Construction of a support vector machine (SVM) for accurately detecting domain boundaries

预测新蛋白质序列的结构域具有重要的实际意义。一个重要的应用领域是开发计算机辅助技术，以低成本识别大规模功能和结构蛋白质组学的新蛋白质结构域靶点。传统的计算机辅助方法依赖于与域数据库（如Pfam、Prosite或SMART）的序列相似性。然而，需要独立于域数据库工作的方法来检测新域。这种方法需要仅从目标蛋白质的氨基酸序列中包含的信息来预测结构域区域。在这个项目中，我们报告了一个支持向量机（SVM），识别域连接器，这是两个结构域分开的环。我们构建了一个基于SCOP和CATH结构域边界定义的多结构域蛋白质数据库。首先，我们选择了不形成域间相互作用的域，如通过域间VdW、H键和SS键的存在所定义的，并且是独立可折叠的。从这个集合中，我们进一步选择了形成DSSP定义的环路的域边界。满足这两个条件的域边界被称为连接器，用于训练和测试SVM。我们开发了一个基于SVMlight的域连接器预测（DLP-SVM）。敏感性为46.8%，特异性为57.1%。这些值分别超过5.1%和6.8%，高于以前报道的方法。DLP-SVM可在以下网址免费获得：http://www.tuat.ac.jp/~ domserv/cgi-bin/DLP-SVM. cgi 2008年6月，日本蛋白质科学学会向博士生Teippei Ebina颁发了旅行补助金，用于在澳大利亚凯恩斯举行的PRICIPS 2008会议上介绍这项研究。

The prediction of structural domains in novel protein sequences is becoming of practical importance. One important area of application is the development of computer-aided techniques for identifying, at a low cost, novel protein domain targets for large-scale functional and structural proteomics. Traditional computer-aided methods rely on sequence similarity to domain databases such as Pfam, Prosite or SMART. However, methods that work independently from domain databases are required for detecting novel domains. Such methods need to predict domain regions solely from the information contained in the amino acid sequence of the protein of interest. In this project, we report a Support Vector Machine (SVM) that identifies domain linkers, which are loops separating two structural domains. We constructed a multi-domain protein database based on SCOP and CATH domain boundary definition. First, we selected domains that do not form inter-domain interactions, as defined by the presence of inter-domain VdW, Hbonds and SS-bonds, and are independently foldable. From this set, we further selected domain boundaries that form loops as defined by DSSP. Domain boundaries that fulfilled both conditions were called linkers and used for training and testing the SVM.We developed a domain linker prediction (DLP-SVM) based on SVMlight. The sensitivity and the specificity were, respectively, 46.8% and 57.1%. These values are over 5.1 and 6.8%, respectively, higher than previously reported methods. DLP-SVM is freely available at : http://www.tuat.ac.jp/~ domserv/cgi-bin/DLP-SVM.cgiA travel grant from the Protein Science Society of Japan was awarded to PhD student Teippei Ebina for presenting this research at the PRICIPS 2008 conference held in Cairns, Australia, June 2008.