Classification of G-protein coupled receptors

Abstract

We developed programs for classifying G-protein coupled receptors (GPCRs) to subfamilies, using several methods. The methods are self-organizing map (SOM), nearest neighbor method, and a feedforward artificial neural network. These programs were trained using GPCR dataset and predicted classification of given GPCR sequences as one of fifteen conventional subfamilies. All of three methods employ the same representation of GPCR sequence, that is, auto-cross covariance derived from z-scales that were computed as principal components of 26 physicochemical properties of amino acids. All of them gave results with high precision for non-aligned GPCR amino acid sequences. Specifically, for test datasets, the nearest neighbor gave the highest precision (99.68%), SOM also gave high precision (97.96%), and the feedforward network gave 92.50%. The output plane of the SOM consists of 50x50 neurons, and learn to respond individually to distinct input patterns so that nearby neurons respond to similar inputs. The feedforward network consists of three layers with 12 hidden layer units and 15 output layer units. Two-fold cross-validation were carried out by selecting 30 training data sets from GPCR dataset. The results suggest that the methods could become useful tool for non-aligned sequence classification.