以迴歸模型檢驗與肺腺癌預後相關之基因組

Abstract

肺癌在全世界的已開發國家都是癌症死亡率之首。即使被分類為早期的肺癌患者，術後復發率也高達30%以上。雖然之前已有許多以檢測基因表現模式為基礎而開發的基因組，宣稱可被用於預測預後；但這些基因組之間重複性極低，使得直接應用這些基因組於臨床治療上的可行性令人疑慮。此外過去在建立預測預後基因組時，經常沒有考慮基因本身的生物意義或基因間的互動關係，使得這些預後基因組其整體生物意義通常是難以解讀的。 為了突破此一困境，我們提出一個全新的方法，同時整合了基因組顯著分析以及存活分析兩種技術。不但將篩選對象由單一基因改為事前由生物傳導途徑和生物實驗結果所定義的基因組，並藉由Cox風險迴歸模型來分析基因組的表現模式和存活結果的關係。這項整合使得結果不僅僅是具有預測病患存活結果能力的基因組，並且能展現生物現象及生物機能和存活結果之間相關性。 本方法由三個演算法所組成：第一步是以兩個虛無假說的檢驗來篩選和存活結果有關的基因組，兩個虛無假說在檢驗後皆被拒絕的基因組才進入下一步分析。第二步是以群集分析，將組成成員類似的基因組分入同組，並從每一組中選出一個做為代表性基因組。第三步是分析這些代表性基因組之間的基因表現模式相似程度，並把結果畫成基因組互動網路。 和肺腺癌預後相關的生物機能可以被分為四大類：和細胞週期調控有關、和細胞能量利用方式變異有關、和轉譯後修飾有關以及和傷口組織癒合有關幾個部分。雖然這些生物機能過去都曾經被文獻報導過，在肺腺癌的惡化過程中扮演了關鍵角色，我們的基因互動網路更顯示這些功能是藉由協同運作而達成促進癌症組織生長的結果。 我們的方法不僅可用於臨床上的預後預測，還可以檢驗任何連續性的資料以及基因組表現模式間的關係是否顯著。藉由同時引入事前定義的基因組和回歸分析，不但可檢定基因組的表現是否和資料有關，還可闡明和這些結果有關的生物意義為何。

Lung cancer has been the leading cause of cancer related death worldwide. Hitherto, more than 30% of early stage lung cancer patients who received complete surgical resection died of relapse. Although lots of prognosis studies based on whole genome profiling had been published, little overlaps of the prognostic genes from different research groups made the utilization in clinical infeasible. Moreover, due to the biological meaning and the interactions of the prognostic genes were not taken into account in previous studies, explaining those results biologically and comprehensively was not easy. To overcome the bottleneck, we proposed a novel method based on gene set enrichment analysis and Cox-hazard regression model. The predefined gene sets which derived from biological pathways, in vitro, and in vivo experiments were utilized as prognosis targets instead of single genes. Cox-hazard regression model was then applied to assess the relation between the survival outcome and the enrichment of gene sets. The integration of predefined gene sets and Cox-hazard regression model provided not only the power of predicting survival outcome, but also the connection between biological functions and prognosis. Our method was composed of three algorithms: first, a two-step hypothesis testing procedure was applied to select gene sets associated with the survival outcome. Second, prognostic genes sets were clustered, and a representative gene set was selected from each cluster. Lastly, the similarities between the gene expression patterns of those representative gene sets were evaluated by kernel matrix, and the results were illustrated as gene set association networks. The biological functions associated with the survival outcome of lung adenocarcinoma can be divided into four categories: cell cycle regulation, energy dysregulation, post-translation modification, and wound healing process. Although they were all reported by previous literatures as essential functions in the progression of lung adenocarcinoma, our data indicated that they functioned in a coordinate manner. Not only for clinical prognosis as we demonstrated, our method can also be applied to assess the relations between other types of continuous variable and gene set. By incorporating predefined gene sets and the regression model, the gene sets associated with the testing variables and the biological interpretation connected with those gene sets could be illuminated by our method.