Feature Selection on Pectin Lyase-Like Enzyme Using Computational Methods

The employment of feature selection algorithms (FSAs) prior to classification has become a necessity due to an enormous growth of public sequence databases and the nature of high dimensionality in protein sequences. This paper provides a comparative framework on four multivariate FSAs for finding minimal feature subsets prior to classification of a protein function from a pectin lyase-like superfamily. The comparative studies for these FSAs are based on four criteria: the accuracy, the area under ROC graph (AUC), the selected features, and the modelling time taken. Classification was performed on a reduced feature set using three state-of-the-art machine learning classifiers, Support Vector Machines, Naïve Bayes and Decision Tree, on the dataset with and without FSAs. Our results suggest the importance of FSAs in improving the classification accuracy and reducing the modelling time.