Improvement of HIV-1 coreceptor tropism prediction by employing selected nucleotide positions of the env gene in a Bayesian network classifier

This study aimed to develop a genotypic method to predict HIV-1 coreceptor usage by employing the nucleotide sequence of the env gene in a tree-augmented naive Bayes (TAN) classifier, and to evaluate its accuracy in prediction compared with other available tools. A wrapper data-mining strategy interleaved with a TAN algorithm was employed to evaluate the predictor value of every single-nucleotide position throughout the HIV-1 env gene. Based on these results, different nucleotide positions were selected to develop a TAN classifier, which was employed to predict the coreceptor tropism of all the full-length env gene sequences with information on coreceptor tropism currently available at the Los Alamos HIV Sequence Database. Employing 26 nucleotide positions in the TAN classifier, an accuracy of 95.6%, a specificity (identification of CCR5-tropic viruses) of 99.4% and a sensitivity (identification of CXCR4/dual-tropic viruses) of 80.5% were achieved for the in silico cross-validation. Compared with the phenotypic determination of coreceptor usage, the TAN algorithm achieved more accurate predictions than WebPSSM and Geno2pheno [coreceptor] (P