HLA-DQA1-HLA-DRB1 variants confer susceptibility to pancreatitis induced by thiopurine immunosuppressants

Abstract

Pancreatitis occurs in approximately 4% of patients treated with the thiopurines azathioprine or mercaptopurine. Its development is unpredictable and almost always leads to drug withdrawal. We identified inflammatory bowel disease patients who had developed pancreatitis within three months of starting these drugs from 168 sites around the world. After detailed case adjudication, we performed a genome wide association study on 172 cases, and 2035 inflammatory bowel disease controls. Strong evidence of association was identified within the class II HLA region with the most significant association at rs2647087 (Odds ratio 2.59, 95% CI 2.07 – 3.26 P=2x10^-16). These findings were replicated in an independent set of 78 cases and 472 IBD controls matched for drug exposure. Fine mapping of the HLA region identified association with the HLA-DQA1*02:01-HLA-DRB1*07:01 haplotype. Patients heterozygous at rs2647087 have a 9% risk of developing pancreatitis after administration of a thiopurine, while homozygous patients have a 17% risk.

Almost one million prescriptions for the thiopurines azathioprine and mercaptopurine were dispensed in England alone in 2011 ¹. Mercaptopurine (MP) and its prodrug azathioprine (AZA) are the most commonly prescribed immunosuppressive agents used to maintain corticosteroid-free remission and to prevent postoperative recurrence in patients with inflammatory bowel disease (IBD) ^2–4. Outside of gastroenterology they are widely used as anti-rejection agents after solid organ transplantation and as a steroid sparing agents for conditions such as rheumatoid arthritis. Despite their widespread use, it has been estimated that 17% of patients taking these medications develop side effects that necessitate drug withdrawal ⁵. Acute pancreatitis after thiopurine therapy is a well recognised, idiosyncratic, dose-independent adverse drug reaction with an incidence of approximately 4 to 7% in patients with IBD ^5,6. The pathogenesis of thiopurine-induced pancreatitis is unknown and the clinical picture is poorly described. Development of pancreatitis, which can be life threatening, precludes the patient from continuing on conventional thiopurine-based therapy and necessitates the use of other agents, which may be less effective or more costly.

Recent studies have confirmed that some rare adverse responses to drug therapy are associated with clinically useful genetic variants that can be identified using a small number of rigorously characterised cases by genome wide association study methodologies. For example the HLA class I allele HLA-B*57:01 has been shown to be a major determinant of flucloxacillin-induced cholestatic hepatitis with an odds ratio of 80, using a cohort of only 51 patients ⁷. This same HLA allele was earlier found to be associated with abacavir hypersensitivity. In Europe, the USA and Australia HLA-B*57:01 testing is now mandatory before prescribing abacavir ⁸. We aimed to characterise the clinical features of thiopurine-induced pancreatitis and identify genetic markers that might predict the development of this serious adverse drug reaction.

In total we recruited 441 patients to the study in two recruitment rounds, the initial GWAS round and a case-control replication cohort. Eight individuals submitted to the study failed to meet the eligibility criteria and were removed leaving 433 patients in the combined cohort. Each case was reviewed by an expert panel of gastroenterologists and a causal link between AZA/MP and pancreatitis development assessed using a modified version of the Liverpool Causality Assessment Tool ⁹. Definite cases required the development of recurrent pancreatitis upon thiopurine re-challenge. Cases classified as probable demonstrated a clear temporal relationship with thiopurine administration, with no other identifiable risk factors for pancreatitis, including the concomitant use of other drugs recognised as causing pancreatitis. Only definite and probable cases were taken forward for analyses. Details of case recruitment, the adjudication process and results are available in the Methods section and Supplementary Information. The number of cases in each adjudication group is displayed in Supplementary Table 1. A summary of the clinical characteristics of the recruited patients is shown in Supplementary Table 2.

We first conducted a genome wide association study with 217 of the 248 patients with thiopurine-induced pancreatitis and 2035 previously genotyped control Crohn’s disease and ulcerative colitis cases ^10,11, matched for disease (Crohn’s disease or Ulcerative Colitis) but unselected for thiopurine exposure or pancreatitis development. Of the initial 217 patients, we restricted the analyses to 177 patients of European descent. After genotyping quality control procedures (see Methods) 172 cases remained. In order to account for the different genotyping platforms used within the control cohort and the case cohort, we first performed an imputation analysis and used genotype dosages from all individuals (0% missingness) for the analysis. The case cohort and the Crohn’s disease and ulcerative colitis control cohorts were imputed separately to the 1000 Genomes European phase1 version 3 (20101123) panel using minimac ^12,13. We restricted our analysis to the 2,819,700 SNPs that had an imputation R²>0.95 across all three cohorts. The genomic inflation factor was 1.03.

We identified an association with pancreatitis development at rs2647047 (Odds Ratio 2.26, P=1.9x10^-12) (Supplementary Table 3). As the association appeared to be within the Class II HLA region, we performed dedicated HLA imputation using SNP2HLA into the T1DGC reference panel of 5,224 individuals that have had classical HLA alleles typed as well as SNPs and indels by immunochip ¹⁴. This enabled us to refine the identified association with thiopurine-induced pancreatitis to the SNPs rs2647087 and rs7745656 within the Class II HLA region (Odds Ratio 2.59, 95% CI 2.07 – 3.26, P=2x10^-16) (Figure 1). This association was robust to principal component correction (Odds ratio 2.62, p=2x10^-15 after adjustment for 20 PCs). The HLA alleles HLA-DQA1*02:01 and HLA-DRB1*07:01 were also significantly associated with pancreatitis development and were partially tagged by rs2647087 (R² 0.49). Significant HLA allele associations are displayed in Table 1, while complete results are displayed in Supplementary Table 4. In order to validate the HLA imputation, we performed classical HLA SSP based PCR genotyping in 160 case samples. The resulting correlation (R² = 0.93 for all HLA-DRB1 alleles) was in line with published observations ¹⁴. There was a 99% concordance between the imputed and directly genotyped HLA-DRB1*07:01 alleles. Full results for imputed and SSP PCR based genotyping are shown in Supplementary Table 5. Including rs2647087 as a covariate removed all the associations in the HLA regions indicating that there is only one HLA association signal (Supplementary Figure 1). rs2647087 remained borderline significant after conditioning the association on HLA-DRB1*07:01 (P=2.0x10^-5). This disparity, together with the low R² between rs2647087 and HLA-DRB1*07:01 indicates that the causal variant may be an unidentified rare allele that is being tagged by these two alleles. The well described association of HLA-B*57:01 with Abacavir hypersensitivity requires only one copy of the allele for the development of hypersensitivity. In this study, however, there was no evidence of a dominant effect (P=0.37).

Figure 1. Genome-wide association results for all SNPs post 1000 Genomes and HLA imputation.

A Manhattan plot

B Q-Q Plot

Table 1. Top imputed HLA association signals (P<5x10^-8).

Risk described for presence of the listed allele

HLA Allele	Position (hg19)	GWAS control risk allele Freq.	GWAS case risk allele Freq.	GWAS OR (95% CI)	GWAS P
HLA-DRB1*07:01	32660042	0.16	0.33	2.55 (2.01, 3.23)	1x10-14
HLA-DQA1*02:01	32716284	0.16	0.33	2.54 (2.00, 3.22)	2x10-14

Including only the definite cases (n=37) increased the odds ratio marginally (Odds Ratio 3.18), but due to the much smaller sample size (n = 37 European cases) the association did not achieve genome-wide significance (HLA-DQA1*02:01-HLA-DRB1*07:01 P=5.88x10^-8). To investigate if smoking, a recognised risk factor for pancreatitis, was associated with rs2647087 genotype, the cases were subdivided by smoking status at the time of their pancreatitis. rs2647087 had an odds ratio of 2.99 (95% CI 2.18 – 4.11) in patients who were smoking at the time of their pancreatitis and 2.19 (95% CI 1.58 – 2.96) in patients who were not (p=0.17), providing no evidence for an interaction between smoking and genotype. rs2647087 remained the most significantly associated marker after including smoking as a covariate in the association test (Odds Ratio 2.50 P=3.4x10^-13). Thiopurine methyltransferase (TPMT) genotype (TPMT*3A, *3C, *2, *4 & *8) was not associated with pancreatitis development at any of the five most common loss of function loci (P=0.99) that predispose patients to bone marrow suppression with thiopurine treatment.

We next attempted to replicate our finding in an independent study of 78 cases and 472 IBD controls treated with thiopurines for at least 12 months without the development of pancreatitis. From the four variants chosen for replication, all of which are in complete linkage disequilibrium with rs2647087, the SNP rs6935723 had the highest genotype success rate (98.2%) and demonstrated robust replication of the association, Odds Ratio 2.21, P=4x10^-6 (Supplementary Tables 6 and 7). There was insufficient statistical power within the replication cohort to attempt to replicate the putative associations observed on chromosomes 3 and 10. The combined results of the GWAS and replication cohort are displayed in Table 2.

Table 2. Top SNP association signals from the GWAS analysis (P<5x10^-8).

All SNPs located on Chromosome 6.

Stage	SNP	Position (hg19)	Risk Allele	Other Allele	Control risk allele Freq.	GWAS case risk allele Freq.	OR (95% CI)	GWAS P
GWAS	rs7745656	32680970	T	G	0.27	0.49	2.59 (2.07,3.26)	2x10-16
Replication					0.26	0.46	2.22 (1.57, 3.14)	6x10-6
Combined							2.47 (2.05, 2.99)	2x10-21
GWAS	rs2647087	32681049	C	A	0.27	0.49	2.59 (2.06,3.26)	2x10-16
Replication					0.26	0.48	2.37 (1.68, 3.34)	1x10-6
Combined							2.52 (2.09, 3.05)	1x10-20
GWAS	rs6935723	32681669	C	T	0.27	0.49	2.59 (2.07,3.26)	2x10-16
Replication					0.26	0.46	2.21 (1.58, 3.10)	4x10-6
Combined							2.46 (2.04, 2.98)	1x10-21
GWAS	rs2647089	32681568	C	T	0.27	0.49	2.59 (2.07,3.25)	2x10-16
Replication					0.26	0.45	2.17 (1.54, 3.06)	9x10-6
Combined							2.45 (2.03, 2.97)	6x10-20

HLA-DRB1*07:01 has previously been described to be associated with ileal Crohn’s disease with a relative risk of 1.6 ¹⁵. To ensure the association was not due to a disparity in the anatomical location of Crohn’s disease between cases and controls (i.e. ileal vs. colonic disease), the association was performed with disease location as a covariate in the regression analysis (rs2647087, Odds Ratio 2.71, 95% CI 2.07 – 3.26, P=8.67x10^-13). In addition we performed the association test using only cases and controls with ulcerative colitis. This yielded consistent evidence of association despite the smaller number of cases (rs6935723, n=38, Odds Ratio 2.23, 95% CI 1.43 – 3.49, P=0.0003). A shared genetic association between ileal Crohn’s disease and thiopurine-induced pancreatitis might direct clinicians to other therapies including surgery for patients with this pattern of disease.

The prevalence of pancreatitis in patients treated with thiopurines is approximately 4-7%. Our results suggest that heterozygotes at rs2647087 will be approximately 2.5 times more likely to get pancreatitis and risk allele homozygotes approximately 5 times more likely to get pancreatitis than common allele homozygotes. In a clinical setting this means that for every 1000 patients tested for rs2647087, 77 risk allele homozygotes will be identified and these individuals will an approximate risk of 17% for the development of pancreatitis (based on a relative risk of 4.31, see Methods). Patients heterozygous at rs2647087 would have a 9% risk of developing pancreatitis. If AZA/MP are subsequently avoided in all homozygote risk allele individuals, this equates to an overall number needed to test of 76 patients to prevent 1 case of pancreatitis.

Pancreatitis was first described in renal transplant recipients treated with AZA in the early 1960s ¹⁶. It wasn’t until 1972, however, that AZA was implicated as the causal agent when a patient with Crohn’s disease was re-challenged with AZA and developed pancreatitis for a second time ¹⁷. Candidate gene case-control studies conducted prior to this study have suggested an association between a polymorphism in ITPA ¹⁸ and thiopurine-induced pancreatitis, however, we were unable to replicate this association (rs1127354, P=0.177).

To date six pharmacogenetic genome wide association studies for adverse drug reactions have demonstrated association with the class II HLA region ¹⁹. The mechanism for these associations has not yet been elucidated, however, the clinical features described here, with an average time from starting the drug to development of pancreatitis of 23.8 days (95% CI 21.2 – 26.4), would be consistent with a delayed immunological or T cell mediated reaction. The extended HLA DRB1*07-DQA1*02 haplotype identified here has previously been described as associated with drug-induced liver injury to both the tyrosine kinase inhibitor Lapatinib and the anticoagulant Ximelagatran although the mechanism of action is unknown ^20,21.

Recent work investigating the HLA-B*57:01 association with abacavir has described how the molecule interacts with the HLA binding pocket to alter the antigen presentation repertoire. This alteration can result in novel “self” antigen presentation or novel presentation of constitutive self-peptides ²². Class II HLA associations are less well established, but Ximelagatran has been demonstrated to competitively bind preferentially to the HLA DRB1*07:01 haplotype ²⁰. We have constructed a computational docking model (Supplementary Figure 2), which predicts that MP is able to bind the HLA DRB1*07:01 molecule; the effects of this on T cell activation, however, are at this stage unknown.

We have undertaken the first large scale clinical and genetic analyses of thiopurine-induced pancreatitis and identified an association with a common variant in the Class II HLA region. This is the first description of any genetic association with a drug-induced pancreatitis. Although in this study the indication for thiopurine use was IBD, these results are likely to be generalisable to many other patient populations where AZA/MP are used. Information on the risk of pancreatitis development may form part of a future panel of genetic tests that could aid clinicians and patients when deciding upon treatment options.

Online Methods

Patient Recruitment and selection

Patients were identified and recruited from 115 NHS hospital UK research sites and 53 international sites. Cases were identified through electronic searches of patient databases, pathology services, directly through gastroenterology clinics and by direct advertising to patients. The protocol was approved by the National Research Ethics Committee South West, Exeter (11/SW/0222) and by all local research and development offices.

Inclusion criteria for patient recruitment included all of the following:

• a)Onset of acute severe abdominal pain within three months of starting mercaptopurine or azathioprine treatment for ulcerative colitis or Crohn’s disease.
• b)Greater than or equal to a two-fold rise in amylase or lipase as defined by the research site’s local laboratory.
• c)Medical opinion implicating thiopurine therapy and subsequent drug withdrawal.

At the patient recruitment visit, a case report form was completed that detailed demographic, clinical and drug history after informed consent was obtained. Two 6ml EDTA blood samples were taken at the same visit for DNA extraction (BD Vacutainer, USA).

To assess patient eligibility, an adjudication panel assessed causality from case report forms using an adapted version of the validated Liverpool Adverse Drug Reaction Causality Assessment Tool displayed in Supplementary Figure 3 ⁹. All cases were reviewed independently by four medically trained experts and assigned a causality category based on the assessment tool. Confounding medications or conditions (listed below) classified the patient as a possible cause of thiopurine-induced pancreatitis:

• Gallstones

• Alcohol

• Hyperlipidaemia (in particular hypertriglyceridaemia)

• Concomitant administration of other medications from Badalov et al ²³

• Infection (e.g. Viruses-mumps, coxsackie, hepatitis B, CMV, varicella-zoster, HSV)

• Post-ERCP

• Ischaemia

• Trauma

The collective results from each panel member were collated and the panel discussed discrepant cases before a final adjudication decision was reached. A set of guidelines was drawn up through the course of these expert meetings to ensure consistent decision-making in borderline cases. These guidelines are displayed in the Supplementary Information and help illustrate the robust phenotype assessment undertaken for all cases. Co-administration of a thiopurine with any drug known to cause drug-induced pancreatitis (Class 1a or class 1b as defined by Badalov et al ²³) within the three months prior to the development of pancreatitis classified the patient as having had “possible” thiopurine-induced pancreatitis. Only patients classified as “definite” or “probable” cases of thiopurine-induced pancreatitis were taken forward to subsequent clinical and genetic analyses.

DNA Extraction and Genotyping

DNA was extracted from EDTA stabilised blood using the Qiagen Autopure LS with Puregene chemistry. 248 samples were genotyped on the Illumina Infinium HumanCoreExome beadchip (Illumina, USA), which contains 264,909 tagging SNP markers, and 244,593 exome focused markers by the Broad Institute (Boston, USA).

472 IBD patients who had been screened for thiopurine use without development of pancreatitis from the Royal Devon and Exeter NHS Trust were used as a control cohort for replication analysis. These 472 samples, together with the 87 thiopurine induced pancreatitis cases were genotyped at four SNPs (rs2647087, rs6935723, rs2647089 and rs7745656) using the KASP™ genotyping assay by LGC Limited, UK. Analysis was limited to the 78 samples that self identified as “White”. The genotyping success rates were > 97% for all SNPs.

Genetic and Statistical Analysis

The study was open for data collection from March 2012 to December 2013. Multivariate linear regression analysis was used to calculate risk variables. Logistic regression was used for categorical variables. Severe pancreatitis was defined as single organ failure or greater. Smoking status was calculated based on number of pack years (number of packs of 20 cigarettes per day multiplied by years smoked). For study entry a raised amylase or lipase was characterised as a greater than or equal to a two fold increase above the upper limit of normal for the research site’s local laboratory. Relative risks for genotypes were calculated from odds ratios based on the formula ²⁴:

$$\text{Relative}\text{risk}=\text{odds}\text{ratio}/(1-\text{baseline}\text{risk}+(\text{baseline}\text{risk}\times \text{odds}\text{ratio}))$$

All statistical analysis was undertaken in R (Version 3.0.2) and Stata (Version 13). Manhattan and QQ plots were created with the qqman R package.

Quality control of SNPs and samples before imputation

Genotyping was performed on 217 cases assigned as having “definite” or “probable” thiopurine-induced pancreatitis using the HumanCoreExome SNP Chip. Genotypes were called using Gencall ²⁵. We excluded SNPs with a HWE P<0.0001 and a genotype success rate <0.99. We excluded indels. Exclusion criteria for case samples were genotyping success rate <0.98 and a heterozygosity rate >4SD (5 samples removed). We used zCall to improve calling of low frequency variants ²⁶. After running zCall, SNPs were excluded if they had a HWE P<0.0001, MAF<0.01 or if they were duplicated. This left 254,457 autosomal SNPs for imputation. The control patients with Crohn’s disease and ulcerative colitis were obtained from the International IBD Genetics Consortium as part of the Wellcome Trust Case Control Consortium (WTCCC 1 for Crohn’s disease and WTCCC 2 for Ulcerative Colitis) ^27,28. There were 1748 CD control samples genotyped on the Affymetrix 500K SNP chip and 2361 UC samples genotyped on the Affymetrix 6 SNP chip available for this analysis. Preliminary QC had already been performed on the 1748 CD and 2361 UC samples ^27,28. From these two control cohorts, we excluded SNPs with a genotyping success rate <0.99, MAF<0.01 and a HWE P<0.0001 – this left 396,255 (Crohn’s disease) and 727,195 (Ulcerative Colitis) autosomal SNPs. To exclude ethnic outliers we performed principal components analysis using GCTA ²⁹. To generate the principal components we used a set of 79,974 SNPs that were imputed with R²>0.99 in the cases (see below) and directly genotyped in the two control cohorts (Supplementary Figure 4). 40 cases and 66 control samples were excluded for being >4SD for the first or second principal components. The case exclusions were consistent with self-reported non-European ancestry. We used KING ³⁰ to test for cryptic relatedness between samples. If a case and control pair of samples had a kinship coefficient >0.2 we excluded the control sample, otherwise we excluded one of the pair of samples at random if both were in the same cohort. Six control samples were excluded because of relatedness to a pancreatitis case patient. Nine Crohn’s disease control samples were excluded because they were diagnosed with pancreatitis and also included in the case cohort. After exclusions this left 172 “probable” and “definite” cases, 80% of which had Crohn’s disease. To match the ratio of Crohn’s disease to ulcerative colitis patients in the control group to that in the case cohort we used all the 1669 Crohn’s disease patents that passed QC and a random 366 samples from the ulcerative colitis cohort in subsequent association analyses.

Genome-wide and HLA Imputation and association analyses

We used minimac ¹³ to impute into the European phase1 version 3 (20101123) panel SNPs and indels reference panel. 76% of the 9,412,474 SNPs with MAF>1% frequency was imputed at RSQ>0.6 in the cases: 75% in the Crohn’s disease (CD) controls and 82% in the ulcerative colitis (UC) controls. Each of the three case and control cohorts used a different SNP genotyping chip. This has been reported to lead to spurious associations ³¹. Therefore, to avoid excessive false positive associations we focused subsequent association analyses on a very conservative subset of 2,819,700 SNPs that had an RSQ>0.95 in all three cohorts. For dedicated imputation of the HLA region we used SNP2HLA ¹⁴ and imputed into the T1DGC reference panel of 5,224 individuals that have had classical HLA alleles typed as well as SNPs and indels by the immunochip. 8321 of the 8961 variants in the T1DGC panel were captured with an INFO score > 0.8. mach2dat ³² was used to perform the association analyses for the genome-wide analysis, and PLINK ³³ was used to perform the association analyses for the HLA imputed analysis. A standard test for dominance deviation analysis was performed whereby we tested if the genotype coded 0/1/0 improved upon a regression model which included a term for the genotype coded 0/1/2.

In-silico peptide binding

Computational docking was performed to probe the binding of mercaptopurine to HLA-DRB1*07:01. As a crystal structure of HLA-DRB1*07:01 is not available, the HLA-DRB1*07:01 was homology modelled from the crystal structure of HLA-DRB1*04:01 (PDB code 4MCY ³⁴). For comparison, docking runs were also performed with HLA-DRB1*01:01 (PDB code 4AH2), HLA-DRB1*04:01 (PDB code 4MCY), HLA-DRB3*01:01 (PDB code 2Q6W), HLA-DP2 (PDB code 3LQZ) and HLA-DQ8 (PDB code 2NNA). AutoDock Tools 1.5.6 ³⁵ was used to assign hydrogens, Gasteiger charges and rotatable bonds to the compounds. Each docking run was performed in the absence of peptide within the binding cleft and utilised the AutoDock Vina software ³⁶ to search a docking grid that encompassed the entire peptide-binding cleft.

Data access

Phenotype and genotype data for cases is freely available upon request from the iSAEC Data Access Committee for users who comply with the Consortium’s Data Release and IP Policy. Data will be available within 12 months of genotype completion. Raw genotype data is freely available to researchers upon request. For further data access details please contact saec@c2b2.columbia.edu.

Genotype data for the WTCCC Ulcerative colitis and Crohn’s disease cases are available from the European Genome-Phenome Archive.