A Large-Scale Genetic Association Study Confirms IL12B and Leads to the Identification of IL23R as Psoriasis-Risk Genes

Overall Strategy

We conducted three sequential case-control studies (i.e., discovery, replication 1, and replication 2), to identify SNPs associated with psoriasis. In the discovery study, DNA samples from individuals with (cases) and without (controls) psoriasis were genotyped for 25,215 SNPs, with the use of disease-phenotype–based pooled DNA samples, to increase genotyping throughput and to minimize DNA consumption. The allele frequency of each SNP was determined as described elsewhere,¹⁵ and SNPs associated with psoriasis at the .05 significance level were evaluated in an independent sample set (replication 1 study), by use of a similar pooling strategy. Non–MHC-linked SNPs that were associated with psoriasis in the replication 1 study (P<.05) and had the same risk allele as in the discovery study were then individually genotyped in both the discovery and the replication 1 sample sets, to verify the results from the pooled DNA studies. SNPs that were still significant (P<.05) were then interrogated in a third independent sample set (replication 2 study). Here, we report the results of the most significant non-MHC SNP.

Subjects

The discovery samples consisted of 467 white individuals with psoriasis and 500 white individuals who served as controls. All subjects are of northern European ancestry and reside in Utah or in adjacent regions of southern Idaho, the area served by the University of Utah Hospital and Clinics. The individuals with psoriasis were recruited from the University of Utah Department of Dermatology clinics, as part of the Utah Psoriasis Initiative (UPI), and had a diagnosis of psoriasis confirmed by a UPI investigator. A full clinical workup and a questionnaire were available for each individual with psoriasis, and samples from first- or second-degree relatives were not included in the study. Samples from 500 white individuals matched for sex and ethnicity were selected from the CEPH Utah (CEU) population (n=112 grandparents) and from other University of Utah research studies (n=388) to serve as controls. The other research studies include those of colon cancer, restless legs, smoking addiction, epilepsy, and retinal disorders, and all control subjects used in the study described here were either control subjects in these other research studies or were unrelated spouses from family studies. The 112 control subjects from the CEU population were ascertained without regard to disease status. For the remaining 388 individuals, information about autoimmune disease status was available from 53%, and any individual with psoriasis or another autoimmune disease was excluded. Information about autoimmune disease status was not available for the remaining 47%; however, we expect the frequency of psoriasis and other autoimmune diseases in this control population to be similar to that of the general Utah population. Note that, if any of these control subjects carry, with increased frequency, alleles for autoimmune diseases that are also found for psoriasis, the significance of the results reported here would increase if those controls were removed.

The replication 1 sample set was collected by the Genomics Collaborative Division of SeraCare Life Sciences (GCI). Samples were included from 498 unrelated white North American individuals with a dermatologist-confirmed diagnosis of psoriasis and a Psoriasis Area and Severity Index (PASI) score and from 498 unrelated white North American individuals without a history of psoriasis or other autoimmune disorders and individually matched to the cases by sex and age. The replication 2 sample set, consisting of samples from 481 unrelated white North American individuals with a dermatologist-confirmed diagnosis of psoriasis and from 424 unrelated white North Americans without psoriasis, was used to confirm results from the two previous sample sets; samples from 293 individuals with psoriasis and 292 individuals without psoriasis were provided by GCI and met the criteria described above. BioCollections Worldwide provided an additional 188 samples from individuals with psoriasis who met the same criteria described above, as well as samples from 132 white North Americans without psoriasis.

All individuals included in this study were 18 years or older at the time they were enrolled in the sample collections. In addition, none of the subjects had undergone a bone-marrow transplant. All protocols were approved by national and/or local institutional review boards, and informed written consent was obtained from all subjects. A detailed breakdown of the clinical characteristics of the discovery and replication 1 sample sets is provided in table 1. Detailed clinical information for all individuals with psoriasis in the replication 2 sample set was limited. The female:male sex ratio was 0.89, and the average (±SD) age of onset was 28 ± 16 years.

Scan SNP Selection

Allele-specific, kinetic PCR assays were developed for a collection of 25,215 genecentric SNPs curated from dbSNP, the Applera Genome Initiative,^16,17 and the literature. SNPs were selected for inclusion if they appeared in more than one database and had a minor-allele frequency (MAF) ⩾1%. Of the SNPs, ∼70% are missense polymorphisms predicted to modify the amino acid sequences encoded by the genes. The majority of the remaining 30% of SNPs were splice-site acceptors or donors, putative transcription-factor binding sites, etc. Of the SNPs, ∼75% have MAFs ⩾5%.

Allele Frequency and Genotype Determination

SNP allele frequencies in pooled DNA samples were determined by allele-specific, kinetic PCR, as described elsewhere.^15,16 In brief, 3 ng of pooled DNA was amplified using allele-specific primers and allele frequencies calculated from the two allele-specific PCR amplification curves, each determined in duplicate. Individual genotyping was performed using allele-specific kinetic PCR on 0.3 ng of DNA, and the data were hand curated before statistical analysis, without knowledge of case-control status. Genotyping calls were made on >99% of samples, for all SNPs genotyped. Previous analyses suggest a genotyping accuracy of >99%.^18,19 Genotyping of the four most interesting SNPs described here (i.e., rs3212227, rs6887695, rs7530311, and rs11209026) in 320 of the psoriasis study samples, with use of an unrelated technology—a multiplexed oligoligation assay—showed >99.8% concordance. HLA-C genotypes were determined for the discovery samples by high-resolution sequence-based typing (performed by Atria Genetics). HLA-C genotypes were not available for the other two sample sets.

Pools were constructed using an orthogonal design in which an individual DNA sample was arrayed into multiple pools on the basis of clinically relevant phenotypes. For the unstratified analyses (i.e., all cases vs. all controls), the allele-frequency measurements in each stratum were combined and averaged on the basis of the formula

where l_j is the number of pools in which sample j appears, p_ij=1 if sample j belongs to pool i and is 0 otherwise, f_i is the allele-frequency estimate in pool i, and n is the number of distinct samples across all pools. By using an orthogonal pooling strategy based on clinically relevant phenotypes and by collapsing all the strata into an analysis of all cases versus all controls, we were essentially taking repeated measurements and were able to reduce measurement error for the comparison of all cases versus all controls.

IL12B Sequencing

To identify novel variants of IL12B, we selected DNA for resequencing from 96 randomly chosen individuals with psoriasis in the replication 1 sample set. Sequence data from all eight annotated exons and from the 5,000 bp upstream of the 5′-most exon of IL12B, which spans ∼17 kb, were extracted from the R27 draft of the Celera human genome sequence (Entrez Nucleotide [accession number NW_922784]). Primers were designed using the Primer3 program²⁰ and included the M13 forward (5′ primer) or reverse (3′ primer) universal sequencing primer-binding sequence at their 5′ ends. PCR amplification and sequencing were performed as described elsewhere.¹⁹

SNP Selection for Follow-Up Genotyping

Tag SNPs for follow-up genotyping were selected by downloading the CEU HapMap genotypes between positions 158,464,177 and 158,828,966 on chromosome 5 from the HapMap Web site (release 12, October 2004) and by running the computer program Redigo.²¹ The approach used by this program maximizes the statistical power to detect truly associated SNPs, while minimizing the number of SNPs that are genotyped, and does not use measures of LD to select the SNPs. For the power calculations, we set the following parameters: 500 cases, 500 controls, 0.95 power threshold, disease prevalence 0.01, and a conservative disease model (additive mode of inheritance with genotypic relative risk [GRR] 1.5). As described by Hu and colleagues,²¹ use of a conservative disease model produces a tag SNP set that can exhibit moderate-to-substantial LD between pairs of selected SNPs and is also robust to changes in disease models.

LD Analysis and Haplotype Estimation

The LD measure r² was calculated from unphased data with use of the LDMAX program in the GOLD package.²² Spotfire software was used to generate graphical representations of the r² LD matrix. A pseudo-Gibbs sampling algorithm from the program SNPAnalyzer version 1.0²³ was used to estimate haplotype and diplotype frequencies from unphased data. Cases and controls were treated separately in this process. The Haplo.Stats package, which calculates global P values (P values that test the null hypothesis of independence between disease status and all the haplotypes) as well as P values for each individual predicted haplotype, was used to test for association between haplotypes and disease status.²⁴ The Haplo.Stats package is particularly useful for unphased data, since it adjusts the variance on the test statistic for haplotype estimation error.

Association Statistics

A test of independence between allelic counts and disease status was performed using Fisher’s exact test.²⁵ For larger contingency tables, the Williams-corrected G test was applied, to evaluate the null hypothesis of independence between genotype counts and psoriasis status,²⁶ except when the data were sparse. In that situation, a Monte Carlo simulation approach was used to obtain appropriate P values. To explore the extent of deviation from Hardy-Weinberg equilibrium (HWE), Weir’s exact test for HWE was calculated.²⁷ Since many disease models will predict a difference in pairwise LD between samples from affected individuals and samples from unaffected individuals, we used the LD contrast test, described by Nielsen and colleagues,²⁸ to test for differences in LD patterns between cases and controls. We employed Fisher’s combined P value to perform a joint analysis across independent studies. A Mantel-Haenszel common odds ratio (OR) was calculated across independent studies.²⁶ GRR estimates (denoted as γ) were calculated using a Bayesian approach modeling (i) the prevalence of disease as a beta distribution having an expected disease prevalence of 2.3% and (ii) the genotype counts as multinomially distributed with parameters estimated from the empirical data. The Breslow-Day statistic²⁹ was used to determine whether SNPs exhibited significant heterogeneity of ORs across subphenotypes. Estimates of the population-attributable fraction were calculated as described by Walter³⁰ and Schlesselman.³¹

Since false-positive results can be problematic in any large-scale experiment in which modest nominal significance levels are used, we performed a Monte Carlo simulation to obtain the distribution of combined P values across the two studies. Two null hypotheses were considered, and the results were presented under each hypothesis. Under the first model, which assumed independence between the cases and controls and a uniform distribution of P values, we generated 26,000 stochastically independent P values, retaining the top 1,500 most significant. A second set of 1,500 P values were then generated under a uniform distribution, to simulate P values for the replication data set, and the two P values for each replicate were combined using Fisher’s combined P value. The entire in silico experiment was repeated 1,000 times. The single most significant P value from each in silico experiment was then used to generate a distribution of the 1,000 most significant experimentwide P values. This is analogous to results obtained under the Dunn-Sidak–corrected significance level (similar to a Bonferroni correction).²⁶ For the second model, we took a similar approach, except that we used the results from the pooled discovery samples to model the distribution of null P values. This approach is typically more conservative and is described in detail by Schrodi.³² The resulting gamma-null model was defined by scale and shape parameters of 1.2032 and 0.8517, respectively. Similar to other empirically based methods used in large-scale studies to obtain null or alternative models,³³ this gamma-model method concurrently adjusts for diffuse effects of population stratification in a way similar to genomic control approaches.³⁴

In an approach similar to the haplotype method,^35,36 we performed tests of association for a specific SNP conditional on genotypes at other SNPs. That is, cases and controls were both partitioned on the basis of genotypes at one SNP, and the counts at an interrogated SNP were used in a test statistic. The statistical significance of the resulting test statistic was assessed using a permutation procedure. This type of test is useful in dissecting the relative contribution of each SNP to disease association for a set of SNPs in LD.

To explore the likelihood of our data under various disease models, we calculated Bayes factor curves,^37,38 using data from all three studies. The results were compared against the likelihood of the data under a null hypothesis that accounts for diffuse population stratification. Mathematically, the Bayes factor can be written as , where R is the allelic relative risk, H_D is the disease model, which is a function of R, and H₀ is the null hypothesis.³²

Discovery Scan and Replication Data

A genecentric set of 25,215 SNPs was genotyped in pooled discovery samples consisting of 466 white North American individuals with psoriasis and 500 white North American control individuals from Utah (see “Subjects and Methods” section and table 1). DNA for 1 of the 467 patients was not available at the time the pools were constructed. This set of SNPs was highly enriched for missense, nonsense, and regulatory polymorphisms.¹⁶ The discovery scan on DNA pools yielded 1,654 markers with nominal P values <.05, 130 of which mapped within or near the MHC. These markers were then genotyped in pooled replication 1 samples consisting of 498 white, unrelated North American individuals with psoriasis and 498 white, unrelated North American controls matched for age and sex. Given that the MHC harbors the major susceptibility locus for psoriasis (the frequency of the psoriasis-risk HLA allele, C*0602, was 22.4% in our discovery cases vs. 8.5% in the controls; OR 3.10 [95% CI 2.37–4.08], P=2.37×10^-17) and given the strong LD known to exist across this region,³⁹ it is not surprising that 71 of the MHC-linked SNPs had P values <.05 in the replication 1 sample set, confirming the validity of this tiered approach to identify disease alleles. The most significant non–MHC-linked marker across both studies was rs3212227, an A→C SNP in the 3′ UTR of IL12B, located on 5q (pooled discovery samples P=5.24×10^-5; pooled replication 1 samples P=1.88×10^-8). This gene encodes the common IL-12p40 subunit of two cytokines, IL-12 and IL-23,¹⁴ and insertions and deletions in this gene have been associated with a rare, recessive form of Mendelian susceptibility to mycobacterial disease (MIM 209950).^40,41

Individual genotyping of rs3212227 in the discovery and replication 1 samples and analyses of HWE showed that we could not reject HWE for this SNP in the cases and controls in either sample set (P=.847 in discovery cases and P=.588 in discovery controls; P=.297 in replication 1 cases and P=.913 in replication 1 controls) and confirmed association of the A (major) allele with risk of psoriasis (discovery sample set OR 1.59, P=1.89×10^-4; replication 1 sample set OR 1.81, P=7.59×10^-7), with a Fisher’s combined allelic P value of 3.39×10^-9 and an OR_common of 1.70 (table 2) (see below for a detailed discussion of experimentwide significance). The common A allele is increased in cases compared with controls (∼86% vs. ∼78%–79%, respectively). The association at rs3212227 appears to be independent of HLA-C, since stratification of the discovery data by presence or absence of the HLA-C*0602 risk allele did not result in significantly different ORs (1.47 and 1.61, respectively; Breslow-Day P=.88).

Genotypic analysis of rs3212227 is presented in table 3. In a combined analysis of both sample sets, we found that the heterozygous genotype confers a modest, although not statistically significant, risk of psoriasis when compared with the minor homozygote genotype (OR_common 1.47 [95% CI 0.86–2.50]), whereas individuals homozygous for the A allele appear to be at greater risk for psoriasis relative to the minor homozygote (OR_common 2.55 [95% CI 1.52–4.28]). We also calculated Bayesian estimates of GRR (γ) for both the AA and CA genotypes (relative to the CC genotype) in each sample set, assuming that disease prevalence was beta distributed, with a mean of 2.3% (table 3). The resulting GRR estimates, which are quite similar to the ORs, are also consistent with a general model of inheritance, in which individuals carrying one copy of the A allele at rs3212227 have an increased risk of psoriasis relative to noncarriers, and individuals carrying two copies have an even greater risk.

Since false-positive results can be problematic in any large-scale experiment, we performed a Monte Carlo simulation, to obtain an experimentwide significance level. Two null hypotheses were considered, and results were presented under each. Under the first null model, which generates results similar to a Dunn-Sidak– or Bonferroni-corrected significance level,²⁶ we assumed independence between the cases and controls, generating a uniform distribution of P values (see the “Association Statistics” subsection). The resulting combined P values, which represent 1,000 independent in silico experiments, ranged from 2.73×10^-4 to 2.58×10^-8 (uniform model) (fig. 1). None of these are as significant as the calculated Fisher’s combined P value of 3.39×10^-9 for the IL12B 3′ UTR SNP, rs3212227. The second model (gamma model) (fig. 1), which uses the results from the pooled discovery samples to model the distribution of null P values and is more conservative,³² yielded 1,000 null combined P values ranging from 1.53×10^-4 to 2.30×10^-9. Only one exceeded the significance of the Fisher’s combined P value of 3.39×10^-9 for the IL12B 3′ UTR SNP. Taken together, these data suggest that association of the IL12B 3′ UTR SNP with psoriasis is unlikely to be a type I error.

IL12B Sequencing in Patients with Psoriasis

To identify novel IL12B SNPs in patients with psoriasis, we sequenced 1,136 bases of 5′ sequence, all exons and intron/exon boundaries, and 1,424 bases of 3′ sequence of IL12B in 96 individuals with psoriasis from the replication 1 sample set. On average, we successfully sequenced 89 individuals (coverage ranged from 70 to 96 individuals) and identified six SNPs, only one of which was not found in public databases. The minor allele of this novel 3′ UTR SNP (C158674941T; ss52085990) was observed on only one chromosome in the 96 individuals sequenced and was not genotyped further in this study.

Single-Marker Analysis of the IL12B Region

To further explore the association signal in the IL12B region, we first chose to determine whether other alleles within the IL12B region were associated with psoriasis by genotyping a combination of tag SNPs, which had the highest average power to detect disease-associated variants under a conservative disease model, and functional SNPs in our discovery and replication 1 sample sets. Of the 32 tag SNPs we selected using the program Redigo,²¹ from a total of 44 CEU phase I HapMap SNPs⁴² in a region of 364.8 kb surrounding IL12B on chromosome 5, 27 produced high-quality individual genotyping data in both the discovery and replication 1 sample sets, with minimal loss in average power (0.89 for 27 tag SNPs vs. 0.90 for 32 tag SNPs). We also individually genotyped a putative promoter polymorphism upstream of IL12B—rs17860508,^43,44 which has been shown to affect IL-12 and IL-12Bp40 protein levels in stimulated cells, albeit inconsistently^44–47—as well as two validated (SeattleSNPs resource) IL12B missense SNPs (rs3213119 and rs3213096) (fig. 2 and table 4), for a total of 30 additional SNPs in the IL12B region. An exact test of HWE of the genotypic data for these 30 variants, with the analysis performed separately for cases and controls in each sample set (data not shown), identified only two instances where a marker was not in HWE at the P<.05 significance level (rs7721001 in the discovery controls, P=.005; rs270654 in the replication cases, P=.020).

All 30 variants were analyzed for single-marker association with psoriasis susceptibility. In addition to the original hit, three other SNPs—rs3212220, rs7709212, and rs6887695, all of which lie upstream of the IL12B ATG start codon—had P values <.05 in an allelic test, with consistent ORs (i.e., the psoriasis-risk allele was the same) in both sample sets (table 4). The results for the putative promoter polymorphism, rs17860508, did not reach statistical significance in the discovery sample set (P=.056) and were only marginally significant in the replication 1 sample set (P=.031). The two missense SNPs, which were relatively infrequent in our sample sets (rs3213119, F298V: allele frequency ∼3%–4%; rs3213096, I33V: allele frequency <1%), were not significantly associated with disease.

LD and Haplotype Analysis of the IL12B-Region SNPs

We examined the LD patterns in this region by calculating the pairwise r² values between all 31 markers separately for the cases and the controls in each study (fig. 3). The original hit, rs3212227, is in high LD with one of the tag SNPs, rs3212220 (r²>0.97 in the cases and the controls of both sample sets), and is in modest LD with rs1897565 (r²=0.25–0.33), rs17860508 (r²=0.12–0.23), rs7709212 (r²=0.28–0.40), and rs6887695 (r²=0.15–0.31) but exhibits little to no LD with the other 25 SNPs. Four other marker pairs show moderate LD (0.50<r²<0.80) in the cases and the controls of both sample sets; however, the majority of the markers exhibited minimal LD between one another.

To focus on regions of interest, we generated three-marker sliding-window haplotypes for the 28 markers with a MAF >5% for the cases and the controls from both sample sets. Global P values were determined using the Haplo.Stats program,²⁴ to assess the overall haplotype-frequency differences between the cases and the controls for each 3-SNP window (fig. 4). Two peaks with highly significant global P values (P<.0001) were observed in both sample sets. The first peak (w14-17) centers on the original marker, rs3212227, and on rs3212220, which, as noted above, are in strong LD with one another. The second peak (w22-24) centers on rs6887695. These three markers also showed the most-significant single-marker association (table 4).

On the basis of these results, we focused subsequent analyses on the 13 SNPs with a MAF >5% and encompassed by these two peaks (rs11744690 to rs4921226 in table 4). Haplotypes for all 13 SNPs across this 144-kb region were predicted, and association between these haplotypes and disease status was determined using Haplo.Stats. This was followed by the stepwise removal of each SNP and reanalysis of the haplotypes and their association with disease, to define the minimum number of SNPs that explained the observed association (data not shown). This systematic approach indicated that 2 of these 13 SNPs—our original hit (rs3212227) and rs6887695, which exhibit some LD (r²=0.15–0.31)—could explain the majority, if not all, of the association with psoriasis observed in both sample sets (table 5). The 2-SNP haplotype containing the major allele of both SNPs (A at rs3212227 and G at rs6887695) appears to be associated with risk in both sample sets (discovery sample set OR 1.42, P=3.27×10^-4; replication 1 sample set OR 1.48, P=1.35×10^-5), whereas a second haplotype containing the minor allele at both SNPs (C-C) confers protection (discovery sample set OR 0.63, P=5.8×10^-4; replication 1 sample set OR 0.46, P=2.56×10^-9). The frequencies of the other two haplotypes, which contain a risk allele at one SNP and a protective allele at the second SNP, did not exhibit substantial differences between cases and controls.

Because these two markers (rs3212227 and rs6887695) exhibited some LD, we performed genotype-conditioned analyses similar to that described by Thomson and colleagues,^35,36 to see if either SNP was associated with disease independent of the other SNP. To do this, we calculated a summary statistic for each sample set, using the genotypes at one marker conditioned on the other marker, and evaluated the significance of these observed results through a permutation test. The results of 20,000 permutations for each sample set demonstrated that association of rs3212227 with psoriasis given rs6887695 (and rs6887695 given rs3212227) was significant at the .001 level in both sample sets, indicating that each SNP contributes independently to psoriasis risk.

Replication of IL12B Results in a Third Sample Set and in Combined Analyses

We then genotyped rs3212227 and rs6887695 in a third independent sample set consisting of 481 white North American individuals with psoriasis and 424 white North American controls. Allele frequencies were comparable across all three sample sets, and single-marker analysis confirmed that both rs3212227 and rs6887695 were associated with psoriasis in this independent sample set (P=.014 and P=.007, respectively) (table 4). The combined analysis for each individual SNP across all three sample sets was highly significant (combined P [P_comb]=7.85×10^-10 and P_comb=4.08×10^-8, respectively). Haplotypes were also estimated in the third sample set, confirming the common IL12B risk haplotype (OR 1.31, P=6.00×10^-3) as well as the less frequent protective haplotype (OR 0.70, P=7.00×10^-3) (table 5). The combined analyses across all three sample sets were highly significant; the frequency of the common risk A-G haplotype is ∼71%–72% in cases versus ∼64% in controls (OR_common 1.40, P_comb=8.11×10^-9), and the frequency of the protective C-C haplotype is ∼10%–12% in cases versus ∼17%–19% in controls (OR_common 0.58, P_comb=5.65×10^-12) (table 5). Importantly, the other two haplotypes carrying a risk allele at one SNP and a protective allele at the other SNP were not associated with psoriasis. Together, these data provide convincing statistical evidence that the IL12B region on 5q31.1-33.1 harbors a psoriasis-susceptibility locus.

Analysis of IL12B-Related Genes

IL12B encodes the common IL-12p40 subunit of two heterodimeric cytokines, IL-12 and IL-23, each with a distinct subunit encoded by the genes IL12A (IL-12p35) and IL23A (IL-23p19). The IL-12 and IL-23 receptors also share a common subunit encoded by IL12RB1, in addition to their unique components—IL23R for the IL-23 receptor and IL12RB2 for the IL-12 receptor. To determine whether variants of these candidate genes might be associated with psoriasis, we individually genotyped 17 SNPs (7 in IL12A, 2 in IL23A, 4 in IL12RB1, 1 in IL12RB2, and 3 in IL23R) from our collection of assays in the discovery sample set. Two SNPs, rs7530511 in IL23R (OR 0.67, P=.006) and rs2914119 in IL12A (OR 0.78, P=.045), were nominally significant in the single-marker analyses (P<.05) (table 6). We also estimated haplotypes for all SNPs in each gene (data not shown) and tested for association with disease with the use of Haplo.Stats. Since IL23R and IL12RB2 lie directly adjacent to each another on chromosome 1, all four SNPs in these two genes were used together in the haplotype analysis of this region. If the global P value for all the SNPs in a particular gene was significant, we then evaluated the contribution of each SNP to this signal, using a systematic stepwise approach in which each SNP was removed and Haplo.Stats was rerun. Using this approach, we found that a second IL23R SNP, rs11209026, in combination with rs7530511, was associated with psoriasis risk (global P=.004) (table 7). The most common haplotype marked by these two SNPs, C at rs7530511 and G at rs11209026, was increased in patients compared with controls (85.2% vs. 79.3%, OR 1.5, P=9.48×10^-4) (table 7) and was more significant than each SNP individually (table 6).

The two IL23R SNPs and the IL12A SNP were then individually genotyped in the replication 1 sample set (table 6). The IL12A SNP did not replicate (P=.955), and a meta-analysis of both discovery and replication 1 samples yielded a Fisher’s combined P value of .178 (data not shown). However, the common IL23R haplotype, C at rs7530511 and G at rs11209026, was associated with psoriasis risk in the replication 1 sample set (OR 1.37, P=.006), as well as in the replication 2 sample set (OR 1.45, P=.003) (table 7), and the combined analysis across all three sample sets was highly significant (OR_common 1.44, P_comb=3.13×10^-6), providing strong statistical evidence for a psoriasis-susceptibility gene on 1p31. It is interesting to note that these two IL23R SNPs are both missense SNPs (rs7530511, L310P; rs11209026, Q381R) with an r² value between them of <0.012 in the cases and controls of all three sample sets (data not shown). The common risk haplotype carries a proline at amino acid 310 and an arginine at amino acid 381. Both of these amino acids are conserved in the chimpanzee, mouse, rat, cow, dog, and chicken.

Additional Statistical Support of IL12B and IL23R Haplotypes

Complementing traditional homogeneity tests, significant differences in LD patterns between cases and controls can also indicate disease association. We applied the LD contrast test described by Nielsen and colleagues²⁸ to pairs of SNPs in IL12B and IL23R. The results show departures between case and control pairwise values of D at rs3212227-rs6887695 for IL12B, as expected under a disease model: P_{cs-dis,ct-dis}=.0129, P_{cs-rep1,ct-rep1}=5.16×10^-8, P_{cs-rep2,ct-rep2}=.0497 (where “cs” denotes cases,“ct” denotes controls, “dis” denotes the discovery sample, and “rep1” and “rep2” denote replication sample sets 1 and 2, respectively), and P_comb=1.05×10^-8. With regard to rs7530511-rs11209026 at IL23R, this test yielded similar significant P values: P_{cs-dis,ct-dis}=.0085, P_{cs-rep1,ct-rep1}=.0019, P_{cs-rep2,ct-rep2}=.0019, and P_comb=5.30×10^-6. Since population stratification and other sampling irregularities can generate departures of pairwise LD measures from their null distribution, we performed the LD contrast test on all combinations of control-control (i.e., discovery controls vs. replication 1 controls, discovery controls vs. replication 2 controls, etc.), case-case, and case-control samples, to better understand the significance of our results. In all instances for both genes, case-control comparisons were statistically significant (P<.05), whereas the case-case and control-control comparisons were not (results not shown).

Bayes factor curves were calculated for the risk rs3212227-rs6887695 IL12B haplotype and the risk rs7530511-rs11209026 IL23R haplotype. Bayes factors were adjusted for population stratification and were combined across the three independent studies.³³ The results show a maximum log₁₀B(R) of 6.85 for the IL12B haplotype data, occurring at R=1.41. Given equal prior probabilities for disease and null models, the probability of the disease hypothesis, given the data, is ∼7-million-fold larger (at a disease model of R=1.41) than the probability of the null hypothesis conditioned on the data. Further, the log₁₀B(R)>4, as R spans 1.15 to 1.73, testifying to the strength, robustness, and specificity of the result. Similar results were observed for the IL23R haplotype data, with a maximum log₁₀B(R) of 5.78 at R=1.47. Log₁₀B(R) remains >4 for the relatively narrow range of 1.20<R<1.80.

Diplotype Analyses

To better understand how the IL12B and IL23R genes influence an individual’s risk of psoriasis, we assessed the impact of diplotypes at each locus independently and then jointly on psoriasis risk. Of the 10 possible 2-SNP IL12B diplotypes, 9 were observed in our samples (table 8). The combined results from all three sample sets suggest that there is a hierarchy of risk depending on an individual’s IL12B diplotype. Psoriasis susceptibility appears to be conferred by two copies of the common risk haplotype (A-G/A-G OR_common 1.52, P_comb=6.14×10^-7), which is found in ∼48%–51% of cases versus 38%–40% of controls. One copy of the risk haplotype (A-G) in combination with a copy of a “neutral” haplotype (A-C or C-G) does not appear to substantially affect psoriasis risk; however, individuals who carry copies of both the risk and the protective haplotypes (A-G/C-C) appear to be protected (OR_common 0.67, P_comb=1.39×10^-4). Individuals homozygous for the protective haplotype (C-C/C-C) are relatively uncommon (∼1% of cases vs. ∼2–3% of controls); however, our data suggest that their risk of psoriasis may be even lower (OR_common 0.38, P_comb=.018), although, given the stochastic variability around this estimate, additional samples are required to fully address this issue. The population-attributable fraction^30,31 summarized across all three sample sets for the homozygous susceptible A-G/A-G IL12B diplotype, relative to all other diplotypes collapsed together, was calculated to be 17.1% (95% CI 11.4%–22.5%).

At the IL23R locus, 7 of the 10 possible 2-SNP diplotypes were observed (table 9). The data from all three sample sets suggest that individuals homozygous for the susceptible C-G haplotype are at increased risk of psoriasis (OR_common 1.48, P_comb=2.24×10^-5), whereas all other diplotypes appear to be neutral or protective. The estimated population-attributable fraction summarized across all three sample sets for the IL23R homozygous risk diplotype (C-G/C-G), relative to all other diplotypes collapsed together, is 23.5% (95% CI 14.9%–31.5%).

To assess the joint impact of the IL12B and IL23R susceptible haplotypes, we determined the distribution of diplotypes at both genes, comparing cases and controls in all three sample sets (table 10). Given the large number of possible diplotypes across these two loci relative to the number of cases and controls, we chose to concentrate our analysis on the risk haplotypes at each locus (A-G for IL12B and C-G for IL23R) and to group the three nonrisk haplotypes at each locus into a single haplotype (indicated as “X” for both loci in table 10). The combined analysis shows that individuals homozygous for both the IL12B and IL23R predisposing haplotypes, ∼35%–37% of cases versus ∼24%–27% of controls, appear to be at increased risk of psoriasis (OR_common 1.66, P_comb=1.33×10^-8). A much larger sample size is needed to thoroughly evaluate the effect of all possible two-locus diplotypes; however, a preliminary meta-analysis of the available data (using a Bayesian approach, assuming no epistasis between loci and an expected prevalence of psoriasis equal to 2% in the general population) suggests at least a threefold difference in risk between individuals who carry one copy of the protective C-C IL12B haplotype and one copy of the protective C-A IL23R haplotype, representing 4.4% of the general population (C-C/X–C-A/X, expected psoriasis risk 0.84%) relative to individuals who are homozygous for risk haplotypes at both loci, representing 25.5% of the general population (A-G/A-G–C-G/C-G, expected psoriasis risk 2.83%).