Determination of highly effective attributes in fold level classification of proteins

Abstract

In this paper it is aimed to determine which of the protein features or attributes is the most significant for classification of proteins according to their folds. Proteins in the database used in this study are represented by six feature groups called attributes and by a 125-dimensional feature vector. The representation of proteins with very high dimensional vectors such as 125 causes increasing computational load of the classification process and extending the process time. In this study “dimension reduction” solution is offered for this negative situation. Hence, with two different approaches, the features and attributes having high classification performance are determined. In the first approach, which attribute gives higher performance is determined by testing separately each of the six attributes. In the second approach, the most significant of the 125 features are determined using Divergence Analysis method. In this study, a classic classifier KNN (K-nearest neighbor) and artificial neural network models GAL (Grow and Learn) and SOM (Self-Organizing Map) networks are used as classifier and classification performance is analyzed for reduced dimension datasets.