A Stacking-Based Classification Approach: Case Study in Volatility Prediction of HIV-1

Human immunodeficiency virus type 1 (HIV-1) is eminent among chronic viruses for the vast number of therapeutics that exist for it. However, a hurdle to a promising long-term antiviral therapy is the error-prone replication of the viruses. The occurrence of mutations in some patients may result in resistance against medications. As a result, this can lead to increased morbidity and the likelihood of transmission to other individuals. Thus, the dissemination of such impervious mutants is of deep concern. In this study, we proposed a stacking-based classification technique to predict the absence or presence of variance in amino acid sequence of the envelope glycoprotein (Env) of HIV-1 based on the sequence variance of the positions within a specific neighborhood. For this purpose, we used sequence data from HIV-1-infected patients that describe the in-host variance in amino acid sequence (volatility) at each position of the Env protein. We tested the method on 4 different datasets, each corresponding to a specific position on Env. We compared the method with the performance of individual classifiers that have been incorporated into the algorithm as the base learners. We utilized a multi-layer perceptron model as the meta-learner in the second stage. Using the proposed method, we observed improvement in the classification metrics for all cases.