Genome-wide gene-diabetes and gene-obesity interaction scan in 8,255 cases and 11,900 controls from Pancreatic Cancer Cohort Consortium and Pancreatic Cancer Case Control Consortium

Abstract

Background.

Obesity and diabetes are major modifiable risk factors for pancreatic cancer. Interactions between genetic variants and diabetes/obesity have not previously been comprehensively investigated in pancreatic cancer at the genome-wide level.

Methods.

We conducted a gene-environment interaction (GxE) analysis including 8,255 cases and 11,900 controls from four pancreatic cancer GWAS datasets (PanScan I-III and PanC4). Obesity (BMI≥30 kg/m2) and diabetes (duration ≥ 3 years) were the environmental variables of interest. Approximately 870,000 SNPs (MAF ≥ 0.005, genotyped in at least one dataset) were analyzed. Case-control (CC), case-only (CO), and joint-effect test methods were used for SNP-level GxE analysis. As a complementary approach, gene-based GxE analysis was also performed. Age, sex, study site and principal components accounting for population substructure were included as covariates. Meta-analysis was applied to combine individual-GWAS summary statistics.

Results.

No genome-wide significant interactions (departures from a log-additive odds model) with diabetes or obesity were detected at the SNP level by the CC or CO approaches. The joint-effect test detected numerous genome-wide significant GxE signals in the GWAS main effects top hit regions but the significance diminished after adjusting for the GWAS top hits. In the gene-based analysis, a significant interaction of diabetes with variants in the FAM63A (family with sequence similarity 63 member A) gene (significance threshold P<1.25 × 10⁻⁶) was observed in the meta-analysis (P_GxE= 1.2 × 10⁻⁶, P_Joint= 4.2 × 10⁻⁷).

Conclusions.

Our current analyses did not find significant GxE interactions at the SNP level but found one significant interaction with diabetes at the gene level. A larger sample size might unveil additional genetic factors via GxE scans.

Impact.

This study may contribute to discovering the mechanism of diabetes-associated pancreatic cancer.

Introduction

Pancreatic cancer is the third leading cause of cancer death, accounting for more than 47,000 deaths each year in the United States (1). It is a highly lethal disease with a 5-year survival rate of 9% (2). Epidemiological studies have shown that 20–25% of pancreatic cancer cases are attributable to cigarette smoking (3). However, the incidence of pancreatic cancer has been rising slightly each year in the United States since 2002; this is unexpected given the decreasing prevalence of cigarette smoking and may be due to the rising prevalence of obesity and diabetes. Accumulating evidence suggests that obesity and long-term type II diabetes are associated with increased risk of pancreatic cancer. For example, a pooled analysis of 14 cohort studies of body mass index (BMI) has shown that obesity (BMI ≥30 kg/M²) was associated with 47% (95% confidence interval of 23%−75%) increased risk of pancreatic cancer (4). A meta-analysis of 23 cohort or case-control studies suggests that the association between BMI and pancreatic cancer is not linear (5). At least four meta-analyses of large datasets from cohort and case-control studies have shown that long-term diabetes was associated with a 1.5- to 2-fold increased risk of pancreatic cancer (6–9). Because only a small portion of obese and diabetic individuals develop pancreatic cancer, understanding how genetic factors affect risk among those individuals could inform targeted interventions or screening. Identifying variants that are only associated with risk of cancer (or have stronger associations) among obese or diabetic individuals is of particular interest.

Genome-wide association studies (GWAS) conducted by the Pancreatic Cancer Cohort Consortium (PanScan) and Pancreatic Cancer Case Control Consortium (PanC4) have identified 21 genetic loci and chromosome regions significantly associated with the risk of pancreatic cancer (10–15). However, these findings explain limited heritability of the disease, i.e, the established GWAS loci explain 2.1% of the heritability of pancreatic cancer in contrast to the estimated heritability of 36% from a large population-based twin study (13,16). Beyond main effects, some genetic factors may contribute to the risk of pancreatic cancer only in the presence of specific risk factors for the disease such as obesity and diabetes, i.e., gene-obesity/diabetes interaction, and broadly referred as gene-environment interaction (GxE) herein. Therefore, a genome-wide GxE scan may help find the missing heritability of pancreatic cancer. Several of the susceptibility genes identified by GWAS, i.e. NR5A2, PDX1, HNF1B, and HNF4G, are important for pancreas development (17). These genes are important components of the transcriptional networks governing embryonic pancreatic development and differentiation as well as maintaining pancreatic homeostasis. Mutations in some of these genes are responsible for maturity onset diabetes of the young (MODY) and common variants of these genes have been associated with BMI and risk of type II diabetes in GWAS (17). Therefore, in addition to their roles in regulating the development and function of the pancreas these genes may contribute to pancreatic cancer, partially through an increased risk of obesity and diabetes. Whether these genes and other unidentified genes have an interactive action with obesity and diabetes in modifying the risk of pancreatic cancer is the focus of the current investigation.

We have previously performed GxE analyses at SNP/gene/pathway-levels using GWAS data from 2,028 cases and 2,109 controls from PanScan I and II. No significant interactions at the SNP or gene levels were observed for diabetes or obesity. At the pathway level, NF-kappaB-mediated chemokine signaling and axonal guidance signaling pathway, respectively, were identified as the top pathway interacting with obesity and smoking in modifying the risk of pancreatic cancer (18,19). These studies were limited by the small sample size, and underpowered for genome-wide GxE analysis (20). To address this limitation, the current analysis is conducted in a much larger combined dataset of PanScan I-III and PanC4 with 8,255 cases and 11,900 controls and further leverages recently developed, more powerful SNP-set/gene-based GxE tests (21,22), to discover novel genetic variants that may modify the association between diabetics/obesity and pancreatic cancer.

Materials and Methods

Study population and datasets

The present genome-wide GxE study includes 8255 cases and 11900 controls of European-ancestry drawn from the PanScan and PanC4 consortia. Cases were patients with known or presumed primary pancreatic ductal adenocarcinoma (ICD-O-3 code C250-C259) and controls were free of pancreatic cancer. Individual studies were approved by the respective institutional review board following the institution’s requirement. Written informed consent was obtained from each study participant. The approaches for data harmonization and meta-analysis were approved by the University of Texas MD Anderson Cancer Center Institutional Review Board.

Genotype data were generated in four previously reported GWAS studies, i.e. PanScan I, II, III and PanC4, and the details of these studies have previously been described (10–13). Genotyping in PanScan I, II, and III was conducted at the Cancer Genomics Research Laboratory (CGR) of the National Cancer Institute (NCI) of the National Institutes of Health (NIH) using the Illumina HumanHap550 Infinium II, Human 610-Quad, and OmniExpress series arrays, respectively. PanC4 employed the HumanOmniExpressExome-8v1 array. Because different genotyping platforms were used in these studies, missing genotypes were imputed using the University of Michigan imputation server (https://imputationserver.sph.umich.edu/index.html) with the Haplotype Reference Consortium (23) as the reference panel or IMPUTE2 with the 1000 Genomes Phase 3 as the reference panel (https://mathgen.stats.ox.ac.uk/impute/impute_v2.html). After imputation, SNPs that were identified by imputation only (not genotyped in any of the four GWAS), having minor allele frequency (MAF) ≤ 0.005, imputation quality score < 0.3, or Hardy-Weinberg equilibrium test p value <1 × 10⁻⁶ in controls were excluded; a total of about 870,000 common SNPs to all four studies were included in the current GxE analysis. The PanScan (I, II, and III) and PanC4 GWAS data are available through dbGaP (accession numbers phs000206.v5.p3 and phs000648.v1.p1, respectively).

Exposure variables

The exposure variables considered in this GxE analysis were obesity (BMI≥30 kg/m² vs. <30 kg/m²) and diabetes (diabetes with ≥3-years of duration vs. non-diabetes). Because diabetes could be a manifestation of occult pancreatic cancer, we excluded diabetes with a short duration (< 3 years) for studies with diabetes duration information to control reverse causality. Covariates for adjustment included age (continuous), sex, study sites and principal components accounting for population substructure. The distribution of demographics and risk factors of participants in each GWAS study included in the current analysis is summarized in Table S1.

Statistical analyses

We applied case-control (CC), case-only (CO), and 2 degrees-of-freedom (df) joint-effect test (24) methods at the SNP level, and the “rareGE” method (21) at the gene level in the genome-wide GxE scan. The 2-df joint-effect test is more powerful in detecting a susceptibility SNP in the presence of strong genetic main effect (SNP), strong interaction effect (SNPxE), or a combination of weak/moderate main and interaction effects (SNP + SNPxE). Thus, the joint-effect test is a useful complementary approach to CC, CO and single-SNP marginal association analysis in identifying disease susceptible loci (20).

The PanScan I-III and PanC4 datasets were analyzed individually using the CC, CO, and joint-effect test at the SNP level. The “rareGE” method was used for gene-based G x E analysis. The summary statistics for each consortium were then subjected to meta-analysis.

SNP-level tests

In order to perform SNP-level analysis, we ran the logistic regression model as follows:

$$\text{Logit\hspace{0.17em}}\left(\text{P}\left(\text{Y}=1\right)\right)={\beta }_0+{\beta }_{\text{E}}\text{E}+{\beta }_{\text{G}}\text{SNP}+{\beta }_{\text{GE}}\text{SNP}*\text{E}+{\beta }_{\text{C}}\text{C},$$ [1]

where Y is the disease status (1 for case; 0 for control), β₀ is the intercept, E is the exposure variable of interest (diabetes or obesity), SNP is the dosage of the genetic variant of interest, coded additively accounting for genotype imputation uncertainty (ranging from 0 to 2), C is the vector of all covariates including age (continuous), sex, study indicators, principal components accounting for population substructures, and either diabetes or BMI (e.g., diabetes serves as the exposure of interest with BMI (continuous) included in the covariate vector). For the CC study design, the null hypothesis to be tested is H₀: β_GE =0. $e^{{\beta }_{\text{GE}}}$ is referred as the interaction odds ratio (OR).

Joint-effect analysis of SNP and SNP*E were run using the approach by Aschard et al (25) by testing the null hypothesis H₀: β_G = β_GE =0, derived from model [1] with a 2-df χ² Wald test. For the CO study design, a logistic regression model was run in the case group only as follows:

$$\text{Logit\hspace{0.17em}}\left(\text{P}\left(\text{E}=1\right)\right)={\beta }_0+{\beta }_{\text{G}}\text{SNP}+{\beta }_{\text{C}}\text{C},$$ [2]

where the coefficients in model [2] are denoted the same as those in model [1].

Gene-level tests

Gene regions were defined according to coordinates of the hg19 assembly, retrieved from the University of California, Santa Cruz (UCSC) Genome Browser (26). About 22,300 genes were downloaded from UCSC server, of which ~20, 000 genes covering ≥ 2 GWAS genotyped SNPs were analyzed in the current study.

We performed gene-based GxE analysis using the “rareGE” method (21) based on common SNPs (MAF ≥ 0.005, located within 20 kb upstream or downstream of a given gene). For a gene with p SNPs, the full model is as follows:

$$\text{Logit\hspace{0.17em}}\left(\text{P}\left(\text{Y}=1\right)\right)={\beta }_0+{\beta }_{\text{E}}\text{E}+{\sum }_{\text{j}=1}^{\text{p}}{\beta }_{\text{Gj}}{\text{SNP}}_{\text{j}}+{\sum }_{\text{j}=1}^{\text{p}}{\beta }_{\text{GEj}}{\text{SNP}}_{\text{j}}*\text{E}+{\beta }_{\text{C}}\text{C},$$ [3]

where β_Gj and β_GEj are the regression coefficients for the genetic main effect and GxE effect for the jth SNP, respectively.

Two tests were implemented in the “rareGE” R package: G x E test with genetic main effects estimated as random effects (P_Int) under the null hypothesis of no G x E, i.e., H₀: β_GE1 = β_GE2 = ⋯ = β_GEp = 0, and a joint test of G and GxE (P_Joint) with H₀: β_G1 = β_G2 = ⋯ = β_Gp = 0 and β_GE1 = β_GE2 = ⋯ = β_GEp = 0 analogous to the 2-df SNP-level joint-effect test.

Meta analyses

We applied fixed-effects meta-analysis in METAL to combine SNP-level GxE results from the CC or CO method across individual consortia (27). Fisher’s meta-analysis was used to combine gene-level G x E p-values from the rareGE method (28).

Statistical thresholds

All tests were two sided. We consider P < 2.5 × 10⁻⁸ and P < 1.25 × 10⁻⁶ as genome-wide significant at the SNP- and gene-level, respectively (29), for each individual study and each meta-analysis, adjusted for 1 million SNPs, 20,000 genes and 2 exposures of interest by the Bonferroni correction at family-wise error rate of 0.05. P < 5.0 × 10⁻² was considered as nominally significant for all analyses.

Statistical power estimation

We used the QUANTO software (version 1.2.4) (30) to perform power estimation for the current GxE scans. With 8,255 cases and 11,900 controls we had 80% power to detect an interaction OR of 1.5 and 1.6, respectively, for obesity (main effect OR = 1.2 with 20% prevalence in controls based on Table S1) and diabetes (main effect OR = 1.7 with 10% prevalence in controls based on Table S1) for a SNP with MAF of 20% at a significance level of 2.5 × 10⁻⁸ by the standard CC test.

Results

We examined the GxE (obesity and diabetes) interactions at the SNP level using the CC, CO and joint tests in each individual GWAS first, followed by meta-analysis of the summary statistics. Figure S1 shows the quantile-quantile (Q-Q) plots for the CC and CO meta-analyses. There was no discernable abnormal behavior in the QQ plots for CC and CO study designs (genomic control λ ranged from 0.942 to 1.023). QQ plots also performed well for meta-analysis of joint-effect tests (λs: 0.94–1.045).

Case-control (CC) and case-only (CO) analyses

No signal at a genome-wide threshold of significance (P < 2.5 × 10⁻⁸) was detected in CC or CO analyses on interactions of gene with diabetes or obesity. Using the CC approach, four SNPs on chromosomes 10, 18 and 20 showed evidence of interactions with diabetes at near genome-wide significance (P <1 × 10⁻⁶) and six SNPs on chromosomes 2, 7, 11 and 16 showed weaker evidence of interactions with obesity (P < 1 ×10⁻⁵) (Table 1). By the CO approach, four SNPs on chromosome 3 and chromosome 10 showed evidence of interactions with diabetes at near genome-wide significance (P < 1 × 10⁻⁶, Table 2). Of these, two SNPs (rs12255372 and rs7901695) were near TCF7L2 and are in linkage disequilibrium (LD) (r²=0.74 and 0.87, respectively) with the lead SNP from a recent GWAS of Type 2 diabetes (rs7903146; P=1×10⁻³⁴⁷) (31). Thus, the CO signals for these two SNPs likely reflect violations of the gene-environment independence assumption rather than evidence for gene-environment interaction. In addition, five SNPs on chromosomes 4, 8, 14 and 17 had possible interactions with obesity at a P value < 1 ×10⁻⁵ (Table 2). In addition, no significant across-consortium heterogeneity was found for the meta-analysis results in Table 1 and 2 (all heterogeneity test P-value > 0.05).

Table 1.

Top SNPs interacting with diabetes and obesity (Case-Control)

SNP	Chr.	Position	Genea	Effect/Ref Allele	MAFb	Meta analysis
						OR (95% CI)	P
Diabetes
rs7505930	18	4092001	*ROCK1P1-SLC35G4	G/A	0.35	1.60 (1.34–1.91)	1.9E-07
rs2777534	10	34109601	*GTPBP4-FGF8	A/G	0.12	2.04 (1.56–2.67)	2.3E-07
rs2812656	10	34116863	*GTPBP4-FGF8	G/A	0.12	0.50 (0.38–0.65)	2.4E-07
rs11086650	20	57183256	APCDD1L_AS1	C/T	0.32	0.61 (0.51–0.74)	5.8E-07
Obesity
rs7802442	7	22736446	*COX19-SLC12A9	C/A	0.31	0.73 (0.65–0.83)	1.2E-06
rs4298423	7	151643909	PRKAG2_AS1-GALNTL5*	A/G	0.34	1.34 (1.19–1.51)	2.3E-06
rs559449	11	55340379	OR4C16	A/G	0.45	1.31 (1.17–1.47)	3.6E-06
rs7608326	2	37903390	*GRHL1-CHST10	C/T	0.07	0.51 (0.38–0.68)	4.2E-06
rs759831	16	82863660	CDH13	A/C	0.31	1.32 (1.17–1.49)	5.5E-06
rs1476483	7	22731199	*COX19-SLC12A9	G/A	0.20	0.72 (0.62–0.83)	8.5E-06

Abbreviations: Chr. for chromosome; MAF for minor allele frequency; OR (95%CI) for interaction odds ratio (95% confidence interval);

^aGene region was defined by the UCSC Genome Browser;

^*marks the nearest gene to the SNP.

^bDerived from the PanC4 dataset.

Table 2.

Top interaction signals from case-only analyses

SNP	Chr.	Position	Genea	Eff/Ref Allele	MAFb	PCO
Diabetes
rs608841	3	138764229	*PRR23C-BPESC1	G/A	0.24	1.6E-07
rs696638	3	138775377	*PRR23C-BPESC1	A/G	0.16	2.2E-07
rs12255372	10	114808902	TCF7L2	A/C	0.28	2.3E-07
exm-rs7903146	10	114758349	TCF7L2	A/G	0.29	4.9E-07
Obesity
rs2018572	17	11599798	BHLHA9-DNAH9*	G/A	0.19	1.3E-07
rs4791473	17	11574959	BHLHA9-DNAH9*	G/T	0.18	1.9E-06
rs4413478	4	48491651	SLC10A4-ZAR1*	A/G	0.25	2.8E-06
rs925611	8	9768690	OR4F21-C8orf49*	T/G	0.097	3.2E-06
rs961044	14	87608094	*LOC283585-GALC	G/A	0.14	6.9E-06

Abbreviations: Chr. for chromosome; MAF for minor allele frequency; P_CO for case-only test p-value;

^aGene region was defined by the UCSC Genome Browser;

^*marks the nearest gene to the SNP.

^bDerived from the PanC4 dataset.

2-df joint-effect test

Meta-analysis of joint-effect tests for SNP and SNP x diabetes or SNP x obesity detected numerous genome-wide significant signals that are all located in the chromosome regions containing previously identified GWAS top hits (Table S2). Conditional analysis adjusting for the GWAS top hits in each region resulted in null findings, indicating that joint effect test signals were all driven by the strong main effects of the SNPs.

Gene-level G x E analysis

Possible interactions of nine genes with diabetes and three genes with obesity at a meta-analysis significance level of P < 1 × 10⁻⁴ in at least one of the interaction-only and joint tests are listed in Table 3. Among these genes, a significant (P <1.25 × 10⁻⁶) interaction of diabetes with FAM63A (family with sequence similarity 63 member A) gene was observed in the meta-analysis (P_Interaction = 1.2 × 10⁻⁶, P_Joint = 4.2 × 10⁻⁷) (Table 3). The SNPs contributing to this gene are listed in Table S3. No individual SNP of this gene showed a significant interaction with diabetes.

Table 3.

Top genes interacting with diabetes and obesity by rareGE method

Gene	Chr.	Meta		PanScan I		PanScan II		PanScan III		PanC4
		PInt	PJoint	PInt	PJoint	PInt	PJoint	PInt	PJoint	PInt	PJoint
Diabetes
FAM63A	1q21.3	1.2E-6*	4.2E-7*	3.8E-2	6.8 E-2	0.024	0.04	2.2E-4	8.8E-6	3.3E-3	8.1E-3
CLTCL1	22q11.21	1.5E-4	5.2E-4	0.85	0.98	4.9E-3	0.01	0.77	0.95	6.0E-5	1.0E-4
MIR561	2q32.1	4.1E-5	6.6E-4	9.3E-4	1.7E-3	0.043	8.5E-2	8.1E-3	3.5E-2	0.13	0.25
GNG2	14q22.1	3.4E-5	1.1E-3	0.76	0.66	3.4E-3	7.9E-3	2.5E-5	5.9E-4	0.51	0.76
ADA	20q13.12	6.8E-5	4.6E-4	0.14	0.27	1.8E-3	3.7E-3	0.47	0.6	0.97	0.38
TP53I3	2p23.2	7.0E-5	1.7E-3	0.31	0.52	0.28	0.36	2.0E-6	3.0E-5	0.46	0.71
SF3B14	2p23.3	6.9E-5	1.9E-3	0.31	0.52	0.28	0.36	2.0E-6	3.6E-5	0.45	0.7
DCAF6	1q24.1	2.7E-2	1.6E-5	2.4E-2	0.05	0.49	2.7E-5	0.21	7.3E-2	0.07	0.14
OR6K2	1q23.1	3.3E-6	4.0E-3	6.4E-2	0.12	0.037	7.3E-2	2.0E-6	2.9E-3	0.45	0.48
MIR4457	5p15.33	0.57	9.9E-6	0.14	2.9E-2	0.93	0.61	0.61	5.8E-4	0.44	7.6E-4
Obesity
CDC42EP3	2p22.2	3.40E-04	2.10E-05	0.18	0.17	0.62	4.3E-3	5.50E-02	1.70E-01	9.10E-05	1.5E-4
FSD1L	9q31.2	6.50E-02	3.60E-05	4.2E-2	0.066	6.1E-2	0.13	8.80E-01	8.70E-06	2.80E-01	0.48
MIR4457	5p15.33	3.10E-01	5.10E-06	0.45	0.0045	0.82	0.32	6.20E-01	6.50E-03	4.00E-02	3.8E-4

Abbreviations: Chr. for chromosome; P_Int, P_Joint for p values, respectively, derived from random-effect GxE interaction test and joint-effect test.

^*Genome-wide significant p-values (<1.25E-6)

Discussion

In this genome-wide G x obesity/diabetes interaction study of pancreatic cancer, no significant departures from a log-linear odds model at the SNP level were identified by the CC or CO approaches. In the gene-based analysis, a significant interaction between variants in the FAM63A gene and diabetes was observed.

FAM63A (family with sequence similarity 63, member A), also known as MINDY-1 (MINDY lysine 48 deubiquitinase 1) is a member of an evolutionarily conserved and structurally distinct family of deubiquitinating enzymes (32), which specifically cleaves K48-linked poly-ubiquitin chain to regulate protein degradation. This distinct deubiquitinase class localizes to DNA lesions, where it plays an important role in genome stability pathways, functioning to prevent spontaneous DNA damage and to promote cellular survival in response to exogenous DNA damage (33). Previous GWAS have associated FAM63A or FAM63A homolog gene variants with risk of primary rhegmatogenous retinal detachment (34) and chronic renal disease (35). Genetic analysis of a diabetes-prone mouse strain has revealed gene regions homologous to FAM63A contributing to diabetes susceptibility (36). Although the role of FAM63A in pancreatic cancer is unknown at present, the observed interaction with diabetes deserves further investigation.

Genome-wide GxE analysis has unique challenges compared with genetic main effects analysis in GWAS. First, GxE analysis requires a much larger sample size to detect a realistic interaction OR than does a GWAS scan for a comparable main effect OR (20,37), largely explaining why few positive findings have been reported in GxE studies (38–40). For example, the current GxE scan with 8,255 cases and 11,900 controls, even though about four times as large as our previous G x obesity/diabetes interaction analysis (18), had 80% power to detect an interaction OR of 1.5 and 1.6, respectively, for obesity and diabetes for a SNP with MAF of 20% at a significance level of 2.5 × 10⁻⁸ by the standard CC test; in contrast, the same sample size had 80% power to detect a genetic main effect OR of 1.18 at the same MAF and significance level. To boost the power for a given sample size, novel statistical and analytical methods have been proposed to leverage a priori biological knowledge in the form of genes, pathways or other functional genomics annotations such as those derived from the ENCODE and NIH Epigenomics Roadmap projects (18,19,41). Second, exposure variability and measurement accuracy play a considerable role in determining the power and reproducibility of GxE studies (42,43). Third, there is no single most powerful statistical method for either SNP or gene-level genome-wide GxE analysis due to the largely unknown patterns of GxE interaction signals and combinations of genetic main and GxE effects (20,22). Therefore, we suggest that the GxE analysis should make use of multiple methods with complementary strengths, as used here and suggested by other investigators (44), to discover the missing heritability of pancreatic cancer (45).

This study identified a statistically significant interaction of diabetes with variants in FAM63A in gene-based GxE analysis but no significant SNP-level GxE interactions with either diabetes or obesity. We note that the absence of interaction on the log-odds scale has potentially important implications for risk modeling, as it typically implies presence of interaction on the risk difference scale, sometimes referred to as “public health interaction” (46). Developing and validating a multifactorial risk model is beyond the scope of this paper, but we note that our results lend support to the common assumption of additive log odds when combining genetic, clinical, and environmental risk factors to predict risk (47,48).

The current study has strengths and limitations. This is by far the largest GxE analysis in pancreatic cancer. Quality control was strictly performed in steps of genotyping, population structure definition, exposure measurement and harmonization. Diabetes was defined as disease with ≥3-year duration, avoiding reverse causality. Along the same line, since it is common for pancreatic cancer patients to experience severe weight loss (43), we avoided using body weight at or close to cancer diagnosis for cases when calculating the BMI. Following the state-of-the-art analysis strategies in large consortium-based GxE scans (49,50), we only adjusted for a “minimum” set of covariates, including age, sex, study sites and principal components accounting for population substructure, in the regression analysis. As shown by the well-behaved QQ plots in Figure S1, there was no indication of uncontrolled confounding effects. Finally, genome-wide significant thresholds based on the Bonferroni correction were applied to reduce false-positive discovery. Nevertheless, relatively small sample sizes curbed the power of genome-wide GxE scan from case-control and case-only study designs. Despite this, the current GxE analysis discovered a novel susceptibility locus for pancreatic cancer using a gene-based GxE test, and may contribute to discovering the mechanism of diabetes-associated pancreatic cancer.

The abbreviations used are:

CC

case-control

CO

case-only

GxE

gene-environment interaction

GWAS

genome-wide association study

PanC4

Pancreatic Cancer Case Control Consortium

PanScan

Pancreatic Cancer Cohort Consortium

MAF

minor allele frequency