利用Stacking的方法預測B細胞表位

Abstract

B細胞表位是在抗原抗體的結合反應中抗原參與結合的部位。在預防醫學中，有效的識別B細胞表位對於設計疫苗和藥物是相當重要的。近年來，由於資料探勘與機器學習的發展，B細胞表位預測研究得到了很大的進步與廣泛的應用。目前預測B細胞表位的方式有針對蛋白質序列和蛋白質結構的預測工具，而基於不同的預測模型與蛋白質屬性，這些預測工具會有不同的特性和預測效能。由於各家的預測方法都不盡相同，且從統計學上的相關性分析，我們發現這些預測工具彼此之間的相關性是低的，其意味著它們有互補的可能。而為了有效運用其互補特性，我們提出一種以stacking為架構的meta-learning方法來預測表位，透過結合底層的預測工具與屬性，在獨立測試資料下，其結果顯示我們得到了比任一預測工具更好的預測結果，並利用多個實驗驗證出我們的方法能結合各預測工具的優點，透過互補的關係有效地預測出B細胞表位。

B-cell epitopes are part of the antigen involved in the antigen-antibody interaction. Effective identification of B-cell epitopes is crucial to the design of vaccines and the research in immune systems. Many B-cell epitope prediction tools adopting the data mining and machine learning techniques have been developed. These tools have different characteristics and performances because they were designed based on various prediction models and protein properties. To exploit the synergy among the tools, we propose a meta-learning approach for epitope prediction that applies stacked generalization. The ablation study has verified the interactions and complementary strengths among the base predictions tools used in the proposed stacking classifiers. By combining base prediction tools and features in a hierarchical architecture, stacked generalization has demonstrated the superior performances when compared with other existing prediction methods in the extensive experiments.