Insights on prediction of patients' response to anti-HIV therapies through machine learning

We collect data from the HIV Resistance Drug Database and, based on CD4+ and viral load measures, together with RNA sequences of the reverse transcriptase and of the protease of the virus, we design models using machine learning techniques MultiLayer Perception (MLP), Radial Basis Function (RBF), and Support Vector Machine (SVM), to predict the patient's response to anti-HIV treatment. In this work we applied the SMOTE Algorithm to deal with the enormous difference between the number of case and control samples, which was crucial for the accuracy of the models. Our results show that the SVM model proved more accurate than the other two, with a ROC curve area of 0.9398. We observe that, from 1000 patients, there are 646 samples for which the three methods delivered correct predictions. On the other hand, for 69 patients all three models fail. We analyzed the data for those patients more carefully, and we identified codons and properties that are important for a response/non-response result. Among the codons that our models identified, there are several with strong support from the literature and also a few new ones. Our analysis offers numerous insights that can be very useful to the prediction of patients' response to anti-HIV therapies in the future.