Protein structural class prediction based on multi-feature and Ma-Ada multi-classifier fusion

Abstract

Protein sequence feature and machine learning algorithm are two important aspects to determine the results of protein structural class prediction. In this study,we established 17-D and 57-D feature information sets through fusing the sequence information,physical and chemical information with the secondary structure information based on the k-word statistical frequency and the k-fragment distribution feature extraction method.By introducing Multi-Agent 's idea into Adaboost. M1 algorithm,a novel method for protein structural class prediction,called Ma-Ada multi-classifier fusion algorithm,was proposed,which fully utilized the information of the single classifier metric layer and the fusion of information among individual classifiers. Four protein datasets including Z277,Z498,1189,D640 were used to validate the performance of the Ma-Ada algorithm.Classification accuracies are 91. 3 %,96. 8 %,85. 3% and 87. 2 % with 57- D features,and 90. 6 %,95.8 %,84. 8 % and 88. 3 % with 17 D features on datasets Z277,Z498,1189 and D640,respectively. The experimental results show better.