Mendelian randomization studies of gastric cancer reveal potential risk factors, promising biomarkers and therapeutic targets.

Introduction
Gastric cancer is the third leading cause of cancer-related death, and although the availability of surgery, chemotherapy and targeted therapies has led to better outcomes, cancer recurrence or metastasis remains a major cause of mortality, with more than 1 million new cases are diagnosed annually and still pose a threat to public health [1, 2]. Therefore, it is essential for researchers to identify potential risk factors and new therapeutic targets for the prevention and treatment of gastric cancer. Given the high cost and long duration of clinical trials and drug development, it is particularly important to explore potential risk factors or biomarkers related to gastric carcinogenesis and progression before clinical trials.
In recent years, genome-wide association studies (GWAS) have achieved great success in revealing the genetic determinants between human diseases and phenotypic traits, and Mendelian randomization (MR) analysis can effectively infer their causal associations by integrating genomic data [3]. MR has been successfully applied to several human diseases to identify potential risk factors and drug therapeutic targets [4–6]. However, it has not been fully explored in gastric cancer and has only been reported in a few related traits, such as body mass index [7], type 2 diabetes [8, 9], and plasma homocysteine levels [10]. Therefore, in this study, we examined approximately 486 traits in 24 subcategories through ordinary MR, reverse MR, and mediating effect analyses to systematically evaluate potential risk factors. This will increase our underlying knowledge of gastric cancer occurrence and help to develop new drug therapeutic targets.

Methods and materials

Study design and data source
As shown in the flow chart of the study design (Fig. 1), on June 16th, 2023, the gastric cancer-related GWAS data was obtained from the BioBank Japan Project (BBJP) [11], including 6563 cases and 195,745 controls. On the MR-Base database [12], a comprehensive selection of potential exposure factors for approximately 42,335 traits was performed to identify potential risk factors and biomarkers. Finally, the obtained 486 traits were divided into 24 subcategories (e.g., amino acid, anthropometric, and behavioral) (Table S1). Additional information, such as the population, sample size, and unit for each trait, is available in the supplementary data.

Genetic instrument selection
To support a causal relationship between potential risk factors and gastric cancer, three hypotheses and specific conditions should be followed. First, the selection of genetic instruments must be related to gastric cancer. A genome-wide significance threshold of P < 5e-8 and genetic instruments that explaining r2 < 0.001 of linkage disequilibrium were used in the MR analysis. Second, the genetic instruments should not be associated with other confounders, including gastric cancer. A harmonization function was applied to each SNP to ensure the same alleles on exposure and outcome. In this step, palindromic SNPs with intermediate allele frequencies were dropped from the analysis. Third, the genetic instruments should not affect the outcome unless it is possible to do so through association with the exposure.
To reduce deviation caused by weak genetic instruments, F-statistics less than 10 were considered weak instruments and were removed from the MR analyses. The variance of genetic instruments that explained variation in exposure was calculated by the formula , and the F-statistics was calculated by the formula , where MAF represents the minor allele frequency of genetic instruments, k represents the number of genetic instruments used, and N represents the total sample size.

Potential risk factors identification
Potential risk factors used as exposures in the MR analysis were restricted to those with genetic instruments SNPs > 2. To avoid false-positive causal associations, according to Egger’s regression model [13], risk factors with P < 0.05 in the pleiotropy test or heterogeneity test, and genetic instruments explaining R2 < 0.1 of variation in exposure, were excluded from the MR results. Then, genetic instruments were further checked among the screened potential risk factors to remove duplicated SNPs that violated the third hypothesis. Directional pleiotropy was considered when P < 0.1 for MR-Egger’s intercept [14]. Finally, bidirectional MR and mediating effect analyses were used to detect reversed causality and the proportion of mediating effects, respectively [15, 16]. The statistical power of MR results was calculated using the online tool mRnd (http://cnsgenomics.com/shiny/mRnd/).
To determine the relationship between gene expression of potential biomarkers and the prognosis of patients with gastric cancer, the KMplot [17], a web-based Cox proportional hazards survival analysis with a log-rank test, was searched in repositories such as The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus. Furthermore, to assess the on-target side effects of promising therapeutic biomarkers, after screening for potential risk factors, we performed an extensive phenotypic MR analysis for 1,403 diseases from PheWeb [18] based on the International Classification of Diseases 10th edition (ICD10). The same proxies of potential risk factors used in gastric cancer were used to analyze causality to improve the interpretability of the results.

Mediation effect calculation
The researchers also evaluated the potential mediators between exposure and gastric cancer using a linear regression model:
Where Y, X, and M represent the outcome, exposure, and mediator, respectively. The direct effect of X on Y was defined by c, and the indirect effect of M on Y was defined by a*b. The proportion of the mediation effect was defined by a*b/(c + a*b).

Drug sensitivity analysis and molecular docking
The drug chemotherapeutic response for each patient with stomach cancer in the TCGA database was predicted based on a pre-trained model through the R package pRRophetic [19]. The half-maximal inhibitory concentration (IC50) assessed by the ridge regression model of each sample was grouped by median gene expression level and compared by Wilcoxon test. In addition, potential chemotherapeutic drugs were searched in the Comparative Toxicogenomics Database (CTD) [20] and had to be supported by interactions with target genes and stomach cancer.
Finally, molecular docking between potential drugs and target genes was performed to evaluate their binding ability and physicochemical properties based on the Lipinski’s rule of Five [21]. The primary molecular docking and visualization were conducted using the Molecular Operating Environment (MOE; 2022.02) software. The affinity score was evaluated through London G scoring and Generalized Born/Volume Integral Weighted by the exposure of Surface Area G scoring. In each docking process, 10,000 poses were generated for placement using the “Triangle Matcher” method, and 500 iterations were made for refinement using the “Induced Fit” method.

Statistical analyses
In this MR analysis, Egger’s regression, inverse-variance weighted (IVW) meta-analyses, and weighted median approaches were mainly used to evaluate small study bias, horizontal or directional pleiotropy, and sensitivity analysis. The confirmation of causality was based on results from at least two methods with P < 0.05. Besides, the results of the above three methods were pooled using the IVW method of a fixed model to ensure the accuracy of genetic instruments and the role of risk factors for gastric cancer. If any of the three methods crossed the reference line, then a suggestive causality was considered. Another sensitivity analysis was conducted by radial MR-Egger’s model, which can detect significant outliers.
The strength of causality for potential risk factors was ranked according to the P-score, which is defined as the mean of all 1 − Pj where Pj, where Pj denotes the one-sided P-value of accepting the alternative hypothesis that risk factor i is worse than one of the competing factors j. Thus, if risk factor i is worse than other risk factors, many of these P-values will be small, and the P-score will be large, and vice versa. A directed acyclic graph was established based on the P-scores of the three methods to evaluate the strength of potential risk factors for gastric cancer. All analyses were performed using `TwoSampleMR` and related packages in R-4.1.3 software (R Core Team, 2022, R Foundation for Statistical Computing, Vienna, Austria, https://www.R-project.org). The reports of molecular docking were generated in the Molecular Operating Environment (MOE; 2022.02, Chemical Computing Group ULC, Montreal, QC, Canada).

Results

Potential risk factors for gastric cancer
Finally, of all 486 traits, 14 possible risk factors (Table 1) were found to be associated with the risk of gastric cancer. There was no heterogeneity (P.het > 0.05) and horizontal pleiotropy effect (P.int > 0.05) in any of them. The F-statistical values ranged from 39.49 to 398727.50, indicating that there were no weak genetic instruments. Nine of them are plasma proteins from an INTERVAL study of the human plasma proteome in healthy blood donors [22]. Other possible risk factors include “androsterone sulfate”, “ulcerative colitis”, “years of schooling”, “type 2 diabetes”, and “ultradistal forearm bone mineral density”. Details of these possible risk factors, as well as other suggestive and removed traits, are provided in the supplementary data.
However, two SNPs including rs2519093 and rs704, were duplicated across five risk factors, which violated the third MR hypotheses. Three outliers including rs72999033, rs10811660 and rs35261542 were found in risk factor “type 2 diabetes” (Fig. S1). Therefore, after the removal of the duplicated or outlying SNPs, MR analyses were carried out again for these potential risk factors. Subsequently, four risk factors, including “zinc fingers and homeoboxes protein 3”, “estrogen sulfotransferase”, “adhesion G protein-coupled receptor F5” and “P-selectin” were removed according to the genetic instruments selection criteria (Table 1).

Causal effect of potential risk factors
The results of the pooled inverse-variance test showed that four potential risk factors positively associated with gastric cancer (Fig. 2A), including ultradistal forearm bone mineral density (OR = 1.30, 95% CI: 1.15–1.48), years of schooling (OR per “SD = 3.71y” increase: 3.64, 95% CI: 1.97–6.72), ulcerative colitis (OR = 1.15, 95% CI: 1.10–1.19), and alpha-2-macroglobulin receptor-associated protein (OR = 1.05, 95% CI: 1.02–1.09).
On the other hand, gastric cancer was inversely association with six potential risk factors (Fig. 2B), including type 2 diabetes (OR = 0.86, 95% CI: 0.80–0.92), androsterone sulfate (OR per log10 unit (0.37) increase: 0.54, 95% CI: 0.36–0.81), vascular endothelial growth factor receptor 2 (VEGFR2, OR = 0.86, 95% CI: 0.82–0.90), promotilin (OR = 0.82, 95% CI: 0.77–0.87), neutrophil collagenase (OR = 0.91, 95% CI: 0.88–0.94), and tyrosine-protein kinase receptor Tie-1 (soluble) (OR = 0.91, 95% CI: 0.88–0.94).

Strength of effect for potential risk factors
To assess the strength of the effect of potential risk factors for gastric cancer, a P-score ranking was computed based on the estimated size and standard error of MR results. All P-scores from the three MR methods for the ten potential risk factors were mapped on a heat-map (Fig. 3A). A directed acyclic graph was constructed based on the P-scores to vividly identify the ranking of risk factors, with the same levels denoting similar strength. As illustrated in Fig. 3B, factors such as “years of schooling” and “ultradistal forearm bone mineral density” received the highest priority, followed by “androsterone sulfate”, “type 2 diabetes”, “promotilin”, “ulcerative colitis”, and “VEGFR2”.

Mediating effect of potential risk factors
Due to the existence of directional pleiotropy [14] effects within the results of MR analyses, a mediation test was conducted using structural equation modeling[16] among the ten potential risk factors and gastric cancer. A total of = 90 possible mediating pathways for gastric cancer were examined, of which only four showed significant effects (Table S2), including “androsterone sulfate -> type 2 diabetes”, “ulcerative colitis -> type 2 diabetes”, “alpha-2-macroglobulin receptor-associated protein -> type 2 diabetes”, and “neutrophil collagenase -> ulcerative colitis”. Interestingly, the mediating effects of these pathways were lower than 5%, suggesting that these mediation pathways may not play an important role in the development of gastric cancer. In addition, the bidirectional MR results of the ten potential risk factors showed that only “gastric cancer -> ultradistal forearm bone mineral density” had significant causality (OR = 0.95, 95% CI: 0.91–0.98), but the F-statistic value was 0.22. Other bidirectional MR analyses showed no significant causality or were not feasible to perform due to too few genetic instruments in the gastric cancer dataset.

Summary of MR results
Taken together, “years of schooling” and “alpha-2-macroglobulin receptor-associated protein” were suggested to be causally associated with gastric cancer risk, and “ulcerative colitis” was considered a confirmed risk factor. Meanwhile, “androsterone sulfate” and “type 2 diabetes” showed suggestive causality for gastric cancer, while “promotilin”, “neutrophil collagenase”, “VEGFR2”, and “tyrosine-protein kinase receptor Tie-1 (soluble)” were identified as confirmed risk factors. In addition, the leave-one-out sensitivity analysis did not show a significant effect on these robust results (Fig. S2). The encoding genes of “promotilin”, “neutrophil collagenase”, “VEGFR2” and “tyrosine-protein kinase receptor Tie-1 (soluble)” are MLN, MMP8, KDR, and TIE1, respectively. The Cox proportional hazards survival analysis was performed on the web-based platform KMplot, which pooled datasets of gastric cancer GSE14210, GSE15459, GSE22377, GSE29272, GSE51105, and GSE62254. As shown in Fig. 4, all Cox proportional hazards survival analysis results showed P-value < 0.05 in overall survival (Fig. 4A), first progression (Fig. 4B), and post progression survival (Fig. 4C). In addition, in the stomach adenocarcinoma dataset of the TCGA database, the overall survival of TIE1, KDR, and MMP8 genes (Fig. 4D) and relapse-free survival of the MLN gene (Fig. 4E) showed P-value < 0.05.
Finally, we explored the extensive phenotypic MR of four identified biomarkers that may be used as therapeutic targets, including “promotilin”, “neutrophil collagenase”, “VEGFR2” and “tyrosine-protein kinase receptor Tie-1”. Two associations with significant detrimental effects were observed for targeting promotilin (Fig. 5A), including celiac disease and intestinal malabsorption (non-celiac), while beneficial effects were observed for targeting neutrophil collagenase (Fig. 5B). Two associations with significant beneficial effects were observed for targeting tyrosine-protein kinase receptor Tie-1 (Fig. 5C), including hemorrhoids and functional digestive disorders. No significant associations were found for targeting VEGFR2 (Fig. S3). Therefore, promotilin, neutrophil collagenase, and tyrosine-protein kinase receptor Tie-1 are more plausible targets for the prevention and treatment of gastric cancer.
In the pooled inverse-variance weighted phenotypic MR analysis, the 1403 non-gastric cancer disease traits are grouped and color-coded by International Classification of Disease 10th edition. The horizontal orange line represents the Bonferroni-corrected significance threshold (P < 3.7×10-5), and the purple line represents the suggestive line P = 0.05. The upward and downward triangles represent deleterious and beneficial effects, respectively.

Chemotherapeutic response difference of potential target genes
In predicting the outcome of chemotherapy response, we assessed the sensitivity of 348 gastric cancer samples to 138 drugs (Fig. 6A). Most of the IC50 values were significantly different below and above the median expression level of target genes. Combined with evidence from CTD, we found that eight common drugs, including axitinib, bortezomib, cisplatin, doxorubicin, gefitinib, sorafenib, sunitinib, and vorinostat, have been validated in vivo or in vitro in gastric cancer-related trials. Therefore, the binding ability of potential drugs to target genes was assessed by molecular docking (Fig. 6B), and drug-like chemicals were assessed by Lipinski’s rules and physicochemical properties (Fig. 6C). The results showed that axitinib and vorinostat had the best docking effects, followed by gefitinib, sorafenib, and sunitinib. For complete information on the molecular docking results, including affinity scores, receptor-ligand interactions, and others, see Supplementary file.pdf.

Discussion
With efforts to achieve robust results during the MR analyses, this study supports a causal linkage between potential risk factors and gastric cancer. Finally, nine risk factors were found to play an important role in gastric cancer, especially six metabolite- or protein-related genes that provide further insight into promising therapeutic targets. To our knowledge, this is the first large-scale systematic analysis of the causal associations between different traits and gastric cancer.
In a subgroup of a recent meta-analysis study [23], ulcerative colitis, a chronic inflammatory bowel disease, was proven to be associated with a higher risk of gastrointestinal tract cancer but not gastric cancer, possibly due to the limited number of included studies. More recently, higher levels of education or longer years of schooling have been found to be associated with an increased incidence and mortality of breast cancer [24, 25], but there have been no reports of an association with gastric cancer to date. In addition, there is no evidence of a causal relationship between alpha-2-macroglobulin receptor-associated proteins and gastric cancer. Therefore, more evidence and verification are needed in future studies.
Consistent with the results of a previous meta-analysis, patients with type 2 diabetes are under a higher risk of developing gastric cancer [26]. Moreover, a previous MR analysis in a Japanese population reported a suggestive association between type 2 diabetes and gastric cancer [9], while another MR study found no causal association [8]. However, both studies involved fewer than 1000 cases (687 and 736, respectively). In our study, using 24 SNPs (25 SNPs used in the study [9]), we obtained a significant association between type 2 diabetes and gastric cancer regardless of whether three outlying SNPs were included. Furthermore, the reverse MR results showed no causality between the two, with the leave-one-out analysis showing no variability across the whole forest plot, and the mediating effect of type 2 diabetes was less than 5%, which was above the significant threshold.
VEGFR2, an alias for the kinase insert domain receptor encoded by the gene KDR, has been identified as a therapeutic target in patients with metastatic gastric/gastroesophageal junction carcinoma [27, 28]. Evidence from a meta-analysis of VEGFR2 expression in gastric cancer showed that its overexpression was significantly associated with the median overall survival, indicating that it could be used as a promising predictive biomarker [29]. In addition, previous randomized, placebo-controlled, double-blind clinical trials have shown that ramucirumab, a monoclonal antibody of VEGFR2 antagonist, significantly extends overall survival in patients with advanced gastric cancer [30, 31]. Consistent with the existing evidence, the present study confirmed the causality between VEGFR2 and gastric cancer under the valid hypothesis of MR analyses.
Promotilin, encoded by the gene motilin (MLN) gene, a regulator of gastrointestinal motility, affects gastric motility by stimulating interdigestive antrum and duodenal contractions [32]. In the results of this MR causal association analyses, plasma protein level of promotilin was shown to have a negative effect on the odds ratio for gastric cancer. However, there has been no report on the causal relationship between the MLN gene and gastric cancer, except for a few studies that have shown that MLN plays an important part in gastroparesis syndrome, gastroesophageal reflux and pylorospasm [33–35]. We also examined the association between selected variants (rs755497 and rs9394169) and their nearest genes for promotilin (Fig. S4). The variants rs755497 (P = 1.38e−12) and rs9394169 (P = 1.78e−89) near the genes LEMD2 and MLN were found to support an association between promotilin and gastric cancer risk, whereas no evidence was found in gastric cancer, which does not violate the MR assumptions. Additionally, phenotypic MR results of promotilin were significantly associated with digestive diseases such as intestinal malabsorption, celiac disease, gastritis, and duodenitis (Fig. S3). Therefore, promotilin may provide valuable information for the prevention and treatment of gastric cancer.
Neutrophil collagenase, an alias for matrix metallopeptidase 8 (MMP8), which is encoded by the gene MMP8, is involved in the breakdown of the extracellular matrix in disease processes, such as the metastasis of gastric adenocarcinoma [36] and the prognosis of gastric cancer [37]. No causal relationship between neutrophil collagenase and gastric cancer has been reported so far; thus, it is necessary to further strengthen the causality because of the residual confounding of potential risk factors. Another plasma protein of tyrosine-protein kinase receptor Tie-1, a member of the tyrosine protein kinase family, is encoded by the gene TIE1, whose extracellular domain cleavage plays an inhibitory role in angiogenesis and vascular stability [38]. In the present study, TIE1 protein exerted a significant impact on gastric cancer, which has been considered both a prognostic biomarker and a therapeutic target in previous studies [39, 40]. The current MR study did not confirm a causal role of plasma TIE1 and MMP8 levels in gastric cancer risk, but we cannot rule out that we may have overlooked a weak association effect.
In addition to sensitivity analysis of potential risk factors, another important step to be performed is the reverse MR, which implies that there is reverse causality in violation of the MR assumptions, such as the “ultradistal forearm bone mineral density” trait in this study. On an individual level, none of the tested traits strongly moderated the causality, and some of the related traits may have co-participated in the mediation effect. Although the mediating effect of a single mediator does not show a significant proportion, multi-mediator MR analysis can still be explored in future studies.
Some limitations of this study should be noted. All MR and mediating effect analyses were based on linear models, which may affect the estimated size of nonlinear factors. The large-scale genetic data from multiple ancestries may bias causality and ignore other traits that were excluded in the initial screening steps of this study. We were unable to assess reverse causality for certain traits (e.g., neutrophil collagenase) of interest due to insufficient genomic data to extract enough genetic instruments. Another drawback is that the exposure dataset came mainly from Europe, whereas the outcome data set came from Japan, which may weaken our findings. The statistical power of the MR outcomes was moderate, suggesting that significant causality may suffer from “false negatives”; therefore, more evidence is needed to validate these results in future studies. Finally, considering the realistic efficacy of the target drug treatment evaluation, future in vivo and in vitro experiments are needed to further validate potential drugs.

Conclusions
In conclusion, our systematic MR study provides plausible causality when investigating potential risk factors for gastric cancer by ordinary MR, reverse MR, and mediating effect assessments, although we did not verify them in additional datasets. To our knowledge, this is the first MR-related analysis to identify novel traits or biomarkers of gastric cancer risk from large-scale GWAS datasets across multiple ancestries. With the availability of new GWAS data, validation of our findings may facilitate the identification of new gastric cancer risk candidates.

Supplementary Information