predMultiLoc-Gneg: Predicting subcellular localization of gram-negative bacterial proteins using feature selection in gene ontology space and resolving the data imbalance issue

Several types of subcellular localization prediction methods have been proposed depending on various classification methods which produce different levels of accuracy. Most of these predictors aim to find the optimal classifier and very few of them consider the issue of simplifying the complexity of biological system. However, there are two important issues that can take place to simplify the complexity of prediction system before developing successful predictor: handling high dimension feature, and overcoming the challenge of large data imbalance in the training data. As a result, in this work, an efficient computational tool named predMultiLoc-Gneg has been constructed to predict the multi-label subcellular localization of gram-negative bacterial proteins by (1) selecting relevant GO (Gene Ontology) terms in order to create a GO subspace which reduces feature dimension in contrast to the whole GO space and extracting GO-based features for a protein by considering only these relevant GO terms, (2) developing a multi-label predictor using support vector machine (SVM) with resolving data imbalance issue. The experimental results demonstrate that the predMultiLoc-Gneg provides remarkably better performance in predicting protein subcellular localization for gram-negative bacterial dataset than the existing top predictors. The web server for the predMultiLoc-Gneg is available at http://research.ru.ac.bd/predMultiLoc-Gneg/