Predicting HIV drug resistance from genomic data using deep learning.
Vinit Jadhav (Supervised by Dr. Daniel Hulme & Dr. Raman Gangakhedkar)
Abstract :
Antiretroviral therapy is the combination of several antiretroviral drugs to reduce the rate at which the Human Immunodeficiency Virus (HIV) multiplies in the human body. HIV however mutates to develop resistance to the antiretroviral drugs and hence hinders the success of antiretroviral therapy. Genotypic and phenotypic tests are used to measure the susceptibility of a particular strain of HIV to various antiretroviral drugs. While genotypic tests look for changes in the HIV genetic sequence compared to the wild type HIV, phenotypic tests provide an accurate measure of antiretroviral drug resistance in a controlled environment. Phenotypic tests are however complex, expensive and time consuming as compared to genotypic tests. In this study 20357 pairs of HIV genomic sequences and their corresponding phenotypic resistance values for 17 antiretroviral drugs were used to construct deep learning models to predict phenotypic drug resistance from genomic data. The accuracy of these models was measured using 10-fold cross-validation. Average prediction accuracy of 90.50% was obtained for eight protease inhibitor antiretroviral drugs. The average accuracy reduced to 84.15% for three non-nucleoside reverse transcriptase inhibitor antiretroviral drugs and further to 81.30% for six nucleoside reverse transcriptase inhibitor antiretroviral drugs. A software tool was developed to harness the predictive capability of these models in predicting the phenotypic drug resistance from the genomic information of an individual’s HIV strain. The tool is now being used by National AIDS Research Institute of India to predict the phenotypic drug resistance of the 17 antiretroviral drugs included in this study and help formulate effective drug regimens to reduce the viral load in infected individuals and help them lead better healthy lives.