HIV acts by attacking the immune system and gradually destroying the TCD4+ defense cells. Without adequate treatment, the carriers develop the most severe form of the infection, AIDS, when the patient can be afflicted by opportunistic diseases that inevitably lead to death. Fortunately, with the advent of the highly active antiretroviral therapy (HAART), the mortality of people with HIV is decreasing. However, mutations can occur in the genotype of the virus, generating drug-resistant phenotypes. Computational methods have been used to predict whether a given strain is drug-resistant, and to which drugs this resistance occurs, thereby increasing the chances of success of the prescribed treatment regimen. However, these methods are not always accurate in their task. In this context, by applying Feature Selection methods and estimating Decision Tree models, we investigated patterns in Protease and Reverse Transcriptase enzyme sequences, as well as in patients' clinical data, which can lead to correct or incorrect computational prediction. As a result, we identified 21 features that are highly informative, 11 which tend to lead the methods to error, and eight that present both behaviors simultaneously, being able to predict the patient's response to therapy and at the same time may lead the predictor's methods to failure.