Study design

The current study was reported according to the STROBE-MR statement [24]. Two-sample MR and SMR analyses were applied to reveal genetic relationships between glucose-lowering drug targets and gastrointestinal cancer risks. The validation of MR analyses’ causal estimation relied on three crucial assumptions: (i) the genetic IVs were closely connected with ten glucose-lowering drug targets; (ii) the independence of genetic IVs from confounding factors; and (iii) no direct effects of genetic IVs on gastrointestinal cancer risk other than through the glucose-lowering drug targets [25]. The random-effect inverse-variance weighted (IVW) method was the primary approach to elucidate the association between glucose-lowering drug targets and gastrointestinal cancer risks in the two-sample MR analyses [26]. We have also integrated validation cohorts for external corroboration of the pertinent results and employed meta-analysis to synthesize the findings from discovery and validation cohorts, guaranteeing the dependability and universality of our research outcomes. Similarly, we utilized a tool (https://shiny.cnsgenomics.com/mRnd/ ) to independently calculate the statistical power of the related analyses in both the discovery and validation cohorts, thereby assuring the efficacy of the analyses. Meanwhile, we conducted SMR analysis, which leverages expression quantitative trait loci (eQTLs) as instruments, to explore and validate the causal relationships between exposures and outcomes at the gene expression level [15]. Additionally, colocalization analysis, employing Bayes factor computation, was performed within a window of ± 500 KB around the gene encoding of each independent glucose-lowering drug target to calculate the posterior probabilities of connection between exposures and outcomes [27]. Furthermore, the mediation effects of some confounding factors (including body mass index (BMI), glucose measurement, and T2DM) and gastrointestinal cancer risks were uncovered through the two-step MR analyses to examine whether the observed relationship was direct. Figure 1 shows the detailed process of this study.

Fig. 1

The flowchart of this study. Abbreviations HEIDI: heterogeneity in dependent instruments; MR-PRESSO: Mendelian Randomization Pleiotropy RESidual Sum and Outlier

All summary data utilized in this study had been approved by the relevant institutional review board of each country on the basis of the Declaration of Helsinki, and all participants involved in these studies had signed the informed consent forms. Separate ethical approval was not required for this study.

Extraction and selection of instrumental variables

Ten targets of seven different glucose-lowering drugs were eventually involved. Thiazolidinediones (TZDs) target PPARG (peroxisome proliferator-activated receptor gamma) precisely, activating it to increase tissue sensitivity to insulin and effectively lower blood sugar levels [28]. Dipeptidyl Peptidase IV Inhibitors, also known as DPP4 inhibitors, act on DPP4, reducing its activity to prevent the breakdown of incretins like GLP-1 [29]. This increase in GLP-1 levels encourages insulin secretion from pancreatic β-cells, helping to reduce blood sugar. Similarly, Glucagon-like Peptide-1 Analogues focus on GLP1R, boosting insulin secretion, hindering glycogenolysis, slowing gastric emptying, and increasing satiety [30]. Concurrently, Insulin and Insulin Analogues, which target INSR, mimic the effects of natural insulin, facilitating glucose uptake and utilization, thereby decreasing blood glucose levels [31]. Sodium-Glucose Cotransporter Inhibitors, or SGLT2 inhibitors, target SLC5A2 and block SGLT2 in the renal tubules, reducing glucose reabsorption and increasing urinary glucose excretion, leading to lower blood glucose levels [32]. Sulfonylureas, targeting ABCC8 and KCNJ11, close ATP-sensitive potassium channels, causing an increase in intracellular calcium and subsequent insulin release from β-cells [33]. Finally, Metformin acts on ETFDH, GPD2, and PRKAB1, activating AMPK, which reduces hepatic glycogenolysis and gluconeogenesis and enhances insulin sensitivity, collectively contributing to the reduction of blood glucose levels [34]. Each of these medications plays a unique and interconnected role in managing blood sugar levels through various mechanisms. Corresponding coding gene targets and pharmacologically active protein targets were searched from ChEMBL (https://www.ebi.ac.uk/chembl) and DrugBank (https://www.drugbank.ca) databases (Supplementary Table S1) [35, 36]. However, we also recognize that other significant genetic variations apart from these known targets might be related to glucose metabolism. To comprehensively solve this problem, we adopted a two-step strategy: (i) broad screening: during the GWAS data selection process, we employed an unbiased strategy to identify all potential genetic variations related to glucose metabolism. (ii) in-depth analysis: for those variations that showed a significant correlation with glucose metabolism in the preliminary screening, we conducted a more comprehensive and in-depth research search and review to ascertain whether they had been reported in previous studies and the potential biological roles they might play in disease [28,29,30,31,32,33,34]. Furthermore, we conducted a comprehensive search of existing GWAS databases. After excluding glucose-lowering drug targets for which adequate data could not be obtained, we incorporated ten targets of seven glucose-lowering drugs to the maximum extent possible, ensuring the comprehensiveness of our study.

For obtaining more effective IVs, the glucose-lowering drug targets were proxied by IVs collected from the summary genetic association data of HbA1c measurement (N = 389,889) in the two-sample MR analyses (Supplementary Table S2) [37]. IVs were constructed within a range of ± 500 KB around the gene encodings for each of the ten glucose-lowering drug targets based on a significance threshold of P-value < 5 × 10− 8. Any IVs in high linkage disequilibrium (LD) with each other (r2 ≥ 0.01) were removed. Genetically proxied perturbation of per standard deviation unit (SD) change in glucose-lowering drug targets were scaled to represent an SD unit of HbA1c reduction.

In SMR analyses, the common (minor allele frequency (MAF) > 1%) eQTLs were identified from eQTLGen Consortium (https://www.eqtlgen.org/) and GTEx Consortium Version 8.0 (https://gtexportal.org/) as significant (P-value < 5 × 10− 8) IVs associated with the expression of DPP4, GPD2, ETFDH, GLP1R, INSR, KCNJ11, PPARG, PRKAB1, and SLC5A2 in blood tissue. Since no significant eQTLs about ABCC8 were found in blood tissue, we collected IVs related to the expression of ABCC8, specifically in muscle and skeletal tissues. Low weak linkage disequilibrium (r2 < 0.01) IVs were selected within the ± 500 KB windows of gene encodings for proxying glucose-lowering drug targets to ensure the high strength of the instruments.

Validation of instrumental variables

The F-statistic method was hired to avoid potential weak IV bias with the criterion of F-value > 10 [38]. T2DM is the original indication of glucose-lowering drugs, and these drugs will ultimately affect the patient’s blood sugar levels [39]. A former meta-analysis indicated that some glucose-lowering drugs contribute to weight gains, such as sulfonylureas, insulin analogues, and thiazolidinediones, whereas GLP-1 analogues cause weight loss, which manifested body weight was another distinct phenotype affected by glucose-lowering drugs [40]. Therefore, BMI (N = 461,460), glucose measurement (N = 400,458), and T2DM (N = 655,666) selected from extensive summary data were utilized as positive controls to validate the strong association of IVs and genetically proxied glucose-lowering drug target perturbation [41, 42].

Outcome data source for gastrointestinal cancer

Seven types of gastrointestinal cancer were selected from three large European ancestry databases (analyzed by Longda Jiang et al., Saori Sakaue et al., and Joshua D Backman et al. based on the UK Biobank and Finnland cohorts, respectively) as the discovery databases, including anal carcinoma (N = 456,348), cardia cancer (N = 456,348), gastric cancer (N = 456,348), HCC (N = 456,276), intrahepatic cholangiocarcinoma (ICC) (N = 456,348), pancreatic cancer (N = 635,945), and rectum cancer (N = 387,797) (Supplementary Table S2) [43,44,45]. In addition, for the purpose of validation, we have engaged additional cohorts from Finland (DF10, Public release: Dec 18, 2023) and other European populations (such as Kaiser Permanente Genetic Epidemiology Research on Adult Health and Aging and UK Biobank cohorts), as analyzed by Sara R Rashkin et al., Saori Sakaue et al., and Longda Jiang et al., respectively [43, 44, 46, 47]. These validation cohorts encompass anal carcinoma (N = 314,291), cardia cancer (N = 411,441), gastric cancer (N = 476,116), HCC (N = 314,693), ICC (N = 315,400), pancreatic cancer (N = 456,276), and rectal cancer (N = 316,683), with further details provided in Supplementary Table S2. Patients with gastrointestinal cancer were clinically or pathologically diagnosed based on the National Comprehensive Cancer Network standards. For no participants overlapping in the exposures and outcomes data, type I error was well avoided to ensure the strength of the MR analyses [48].

Colocalization analysis

To scrutinize the alignment between the exposures and outcomes (specifically, the antihyperglycemic drug target and the susceptibility to gastrointestinal cancer) and to distinguish any confounding arising from linkage disequilibrium potentially ascribed to a shared causal variant, we employed colocalization analysis. This method leverages the computation of approximate Bayes factors to yield posterior probabilities, facilitating a more sophisticated interpretation of the interrelationships involved [27]. The colocalization analysis encompassed five hypotheses: (i) H0: neither the glucose-lowering drug targets nor gastrointestinal cancer possessed a causal variant within the genomic locus; (ii) H1: only the glucose-lowering drug targets harbored a causal variant; (iii) H2: only gastrointestinal cancer had a causal variant; (iv) H3: each of the glucose-lowering drug targets and gastrointestinal cancer had distinct causal variants; (v) H4: a shared causal variant was present for both glucose-lowering drug targets and gastrointestinal cancer [23]. To facilitate a thorough exploration of the genomic terrain encircling these critical regions, the colocalization analysis was executed by generating ± 500 kb windows surrounding the gene responsible for encoding each respective glucose-lowering drug target [27]. We utilized default parameters to conduct the colocalization, setting p1 = 1 × 10− 4 (the prior probability that a SNP is linked with the glucose-lowering drug target), p2 = 1 × 10− 4 (the prior probability that a SNP is linked with gastrointestinal cancer), and p12 = 1 × 10− 5 (the prior probability that a SNP is concurrently linked with both the glucose-lowering drug target and gastrointestinal cancer) [49]. A posterior probability for H4 (PP4) surpassing 0.8, under a variety of priors and windows, was construed as compelling evidence, signifying colocalization. The “coloc” (Version 5.2.1) and “LocusCompareR” (Version 1.0.0) packages were harnessed in the colocalization analysis to compute and graphically represent the outcomes.

Sensitive analysis

In the two-sample MR analyses, the IVW method offers an unbiased estimate of causality, provided that all IVs are valid and free from pleiotropy [50]. To assess the heterogeneity of IVs and detect any pleiotropic effects, we conducted the Cochran’s Q test, with a significance threshold set at a P-value for heterogeneity < 0.05 [50]. As for potential horizontal pleiotropy, we performed additional analyses including weighted median and MR-Egger. the MR-Egger method allows for robust causal inference even when all IVs are potentially invalid and indicates the presence of unbalanced pleiotropy when the P-value for the intercept is < 0.05 [51]. Additionally, we applied the MR Pleiotropy RESidual Sum and Outlier (MR-PRESSO) global test to identify and adjust for outliers. Furthermore, in the SMR analysis, we employed the heterogeneity in dependent instruments (HEIDI) test, leveraging multiple SNPs within a genomic locus, to distinguish between associations of glucose-lowering drug targets with gastrointestinal cancer risk that are attributable to a shared genetic variant as opposed to genetic linkage [52]. A HEIDI test with a P-value > 0.01 indicated that the observed association between glucose-lowering drug targets and gastrointestinal cancer risk was not confounded by linkage disequilibrium [53].

Statistical analysis

All analyses were performed using R software (Version 4.2.3) and SMR software (Version 1.3.1) [15]. R packages, including “TwoSampleMR” (Version 0.5.6) and “MR-PRESSO” (Version 1.0) were utilized. The Bonferroni-corrected significance level of P-value < 0.0007 (0.05/70, ten glucose-lowering drug targets and seven types of gastrointestinal cancer) was utilized to avert bias [54]. Associations with a P-value between 0.0007 and 0.05 was considered suggestive, while a P-value > 0.05 indicated no statistical association between glucose-lowering drug targets and gastrointestinal cancer.