基於支持向量機器方法之蛋白質β-turn預測

Abstract

本研究是利用支持向量機器的方法來預測蛋白質中β-turn的位置。在僅有蛋白質之胺基酸序列的情況下找尋有用的特徵向量，並將這些資訊輸入支持向量機器中，以此方法來預測蛋白質中哪些殘基會形成β-turn。本研究使用426條非同源蛋白質，做7倍的交叉認證以驗證預測的準確率。由結果發現除了前人研究提及的多重序列比對及二級結構資訊外，殘基暴露於溶劑的程度亦可提供有用的資訊；而胺基酸的體積及親水程度則對β-turn的預測無明顯助益。 本研究整合了多重序列比對所產生的位置加權矩陣，二級結構預測資訊，以及殘基暴露於溶劑之程度的預測等三種特徵向量，則總準確率可達79.6%，MCC值可達0.48，皆高於其他β-turn預測方法。

In this study, we use support vector machine approach to predict β-turns in protein. With only the information of protein sequence, we try to find useful feature vectors based on amino acid, and import the information to SVM to predict which residue would be in β-turn. We use 426 non-homologous proteins as dataset, and 7-folded cross validation to examine the prediction performance. In addition to multiple sequence alignment and secondary structure information, we found that relative solvent accessibility could also provide useful information in β-turn prediction. In this work, import multiple feature vectors of multiple sequence alignment information (PSSM), secondary structure prediction, and relative solvent accessibility prediction, the Qtotal could reach 79.6% and the MCC value is 0.48. Both these two measure performance are better than other previous methods.