Improving prediction of interdomain linkers in protein sequences using a consensus approach

Abstract

A two pronged strategy, one involving consensus approach and the Multi Layer Perceptron (MLP) as the classifier and the other including physicochemical properties as additional features, is proposed and implemented here for improved prediction of interdomain linkers in protein chains. The software is tested on proteins of the CASP6 experiment in order to measure its prediction accuracy using three-fold cross validation. Finally, our consensus approach combines results of 28 different neural networks. We observe significant improvements of AUC scores by 9.4% on average in comparison to the corresponding most successful single artificial neural networks. © 2012 Springer-Verlag GmbH Berlin Heidelberg.