Genes–Environment Interactions in Obesity- and Diabetes-Associated Pancreatic Cancer: A GWAS Data Analysis

Pancreatic cancer is the fourth leading cause of cancer-related deaths, accounting for more than 37,600 deaths each year in the United States (1). Epidemiologic studies have identified cigarette smoking as the major modifiable risk factor for this disease. Obesity and long-term history of diabetes mellitus may also affect risk and are also modifiable (2, 3). Genetic factors are known to play a role in pancreatic cancer development. Although genome-wide association studies (GWAS) have identified a few loci and chromosome regions that are significantly associated with the risk of pancreatic cancer (4, 5), these findings explain only a portion of the heritability of this disease. Because of the limitations of single marker analysis on GWAS data, there have been increasing efforts recently on GWAS pathway analysis, which uses prior biologic knowledge of gene function and aims at combining moderate signals of single-nucleotide polymorphism (SNP) and obtaining biologically interpretable findings (6, 7). Despite its great promise in providing insights into disease mechanisms, current GWAS pathway analysis has some caveats including being limited to enrichment of marginal genetic effects in biologic pathways without considering possible interactions between pathways and environmental factors (8). On the other hand, environmental factors are likely to interact with multiple genes through various biologic pathways, contributing to the susceptibility of complex human diseases. Although current GWAS top hits account for only limited heritability, gene–environment interactions may account for some of the missing heritability of pancreatic cancer (9).

We have previously conducted pathway analyses of the GWAS data in pancreatic cancer. Several novel pathways significantly associated with risk were identified (10, 11). For example, the pancreas development pathway [Mature Onset Diabetes of the Young (MODY) pathway] was identified as a top pathway in pancreatic cancer etiology. One possible mechanism related to this pathway is through obesity and diabetes (12–14). However, our previous study did not detect a significant interaction of obesity or diabetes with the NR5A2 gene, a GWAS top hit and a major contributing gene to the pancreas development pathway (MODY pathway), in modifying the risk of pancreatic cancer. On the other hand, we detected a strong interaction of the fat mass and obesity-associated (FTO) gene with overweight for pancreatic cancer risk, even though the gene did not show a marginal effect (15). A recent post-GWAS analysis of diabetes-related genes also failed to find strong evidence that common variants underlying type 2 diabetes or related phenotypes interact with diabetes in modifying the risk of pancreatic cancer (16). These observations suggest that there are unidentified genes contributing to obesity- and diabetes-related pancreatic cancer. Taking advantage of the existing GWAS data and exposure variables from the Pancreatic Cancer Case Control Consortium (PanC4; ref. 17), we conducted a comprehensive gene–environment (G × E) interaction analysis of genetic factors that may modify the associations of obesity and diabetes with pancreatic cancer.

The study population was drawn from seven studies participating in the previously conducted GWAS of the Pancreatic Cancer Cohort Consortium (PanScan) and the PanC4 Consortium (4, 5), including six case–control studies conducted at the MD Anderson Cancer Center (Houston, TX), Mayo Clinic (Rochester, MN), Yale University (New Haven, CT), University of California (San Francisco, CA), Memorial Sloan-Kettering Cancer Center (New York, NY), and University of Toronto (Toronto, ON, Canada), and one nested case–control study from the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort. Cases were defined as patients diagnosed with primary adenocarcinoma of the pancreas; in each study center, controls were matched to cases according to birth year, sex, and self-reported race/ethnicity and were free of pancreatic cancer at the time of recruitment.

GWAS scanning was performed at the National Cancer Institute Core Genotyping Facility using the Illumina HumanHap550 and HumanHap550-Duo SNP arrays and Illumina Human 610-Quad arrays (4, 5). Genotype data covering 562,000 and 621,000 SNPs from 4,195 study subjects (2,163 cases and 2,232 controls) were downloaded from the Database of Genotypes and Phenotypes website (http://www.ncbi.nlm.nih.gov/gap) with the approval of MD Anderson Institutional Review Board (IRB). We first conducted data cleaning and quality control by removing SNPs with minor allele frequency (MAF) less than 5%, deviating from the Hardy–Weinberg equilibrium (P < 0.001) or not genotyped in both SNP array platforms, resulting in a final dataset of 468,114 SNPs. According to the International HapMap Project genotype data (phase III release #3, NCBI build 36, dbSNP b126, 2010-05-28, MAF > 5%) for CEU (European), JPT (Japanese)/CHB (Chinese), and YRI (Yoruba) (18), we used 10,155 high-quality markers (r² < 0.004) in STRUCTURE (19) and identified a total of 4,137 individuals (2,028 cases and 2,109 controls) with 0.75–1.00 similarity to CEU as the study subjects in the current analysis. Then, we derived the top five principal components for population substructure in the Caucasian subjects using the EIGENSTRAT (20).

The pathways and genes used in the current analysis are defined as previously described (11). A total of 214 human biologic pathways were downloaded from the KEGG (Kyoto Encyclopedia of Genes and Genomes) website (21). Of these, 197 pathways with 10 to 500 genes each were considered in the analyses. Gene lists were downloaded from the human genome database version18 (hg18) using the UCSC Table Browser data retrieval tool (22). SNPs within 20 kb upstream or downstream of genes were included. In total, 5,127 genes annotated in the 197 pathways, covering 82,881 SNPs, were tested for interactions with risk factors.

Exposure variables without personal identifiers were provided by each participating institution to MD Anderson under IRB approvals and MTA (Material Transfer Agreement). Exposure variables included age, sex, race/ethnicity, adulthood body mass index (BMI, weight/height²), history of cigarette smoking, history of diabetes, and family history of cancer. All data were coded according to a uniform data dictionary. Missing pack-years of smoking were imputed based on study–age–sex means in 228 smokers. After merging and cleaning the data, we defined the variables in this G × E analysis as follows: obesity (BMI ≤30 kg/m² vs. >30 kg/m²) and diabetes (yes vs. no). Other exposure variables that are adjusted in the multivariable models included age (continuous), sex, and smoking (0, <20, and ≥20 pack-years). Because of a large number of missing values for family history of cancer, this variable was not considered in the model.

We used principal component analysis (PCA) to reduce the dimension of SNPs within a gene or pathway before the interaction analysis (11, 23). Briefly, PCA was performed to decompose the genetic variation in a gene into orthogonal components called eigenSNPs; the eigenvalues were calculated to identify principal components (eigenSNPs) that explained at least 85% of the observed genetic variation within a gene. Before pathway-by-environment interaction analyses, we used the global likelihood-ratio test (LRT) to determine if genes represented by the eigenSNPs were marginally associated with disease status, and only those genes with nominal P values ≤ 0.10 were retained in the pathway (PCA-LRT) screening. The eigenSNPs of genes with marginal effects were included in the pathway–environment interaction analyses, along the same line as the two-step approach for SNP × SNP/SNP × environment interaction analysis proposed in refs. (24) and (25). The gene/pathway and environment interaction was analyzed using LRT in nested logistic regression models. The full model included age (continuous), sex, study sites (categorical), five principal components (quantitative) capturing population structure, smoking (pack-years), genetic factors (eigenSNPs), the risk factor of interest, and the interaction terms (the products of risk factor of interest and eigenSNPs). The interaction terms were removed from the reduced model.

For G × E analysis at the pathway level, in total 172 pathways having at least two genes with marginal effect were identified through the PCA-LRT screening (Supplementary Table S1). Genes with a P_{G × E} value < 0.05 in the interaction analysis were considered as the major contributing genes to the pathway. We also performed a simulation study to demonstrate that the LRT method can effectively control the Type I error for the interaction analysis (Supplementary Text).

For G × E analysis at the gene level, a total of 5,127 genes were tested using LRT and logistic regression. SNPs with a P_{G × E} value of <0.05 were defined as the contributing SNPs to a gene. After screening all 5,127 genes, we also took the “gene to pathway” approach by conducting Ingenuity Pathway Analysis (IPA) on the genes with a P_{G × E} value of ≤0.05 to identify overrepresented canonical pathways (Ingenuity Systems; www.ingenuity.com).

For G × E analysis at the SNP level, we analyzed the interactions of 82,881 SNPs with obesity or diabetes on the risk of pancreatic cancer using LRT in nested logistic regression model. SNPs were coded as 0, 1, or 2 for counts of the minor allele.

To control the problem of false-positive findings associated with multiple testing, we applied the Bonferroni correction for G × E interaction analysis at the pathway level. P values of <1.45 × 10⁻⁴ [0.05/(2 × 172)] were considered statistically significant at the pathway level. Because of the large number of genes/SNPs, we used the q value method with false discovery rate (FDR) at 0.10 as the significance threshold for G × E analysis at the gene/SNP level (26).

The characteristics and exposure variables of the study populations are described in Table 1. There are no significant case–control differences in the distributions of age, race, and sex (all P > 0.10). More than 99% of participants were self-reported non-Hispanic Whites. The case–control association did not suggest any population stratification (adjusted λ = 0.999; ref. 27). The prevalence of obesity (BMI > 30 kg/m²) was 21.1% versus 16.6%, and diabetes was 20.3% versus 9.5% in cases and controls, respectively. Obesity, diabetes, and smoking (≥20 pack-years) were significantly associated with increased risk of pancreatic cancer, with adjusted odds ratios (AOR) and 95% confidence intervals (95% CI) 1.22 (1.02–1.47), 2.35 (1.94–2.84), and 1.60 (1.38–1.86), respectively.

Among the 172 pathways tested, 40 pathways showed nominal interactions (P < 0.05) with obesity (Supplementary Table S2) and 18 with diabetes (Supplementary Table S3). One pathway (contributing genes) remained statistically significant and one nearly so after Bonferroni correction: the chemokine signaling pathway (GNGT2, RELA, and TIAM1) interacting with obesity (P = 3.29 × 10⁻⁶), the calcium signaling pathway with diabetes (GNAS; P = 1.57 × 10⁻⁴; Table 2). In addition, four additional top pathways, i.e., interaction of obesity with pathways in cancer, cytokine–cytokine receptor interaction pathway, as well as interaction of diabetes with mitogen-activated protein kinase (MAPK) signaling pathway, and pathways in cancer, are also shown in Table 2. We checked the sensitivity of the statistical method (LRT) to pathway size and found that the significance levels were unrelated to pathway size (data not shown). Furthermore, as a complementary approach to the above PCA-LRT analysis, we performed IPA analysis on nominally significant genes (P < 0.05) in G × E interactions at the gene level (next section). Several pathways that were highly significant at P < 10⁻⁸ were identified: The role of RIG1-like receptors in antiviral innate immunity canonical pathway and the role of phosphoinositide 3-kinase (PI3K)/AKT signaling in the pathogenesis of influenza were most overrepresented in obesity-interacting genes, whereas molecular mechanisms of cancer pathway were most overrepresented in diabetes-interacting genes (Table 3).

Among the 5,127 genes tested, 335 and 263 genes showed nominal interactions with obesity and diabetes, respectively (P < 0.05; Supplementary Tables S4 and S5). After adjusting for multiple comparisons, none of these interactions remained statistically significant. Twelve genes with the smallest P values (<0.001) are listed in Table 4, including seven genes interacting with obesity and five genes interacting with diabetes. To overcome the reverse causality problem, we analyzed gene interactions with diabetes after excluding subjects with new onset diabetes (≤2 years), but no significant change in the results was observed (data not shown).

A total of 3,859 and 3,551 SNPs exhibited nominal interactions with obesity and diabetes, respectively (P < 0.05), which were identified as the contributing SNPs to the genes with nominal interactions. Of these SNPs, 810 interactions with obesity and 758 interactions with diabetes at the level of P < 0.01 are presented in Supplementary Tables S6 and S7, respectively. There are seven interactions with obesity and six with diabetes that had a P value of <10⁻⁵ (Table 5). One SNP (rs10818684) of the PTGS1 (aka COX1) gene actually had a q value of 0.06, which was significant at an FDR of less than 10%. Notably, all of the top 13 SNPs displayed a differential effect on risk of pancreatic cancer between exposed (obese or diabetic) and unexposed (nonobese or nondiabetic) groups and none of them had marginal effect on risk of pancreatic cancer when the analysis was conducted in the combined dataset of exposed and unexposed individuals (Table 5).

There are a total of 120 SNPs genotyped for GNGT2 (8 SNPs), RELA (5 SNPs), and TIAM1 (107 SNPs), the three major contributing genes in the chemokine signaling pathway. We first conducted LRT in the logistic regression model for each SNP and found that 17 SNPs were significant at the 0.05 level. We further analyzed the interaction pattern of the 17 SNPs using standard interaction analysis method and identified 8 synergisms, 4 antagonisms, and 5 undefined (Supplementary Table S8).

In light of the strong linkage between obesity and diabetes, we investigated the overlap between genes/SNPs interacting with obesity and diabetes on the risk of pancreatic cancer, as well as the overlap between genes/SNPs marginally associated with these two risk factors (Supplementary Table S9). At the significance level of 0.001, there were no overlapping genes/SNPs between obesity and diabetes; at a less stringent significance level of 0.01, there was a moderate 1% to 3% overlapping genes/SNPs. As a result, our analyses here did not support strong overlap between genetic factors interacting with obesity and diabetes on the risk of pancreatic cancer.

In this large G × E analysis in pancreatic cancer, two approaches, from pathway to gene and from gene to pathway, suggest consistent findings that highlight the interactions of inflammatory response pathways with obesity and insulin resistance or cancer-related pathways with diabetes in modifying the risk of pancreatic cancer. We also observed that SNPs without marginal effects had strong differential effects on cancer risk between exposed and unexposed individuals. These preliminary findings underscore the potential value and the challenges of comprehensive G × E analysis in revealing molecular mechanisms that may underlie complex disease.

Using LRT-logistic regression analysis, the current study identified a statistically significant interaction of the chemokine signaling pathway with obesity in modifying the risk of pancreatic cancer. This association was supported by findings from another statistical approach, i.e., IPA analysis. The major contributing genes to the chemokine signaling pathway and the top two canonical pathways identified by IPA (Table 3), e.g., RELA, GNGT2, NFKB1, NFKB2, and IFNA or interleukin genes, suggest a central role of the NF-κB signaling–mediated inflammatory and immune responses in obesity-related pancreatic cancer (2). GNGT2 [guanine nucleotide binding protein (G protein), gamma transducing activity polypeptide 2] has been shown to mediate β-arrestin 1-induced Akt phosphorylation and NF-κB activation (28). The RELA gene encodes the p65 protein, which binds to NF-κB1, forming the most abundant form of NF-κB (29). NF-κB is activated by many proinflammatory cytokines and is constitutively activated in pancreatic cancer. Increased NF-κB activity inhibits apoptosis and promotes growth, tumorigenesis, angiogenesis, invasion, and metastasis (30). Observations from this study suggest that genetic variations conferring proinflammatory responses may act in concert with the chronic inflammatory state of obesity in increasing the risk of pancreatic cancer.

The current study found a nearly significant interaction of diabetes with calcium signaling pathway in modifying the risk of pancreatic cancer. In addition, nominal interactions of the MAPK signaling pathway and pathway in cancer with diabetes were also observed. IPA analysis found that genes in molecular mechanisms of cancer were most overrepresented among diabetes-interacting genes. The physiologic and biochemical roles of calcium signaling range widely, and how this pathway interacts with diabetes in modifying the risk of pancreatic cancer remains unclear. The single significant gene contributing to this pathway was GNAS (GNAS complex locus, aka adenylate cyclase-stimulating G α protein), which encodes the G protein α unit that couples receptors to the generation of intracellular cyclic AMP (cAMP). GNAS mutations have been reported in multiple types of endocrine neoplasms (31). High frequency of GNAS mutations was also found in intraductal papillary mucinous neoplasm of the pancreas, but not in pancreatic ductal cancer (32). However, studies in mice indicate that mutations of this gene lead to obesity, glucose intolerance, and insulin resistance (33). It is possible that GNAS variants contributed to diabetes-associated pancreatic cancer via the mechanism of altered cAMP signaling transduction or enhanced insulin resistance. In addition to the calcium signaling pathway, the most notable interaction of diabetes was with genes or pathways involved in cancer, which was consistently identified by both PCA-LRT and IPA approaches. The major contributing genes to these interactions included the oncogenic FOS (FBJ murine osteosarcoma viral oncogene homolog), the tumor-promoting gene EPAS1 (endothelial PAS domain protein 1, aka hypoxia-inducible factor 2 α), a tumor suppressor DAPK3 (death-associated protein kinase 3), and MAP2K7 (aka MEK7, JNKK2, and SKK4; ref. 34). MAP2K7 mediates the cellular responses to proinflammatory cytokines and environmental stresses with a strong preference for activation of the c-jun-NH2-kinase (JNK) pathway (35). JNK signaling plays a central role in obesity and insulin resistance (36) as well as in regulating apoptosis (37). FOS proteins can dimerize with c-jun, thereby forming the transcription factor complex AP-1 that regulates cell proliferation, differentiation, and transformation as well as apoptosis (38). Overall, the results of our study highlight pathways and genes that have been implicated in cancer development, especially those associated with insulin resistance and apoptosis, in diabetes-related pancreatic cancer.

Several studies have previously suggested the possibility to increase statistical power of G × E analyses by focusing on genes with marginal effects only (24). Our findings that SNPs with the smallest P value for interaction were those without any marginal effect suggest that G × E analysis limited to such genes/SNPs may miss genetic variants that have a true impact on disease risk among exposed individuals only, consistent with a recently reported SNP by alcohol intake interaction influencing the risk of esophageal squamous-cell carcinoma (39, 40). Thus, a comprehensive G × E analysis of GWAS data using multiple analytic methods with complementary strengths as undertaken here and suggested by previous research (41) may be a necessary and useful approach to unveiling missing heritability of a complex disease such as pancreatic cancer (42). It would be interesting to develop hybrid strategies, in line with that of ref. (25), for pathway by environment interaction analysis in the future.

Our study has several strengths and limitations. This is the largest G × E analysis in pancreatic cancer with the most comprehensive analysis of all biologic pathways identified from KEGG using an agnostic approach. We used a PCA approach to reduce the dimensionality of the GWAS data and increased the probability of finding useful information. The analysis was based on high-quality genotype and exposure data with extensive quality control measures. We also applied stringent criteria to control false-positive reporting. However, our sample size is still relatively small for a full G × E GWAS analysis. Our findings cannot be replicated due to the lack of available datasets. Thus, the possibility that some associations are spurious findings cannot be excluded, which limits the generalization of the results. Nevertheless, the pathways and genes found interacting with obesity and diabetes are highly relevant to pancreatic cancer and are supported by other experimental evidence. Our results underscore the interactions of inflammation-related genes with obesity and insulin resistance or cancer-related genes with diabetes in modifying the risk of pancreatic cancer. G × E analysis offers an opportunity to identify genetic factors linking obesity and diabetes to pancreatic cancer. Such information would provide scientific rationales for the development of novel strategies in personalized prevention of pancreatic cancer.

No potential conflicts of interest were disclosed.

Conception and design: H. Tang, P. Wei, H.A Risch, H.B. Bueno-de-Mesquita, E.A Holly, M. Jenab, A. Tjonneland, P.H.M. Peeters, K.-T. Khaw, D. Li

Development of methodology: H. Tang, P. Wei, E.A Holly, P.H.M. Peeters

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E.J. Duell, H.A Risch, S.H. Olson, H.B. Bueno-de-Mesquita, S. Gallinger, E.A Holly, G.M. Petersen, R.R. McWilliams, A. Tjonneland, M.-C. Boutron-Ruault, R. Kaaks, D. Trichopoulos, M. Sund, P.H.M. Peeters, K.-T. Khaw, D. Li

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): H. Tang, P. Wei, H.A Risch, M. Jenab, P.H.M. Peeters, C.I. Amos, D. Li

Writing, review, and/or revision of the manuscript: H. Tang, P. Wei, E.J. Duell, H.A Risch, S.H. Olson, H.B. Bueno-de-Mesquita, E.A Holly, G.M. Petersen, P.M. Bracci, M. Jenab, E. Riboli, A. Tjonneland, M.-C. Boutron-Ruault, R. Kaaks, D. Trichopoulos, M. Sund, P.H.M. Peeters, K.-T. Khaw, C.I. Amos, D. Li

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H. Tang, S. Gallinger, P.H.M. Peeters, K.-T. Khaw, D. Li

Study supervision: H. Tang, P. Wei, H.A Risch, E.A Holly, D. Li

This study was supported from the NIH through a supplemental grant R01 CA98380-05 (to D. Li), an MD Anderson Cancer Center Support Grant CA016672, NIH grants R01 CA169122 and R01 HL106034 (to P. Wei), and the Sheikh Ahmed Center for Pancreatic Cancer Research Funds (to D. Li). S. Gallinger was supported by NIH grant R01 CA09075. P.M. Bracci was supported by NIH grants R01 CA59706, R01 CA109767, R01 CA109767S1, R01 CA154823, and R03 CA89726. K.-T. Khaw was supported by Medical Research Council UK and Cancer Research UK.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.