Our systematic MR analysis of 14 previously reported molecular risk factors and BMI in 12,906 endometrial cancer cases and 108,979 controls provided evidence for roles of elevated BMI, fasting insulin, total and bioavailable testosterone and SHBG in risk of overall and endometrioid endometrial cancer. In mediation analyses, we found evidence that fasting insulin, bioavailable testosterone concentrations and SHBG partially mediated the effect of BMI on overall endometrial cancer risk. When combining pairs of mediators together into a single model, we found evidence that an effect of fasting insulin on endometrial cancer was partially mediated by SHBG levels and that an effect of SHBG on endometrial cancer was largely mediated by bioavailable testosterone levels. An effect of fasting insulin on endometrial cancer risk was also strongly attenuated upon adjustment for bioavailable testosterone levels which could reflect mediation of this effect by bioavailable testosterone or conditionally weak instrument bias for fasting insulin concentrations in this analysis. Our analyses found little evidence that several previously reported molecular risk factors, including several metabolic factors (e.g. LDL cholesterol, HDL cholesterol, IGF-1, adiponectin, leptin) and inflammatory markers (CRP, IL-6), were causally implicated in overall or endometrioid endometrial cancer risk.
Several of the findings in this analysis are consistent with evidence from prior conventional observational and MR analyses. For example, the effect of BMI on endometrial cancer risk and the stronger evidence of an effect on endometrioid, as compared to non-endometrioid, endometrial cancer is well-established in the literature. Additionally, this has been shown previously in an MR analysis that used an alternative strategy for instrument construction to our own [74]. Our findings supporting a causal effect of BMI on endometrial cancer risk (OR 1.88, 95% CI 1.69 to 2.09 per SD (4.7 kg/m2) increase) are larger in magnitude than those from pooled analyses of conventional observational analyses (e.g. the World Cancer Research Fund (WCRF) pooled analysis of 26 prospective studies: relative risk (RR) per 5.0 kg/m2 increase 1.50, 95% CI 1.42 to 1.59), consistent with previous comparisons of observational and MR estimates across other cancer sites [75, 76]. Smaller magnitudes of effect in observational analyses may reflect regression dilution bias from single time-point measurements of BMI and/or reverse causation from cancer-induced weight loss, whereas MR estimates reflect accumulated exposure across the life-course and are unlikely to be influenced by reverse causation [77].
In agreement with previous MR analyses, our results suggest a causal role of fasting insulin, total and bioavailable testosterone and SHBG in endometrial cancer risk, although these previous reports either employed smaller sample sizes than this analysis (e.g. fasting insulin analyses were performed in 1287 endometrial cancer cases vs 12,906 cases in our analysis) or used somewhat differing methods to examine instrumental variable assumptions [27,28,29]. The restriction of an effect of BMI to bioavailable (and not total) testosterone is in agreement with previous observational studies which have suggested that BMI influences testosterone levels through decreased production of SHBG rather than a direct effect on testosterone production [78,79,80,81,82]. Additionally, important mediating roles of fasting insulin, bioavailable testosterone and SHBG in the relationship between BMI and endometrial cancer are consistent with studies of bariatric surgery which have suggested protective effects of this procedure against endometrial cancer risk, along with reductions in insulin and bioavailable testosterone levels, and increases in SHBG levels [83,84,85,86,87,88,89,90,91]. Our findings supporting a role of BMI on these traits are also consistent with the important endocrine function of adipose tissue, which is involved in sex steroid metabolism [80, 92,93,94,95,96,97].
Potential aetiological roles of the molecular mediators identified in this analysis are consistent with the “unopposed oestrogen” hypothesis which postulates that endometrial carcinogenesis is driven by excess endogenous or exogenous oestrogen levels that are unopposed by progesterone [98,99,100]. We were unable to incorporate oestrogen into this analysis as we were unable to identify reliable genetic instruments for this trait. All three of the molecular mediators highlighted in this analysis, however, are known to influence oestrogen: bioavailable testosterone is aromatized to oestradiol; SHBG binds with high-affinity to both oestradiol and bioavailable testosterone [100,101,102,103,104,105]; and insulin increases androgen and decreases SHBG production [106,107,108,109]. We found the inverse effect of SHBG on endometrial cancer risk was largely attenuated upon adjustment for bioavailable testosterone, suggesting a protective effect of SHBG may be driven via binding of biologically active fractions of circulating testosterone. The attenuation of an effect of fasting insulin on endometrial cancer upon adjustment for bioavailable testosterone could reflect mediation of this effect or the presence of conditionally weak instrument bias in this model. In support of the latter explanation, there is biological evidence that hyperinsulinemia and insulin resistance influence endometrial cancer via oestrogen-independent pathways. For example, insulin has been shown to bind directly to endometrial cells and promote proliferation, and can activate two pathways known to have an important role in carcinogenesis—the phosphatidylinositol-3-kinase-protein kinase B/Akt (PI3K-PKB/Akt) and Ras/Raf/mitogen-activated protein kinase (Ras/Raf/MAPK) pathways [109,110,111,112,113,114].
Some findings from this MR analysis differ from those of prior conventional observational studies. For example, our analyses found little evidence to support causal roles of several metabolic traits (e.g. circulating HDL cholesterol, triglycerides, adiponectin, leptin) and inflammatory markers (CRP, IL-6) in endometrial cancer risk, despite these traits being linked to endometrial cancer risk in conventional observational analyses [18,19,20,21,22]. Several of these traits (e.g. HDL cholesterol, LDL cholesterol, triglycerides) represent highly correlated metabolic perturbations associated with the obese phenotype which may be too clustered to disentangle using conventional multivariable regression methods [115]. Consequently, some of the divergence in findings across previous conventional observational studies and this MR analysis could reflect residual confounding in the former. Another potential explanation for divergence in findings is the susceptibility of conventional observational studies to reverse causation (i.e. latent, undiagnosed endometrial cancer influencing levels of a presumed exposure). For example, a previously reported association of circulating IL-6 concentrations with endometrial cancer risk could reflect IL-6 secretion by endometrial cancer-associated fibroblasts rather than a role of IL-6 in endometrial cancer development [116, 117]. Similarly, reverse causation could explain the previously reported associations between CRP, a nonspecific indicator of inflammation, and endometrial cancer risk, as early stages of endometrial carcinogenesis may induce an inflammatory response, leading to elevated levels of CRP [118, 119].
We were unable to replicate a previously reported MR-based inverse association of LDL cholesterol levels and endometrial cancer risk in the ECAC (IVW OR per SD increase in LDL cholesterol 0.90, 95% CI 0.85 to 0.95, P = 8.39 × 10−5). In the previous analysis, SNPs were permitted to be in weak LD (pairwise correlation r2 < 0.05 vs r2 < 0.001 in our analysis) and a Heterogeneity in Dependent Instruments (HEIDI) test was performed to identify potentially pleiotropic SNPs, resulting in the removal of 6 such SNPs from the 146 SNPs initially used as an instrument. We attempted to replicate these previously reported findings using a more stringent r2 threshold (i.e. r2 < 0.001) followed by use of the HEIDI test (resulting in the removal of 2 potentially pleiotropic SNPs) which resulted in a causal estimate that was closer in magnitude to that previously reported (IVW OR 0.93, 95% CI 0.86 to 1.00, P = 4.10 × 10−2) (Additional file 1: Table S48). However, there was greater imprecision in our estimate compared to this previous analysis which could reflect the more liberal LD threshold employed in the earlier analysis.
Our MR analysis provides key insights into potential molecular pathways linking excess adiposity to endometrial cancer risk. This analysis has several strengths including the use of a systematic approach to collate previously reported molecular risk factors for endometrial cancer; the appraisal of their causal relevance in overall and endometrioid endometrial cancer aetiology using an MR framework which should be less prone to conventional issues of confounding and cannot be influenced by reverse causation; the employment of several complementary sensitivity analyses to rigorously assess for violations of MR assumptions; and the use of a summary data-based MR approach which permitted us to leverage large-scale GWAS data from several studies, enhancing statistical power and precision of causal estimates.
There are several limitations to our analysis. First, we were unable to evaluate the role of six previously reported molecular risk factors for endometrial cancer due to the absence of reliable genetic instruments for these traits. These risk factors included oestradiol which is believed to be an important molecular mediator of the effect of BMI on endometrial cancer risk [9]. Second, some of the effect estimates for SNPs included in genetic instruments were obtained from discovery GWAS and have not been replicated in an independent sample which can result in “Winner’s curse” bias. There was sample overlap in this analysis across certain traits. However, the use of conventionally strong (P < 5.0 × 10−8) instruments for these traits and general consistency of results across sensitivity analyses examining their robustness to potential Winner’s curse bias suggests that this phenomenon was unlikely to have substantial influence in this analysis. Third, although sex-specific sensitivity analyses were performed where data were available, some prior GWAS used in this analysis did not examine for heterogeneity of SNP effects by sex which prevented evaluation of the effect of certain traits on endometrial cancer risk using sex-specific instruments. Fourth, univariable and multivariable MR analyses presented here assume that relationships between exposures and outcomes are linear, although it has been previously suggested that the relationship between BMI and endometrial cancer may best be explained by a non-linear model [12, 120]. Multivariable MR additionally assumes no exposure-mediator interaction. While methods exist to examine interaction in an individual-level setting, these do not currently exist for analyses using summary-level data [121]. Fifth, our analysis was almost exclusively restricted to individuals of European ancestry to minimize bias from population stratification, which may limit the generalizability of our findings to non-European populations. Sixth, we only investigated a single measure of adiposity (i.e. BMI) in our analyses. Though widely used as a measure of overall adiposity, BMI may fail to capture the independent contribution of central adiposity and/or body fat distribution on endometrial cancer risk. Seventh, our use of two-sample MR with summary data precluded performing subgroup analysis and assessment of potential effect modification. Eighth, one instrument in our primary analysis (i.e. leptin) and one in a cis-variant-specific sensitivity analysis (i.e. IGF-1) consisted of a single SNP. While we found little evidence of association of these traits with endometrial cancer risk, we were unable to employ various “pleiotropy-robust” models to evaluate exclusion restriction assumptions and therefore cannot rule out the possibility of horizontal pleiotropy biasing causal estimates obtained toward the null. Finally, while various sensitivity analyses were performed to examine violations of exchangeability and exclusion restriction criteria, these assumptions are unverifiable.
With the global incidence of overweight and obesity projected to increase and challenges in implementing successful weight loss strategies, a greater understanding of the molecular mechanisms by which obesity increases risk of disease, including endometrial cancer, is vital [122,123,124,125,126]. Type 2 diabetes and obesity are highly comorbid, with 75% of adults in the UK who have received a diabetes diagnosis being prescribed some form of anti-diabetic medication [127]. Our findings suggest that use of such medications may confer a favourable secondary effect of reducing endometrial cancer risk among these high-risk groups. Among various approved anti-diabetic medications, metformin in particular could plausibly offer the most pronounced endometrial cancer risk-reducing effect as it has been shown to not only increase insulin sensitivity, thus reversing insulin resistance and lowering fasting insulin levels, but also inhibit endometrial proliferation [9, 128]. In addition, unlike some other oral hypoglycaemic medications, metformin users show a tendency toward sustained weight loss [129]. Bioavailable testosterone and SHBG also present potential pharmacological targets, though the multifaceted function of these hormones means that targeting these traits may result in adverse effects [130,131,132,133,134,135]. Phase II clinical trials examining the efficacy of a combination of contraceptive intrauterine devices, metformin and weight loss interventions as a non-invasive treatment option for individuals with obesity with early-stage endometrial cancer have had encouraging results [136]. Additionally, weight loss has been shown to improve oncological outcomes in women with endometrial cancer undergoing progestin treatment [137].
Future studies should aim to “triangulate” these findings using alternate epidemiological study designs with orthogonal (i.e. non-overlapping) sources of bias, for instance using directly measured insulin, SHBG and bioavailable testosterone in a large-scale cohort study, such as UK Biobank [138]. Another possible future direction for this work is to explore the effects of excess adiposity at different life stages, for instance, comparing pre- and post-menopausal BMI, in order to evaluate any potentially independent effects of excess adiposity on endometrial cancer risk across the life-course.
Our systematic evaluation of 14 previously reported candidate mediators of the effect of BMI on endometrial cancer risk identifies fasting insulin, bioavailable testosterone and SHBG as plausible mediators of this relationship. While we were unable to entirely disentangle the independent effects of these three traits, identification of a potential mediating role of these traits (and, in particular, fasting insulin) in endometrial carcinogenesis is nonetheless informative for the development of pharmacological interventions targeting these traits for cancer prevention. In this respect, future assessment of the effect of drugs which target molecular mediators identified in this analysis using a “drug-target Mendelian randomization” approach could inform on the potential efficacy of the repurposing of medications for endometrial cancer prevention.