基于一种动态特征选择融合算法的蛋白质结构类预测

本文根据氨基酸理化性质，基于氨基酸组成成分与自相关函数相结合特征提取法从非同源蛋白质序列中提取七个特征集，采用局部正确性的动态特征选择算法进行多特征组合来预测蛋白质结构类，并与各个特征集进行了比较。结果表明，DFS_LA算法的预测总精度较各个特征集均有不同程度的提高。Jackknife检验下，DFS_LA算法的预测总精度为82．80％，比COMP特征集提高8．91％；独立测试检验下，DFS_LA算法的预测总精度为86．67％，比COMP特征集提高11．67％。这说明DFS_A算法可有效提高结构类预测精度，多特征组合能在一定程度上更多地反映蛋白质的空间结构信息。

Prediction of Protein Structural Classes Based on one DFS Algorithm

According to physicochemical properties of amino acid,the approach of feature extraction of incorporating amino acid composition with different auto-correlation functions has been introduced to predict non-homologous protein structural classes and seven feature sets could be gained.We have combined multiple features using Dynamic Feature Selection with Local Accuracy(DFS-LA) algorithm.The comparisons of the predictive results from combination of multiple features and each parameter data set show that the total predictive accuracy are remarkably improved by using DFS-LA algorithm.In jackknife test,the total predictive accuracy using DFS-LA algorithm is 82.80%, which is 8.91 percentile higher than that of COMP parameter data set.In independent test,the total predictive accuracy using DFS-LA algorithm is 86.67%,which is 11.67 percentile higher than that of COMP parameter data set.These results show that the predictive accuracies of protein structural classes can be effectively improved by using DFS-LA algorithm.To some extent,combination of multiple features can reflect more protein spatial information.