Using multi level nearest neighbor classifiers for g-protein coupled receptor sub-families prediction

Abstract

Prediction based on the hydrophobicity of the protein yields potentially good classification rate as compared to the other compositions for G-Proteins coupled receptor (GPCR's) families and their respective subfamilies. In the current study, we make use of the hydrophobicity of the proteins in order to obtain a fourier spectrum of the protein sequence, which is then used for classification purpose. The classification of 17 GPCR subfamilies is based on Nearest Neighbor (NN) method, which is employed at two levels. At level-1 classification, the GPCR super-family is recognized and at level-2, the respective sub-families for the predicted super-family are classified. As against Support Vector Machine (SVM), NN approach has shown better performance using both jackknife and independent data set testing. The results are formulated using three performance measures, the Mathew's Correlation Coefficient (MCC), overall accuracy (ACC) and reliability (R) on both training and independent data sets. Comparison of our results is carried out with the overall class accuracies obtained for super-families using existing technique. The multilevel classifier has shown promising performance and has achieved overall ACC and MCC of 97.02% and 0.95 using jackknife test, and 87.50% and 0.85 for independent data set test respectively. © Springer-Verlag Berlin Heidelberg 2007.