High-density lipoprotein, low-density lipoprotein and triglyceride levels and upper gastrointestinal cancers risk: a trans-ancestry Mendelian randomization study

Abstract

OBJECTIVE:

This study was conducted to explore the causal associations of high-density lipoprotein (HDL), low-density lipoprotein (LDL) and triglyceride (TG) with the risk of upper gastrointestinal cancers (esophageal cancer [EC] and gastric cancer [GC]).

METHODS:

A total of 5623 Chinese and 4133 Europeans afforded the individual-level genotyping data, and 203,608 Japanese from Biobank Japan project and 393,926 Europeans from UK Biobank supported summary statistics of cancer genetic associations. Mendelian randomization (MR) analyses, including weighted genetic risk scores (wGRSs), inverse-variance weighted (IVW), weighted median and Egger-regression, were utilized to evaluate the causal effects of three blood lipids on upper gastrointestinal cancers risk.

RESULTS:

There was no significantly causal relationships between three blood lipids and EC or GC risk among Chinese or Europeans but a potential causal association between TG and GC risk among Japanese (IVW: odds ratio [OR] = 1.11, P = 0.034; P_het = 0.679). In stratified subgroups, higher genetically predicted TG levels were causally associated with an increased risk of GC among Chinese males (wGRS: OR = 1.61, P = 0.021; IVW: OR = 1.57, P = 0.009; P_het = 0.653) and Japanese females (IVW: OR = 1.33, P = 0.024; P_het = 0.378).

CONCLUSION:

This trans-ancestry MR study suggested null significant causality between serum HDL, LDL or TG and the risk of upper gastrointestinal cancers among Chinese and Europeans, but provided evidence for a causal role of TG involved in GC etiology in Japanese (especially females), which would support a prevention guide for high-risk groups of GC. Further research with more comprehensive information is needed to explore the underlying mechanism.

INTRODUCTION

Despite a decrease in proportion of gastric cancer (GC) and esophageal cancer (EC) among all cancers during recent years, the two common upper gastrointestinal cancers remain a leading cause of cancer deaths worldwide [1]. Regionally, upper gastrointestinal cancers occur more frequently in Eastern Asia including China [1]. In terms of tumor types, esophageal squamous cell carcinoma (ESCC) and intestinal type GC are more common in Asia, whereas esophageal adenocarcinomas (EA) and poorly differentiated GC are more common in Europe [1, 2]. Although many risk factors are associated with increased incidence of upper gastrointestinal cancers [3, 4], the causal relationship with these factors remains obscure.

Blood lipids that include high-density lipoprotein (HDL), low-density lipoprotein (LDL) and triglyceride (TG) are important indicators of human health and are usually targeted as objects relevant to various diseases [5–7]. Epidemiological studies suggest that higher cancer risk occurs along with lower serum HDL levels [8], higher serum LDL [9] or higher TG levels [10]. Several studies have proposed the specific role of these lipid traits in the risk of upper gastrointestinal cancers, but the findings remain inconclusive. Lin et al. revealed no statistically significant associations of neither TG nor HDL with EC risk upon a large cohort in Norway [11]. In a Chinese population, higher TG and lower HDL were observed in GC cases compared with controls [12], while a Korean population study demonstrated that higher LDL was associated with an increased risk of gastric dysplasia, a penultimate stage of GC [13]. These findings indicate less evidence to confirm the relationship between these lipid traits and upper gastrointestinal cancers across different ancestries. To reduce bias from potential confounding factors and reverse causality, we aim to identify the relationship between HDL, LDL and TG and upper gastrointestinal cancers through Mendelian randomization (MR) analytic framework.

MR is an epidemiological method that determines a causal relationship between an exposure and an outcome using instrumental variables (IVs, such as genetic variants [14]). As far as we are aware, MR analysis has not been applied to investigate the role of blood lipids in upper gastrointestinal cancers. In this study, we aim to use MR analyses to clarify whether lower serum HDL levels, higher serum LDL or higher serum TG levels are causal risk factors associated with the risk of upper gastrointestinal cancers in East Asian and European populations, respectively. We first collect IVs of the three lipids for different ancestries through a literature review. Second, we derive genetic associations of these IVs with GC or EC risk across ancestries from both individual-level and summary GWAS datasets. Ultimately, MR techniques including weighted genetic risk scores (wGRSs), Inverse-variance weighted (IVW), weighted median and Egger-regression are utilized to perform causal inference. Considering that lipid levels are easily accessible indicators and can be controlled well by medicine and diet treatment, identification of the underlying causality between dyslipidemia and upper gastrointestinal cancers might help understand the mechanistic developmental processes of upper gastrointestinal cancers and shed new lights on prevention.

METHODS

Participants

In order to obtain individual-level genotype data, we included a total of 5623 Chinese individuals and 4133 European individuals to this study (Supplementary Table S1). Specifically, 1898 ESCC cases, 1625 GC cases and 2100 controls of Chinese ancestry with individual genotyping data were obtained from the studies of the Shanxi Upper Gastrointestinal Cancer Genetics Project (Shanxi) and the Linxian Nutrition Intervention Trial (NIT), deposited in the database of Genotypes and Phenotypes (dbGaP; phs000361.v1.p1) [15]. Moreover, 2268 EA cases and 1865 controls of European ancestry with genotypes were derived from Barrett’s and Esophageal Adenocarcinoma Genetic Susceptibility Study (BEAGESS; phs000869.v1.p1) [16].

In addition to the individual-level data, summary statistics of genetic associations were also included in this study (Supplementary Table S1). For East Asians, summary statistics of genetic associations with ESCC (1300 cases and 195,745 controls) and GC risk (6563 cases and 195,745 controls) were extracted from Biobank Japan project (BBJ), a cohort of ~200,000 patients diagnosed with 47 common diseases from 2003 (http://jenger.riken.jp/en/result) [17, 18]. For Europeans, summary statistics of genetic associations with GC risk (including 554 cases and 393,372 controls) were downloaded from UK Biobank single variant association analysis results at Lee Lab (https://www.leelabsg.org/resources). UK Biobank is a large cohort study with more than 500,000 adults (from 37 to 73 years old at baseline) involved in between 2006 and 2010 [19]. Lee’s team analyzed about 1400 binary phenotypes on 28 million imputed variants among 400,000 white British UK Biobank samples and provided summary statistics online [20].

Genetic instrument selection

We performed a literature review in PubMed with the keywords “blood lipids AND polymorphism” to identify IVs strongly associated with HDL, LDL and TG (search from January 1, 2013 to December 15, 2018). Among the 3123 literatures totally retrieved, we selected the SNPs underlying the criteria: (1) human study and published in English; (2) observational (casecontrol or cohort) design; (3) available effect estimate (β) and standard error (SE) of the relevant SNPs; (4) the association with serum HDL, LDL or TG reaching genome-wide significance level (P < 5 × 10⁻⁸). Finally, 218 lipid-related SNPs [21–23] in Asians (Supplementary Note) and 165 lipid-related SNPs [24] in Europeans were selected for further analysis.

We matched those IVs with our outcome data and performed quality control for IVs using the following criteria, which was constructed for Asians and Europeans, separately: (1) minor allele frequency (MAF) ≥ 0.05; (2) P value for Hardy–Weinberg equilibrium (HWE) in controls ≥1 × 10⁻⁶, (3) call rate ≥95% and (4) linkage disequilibrium (LD) analysis (r² ≥ 0.5).

Finally, a total of 137 SNPs (70 for HDL, 56 for LDL and 42 for TG) and 139 SNPs (72 for HDL, 56 for LDL and 42 for TG) were retained as IVs to evaluate effect of the three lipid traits on ESCC and GC risk, respectively, in Chinese; 155 SNPs (84 for HDL, 62 for LDL and 48 for TG) and 152 SNPs (84 for HDL, 62 for LDL and 48 for TG) were retained to evaluate the effect of the three lipid traits on ESCC and GC risk, respectively, in Japanese; 117 SNPs (60 for HDL, 53 for LDL and 36 for TG) and 157 SNPs (81 for HDL, 71 for LDL and 48 for TG) were retained to evaluate the effect of the three lipid traits on EA and GC risk, respectively, in Europeans (Fig. 1 and Supplementary Tables S2 and S3).

Fig. 1. Flow chart for process of our study.

SNP single-nucleotide polymorphism, ESCC esophageal squamous cell carcinoma, GC gastric cancer, EA esophageal adenocarcinoma, β SNP effect size, SE standard error, HDL high-density lipoprotein, LDL low-density lipoprotein, TG triglyceride, MAF minor allele frequency, HWE Hardy–Weinberg equilibrium, LD linkage disequilibrium, IVs instrumental variables, MR mendelian randomization, EC esophageal cancer.

Statistical analysis

MR has been widely applied to estimate the association between the exposure and outcome through IVs. As genotype is presumed to be randomly allocated during meiosis, confounding factors are considered to be equally distributed among different genotypes. Therefore, this technique effectively avoids the confounding bias compared to traditional epidemiological studies, and is economical, time-saving and labor-saving compared to the randomized trials. MR analysis requires several assumptions (Supplementary Fig. S1): (i) IVs are associated with the exposure; (ii) IVs influence the outcome only via the exposure; and (iii) IVs are independent of factors (measured and unmeasured) that confound the exposure-outcome relationship [25]. Here, we utilized MR approaches including wGRSs, IVW, weighted median and Egger-regression, to evaluate the causal associations between the three blood lipids and upper gastrointestinal cancers risk. Logistic regression with adjustment of age and sex was applied in calculating the effect estimates of IVs and outcomes.

wGRS analysis

The primary analysis was performed using individual-level genotypes. We generated wGRS for each of the three lipid traits (HDL, LDL and TG), calculating the cumulative effect of multiple genetic variants using the following formula:

$$w{\mathit{\text{GRS}}}_j=\sum _{i=1}^n{\beta }_i{X}_{ij}$$

where wGRS_j was the genetic risk score for individual j; n was the number of SNPs acting as IVs for blood lipids reported previously; β_i was the effect estimate (β coefficient) of the ith SNP; X_ij was the number of the risk alleles for the ith SNP (0, 1, 2 for wild-type homozygous, heterozygous and homozygous, respectively). A higher wGRS value for an individual corresponds to a higher genetically inferred lipid level. Associations between lipid trait wGRSs and the risk of upper gastrointestinal cancers (ESCC, GC and EA, separately) were assessed using logistic regression with adjustment of age and sex.

IVW analysis.

In addition to GRS-based method, we used IVW approach to calculate causal effects using summary association estimates. This method combines the effect estimate of SNP-exposure and SNP-outcome using the following formula in a fixed-effect model [26]:

$${\widehat{\beta }}_{\mathit{\text{IVM}}}=\frac{{\sum }_{i=1}^n{X}_i{Y}_i{\delta }_{Y_i}^{-2}}{{\sum }_{i=1}^n{X}_i^2{\delta }_{Y_i}^{-2}}$$

Assuming n was the number of SNPs; X_i and Y_i were the regression coefficients of SNP-exposure and SNP-outcome of the ith SNP; δ_Xi, δ_Yi were standard errors (SEs) corresponding with X_i and Y_i, respectively; taking the weight of variance as reciprocal, the ratio of Y_i to X_i of each SNP was combined, which then made up the causal effect estimation of the weighted linear regression [27]. The ratio of causal effect estimation of the ith SNP was Y_i/X_i, of which SE could be approximated as se $\left(\begin{array}{c}\widehat{\beta }\\ \mathit{\text{IVW}}\end{array}\right)=\sqrt{\frac{1}{{\sum }_{i=1}^n{X}_i^2{\delta }_{Y_i}^{-2}}}$ by delta method [28]. Random-effect model in IVW analysis was utilized when P < 0.05 in the heterogeneity test.

Sensitivity analyses.

Two additional MR approaches, weighted median method and Egger-regression, were used as sensitivity analyses to test the robustness of the findings. Weighted median regression provides a consistent estimate of the causal effect even when up to 50% of the information contributing to the analysis comes from genetic variants that are invalid IVs [29]. Egger-regression provides a valid test of directional (unbalanced) pleiotropy, and a valid test of the causal null hypothesis, giving a consistent causal effect estimate even when all the genetic variants are invalid IVs. [26]. P < 0.05 in the test for intercept of Egger-regression indicates the existence of pleiotropy.

Given that lipid profiles differ between men and women and are changing with the increase of age [30], stratification analyses were also carried out in populations with relevant information. All statistical analyses were performed by R version 3.5.1. MR analysis was conducted using MendelianRandomization R package. All statistical tests under two-sided with P < 0.05 were considered statistically significant.

RESULTS

Based on individual-level genotype data of the Chinese population, we observed no significant association between HDL (wGRS: OR = 1.04, P = 0.362; IVW: OR = 1.00, P = 0.927; P_het = 0.535), LDL (wGRS: OR = 0.99, P = 0.696; IVW: OR = 1.02, P = 0.510; P_het = 0.625), TG (wGRS: OR = 0.98, P = 0.925; IVW: OR = 1.19, P = 0.220; P_het = 0.062) and ESCC risk (Tables 1 and 2 and Supplementary Fig. S2). Similarly, no causal relationship was found between HDL (wGRS: OR = 1.00, P = 0.951; IVW: OR = 1.01, P = 0.791; P_het = 0.050), LDL (wGRS: OR = 1.03, P = 0.360; IVW: OR = 0.99, P = 0.759; P_het = 0.023), TG (wGRS: OR = 1.39, P = 0.062; IVW: OR = 1.19, P = 0.233; P_het = 0.164) and GC risk neither using the wGRS nor IVW method (Tables 1, 2 and Supplementary Fig. S2). However, in the subgroup analysis, we found a significantly increased risk of GC with the increased TG levels in males (wGRS: OR = 1.61, P = 0.021; IVW: OR = 1.57, P = 0.009; P_het = 0.653; Tables 1 and 2 and Supplementary Figs. S3 and S4). In addition, a marginal causal relationship between TG and GC risk was found in IVW method among Chinese younger than 60 years old (wGRS: OR = 1.30, P = 0.306; IVW: OR = 1.54, P = 0.047; P_het = 0.668; Tables 1 and 2 and Supplementary Figs. S3 and S4). In Europeans, we did not observe any significant association between HDL (wGRS: OR = 0.95, P = 0.781; IVW: OR = 0.96, P = 0.828; P_het = 0.047), LDL (wGRS: OR = 1.16, P = 0.354; IVW: OR = 0.79, P = 0.168; P_het = 0.012), TG (wGRS: OR = 1.41, P = 0.071; IVW: OR = 0.82, P = 0.407; P_het < 0.001) and EA risk neither in the wGRS nor IVW framework (Tables 1 and 2 and Supplementary Figs. S3, S4 and S5).

Table 1.

Associations between lipid-trait wGRSs and the risk of upper gastrointestinal cancers in overall and subpopulations of individual genotyping data.

Lipid trait	Subgroup	ESCC-Chinese		GC-Chinese		EA-European
		OR (95% CI)	P	OR (95% CI)	P	OR (95% CI)	P
HDL	Overall	1.04 (0.95, 1.14)	0.362	1.00 (0.91, 1.11)	0.951	0.95 (0.65, 1.38)	0.781
	Sex
	Male	1.08 (0.96, 1.21)	0.198	1.03 (0.92, 1.16)	0.568	0.96 (0.64, 1.45)	0.850
	Female	0.99 (0.84, 1.15)	0.850	0.92 (0.75, 1.13)	0.423	0.88 (0.35, 2.22)	0.791
	Age (years)
	<60	0.96 (0.84, 1.10)	0.580	0.96 (0.83, 1.11)	0.565	1.43 (0.75, 2.72)	0.273
	≥60	1.12 (0.99, 1.27)	0.075	1.05 (0.92, 1.19)	0.514	0.76 (0.48, 1.21)	0.253
LDL	Overall	0.99 (0.94, 1.04)	0.696	1.03 (0.97, 1.08)	0.360	1.16 (0.84, 1.61)	0.354
	Sex
	Male	0.98 (0.92, 1.05)	0.611	1.02 (0.95, 1.09)	0.577	1.16 (0.82, 1.64)	0.413
	Female	1.00 (0.92, 1.09)	0.977	1.05 (0.94, 1.17)	0.394	1.18 (0.53, 2.66)	0.683
	Age (years)
	<60	0.99 (0.91, 1.07)	0.753	1.06 (0.97, 1.14)	0.190	1.44 (0.83, 2.48)	0.194
	≥60	0.99 (0.92, 1.07)	0.814	1.00 (0.93, 1.08)	0.977	1.04 (0.70, 1.54)	0.856
TG	Overall	0.98 (0.71, 1.37)	0.925	1.39 (0.98, 1.96)	0.062	1.41 (0.97, 2.06)	0.071
	Sex
	Male	1.03 (0.69, 1.56)	0.874	1.61 (1.07, 2.40)	0.021	1.39 (0.92, 2.09)	0.118
	Female	0.90 (0.52, 1.55)	0.702	0.93 (0.47, 1.83)	0.823	1.66 (0.64, 4.28)	0.294
	Age (years)
	<60	1.08 (0.67, 1.75)	0.743	1.30 (0.79, 2.16)	0.306	1.51 (0.80, 2.85)	0.209
	≥60	0.90 (0.57, 1.42)	0.653	1.48 (0.92, 2.37)	0.108	1.39 (0.87, 2.21)	0.169

wGRS weighted genetic risk score (adjusted for sex and age), ESCC esophageal squamous cell carcinoma, GC gastric cancer, EA esophageal adenocarcinoma, HDL high-density lipoprotein, LDL low-density lipoprotein, TG triglyceride, OR odds ratio, CI confidence interval.

Table 2.

IVW analyses of causal associations between blood lipids and the risk of upper gastrointestinal cancers in overall and subpopulations of individual genotyping data.

Lipid trait	Subgroup	ESCC-Chinese			GC-Chinese			EA-European
		OR (95% CI)	P	P het	OR (95% CI)	P	P het	OR (95% CI)	P	P het
HDL	Overall	1.00 (0.92, 1.08)	0.927	0.535	1.01 (0.93, 1.10)	0.791	0.050	0.96 (0.67, 1.38)	0.828	0.047
	Sex
	Male	1.02 (0.93, 1.13)	0.632	0.792	1.02 (0.93, 1.13)	0.650	0.111	1.00 (0.72, 1.41)	0.981	0.211
	Female	0.97 (0.85, 1.11)	0.684	0.541	0.94 (0.80, 1.11)	0.490	0.377	1.02 (0.47, 2.20)	0.970	0.071
	Age (years)
	<60	0.96 (0.86, 1.08)	0.500	0.873	0.97 (0.85, 1.10)	0.596	0.476	1.03 (0.61, 1.76)	0.909	0.092
	≥60	1.05 (0.94, 1.17)	0.414	0.201	1.03 (0.92, 1.15)	0.599	0.064	0.99 (0.68, 1.45)	0.953	0.356
LDL	Overall	1.02 (0.97, 1.07)	0.510	0.625	0.99 (0.93, 1.05)	0.759	0.023	0.79 (0.56, 1.10)	0.168	0.012
	Sex
	Male	0.99 (0.93, 1.05)	0.753	0.656	1.02 (0.96, 1.08)	0.527	0.135	0.97 (0.72, 1.31)	0.856	0.070
	Female	0.99 (0.92, 1.07)	0.815	0.900	1.04 (0.94, 1.14)	0.465	0.208	1.26 (0.54, 2.95)	0.589	0.004
	Age (years)
	<60	0.98 (0.91, 1.05)	0.555	0.581	1.04 (0.96, 1.11)	0.327	0.589	1.19 (0.65, 2.20)	0.568	0.001
	≥60	1.00 (0.94, 1.07)	0.984	0.680	1.01 (0.93, 1.10)	0.788	0.006	0.93 (0.67, 1.30)	0.678	0.053
TG	Overall	1.19 (0.90, 1.56)	0.220	0.062	1.19 (0.89, 1.59)	0.233	0.164	0.82 (0.51, 1.31)	0.407	< 0.001
	Sex
	Male	1.11 (0.79, 1.56)	0.556	0.171	1.57 (1.12, 2.19)	0.009	0.653	1.27 (0.80, 2.01)	0.314	0.001
	Female	0.96 (0.61, 1.52)	0.865	0.554	1.06 (0.60, 1.87)	0.844	0.800	1.22 (0.48, 3.10)	0.683	0.013
	Age (years)
	<60	1.31 (0.88, 1.95)	0.180	0.426	1.54 (1.01, 2.35)	0.047	0.668	1.26 (0.64, 2.50)	0.499	0.003
	≥60	0.86 (0.59, 1.26)	0.441	0.120	1.31 (0.88, 1.95)	0.176	0.099	1.26 (0.77, 2.05)	0.361	0.005

Fixed-effect model was used when P_het > 0.05 and random-effect model was used when P_het < 0.05.

P_het P value of heterogeneity test, MR mendelian randomization, ESCC esophageal squamous cell carcinoma, GC gastric cancer, EA esophageal adenocarcinoma, HDL high-density lipoprotein, LDL low-density lipoprotein, TG triglyceride, IVW inverse-variance weighted method, OR odds ratio, CI confidence interval.

Interestingly, IVW analyses based on summary statistics in Japanese showed that GC risk became higher with increasing serum TG levels (OR = 1.11, P = 0.034; P_het = 0.679; Table 3 and Supplementary Fig. S6), especially in the female subgroup (OR = 1.33, P = 0.024; P_het = 0.378; Table 3 and Supplementary Fig. S7). Consistent with the results in Chinese, there was no causal effect on ESCC risk (IVW_HDL: OR = 1.02, P = 0.146, P_het < 0.001; IVW_LDL: OR = 1.00, P = 0.956, P_het < 0.001; IVW_TG: OR = 0.71, P = 0.065, P_het < 0.001; Table 3 and Supplementary Fig. S6), and neither HDL nor LDL acted as a causal factor on GC risk in Japanese (IVW_HDL: OR = 1.01, P = 0.258, P_het = 0.068; IVW_LDL: OR = 0.99, P = 0.395, P_het < 0.001; Table 3 and Supplementary Fig. S6). In line with the results in Europeans with genotype data, none of these three lipids was associated with the GC risk using the UKBB summary statistics (IVW_HDL: OR = 1.27, P = 0.132; P_het = 0.154; IVW_LDL: OR = 0.74, P = 0.055; P_het = 0.896; IVW_TG: OR = 0.72, P = 0.085; P_het = 0.065; Table 3, Supplementary Fig. S5).

Table 3.

IVW analyses of causal associations between blood lipids and the risk of upper gastrointestinal cancers in overall and subpopulations of summary statistics.

Lipid trait	Subgroup	ESCC-Japanese			GC-Japanese			GC-European
		OR (95% CI)	P	P het	OR (95% CI)	P	P het	OR (95% CI)	P	P het
HDL	Overall	1.02 (0.99, 1.05)	0.146	<0.001	1.01 (0.99, 1.03)	0.258	0.068	1.27 (0.93, 1.74)	0.132	0.154
	Sex
	Male	1.03 (0.99, 1.06)	0.122	0.003	1.01 (0.99, 1.03)	0.385	0.232
	Female	0.92 (0.82, 1.03)	0.151	0.692	0.99 (0.96, 1.01)	0.279	0.336
LDL	Overall	1.00 (0.98, 1.02)	0.956	<0.001	0.99 (0.98, 1.01)	0.395	<0.001	0.74 (0.55, 1.01)	0.055	0.896
	Sex
	Male	1.00 (0.97, 1.03)	0.860	<0.001	1.00 (0.98, 1.01)	0.693	0.002
	Female	1.02 (0.95, 1.03)	0.570	0.226	1.00 (0.98, 1.02)	0.768	0.002
TG	Overall	0.71 (0.50, 1.02)	0.065	<0.001	1.11 (1.01, 1.22)	0.034	0.679	0.72 (0.49, 1.05)	0.085	0.065
	Sex
	Male	0.73 (0.50, 1.07)	0.110	<0.001	1.06 (0.93, 1.19)	0.386	0.453
	Female	0.64 (0.35, 1.16)	0.144	0.500	1.33 (1.04, 1.70)	0.024	0.378

Fixed-effect model was used when P_het > 0.05 and random-effect model was used when P_het < 0.05.

P_het P value of heterogeneity test, MR mendelian randomization, ESCC esophageal squamous cell carcinoma, GC gastric cancer, EA esophageal adenocarcinoma, HDL high-density lipoprotein, LDL low-density lipoprotein, TG triglyceride, IVW inverse-variance weighted method, OR odds ratio, CI confidence interval.

Several sensitivity analyses were also conducted to assess the validity of our results. As shown in Supplementary Tables S4 and S5, most of the overall and stratified results of weighted median and Egger-regression were consistent with our primary results. Focusing on the comparation with our main findings among Chinese and Japanese, we found unstable results in the sensitivity analyses. In the Chinese population, our main results that increasing TG levels caused higher GC risk among males and younger subpopulation was not validated (Males: weighted median: OR = 1.58, P = 0.062; Egger-regression: OR_β = 1.31, P_β = 0.260, Pintercept = 0.276. <60 years: weighted median: OR = 1.55, P = 0.163; Egger-regression: OR_β = 1.45, P_β = 0.217, P_intercept = 0.771). Although the result of weighted median for the association between TG and risk of ESCC in Chinese was significant (OR = 1.49, P = 0.040), it attenuated to the null in Egger-regression (OR_β = 1.10, P_β = 0.673, P_intercept = 0.651), in line with results of wGRS and IVW methods. In Japanese, the causal relationship between TG and GC risk was also supported by weighted median method in overall population (weighted median: OR = 1.15, P = 0.034; Egger-regression: OR_β = 1.06, P_β = 0.409, P_intercept = 0.403). However, the causal relationship among females attenuated to borderline significance in weighted median method (weighted median: OR = 1.42, P = 0.066; Egger-regression: OR_β = 1.25, P_β = 0.269, P_intercept = 0.702). We additionally obtained two significant results including the association between HDL and ESCC risk among males (weighted median: OR = 1.04, P = 0.015; Egger-regression: OR_β = 1.03, P_β = 0.068, P_intercept = 0.241), and the association between TG and ESCC risk in overall population (weighted median: OR = 0.70, P = 0.102; Egger-regression: OR_β = 0.54, P_β = 0.037, P_intercept = 0.236) and female subjects (weighted median: OR = 0.70, P = 0.427; Egger-regression: OR_β = 0.39, P_β = 0.031, P_intercept = 0.112). However, these associations derived from only one method were not reliable enough to make causal inference, indicating the existence of bias or unknown confounding factors.

DISCUSSION

In this study, we evaluated the causal effect of three serum lipids (HDL, LDL and TG) on risk of upper gastrointestinal cancers (EC and GC) through MR analyses. Overall, the results did not support any causal association between the three lipids and the risk of upper gastrointestinal cancers in European populations. Nevertheless, we found evidence for a causal relationship between TG and GC risk among East Asian populations.

To the best of our knowledge, there was no causal association between the three lipids and upper gastrointestinal cancers in Europeans, which reached a consensus with previous cohort studies [10, 11, 31, 32]. However, a MR meta-analysis reported that reduction in circulating HDL might potentially increase overall cancer risk in Asians [33]. Specifically, inverse associations were demonstrated between HDL and GC risk in Chinese case-control studies and a Korean cohort study [12, 34, 35]. Besides, Wang et al. found that pre-therapy serum HDL levels were significantly lower in ESCC patients than in normal controls [36], while no significant association was found in the Chinese population of this study. MR analyses in this study did not show any causal association between LDL and the risk of the two cancers in Chinese, either. Considering LDL is the key to the treatment for dyslipidemia related diseases [37–40], we could not rule out the possibility that individuals with genetically high LDL levels had already received medical intervention and diet changes before lipid measurement [21–24], which might result in biased judgment of the representativeness of genetic variants to LDL levels.

Previous studies have reported that serum TG involved in the pathogenesis of lung, rectal, thyroid, prostate, and gynaecological cancers [32], but there is still lack of evidence on the causal impact of TG on EC or GC risk. According to the results in this study, serum TG had genetically causal impact on GC risk among East Asians but not Europeans. Different genetic architecture of TG among ethnic groups [24] may be one reason and Asian populations (including Chinese and Filipino men and women, Japanese women, and Asian Indian men) were reported to have greater odds of suffering from hypertriglyceridemia compared to Europeans [41]. Meanwhile, GC and EC were more common and brought higher burden in Asia than in other parts of the world [42]. This phenomenon due to combined influence of race, region, culture, living habits, etc. might be responsible for our different results between East Asians and Europeans. Based on IVW and weighted median methods using summary statistics from BBJ, we found a potential causal relationship that higher serum TG levels might give rise to higher risk of GC among Japanese. After stratification, this relationship only existed in IVW framework among females. This contrary causal relationship compared to that in Chinese population in our primary analyses seems because of delicate genetic difference as well as distinct environmental exposures, diet, and lifestyle factors between Japanese and Chinese populations. Marginal causal relationships between TG and GC risk in Chinese younger than 60 years old, and ESCC risk in the overall Chinese population were obtained from only one of the MR approaches without support from individual-level analysis (wGRS). Meanwhile, Egger-regressions showed no directional pleiotropy generated by multiple IVs. Taking into consideration these results, we conjectured that there might be underlying bias and confounding factors. One possible bias might arise from selection of samples. For example, since the Chinese population in our study consisted of participants from two provinces—Shanxi and Henan, and the burden of lipid-related diseases in the two places differed to a certain extent [43], hence the distribution of genetic variants in cases and controls might not be perfectly random. What’s more, factors associated with both serum TG levels and GC risk, such as sex hormones and helicobacter pylori (Hp) infection, may act as confounders in our study. As reported by previous studies, premenopausal women have lower serum TG levels and lower risk of cardiovascular diseases than men due to sex hormones and chromosomal effects [44–46]. Sex hormones have also been hypothesized to be associated with the incidence of upper gastrointestinal cancers [47–49]. This is a factor worth considering in stratification, while precise data on menopause in women of this study was not available to determine the role of sex hormones in the association between TG and GC risk. Interestingly, higher serum TG level was reported often accompanied by higher risk of Hp infection, which occurred more frequently in males [50–52]. Hp has been proven to be an essential risk factor of GC [53, 54]. Though the mechanism that accounts for the association between TG and Hp has not been explicitly elucidated, inflammation and insulin resistance are proposed as the possible link [52, 55]. Therefore, more attention could also be paid to Hp when digging into the relationship between TG and GC risk.

So far, the mechanism of serum TG increasing the incidence of GC remains an enigma. However, hypertriglyceridemia has been hypothesized to lead to low-grade inflammation due to the detrimental effect of free fatty acids and monoacylglycerols in decomposition of TG [56, 57]. On the other hand, high serum TG levels can also result in impaired blood flow, ischemia and inflammation by increasing plasma viscosity [58, 59]. Moreover, some key lipid metabolism-derived signals in several signaling pathways and adipogenesis may promote oncogenesis [60, 61]. Y Gong brought out the similarities between adipogenesis and oncogenesis and proposed two underlying factors (malonyl-CoA and TG/free fatty acid cycling) which might cause uncontrolled cell proliferation leading to cancer [60]. Therefore, we propose a similar probable hypothesis in stomach that when patients with high serum TG levels are in a state of susceptibility such as gastric epithelial injury or infection, the toxic products from the decomposition of TG reach the gastric epithelium through the injured blood vessels accelerating the development of local inflammation. Also, with the release of related cytokines, inflammatory infiltration develops and eventually leads to GC.

The strengths of this study lie in several aspects. First, this is the first study to explore associations between serum HDL, LDL and TG levels and upper gastrointestinal cancers risk in multiple ethnic groups by MR analyses. Second, our causal inference was derived from the results of both individual-level and summary statistics. Third, bias due to population stratification was reduced since we conducted these analyses respectively in Chinese, Japanese and European populations. However, limitations can be derived from this study. First, our analyses for GC risk among Europeans and upper gastrointestinal cancers risk among Japanese were limited in summary statistics, which might influence the rigor and persuasiveness of the result. Second, MR assumptions (ii) and (iii) are not possible to definitely validate, and heterogeneity existed in part of the analyses, but we performed complementary sensitivity approaches to increase the robustness of our results. In addition, more information from individuals is needed to be incorporated for further investigation, such as menopause in women, medicine use, diet change and histological classification.

To conclude, our findings did not imply a causal role of HDL, LDL or TG in the risk of upper gastrointestinal cancers in Europeans and Chinese, but provide evidence on a causal association between TG and GC risk among Japanese (especially females). Further investigations are needed to confirm our conclusions and increase the generalizability.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41430-022-01078-6.

DATA AVAILABILITY

Individual genotyping data of Chinese and European ancestry were obtained from dbGaP datasets (phs000361.v1.p1 and phs000869.v1.p1, respectively). Summary statistics of genetic associations with gastric cancer risk were extracted from UK Biobank single variant association analysis results at Lee Lab (https://www.leelabsg.org/resources). 1Shanxi and Linxian NIT: dbGaP, phs000361.v1.p1. 2BEAGESS: dbGaP, phs000869.v1.p1. 3BBJ GWAS summary statistics: http://jenger.riken.jp/en/result. 4UK Biobank GWAS summary statistics at Lee Lab: https://www.leelabsg.org/resources.