以中介效應分析評估吸菸誘導之人類肺腺癌及肺部鱗狀細胞 癌的差異基因表達模式

Abstract

背景：肺癌的常見亞型有鱗狀細胞癌、腺癌、小細胞癌及大細胞癌，許多研究發現在從未抽菸者與抽菸者的兩族群中，肺癌型態分布有很明顯的差異，抽菸者與從未抽菸者間基因的表達差異與突變也被某些研究發現有相異之處，然而，問題在於這些被找到的基因是否參與在因果途徑(causal pathway)中，抑或僅是抽菸習慣伴隨的附帶現象，本論文的目的即是找到抽菸行為導致肺癌亞型不同的因果基因途徑。 材料與方法：我們利用美國癌症基因圖譜計畫(The Cancer Genome Atlas，簡稱TCGA)的資料進行插補法中介效應分析探討可能與抽菸型態及肺癌亞型有關的失調基因，利用肺癌型態為依變項，抽菸狀態為自變項的羅吉斯迴歸作為反事實(counterfactual)資料的插補模型，並納入年齡、性別、種族與腫瘤分期作為共變量，抽菸狀態為暴露，基因表現量為中介因子，肺癌亞型為結果，為了計算族群平均效應(population-average effect) ， 我們使用了倒數機率權重(Inverse probability weighting)，並以偏誤加速校正之靴拔法(bias-corrected and accelerated bootstrap)計算效應之信賴區間，本分析使用軟體R 內的套件”medflex”，所有模型假設並無未測量的干擾因子，為驗證結果，亦從Gene Expression Omnibus(GEO)下載驗證資料集：GSE41271、GSE50081 及GSE81089。 結果：我們找到五個自然間接效應(natural indirect effect)在三組皆有顯著(p <0.05)的基因SIAH2、ABCC5、ABCF3、SKP2 以及MAP6D1，其自然間接效應的勝算比與信賴區間分別為2.94(95%信賴區間：2.07-4.17)、2.46(95%信賴區間：1.74-3.49)、2.44(95%信賴區間：1.74-3.42)、2.27(95%信賴區間：1.65-3.11)、2.18(95%信賴區間：1.64-2.90)，進一步探討多中介因子的間接效應發現，在所有可能的中介因子組合中，SIAH2、ABCC5、SKP2 及MAP6D1 四個基因的間接效應最大，其自然間接效應的勝算比為2.94(95%信賴區間：1.66 - 5.06)，敏感度分析說明未觀察到的干擾因子效應需要非常強才會使SIAH2、ABCC5、SKP2 及MAP6D1 四個基因之真實自然間接效應為零(null)或與本論文所得之估計值方向相反。 結論：SIAH2、ABCC5、ABCF3、SKP2 以及MAP6D1 可能是抽菸導致肺癌亞型不同的關鍵基因，其可能的生物機轉有待未來實驗去驗證與釐清。

Introduction：The most common morphological types of lung cancers are squamous cell carcinoma, adenocarcinoma, small cell undifferentiated carcinoma, and large cell undifferentiated carcinoma. Differences have been observed in the expression patterns and mutations for genes when never-smokers are compared to smokers. However, the key question is whether these identified genes involved in the causal pathway, or whether they were simply epiphenomena of smoking habits. The aim of this study is to identify the causal pathway between tobacco smoking and the types of lung cancer. Materials and methods：We performed imputation based mediation analysis to identify dysregulated genes associated with tobacco smoking and lung cancer subtypes by using the datasets from The Cancer Genome Atlas (TCGA). Logistic regression model which lung cancer type as outcome regressed on smoking status was used to impute for counterfactual data. Covariates included in the models are age, sex, and cancer stage. For population-average mediation effect, we use inverse probability weighting method. Mediation analyses was conducted using the “medflex” package in R with smoking as the “exposure”, gene expression as the “mediator”, and lung cancer type as the “outcome”. bias-corrected and accelerated (BCa) bootstrap were then used to model the indirect association between smoking status and lung cancer subtype through changes in gene expression. The models assumed no unmeasured confounding. To validate the findings, two additional datasets from Gene Expression Omnibus (GEO): GSE41271, GSE50081, and GSE81089 were analyzed. Results：We found five genes including SIAH2, ABCC5, ABCF3, SKP2, and MAP6D1 that had significant (p <0.05) natural indirect effect in three datasets. The odds ratios for natural indirect effect are 2.94(95% confidence interval: 2.07-4.17), 2.46(95% confidence interval: 1.74-3.49), 2.44(95% confidence interval: 1.74-3.43), 2.27(95% confidenceinterval: 1.65-3.11), and 2.18(95% confidence interval: 1.64-2.90), respectively. By performing multiple-mediators mediation analysis, we found the combination of SIAH2, ABCC5, SKP2, and MAP6D1 had the largest indirect effect which was 2.94 odds ratio (95% CI：1.66 - 5.06). The sensitivity analysis for the indirect effect of the combination of SIAH2, ABCC5, SKP2, and MAP6D1 shows that the effects of the unmeasured confounders need to be very strong.to make the point estimate be null or in the opposite direction. Conclusion：In conclusion, SIAH2, ABCC5, ABCF3, SKP2 and MAP6D1 may be the key genes involving in the pathway that the effect of smoking on lung cancer type. Further biological mechanisms have to be validated in the future.