利用動態路徑分析法進行基因調控網路鑑別與修復

Abstract

本研究延伸路徑分析演算法(Path Analysis)，並引用微陣列晶片之動態資料，應用於資料庫中基因調控網路之鑑別上，以提供使用者判定基因調控網路可信度之參考。此外本研究亦利用演算法所提供的模型修正指標，適度地修正缺失的網路。 貝式模型、布林模型、結構方程模式與微分方程模型為現今較常見的基因調控網路模型。這些模型的目的係重建基因調控網路。本研究與上述方法不同之處在於著重於現有基因網路的鑑別。方法為根據KEGG的靜態網路產生多組候選的靜態網路，並拓展成N階動態網路。再計算這些動態網路之四組網路評估指標，以獲得最佳評估指標數目最多之網路，將之認定為最佳網路。 為了驗證本方法的可行性，本研究初步擷取十組基因調控網路進行鑑別。所採用的基因調控網路與微陣列晶片資料分別取自於KEGG與NCBI資料庫。網路為cell cycle –yeast 之子網路，包括regulation of autophagy 、MAPK signaling networks，資料為Segal 等人所提供的微陣列晶片資料。此外，為了進一步的驗證本方法的可靠性，本研究也與SSEM演算法以及貝式建模相比較。 實驗結果顯示，十組測試基因調控網路中，共有七組基因調控網路被判定為最佳網路，鑑別率達70%。另在針對該七組最佳網路，分別產生一些缺陷網路進行修正，其修正率為43%。此外，本研究方法也與SSEM演算法及貝式建模比較，針對相同網路，本研究方法皆有較佳的網路鑑別性，對於各網路之有向性連結之判斷準確性至少達60％以上。

The study expands Path Analysis (PA), and adopts microarray time course data to identify gene regulatory networks (GRNs). It provides users degrees of confidence on GRNs in databases. In addition, defective networks can be modified based on modification indices in PA. A couple of approaches, such as Bayesian, Boolean, structural equation modeling and differential equations model, are used for the reconstruction of gene regulatory networks in databases. We generate several alternative networks as candidates from original networks in KEGG database for comparison. Furthermore, the static networks are expanded to dynamic form (multiple orders). Finally, path analysis may suggest the best one in the network pool based on various performance indices. In other words, this approach can evaluate the existing networks in gene networks databases and provide users degrees of confidence on each network. The gene regulatory networks are form KEGG in this study, including sub-networks of cell cycle –yeast, e.g., regulation of autophagy and MAPK signaling networks, and corresponding microarray time course data are adopted from NCBI database. Furthermore, we compare our approach with SSEM algorithm and dynamic Bayesian model method. Seven out of ten original GRNs in KEGG are ranked the best networks by our approach, and 43 percent of defective networks generated from seven best original GRNs can be correctly modified. Besides, we obtain better results than SSEM algorithm and the dynamic Bayesian model method do to the same networks. The true positive rate on the directed links of networks is at least 60 percent.