Study Participants

We obtained data from the 20 studies in the PanScan collaboration who agreed to participate in this pathway analysis. PanScan phases I-III have been previously described [14–16]. Our primary analysis included genotype data from 10 cohort studies and 10 case control studies in the PanScan collaboration. Participating cohorts included the Agricultural Health Study, Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study (ATBC), Give us a Clue to Cancer and Heart Disease Study (CLUE II), Cancer Prevention Study II (CPS-II), Melbourne Collaborative Cohort Study (MCCS), Multiethnic Cohort (MEC), New York University Women’s Health Study (NYU-WHS), Prostate Lung Colorectal and Ovarian Cancer Screening Trial (PLCO), Selenium and Vitamin E Cancer Prevention Trial (SELECT) and the VITamins and Lifestyle cohort (VITAL). The included case-control studies were the Mayo Clinic Molecular Epidemiology of Pancreatic Cancer Study, University of California San Francisco, Yale University, MD Anderson Cancer Center, University of Toronto, Johns Hopkins University, Memorial Sloan-Kettering Cancer Center, PACIFIC Study of Group Health and Northern California Kaiser Permanente, Spanish Pancreatic Cancer Study (PANKRAS II) [17], and PANcreatic Disease ReseArch (PANDoRA) (Heidelberg, Germany) [18]. All cases were diagnosed with primary pancreatic adenocarcinoma (ICD-O-3 code C250-C259 or C25.0-C25.3, C25.7-C25.9). In short, PanScans-I and II used a nested case-control study design for the cohort studies. Cohort cases were confirmed through cancer registries, death certificates or review of medical records by medical personnel. Cohort controls for PanScan-I were incidence density sampled with a 1:1 ratio and were alive and cancer free at the time of diagnosis of the matched case. In all case-control studies, matching criteria included calendar year of birth within five years, gender, race and ethnicity, while some cohorts also matched on age at baseline or blood draw, smoking, date/time of blood draw, fasting status at time of blood draw, and length of follow-up. All data was de-identified before genotyping and before samples were sent to NCI. Genotyping was performed at the National Cancer Institute’s (NCI’s) Cancer Genomics Research Laboratory (formerly known as the Core Genotyping Facility) using the Illumina HumanHap550 array for PanScan-I, the Illumina Human 610-Quad array for PanScan II, and the Illumina Human 770-Quad chip for PanScan III. In PanScan III controls were previously genotyped using second-generation Illumina SNP microarrays (e.g. OmniExpress, Omni 1M or Omni 2.5M) and drawn from PanScan III prospective cohorts and a Spanish case-control study and thus were not matched to the cases [19]. SNPs reported here were limited to those with minor allele frequencies (MAF) ≥5%. In total, we used data from 3,583 pancreatic cancer cases and 7,053 controls (1,108 cases and 4,353 controls from the cohorts and 2,475 cases and 2,700 controls from case-control studies) of European descent. The age and sex distribution of cases and controls is described in S1 Table. Written consent was obtained from all study participants. Each participating study was reviewed and approved by their local IRB for appropriateness in PanScan participation (S2 Table) [15].

Vitamin D measurement

A subset of the cohort participants with GWAS data also had 25(OH)D measured in serum as part of the Vitamin D Pooling Project (713 cases and 878 controls) [20]. These subjects were from the following cohorts: ATBC, CLUE-II, CPS-II, NYU-WHS, and PLCO. Methods for assaying 25(OH)D have been previously described [10]. In short, Heartland Assays, Inc. (Ames, Iowa) performed assays for 25(OH)D for samples from CLUE-II, CPS-II, NYU-WHS, PLCO, and a subset of ATBC samples using the DiaSorin LIAISON 25 OH Vitamin D TOTAL Assay (Diasorin, Inc., Stillwater, Minnesota). The remaining ATBC samples were assayed previously using a similar method in the laboratory of Dr. R. Vieth [21]. The methods and coefficients of variation percentages for the blinded quality control samples of the Heartland 25(OH)D measures have been previously reported (10). Using a nested components of variance analysis with logarithmically transformed quality control measures across batches, the overall intra- and interbatch coefficients of variation were 16.5% and 4.7% for the previously assayed concentrations in the ATBC and PLCO studies, respectively.

Vitamin D-related functions of included genes

DHCR7 (DHC-7 reductase) converts pro-vitamin D₃ (7-dehydrocholesterol) in the skin to cholesterol. Alternatively, pre-vitamin D is formed from 7-dehydrocholestrol following dermal UVB exposure. Vitamin D is also derived from diet or supplements in the form of cholecalciferol (D₃) or ergocalciferol (D₂). The vitamin D binding protein (DBP, also known as GC), transports provitamin D to the liver, as well as other vitamin D compounds to target tissues. In the liver 25-hydroxylases CYP2R1 and CYP27A1 convert vitamins D₂ and D₃ from diet and sun exposure to 25(OH)D (calcidiol), the major circulating vitamin D metabolite. Calcidiol is then converted to the active form 1,25(OH)₂D₃ (calcitriol) by 1α-hydroxylase (CYP27B1) in the kidney and other organs. 1α-hydroxylase (CYP27B1) is localized on the inner mitochondrial membrane where it produces the active form of vitamin D which binds the vitamin D receptor (VDR) with substantially higher affinity than 25(OH)D. The vitamin D binding protein, GC, transports vitamin D metabolites to target organs, where calcitriol binds to VDR and forms a heterodimer with RXRA (retinoid x-receptor alpha). This heterodimer attaches to vitamin D response elements on various target genes, some of which are thought to have anti-carcinogenic properties [1]. Vitamin D is catabolized by 24-hydroxylase (CYP24A1) to inactive forms [22].

Three additional genes, cubulin (CUBN), megalin (LRP2) and Calcium Sensing Receptor (CASR) were added to our analysis to be consistent with the previous study on vitamin D-related genes and pancreatic cancer [13]. Both cubulin and megalin are plasma membrane receptors that in combination mediate an endocytic update of GC-bound vitamin D. The CASR membrane protein binds calcium in the extracelluar matrix and plays an important role in calcium homeostasis.

In total, genotype data from 213 tag SNPs with MAF > 0.05 in 11 genes involved in the synthesis (DHCR7, CYP27A1, CYP2R1), transport (GC, CASR), metabolism (CYP27B1, LRP2, CUBN), signal transduction (VDR, RXRA) or catabolism (CYP24A1) of endogenous vitamin D were used in our analysis [1]. These SNPs are located within a span of 20kb 5′ upstream and 10kb 3′ downstream of the gene coding region as defined by the National Center for Biotechnology Information’s human genome build 36.3. SNPs and associated genes are listed in S3 Table.

Statistical analysis

We used unconditional logistic regression to test the association between individual SNPs and pancreatic cancer risk, adjusting for age (≤50, 51–60, 61–70, 71–80, ≥81 years), sex, study, and 5 eigenvectors capturing ethnic ancestry. We performed the pathway and gene analyses using the R package AdaJoint and the adaptive rank truncated product (ARTP) statistic [23]. In each analysis 1,000,000 permutations were conducted. This statistic accounts for gene or pathway size and linkage disequilibrium and summarizes joint association signals within a gene or pathway. Analyses were restricted to Caucasians. We also tested for heterogeneity between the three phases (PanScan I, PanScan II and PanScan III) using the R package for fixed effects meta-analysis. We adjusted for multiple comparisons in the p-heterogeneity analysis considering a p-value<0.0002 as significant. Analyses stratified by the three PanScan phases were performed to explore possible differences in associations by phase.

To test possible effect modification by vitamin D status, we performed stratified analyses in the subset of cohort studies with measured 25(OH)D; we created a dichotomous variable for circulating vitamin D (≤50 nmol/L or >50 nmol/L), as above this threshold is considered to be adequate for bone and overall health in national recommendations [24,25] and was close to the median for controls in our study population (control median 25(OH)D = 51.5 nmol/L). In this subset analysis we additionally adjusted for smoking (never, former, current), body mass index (BMI), and season of blood draw (fall, winter, spring, summer). To test for multiplicative interaction, we created an interaction term between circulating vitamin D as a dichotomous variable and individual SNPs. Using a Bonferroni correction for multiple comparisons, genes with a p-value<0.006 and SNPs with a p-value<0.0002 were considered statistically significant. We additionally evaluated whether SNPs associated with vitamin D levels in published GWAS (rs2282679, rs12785878, rs10741657, rs6013897) on circulating vitamin D [26,27], or representative tag SNPs, were associated with vitamin D in a subset of our study sample.