HIVM, An Application for Prediction of HIV Drug Resistance using Support Vector Machine Algorithm


Edward Otis Johnson III

Oral Defence Date: 

Friday, December 1, 2006 - 17:30


TH 935


5:30 PM


Prof. Petkovic, Prof. Singh, Prof. Buturovic


Drug resistance is a critically important open problem in HIV treatment today. Current laboratory methods for determining drug resistance have many shortcomings such as high cost, long waits for results, and inability to predict resistance in many circumstances. We developed a pattern classification software, hivm, that overcomes many of these obstacles by using a classification algorithm known as Support Vector Machine (SVM). hivm is a cross-platform, open-source application written in C++ using test driven development techniques. The idea is to train the learning machine (SVM) using known patterns of patient drug resistance for various virus genotypes. The algorithm models the relationship between drug resistance and virus genotype. The learning process creates an optimal model and estimates its accuracy. The model can then be applied to classify previously unseen patients (represented by their viral genotype signatures) into the response/no-response categories. In order to adapt the problem domain data to SVM compatible format, we performed pairwise comparisons of the HIV amino acid sequences using a variation of the Smith-Waterman algorithm known as Local Alignment. We then performed SVM model selection and validation using phenotypic drug resistance data provided by the Stanford HIV Drug Resistance Database. The results show promise for several of the drugs we tested. However, for hivm's true effectiveness to be measured, we need access to patient response datasets from clinical trials. This data was unavailable for use during implementation of the project. The long-term goal of the hivm project is prediction of patient response to HIV drug therapy given a genotype of the virus found in the patient. Ultimately, this would lead to a diagnostic test which may aid a physician in prescribing a most efficient HIV drug regimen for a particular patient.

Edward Otis Johnson III

Support Vector Machine, Machine Learning, HIV, Drug Resistance