Support Vector Machines for Predicting Protein Structural Classes via Pseudo Images Derived From Amino Acid Sequences

Abstract

SVM is one of the most widely used and powerful classification algorithms to predict protein structural classes. Via radial base functions, SVM maps the linearly non separable input data to a higher dimensional space where it is almost separable by a hyperplane. First using the training data, six hyperplanes that separate pair wise the four classes training data are constructed. Then the expertise of these six SVMs genuinely aggregated to classify the testing data into four classes. The validation is performed by a boot strap technique. The 33 dimensional data that represents proteins of the data set is derived from pseudo images of proteins that stems from their amino acid sequences. In spite of the simplicity of the features, a Q3 accuracy around 75% is achieved.