A genome-wide association study of atopic dermatitis identifies loci with overlapping effects on asthma and psoriasis

Atopic dermatitis (AD) is the most common dermatological disease of childhood. Many children with AD have asthma and AD shares regions of genetic linkage with psoriasis, another chronic inflammatory skin disease. We present here a genome-wide association study (GWAS) of childhood-onset AD in 1563 European cases with known asthma status and 4054 European controls. Using Illumina genotyping followed by imputation, we generated 268 034 consensus genotypes and in excess of 2 million single nucleotide polymorphisms (SNPs) for analysis. Association signals were assessed for replication in a second panel of 2286 European cases and 3160 European controls. Four loci achieved genome-wide significance for AD and replicated consistently across all cohorts. These included the epidermal differentiation complex (EDC) on chromosome 1, the genomic region proximal to LRRC32 on chromosome 11, the RAD50/IL13 locus on chromosome 5 and the major histocompatibility complex (MHC) on chromosome 6; reflecting action of classical HLA alleles. We observed variation in the contribution towards co-morbid asthma for these regions of association. We further explored the genetic relationship between AD, asthma and psoriasis by examining previously identified susceptibility SNPs for these diseases. We found considerable overlap between AD and psoriasis together with variable coincidence between allergic rhinitis (AR) and asthma. Our results indicate that the pathogenesis of AD incorporates immune and epidermal barrier defects with combinations of specific and overlapping effects at individual loci.


INTRODUCTION
Atopic dermatitis (AD) is a chronic relapsing inflammatory skin disease that may affect 20% of children and 3% of adults. Up to 60% of children with severe AD will develop asthma (1), and infantile AD is a strong predictor of subsequent asthma development (2). These observations are consistent with the genetic correlation of 0.35 -0.55 (3,4) found between these two strongly heritable (5 -10) diseases, and indicates some shared elements of aetiology.
Despite this apparent overlap, few genes implicated in AD have been confirmed to play a role in asthma. The positionally cloned (11,12) and highly replicated (13) FLG locus may contribute towards the pathogenesis of asthma, but only within the context of AD (11,14). FLG mutations have not been found to be a risk factor for asthma alone, suggesting that 'AD plus asthma' is a genetically distinct form of asthma or a specific endophenotype of AD (11,14,15). In contrast, far greater levels of genomic coincidence have been observed between AD and psoriasis (1), another common familial inflammatory skin disease. While not mutually exclusive, AD and psoriasis rarely co-exist. They share a number of similar features, particularly in the chronic phase of AD, including epidermal hyperplasia and altered terminal differentiation of keratinocytes (16,17), as well as a number of highly divergent features, such as the level of cornification and susceptibility towards recurrent bacterial, viral or fungal infections.
In this study, we consequently have set out to map the genetic architecture of AD in a Northern European population and to explore the genetic overlap between AD and asthma in the same dataset (Fig. 1). Using existing evidence from GWAS of psoriasis, we also sought to identify genetic communalities between these two common inflammatory skin diseases.

RESULTS
The discovery phase comprised 1563 childhood onset moderatesevere AD cases of European descent recruited through secondary and tertiary referral centres and 4054 population controls of matched ancestry. The subject set incorporated independent cohorts from the UK and Ireland (881 cases, 3442 controls) and Germany (682 cases, 612 controls). Overall, 36.4% of the cases also had a doctor diagnosis of asthma (Table 1; Supplementary Material, Table S1). A total of 268 034 observed and 2 406 139 imputed single nucleotide polymorphisms (SNPs) passed quality control filters. There was little evidence of inflation of the test statistic due to population stratification, but a marked enrichment for significant values at the tail of the distribution. Genomic control (GC) inflation factors for the UK and German portions of the cohort with 10 principal components as covariates were 0.995 and 1.010, respectively, for AD, 1.007 and 1.005 for AD and asthma, and 0.977 and 0.997 for AD not asthma. For the data with GC factor ≤1.00, no further adjustments were made. For the data with GC factor .1.00, genomic control was applied prior to meta-analysis.
The association signal on chromosome 1 coincided with the FLG locus (13), the most robust known genetic risk factor for AD (22), and one of the strongest known genotype-phenotype associations of any complex trait. Independence among the chromosome 1 SNPs and the most common FLG AD risk alleles (R501X and 2282del4) was assessed by logistic regression (Supplementary Material, Table S2). Known FLG AD risk alleles accounted for the majority of variance observed at this site with no significant residual effects attributable to other associated SNPs. The association at this site was more pronounced in individuals with both AD and asthma, as compared with those with AD alone (Fig. 2). This observation is consistent with data, suggesting that the risk for asthma conferred by FLG variants is limited to asthma occurring within the context of AD (23,24).
Inclusion of asthma disease status into the GWAS model (AD plus asthma, AD no asthma) did not lead to the identification of any additional loci significant at a genome-wide threshold (P ≤ 5.0E208, Fig. 2). All 5 loci achieved peak significance when asthma status was not taken into account in the model most likely reflecting increased power through increased case numbers. Nevertheless, considerable variability was observed between loci; with the loci on chromosomes 1 and 5 achieving markedly greater significance in the 'AD plus asthma' group compared with the 'AD no asthma' group. The reverse scenario was observed for chromosome 6 ( Fig. 2).  We sought replication of our findings by selecting the highest ranking SNPs with a consistent direction of effect in UK -Irish and German cohorts from the GWA discovery stage (P , 1.0E205). Multiple SNPs ranked high from a single chromosomal region were filtered to exclude those in LD (r 2 . 0.3) with the highest ranking SNP. Potentially functional Figure 2. Manhattan plot of GWAS results. Manhattan plots relating to each strata of the GWA analysis presented in the following order: AD (all), AD and asthma, AD no asthma. The threshold for genome-wide significance (P ≤ 5.0E208) is shown on each plot as a solid red line. A suggestive threshold of 1.0E205 is shown as a solid blue line. Abbreviations: 2log10(P-value), the base 10 logarithm of the probability (P) values; AD, atopic dermatitis.
SNPs from LD bins were retained. SNPs from the known FLG locus on chromosome 1 were not included due to prior replication by others. A total of 47 SNPs (Supplementary Material,   Table S3) were carried forward to the replication cohorts, which contained 2286 cases and 3160 controls derived from Ireland, Scotland, Sweden and Germany.  Meta-analysis of replication samples using a random-effects model. c Meta-analysis of the screen and the three replication cohorts using a random-effects model.
Sixteen SNPs from three different genomic regions showed evidence of replication after Bonferroni correction for the 47 tests (P , 0.001). Associations were with the same allele and in the same direction as the discovery phase. Fourteen of the 16 SNPs achieved genome-wide significance in a meta-analysis incorporating all study collections (P , 5.0E208, Table 2). These SNPs were positioned on chromosomes 5q31 (5 SNPs), 6p21 (7 SNPs) and 11q13 (2 SNPs).

Chromosome 6p21-the MHC locus
We identified a strong signal in the major histocompatibility complex (MHC) region on chromosome 6 with a number of weakly correlated (r 2 , 0.4) markers reaching genome-wide significance. Nine of the 24 SNPs achieved significance in a meta-analysis of the replication cohorts following a correction for multiple testing. Of these, seven SNPs achieved a P-value of ,5.0E208 when tested for association in the combined discovery and replication datasets. A peak association P-value of 2.96E214 was achieved at marker rs12153855 in the class III region of the MHC. This SNP was intronic to the gene Tenascin XB (TNXB; a gene previously implicated in systemic lupus erythematosus (25)), and positioned just 83 kb from the previously published peak of association for AD within the MHC (rs176095) (26). Rs176095 achieved nominal significance in our dataset (Table 4, P ¼ 0.0031). Independence among these signals was assessed by conditional regression. When conditioned on rs176095, the effect of rs12153855 was weakened, but not eliminated (b ¼ 0.3741, SE ¼ 0.0980 and The MHC is characterized by high levels of duplication, polymorphism and LD. We applied two strategies in order to extract maximal information from the observed association. First, we used existing expression quantitative trait locus (eQTL) data derived from 496 subjects from the UK portion of the discovery cohort (27) in order to search for eQTL coincident with the seven replicated SNPs. These data were generated using RNA prepared from Epstein-Barr virus (EBV) -transformed lymphoblastoid cell lines followed by expression analysis with Illumina Human 6 BeadChips (27). The proximal genomic region (31 -33 Mb) encapsulating the seven replicated SNPs was examined for cis-or trans-acting eQTL capable of affecting any transcript represented on the Illumina 6 Human BeadChip (including .33 000 RefSeq transcripts). The seven SNPs of interest were not found to align with the most significant eQTL in the region, although effects in other cell types or physiological conditions cannot be excluded. The SNPs nevertheless did show moderate levels of association with a number of gene expression traits (Fig. 4). The most significant hits were the natural cytotoxicity triggering receptor NCR3 (rs2844509, LOD ¼ 14.7), ATPase ATP6V1G2 (rs2844509, LOD ¼ 10.5), activating transcription factor CREBL1 (four SNPs, LOD range ¼ 9.2 -8.6), and HLA-DRB5 (five SNPs, LOD range ¼ 9.1-3.5).
Polymorphisms in the DR beta chains of the Class II HLA molecule specify peptide-binding affinities. The HLA-DRB1 * 0701 allele found here to associate with AD has previously been associated with a range of T-cell mediated inflammatory diseases including podoconiosis (31) (an inflammatory disease of the lower legs that occurs after long-term skin exposure to mineral microparticles), psoriasis (32)(33)(34) and vitiligo (35). In each of these disorders, the HLA-DRB1 * 0701 allele increases the disease risk, suggesting that it may have particular relevance in response to cutaneously exposed antigens.
We found that the locus did not have an effect on the AD plus asthma phenotype. Of the 24 SNPs that attained a P-value of ,1.0E205 for AD in the discovery phase, 11 achieved P-values below this threshold when the case status was limited to individuals without asthma compared with only two SNPs in cases with asthma (Supplementary Material, Table S1). In addition, the signal reported here did not coincide with that identified in a large-scale GWAS of asthma (36), where the asthma association was positioned in the HLA-DQ region (rs9273349), and showed low levels of correlation (r 2 , 0.2) with the markers found to associate with AD in this dataset.

Chromosome 5q13
We identified a second major site of association on chromosome 5q31. Eight SNPs in this region achieved P , 1.0E205 for AD in the discovery phase. In contrast with the MHC, half of these markers (four out of eight) had P-values below this threshold when the case status was limited to the 'AD plus asthma' subgroup, whereas only one SNP retained significance in the 'AD no asthma' subgroup (Supplementary Material, Table S3). These data suggest that the chromosome 5 locus may exert overlapping effects, affecting both AD and asthma. Consistent with these observations, markers in this region (spanning between the RAD50 locus to the IL13 locus) have previously been implicated in the aetiology of asthma by GWA (36). They have also shown genome-wide association with the intermediate phenotypes of serum IgE and eosinophil numbers (37,38).
Five out of eight discovery phase SNPs retained significance in the replication data following correction for multiple testing. All five SNPs achieved a P-value of ,5.0E208 in the combined discovery and replication datasets (with a peak association P-value of 1.65E217 at rs1295686). Two distinct, but incompletely independent, association signals were observed with one block containing three of the SNPs rs2897433, rs6871536, rs2158177 and the other block containing rs1295686 and rs20541 (LD, r 2 . 0.89 within the two blocks, HapMap III, release 2 CEU + TSI), suggesting that there may be more than one susceptibility locus within the region.
One of the most strongly associated SNPs (rs20541) was a missense variant in the gene encoding Interleukin 13 (IL13 + 2044G . A); a cytokine predominantly secreted by Th2 cells and an important factor in the regulation of allergic inflammation. This variant has been repeatedly associated with asthma as well as being shown to result in a distinct isoform of IL-13 with increased biological activity (39).
A previous GWAS of AD has identified a peak marker (rs2897442) located in an intron of KIF3A in the 5q31 genomic region (18) This signal is located .50 kb away from the site of association reported here, yielded P-values of lower significance (Table 4), and demonstrates only minimal levels of LD (r 2 , 0.1, D' , 0.4 based on the CEU 1000 Genomes dataset) with the five replicated SNPs in this current study. Subsequent high-resolution mapping of the region, however, revealed a second independent signal (rs848), in only weak LD with rs2897442, positioned in IL13 and showing strong LD with the functional IL13 variant found to be associated here (rs20541). These results are consistent with the data that we present here, indicating that more than one effect may reside within this region.
Chromosome 11q13.5 We mapped a third genome-wide significant signal to an intergenic interval on chromosome 11q13.5 between the uncharacterized protein coding gene C11orf30 and the gene encoding the leucine-rich type I membrane protein, LRRC32.This region has previously been associated with AD (20). Two markers at this site (rs7130588 and rs2155219) demonstrated strong and consistent patterns of association with AD in the discovery and replication panels ( Table 2). P-values did not appear to differ systematically between the 'AD plus asthma' and the 'AD no asthma' strata of the analysis. Examination of the HapMap CEU and TSI datasets revealed high levels of LD between marker rs7130588 and the rs7927894 SNP previously shown to be associated with AD (20) (r 2 ¼ 0.91, HapMap III, release 2), indicating that these SNPs are likely to represent a single risk locus.

Analysis of previously identified risk alleles for atopic disease and psoriasis
A composite list of variants associated with atopic diseases [asthma, AD and allergic rhinitis (AR)], and the chronic inflammatory skin disease psoriasis, by means of GWAS, was compiled from the catalogue of Published GWAS (40) and associated reference lists. Where possible, we examined these variants for association with AD in our discovery dataset, either directly or by imputation. Consistent with historical evidence of genomic correspondence (1), 13 variants previously shown to associate with psoriasis by GWAS achieved low P-values in our dataset (Table 5). Approximately two-thirds of these variants showed opposite risk profiles for AD versus psoriasis. The remaining one-third exhibited an association with the same allele. These patterns may reflect genetic contributions towards the divergent and shared features of these two common dermatological diseases. A propensity towards opposing risk profiles is consistent with the infrequency of the two diseases' co-occurrence (41).
Two SNPs attained significance at a genome-wide level. Both exhibited opposing effects in AD and psoriasis, and localized to the gene IL13 (rs1295685, P ¼ 4.94E209 and rs20541, P ¼ 1.786E208). IL13 is an immunoregulatory cytokine primarily secreted by activated T helper type II (Th2) cells, triggering B lymphocytes to switch towards IgE production. An inverse association at this site may therefore contribute towards the distinct immune profiles of psoriasis and AD (type I and type II, respectively). Similarly, suggestive opposing signals were also seen at several markers positioned in and around the cytokine interleukin 12B (IL12B), a key cytokine in Th1 lineage development and survival (42)(43)(44). AD and psoriasis also differ in their susceptibility towards cutaneous bacterial and viral infection (41). Consistent with this, suggestive opposing signals were seen at a single marker proximal to interleukin 28 receptor alpha (interferon, lambda receptor, IL28RA); a gene previously implicated in viral infection and resolution (45), and at a missense mutation in TRAF3IP2; an evolutionarily conserved component of the host antiviral signalling pathway (46).
Four variants showed equivalent risk profiles in AD and psoriasis, such that both diseases associated with the same alleles. These included a common, damaging missense variant (47,48) in the gene CARD14. CARD14 is predominantly expressed in the skin, where it activates the NF-kB pathway augmenting the keratinocyte response to inflammatory cytokines.
A single variant located in the c11orf30-LRRC2 region has previously been shown to associate with AR by GWAS (rs2155219) (49). This variant yielded genome-wide significant evidence of association with AD in our discovery phase (AD P ¼ 8.17E209, OR 1.36); an effect of equivalent direction to that previously reported in AR, and of similar significance in AD patients with (P ¼ 3.67E206) and without (P ¼ 6.01E206) asthma (Table 6).
We also tested in our dataset 21 variants previously implicated in the aetiology of asthma by GWA. While none of these achieved criteria for genome-wide significance, suggestive signals were observed on chromosome 11 around the c11orf30-LRRC2 locus described above for its role in AR (AD P ¼ 5.81E207, rs7130588). The effect at this site matched the direction of that reported for asthma, suggesting that this region may harbour mechanisms capable of moderating multiple atopic disease processes.
Lower magnitude signals were observed at a number of other positions, including the chromosome 17 GSDMB-ORMDL3-GSDMA locus, SMAD3, the interleukin 6 receptor (IL6R), FLJ44477-USP38 and SUOX-IKZF4 (Table 7). These all demonstrated a common pattern of association, achieving the most significant P-values in 'AD plus asthma' cases and showing a complete absence of association in 'AD no asthma' cases. This implies that association at these sites may be primarily driven by asthma status.

DISCUSSION
We have presented here a GWAS of AD in a large panel of Northern European cases with childhood onset disease and known asthma status. We identified loci involved in the aetiology of AD at the well-established FLG locus on chromosome 1, the MHC on chromosome 6p21, the RAD50/IL13 locus on chromosome 5q31 and an 11q13.5 locus in strong LD with a previously characterized AD susceptibility site (20). We show that variation exists between these loci in terms of their contribution towards co-morbid asthma, and furthermore, that genomic overlap between AD and psoriasis can be explained by a pattern of systematic association with the same markers, but not necessarily the same alleles.
The MHC has only previously been implicated in AD in a single cohort of Japanese origin (21). Here, we observe an independent association signal in close proximity with that previously reported. The associated variants did not exhibit pronounced direct effects on gene expression in the cell type studied, indicating an alternative mode of action, indirect association or tissuespecific effects. By imputation, we are able to show that this signal, identified here in a Caucasian population of Northern European origin, is driven by independent effects of two classical HLA alleles, HLA-DRB * 0701 and HLA-B * 4402. The MHC complex exhibits high levels of population divergence (50,51) reflecting, among other factors, local adaptation to parasite and pathogen communities. The region demonstrates considerable geographical variation in LD (52) and allele frequencies, as well as regional differences in the allelic spectrum (53), implying the presence of distinct core haplotypes. It remains unclear the extent to which the deviation in the patterns of association between Caucasian and Asian populations at this locus can be attributed to differences in the underlying LD structure (and the same causal alleles) or separable, co-localized population-specific effects.
Pleiotropy is a common feature emerging from recent GWAS. Autoimmune diseases in particular have been mapped to many shared loci with variable specificities for different diseases (54). Attempts to model this phenomenon have indicated that while the number of traits affected by any one gene is likely to be small, those genes associated with higher levels of pleiotropy are likely to exert larger effects (55). As such, designation of shared loci may represent a useful metric for the prioritization of genes for further detailed mapping and functional analysis.
Here, by examination of the endophenotypes 'AD and asthma' and 'AD no asthma' and by comparison of our results with those of existing GWAS for other related phenotypes, we have found that AD is determined by a combination of specific and shared effects, with the c11orf30-LRRC2 locus on chromosome 11 in particular showing effects on AD, asthma and AR. Co-localization between AD, asthma and rhinitis was largely mediated by the same alleles, whereas in the sites of co-localization between AD and psoriasis association was mediated by a combination of shared and opposing alleles, with the most significant effects operating in opposing directions. These data are consistent with observations of clinical co-occurrence between AD, rhinitis and asthma, and relative clinical independence between AD and psoriasis.
While this study is well powered for the detection of small, common effects underpinning the genetic origins of AD, it suffers from several limitations. Asthma status was not available in all of our population-based cohorts, increasing the potential for misclassification bias. This may have resulted in a modest loss of statistical power, given the prevalence of the phenotypes studied. This effect has been mitigated through the use of strict diagnostic criteria for cases status (see Materials and Methods). Inclusion of cohorts from multiple geographic origins may lead to unrecognized population heterogeneity and false-positive errors. We have minimized this influence by removal of non-European individuals; by principal components analysis and inclusion of the top 10 components in the association model and by separate analysis of the UK -Irish and German portions of the cohort, followed by meta-analysis with application of a genomic control estimate.
Until recently, the genetic origins of AD have been characterized by a single robust major locus (FLG) and a multitude of small, poorly replicated effects. Our large GWAS has confirmed a number of known effects and demonstrated a marked variation in the phenotypic contribution of particular loci. The data presented here thus support a complex model of AD that incorporates both epidermal barrier and immune mechanisms (56), and a combination of shared and AD-specific effects.

Participants
The discovery dataset comprised 5617 individuals of European ancestry, including 1563 AD cases, and 4054 population-based controls. Individuals with AD were recruited from tertiary dermatology clinics across three centres (Great Ormond Street Hospital, London; Our Lady's Children's Hospital, Dublin; and University Hospital rechts der Isar, München). The diagnosis of AD was made by experienced dermatologists or paediatricians according to the UK diagnostic criteria (57). All AD cases had a reported disease onset in childhood and suffered from moderate-to-severe disease at the time of recruitment. AD-free controls were drawn from three population-based studies; the cross-sectional International Study of Asthma and Allergies in Childhood (ISAAC) Phase II (58), the Avon Longitudinal Study of Children and Parents (ALSPAC) (59) and the 1958 Birth Cohort (60). Asthma status (defined elsewhere (36)) was available for all datasets with the exception of the 1958 birth cohort. The 4,054 population controls were not selected to be asthma-free, in order to allow unbiased analysis of the three endophenotypes; AD, AD and asthma, AD no asthma. The replication dataset comprised 5446 individuals of European Ancestry, including 2286 AD cases, and 3160 AD-free controls. AD cases were dermatologist-diagnosed, recruited from dermatology clinics at five sites across Germany, Ireland and Sweden. Samples utilized in the replication phase were drawn from the KORA S4/F4 population-based survey (61) ISAAC Phase II, the Swedish longitudinal birth cohort BAMSE (62) and a previously described set of unselected Irish population controls derived from the Trinity College Dublin Biobank (63). Controls of German and Swedish origin were both ADand asthma-free. Disease status was not available for the Irish Biobank controls.
The study was approved by the institutional review boards at each of the participating study sites. A written informed consent for the collection and genotyping of DNA was obtained from all study participants and/or their parents/guardians where appropriate.

Genotyping
Genome-wide genotyping was performed on bar-coded LIMS (Laboratory Information Management System) tracked samples using the Illumina Sentrix HumanHap300 Genotyping BeadChip (ALSPAC, UK families, German samples, ISAAC) and the Illumina Human 610-Quad BeadChip (Irish samples, 1958BC) (Illumina, San Diego). BeadChips were processed within an automated BeadLab at the Centre National de Genotypage as per the manufacturer's instructions. Samples were subject to strict quality control criteria including assessment of concentration, fragmentation and response to PCR. A total of 20 ml of DNA aliquoted to a concentration of 50 ng/ml was used for each array. Replication phase genotyping was performed using matrix-assisted laser desorption/ionization time-of flight (MALDI-TOF) mass spectrometry (http://www. sequenom.com), or Applied Biosystems TaqMan probes (http:// www.appliedbiosystems.com/). Genotyping of Filaggrin variants R501X, 2282del4, R2447X and S3247X was performed as previously described (12,15,23). FLG genotype availability and distributions are given in Supplementary Material, Table S2.

Statistical analysis
In the discovery phase, genome-wide genotypes were used for controlling the quality of the samples. First individuals with call rates ,95% or sex inconsistencies were excluded. Next principal components analysis and identity-by-descent (IBD) estimation were done using EIGENSOFT and PLINK version 1.07, respectively. Duplicates and individuals who were possibly non-European were removed, and from each group of subjects who exhibited a strong relationship only one subject was randomly chosen (full-sibs or parent-offspring). By using this filtered sample set, we calculated quality control statistics, and SNPs with call rates ,98% or SNPs with a Hardy -Weinberg equilibrium test P -value ,1.0E206 were excluded. Finally, 268 034 experimentally genotyped SNPs with a minor allele frequency .1% were used for imputation analysis. Imputation was done using Mach and minimac software, following the instructions provided by the author (http://genome.sph.umich.edu/ wiki/Minimac). HapMap CEU release 22 reference panel was used, and the imputed genetic dosage with R-squared statistics .0.3 was used for the association analysis. As a result, 2 543 887 SNPs were imputed, and association tests were performed for 2 406 139 SNPs with an imputation quality metric R-squared value of .0.3 and a minor allele frequency of .0.01. Association was tested separately in the UK -Irish and German portions of the cohort using mach2dat software (64,65), incorporating the top 10 principal components as covariates. These data were then combined by fixed-effects meta-analysis after genomic control was applied to both datasets.
Since null mutations in the FLG gene are known to be associated with AD, we performed additional regression analysis for all SNPs around the FLG gene conditioned on the two most common FLG null alleles, R501X and 2282del4.
We also performed imputation of HLA alleles by using HLA * IMP software (29,30). To analyse imputed HLA alleles, we treated post-probabilities as genetic dosages of the alleles and estimated their effects in the regression model in conjunction with principal components as used in genome-wide analysis. The frequencies of HLA alleles were calculated as the sum of genetic dosage divided by the number of chromosomes. The logistic regression analysis was performed by R (release 2.13.0) (66). In this analysis, significance levels were Bonferronicorrected at each test group, i.e. nominal P-value of 5.0E202 divided by the number of tests carried out for one HLA locus. From the results of association analysis, we chose SNPs which showed P-values ,1.0E205 and evaluated these signals in independent replication samples. SNPs in LD (r 2 . 0.3 in Hap-Map_CEU) with another marker with a lower P-value, were not included to avoid redundancy. As a result, we chose 47 SNPs for the replication stage. In addition, SNPs categorized as candidate functional SNPs were selected from each LD bin. Association was tested in the replication sets by meta-analysis using a random-effects model as implemented in R (release 2.13.0). In order to control for known AD susceptibility alleles located within the EDC on chromosome 1, significant markers were included in a logistic regression incorporating the two most common FLG null alleles.