Identifying significant features in HIV sequence to predict patients' response to therapies

The Human Immunodeficiency Virus (HIV) is a retrovirus that attacks the human immune system reducing its effectiveness. Combinations of antiretroviral drugs are used to treat the infection by HIV. However, the high mutation rate in the HIV virus makes it resistant to some antiretroviral drugs and leads to treatment failure. Nowadays, there are computational methods based on machine learning that try to predict the patients' response to therapies. In this bioinformatics study we deal with data preprocessing techniques to find significant features in HIV sequences that can be interesting for the prediction of patients' short-term progression. Experiments were conducted trough four classification methods using datasets composed by different sets of attributes. Classifiers trained with a dataset including solely viral load, CD4+ cell counts and information about mutations in the viral genome achieved accuracies ranging from 50.29% to 63.87%. Nevertheless, the addition of attributes (antiretroviral drug resistance levels, HIV subtype, epitope occurrence and others) in the dataset has improved the accuracy of the classifiers in almost all tests executed in this work, indicating its relevance to the prediction task discussed here.