) versus chromosome position. Unconditional single locus analyses are presented for loci typed in up to 850 families (a) and in up to 2,049 cases and 1,125 controls (b, first case–control set). c, d, Analyses conditional on HLA-DRB1 and HLA-DQB1 in the families (c) and in the first case–control set (d). Results are listed in Supplementary Tables 1 and 2.

There was evidence for a secondary peak of type 1 diabetes association around HLA-B (P=3.44×10^-30 and 3.59×10^-42, in the families and the case–control set, respectively; Fig. 1 and Supplementary Table 2). To test whether these were MHC-class-II-independent effects, or merely reflected linkage disequilibrium with class II, we had to use a method that takes into account the complex multi-allelic effects of the highly disease-associated HLA-DQB1 and HLA-DRB1 genes. We compared three strategies for grouping class II genotypes in our families (Supplementary Results). The P-value for the test locus, conditional on the class II genotypes, was, at some loci, dependent on the method adopted for grouping the class II loci (Supplementary Results and Supplementary Table 3). Hence, these methods were unsatisfactory and we adopted a classification tree approach, namely, recursive partitioning^20-22 (http://cran.r-project.org). This is a risk-categorization method of grouping that differs from other risk-based grouping methods because it does not require the risk to be known a priori. The method classifies individuals as affected or unaffected using their class II genotypes by carrying out a series of binary splits on the basis of those class II genotypes, such that homogeneity with respect to disease status (risk) is maximized for each group while retaining good statistical power (see Methods). In contrast to other grouping methods considered, the recursive partitioning model provided consistent results (Supplementary Table 3).

In the families, using the optimized tree model, we found evidence of an additional effect of HLA-B (P=4.19×10^-7), HLA-DPB1 (P=2.21×10^-5) and of a SNP in the TAP2 gene (rs241448, P=5.29×10^-5; Fig. 1 and Supplementary Table 2). In the case–control set, we again found evidence of an independent effect of HLA-B (P=1.74×10^-7) over and above the combined effect of HLA-DQB1 and HLA-DRB1, as well as an independent effect of HLA-A (P=1.67×10^-10). No evidence was obtained for independent effects of the 169 non-synonymous (ns)SNPs, or the seven candidate SNPs (P>0.001, Fig. 1 and Supplementary Table 2). Specifically, no evidence was found for association of the TAP2 SNP (rs241448, P=0.074), the recently reported ITPR3 SNP²³ (rs2296336), nor for the UBD and MAS1L gene regions²⁴ (Supplementary Results and Supplementary Table 2).

After conditioning on HLA-B, HLA-DRB1 and HLA-DQB1, we had 49% power to find an effect of odds ratio 2.0 in the first case–control set, assuming a minor allele frequency of 0.1 at α=1×10^-5 with HLA-B, HLA-DRB1 and HLA-DQB1 in the model (Supplementary Methods). Hence, we conditioned on HLA-B, HLA-DRB1 and HLA-DQB1, obtaining evidence that HLA-A was independently associated with type 1 diabetes (P=2.31×10^-7), as was rs4151651 (P=8.13×10^-5), a nsSNP in the complement factor B (CFB) gene. However, we only had 13% power to test for additional associations to HLA-B, HLA-DRB1 and HLA-DQB1 in our 850 families, probably accounting for our failure to detect the HLA-A association in these families. So we sought to replicate the HLA-A result in an independent 1,050 cases and 1,125 controls (Table 1, second case–control set), obtaining convincing confirmatory evidence at P=1.77×10^-5 after conditioning on both HLA-DRB1 and HLA-DQB1 (Supplementary Table 4).

Having taken into account the combined effect of HLA-DQB1 and HLA-DRB1, as above, we found that the HLA-B*39 allele (where * represents the allele) was consistently associated with type 1 diabetes susceptibility (relative risk=3.55 (95% confidence interval 2.21–5.72) in the families; odds ratio=2.41 (95% confidence interval 1.49–3.89) in the first case–control set; Table 2 and Supplementary Table 5). Moreover, HLA-B*39 was also associated with a lower age-at-diagnosis of type 1 diabetes in the families (P=0.0022) and in the cases from the first case–control set (P=0.0021; Supplementary Table 5). Once the association of HLA-B*39 was taken into account, there was no association of other HLA-B alleles in the families (P=0.047). Nevertheless, in the first case–control set HLA-B*18 conferred susceptibility to and HLA-B*27 protection from type 1 diabetes (Table 2 and Supplementary Table 5). These HLA-B allele associations were still present after conditioning on HLA-A as well as HLA-DRB1 and HLA-DQB1 combined (Table 2).

Table 2.

Type 1 diabetes association of the HLA-B alleles (with frequencies >0.015) conditioned on the MHC class II genes and HLA-A

HLA-B allele	Allele frequency*		Families (RR (95% CI))	Case–control set (OR (95% CI))
	Number of cases (%)	Number of controls (%)	Conditioning on HLA-DQB1 and HLA-DRB1	Conditioning on HLA-DQB1 and HLA-DRB1	Conditioning on HLA-DQB1, HLA-DRB1 and HLA-A
HLA-B*39	143 (4.7)	79 (2.4)	3.55 (2.21–5.72)	2.41 (1.49–3.89)	1.92 (1.16–3.19)
HLA-B*18	202 (6.7)	128 (3.8)	1.77 (1.24–2.53)	1.83 (1.19–2.82)	1.95 (1.20–3.15)
HLA-B*13	43 (1.4)	56 (1.7)	1.17 (0.65–2.12)	1.94 (0.98–3.85)	1.88 (0.90–3.90)
HLA-B*08	788 (26.1)	461 (13.8)	1.26 (0.96–1.65)	0.95 (0.72–1.27)	1.24 (0.87–1.76)
HLA-B*55	37 (1.2)	71 (2.1)	0.73 (0.38–1.39)	1.28 (0.64–2.56)	1.19 (0.58–2.47)
HLA-B*07	249 (8.2)	465 (13.9)	1.14 (0.84–1.55)	1.25 (0.88–1.78)	1.11 (0.76–1.61)
HLA-B*44	344 (11.4)	567 (17.0)	1.00 (reference)	1.00 (reference)	1.00 (reference)
HLA-B*51	78 (2.6)	109 (3.3)	0.96 (0.61–1.52)	0.93 (0.54–1.62)	0.93 (0.52–1.66)
HLA-B*15	388 (12.8)	267 (8.0)	1.24 (0.91–1.69)	0.98 (0.70–1.36)	0.87 (0.61–1.23)
HLA-B*35	130 (4.3)	197 (5.9)	1.00 (0.70–1.43)	0.82 (0.55–1.22)	0.85 (0.55–1.31)
HLA-B*40	257 (8.5)	232 (6.9)	1.17 (0.86–1.61)	0.87 (0.61–1.22)	0.85 (0.59–1.22)
HLA-B*37	21 (0.69)	51 (1.5)	1.03 (0.51–2.08)	0.67 (0.28–1.58)	0.71 (0.28–1.79)
HLA-B*14	54 (1.8)	159 (4.8)	1.16 (0.75–1.80)	0.66 (0.39–1.12)	0.67 (0.38–1.17)
HLA-B*57	26 (0.86)	143 (4.3)	0.76 (0.37–1.55)	0.50 (0.25–1.00)	0.52 (0.26–1.07)
HLA-B*27	113 (3.7)	145 (4.3)	1.02 (0.69–1.52)	0.52 (0.34–0.80)	0.51 (0.33–0.80)

Alleles are ordered by risk in the case–control set (once HLA-DQB1, HLA-DRB1 and HLA-A have been accounted for). The most common allele, HLA-B*44, gives the tightest 95% confidence intervals, so is used as a reference. Results are given for the families (736) and the first case–control set (1,451 type 1 diabetes patients and 1,628 controls) that were successfully typed at all four classical HLA loci. Note that HLA-B*08 is not a primary effect in type 1 diabetes and is only elevated in frequency in type 1 diabetes cases because of its strong linkage disequilibrium with HLA-DRB1*03 (D′ 0.8). CI, confidence interval; OR, odds ratios; RR, relative risks.

^*Allele frequencies in the families are shown in Supplementary Table 1.

In the first case–control set, having conditioned on HLA-DQB1, HLA-DRB1 and HLA-B using allele HLA-A*02 as a reference, HLA-A*01, HLA-A*11 and HLA-A*31 were protective and HLA-A*24 was predisposing for type 1 diabetes; HLA-A*03 was more predisposing than HLA-A*11 and HLA-A*31 (Supplementary Table 4). Once these alleles were accounted for, there was no further detectable HLA-A effect in the case–control set (P=0.15). In the second case–control set, having conditioned on HLA-DRB1 and HLA-DQB1, both HLA-A*01 and HLA-A*11 were again more protective than HLA-A*02. HLA-A*24 was still the most predisposing for type 1 diabetes and may also be associated with an earlier age-at-diagnosis (P=0.01; Supplementary Tables 4 and 5).

Finally, the SNPs from the WTCCC¹⁷ scan were analysed for association with type 1 diabetes. The 20 most associated SNPs all lay within the MHC class II region, with the most associated locus, rs9273363, close to HLA-DQB1 (P=4.29×10^-298 in 1,964 cases and 2,923 controls; Fig. 2 and Supplementary Table 6; see also http://dil.t1dbase.org/page/poster/mhc_association). Once the effects of HLA-DRB1 and HLA-DQB1 are accounted for, the polymorphisms in the MHC class I region provide the strongest signals of association in the 1,281 cases and 860 controls genotyped at HLA-DRB1, HLA-DQB1 and the WTCCC SNPs (Fig. 2 and Supplementary Table 6). The most associated locus became rs3130531, located ~40 kilobases telomeric of HLA-C (P=6.74×10^-7 compared with P=0.0056 before conditioning). Once HLA-DRB1, HLA-DQB1 and HLA-B were conditioned on, however, none of the WTCCC SNPs were convincingly associated with type 1 diabetes (Fig. 2 and http://dil.t1dbase.org/page/poster/mhc_association).

Figure 2. Association analyses of 1,475 SNPs across the MHC.

a, Unconditional single locus analysis in up to 1,964 cases and 2,923 controls—the limits of the association are at 25.9 Mb (rs1324088 P=4.65×10^-6) and 34.0 Mb (rs6941621 P=9.95×10^-6). b, c, Results are presented in up to 1,281 cases and 860 controls for analyses conditioned on HLA-DRB1 and HLA-DQB1 combined (annotated SNPs were followed up in a larger case–control set; Supplementary Table 6) (b), and for analyses conditioned on HLA-DRB1 and HLA-DQB1 combined and the alleles of HLA-B with frequency >0.01 (c).

Nevertheless, to increase our statistical power, we genotyped the eight WTCCC SNPs most associated with type 1 diabetes, after HLA-DRB1 and HLA-DQB1 conditioning, in a larger set of 2,484 cases and 2,019 controls with complete HLA-DRB1 and HLA-DQB1 genotyping. The most associated locus was rs9268831 (P=6.95×10^-8; Supplementary Table 6); note that this was less significant than either HLA-B or HLA-A, which remained the most associated loci in the same data set after class II genes were accounted for (P=3.80×10^-17 and 4.59×10^-15, respectively; Supplementary Table 6). This SNP, located ~15 kb centromeric of HLA-DRA, was still associated once HLA-B was included in the model (P=5.44×10^-6). In contrast, the SNP rs3130531 at 31.3 Mb was not associated in this data set after conditioning on HLA-DRB1, HLA-DQB1 and HLA-B (P=0.16; Supplementary Table 6).

Our results indicate that, once the effect of the MHC class II genes has been accounted for, most of the detectable residual association is attributable to HLA-B and HLA-A (combining all data sets, P_combined = 2.01×10^-19 and 2.35×10^-13, respectively). We conclude that the existence of other major type 1 diabetes genes in the extended MHC is unlikely. Smaller independent effects, however, might still exist, necessitating future studies including analysis of rs9268831 (HLA-DRA), rs4151651 (CFB), HLA-C, HLA-DQA1 and the HLA-DP loci. The HLA-B and HLA-A alleles have previously been associated with type 1 diabetes^4,5,8,10,11, but unlike these previous studies, our results localize the effects to these specific loci and alleles, thereby implicating them directly in disease aetiology.

In the nonobese diabetic mouse model of type 1 diabetes, MHC class I molecules and class-I-restricted CD8⁺ T cells are central to the development of autoimmune diabetes^12-14. This correlates with the observations that in type 1 diabetes patients, cells infiltrating pancreatic islets are predominantly CD8⁺ and islet cells hyperexpress MHC class I molecules^15,16. Taken together with our results, we conclude that class-I-mediated anti-islet β-cell responses are critical in type 1 diabetes and may accelerate disease onset. This might involve both the innate and adaptive immune system^25,26. The HLA-A*02 allotype has been functionally and directly linked to T-cell autoreactivity to insulin²⁷, and our results now justify investigation of the naturally processed peptides that bind the HLA-B*39 allotype as a first step towards future evaluation of inducing tolerance to such peptides in attempts to prevent type 1 diabetes.

METHODS SUMMARY

A detailed description of the methods is given in Methods and Supplementary Information. All subjects were of white ethnicity: 850 families came from established collections; 4,126 cases were Juvenile Diabetes Research Foundation/Wellcome Trust type 1 diabetes cases and 4,394 were British 1958 Birth Cohort controls. The classical loci were typed using Dynal RELI SSO assays. All loci conformed to Hardy–Weinberg equilibrium in unaffected subjects. Family data were analysed using cases with matched pseudo-controls in regression models. Stepwise logistic regression²⁸ was used to test for effects independent of HLA-DRB1 and HLA-DQB1. The class II genotypes (rather than the alleles of HLA-DRB1 and HLA-DQB1 which do not behave multiplicatively in conferring type 1 diabetes risk) were modelled using a recursive partitioning approach^20-22 (http://cran.r-project.org). These groups created by recursive partitioning, define strata within which additional loci can be tested. Pruning of the tree, that is, assessing how much of the tree to retain, was done by cross-validation²⁹. Although when to stop pruning the tree was unclear, the number of leaves (HLA-DQB1/HLA-DRB1 groups) must be sufficient to prevent residual confounding, but not so great as to become inestimable. The recursive partitioning model was assessed using a third locus. The non-class-II loci were modelled as multiplicative effects of alleles. All analyses of the case–control set were stratified by broad geographical region^17,30. Power to detect effects independent of HLA-DRB1 and HLA-DQB1 was assessed using data sets simulated from the case–pseudo-control data (and separately from the case–control set). Crucially the groupings from the recursive partitioning model were retained. A total of 100,000 replicates were performed. The Wald test was used to assess significance and subsequently the power of the study. We had 49% power to detect odds ratio=2.0 for an allele frequency of 0.1 at α=1×10^-5 in the case–control set with HLA-DRB1, HLA-DQB1 and HLA-B in the model. We set a threshold of P<0.0001 for rejection of H₀.

APPENDIX

METHODS

Subjects

The family set comprised 850 type 1 diabetes families of white ethnicity with both parents and at least two affected children in each family, including 472 Warren type 1 diabetes families from the United Kingdom (11 with only one affected offspring) and 378 Human Biological Data Interchange (HBDI) type 1 diabetes families from the United States (five with only one affected offspring). The case–control sample comprised 4,126 type 1 diabetes patients collected as part of the JDRF/WT DIL British type 1 diabetes case collection (http://www-gene.cimr.cam.ac.uk/ucdr/grid.shtml) and 4,394 controls selected from the British 1958 Birth Cohort of people born in England, Scotland and Wales during 1 week in 1958 (http://www.b58cgene.sgul.ac.uk). Of these individuals, 2,049 cases and 1,912 controls were typed at the classical MHC loci. A further 1,445 controls were from the WTCCC’s UK Blood Service samples¹⁷. The relevant research ethics committees approved the study, and written informed consent was obtained from the participants, or their parents/guardian for those too young to consent.

Grouping of HLA-DRB1 and HLA-DQB1 alleles and genotypes: recursive partitioning

In the families, cases and matched pseudo-controls were generated²⁸ and the matching discarded so as to be able to run recursive partitioning in the recursive partitioning library in R (http://cran.r-project.org; refs ²⁰​-​^​22). Cases and pseudo-controls were matched when doing later analyses. The alleles of HLA-DRB1 and HLA-DQB1 do not behave multiplicatively in conferring type 1 diabetes risk, so, to allow for dominance effects, the genotypes of the individual loci were modelled, as this does not assume a specific mode of inheritance. All possible binary splits of the data corresponding to presence or absence of various different genotypes at HLA-DRB1 and HLA-DQB1 were considered. The split that best categorized the data as cases and pseudo-controls, which corresponds to the split that maximizes the reduction in impurity (or maximizes the homogeneity of the cases or pseudo-controls within groups), was chosen. We chose the information index, which has the form f(p) = -2plog(p) (where p = the proportion of observations in a node that for future samples belong to a different class) as the impurity measure because it is likelihood based. This process was repeated until no further improvement could be made or the minimum group size was met. The terminal ‘leaves’ of the tree represent optimized groups of the HLA-DQB1/HLA-DRB1 genotypes and so define strata within which additional MHC loci can be tested. However, the trees generated are often complex and need to be pruned. Pruning of the tree, that is, assessing how much of the tree to retain, was done by cross-validation²⁹. Nevertheless, when to stop pruning the tree was unclear, the number of leaves (HLA-DQB1/-DRB1 groups) must be sufficient to prevent residual confounding, but not so great as to become inestimable.

Pruning and evaluation of the HLA-DQB1/HLA-DRB1 trees

Initially, we used the TAP2 SNP rs241448 to assess the appropriateness of the HLA-DRB1 and HLA-DQB1 tree models and subsequently used an additional 14 loci (Supplementary Table 3). As the number of terminal leaves (HLA-DRB1/-DQB1 groups) in the model increases, we expect the effect size of the additional locus to decrease as confounding is reduced, while the 95% confidence interval will become larger (Supplementary Fig. 1). This information can be used to optimize the choice of tree by considering the best compromise between number of terminal leaves (groups of HLA-DRB1 and HLA-DQB1 genotypes), effect size and 95% confidence interval. By adding the TAP2 SNP rs241448, we tested each of the possible six pruned HLA-DRB1/HLA-DQB1 trees (that corresponded to different complexity parameters) in the family data set. The greatest disparity in effect size was seen between the two models with the minimum number of terminal leaves (four and seven groups) with relative risk = 0.7 (95% confidence interval 0.5–0.9) and the remainder with between 12 and 22 groups, relative risk = 0.6 (95% confidence interval 0.5–0.8). The effect size and 95% confidence interval at this locus was stable with respect to the number of terminal leaves (that is, HLA-DRB1/-DQB1 groups). Therefore, using the TAP2 SNP only, the 12 group model appeared to be the best compromise between effect size, 95% confidence interval and number of groups (that is, complexity of the model). Nevertheless, to verify that the number of groups used for the MHC class II model would not affect the interpretation of results, we tested the remaining 14 loci used to evaluate other grouping methods (Supplementary Results) for association, conditioning on class II effects using each of the four tree models with 12 or greater terminal leaves (Supplementary Table 3). The 12 group model exhibited one inconsistent result at MICA compared to all the other recursive partitioning models (P = 0.0005 versus P > 0.05), otherwise all loci tested were stable with respect to the number of groups in the model (Supplementary Table 3). Hence, the model with 16 terminal leaves was chosen to model the confounding effects of HLA-DRB1 and HLA-DQB1: this model had the minimum number of groups that gave results consistent with both the 18 group model and the 22 group model at all loci.

We did not use this 16 group model for the case–control collection because the way in which the two sample sets were ascertained could affect their MHC associations. The ASP families are likely to be enriched for HLA susceptibility haplotypes compared to isolated cases. They were also collected over 10 years earlier, during which time the incidence of type 1 diabetes has increased, and have a higher average age-at-diagnosis (12 years) compared to the British cases (7 years). We did, however, use the same approach to choosing the optimal tree model for the cases and controls as for the families. The MHC class II genotypes were put into the recursive partitioning library. The maximum number of groups obtained was 14 (which corresponded to a complexity parameter of 0). Of the five pruned HLA-DRB1/HLA-DQB1 trees possible (corresponding to different complexity parameters) the model with 12 terminal leaves was the best compromise between effect size and 95% confidence interval, gave consistent results across loci and was very similar to the tree with the maximum 14 terminal leaves.

The effect sizes with corresponding 95% confidence intervals are given in Supplementary Table 7 for the case–control model and the model used for the families, using an approximately neutral group as reference. Note that although the tree used for the families has 16 groups, one of these groups only contains pseudo-controls and so is not used for the analysis. Similarly one of the 12 groups used for the case–control set consisted of just 13 cases, which are dropped from the association analysis.

We then assessed the effectiveness of the HLA-DRB1/HLA-DQB1 tree model. Hence, we generated 1,000 bootstrap sample data sets, with replacement, within geographical and case–control strata. Four loci were used as the non-class-II test locus and analysed in each data set: the TAP2 SNP rs241448; the UBD SNP rs389419; the HLA-DRA SNP rs9268831; and the HLA-B Bw4/Bw6 epitope polymorphism. Supplementary Fig. 1 shows a plot of the regression coefficient for the test locus rs9268831 against number of groups for each bootstrap data set. Notably, although the effect size decreases (that is, the regression coefficient increases) with number of groups, this decrease was very modest, indicating that our effect size estimates are good. We then used the bootstrap samples to calculate 95% confidence intervals for each test locus. Reassuringly, all bootstrap 95% confidence intervals were consistent with the original 95% confidence intervals (rs241448, 95% CI_orig = 0.66–1.02 and 95% CI_boot = 0.71–1.17; rs389419, 95% CI_orig = 1.13–1.59 and 95% CI_boot = 1.17–1.84; rs9268831, 95% CI_orig = 0.62–0.83 and 95% CI_boot = 0.57–0.80; HLA-B Bw4/Bw6, 95% CI_orig = 1.02–1.45 and 95% CI_boot = 1.003–1.45) and do not lead to a different interpretation of results. Hence, we believe that the trees are effective models for the HLA-DRB1/HLA-DQB1 effects.

Testing for associations at non-class-II loci in families

We specifically wished to test the hypothesis that loci within the MHC were associated with type 1 diabetes independently of the highly associated class II genes HLA-DRB1 and HLA-DQB1. Owing to the complex relationship between these two genes, extensive linkage disequilibrium and epistatic interaction effects², we believed that a joint model was required to explain the observed association. This approach was justified because both loci were necessary to partition the data within recursive partitioning.

Forward stepwise conditional logistic regression was used to test whether any of the 83 loci typed in the MHC had an effect in addition to the HLA class II DRB1/DQB1 effect²⁸. Only individuals typed at both the class II loci and the test locus were used for the stepwise analysis. The HLA-DRB1/HLA-DQB1 loci (modelled using the recursive partitioning method described above) were placed in the regression model as confounders and other loci added; whether or not a non-HLA-DRB1/HLA-DQB1 locus improved on the model was tested by a Wald test where robust variance estimates could be applied, or else by a likelihood ratio test. The non-HLA-DRB1/HLA-DQB1 loci were modelled as alleles when the multiplicative model was appropriate, and genotypes otherwise.

Testing for non-HLA-DRB1/HLA-DQB1 loci in the case–control collection

As with the family data set, HLA-DRB1 and HLA-DQB1 were grouped by recursive partitioning (detailed above) and placed in the logistic regression model as confounders. A likelihood ratio test was used to test whether other loci added to the regression model. P < 0.0001 was considered significant. The analysis was stratified both by broad geographical region³⁰ and by the HLA-DRB1/HLA-DQB1 groups. The most significant locus was added as alleles or genotypes to the grouped class II loci and other loci added to them to test for additional effects.

The Wellcome Trust Case Control Consortium

Management committee Paul R. Burton¹, David G. Clayton², Lon R. Cardon³, Nick Craddock⁴, Panos Deloukas⁵, Audrey Duncanson⁶, Dominic P. Kwiatkowski^3,5, Mark I. McCarthy^3,7, Willem H. Ouwehand^8,9, Nilesh J. Samani¹⁰, John A. Todd² & Peter Donnelly (Chair)¹¹

Analysis committee Jeffrey C. Barrett³, Paul R. Burton¹, Dan Davison¹¹, Peter Donnelly¹¹, Doug Easton¹², David Evans³, Hin-Tak Leung², Jonathan L. Marchini¹¹, Andrew P. Morris³, Chris C. A. Spencer¹¹, Martin D. Tobin¹, Lon R. Cardon (Co-chair)³ & David G. Clayton (Co-chair)²

UK blood services and University of Cambridge controls Antony P. Attwood^5,8, James P. Boorman^8,9, Barbara Cant⁸, Ursula Everson¹³, Judith M. Hussey¹⁴, Jennifer D. Jolley⁸, Alexandra S. Knight⁸, Kerstin Koch⁸, Elizabeth Meech¹⁵, Sarah Nutland², Christopher V. Prowse¹⁶, Helen E. Stevens², Niall C. Taylor⁸, Graham R. Walters¹⁷, Neil M. Walker², Nicholas A. Watkins^8,9, Thilo Winzer⁸, John A. Todd² & Willem H. Ouwehand^8,9

1958 birth cohort controls Richard W. Jones¹⁸, Wendy L. McArdle¹⁸, Susan M. Ring¹⁸, David P. Strachan¹⁹ & Marcus Pembrey^18,20

Bipolar disorder Gerome Breen²¹, David St Clair²¹ (Aberdeen); Sian Caesar²², Katherine Gordon-Smith^22,23, Lisa Jones²² (Birmingham); Christine Fraser²³, Elaine K. Green²³, Detelina Grozeva²³, Marian L. Hamshere²³, Peter A. Holmans²³, Ian R. Jones²³, George Kirov²³, Valentina Moskvina²³, Ivan Nikolov²³, Michael C. O’Donovan²³, Michael J. Owen²³, Nick Craddock²³ (Cardiff); David A. Collier²⁴, Amanda Elkin²⁴, Anne Farmer²⁴, Richard Williamson²⁴, Peter McGuffin²⁴ (London); Allan H. Young²⁵ & I. Nicol Ferrier²⁵ (Newcastle)

Coronary artery disease Stephen G. Ball²⁶, Anthony J. Balmforth²⁶, Jennifer H. Barrett²⁶, D. Timothy Bishop²⁶, Mark M. Iles²⁶, Azhar Maqbool²⁶, Nadira Yuldasheva²⁶, Alistair S. Hall²⁶ (Leeds); Peter S. Braund¹⁰, Paul R. Burton¹, Richard J. Dixon¹⁰, Massimo Mangino¹⁰, Suzanne Stevens¹⁰, Martin D. Tobin¹, John R. Thompson¹ & Nilesh J. Samani¹⁰ (Leicester)

Crohn’s disease Francesca Bredin²⁷, Mark Tremelling²⁷, Miles Parkes²⁷ (Cambridge); Hazel Drummond²⁸, Charles W. Lees²⁸, Elaine R. Nimmo²⁸, Jack Satsangi²⁸ (Edinburgh); Sheila A. Fisher²⁹, Alastair Forbes³⁰, Cathryn M. Lewis²⁹, Clive M. Onnie²⁹, Natalie J. Prescott²⁹, Jeremy Sanderson³¹, Christopher G. Mathew²⁹ (London); Jamie Barbour³², M. Khalid Mohiuddin³², Catherine E. Todhunter³², John C. Mansfield³² (Newcastle); Tariq Ahmad³³, Fraser R. Cummings³³ & Derek P. Jewell³³ (Oxford)

Hypertension John Webster³⁴ (Aberdeen); Morris J. Brown³⁵, David G. Clayton² (Cambridge); G. Mark Lathrop³⁶ (Evry, France); John Connell³⁷, Anna Dominiczak³⁷ (Glasgow); Nilesh J. Samani¹⁰ (Leicester); Carolina A. Braga Marcano³⁸, Beverley Burke³⁸, Richard Dobson³⁸, Johannie Gungadoo³⁸, Kate L. Lee³⁸, Patricia B. Munroe³⁸, Stephen J. Newhouse³⁸, Abiodun Onipinla³⁸, Chris Wallace³⁸, Mingzhan Xue³⁸, Mark Caulfield³⁸ (London); Martin Farrall³⁹ (Oxford)

Rheumatoid arthritis Anne Barton⁴⁰, The Biologics in RA Genetics and Genomics Study Syndicate (BRAGGS) Steering Committee*, Ian N. Bruce⁴⁰, Hannah Donovan⁴⁰, Steve Eyre⁴⁰, Paul D. Gilbert⁴⁰, Samantha L. Hider⁴⁰, Anne M. Hinks⁴⁰, Sally L. John⁴⁰, Catherine Potter⁴⁰, Alan J. Silman⁴⁰, Deborah P. M. Symmons⁴⁰, Wendy Thomson⁴⁰ & Jane Worthington⁴⁰

Type 1 diabetes David G. Clayton², David B. Dunger^2,41, Sarah Nutland², Helen E. Stevens², Neil M. Walker², Barry Widmer^2,41 & John A. Todd²

Type 2 diabetes Timothy M. Frayling^42,43, Rachel M. Freathy^42,43, Hana Lango^42,43, John R. B. Perry^42,43, Beverley M. Shields⁴³, Michael N. Weedon^42,43, Andrew T. Hattersley^42,43 (Exeter); Graham A. Hitman⁴⁴ (London); Mark Walker⁴⁵ (Newcastle); Kate S. Elliott^3,7, Christopher J. Groves⁷, Cecilia M. Lindgren^3,7, Nigel W. Rayner^3,7, Nicholas J. Timpson^3,46, Eleftheria Zeggini^3,7 & Mark I. McCarthy^3,7 (Oxford)

Tuberculosis Melanie Newport⁴⁷, Giorgio Sirugo⁴⁷ (Gambia); Emily Lyons³, Fredrik Vannberg³ & Adrian V. S. Hill³ (Oxford)

Ankylosing spondylitis Linda A. Bradbury⁴⁸, Claire Farrar⁴⁹, Jennifer J. Pointon⁴⁸, Paul Wordsworth⁴⁹ & Matthew A. Brown^48,49

Autoimmune thyroid disease Jayne A. Franklyn⁵⁰, Joanne M. Heward⁵⁰, Matthew J. Simmonds⁵⁰ & Stephen C. L. Gough⁵⁰

Breast cancer Sheila Seal⁵¹, Breast Cancer Susceptibility Collaboration (UK)*, Michael R. Stratton^51,52 & Nazneen Rahman⁵¹

Multiple sclerosis Maria Ban⁵³, An Goris⁵³, Stephen J. Sawcer⁵³ & Alastair Compston⁵³

Gambian controls David Conway⁴⁷, Muminatou Jallow⁴⁷, Melanie Newport⁴⁷, Giorgio Sirugo⁴⁷ (Gambia); Kirk A. Rockett³ & Dominic P. Kwiatkowski^3,5 (Oxford)

DNA, genotyping, data QC and informatics Claire Bryan⁵, Suzannah J. Bumpstead⁵, Amy Chaney⁵, Kate Downes^2,5, Jilur Ghori⁵, Rhian Gwilliam⁵, Sarah E. Hunt⁵, Michael Inouye⁵, Andrew Keniry⁵, Emma King⁵, Ralph McGinnis⁵, Simon Potter⁵, Rathi Ravindrarajah⁵, Pamela Whittaker⁵, David Withers⁵, Panos Deloukas⁵ (Wellcome Trust Sanger Institute, Hinxton); Hin-Tak Leung², Sarah Nutland², Helen E. Stevens², Neil M. Walker² & John A. Todd² (Cambridge)

Statistics Doug Easton¹², David G. Clayton² (Cambridge); Paul R. Burton¹, Martin D. Tobin¹ (Leicester); Jeffrey C. Barrett³, David Evans³, Andrew P. Morris³, Lon R. Cardon³, Niall J. Cardin¹¹, Dan Davison¹¹, Teresa Ferreira¹¹, Joanne Pereira-Gale¹¹, Ingeleif B. Hallgrimsdóttir¹¹, Bryan N. Howie¹¹, Jonathan L. Marchini¹¹, Chris C. A. Spencer¹¹, Zhan Su¹¹, Yik Ying Teo^3,11, Damjan Vukcevic¹¹ & Peter Donnelly¹¹ (Oxford)

Primary investigators David Bentley⁵†, Matthew A. Brown^48,49, Lon R. Cardon³, Mark Caulfield³⁸, David G. Clayton², Alistair Compston⁵³, Nick Craddock²³, Panos Deloukas⁵, Peter Donnelly¹¹, Martin Farrall³⁹, Stephen C. L. Gough⁵⁰, Alistair S. Hall²⁶, Andrew T. Hattersley^42,43, Adrian V. S. Hill³, Dominic P. Kwiatkowski^3,5, Christopher G. Mathew²⁹, Mark I. McCarthy^3,7, Willem H. Ouwehand^8,9, Miles Parkes²⁷, Marcus Pembrey^18,20, Nazneen Rahman⁵¹, Nilesh J. Samani¹⁰, Michael R. Stratton^51,52, John A. Todd² & Jane Worthington⁴⁰

*See Supplementary Information for details.