An Ensemble Learning Algorithm for Predicting HIV-1 Protease Cleavage Sites

Abstract

Understanding the substrate specificity of human immunodeficiency virus 1 (HIV-1) protease plays a significance role in the design of effective HIV-1 protease inhibitors. During the past two decades, a variety of machine learning models have been developed to predict the existence of HIV-1 protease cleavage sites. However, since the acquisition of cleavable octapeptides requires expensive and time-consuming experiments, and uncleavable octapeptides are usually generated by artificial strategies, the number of cleavable octapeptides in the existing data set is far less than that of uncleavable octapeptides. This phenomenon of unbalanced datasets may cause the prediction performance of the classification model to be inaccurate. In this work, we combine the idea of asymmetric bagging and the support vector machine (SVM) classifier to propose an ensemble learning algorithm, namely AB-HIV, for an effective treat the dataset imbalance problem in predict HIV-1 protease cleavage sites. In order to make full use of the information of the substrate sequence, AB-HIV uses three different coding schemes (amino acid identities, chemical properties and variable length coevolutionary patterns) to construct the feature vector. By using asymmetric bagging to resample a set of balanced training subsets from the training set, and then a set of SVM classifiers can be built for integration to complete the prediction task. Experiments on three independent benchmark datasets indicate that the proposed ensemble learning method outperforms the existing prediction methods in terms of AUC, PR AUC and F-measure evaluation criteria. Therefore, AB-HIV can be regarded as an effective method to deal with the dataset imbalance problem in predict HIV-1 protease cleavage sites.