The Mediation Analysis Of Serum Metabolites On The Relationship Between Smoking Or Drinking And Esophageal Squamous Epithelial Lesion

Backgrounds:Esophageal cancer is a malignant digestive tract tumor of the esophageal mucosa,caused by abnormal hyperplasia of the esophageal squamous or glandular epithelium.China is a high incidence area of esophageal cancer,accounting for about half of the number of cases or deaths in the world,and more than 95%of esophageal cancer in China is esophageal squamous cell carcinoma(ESCC).Due to the lack of effective early diagnostic markers,most of the patients were found in the late clinical stage with poor prognosis,and the 5-year survival rate was only 21%,which is one of the major public health problems that need to be solved in China.ESCC is a gradual process that the development of precancerous lesions to cancer typically may need years or even more than ten years.In addition,the precancerous stage of ESCC is a precious period for its diagnosis and treatment.Early diagnosis and early treatment can reverse the disease and greatly reduce the probability of its cancerization.Epidemiological studies have shown that smoking and alcohol drinking are important risk factors of its occurrence and development,but its pathogenesis is still unclear.Therefore,the exploration of biomarkers of esophageal squamous epithelial lesions(ESEL)caused by smoking and drinking is of important reference value for early diagnosis and early treatment of esophageal cancer in people with high incidence areas.Metabolomics is the detection of metabolic fingerprints of biological samples such as serum,urine or tissue by targeted or untargeted strategies,which provides a powerful platform for the study of pathobiological mechanisms of diseases.Small molecule metabolites detected by metabolomics technology are not only the material basis of organism’s life activities and biochemical metabolism,but also reflect the organism’s response to external environmental stimuli.For example,smoking and alcohol drinking can lead to changes of metabolites in the body.Similarly,metabolic disorders of some amino acids,lipids,and energy metabolites have been found in the metabolomics of esophageal cancer.These results indicate that metabolomics has potential for early diagnosis and early treatment of esophageal cancer,so we assumed that metabolites were the mediating variables in the occurrence of ESEL caused by smoking and drinking exposure,and conducted an exploratory study.However,as metabolomics data belong to high-dimensional data with the characteristic of n<p,that is,there are many mediators,such data cannot be studied by univariate mediation analysis method.Therefore,some researchers proposed high-dimensional mediation analysis method.At present,the high-dimensional mediation analysis method is still in the exploration stage,which can be divided into two types:one is regularization based on the linear structural equation model(LSEM);The other is to reduce the dimension based on principal component analysis(PCA)to screen the mediators.Moreover,both methods need to fit the model for each mediator and outcome variable respectively,and conduct effect estimation through various methods.The high-dimensional mediation analysis method based on PCA can obtain a series of linear combination of mediators,so it can provide more mediators without considering the correlation between the mediators.But it is also difficult to interpret because it is a linear combination of a series of mediators,rather than a specific mediator,which is not suitable for research whose purpose is to screen out specific mediating variables.Most of the high dimensional mediation analysis methods based on LSEM implement the purpose of variable selection by regularizing the regression coefficient.Penalty by convex functions such as Lasso and ridge regression is the most common,which has the advantages of rapid and continuous sparse estimation.However,since the loss function is also convex function,the effect estimation has only one minimum point,which is biased and does not meet the Oracle property.Therefore,some studies have proposed methods using concave penalty,such as minimax concave penalty(MCP)method,which is consistent in model selection,approximately unbiased,and alleviates the problems of local optimality and instability of coefficient estimation of general concave penalty methods.Since there are many regularization methods,this study intends to compare the accuracy of various penalty methods for selecting true mediators through simulation experiments And according to the simulation results to investigate whether the relationship between smoking or drinking exposures and ESEL is mediated by serum metabolites.Methods:In this study,the accuracy of effect estimation was compared between univariate mediation analysis and three high-dimensional mediation analysis methods(including Lasso,smoothly clipped absolute deviation(SCAD)and MCP regularization)by simulation.Each method was repeated 500 times,and the estimation of the mediation effect,the mean square error(MSE),the family wise error rate(FWER)and the test power were calculated to evaluate the accuracy and robustness of the mediation effect estimation of different methods.Metabolomics and pathological examination were performed on the subjects collected from 2013 to 2014 in the esophageal cancer screening platform in Feicheng,Shandong Province.Multiple logistic regression was used to assess the relationship between exposure to cigarette smoking or alcohol drinking and ESEL risk,with OR values and 95%confidence intervals.The general linear regression model was used to evaluate the relationship between various exposures and metabolites and between metabolites and ESEL outcome.The FDR correction is also performed to for the multiple comparison problem.Then,a high-dimensional mediation analysis based on selected method was used to screen potential mediation metabolites between various exposures and ESEL risks.And the natural indirect effects(NIE)of each intermediate metabolite were calculated by univariate causal mediation analysis to verify the selected intermediate metabolites.Results:The simulation study showed:(1)In univariate mediation analysis,the estimated value of mediation effect has little relationship with sample size and the dimension of mediators.MSE tended to decrease gradually with the increase of sample size,but the change was not obvious.For the three high-dimensional mediation analysis methods,with the increase of sample size,the estimated value of mediation effect is close to the true value and the MSE decreases gradually.Mediator dimension has little influence on effect estimation and MSE.(2)The univariate mediation analysis has poor type I error control,and the test power is lower.The three high-dimensional mediation analysis methods have reasonably well controlled type I error,while a little conservative,and showing a decreasing trend with the increase of sample size.Similarly,the test power also shows an increasing trend,and which of the MCP is the highest.Based on above results,we believe that the comprehensive evaluation of MCP method is better,so it is selected for the high-dimensional mediation analysis of the example part.The case study showed:Smoking was associated with a 3.11-fold increased risk of ESEL(OR=3.11,95%CI 1.63-6.05);Risk for ESEL increased by 56%for each one-unit increase in smoking index(OR=1.56,95%CI 1.18-2.13);Alcohol drinking was associated with a 1.97-fold increased risk of ESEL(OR=1.97,95%CI 1.05-3.77).Under the standard of FDR<0.05,there were 3 metabolites associated with smoking,and 5 metabolites associated with smoking index.There were 16 metabolites associated with alcohol drinking,8 were up-regulated and 8 were down-regulated.134 metabolites associated with ESEL,and most presented down-regulated trend.High-dimensional mediation analysis showed that there were 3 mediator metabolites in smoking induced ESEL,including carnitine(9:0),L-histidine and L-glutamine,and the corresponding mediation effects were 13.53%,24.80%and 26.58%,respectively.Five potential mediator metabolites were screened for smoking index.In addition to the three metabolites screened by smoking,PG(14:1/7:0)and cholic acid were also selected.The mediation effects of five metabolites were 7.87%,35.51%,29.54%,10.78%and 21.01%,respectively.A total of four metabolites were screened for drinking exposure,which were PC(P-16:0/14:l),L-histidine,asparagine-phenylalanine and L-glutamine.Their mediation effects were 51.42%,63.62%,-19.91%and 14.14%,respectively.Univariate mediation analysis further confirmed that L-histidine and L-glutamine were the mediator metabolites of smoking-induced ESEL.In terms of smoking index,the mediation effects of L-histidine and L-glutamine were similar to smoking,and the mediation effect of cholic acid was also statistically significant.For alcohol exposure,PC(P-16:0/14:1),L-histidine and L-glutamine had significant natural indirect effectsConclusions:(1)Compared with univariate mediation analysis,Lasso and SCAD,the MCP-based high-dimensional mediation analysis method has a higher test power and better control of type I error.Therefore,it is more suitable for the screening of high-dimensional omics mediators;(2)Smoking significantly increases the risk of ESEL,and this process is mediated by L-glutamine,L-histidine,and cholic acid;(3)Alcohol drinking is associated with an increased risk of ESEL,which is mediated by three serum metabolites(PC(P-16:0/14:1),L-histidine and L-glutamine).