Avaliação de parâmetros de dados para construção de modelo classificador de tropismo de HIV-1

Abstract

AIDS is a disease of worldwide importance, caused by the HIV-1 virus. Of the several existing subtypes, the most prevalent are the B and C subtypes. Although it has no cure, several drugs have been developed over time to reduce its spread in the body. For example, Maraviroque® administration requires determining that the virus has tropism by the CCR5 receptor. There are highly accurate phenotypic tests, but they are very expensive and not agile for use in clinical routine. As an alternative, artificial intelligence models were developed to determine tropism by observing the sequence of 35 amino acids from gp120 region V3 of the virus. The models face difficulties in classifying correctly non-R5 tropism virus. In this work, we evaluated automatic variable selection and data balancing steps for classifiers performance. We used the random forest algorithm to develop separately trained models with 1,622 sequences of subtype B and 560 sequences of subtype C. The models were compared with the already established geno2pheno and T-CUP 2.0 classifiers. For subtype B, the AUC of all models presented values close to 0.95 and presented performance parity with the established predictors. For subtype C, the models presented AUC variants, but with higher performance than the established classifiers. The models presented positions that, despite having little variability, proved to be particularly important for the model. It was concluded that data balancing did not bring improvements and the selection of variables is a desirable step, but it should be performed considering previous information obtained empirically.