Genotypic Prediction of HIV-1 CRF01-AE Tropism

Abstract

HIV-1 subtype CRF01-AE predominates in south Asia and has spread throughout the world. The virus tropism must be determined before using CCR5 antagonists. Genotypic methods could be used, but the prediction algorithms may be inaccurate for non-B subtypes like CRF01-AE and the correlation with the phenotypic approach has not been assessed. We analyzed 61 CRF01-AE V3 clonal sequences of known phenotype from the GenBank database. The sensitivity of the Geno2pheno10 genotypic algorithm was 91%, but its specificity was poor (54%). In contrast, the combined 11/25 and net charge rule was highly specific (98%) but rather insensitive (64%). We thus identified subtype CRF01-AE determinants in the V3 region that are associated with CXCR4 use and developed a new simple rule for optimizing the genotypic prediction of CRF01-AE tropism. The concordance between the predicted CRF01-AE genotype and the phenotype was 95% for the clonal data set. We then validated this algorithm by analyzing the data from 44 patients infected with subtype CRF01-AE, whose tropism was determined using a recombinant phenotypic entry assay and V3-loop bulk sequencing. The CRF01-AE genotypic tool was 70% sensitive and 96% specific for predicting CXCR4 use, and the concordance between genotype and phenotype was 84%, approaching the concordance obtained for predicting the tropism of HIV-1 subtype B. Genotypic predictions that use a subtype CRF01-AE-specific algorithm appear to be preferable for characterizing coreceptor usage both in pathophysiological studies and for ensuring the appropriate use of CCR5 antagonists.

INTRODUCTION

The tropism of HIV-1 is defined by the capacity of the virus to use the chemokine receptors CCR5 and/or CXCR4 to enter CD4-expressing cells. A virus may use only CCR5 (R5 virus), only CXCR4 (X4 virus), or either coreceptor (R5X4 virus) (1). Determination of HIV-1 tropism is crucial for optimizing patient selection prior to coreceptor antagonist use and for studying the epidemiology and pathophysiology of HIV infection (1, 2). Genotypic determination of HIV tropism relies on sequencing the V3 env region and using prediction algorithms developed with reference to the in vitro phenotype (3, 4). HIV-1 is characterized by a high genetic diversity, especially in the env gene encoding the gp120 protein that harbors differences within and between subtypes, which can be as great as 20% (within) and 35% (between) (5). Genetic variations in the envelope gene influence the genotypic prediction of HIV-1 tropism that relies on data on the correlation between genotype and phenotype obtained in countries where subtype B viruses predominate. But the performance of genotypic algorithms can differ from one HIV-1 subtype to another and must be validated for each particular subtype (6–11).

The circulating recombinant form CRF01-AE was first isolated in Thailand and is now the dominant subtype in south and southeast Asia (excluding India), where its prevalence is 84% (12). The global prevalence of CRF01-AE is 5%, and it is responsible for up to 4% of the HIV-1 infections in Europe according to the SPREAD study (13). Previous studies have employed either phenotypic or genotypic prediction of coreceptor usage by HIV-1 CRF01-AE, but the correlation between the two methods has not been investigated (14–17).

We have therefore evaluated the performance of the genotypic algorithms for predicting HIV-1 subtype CRF01-AE tropism. We used the data from 61 CRF01-AE virus clones from GenBank to compare V3 sequences and phenotypes. This resulted in the definition of a new genotypic algorithm for the CRF01-AE subtype that was validated using data obtained from 44 patients.

MATERIALS AND METHODS

GenBank clonal data set of HIV-1 subtype CRF01-AE viruses.

The V3 sequences of HIV-1 subtype CRF01-AE viruses for which the entry phenotype was known were selected from the GenBank database. We selected 61 sequences from independent individuals resulting exclusively from clonal analysis. HIV-1 phenotypes were determined using a recombinant virus assay and by blocking pseudotyped viruses by the use of coreceptor antagonists (16, 18).

Subjects and samples.

We studied 44 samples of HIV-1 subtype CRF01-AE obtained at Toulouse University Hospital and Limoges University Hospital in France. The median age of the patients was 33 years (interquartile range [IQR], 30 to 41), and 69% were men. The median HIV-1 virus load was 4.9 log copies/ml (IQR, 4.3 to 5.5). The median CD4 cell count was 78 cells/mm³ (IQR, 49 to 188). None of the patients were at the primary infection stage. Most of the patients were naïve with respect to treatment (73%), whereas 5 patients had previously received an antiretroviral treatment.

All viruses were identified as HIV-1 subtype CRF01-AE by analysis of the env sequence using the NCBI genotyping tool (http://www.ncbi.nlm.nih.gov/projects/genotyping/formpage.cgi). We confirmed that these viruses belonged to the CRF01-AE subtype by neighbor-joining phylogenetic analysis of the sequences studied here, together with HIV-1 subtype reference sequences from the Los Alamos National Laboratory (http://www.hiv.lanl.gov/content/index).

Phenotypic characterization of HIV-1 coreceptor usage.

We determined the HIV-1 tropism with the Toulouse Tropism Test (TTT) phenotypic assay (19). Briefly, a fragment encoding the gp120 and the ectodomain of gp41 was amplified by reverse transcription-PCR (RT-PCR) using HIV-1 RNA isolated from the plasma. The PCR products were then subjected to nested PCR. Two amplifications were performed in parallel on aliquots of each sample; the amplified products were then pooled to prevent sampling bias of the virus population. The phenotype of HIV-1 coreceptor usage was determined using a recombinant virus entry assay with the pNL43-Δenv-Luc2 vector. 293T cells were transfected with both the NheI-linearized pNL43-Δenv-Luc2 vector DNA and the product of the nested PCR obtained from the challenged HIV-1-containing sample. The chimeric recombinant virus particles released into the supernatant were used to infect U87 indicator cells bearing CD4 and either CCR5 or CXCR4. Virus entry was assessed by measuring the luciferase activity (quantified in relative light units [RLU]) in lysed cells. Minor X4 variants were detected when they accounted for 0.5% or more of the total population as assessed in experiments performed on mixtures of various X4/R5 DNA ratios (19).

Genotypic prediction of HIV-1 coreceptor usage.

The V3 region was directly sequenced from bulk env PCR products in both directions by the dideoxy chain-termination method (BigDye Terminator v3.1; Applied Biosystems, Courtaboeuf, France) on an ABI 3130 DNA sequencer. The two primer pairs used for sequencing have been previously described (3). Results were analyzed with Sequencher (Genecodes, Ann Arbor, MI) by an operator blinded to the phenotype. Minority species were detected when the automated sequencer electropherogram showed a second base peak reaching at least 15% of the height of the upper peak. Multiple alignments were performed with CLUSTALW 1.83, and sequence alignments were manually edited with BioEdit software. Phylogenetic analyses excluded any possibility of sample contamination (see Fig. S1 in the supplemental material).

We used a combination of criteria from the 11/25 and net charge rules to predict HIV-1 tropism from the V3 genotype (3). One of the following criteria is required for predicting CXCR4 coreceptor usage: (i) 11R/K and/or 25K in V3; (ii) 25R in V3 and a net charge of ≥+5; and (iii) a net charge of ≥+6. All possible permutations were assessed when mixtures of amino acid were found at some codons of V3. The combination resulting in the highest net charge was used to predict the tropism. In the presence of mixtures, we chose the combination of amino acids leading to a predicted X4 sequence. We also used the geno2pheno tool in its original version and the new Geno2pheno-C_NGS-Sanger tool to predict HIV-1 coreceptor usage. Geno2pheno tools are available at http://coreceptor.bioinf.mpi-sb.mpg.de/cgi-bin/coreceptor.pl (September 2012).

Statistical methods.

The kappa coefficient was measured using STATA SE 9.2 to assess the agreement between the genotypic algorithms for predicting HIV-1 tropism and the phenotypic assay. The correlation between two tests is usually considered good when the kappa coefficient is greater than 0.60 with P < 0.05. The performances of the genotypic algorithms were compared using McNemar's chi-square test. A P value of less than 0.05 was defined as a statistically significant difference.

Nucleotide sequence accession numbers.

The sequences reported here have been given GenBank accession numbers JX535766 to JX535808.

RESULTS

Correlation between phenotype and V3 genotype for determining subtype CRF01-AE coreceptor usage.

The GenBank database contained 61 V3 clonal sequences of CRF01-AE with the corresponding phenotype. A total of 50 clones exclusively used CCR5, and 11 clones used CXCR4. The V3 genotype was predicted using the algorithms built for HIV-1. Geno2pheno10 predicted 33 X4 viruses (23 were mispredicted as X4) and 28 R5 viruses. The combined 11/25 and net charge rule predicted 8 X4 viruses and 53 R5 viruses, but 4 of them were mispredicted as R5 (Table 1). Thus, geno2pheno10 was 91% sensitive and 54% specific and the combined rule was 64% sensitive and 98% specific for predicting CXCR4 usage by HIV-1 subtype CRF01-AE clones (Table 1).

Table 1.

Comparison of genotypic prediction of HIV tropism and the observed phenotype of 61 HIV-1 subtype CRF01-AE virus clones in GenBank

Genotypic tool	Virus	No. of clones with indicated phenotype		Genotypic predictiona
		R5	R5X4/X4	% Sen	% Spe	% PPV	% NPV	Concordance (%)
Geno2pheno10	R5	27	1	91	54	29	97	61 (κ = 0.29; P < 0.001)
	X4	23	10
Geno2pheno5	R5	37	1	91	74	50	98	77 (κ = 0.46; P < 0.0001)
	X4	13	10
Combined 11/25 and net charge rule	R5	49	4	64	98	87	92	92 (κ = 0.69; P < 0.0001)
	X4	1	7

^aSen, sensitivity for detecting CXCR4-using viruses, calculated as the number of concordant X4/R5X4 results divided by the number of viruses phenotyped as R5X4/X4; Spe, specificity for detecting exclusive CCR5-using viruses, calculated by the number of concordant R5 results divided by the number of viruses phenotyped as R5; PPV, positive predictive value, calculated by the number of number of concordant X4/R5X4 results (true positive) divided by the number of viruses predicted X4 by the genotypic algorithm (number of positive tests); NPV, negative predictive value, calculated by the number of number of concordant R5 results (true negative) divided by the number of viruses predicted R5 with the genotypic algorithm (number of negative tests). Concordance between each genotypic algorithm and the phenotype was calculated as follows: number of samples with a concordant R5 genotype and phenotype plus number of samples with a concordant R5X4/X4 genotype and phenotype, all divided by the total number of tested samples. κ, kappa coefficient.

Genotypic determinants predicting CXCR4 use by HIV-1 subtype CRF01-AE viruses using clonal data.

We looked for V3 genotypic determinants known to be associated with CXCR4 usage by subtype B viruses such as the residues at positions 11 and 25, N-linked glycosylation sites, and the global net charge of the V3 region (Fig. 1 and Table 2). Among the 11 CXCR4-using clones, 7 had determinants of the combined 11/25 and net charge rules. The other four CXCR4-using clones had no “R” or “K” at positions 11 or 25 and a net charge <+6, but three of them had mutations leading to loss of the N-linked glycosylation site at the beginning of V3 in addition to a net charge of ≥+4. This site was also lost by 6 virus clones with an R5 phenotype, but 5 of them had a net charge of <+4 (Table 2). Thus, the loss of the N-linked glycosylation site combined with a net charge of ≥+4 is an independent criterion for predicting CXCR4 usage by CRF01-AE HIV-1 viruses.

Fig 1.

V3 amino acid sequence alignments and matched phenotypes of the 11 CRF01-AE CXCR4-using clones from the GenBank data set. V3 amino acid sequence alignments and corresponding phenotypes were selected from the GenBank database. These sequences are shown with the following abbreviation with reference to the consensus sequence: dash, gap inserted to maintain alignment. Residues at positions 11 and 25 and mutated N-linked glycosylation sites are boxed to highlight the substitutions noted. The V3 net charge (calculated by subtracting the number of negatively charged amino acids [D and E] from the number of positively charged ones [K and R]), the number of amino acids in V3, and the genotype predicted by the combined 11/25 and net charge rules built for subtype B viruses and by Geno2pheno10 are shown, together with the phenotype. Discordant genotypic predictions with reference to the phenotype are boxed.

Table 2.

Genotypic determinants in the sequences of the GenBank HIV-1 subtype CRF01-AE virus clones predicting their coreceptor use

Criteria observed in V3		No. of clonal sequences with indicated phenotype
Amino acid(s)	Net charge	R5	R5X4/X4
N-glycosylation site (NXT/S) mutation	<+5	5	3
	≥+5	1a	5
“R” or “K” at 11	<+5	0	5
	≥+5	1	5
“K” at 25	<+5	0	0
	≥+5	0	0
“R” at 25	<+5	0	0
	≥+5	0	0
No “R” or “K” at 11 or 25	<+6	49	4
	≥+6	0	1
No “R” or “K” at 11 or 25 but NXT/S mutation	<+4	5	0
	≥+4	1a	3

^aThe virus harboring a mutation of the N-glycosylation site had no “R” or “K” at position 11 or 25 and a net charge of +5.

We therefore designed a genotypic rule based on the 11/25 and net charge rules and mutation of the N-linked glycosylation site for determining the tropism of CRF01-AE HIV-1. One of the following criteria was required for predicting CRF01-AE CXCR4 coreceptor usage: (i) 11R/K and/or 25K in V3; (ii) 25R in V3 and a net charge of ≥+5; (iii) a net charge of ≥+6; and (iv) loss of the N-linked glycosylation site in V3 and a net charge of ≥+4. The genotypic and phenotypic approaches using this rule were 95% concordant on the clonal data set (Table 3).

Table 3.

Genotypic prediction of the tropism of the GenBank HIV-1 subtype CRF01-AE virus clones using a specific algorithm

Genotypic tool	Virus	No. of clones with indicated phenotype		Genotypic predictiona
		R5	R5X4/X4	% Sen	% Spe	% PPV	% NPV	Concordance (%)
Combined NXT/S mutation, 11/25, and net charge rule for subtype CRF01-AE	R5	48	1	91	96	83	98	95 (κ = 0.84; P < 0.0001)
	X4	2	10

^aSen, sensitivity for detecting CXCR4-using viruses; Spe, specificity for detecting exclusive CCR5-using viruses; PPV, positive predictive value; NPV, negative predictive value; κ, kappa coefficient.

Comparison of the performances of the combined rule and geno2pheno on the clonal data for CRF01-AE tropism prediction.

The sensitivity of geno2pheno10 was identical to the sensitivity of the combined rule (91%) for CRF01-AE for predicting CXCR4 use on the GenBank data set. However, the combined rule was more specific (96%) than geno2pheno10 (54%, P < 0.0001). We examined the performance of geno2pheno at all possible cutoffs using the receiver operating characteristic (ROC) curve (Fig. 2a). We also determined the ROC curve for geno2pheno-C_NGS-Sanger. The performance of the combined rule for CRF01-AE appears as a single point.

Fig 2.

Receiver operating characteristic curves of the genotypic algorithms versus the phenotype. (a) ROC curve for the data obtained on the GenBank clones. (b) ROC curve for the data obtained on the set of clinical samples. The black line represents the performance of the original geno2pheno algorithm, and the gray line represents the performance of the geno2pheno-C_NGS-Sanger algorithm, both referring to the phenotype. The point represents the performance of the combined rule for CRF01-AE.

Validation of the subtype CRF01-AE genotypic algorithm.

The 44 individuals infected with HIV-1 subtype CRF01-AE included 24 infected with virus populations with an R5 phenotype, 14 with a dual/mixed R5X4 phenotype, and 6 with a pure X4 phenotype. We sequenced the V3 region from the bulk env PCR products of the viruses from these 44 patients (Fig. 3). Geno2pheno10 predicted 28 X4 viruses (13 were mispredicted as X4) and 16 R5 viruses (Table 4). Geno2pheno5 predicted 18 X4 viruses (4 were mispredicted as X4) and 26 R5 viruses. The specific combined rule for subtype CRF01-AE predicted 15 X4 viruses (one was mispredicted as X4) and 29 R5 viruses (6 were mispredicted as R5). Thus, the combined rule for subtype CRF01-AE and geno2pheno10 were similarly sensitive (70% and 75%, respectively) but the combined rule was more specific (96%) than geno2pheno10 (46%) for predicting CXCR4 usage by HIV-1 subtype CRF01-AE (P < 0.001). Geno2pheno5 was 70% sensitive and 83% specific in this data set, and the performance was not different from the performance of the combined rule (P = 0.08). The performances of geno2pheno, geno2pheno-C_NGS-Sanger, and the combined rule for CRF01-AE are compared on the ROC graph (Fig. 2b).

Fig 3.

V3 amino acid sequence alignments and matched phenotypes of the 44 subtype CRF01-AE viruses. V3 amino acid sequence alignments were obtained by bulk sequencing env PCR products from the 44 subtype CRF01-AE-infected patients. These sequences are shown with the following abbreviations with reference to the CRF01-AE consensus sequence: dot, identity with amino acid baseline sequence; dash, gap inserted to maintain alignment; slash, amino acid position related to dual-virus population. Replacements are indicated by the appropriate code letters. Residues at positions 11 and 25 and mutated N-linked glycosylation sites are boxed to highlight the substitutions noted. The V3 net charge (calculated by subtracting the number of negatively charged amino acids [D and E] from the number of positively charged ones [K and R]), the number of amino acids in V3, and the genotype predicted by the combined 11/25 and net charge rules built for subtype B viruses and by Geno2pheno10 are shown, together with the observed phenotype. Discordances between the genotypic predictions and the phenotype are boxed.

Table 4.

Genotypic prediction of the tropism of 44 HIV-1 subtype CRF01-AE infections compared to the TTT phenotypic assay

Genotypic tool	Virus	No. of samples with indicated phenotype in TTT assay		Genotypic predictiona
		R5	R5X4/X4	% Sen	% Spe	% PPV	% NPV	Concordance (%)
Geno2pheno10	R5	11	5	75	46	54	69	59 (κ = 0.20; P = 0.08)
	X4	13	15
Geno2pheno5	R5	20	6	70	83	78	77	77 (κ = 0.54; P < 0.001)
	X4	4	14
Combined NXT/S mutation, 11/25, and net charge rule for subtype CRF01	R5	23	6	70	96	93	79	84 (κ = 0.67; P < 0.0001)
	X4	1	14

^aSen, sensitivity for detecting CXCR4-using viruses; Spe, specificity for detecting exclusive CCR5-using viruses; PPV, positive predictive value; NPV, negative predictive value; κ, kappa coefficient.

DISCUSSION

HIV-1 subtype CRF01-AE is the major subtype responsible for the pandemic in southeast Asia, where about 4 million people are infected with HIV-1 (http://www.unaids.org). A tropism test that is easy to use in clinical practice would facilitate the use of CCR5 antagonists for treating CRF01-AE-infected patients. Previous studies have used recombinant phenotypic assays that may be unsuitable for clinical use, particularly in resource-limited countries (19–21). Genotypic assays have been developed to determine the tropism of HIV-1, but they may be inappropriate for predicting coreceptor usage by the subtype CRF01-AE because of the great variability of the env gene. We have now analyzed the correlations between phenotypic and genotypic approaches for determining the tropism of HIV-1 subtype CRF01-AE.

We analyzed the performance of the genotypic algorithms using data from 61 subtype CRF01-AE clones listed in GenBank. The genotypic algorithms validated for HIV-1 subtype B were inadequate for predicting CXCR4 use by subtype CRF01-AE viruses. Geno2pheno10 lacked specificity (55%), while the combined rule lacked sensitivity (64%). Geno2pheno with a 10% cutoff has been recommended for V3 genotyping (22). Recent evaluations of genotypic algorithms for predicting tropism of non-B subtype viruses showed good agreement with the phenotype, but the CRF01-AE subtype was not included (6, 23). Moreover, we have shown that these genotypic tools are not valid for predicting the tropism of subtype D viruses (8). In contrast, the combined 11/25 and net charge rule was efficient for predicting CXCR4 use by HIV-1 subtype B, subtype C, and CRF02-AG (3, 9, 10).

We therefore looked for a genotypic tool that was more suitable for subtype CRF01-AE using data from GenBank. Analysis of the V3-loop sequences and the corresponding phenotypes showed that the 11/25 criteria and the net charge rule did not identify all the CXCR4-using clones. The loss of the N-linked glycosylation site at the beginning of V3 was an independent determinant of CXCR4 use by the CRF01-AE virus clones. This mutation is known to be associated with CXCR4 use by subtype B and subtype C viruses (10, 24–26). We therefore designed a genotypic algorithm specific for HIV-1 subtype CRF01-AE that combined the criteria of 11/25, net charge, and NXT/S mutation. Genotypic predictions for the GenBank virus clones using this CRF01-AE algorithm were 95% concordant with their phenotypes.

We validated this new rule by analyzing the data for 44 patients infected with the CRF01-AE subtype. The TTT phenotypic assay identified 20 R5X4/X4 viruses. Thus, CXCR4-using viruses were highly (45%) prevalent in this set of patients. This was to be expected, as all our patients were at an advanced stage of HIV-1 infection, with low CD4 cell counts. The prevalences of X4 in the GenBank set of samples and in the validation cohort were different (20% versus 45%), but the GenBank clones were not suitable for evaluation of X4 prevalence. Because genotypic tools perform differently according to the study population, the rule combining criteria from the 11/25 and the net charge rules and the loss of the N-linked glycosylation site should be evaluated in other cohorts.

Geno2pheno10 was inadequate for predicting CXCR4 use by HIV-1 subtype CRF01-AE. The cutoff could be modified, but the ROC curves established on GenBank and patient data sets showed discordant cutoffs. Moreover, these cutoffs were very low (1.8% to 3%) and could lead to a lack of sensitivity for predicting CXCR4 use in another data set, as previously shown for HIV-1 subtype B (27–29). A new geno2pheno algorithm has been developed recently to improve genotypic prediction on bulk sequencing using next-generation sequencing data (30). We evaluated geno2pheno-C_NGS-Sanger, but this tool was less efficient that the original geno2pheno for predicting CRF01-AE tropism.

Finally, the best concordance with the phenotype was obtained with the CRF01-AE combined rule (sensitivity, 70%; specificity, 96%), which was in good agreement with the phenotypic assay (kappa coefficient of 0.67). The sensitivity for predicting CXCR4 use is crucial to avoid inappropriate administration of CCR5 antagonists. Genotypic prediction of HIV-1 tropism using bulk sequencing is less sensitive than phenotypic approaches, but clinical trials of CCR5 antagonists have shown that these methods are suitable for clinical practice (31). Genotypic prediction of HIV-1 tropism also requires high specificity in order to allow the use of CCR5 antagonists in patients with limited options and to analyze the impact of CXCR4-using viruses on the disease progression. The genotypic rule was less sensitive for these patients than it was for the GenBank clonal samples (93%) because bulk sequencing detects variants accounting for at least 20% to 25% of the virus quasispecies. Minor variants detected by the sensitive phenotypic assay may be missed by bulk genotyping. A crucial strategy to improve the sensitivity of the bulk sequencing was to consider amino acid mixtures and the combination resulting in the highest net charge (3, 32). Sensitivity could be increased by using ultrasensitive sequencing methods, but they also require reliable genotypic algorithms for predicting tropism.

The global concordance for the subtype CRF01-AE genotypic rule (84%) was similar to those of the genotypic algorithms for subtype B (91%) and subtype D (92%) viruses (3, 8). This simple rule may also be convenient for identifying patients eligible for treatment with a CCR5 antagonist, but clinical outcome data are needed to confirm that this approach is valid.

In conclusion, we have developed a specific approach for the genotypic prediction of subtype CRF01-AE tropism. The performance of this rule needs to be confirmed in other cohorts with clinical outcome. This genotypic method could facilitate the clinical use of CCR5 antagonists and help to determine the impact of virus tropism on disease progression. Analysis of other major subtypes responsible for the HIV-1 pandemic could also benefit from optimized genotypic algorithms.