Prediction of protein functional sites by multivariate analysis of amino acid sequences

为了从蛋白质的氨基酸序列中预测蛋白质的功能位点，我们发展了多元分析等方法，并建立了预测数据库。我们的多变量分析方法的出发点是用一系列反映氨基酸残基的各种生物物理化学方面的数值来表示氨基酸序列。为此，我们通过收集已公布的疏水性和其他性质的数据，组织了一个氨基酸指数数据库。由于许多已报道的指数高度相关，我们进行了聚类分析进行分组。然后，利用判别分析，我们设计了从一组从氨基酸序列数据中定义的众多变量中选择表征功能位点的重要变量的程序。这些方法被应用于蛋白质二级结构片段的预测以及糖基化和磷酸化位点的预测。为了预测抗原性决定位点，我们组织了一个数据库，其中包含已发表的多肽片段和NBRF蛋白质序列数据库中相应条目的交叉引用。然而，我们的变量选择程序并没有产生令人满意的预测。由于在多变量分析中仅用变量来表示序列特征是一个严重的限制，因此我们研究了更灵活的方法。因此，我们应用人工智能方法，开发了一个专家系统。专家系统更有优势，因为它可以结合各种观察结果，包括多变量分析方法的结果。我们用这种专家系统的方法研究了预测细胞中蛋白质转位位置的问题。总而言之，这里开发的多变量分析方法本身就是有用的工具，但当它们与其他方法结合时，它们可能会更有效。专家系统似乎最适合实际应用。

In order to predict functional sites of proteins from their amino acid sequences, we have developed multivariate analysis and other methods and constructed databases for prediction. The starting point of our multivariate analysis method is to represent the amino acid sequence by a series of numerical values reflecting various biophysicochemical aspects of amino acid residues. For this purpose, we organized a database of amino acid indices by collecting published data for hydrophobicity and other properties. Since many of the reported indices were highly correlated, we performed a cluster analysis for grouping. Then, using discriminant analysis, we designed procedures to select important variables characterizing functional sites from a set of numerous variables defined from amino acid sequence data. The procedures were applied to the prediction of protein secondary structure segments and also to the prediction of glycosylation and phosphorylation sites. For the prediction of antigenicity determining sites, we organized a database with cross references of published peptide fragments and corresponding entries of the NBRF protein sequence database. However, our variable selection procedure did not produce satisfactory prediction. Because it was a severe limitation to represent sequence characteristics only by variables for multivariate analysis, we investigated more flexible methods. Thus, we applied an artificial intelligence method and developed an expert system. An expert system is more advantageous because it can incorporate various observations including results from multivariate analysis methods. We investigated the problem of predicting protein translocation sites in cells with this expert system approach. In summary, the multivariate analysis methods developed here are useful tools by themselves, but they can be more effective when combined with other approaches. Expert systems seem most suitable for practical applications.