Diabetes and cancer risk: A Mendelian randomization study.

Earlier cohort studies using conventional regression models have consistently shown an increased cancer risk among individuals with type 2 diabetes. However, reverse causality and residual confounding due to common risk factors could exist, and it remains unclear whether diabetes per se contributes to cancer development. Mendelian randomization analyses might clarify the true association between diabetes and cancer risk. We conducted a case-cohort study with 10,536 subcohort subjects and 3,541 newly diagnosed cancer cases derived from 32,949 eligible participants aged 40-69 years within the Japan Public Health Center-based Prospective Study. With 29 known type 2 diabetes susceptibility variants, we used an inverse variance-weighted method to estimate hazard ratios for the associations of diabetes with risks of total and site-specific cancers. The hazard ratios of cancer per doubling of the probability of diabetes were 1.03 (95% confidence interval [CI], 0.92-1.15) overall, 1.08 (95% CI: 0.73-1.59) for the pancreas, 0.80 (95% CI: 0.57-1.14) for the liver and 0.90 (95% CI: 0.74-1.10) for the colorectum. Additional analyses, using publicly available large-scale genome-wide association study data on colorectal cancer in Japan, resulted in a narrower CI (hazard ratio: 1.00; 95% CI: 0.93-1.07). In this prospective Mendelian randomization study with a large number of incident cancer cases, we found no strong evidence to support associations between diabetes and overall and site-specific cancer risks. Our findings suggest that there is little evidence to support the genetic role of type 2 diabetes in cancer development in the Japanese population.

BioBank Japan

body mass index

confidence interval

Fundamental Innovative Oncology Core

genome‐wide association study

International Classification of Diseases for Oncology, Third Edition

inverse variance‐weighted

Japan Public Health Center‐based Prospective

single nucleotide polymorphism

Introduction

Over the decades, a number of cohort studies reported associations between type 2 diabetes and increased cancer risks, especially liver, colorectal and pancreatic cancers.1, 2 More recently, studies have shown that high levels of blood glucose3, 4 or glycated hemoglobin A1c5 are associated with cancer incidence. These associations are biologically plausible, given that hyperglycemia increases mitochondrial glucose oxidation, thereby promoting DNA damage through oxidative stress. Because diabetes or high glucose levels could be prevented through lifestyle modifications or medications, these could be potential strategies for cancer prevention.

However, the causal relationship between diabetes and cancer risk has not been established yet. Diabetes and cancer have many common risk factors such as unhealthy diet, physical inactivity, smoking and adiposity, which raise the likelihood of residual confounding. If diabetes occurs secondary to precancerous state or subclinical cancer, reverse causation could also be present.

A Mendelian randomization (MR) approach may overcome the limitations of traditional observational study designs by using Mendel's law. The MR approach utilizes the random allocation of genotypes at conception, which makes genotypes to be independent of potential confounders and also avoids reverse causation.6

Large‐scale MR studies have been increasingly reported in recent years.7, 8 However, these studies often consist of mixed populations mainly from traditional case–control studies, with high risk of selection bias and heterogeneity, and lack important information about lifestyle factors prior to the onset of cancer. Thus, MR studies from prospective cohort studies with detailed covariates at baseline would probably give less biased estimates than those from case–control studies. Therefore, we performed an MR analysis to clarify the true association between diabetes and cancer risk using genotyped and imputed genetic data from the Japan Public Health Center‐based Prospective (JPHC) Study.

Methods

Study population

The JPHC study was initiated in 1990 (cohort I) and 1993–1994 (cohort II). All subjects were Japanese from 11 public health center (PHC) areas and were aged 40–59 years in 1990 (cohort I) and 40–69 years in 1993 (cohort II) at the time of their first survey. The JPHC Study has been described in detail previously.9 Our study population was derived from a cohort of 33,736 subjects from nine PHC areas across Japan, who responded to the baseline questionnaire and provided blood samples during the health checkup. We excluded 787 participants who had past history of cancer or cancer incidence before their return of the self‐administered questionnaire. Our study was approved by the institutional review board of the National Cancer Center, Japan.

Follow‐up and case–cohort selection

We followed up the subjects from the start of the study period until December 31, 2009. Changes in residence status and deaths were ascertained through the residential registry. Among the study population of 32,949 subjects, 1,946 (5.9%) died, 1,085 (3.3%) moved out of the study area and 84 (0.3%) were lost to follow‐up. Cancer occurrence was documented through active notifications from the major hospitals in the study areas and data linkage with population‐based cancer registries. Death certificates were also used as a supplementary information source. The site and histological features of each cancer case were coded according to the International Classification of Diseases for Oncology, Third Edition (ICD‐O‐3).10 For the registry system used, 4.1% of the cases only had information available from the death certificates. For analysis, the earliest date of diagnosis was considered for cases with multiple primary cancers occurring at different times. For the present case–cohort study, all incident cancers were included as case subjects. Cancer was newly diagnosed in 3,750 subjects (2,028 men and 1,722 women) during a median follow‐up of 15.9 years. A subcohort sample of 13,024 participants was randomly selected from the base cohort participants irrespective of the cancer and follow‐up status, comprising approximately 40% of the base cohort. We further excluded 298 subjects with past history of cancer and four with missing follow‐up times, from the subcohort sample, leaving 12,722 subcohort subjects and 3,750 incident cancer cases potentially eligible for the analyses.

Laboratory assays

For the potential gain in power or the reduction in variance, given a fixed maximum cost for our study, a case–cohort design was chosen for the present analyses on genetically predicted diabetes in relation to cancer risk. For each participant, DNA was extracted from a buffy coat of white blood cells using FlexiGene DNA kits (Qiagen, Hilden, Germany). DNA samples of the participants were analyzed using the HumanOmniExpressExome‐8 v1.2 BeadChip, the HumanOmniExpress‐12 BeadChip, or the HumanOmni2.5‐8 BeadChip arrays (Illumina Inc., San Diego, CA). The genotyping for our study was conducted at the Genetics Division or the Department of Clinical Genomics, Fundamental Innovative Oncology Core (FIOC), National Cancer Center Research Institute or at the RIKEN Center for Integrative Medical Sciences. Among the potentially eligible subjects, we excluded those with inappropriate samples, no calls, call rate per sample of <0.99, inconsistent sex, non‐Japanese samples based on principal component analysis using the smartpca,11 or contaminated samples. Applying these sample exclusion criteria, genotype data were available for 12,158 subcohort subjects and 3,680 incident cases. We further excluded closely related samples using concordance rate, leaving 10,536 subcohort subjects and 3,541 incident cases. For imputation, we excluded single nucleotide polymorphisms (SNPs) using standard SNP quality control (call rate per SNP of <0.99, minor allele frequency of <0.01, p‐value of Hardy–Weinberg equilibrium in the control group of <1 × 10−4, a large allele frequency difference between the study sample and the reference panel, or SNPs on the Y chromosome). Then, we performed imputation of ungenotyped SNPs using SHAPEIT and IMPUTE2 with the 1,000 Genome Project Phase 3 (all ethnic groups) as a reference panel.

We selected a total of 29 SNPs, identified through published genome‐wide association studies (GWASs), up until July 2014, and replicated in Asian populations as follows: GCKR, rs78009412; IRS1, rs294364113; DNER, rs186161214; UBE2E2 rs678056915; PSMD6, rs83157116; IGFBP2, rs440296017; MAEA, rs681546416; CDKAL1, rs775699218; ZFAND3, rs947079416; KCNK16, rs153550016; JAZF1, rs86474519; GCC/PAX4, rs646713616; ARF5/PAX4/SND1, rs1022958320; MIR129/LEP, rs79159521; ANK1, rs51507122; SLC30A8, rs1326663418; GLIS3, rs704184716; CDKN2A/B, rs238320823; GPSM1, rs1178779221; CDC123/CAMK1D, rs1277979019; HHEX, rs111187517; TCF7L2, rs790314617; KCNQ1, rs223789224; KCNJ11, rs521917; FTO, rs805013625; C2CD4A, rs717243215; SLC16A13, rs31245721; PEPD, rs378689716; FITM2/R3HDML/HNF4A, rs6017317.16 Among the 29 SNPs, 11 were genotyped and 18 were imputed (Supporting Information Table S1). With the exception of rs7041847 (INFO as a measure of imputation accuracy of 0.5168), all imputed SNPs showed high imputation quality (INFO > 0.9). Thus, we performed an analysis excluding rs7041847 on GLIS3; however, the results did not materially change, and rs7041847 was included in the main results.

Definition of diabetes

Diabetes was defined by (i) a self‐reported diabetes diagnosis and/or use of a glucose‐lowering medication via a self‐administered questionnaire at baseline or (ii) high blood glucose levels during the health checkups at baseline: ≥ 126 mg/dl for those who fasted for ≥8 hr or ≥200 mg/dl for nonfasting subjects26 (defined by <8 hr or missing data on fasting hours). Blood glucose levels were available for 8,659 subjects among the subcohort and 2,815 incident cancer cases. In a validation study conducted in a subpopulation of the JPHC Study, 94% of subjects with self‐report of diabetes were confirmed to have diabetes, after a medical chart review.27

Statistical methods

To investigate whether the reported diabetes susceptibility SNPs were associated with diabetes in our study, we fitted the logistic regression model with adjustment for age, sex and study areas under the additive genetic model among the 10,536 subcohort subjects (Supporting Information Table S2). Next, we examined the impact of diabetes susceptibility SNPs on total cancer and cancer risks at major sites: stomach (ICD‐O‐3: C16), colorectum (C18–C20), liver (C22), pancreas (C25), lung (C34), breast (C50) and prostate (C61) (Supporting Information Tables S3–S10). We further divided the participants into five groups based on quintiles of the total number of risk alleles of the diabetes susceptibility SNPs (Supporting Information Table S1) among the subcohort subjects. Then, we fitted the logistic regression model with adjustment for age, sex and study areas to estimate the odds ratios for diabetes by each quintile among the 10,536 subcohort subjects. We also fitted the weighted Cox model28 with stratification by study areas to examine hazard ratio for cancer by each quintile, with adjustment for age (continuous) and sex (men vs. women) among the 10,536 subcohort subjects and 3,541 incident cases. Furthermore, to compute p values for linear trend, we included the total number of risk alleles as a continuous variable to the logistic and Cox models.

For conventional analyses on associations between diabetes and total cancer risk and risks of cancer at major sites, we used the weighted Cox proportional hazards model with stratification by study areas and adjustment for age (continuous), sex (men vs. women), body mass index (BMI; continuous), smoking (never, past or current smokers; categorical), alcohol intake (none, occasional or regular: 1–150 g/week; or regular drinkers: >150 g/week; categorical), coffee consumption (almost none, 1–2 days/week, 3–4 days/week, 1–2 cups/day, 3–4 cups/day or ≥5 cups/day; categorical) and physical activity (none, 1–3 days/month or ≥1/week; categorical). To address missing data on BMI, smoking, alcohol intake, coffee consumption and physical activity, we performed multivariate normal imputation (the SAS PROC MI procedure) with 100 rounds of multiple imputations. We included all covariates, follow‐up length and outcome status to account for missing data. We then averaged the estimates from each imputed dataset according to Rubin's rules using the SAS PROC MIANALYZE procedure.

For instrumental variable analyses, we used estimates for each SNP‐cancer association (the first stage model: Supporting Information Tables S3–S10) and estimates for each SNP‐diabetes association (the second stage model: Supporting Information Table S2) as described above. To increase the statistical power of detecting a significant association, we also utilized the publicly available results for SNP‐colorectal cancer associations from the BioBank Japan (BBJ) Project (sex‐combined: 6,692 colorectal cancer patients and 27,178 controls; Supporting Information Table S11).29 Using the estimates for 29 SNPs, we employed an inverse variance‐weighted (IVW) method, a summary data‐based MR analysis,30, 31 to estimate MR estimates for the association between diabetes and cancer risk. The original MR estimate represents the change in the log hazard for cancer per unit increase in the log odds of diabetes. A one‐unit increase in the log odds of diabetes corresponds to an exp (1) = 2.72‐fold multiplicative increase in the odds of diabetes. To enable clearer interpretation of this relation, we have multiplied the original MR estimate by loge2 (= 0.693), yielding the change in log hazard per twofold increase (doubling) in the odds of diabetes. Because diabetes was a rare disease in Japan during the study period, the odds approximates the probability. We then exponentiated the estimates to obtain the hazard ratios for cancer per doubling in the probability of diabetes.32

MR analyses derive valid estimates where the following assumptions are met33: (i) the SNPs are correlated with diabetes, (ii) the SNPs affect cancer risk only through their effects on diabetes and (iii) the SNPs are independent of any confounding factors for the association between diabetes and cancer risk. For assumption (i), because we selected SNPs that were identified through the GWAS and replicated in Asian populations, such SNPs are likely to be correlated with diabetes in the source population.

For assumptions (ii) and (iii), we searched the PhenoScanner database (available at http://www.phenoscanner.medschl.cam.ac.uk/phenoscanner)34 and the Haploreg database (available at: http://archive.broadinstitute.org/mammals/haploreg/haploreg.php) to examine whether the chosen SNPs had pleiotropic effects on other phenotypes.35 Some SNPs were associated with traits such as fasting insulin levels at the GWAS significance level (p < 5.0 × 10−8); rs780094 on the GCKR was significantly associated with fasting insulin‐related traits;12 rs7756992 on the CDKAL1 was associated with insulin‐related traits;36 rs864745 on JAZF1 was associated with Crohn's disease37 but not likely to directly affect diabetes status; rs7903146 on TCF7L2 was associated with proinsulin and insulin levels;38 and rs2237892 on KCNQ1 was associated with BMI.39 Furthermore, among the subcohort subjects of our study, we examined whether SNPs were significantly associated with established risk factors for cancer, including BMI, smoking, alcohol intake and physical inactivity after applying the Bonferroni correction by dividing the p‐value of 0.05 by the number of SNPs examined (p < 0.05/29 ≈ 1.72 × 10−3).

Manual pruning of SNPs that potentially have horizontal pleiotropic effects is not generally recommended, because such pruning might lead to an instrument that is not biologically meaningful.40 Instead, as a sensitivity analysis, we conducted an MR‐Egger regression, which provides valid estimate even in the presence of unbalanced pleiotropy.41 We further performed the leave‐one‐out analysis by recalculating the MR estimates with the IVW method by sequentially dropping one SNP at a time to examine whether a single SNP that might have a large horizontal pleiotropic effect would influence the MR estimates.31

Thresholds for nominal significance were set at a two‐sided p‐value of <0.05. Statistical analyses were conducted using SAS (version 9.4; SAS Institute, Cary, NC) or R (version 3.4.2; R Foundation for Statistical Computing, Vienna, Austria).

Results

Among the subcohort subjects (n = 10,536), 5.8% (n = 613) had diabetes at baseline. Among them, 447 were known to have diabetes, while 166 were identified to have diabetes by blood glucose testing during the health checkups. Among those with incident cancer (n = 3,541), 8.4% (n = 298) had diabetes at baseline. Of them, 230 were known to have diabetes and 68 were identified to have the condition by blood glucose testing. Compared to subjects without diabetes, those with diabetes tended to be older individuals and smokers, had a higher BMI, engaged in more physical activity, drank alcohol more frequently and consumed coffee less frequently (Table 1).

Table 1

Baseline characteristics among subcohort subjects (n = 10,536)

	All subjects	Subjects without diabetes	Subjects with diabetes
Characteristic1	n = 10,536	n = 9,923	n = 613
Age, years	553.9 ± 7.9	53.7 ± 7.9	56.6 ± 7.7.6
Men, %	35.1	33.7	57.3
BMI, kg/m2	23.6 ± 3.0	23.6 ± 3.0	24.2 ± 3.3
Missing data, %	0.9	0.8	1.6
Smoking status
Current smoking, %	17.9	17.2	29.2
Past smoking, %	10.0	9.7	16.2
Missing data, %	0.5	0.5	0.3
Physical activity
≥1 day/week, %	17.5	17.4	20.4
Missing data, %	1.4	1.4	1.3
Alcohol consumption
Current drinking,1 %	28.1	27.2	41.4
Missing data, %	2.7	2.6	3.4
Coffee consumption
≥1 day/week, %	30.9	31.4	21.9
Missing data, %	1.0	0.9	1.1

Data are presented as mean ± standard deviation.

Alcohol consumption at least once per week.

Abbreviation: BMI, body mass index.

Among the 29 SNPs, four (rs7756992, rs2383208, rs1111875 and rs2237892) were nominally significantly associated with diabetes in our study (Supporting Information Table S2). Quintiles of the total number of risk alleles were positively associated with diabetes in a dose–response manner (p‐value for liner trend = 1.25 × 10−5) but were not significantly associated with cancer risk (overall: p‐value for linear trend = 0.95; stomach: 0.44; colorectum: 0.18; liver: 0.94; pancreas: 0.38; lung: 0.89; breast: 0.35; prostate: 0.52; Fig. 1).

Figure 1

Diabetes/cancer risk by quintiles of total number of risk alleles for diabetes. Abbreviation: CI, confidence interval. Hazard ratios were estimated using the weighted Cox model to examine cancer risk by each quintile comparing to the lowest quintile with adjustment for age (continuous), sex (except for prostate and breast cancer), and study areas among the 10,536 subcohort subjects and 3,541 incident cases. Odds ratios were estimated using the logistic regression model with adjustment for age, sex, and study areas to examine the association of diabetes by each quintile comparing to the lowest quintile among the 10,536 subcohort subjects.

In contrast to conventional analyses showing a positive association between diabetes and cancer risks, the MR analysis indicated no strong evidence to support associations between diabetes and risks of total cancer and cancer subtypes (Figs. 2 and 3). The IVW hazard ratios per doubling of the probability of diabetes for cancer were 1.03 (95% confidence interval [CI], 0.92–1.15) overall, 1.08 (0.73–1.59) for the pancreas, 0.80 (0.57–1.14) for the liver and 0.90 (0.74–1.10) for the colorectum (Fig. 3). The Egger estimates resulted in slightly wider confidence intervals. We also examined whether the selected variants were known associated risk factors of cancer. After the Bonferroni correction, statistically significant findings were observed between three SNPs and associated risk factors: rs780094 at the GCKR and BMI, p‐value = 5.8 × 10−4; rs8050136 at the FTO and BMI, p‐value = 6.87 × 10−12; and rs5219 at the KCNJ11 and regular alcohol drinking, p‐value = 1.64 × 10−3; data not shown). However, p values for intercepts from the Egger regression did not suggest the possibility of pleiotropic effects (Fig. 3). Furthermore, the leave‐one‐out analyses, performed by sequentially dropping one SNP at a time, resulted in similar findings (Fig. 4). Additional analyses, using publicly available genetic data on SNP‐colorectal cancer associations from the BBJ Project (29 SNPs), resulted in a narrower confidence interval (hazard ratio for colorectal cancer = 1.00 [0.93–1.07]; Fig. 3 and Supporting Information Table S11).

Figure 2

Scatter plot showing the associations of the SNP effects on the diabetes (log odds ratio) against the SNP effects on the total cancer (log hazard ratio). Abbreviations: MR, Mendelian randomization; SNP, single nucleotide polymorphisms. The inverse variance‐weighted estimates are represented by a dotted light blue line, and the Egger regression estimates by a dark blue line.

Figure 3

Association between diabetes and cancer risk. Abbreviations: BBJ, BioBank Japan; CI, confidence interval; HR, hazard ratio; IVW, inverse variance‐weighted; JPHC, Japan Public Health Center‐based Prospective; MR, Mendelian randomization; NA, not applicable; SNP, single nucleotide polymorphism. *Number of cases/controls are shown for the colorectum (two‐sample). The conventional HRs for cancer comparing people with diabetes to those without diabetes were estimated using the weighted Cox model with adjustment for age, sex (except for breast and prostate cancer), study areas, body mass index, smoking, alcohol intake, coffee intake and physical activity. The IVW HRs per doubling of the probability of diabetes were estimated using the IVW method with the estimates from the first and second stage models for 29 SNPs. The Egger HRs per doubling of the probability of diabetes and p‐values for the Egger intercept were estimated using the MR‐Egger method with the estimates from the first and second stage models for 29 SNPs. All results, except for the colorectum (two‐sample), are summary data‐based MR analyses in one sample. For the colorectum (two‐sample), we employed a summary data‐based MR in two samples (i.e., the JPHC data for SNP‐diabetes associations and the BBJ data for SNP‐cancer associations).

Figure 4

Leave‐one‐out analysis. Abbreviations: CI, confidence interval; HR, hazard ratio; IVW, inverse variance‐weighted; MR, Mendelian randomization; SNP, single nucleotide polymorphism. The leave‐one‐out analysis was performed by recalculating the MR estimates using the IVW method, by sequentially dropping one SNP at a time to examine whether a single SNP that might have a large horizontal pleiotropic effect would influence the MR estimates. The IVW HRs per doubling of the probability of diabetes were estimated using the IVW method with the estimates from the first and second stage models for 29 SNPs. [Color figure can be viewed at http://wileyonlinelibrary.com]

Discussion

In this population‐based prospective study, we utilized the MR methodology and examined the relation between diabetes and cancer risk in a large number (3,541) of incident cancer cases from 32,949 people as a source population, during a median follow‐up of 15.9 years. In our MR analysis, no strong evidence was found to support associations between diabetes and the risks of total cancer, colon cancer, pancreatic cancer or liver cancer. These findings were robust to a number of sensitivity analyses. To further complement our results, we utilized publicly available large‐scale GWAS results of colorectal cancer (the BBJ Project) and confirmed that genetically predicted diabetes was not associated with colorectal cancer. The results of our study support the notion that diabetes per se might not be responsible for the reported positive association between diabetes and cancer. To the best of our knowledge, this is the first prospective MR study that examined the association between type 2 diabetes and risks of total cancer and site‐specific cancer.

Our results are in general agreement with those reported in earlier MR studies with case–control data mainly conducted among European populations using a case–control strategy.7, 8 A recent MR study led by researchers from the International Agency for Research on Cancer investigated the associations of type 2 diabetes and metabolic factors with pancreatic cancer among 7,110 pancreatic cancer patients and 7,264 controls.7 Using 43 of the susceptibility SNPs for type 2 diabetes, their MR analysis indicated that diabetes was not significantly associated with pancreatic cancer (odds ratio = 1.03; 95% CI = 0.85–1.11; p = 0.47). In contrast, their MR analysis found BMI and hyperinsulinemia to be significantly associated with pancreatic cancer. These findings and our prospective findings seem to suggest that there is little evidence to support the genetic role of type 2 diabetes in cancer development.

The null findings observed in our prospective MR analyses could be explained by several interpretations. One possibility is that our study did not have sufficient power to detect a significant association between type 2 diabetes and cancer risk. As shown in post hoc power calculations, we had 80% power to detect relatively small effect sizes for total cancer but not for cancer sites (Supporting Information Table S12). To complement the analysis, we used the two‐sample MR analysis by using the JPHC data for SNP‐diabetes associations and the BBJ data for SNP‐colorectal cancer associations, which had 80% power to detect a small effect size (Supporting Information Table S11); however, the findings were consistent overall. Given the relatively wide confidence intervals for other cancer sites, a small effect of diabetes on these cancer sites cannot be ruled out. Another possibility is that type 2 diabetes susceptibility variants tended to be associated with both lower insulin levels and higher glucose levels. This may lead to vertical pleiotropy in which such genetic variants are associated with insulin and glucose levels on the same biological pathway from diabetes to cancer, thus not violating the MR assumptions. However, such genetic variants could also be associated with cancer through discrete pathways, which may lead to horizontal pleiotropy that violates the MR assumptions. For instance, variants in the KCNQ1 gene might affect insulin secretion, leading to hyperglycemia and hypoinsulinemia, both of which could affect cancer risks. In this case, a possible carcinogenicity of hyperglycemia might have been wiped out by lower insulin levels, leading to null associations. However, the leave‐one‐out analysis also indicated null associations (Fig. 4). Furthermore, the Egger regression did not suggest the possibility of pleiotropic effects (Fig. 3). Finally, the impact of diabetes per se on cancer development, if any, may be smaller than we thought. The conventional regression analyses might have overestimated the true association, possibly due to uncontrolled confounding by common risk factors or reverse causation.

Our MR analysis provides evidence that the genetic mechanisms responsible for type 2 diabetes may not play major roles in cancer development. However, from a public health perspective, it is certainly important to control such shared risk factors for the prevention of both diseases, because diabetes and cancer share a number of established modifiable risk factors such as obesity, physical inactivity and smoking.

The strengths of our study include its prospective, nationwide, population‐based design and high rate of follow‐up. Our findings might be generalizable to non‐Japanese populations, such as Asian populations with similar genetic and environmental background. Nevertheless, several limitations need to be addressed. First, we were unable to verify assumptions (ii) and (iii) of MR analyses listed in the “Statistical Methods” section. Although we exerted our best efforts to address potential pleiotropic or confounding effects, the possibility remains. Most importantly, as stated above, some variants included in our study are known to be associated with traits related to insulin rather than glucose, which may lead to horizontal pleiotropy. This potentially violates the assumption (ii) that SNPs affect cancer risk only through their effects on diabetes. Instead of using diabetes susceptibility variants, the use of variants related to glucose levels with such known function as glucose transporters might be better to meet this assumption. Second, although we included a large number of incident cases, the number of cases for each cancer site was limited. Therefore, we were probably underpowered to detect small effects for several cancer sites. However, the use of a publicly available large‐scale GWAS on colorectal cancer from the BBJ Project resulted in a similar finding. Thus, similar analyses with the use of large‐scale GWAS results among Japanese or Asian populations for other cancer sites should be performed, once the SNP‐outcome associations become available.

In conclusion, in this prospective MR study with a large number of incident cancer cases, we found no strong evidence to support the associations between diabetes and overall and site‐specific cancer risks. Our findings suggest that there is little evidence to support the genetic role of type 2 diabetes in cancer development in a Japanese population.

Table S1 Characteristics of SNPs included in the analysis among subcohort participants (n = 10,539)

Table S2 Associations for prevalent diabetes by SNP among subcohort participants under the additive genetic model (among 10,539 subcohort subjects; 3,699 men and 6,837 women)

Table S3 Associations for total cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 3,541 incident cases)

Table S4 Associations for stomach cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 687 incident cases)

Table S5 Associations for colorectal cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 664 incident cases)

Table S6 Associations for liver cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 168 incident cases)

Table S7 Associations for pancreatic cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 129 incident cases)

Table S8 Associations for lung cancer risk by SNP under the additive genetic model (10,536 subcohort subjects and 437 incident cases)

Table S9 Associations for breast cancer risk by SNP under the additive genetic model (6,837 women in the subcohort and 239 women with incident cancer)

Table S10 Associations for prostate cancer risk by SNP under the additive genetic model (3,699 men in the subcohort and 294 men with incident cancer)

Table S11 Associations for colorectal cancer by SNP under the additive genetic model in the Biobank Japan Project (6,692 cases and 27,178 controls)

Table S12. Power calculation.

Click here for additional data file.