以文字雲探勘肺癌預後相關基因

Abstract

肺癌近年來已經成為全球主要癌症死因之首。根據世界衛生組織2020 年最新統計數據顯示，在不考慮年齡和性別情況下，肺癌占所有癌症死亡率18%。另外根據台灣衛生福利部統計資料，肺癌已經連續10 年位居癌症死因第一位，死亡率約為19%。而肺癌之中，非小細胞肺癌發生率高達90%，更其中以肺腺癌約70%佔大多數。 背景： 透過文字雲快速歸納所有學術論文中討論最廣泛的基因，並使用分子檢測技術找出生物標(biomarker) 以預測肺腺癌患者之存活率或預後情形。本研究目標：1.找出與肺腺癌存活率或預後相關的基因生物標記分子。2. 找出一種更好的方法分析患者臨床數據並評估風險模型3. 探勘學術論文討論之基因與臨床病人表現差別。 方法： 從PubMed 網站擷取期刊論文的摘要，並建立基因資料庫。其中一組數據取自美國國家癌症研究所，其中有443 個病人，來自4 個不同研究機構。此組數據，再根據數據特性及原作者使用建議，分成訓練組以及測試組。此外，同時再納入2 組已公開發表的數據為測試組。前面這3 組都是核糖核酸微陣列 (RNA microarray) 數據。最後我們另外使用一組核糖核酸定序 (RNA sequencing) 同樣作為測試組，驗證風險預測模型。 結果： 目前學說上關於肺癌預後或者病人存活率討論度最高的5 個基因為：EGFR、KRAS、TP53、ALK 和AKT1 。另外我們使用3 種不同方法，並透過調整風險分數(risk score) 參數、風險分數切點、或者基因表現切點等共計38 種分析組合，其中有2 種組合結果能同時使訓練組以及測試組的結果呈現顯著，亦即病人的存活或預後與風險分數有顯著關係。第一種是將測試組的患者基因以中位數分成高表現量與低表現量，接著調整成其各自適用的參數值，最後將風險分數以maxstat 分成高分群與低分群，並可以發現這個組合的基因分子標記與病人生存或預後有顯著相關性(P < 0.0001, P = 0.00013, P = 0.00015, P = 0.011, P < 0.0001, P = 0.0017; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 和 TCGA_LUAD 依序對應排列)。另一種方式則不將病人預做分群，而是使用連續型Cox 迴歸分析，得到其參數後計算風險分數。此時，不論是用中位數抑或使用maxstat 將風險分數分群，皆能獲得具有顯著相關性的結果 (以中位數分群: P < 0.0001, P = 0.024, P < 0.0001, P = 0.0074, P <0.0001, P = 0.00012; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 和TCGA_LUAD 依序對應排列)、 (以maxstat 分群: P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 和 TCGA_LUAD 依序對應排列)。 結論： 總結來說，最後一種方法，也就是不要將病人預先分組，所得的結果會比較樂觀並可能較適合往後不同型態的研究。17 種基因生物標記分子(SP1、E2FF3、SPP1、GSTP1、ALK、KIF5B、CCL2、RHOB、AXL、GATA3、RET、CCL19、FGFR4、POU5F1、ID1、TWIST1、ROR1)組成的風險模型能預測其存活率或者預後情況，提供給病人後續治療之參考。

Lung cancer has been the top rank of the leading cancer-caused deaths worldwide in recent years. In 2020, according to the World Health Organization International Agency for Research on Center, the global mortality of all cancer types, lung cancer accounted for 18% in both sex and all ages. In 2019, based on the Ministry of Health and Welfare, the lung cancer mortality in Taiwan was 19.3%, which was the highest rate among other cancer types. Among all histologic subtypes of lung cancer, adenocarcinoma is the most commonly diagnosed either in men or women. Background: Drawing the word-cloud is to show the most discussed genes in previous reports. Using molecular profiling approach to build a gene signatures model may provide a better way to predict lung adenocarcinoma patients’ prognosis and overall survival. The aims of this thesis were: 1. Gene expression signatures related to lung adenocarcinoma overall survival or prognosis. 2. A better approach to deal with patients’ clinical and pathological data in estimating a risk score and building up a model. 3. The gene differences between previous papers and clinicopathological results. Methods: Extract abstracts from the PubMed website, and establish a gene library. Plotting the word-cloud could quickly summarize the gene frequency results. 4 cohorts included 443 patients RNA microarray data were from National Cancer Institute Director’s Challenge Consortium (DCC). We separated 4 cohorts into training and testing datasets. The other 2 testing datasets were collected from the publicly available published data from Gene Expression Omnibus (GEO) database. Another RNA sequencing dataset was from The Cancer Genome Atlas (TCGA). Results: The top 5 highly-discussed genes were EGFR, KRAS, TP53, ALK, and AKT1. We used 3 different ways to take part patients into groups and tested different combinations in either cutoff point or coefficient estimation. 2 of these combinations obtained good results. One of the results was to separate each patient’s gene expression into high- and low-levels by median, and re-estimate the risk score coefficient and risk score model by maxstat. All the testing datasets had significant relationship between gene signature expressions and overall survival or prognosis (P < 0.0001, P = 0.00013, P = 0.00015, P = 0.011, P < 0.0001, P = 0.0017; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 and TCGA_LUAD dataset in order). The other one of the results was not to separate patients into high- and low-levels first, but performed continuous variable Cox regression analysis. No matter using median or maxstat, the risk score model showed significance in gene signatures expression and overall survival or prognosis (risk score in median: P < 0.0001, P = 0.024, P < 0.0001, P = 0.0074, P < 0.0001, P = 0.00012; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 and TCGA_LUAD dataset in order), (risk score in maxstat: P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001, P < 0.0001; CAN/DF & MSK, CAN/DF, MSK, GSE14814, GSE13213 and TCGA_LUAD dataset in order). Conclusions: Overall the last method should be the better way for further studies in mention to patients’ gene expression and clinical characteristics. 17 gene signatures (SP1, E2FF3, SPP1, GSTP1, ALK, KIF5B, CCL2, RHOB, AXL, GATA3, RET, CCL19, FGFR4, POU5F1, ID1, TWIST1, ROR1) could predict survival or prognosis in lung adenocarcinoma patients.