Robust associations of four new chromosome regions from genome-wide analyses of type 1 diabetes

Abstract

The Wellcome Trust Case Control Consortium (WTCCC) primary genome-wide association (GWA) scan¹ on seven diseases, including the multifactorial, autoimmune disease, type 1 diabetes (T1D), shows significant association (P < 5 × 10⁻⁷ between T1D and six chromosome regions: 12q24, 12q13, 16p13, 18p11, 12p13 and 4q27. Here, we attempted to validate these and six other top findings in 4,000 individuals with T1D, 5,000 controls and 2,997 family trios that were independent of the WTCCC study. We confirmed unequivocally the associations of 12q24, 12q13, 16p13 and 18p11 (P_follow-up ≤ 1.35 × 10⁻⁹; P_overall ≤ 1.15 × 10⁻¹⁴), leaving eight regions with small effects or false-positive associations with T1D. We also obtained evidence for chromosome 18q22 (P_overall = 1.38 × 10⁻⁸) from a genome-wide association study of nonsynonymous SNPs. Several regions, including 18q22 and 18p11, showed association with autoimmune thyroid disease. This study increases the number of T1D loci with compelling evidence from six to at least ten.

There is convincing evidence for association of six loci with T1D: the first, discovered over 25 years ago and having by far the largest effect, are the HLA class II genes on chromosome 6p21 in the major histocompatibility complex (MHC). Other loci include the gene encoding insulin (INS) on 11p15, CTLA4 on 2q33, PTPN22 on 1p13, the interleukin-2 receptor α chain (IL2RA, also known as CD25) region on 10p15 and, most recently, the IFIH1 (also known as MDA5) region on 2q24 (ref. ², ref. 3). These genes explain only some of the familial clustering of T1D (Supplementary Table 1 online). We have assumed for T1D³ the classical model of a small number of genes with large effects and a large number of genes with small effects^4,​⁵. If this genetic model is correct, notwithstanding a major role for (unknown) environmental factors^6,​⁷, there should be many more new genes (and pathways) to be discovered, provided sample sizes, study design and genotyping technology suffice^2,​^3,​^8-​¹³.

Here, we followed up on the most statistically significant results from two GWA studies: a nonsynonymous SNP (nsSNP) case-control study of 13,378 SNPs in 3,400 affected individuals and 3,300 controls and the WTCCC study using an Affymetrix 500K Mapping Array GWA GeneChip on 2,000 cases and 3,000 controls¹. There was a substantial overlap of samples (1,834 cases and 1,134 controls) between these studies, but we still had independent samples available for follow-up (up to 4,000 affected individuals and 5,000 controls available from the same DNA collections and 2,997 parent-child trios).

Based on the WTCCC GWA study¹, we initially genotyped 11 SNPs with TaqMan technology that had shown association with P ≤ 1.64 × 10⁻⁵ (with five having P values < 5 × 10⁻⁷) from 11 chromosome regions not previously associated with T1D. We genotyped samples from 4,000 affected individuals and 5,000 controls and from 2,997 parent-child trios that were independent of the WTCCC study (Table 1 and Supplementary Table 2 online). Four of these regions showed convincing evidence of disease association: chromosomes 12q24, 12q13, 16p13 and 18p11 in independent cases and controls (P ≤ 1.82 × 10⁻⁶), in families (P = 5.23 × 10⁻³ to 1.07 × 10⁻⁶) and overall (P = 1.15 × 10⁻¹⁴ to 1.52 × 10⁻²⁰) (Table 1, Supplementary Table 2 and Supplementary Fig. 1 online). Results from SNPs in the T1D-associated MHC region will be presented elsewhere and were excluded from the analyses presented here.

Table 1.

Follow up analysis of the Wellcome Trust Case Control Consortium genome-wide association study of 500,000 random SNPs in type 1 diabetes

Chr	Gene regiona	SNP	Cases and controls							Families		Overall P values
			WTCCC results (2,000 affected individuals, 3,000 controls)			Follow-up (4,000 affected individuals, 5,000 controls)		All (6,000 affected individuals, 6,200 controlsb)		Follow-up (2,997 parent-child trio samples)		Follow-up samples	All samples
			MAF	P (1-d.f. test)	OR (95% c.i.)	P (1-d.f. test)	OR (95% c.i.)	P (1-d.f. test)	OR (95% c.i.)	TDT Pc	RR (95% c.i.)	P	P
1p13	PHTF1– PTPN22	rs6679677d C>A	0.0962	8.03 × 10−24	1.89 (1.67– 2.13)	Confirmed previously
12q24	C12orf30	rs17696736 A>G	0.423	7.27 × 10−14	1.37 (1.27– 1.49)	1.82 × 10−6	1.16 (1.09– 1.23)	1.73 × 10−13	1.22 (1.15– 1.28)	9.20 × 10−5	1.16 (1.07– 1.25)	1.35 × 10−9	2.31 × 10−16
12q13	ERBB3 e	rs2292239 C>A	0.34	1.49 × 10−9	1.30 (1.20– 1.42)	1.89 × 10−14	1.28 (1.20– 1.36)	6.46 × 10−19	1.28 (1.21– 1.35)	2.33 × 10−4	1.15 (1.06– 1.24)	3.83 × 10−16	1.52 × 10−20
16p13	KIAA0350	rs12708716 A>G	0.322	1.28 × 10−8	0.77 (0.70– 0.84)	7.07 × 10−9	0.83 (0.78– 0.89)	7.43 × 10−14	0.81 (0.77– 0.86)	1.07 × 10−6	0.82 (0.76– 0.89)	20.8 × 10−13	2.57 × 10−18
18p11	PTPN2	rs2542151 A>C	0.163	8.40 × 10−8	1.33 (1.20– 1.49)	3.36 × 10−10	1.29 (1.19– 1.40)	1.49 × 10−14	1.30 (1.22– 1.40)	5.23 × 10−3	1.13 (1.03– 1.25)	1.23 × 10−10	1.15 × 10−14
11p15	INS f	rs3741208 C>T	0.379	2.28 × 10−7	1.25 (1.15– 1.35)	Confirmed previously
4q27	Tenr–IL2	rs17388568 G>A	0.261	6.35 × 10−7	1.27 (1.15– 1.39)	0.0231	1.08 (1.01– 1.15)	2.94 × 10−4	1.11 (1.05– 1.18)	0.0307	1.08 (1.00– 1.17)	3.32 × 10−3	5.57 × 10−5
5q14	Q8WY63	rs7722135 G>A	0.241	4.24 × 10−6	0.79 (0.71– 0.87)	0.0474	0.92 (0.86– 1.00)	1.23 × 10−3	0.90 (0.85– 0.96)	0.0653	0.94 (0.06– 1.02)	0.0149	7.64 × 10−4
2q11	AFF3– LOC150577	rs9653442 A>G	0.458	4.78 × 10−6	1.21 (1.11– 1.32)	0.0213	1.07 (1.01– 1.14)	6.23 × 10−5	1.11 (1.05– 1.17)	0.0139	1.09 (1.01– 1.18)	1.43 × 10−3	5.04 × 10−6
2p13	DQX1	rs6546909 T>A	0.132	8.53 × 10−6	1.31 (1.16– 1.47)	0.71	1.02 (0.93– 1.11)	0.0537	1.07 (1.00– 1.16)	N/A
10p11	NRP1	rs2666236 G>A	0.414	1.05 × 10−5	1.21 (1.11– 1.31)	0.126	1.05 (0.99– 1.12)	1.76 × 10−4	1.10 (1.05– 1.16)	8.89 × 10−3	1.10 (1.02– 1.19)	7.78 × 10−3	9.77 × 10−6
	2-d.f. testg					3.07 × 10−3		6.08 × 10−6		0.0236		1.46 × 10−4	6.83 × 10−8
1q32	PIK3C2B	rs12061474 G>A	0.122	1.64 × 10−5	0.75 (0.65– 0.85)	0.934	1.00 (0.91– 1.10)	0.0316	0.91 (0.84– 0.99)	N/A
10p15	RBM17– CD25	rs12251307 C>T	0.123	3.73 × 10−5	0.75 (0.66– 0.86)	Confirmed previously
2q33	CTLA4	rs3087243 G>A	0.446	8.89 × 10−5	0.85 (0.78– 0.92)	Confirmed previously
Multilocus imputation analysis
12p13	CLEC2D	rs3764021 C>T	0.471	7.19 × 10−5	0.64 (0.56– 0.72)	0.0267	0.93 (0.88– 0.99)	1.85 × 10−5	0.89 (0.86– 0.94)	0.124	0.96 (0.89– 1.04)	0.0246	4.77 × 10−5
	2-d.f. testh			5.8 × 10−8		0.0646		4.11 × 10−6		0.256		0.0330	2.11 × 10−6

The results for the 500,000-SNP scan were taken from the WTCCC¹ and are used here as reference points. We used the analyses stratified by geographical region (see Methods). Chr = chromosome; MAF = minor allele frequency in control samples; N/A = not attempted; OR = odds ratio for minor allele; 95% c.i. = 95% confidence intervals and RR = relative risk for minor allele. Confirmed previously = these regions have been published previously and thus were not regenotyped in this study.

^aFor any disease-associated region where a particular gene is named, this does not imply that this gene is causal but that it contains, based on currently available sequence and genotype information, the SNP with the highest association to disease. Further studies are required to localize causal variants with each region.

^bAll WTCCC samples, except for 1,500 blood donor samples, were used in the follow-up using TaqMan genotyping technology.

^cTDT P values are based on one-tailed tests (in other words, the null hypothesis was not rejected unless the effect was in the same direction as the original study).

^dFor chromosome 1p13, it is believed that nsSNP R620W (rs2476601) is the causal variant²⁰ for T1D susceptibility. However, note that there is a SNP, rs6679677, that is in perfect LD with R620W (r² = 1), making it and gene PHTF1 an additional candidate.

^eWe also genotyped rs11171739 from ERBB3 but found it not as significant as rs2292239 (P = 3.48 × 10⁻¹⁶; OR = 1.22 (95% c.i. = 1.17–1.29).

^fIn T1D case and control subjects, INS rs3741208 passed the WTCCC call rate filter (call rate ≥ 0.95), but as the WTCCC required all disease and control samples to pass the call rate filter as a whole, INS rs3741208 did not make the P < 5 × 10⁻⁷ level. INS rs3741208 is in LD with the INS VNTR.

^gFor chromosome 10p11, there was a significant difference between the full genotype model and the multiplicative allelic effects model (P = 0.00161) for cases and controls but not for families (P = 0.479).

^hFor chromosome 12p13, there was a significant difference between the full genotype model and the multiplicative allelic effects model (P = 0.0110) for cases and controls but not for families (P = 0.989).

We developed and applied a strategy for follow-up genotyping as a first step toward defining the disease association of the region. Our aims were to explore in a preliminary way (i) whether there were SNPs even more strongly associated with T1D in a region, (ii) whether the T1D association was due to one or more causal variants and (iii) more precisely where those variants might be within the region (Supplementary Note online).

On chromosome 18p11, the 114-kb region of strong linkage disequilibrium (LD)¹⁴ contained only one gene: PTPN2 (encoding T-cell protein tyrosine phosphatase) (Supplementary Fig. 1). We selected 11 SNPs from this interval for genotyping based on their pattern of LD with the original SNP found to be associated in the WTCCC study (rs2542151); two SNPs in introns 3 (rs1893217) and 7 (rs478582) of PTPN2 were more associated with T1D than the original WTCCC SNP and were independently associated with disease (Supplementary Table 3 online). We also resequenced nine of the ten exons of PTPN2 and 3 kb of each of the 3′ and 5′ regions, uncovering 19 new SNPs and 7 new deletion-insertion polymorphisms. We did not identify any coding variants or obvious splice mutations (Supplementary Note). However, noncoding variants could alter expression of the alternative PTPN2 45-kDa isoform, which is known to dephosphorylate STAT1 (signal transducer and activator of transcription), a major regulator of immune signaling, including in the IL-2 pathway¹⁵.

On chromosome 12q24, the most WTCCC-associated SNP , , rs17696736 (ref. ¹), is located within a large (>1.2-Mb) LD block¹⁴ that contains several genes of possible functional relevance to T1D (Table 1 and Supplementary Fig. 1 (ref. ¹)). We genotyped four SNPs for which the LD r² values with rs17696736 ranged from 0.59 to 0.82; rs3184504, an nsSNP in exon 3 of SH2B3 encoding a pleckstrin homology domain (R262W), had the highest association (P = 1.73 × 10⁻²¹; odds ratio (OR) = 1.33, 95% confidence interval (c.i.) = 1.26–1.42). This single nsSNP was sufficient to model the association of the entire region (Supplementary Table 3).

In the 16p13 region, SNP rs12708716, which was found to be associated with T1D in the WTCCC study (ref. ¹), remained the most associated after genotyping of additional SNPs (Supplementary Note). LD between HapMap SNPs and rs12708716 localized the association to intron 18 of KIAA0350 (Supplementary Fig. 1). The KIAA0350 LD block is flanked by two strong functional candidate genes, CIITA (activator of the MHC class II gene transcription) and SOCS1 (suppressor of cytokine signaling). We resequenced exonic and flanking sequences and genotyped SNPs from these two genes, but neither was responsible for the observed association in KIAA0350 (Supplementary Note). We resequenced the 24 exons and potentially regulatory 5′ and 3′ sequences of KIAA0350 and found 12 new SNPs, none of which were an obvious functional candidate (Supplementary Note). We also note that the dexamethasone-induced transcript (DEXI) may also be in the LD-defined region; further resequencing and genotyping of the entire region is required.

KIAA0350 is a widely expressed and highly conserved transcript of unknown function with a recognized putative C-type lectin domain encoded by exon 14 (according to Ensembl; see URL below). However, alignment of the domain across species suggested that this domain cannot be considered functional based on homology alone¹⁶. Further bioinformatics analyses showed that exon 12 may encode an immunoreceptor tyrosine-based activation motif (ITAM) (Supplementary Fig. 2 and the T1DBase PosterPages (see URL below) ). ITAMs bind proteins such as SH2B3 (SH2B adaptor protein 3) (also known as LNK, Lymphocyte adaptor protein) that contain SH2 signaling domains. We also noted that SH2B3 binds ERBB3 (v-erb-b2 erythroblastic leukemia viral oncogene homolog 3 (avian)) (ref. ¹⁷), which has the highest association with T1D in the other chromosome 12 region, 12q13. Therefore, we identified potential functional links between the new candidate genes and previously identified loci in interactions between antigen presenting cells (for example, dendritic cells) and T lymphocytes during T cell repertoire formation and immune inflammatory events leading to autoimmune pancreatic β-cell destruction in T1D¹³.

Of the remaining loci, two are probably false positives (2p13 and 1q32; P > 0.7 in the new 4,000 cases and 5,000 controls), and five could be true effects (P < 0.05; Table 1). We followed up on the SNPs in chromosome regions 4q27, 5q14, 2q11, 10p11 and 12p13 in the families, obtaining weak (P ≤ 0.0307) or no (P ≥ 0.0653 for 5q14 and 12p13) support for disease association (Table 1 and Supplementary Table 2).

We carried out further genotyping of the 4q27 region (Supplementary Fig. 1) because (i) it contains the IL2 gene, which has been identified as a susceptibility gene in the nonobese diabetic (NOD) mouse model of T1D⁹; (ii) the chromosome 10p15 region, containing the gene encoding the IL-2 receptor, IL2RA, is associated with T1D¹⁸ and autoimmune thyroid disease¹⁹ and (iii) using imputation, the WTCCC study reported a SNP (rs6534347) in the 4q27 region with an apparently strong association with T1D (OR = 1.30, 95% c.i. = 1.10–1.55; P = 4.48 × 10⁻⁷)¹. We resequenced the region encompassing genes IL2 through IL21 and found 178 new SNPs but observed neither IL2 and IL21 coding variants nor obvious regulatory or splice variants (Supplementary Note).Follow-up genotyping provided some support for association of this region with T1D, but finer localization within the 200-kb region on chromosome 4q27 was not possible owing to strong LD (Supplementary Fig. 1). We did not obtain support for the presence of an effect as large as OR = 1.3 in the region from IL2 to IL21; our most associated SNP was rs3136534, 3′ of IL2 (OR = 1.11, 95% c.i. = 1.05–1.18; P_{all cases and controls} = 1.62 × 10⁻⁴; Supplementary Note).

The IL-2 receptor, which is critical for immune function and regulation, is a trimeric molecule of α (IL2RA), β (IL2RB, also known as CD122) and γ (IL2RG) chains. We noted that SNP rs3218253 in intron 1 of IL2RB shows evidence of T1D association in the WTCCC study¹ (P = 1.59 × 10⁻⁴), but we found no convincing support for T1D association (Supplementary Note). This suggests that the WTCCC result was a false positive, emphasizing, along with other findings presented here, the fact that most results in a GWA study at P < 10⁻⁶ are false positives, even in a sample as large as that used in the WTCCC study.

Using 2,700 case and 3,500 control follow-up samples, we genotyped 14 out of 7,446 nsSNPs from the nsSNP GWA study that had minor allele frequencies (MAF) ≥ 0.01 and P values <1 × 10⁻³ (Table 2 and Supplementary Table 2).In addition to the previously reported PTPN22 and IFIH1 region associations^2,20,21, we found one other locus with consistent statistical support for a T1D association: rs763361 in the T lymphocyte costimulation gene CD226 (ref. ²²) on chromosome 18q22 (P_follow-up = 9.46 × 10⁻⁶ and P_overall = 1.38 × 10⁻⁸; Table 2 and Supplementary Fig. 1). The CD226 nsSNP could alter splicing of exon 7 of the gene (Supplementary Note).

Table 2.

Follow up analysis of the genome-wide association scan of 13,378 nonsynonymous SNPs in type 1 diabetes

Chr	Gene	SNP	Cases and controls							Families		Overall P-values
			nsSNP results (3,400 affected individuals, 3,300 controlsa)			Follow-up (2,700 affected individuals, 3,500 controls)		All (6,100 affected individuals, 6,800 controls)		Follow-up (2,997 parent-child trios)		Follow-up samples	All samples
			MAF	1-d.f. test P	OR (95% c.i.)	1-d.f. test P	OR (95% c.i.)	1-d.f. test P	OR (95% c.i.)	TDT Pb	RR (95%c.i.)	P	P
1p13	PTPN22	rs2476601 (R620W)	0.0939	3.26 × 10−37	1.99 (1.79– 2.22)	1.69 × 10−32	2.04 (1.81– 2.29)	2.71 × 10−63	1.98 (1.82– 2.15)	2.79 × 10−19	1.66 (1.48– 1.86)	1.33 × 10−48	2.07 × 10−80
5p13	CAPSL a	rs1445898 (R75Q)	0.452	1.45 × 10−8	0.81 (0.76– 0.87)	0.114	0.94 (0.87– 1.01)	9.61 × 10−6	0.89 (0.84– 0.94)	0.0885	0.95 (0.88– 1.02)	0.0379	8.11 × 10−6
2q24	IFIH1	rs1990760 (A946T)	0.398	2.28 × 10−7	0.82 (0.77– 0.89)	3.84 × 10−3	0.89 (0.83– 0.89)	3.27 × 10−9	0.85 (0.81– 0.90)	4.66 × 10−4	0.88 (0.81– 0.95)	1.34 × 10−5	1.77 × 10−11
20q13	C20orf168 d	rs380421e	0.387	8.84 × 10−7	0.83 (0.78– 0.90)	0.738	0.99 (0.91– 1.07)	8.25 × 10−4	0.91 (0.87– 0.96)	7.76 × 10−3	0.91 (0.84– 0.99)	0.0554	3.99 × 10−5
20q13	SPINT4 d	rs6017667 (G73S)	0.386	1.78 × 10−6	0.84 (0.78– 0.90)	N/A
18q22	CD226	rs763361 (S307G)	0.465	2.16 × 10−5	1.17 (1.09– 1.25)	1.55 × 10−5	1.18 (1.10– 1.27)	2.82 × 10−8	1.16 (1.10– 1.22)	0.0281	1.08 (1.00– 1.16)	9.46 × 10−6	1.38 × 10−8
5p13	IL7R a	rs3194051 (I356V)	0.249	1.08 × 10−4	1.17 (1.08– 1.27)	0.0199	1.11 (1.02– 1.21)	2.06 × 10−4	1.12 (1.05– 1.19)	N/A
19p13	PDE4A	rs1051738 (A497E)	0.196	1.51 × 10−4	0.84 (0.76– 0.92)	0.603	0.97 (0.88– 1.07)	3.55 × 10−3	0.91 (0.85– 0.97)	N/A
1p32	ZMYM4 c	rs12094543e	0.0122	1.92 × 10−4	1.84 (1.32– 2.55)	0.988	1.00 (0.75– 1.32)	0.0132	1.27(1.05– 1.55)	N/A
7q31	CFTR	rs213950 (V470M)	0.396	2.18 × 10−4	1.14 (1.06– 1.23)	0.0601	1.08 (1.00– 1.16)	6.93 × 10−4	1.09 (1.04– 1.15)	4.27 × 10−3	1.11 (1.03– 1.20)	1.49 × 10−3	1.95 × 10−5
5p13	IL7Ra,c	rs6897932 (T244I)	0.285	2.19 × 10−4	0.81 (0.72– 0.91)	0.0959	0.93 (0.85– 1.01)	8.07 × 10−5	0.89 (0.84– 0.94)	0.0139	0.91 (0.84– 0.99)	6.54 × 10−3	7.77 × 10−6
3q25	MED12L	rs3732765 (R1210Q)	0.378	2.87 × 10−4	0.87 (0.81– 0.94)	0.298	0.96 (0.89– 1.04)	1.26 × 10−3	0.92 (0.87– 0.97)	N/A
20q11	LBP c	rs2232613 (L333P)	0.0917	4.43 × 10−4	0.77 (0.67– 0.89)	0.738	0.98 (0.85– 1.12)	0.0164	0.89 (0.81– 0.98)	N/A
16q23	WWOX c	rs7499843e	0.303	6.33 × 10−4	1.16 (1.07– 1.27)	0.497	1.03 (0.95– 1.11)	0.0354	1.06 (1.00– 1.12)	N/A

The genotyping of 13,378 nsSNPs, which was every possible nsSNP sequence across the entire genome to which a genotyping assay could be designed for, was carried out using molecular inversion probe technology⁸,³⁰, resulting in 7,446 nsSNPs with a MAF > 0.01 scored successfully²⁷ Chr = chromosome; MAF = minor allele frequency in control samples; N/A = not attempted; OR = odds ratio for minor allele; 95% c.i. = 95% confidence intervals; RR = relative risk for minor allele .

^aD′ = 0.99 and r² = 0.13 between IL7R SNPs rs3194051 and rs6897932. CAPSL nsSNP rs1445898 is in the same region of LD as the IL7R nsSNPs and D′ = 0.81 and r² = 0.23 and D′ = 0.95 and r² = 0.43 between rs1445898 and IL7R SNPs rs3194051 and rs6897932, respectively.

IL7R is also called CD127. IL7R nsSNP rs6897932 is in the transmembrane domain and has been associated previously with multiple sclerosis (Supplementary Note).

^bTDT P values are based on one-tailed tests (that is, the null hypothesis was not rejected unless the effect was in the same direction as the original study).

^cDifferent numbers of samples were genotyped in the GWA study for rs12094543 (ZMYM4), rs2232613 (LBP) and rs7499843 (WWOX, 2,641 affected individual and 2,484 control samples) and rs6897932 (IL7R, 1,712 affected individual and 1,529 control samples).

^dD′ = 1.00 and r² = 0.99 between rs380421 (C20orf168) and rs6017667 (SPINT47), so only rs380421 was followed up.

^eThese SNPs are no longer nsSNPs according to dbSNP build 36.

In addition to CD226, we found evidence (P_{all cases and controls} ≤ 8.25 × 10⁻⁴) for nsSNPs rs1445898 (in CAPSL on 5p13), rs380421 (in C20orf168 on 20q13), rs3194051 and rs6897932 (in IL7R on 5p13) and rs213950 (in CFTR on 7q31) (Table 2).In the family collection (2,997 parent-child trios), we obtained consistent evidence of disease association for all of these nsSNPs (that is, P < 0.05 and allelic ORs in the same direction as the original study) except rs1445898 (in CAPSL; P = 0.0885) (Table 2 and Supplementary Table 2). Confirmation of these potential associations will require further studies.

The SH2B3 nsSNP rs3184504 was originally excluded from our nsSNP GWA analysis, as the genotype clustering was of marginal quality^2,8. Recently, we attempted to recover additional poorly clustered nsSNPs from the nsSNP GWA study by identifying for each nsSNP the batches of cases or controls lowering the quality of the fluorescent signal and excluding them. Although it reduced the sample size, this exclusion improved the clustering of nsSNP rs3184504 in SH2B3 (P = 2.0 × 10⁻¹²; OR = 1.30, 95% c.i. = 1.20–1.39 in 3,712 cases and 2,682 controls), making it the nsSNP with the second highest association with T1D in the study, after PTPN22 (Table 2).

One other outcome of the nsSNP scan analyses, regarding geographical variability in nsSNP allele frequencies, pertains to two potential questions. First, does population structure increase the false-positive rate in case-control association studies^1,8 (see Methods)? Second, are the allelically variable regions of genes responsible for host resistance to infectious disease, which are subject to selection pressures, also candidate susceptibility loci for autoimmune disease¹³. For example, the IFIH1 nsSNP rs1990760 (ref. ²), which is associated with T1D² and autoimmune thyroid disease (Tables 2 and 3), showed some variability in frequency across Great Britain (3.11% from north to south; P = 6.33 × 10⁻⁴; Supplementary Table 4) and is known to function as the pathogen recognition receptor (PRR) for picornavirus and enterovirus RNA molecules²³. We analyzed the nsSNPs for allele frequency differences between geographical regions of Great Britain (Supplementary Table 4). The most geographically variable nsSNP was in the PRR Toll-like receptor 1 (TLR1), N248S (rs4833095) on chromosome 4p14 (Supplementary Table 4). This region also showed extreme geographical variation in the WTCCC study¹. The TLR1 nsSNP was even more stratified than the three nsSNPs analyzed in the well-established geographically variable lactase persistence gene (LCT) (Supplementary Table 4, which has been under recent selection to allow adult consumption of cows' milk²⁴. In a heterodimeric receptor with TLR2, TLR1 recognizes lipopeptides from Mycobacteria, the causes of leprosy and tuberculosis (Supplementary Note). The SNP in TLR1 causing the N248S variantand/or other variants in LD with it in the neighboring TLR6 and TLR10 genes could have been under selection for resistance to these and other infectious diseases (Supplementary Note and Supplementary Table 4), thus helping to explain their extreme geographical variation across Great Britain and Europe and between the major ethnic groups (Supplementary Note). However, the SNP causing the TLR1 N248S variant (rs4833095) was not associated in any convincing way with T1D (Supplementary Table 4).

Table 3.

Association study of type 1 diabetes associated SNPs in 2,200 individuals with Graves' disease and 3,600 geographically-matched controls

Chr	Gene region	SNP	MAF	OR (95% c.i.)	P (1-d.f. test)
2q11	AFF3–LOC150577	rs9653442 A>G	0.467	1.10 (1.01–1.19)	0.0221
2q23	IFIH1	rs1990760 A>G	0.398	0.91 (0.84–0.99)	0.0265
4q27	Tenr–IL2	rs17388568 G>A	0.293	0.87 (0.79–0.95)	1.81 × 10−3
5p13	IL7R	rs6897932 G>A	0.268	0.96 (0.88–1.05)	0.363
5p13	CAPSL	rs1445898 G>A	0.446	0.88 (0.81–0.96)	2.72 × 10−3
10p12	NRP1	rs2666236 G>A	0.421	1.03 (0.95–1.12)	0.450
12q13	ERBB3	rs2292239 C>A	0.352	0.99 (0.92–1.08)	0.899
12q24	C12orf30	rs17696736 A>G	0.438	1.04 (0.96–1.13)	0.332
12q24	SH2B3	rs3184504 A>G	0.488	1.07 (0.98–1.16)	0.127
16p13	KIAA0350	rs12708716 A>G	0.359	0.97 (0.89–1.06)	0.497
18p11	PTPN2	rs1893217 A>G	0.167	1.13 (1.02–1.25)	0.0251
18p11	PTPN2	rs478582 A>G	0.460	0.91 (0.84–0.99)	0.0239
18q22	CD226	rs763361 G>A	0.465	1.10 (1.02–1.20)	0.0182
11p15	INS	rs689 A>T	0.300	0.95 (0.87–1.05)	0.305

MAF = minor allele frequency in control samples, OR = odds ratio for minor allele, 95% c.i. = 95% confidence interval. We can conclude that these potential associations with Graves' disease are not due to the presence of T1D in a few individuals with Graves' disease, because some SNPs (for example, rs12708716 and the very strongly T1D associated SNP rs689 in INS) do not show any evidence of association with disease in these individuals with Graves' disease. We note that for the Tenr–IL2 region SNP rs17388568, the minor allele is associated with reduced risk in Graves' disease but with susceptibility in T1D.

As the autoimmune thyroid disease Graves' disease is known to share genetic susceptibility with T1D^13,19,21, we genotyped 13 T1D-associated SNPs in 2,200 individuals with Graves' disease. We found some evidence of association for 2q11 (rs9653442, intergenic AFF3 to LOC150577), 4q27 (rs17388568, in Tenr-IL2), 5p13 (rs1445898, in CAPSL), 18p11 (rs1893217 and rs478582, in PTPN2) and 18q22 (rs763361, in CD226) (Table 3 and Supplementary Table 5 online). Except for the SNP in the Tenr-IL2 region, all alleles were associated in the same direction as in T1D. We note that the IFIH1 nsSNP, rs1990760 (ref. ²), also showed some evidence of association with Graves' disease (Table 3). These data suggest that these genes may be acting as more generalized susceptibility loci for autoimmune disease.

Some²⁵, but not all¹⁰, authors predict that in human association studies, the distribution of genotypes between unlinked disease loci will deviate from a multiplicative model, and hence, statistical power could be improved in the detection of novel loci using gene-gene interaction analyses²⁵. In case-only gene-gene interaction analyses between the new candidates and the known T1D loci, we did not find any evidence of deviation from the model of multiplicative (random) effects, sex effects or age-at-diagnosis effects (Table 4). We can model that the previously identified and newly associated SNPs account for approximately 48% of familial clustering of T1D, compared with an estimated 41% for the MHC region alone. Together, and estimating an environmental contribution of approximately 20% (ref. ⁶), about one-third remains unexplained. This residual could be due to numerous as-yet-undetected susceptibility loci, which we expect to range in relative risk effect size up to 20 %, consistent with the expected and emerging L-shaped distribution of allelic effect sizes for the ten loci so far confirmed (Fig. 1 and Supplementary Table 1). Rare causal variants will also have a role.

Table 4.

Analysis of gene-gene interactions of new type 1 diabetes loci with known disease loci

		New T1D loci
Gene	rs number	12q24 SH2B3 rs3184504	18p11 PTPN2 rs1893217	18p11 PTPN2 rs478582	12q13 ERBB3 rs2292239	16p13 KIAA0350 rs12708716
HLA class II	(11d.f.)	0.780 (2,347)	0.811 (2,527)	0.973 (2,488)	0.307 (2,551)	0.142 (2,554)
INS	rs689	0.602 (3,839)	0.333 (4,081)	0.299 (4,077)	0.904 (3,616)	0.233 (4,174)
PTPN22	rs2476601	0.529 (4,970)	0.631 (5,270)	0.537 (5,286)	0.0240 (4,695)	0.659 (5,380)
CD25	ss52580109	0.264 (4,816)	0.638 (5,076)	0.752 (5,087)	0.747 (4,519)	0.244 (5,188)
CD25	rs11594656	0.407 (4,764)	0.341 (5,023)	0.487 (5,050)	0.471 (4,481)	0.457 (5,133)
CTLA4	rs3087243	0.600 (3,806)	0.955 (4,043)	0.989 (4,044)	0.359 (3,580)	0.922 (4,137)
IFIH1	rs1990760	0.777 (5,147)	0.362 (5,463)	0.825 (5,484)	0.877 (5,075)	0.104 (5,589)
	Sex	0.355 (5,338)	0.422 (5,674)	0.849 (5,699)	0.683 (5,371)	0.776 (5,804)
	Age at diagnosis	0.0153 (5,338)	0.450 (5,674)	0.396 (5,699)	0.139 (5,371)	0.715 (5,804)

Data represent P values (with the number of affected individuals in parentheses).

Figure 1.

Odds ratios for the susceptibility allele for the ten independent type 1 diabetes associated genes or regions. The filled black bars indicate previously known associated genes and regions. The open bar indicates the IFIH1 region identified by the nsSNP genome scan² (Table 2), and the filled gray bars were identified by the WTCCC Affymetrix 500K scan¹ and confirmed by the studies reported here (Table 1). The HLA class II SNP (rs3129934) was the marker with the highest association with T1D in the MHC 25-35 Mb region in the WTCCC study¹.

Our results place the genetic basis of T1D in a genome-wide context.The known genes and the new candidates (such as PTPN2 and CD226) indicate that T1D is caused, in a permissive environment^6,7, by a combination of immune recognition of pancreatic islet antigens (including insulin), T cell repertoire development, immune regulation¹³ and other unknown pathways (for example, a pathway including the potential candidate KIAA0350 protein) that have common functional variation.

METHODS

Subjects

The 6,800 affected individuals were recruited as part of the Juvenile Diabetes Research Foundation/Wellcome Trust (JDRF/WT) Diabetes and Inflammation Laboratory's JDRF/WT British case collection (Genetic Resource Investigating Diabetes), which is a joint project between the University of Cambridge Departments of Paediatrics at the Addenbrooke's Hospital and Medical Genetics at the Cambridge Institute for Medical Research. Most affected individuals were <16 years of age at the time of collection; all were under age 17 years at diagnosis and all resided in Great Britain. The 7,000 control samples were obtained from the British 1958 Birth Cohort (B58C), an ongoing study of all people born in Great Britain during one week in 1958 (see URL below). All cases and control were of self-reported white ethnicity, with the exception of 18 cases for whom the WTCCC study found genotype evidence for non-white ethnic group status¹.

All families were of reported or self-reported white ethnicity and of European descent, with two parents and at least one affected child. The family collection consisted of 458 families from the UK Diabetes UK Warren 1 repository, 328 families from USA Human Biological Data Interchange, 250 families from Northern Ireland, 951 Finnish families, 360 Norwegian families, 412 Romanian families and 80 families from Yorkshire, UK (Supplementary Table 6). All DNA samples were collected after approval from the relevant research ethics committees, and written informed consent was obtained from the participants or their guardians.

As part of the AITD Autoimmune thyroid disease (AITD) UK National Collection, 2,200 unrelated, reported white individuals with Graves' disease were recruited. Participants were recruited from centers across the UK, including Birmingham, Bournemouth, Cambridge, Cardiff, Exeter, Leeds, Newcastle and Sheffield (Supplementary Table 6). Affected individuals were defined by the presence of biochemical hyperthyroidism together with at least one of the following: (i) a diffuse goiter on a scan, (ii) positive autoantibodies to the thyrotropin receptor (TSHR), (iii) diffuse goiter on palpation, along with thyroglobulin or thyroid peroxidase autoantibodies or (iv) thyroid eye disease (NOSPECS classification score of 2–6).

Sequencing

Polymorphisms in MHC2TA, SOCS1, KIAA0350 and PTPN2 were identified by resequencing 32 CEPH DNA samples (from Utah residents with northern and western European ancestry) in common with HapMap¹⁴. The sequencing reactions were performed using Applied Biosystems' BigDye (version 3.1) chemistry and the sequences resolved using an ABI 3700 Genetic Analyzer. Analyses of the sequence traces were performed using the Staden package, and traces were scored independently by a second operator by hand. Annotations for MHC2TA, SOCS1, KIAA0350 and PTPN2 are available from T1DBase (available only from the UK mirror site; see URLs below), together with sequence and polymorphism data the T1DBase PosterPages (see URL below) . For IL2 and IL21 and the flanking regions, polymorphisms were identified by resequencing samples from 32 individuals with T1D.

Genotyping

Follow-up SNPs in the nsSNP and WTCCC studies were genotyped using TaqMan (Applied Biosystems). All genotyping data were scored twice to minimize error; the second operator was unaware of case-control status or and family structure. Concordance data between the two GWA studies and TaqMan genotyping are shown in Supplementary Table 7. All SNPs genotyped in controls did not significantly deviate from Hardy-Weinberg disequilibrium.

Statistical analyses

All statistical analyses were performed in the Stata or R statistical systems (see URLs below) and information about the R package SNP Matrix can be found in ref. 26.

Genome-wide association nsSNP genotyping

In the nsSNP GWA study, we developed and used a clustering method to call genotypes automatically²⁷. As two research and development chips had been used in the study, we analyzed 7,446 nsSNPs (MAF ≥ 0.01) that had been on both chips or introduced on the second chip, as these had been attempted in at least 2,908 case and 2,664 control samples. We excluded 172 HLA nsSNPs from this study. Poor clustering was defined as a cluster quality score <2.8 (ref. ²⁷) or extreme deviation from Hardy-Weinberg equilibrium (χ₁² > 16; 165 SNPs dropped)⁸. GWA study data were analyzed using the R package snpMatrix²⁶, and follow-up analyses used Stata.

Logistic regression analyses

Logistic regression models were used for all case-control association tests. As the T1D cases and controls were chosen to be well matched geographically, we were able to stratify by the 12 subregions of England, Scotland and Wales to exclude the possibility of confounding by geography with little loss of power. We note that the WTCCC study shows that SNPs with significant geographical variation are limited to a small numbers of chromosome regions¹, including the TLR region on chromosome 4p14 described in the present report.

In the logistic regression analysis of a SNP, we performed a one–degree of freedom (1-d.f.) likelihood ratio test to determine whether a 1-d.f. multiplicative allelic effects model or a 2-d.f. full genotype model was more appropriate²⁸. We assumed a multiplicative allelic effects model, as it was not significantly different from the full genotype model, except for rs2666236 (NRP1). In the forward logistic regression analysis, we started by assessing the evidence against the most significant SNP being the sole variant in the region (in other words, whether this SNP alone was sufficient to model the association). For the purposes of this analysis, we did not assume any specific mode of inheritance for the most associated SNP (A>a) or for any additional SNP with significant independent effects on T1D, so genotype risks of A/A and A/a were modeled relative to the a/a genotype. We then used a 1-d.f. test for adding each of the remaining SNPs to the model by assuming multiplicative allelic effects for the additional SNPs.

2-d.f. locus-based test for pairs of SNPs

To estimate the joint effects of the two independently associated PTPN2 SNPs from the 18p11 region (rs1893217 and rs478582), we performed a 2-d.f. test by simply entering both genotypes into the logistic regression model as numerical indicator variables coded 0, 1 or 2 (in other words, as multiplicative allelic effects), representing the number of occurrences of the minor alleles A and A. When compared with the basic model, this 2-d.f. likelihood ratio test corresponds to the ‘locus-based’ score test described in ref. 11.

3-d.f. haplotype-based test

To test for a haplotype-specific effect, we compared a 3-d.f. haplotype-based test with the 2-d.f. locus-based test. The 3-d.f. haplotype-based test was performed by adding a numerical indicator variable for the ‘interaction’ term to the 2-d.f. locus-based model: coding the indicator variable as 0, 1 or 2, representing the number of occurrences of the G.G haplotype. However, this interaction term often depends on the (unobserved) haplotype phase, so for the case-control analysis, we replace this indicator variable by its expectation under the null hypothesis, θ / (1 + θ), where θ is the odds ratio measure of association between the rs1893217(G>A) and rs478582(G>A).

In the 3-d.f. haplotype-based test, the haplotype phase required by the interaction term was resolved in cases and controls together—consistent with the null hypothesis that case and control haplotypes were drawn from the same population. The interaction term was estimated using the EM algorithm without the imputation of missing genotypes.

Combined test

A score test was used to combine evidence from cases, controls and families²¹.

Gene-gene interaction

The case-only gene-gene interaction analysis, defined as deviation from the multiplicative model for the joint effects of the two genotypes, was performed using a regression model as a score test for association between genotypes in case subjects²¹. Affected sib pairs were not tested, as they are not independent. The HLA class II loci were grouped according to their genotypes using a risk-based method, rpart (S.N., J.M.M.H. and J.A.T., unpublished data; see URL below).

Geographically variable SNPs

To test for allele frequency differences between geographical regions, we used the R function snp.lhs.tests, which is part of the snpMatrix package and described in ref. 26. The SNP genotype was treated as the dependent variable (a binominal variate with two ‘trials’). Case-control status was fitted as a covariate, and region, the term to be tested, was fitted as a factor. This results in an 11-d.f. test for allele frequency differences between geographical regions.

Linkage disequilibrium

Measures of linkage disequilibrium, D′ and r², were calculated using the Haploview package, and the plots were subsequently generated and displayed through gbrowse (URLs given below) within T1DBase²⁹.

URLs

Ensembl: http://www.ensembl.org; British 1958 Birth Cohort: http://www.b58cgene.sgul.ac.uk/; T1DBase: http://t1dbase.org (and UK mirror site, http://dil.t1dbase.org); Stata: http://www.stata.com/; R: http://www.r-project.org/; rpart: http://cran.r-project.org/; David Clayton's Software: http://www-gene.cimr.cam.ac.uk/clayton/software/; Haploview: http://www.broad.mit.edu/mpg/haploview/; gbrowse: http://www.gmod.org/. T1DBase PosterPages : https://dil.t1dbase.org/page/PosterAdhoc

Accession codes

All genes are referred to by their HUGO symbol, except for Tenr on 4q27 (Entrez GeneID 132612, alias FLJ32741) and DEXI on 16p13 (Entrez GeneID 28955, alias MYLE)