Learning algorithms for predicting HIV-1 phenotypic drug resistance

Abstract

The limitation of HIV treatment is the decrease of the viral sensitivity to the drug that is called drug resistance. The cause of drug resistance is from the mutations in the reverse transcriptase and protease enzymes of HIV. Thus, resistance testing plays an important role in HIV treatment. In the medical area, there are two methods of resistance testing: genotyping and phenotyping. The advantages of genotypic testing are faster and cheaper than phenotyping. On the other hand, the results of phenotypic method are easier to interpret than genotypic testing. This thesis applied four learning algorithms, which are the Support Vector Machine (SVM), the Radial Basis Function Network (the RBF network), k-Nearest Neighbor (k-NN), and Classification Based on Association (CBA), to construct the models for classifying HIV-1 drug resistance from HIV-1 genotypic data. Further, the predictive behavior of each classification model was studied. Finally, a new dynamic classifier combination method was proposed to construct the composite classifier from these single models. The predictive performances of the learning algorithms were compared with two online drug resistance prediction systems: HIVdb and Geno2Pheno. Our experimental results demonstrated that all learning algorithms yielded the higher average accuracy than that of the online systems. To evaluate the predictive performance of the proposed dynamic classifier combination method, we compared the accuracy with two classifier combination methods which are majority voting and Naïve Bayes. The results showed that our proposed method provided the best average predictive performance.