基於結合模序配對探勘之蛋白質交互作用預測

Abstract

蛋白質交互作用是生物體執行功能的基礎。透過蛋白質交互作用的研究，可以理解細胞運作的基本原理，進而開發、設計藥物，並針對疾病進行治療，因此了解蛋白質的交互作用不論在基礎或臨床的研究上都非常重要。由於以生物實驗驗證蛋白質交互作用需要耗費過多的時間與金錢，因此開發計算方法輔助，減少研究蛋白質交互作用消耗的資源，為現今系統生物學研究的首要工作之一。 計算方法中，「以模序為基礎 (motif-based)」的方法利用樣式探勘 (pattern mining) 演算法找出結合模序 (binding motif)，再透過樣式比對 (pattern matching) 來預測蛋白質交互作用。其優勢在於只需要蛋白質序列即可進行分析與預測，而且可以得知蛋白質產生交互作用之區域。傳統以模序為基礎之方法大多是利用功能序列在同源家族蛋白質中具有高度保留性的原理，針對同源蛋白質序列進行樣式探勘，試圖挖掘出結合模序。不過以此方法所找出的具保留性之序列不一定與交互作用有關，可能是用來維持結構或其他功能。為了克服這個問題，近年來有學者提出應該針對具備相同交互作用機制的蛋白質進行樣式探勘，並提出透過偵測全對全作用網路 (all-versus-all interaction network) 的方式來蒐集探勘樣式所需之蛋白質。然而，近年來已發表的研究中並沒有將這套方法所探勘之結合模序運用於蛋白質交互作用之預測，以精確評估這種機制的效果以及採用不同樣式探勘演算法的效應。 本研究提出以一個先進的蛋白質序列樣式探勘演算法為基礎，搭配全對全作用網路，以進行蛋白質交互作用預測。此蛋白質序列樣式探勘演算法之特色在於能夠探勘出由數個短樣式組合而成的長樣式，這種特性相當符合蛋白質結合介面 (binding interface) 通常是由許多序列片段所組成之特性。本研究的結果顯示，樣式探勘演算法對於結合模序挖掘與蛋白質交互作用預測影響皆非常顯著。本研究所採用的先進演算法，不論結合模序之正確性以及蛋白質交互作用預測之準確率都較其他方法來得優越。

Protein-Protein interactions (PPIs) are essential to various biological functions in living organisms. Studying PPI not only provides critical clues for understanding how a cell operates but also may lead to development of advanced diagnoses and therapies. In this regard, as it requires huge amounts of time and resources to confirm protein-protein interactions with molecular biology experiments, design of computational approaches to predict possible protein-protein interactions is of scientific significance for advances in systems biology. One existing approach to predict protein-protein interactions is based on the binding motifs extracted by pattern mining algorithms. Motif-based approaches are favored by biologists who want to conduct in-depth analyses on how the concerned proteins interact, instead of just knowing whether these proteins interact with each other or not. With respect to motif-based prediction of protein-protein interactions, there exist two major categories of approaches. One category of approaches simply resorts to analysis of the polypeptide sequences, while another category of approaches further refers to the tertiary structures of proteins. As the availability of the tertiary structures of proteins is still limited to certain groups of proteins, sequence-based approaches are more generally applicable. The conventional motif-based approaches extract binding motifs through identifying evolutionally conserved regions in polypeptide sequences. However, evolutional conservation is just a necessary condition and is not a sufficient condition for presence of interaction sites. Certain regions in a protein chain may be conserved in order to maintain a conformation. Therefore, in recent years, researchers have proposed a novel approach to identify protein-protein interaction motifs through analysis of interaction networks. Nevertheless, latest studies did not report a comprehensive analysis on the quality of the interaction motifs identified, let alone the effects with alternative pattern mining algorithms. The study reported in this thesis has followed the recent development and has employed a state-of-the-art pattern mining algorithm to deliver superior performance in identifying protein-protein interaction motifs. The most distinctive feature of the pattern mining algorithm employed in this study is its capability in identifying patterns composed of several short gapped segments. Experimental results reveal that the predictor designed in this study really outperforms the predictors that incorporate other pattern mining algorithms.