Poor Performance of Bioinformatics Programs for Genotypic Prediction of Coreceptor Usage of HIV-1 Group O Isolates

To the Editors: Coreceptor-specific pharmacotherapy requires reliable determinations of coreceptor usage of patients' HIV-1 isolates. We examined the performance of 4 major coreceptor prediction programs (geno2pheno[coreceptor], WebPSSM, SVM-SK, WetCat), which primarily utilize the envelope-V3 loop sequence, for HIV-1 O isolates. The specificity of the sequence-based prediction was low, ranging from 0% to 28.6% relative to the biologic coreceptor phenotype. Consequently, coreceptor usage in HIV-1 O-infected individuals should at present be exclusively assessed phenotypically until prediction programs have been educated and optimized for this divergent HIV-1 group. The reliable determination of HIV-1 coreceptor usage has recently become an issue of clinical importance because the repertoire of licensed antiretroviral drugs has been expanded by the CCR5 coreceptor antagonist maraviroc (celsentri). This new class of entry inhibitors has shown low-nanomolar activity in vitro against most R5 HIV-1 strains and primary isolates, including several HIV-1 group O viruses.1 Maraviroc has demonstrated high efficacy in HIV-1-infected individuals harboring R5 viruses.2 Coreceptor tropism testing is mandatory before prescribing coreceptor-specific drugs and useful for therapy monitoring. Thus far, treatment failure was mostly associated with lack of X4 virus detection before initiation of therapy and subsequent in vivo expansion of preexisting X4 viruses that were insensitive to the CCR5 inhibitor.3 Identification of such low-frequency X4 variants from HIV-1 group M in clinical samples can apparently be improved by “ultradeep” sequencing approaches.4 The most commonly used genotypic prediction programs utilize clonal sequence data of the third hypervariable (V3) loop of the HIV-1 envelope glycoprotein, are built around statistical learning algorithms, and are educated by data training sets. These bioinformatics prediction engines include simple amino acid charge rules, decision trees, neural networks, more advanced support vector machines (SVM), position-specific scoring matrices, and SVM-based approaches that integrate viral sequences with structural information of the envelope glycoprotein and clinical data.5 The specificity of coreceptor prediction for clinical HIV-1 M isolates by these V3 loop-based bioinformatics programs currently ranges between 83% and 97%5-8 compared with phenotypic, tissue culture-based assays which determine the biologic coreceptor tropism.9 Although such phenotypic assays are still commonly used for clinical samples, their reliability has recently been questioned based on interassay concordance analyses10 and genotypic predictions are rapidly improving in fidelity, efficiency, and cost.5 In contrast to the pandemic HIV-1 group M, infections with the highly divergent HIV-1 group O are found primarily in the Republic of Cameroon, Equatorial Guinea, and Gabon with a current seroprevalence of 1%-2%.11,12 Outside of Africa, infections are very rare and have been reported in Belgium, France, Germany, Spain, Norway, and the United States.13 Contrasting their relative contribution to the pandemic, disease progression in patients infected with HIV-1 O and M seems to be comparable.14,15 In the current study, we tested 4 web-based coreceptor prediction programs for their ability to determine the coreceptor usage of HIV-1 O primary isolates based on the amino acid sequence of the V3 loop. Viral sequences were obtained from DNA isolated from infected cells and aligned relative to the prototypic group M (clade B) strain NL4-3 (Fig. 1A). The biologic phenotype (Fig. 1B), which describes the utilization of the major coreceptors CCR5 and/or CXCR4, was recently determined for this panel of HIV-1 O isolates on TZM-bl reporter cells.16FIGURE 1: Prediction of coreceptor usage of HIV-1 O primary isolates by commonly used bioinformatics programs. A, Alignment of V3 loop sequences of 2 prototypic HIV-1 M strains and 7 HIV-1 O primary isolates. Gray boxes indicate identity with the NL4-3 sequence and dashes the gaps introduced to optimize the alignment. The total number of amino acids in an isolate's V3 loop is indicated. B, Results for HIV-1 isolates' biological, phenotypic coreceptor usage test performed in a TZM-bl reporter assay in the presence or absence of AMD3100 or SCH-D,16 and V3 loop sequence-based prediction performed by the indicated bioinformatics programs. Bold font indicates concordance of the prediction with the isolate's biological phenotype, regular font indicates discordance. The false positive rate given for geno2pheno[coreceptor] indicates the probability of classifying an R5-virus falsely as X4 and vice versa. WebPSSM can use 3 different matrices based on clade (B or C), coreceptor usage (Bx4r5), or syncytia induction (Bsinsi or Csinsi) for its prediction. No predic. = no prediction; CXCR4/CCR5 = mixed usage of both coreceptors predicted.Geno2pheno[coreceptor] identified 4 out of 7 V3 sequences (Fig. 1A) as belonging to HIV-1 O and accurately predicted the coreceptor usage of 2 isolates (28.6% specificity; Fig. 1B). In contrast, the program correctly predicted the coreceptor tropism of NL4-3, Ba-L (Fig. 1A), and several other primary M isolates (not shown), in line with recently published specificity rates.5 All 3 WebPSSM matrices performed well for the 2 clade B prototypes, but predicted all 7 O-type sequences to be X4-tropic, which was correct only for one (MVP-2901-94; 14.2% specificity) (Fig. 1B). Results from SVM-SK analyses predicted the inconsistent usage of CCR5 and CXCR4, but not dual-tropism, for all O-types (0% specificity, Fig. 1B). The fourth prediction program, WetCat, did not recognize the O-type sequences as V3 loop derived and was consequently unable to predict any coreceptor usage (Fig. 1B). In conclusion, 4 major web-based in silico prediction programs for coreceptor usage based on bioinformatic V3 loop analyses performed below chance level for HIV-1 O viruses. In light of the increasing importance of reliable coreceptor predictions in guiding treatment selection and prognosis in HIV/AIDS patients, coreceptor usage in the small number of individuals infected or coinfected with HIV-1 O viruses should at present be exclusively assessed by phenotypic assays and, ideally, sensitivity of these highly divergent isolates to CCR5 antagonists should be predetermined in vitro. Furthermore, all online prediction algorithms should clearly indicate their data training sets, predicted group/clade of the V3 sequences, and confidence of their assignments, and need to be educated and optimized for HIV-1 group O. Daniel Rupp Silvia Geuenich, PhD Oliver T. Keppler, MD Department of Infectious Diseases, Virology, University of Heidelberg, Germany