New Method Combining Feature Weighting and Feature Selection for Protein Classification

The primary biological data structure is represented as a string of characters. Several issues in bioinformatics are interested in handling this type of data. Such as the alignment of biological sequences, the 2D/3D structure prediction and detection of anomalies in the genes. Our work is integrated into the framework of knowledge discovery from biological data (KDBD) process and specifically the pre-processing phase. We are interested in protein classification using features extracted from their primary structures. Data mining techniques require a data matrix: Individual X Features (sequence X n-grams, in our case), so we understand the importance of the type of features used, their number and their weighting in the protein classification problem. In this paper we first present a new method for feature weighting based on dynamic programming of the Smith/Waterman local alignment algorithm, then a new method of feature selection. We used the SVM classifier to calculate the error rates. The results have shown the effectiveness of this work, especially by comparing it to previous works.