Diabetes mellitus in relation to colorectal tumor molecular subtypes: A pooled analysis of more than 9000 cases

Abstract Diabetes is an established risk factor for colorectal cancer. However, colorectal cancer is a heterogeneous disease and it is not well understood whether diabetes is more strongly associated with some tumor molecular subtypes than others. A better understanding of the association between diabetes and colorectal cancer according to molecular subtypes could provide important insights into the biology of this association. We used data on lifestyle and clinical characteristics from the Colorectal Cancer Family Registry (CCFR) and the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO), including 9756 colorectal cancer cases (with tumor marker data) and 9985 controls, to evaluate associations between reported diabetes and risk of colorectal cancer according to molecular subtypes. Tumor markers included BRAF and KRAS mutations, microsatellite instability and CpG island methylator phenotype. In the multinomial logistic regression model, comparing colorectal cancer cases to cancer‐free controls, diabetes was positively associated with colorectal cancer regardless of subtype. The highest OR estimate was found for BRAF‐mutated colorectal cancer, n = 1086 (ORfully adj: 1.67, 95% confidence intervals [CI]: 1.36‐2.05), with an attenuated association observed between diabetes and colorectal cancer without BRAF‐mutations, n = 7959 (ORfully adj: 1.33, 95% CI: 1.19‐1.48). In the case only analysis, BRAF‐mutation was differentially associated with diabetes (P difference = .03). For the other markers, associations with diabetes were similar across tumor subtypes. In conclusion, our study confirms the established association between diabetes and colorectal cancer risk, and suggests that it particularly increases the risk of BRAF‐mutated tumors.


Abstract
Diabetes is an established risk factor for colorectal cancer. However, colorectal cancer is a heterogeneous disease and it is not well understood whether diabetes is more strongly associated with some tumor molecular subtypes than others. A better understanding of the association between diabetes and colorectal cancer according to molecular subtypes could provide important insights into the biology of this association. We used data on lifestyle and clinical characteristics from the Colorectal Cancer Family Registry (CCFR) and the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO), including 9756 colorectal cancer cases (with tumor marker data) and 9985 controls, to evaluate associations between reported diabetes and risk of colorectal cancer according to molecular subtypes. Tumor markers included BRAF and KRAS mutations, microsatellite instability and CpG island methylator phenotype.
In the multinomial logistic regression model, comparing colorectal cancer cases to cancer-free controls, diabetes was positively associated with colorectal cancer regardless of subtype. The highest OR estimate was found for BRAF-mutated colorectal cancer, n = 1086 (OR fully adj : 1.67, 95% confidence intervals [CI]: 1.36-2.05), with an attenuated association observed between diabetes and colorectal cancer without BRAF-mutations, n = 7959 (OR fully adj : 1.33, 95% CI: 1. 19-1.48). In the case only analysis, BRAF-mutation was differentially associated with diabetes (P difference = .03). For the other markers, associations with diabetes were similar across tumor subtypes. In conclusion, our study confirms the established association between diabetes and colorectal cancer risk, and suggests that it particularly increases the risk of BRAF-mutated tumors.

What's new?
Diabetes is a well-known risk factor for colorectal cancer, but colorectal cancer varies widely among patients. To better understand the association between diabetes and particular molecular subtypes of colorectal cancer, these authors analyzed data from 9,756 colorectal cancer cases and 9,985 controls. They found that diabetes appears to increase the risk of tumors with BRAF mutations, which generally have poorer outcomes. The large pooled dataset allowed detection of even small variations among subtypes, but the study also was not able to account for some potentially relevant factors, such as metformin use.

| INTRODUCTION
Metabolic health, and excess body fat in particular, are involved in the development of colorectal cancer. 1-3 Individuals with diabetes mellitus, especially those with type 2 diabetes, have an increased risk of developing colorectal cancer. 4 This connection is likely independent of shared risk factors between diabetes and colorectal cancer. 5 Instead, the association between diabetes and colorectal cancer may depend on other mechanisms such as alterations to the gut microbiome, increased inflammation in the gut, hyperinsulinemia in early stage type 2 diabetes and activation of cancer promoting pathways. 6 Colorectal cancer is a heterogeneous disease, displaying considerable differences in molecular markers, which correlate with anatomical tumor location and other clinical and patient characteristics.
For example, BRAF-mutations are more common in proximal colon cancer, older patients and women and often co-occur with highlevel microsatellite instability (MSI). 7 This raises the question of whether the risk factors for colorectal cancer also vary by tumor molecular markers. Investigations into potentially variable associations of metabolic factors and molecular subtypes of colorectal cancer have been inconsistent, but some reports suggest a possible association between adiponectin and lower risk of KRAS-mutated colorectal cancer. 8,9 Adiponectin is known to have anti-inflammatory effects and has been suggested as a potential treatment for obesity and type 2 diabetes. 10 Although studies have examined associations between body mass index (BMI) and different colorectal cancer phenotypes, 11,12 to our knowledge, no studies have investigated diabetes in relation to the risk of molecular subtypes of colorectal cancer, despite the fact that high BMI is one of the strongest known predictors of type 2 diabetes risk. 13 The aim of the present study was to investigate self-reported dia-  17 We created two CIMP categories for this analysis: CIMP-high and CIMP-low/negative. In instances in which studies categorized tumors as CIMP-high, CIMP-low and CIMP-negative, we combined CIMP-low and CIMPnegative into the CIMP-low/negative category.

| Exposure data
Data collection and harmonization of GECCO and CCFR epidemiologic data have been described elsewhere. 14,15,21 Briefly, demographic and environmental risk factor data were self-reported at in-person interviews or via structured self-administered questionnaires. Data were collected at study entry, or 1 to 2 years prior to sample ascertainment. A multistep iterative data-harmonization procedure was applied, reconciling each study's unique protocols and data collection instruments. Multiple quality-control checks were performed, and outlying values of variables were truncated to the minimum or maximum value of an established range for each variable. Variables were combined into a single dataset with common definition, standardized coding and standardized permissible values. Diabetes status was obtained through self-reported answers to questions about diabetes diagnoses (summarized in Table S1) and includes, but does not distinguish between, both type 1 and type 2. We defined age at the time of a colorectal cancer diagnosis for cases and time of enrolment for controls. Missing covariate data were assumed to be missing at random, conditional on observed data and were imputed using mean imputation.

| Statistical analyses
We used multinomial models to estimate odds ratios (OR) and 95% confidence intervals (CIs) for the association between diabetes and the risk of each molecular tumor marker among colorectal cancer cases, defined as MSI-high vs non-MSI-high, CIMP-high vs low/negative and BRAF or KRAS mutated vs nonmutated. To test for differences related to subtype within the case-only analysis we used unconditional logistic regression. In the combined marker analysis, Type 4 (non-MSI-high, CIMP-low/negative, BRAF-wildtype, KRAS-wildtype) was used as a reference group in the case-only analysis, whereas in the polytomous analysis, cancer-free controls were used as the reference group. Both analyses used multinomial logistic regression to compare each molecular pathological subtype to the reference group. We also used multinomial logistic regression to estimate the association between diabetes and risk of colorectal cancer stratified by tumor location (colon, rectum, proximal colon, distal colon) and sex (male and female), and compared casecombinations of marker and tumor site with controls. For case-only analyses, we compared marker combinations stratified by tumor site.
Minimally adjusted models (presented in the Supplementary Materials) included study, age and sex as covariates, and fully adjusted models additionally included energy intake, family history of colorectal cancer, BMI, red and processed meat consumption, vegetable consumption, fruit consumption, alcohol use, smoking status, exercise and aspirin/NSAID use. Variables were first selected based on their theoretical relevance as potential confounders of an association between diabetes and colorectal   KRAS and CIMP-status), we considered a two sided P-value of <.05 to be significant (in both minimally and fully adjusted models). However, when testing associations between diabetes and combined marker subtypes, we used the alpha of 0.5% as recommended by Benjamin et al. 22 All analyses were performed using R version 4.0.0 (R Foundation for Statistical Computing, Vienna).

| RESULTS
The main characteristics of the study participants are described in Table 1. Individuals reporting a diabetes diagnosis were generally more likely than those not reporting diabetes to be male (P < .01) and to use aspirin regularly (P < .01), and less likely to have a reported family history of colorectal cancer, especially among colorectal cancer cases (P < .01). Participants with diabetes were also more often nondrinkers (P < .01), were more likely to exercise (P < .01), had a history of smoking (P < .01) and were more often obese (P < .01). These relationships were consistent among both colorectal cancer cases and controls.
In the case-control analysis, we observed an association Analyses were then further stratified by tumor subsite and sex (Table S6).
We observed significant differences between the association of diabetes and colorectal cancer by BRAF-mutation status for colon (P difference = .04) and proximal colon cancers (P difference = .02). The difference between BRAF-mutated and wild-type OR point estimates was also retained for rectal tumors (but not for distal tumors) and among both men and women, but without reaching significance (Table   S6). However, the number of BRAF-mutated tumors in the rectum was much lower than in the colon, which may explain why the observed difference in the association between diabetes and colorectal cancer risk by BRAF-mutation status was not statistically significant. The association between diabetes and colorectal cancer risk did not differ by other tumor markers in any of the stratified analyses.
Results from analyses of combined marker subtypes are presented in Figure 1 and Table S7. Among the 10 subtype combinations tested, diabetes was statistically significantly associated with risk of five different types (1, 3-5, and 9) in case-control analyses, but only type 4 remained significant after adjustments for multiple compari-

| DISCUSSION
This large collaborative effort is the first study to investigate the impact of diabetes on subtypes of colorectal cancer, while also considering both distinct molecular markers and tumor subtypes based on marker combinations. Our primary finding was a statistically significant difference in the strength of the association between reported diabetes status and colorectal cancer by BRAF status, with a stronger association for BRAF-mutated than nonmutated tumors. This was consistent in both our minimally and fully adjusted models. Reporting a diabetes diagnosis was more strongly associated with BRAF-mutated tumors in both the proximal colon and rectum indicating that tumor location does not explain this result.
BRAF mutations, found in 8% to 12% of colorectal cancers, are generally associated with a poor prognosis and are more common in proximal tumors. [23][24][25] BRAF is a serine-threonine kinase activated by KRAS as part of the MAPK signaling cascade. 26 Both BRAF and KRAS are oncogenes that are commonly mutated in colorectal cancer, and mutations in these genes are often considered mutually exclusive.
Previous studies have found some evidence that medical drugs can differentially affect the risk of different molecular subtypes. For example, a study from 2013 27 found that aspirin use (which has consistently been shown to decrease colorectal cancer risk 28 ) seemed to specifically lower the risk of BRAF-nonmutated colorectal cancer but not BRAF-mutated colorectal cancer. More related to diabetes, a study from 2012 29 found that metformin, an antidiabetic drug previously shown to have antitumor activity in several different cancers, 30 did not affect BRAF-mutated melanoma cells, but instead seemed to accelerate their growth in a xenograft mice model of melanoma. The authors suggested that the accelerated growth of the BRAF-mutated tumors could be attributed to improved angiogenesis. Although this is in line with our own results, where individuals reporting diabetes had a higher incidence of BRAF-mutated tumors compared to other molecular markers, later studies have been unable to replicate the findings. 31 In addition, at least one study 32 specifically examining the association between metformin and colon or colorectal cancer, but not by subtype, did not report any evidence of metformin acting differently depending on tumor site. It should also be noted that although metformin is commonly prescribed to individuals with type 2 diabetes mellitus, we lacked information about metformin use among our participants.
Colorectal cancer is known to often develop through a specific number of events along the so-called conventional pathway, which is a multistep process initiated by mutations in APC, KRAS or BRAF genes. It is also well known that other important pathways exist, such as the serrated pathway and the alternate pathway, which have other characteristics as well as distinct risk factors. 33 For example, the serrated pathway is associated with older age at onset, female sex and smoking and is also characterized primarily by BRAF mutations and CIMP-high status. 34,35 Previous studies originating from the GECCO consortia have investigated whether other established risk factors for colorectal cancer are associated with specific molecular subtypes, sometimes supporting development through distinct pathways. One recent study focused on dietary factors did show some evidence of heterogeneity related to fruit and especially fiber intake. 15 In polytomous analyses, higher fruit intake was associated with a decreased risk of developing BRAF-mutated tumors. High fiber intake, on the other hand, was associated with decreased risk of non-MSI-high, T A B L E 2 Associations between diabetes and risk of different molecular subtypes of colorectal cancer CIMP-low/negative, BRAF-wildtype and KRAS-wildtype subtypes, although none of these associations retained significance in case-only analyses. In another study, it was found that smoking was strongly associated with subtype combinations that included CIMP-high and MSI-high tumors. 36 As a result of these findings, the authors suggest that smoking might specifically increase the risk of tumors developing through the serrated pathway. Our study however, despite showing an increased risk of developing BRAF-mutated tumors, did not find any evidence of diabetes resulting in any pathway specific development. Both previous studies and ours underline the importance of using large enough sample sets to be able to adequately assess patterns of associations that differ depending on molecular subtype.
Metabolic dysfunction, often defined as the presence of three or more of the criteria for metabolic syndrome (central obesity, hypertension, dysglycemia and dyslipidemia), is associated with an increased risk of developing both diabetes (type 2) and colorectal cancer. Several studies have aimed to assess the relationship between metabolic abnormalities (such as high BMI, hypertension and dysglycemia) and colorectal cancer, focusing on metabolic syndrome, inflammation and specific colorectal cancer subtypes. [37][38][39] One of several plausible links between diabetes and colorectal cancer could be related to insulin resistance and hyperinsulinemia, a state of heightened insulin levels and a hallmark of untreated type 2 diabetes in its earlier natural history. High insulin levels have been linked to an increased risk of multiple cancers, including colorectal cancer. 40 This may relate to its growth promoting effects as well as its ability to increase circulating IGF1 levels. 41 Several studies have also identified a link between markers of heightened insulin levels (eg, C-peptide) and colorectal cancer risk, 42,43 but without finding any evidence of heterogeneity. 44 However, the most important metabolic factor likely to affect both diabetes and colorectal cancer risk is obesity and the relationship between BMI and colorectal cancer has also shown consistent directions of associations across studies, although just as for hyperglycemia, evidence of heterogeneity between subtypes has been somewhat inconsistent. 11,[45][46][47][48][49][50][51] In the current study however, we find some evidence of diabetes increasing the risk of specifically BRAF-mutated tumors, but taking into account previous studies, this subtype-specific risk difference is probably not related to the shared metabolic risk factors connecting diabetes and colorectal cancer.
An important strength of our study is the ability to pool data from multiple observational studies with readily available information on diabetes status. This pooling of datasets enabled us to detect even modestly differential associations between diabetes status and risk of colorectal cancer by molecular subtypes. However, there are also limitations that have to be taken into account when interpreting the study results. First, we were not able to distinguish between type 1 and type 2 diabetes, or to take into account use of specific medications such as metformin. Time and duration of diabetes diagnosis was also lacking. These are all factors that can have important implications and result in misclassification bias. However, the size of our study, and the fact that type 2 diabetes makes up more than 90% of all diabetes cases, 52 makes it less likely that these have substantially distorted our results. There may also be selection bias related to the cases included in the pooled analysis or to tissue availability potentially depending on tumor stage and size, 53 a limitation that has been previously described. 15 Finally, we did not apply any formal adjustment for multiple comparisons to our primary analyses. Although adding such adjustments would not affect the associations of risk, as they are all highly significant, the reported difference in OR estimates between BRAF-mutated and nonmutated tumors would fall slightly below the significance threshold (accounting for tests of four different subtypes). This should be considered when interpreting the findings. It is worth noting that individuals reporting diabetes had a higher risk of all subtypes of colorectal cancer and the difference related to BRAFmutated tumors only affected the magnitude of the risk, but the direction remained the same.
In summary, our study confirms the established association between diabetes and colorectal cancer risk and suggests that it espe-