Evaluation of eight different bioinformatics tools to predict viral tropism in different human immunodeficiency virus type 1 subtypes

Abstract

Human immunodeficiency virus type 1 (HIV-1) tropism can be assessed using phenotypic assays, but this is quite laborious, expensive, and time-consuming and can be made only in sophisticated laboratories. More accessible albeit reliable tools for testing of HIV-1 tropism are needed in view of the prompt introduction of CCR5 antagonists in clinical practice. Bioinformatics tools based on V3 sequences might help to predict HIV-1 tropism; however, most of these methods have been designed by taking only genetic information derived from HIV-1 subtype B into consideration. The aim of this study was to evaluate the performances of several genotypic tools to predict HIV-1 tropism in non-B subtypes, as data on this issue are scarce. Plasma samples were tested using a new phenotypic tropism assay (Phenoscript-tropism; Eurofins), and results were compared with estimates of coreceptor usage using eight different genotypic predictor softwares (Support Vector Machine [SVM], C4.5, C4.5 with positions 8 to 12 only, PART, Charge Rule, geno2pheno coreceptor, Position-Specific Scoring Matrix X4R5 [PSSMX4R5], and PSSMsinsi). A total of 150 samples were tested, with 115 belonging to patients infected with non-B subtypes and 35 drawn from subtype B-infected patients, which were taken as controls. When non-B subtypes were tested, the concordances between the results obtained using the phenotypic assay and distinct genotypic tools were as follows: 78.8% for SVM, 77.5% for C4.5, 82.5% for C4.5 with positions 8 to 12 only, 82.5% for PART, 82.5% for Charge Rule, 82.5% for PSSMX4R5, 83.8% for PSSMsinsi, and 71.3% for geno2pheno. When clade B viruses were tested, the best concordances were seen for PSSMX4R5 (91.4%), PSSMsinsi (88.6%), and geno2pheno (88.6%). The sensitivity for detecting X4 variants was lower for non-B than for B viruses, especially in the case of PSSMsinsi (38.4% versus 100%, respectively), SVMwetcat (46% versus 100%, respectively), and PART (30% versus 90%, respectively). In summary, while inferences of HIV-1 coreceptor usage using genotypic tools seem to be reliable for clade B viruses, their performances are poor for non-B subtypes, in which they particularly fail to detect X4 variants.