Possible Footprints of APOBEC3F and/or Other APOBEC3 Deaminases, but Not APOBEC3G, on HIV-1 from Patients with Acute/Early and Chronic Infections

ABSTRACT

Members of the apolipoprotein B mRNA-editing enzyme-catalytic polypeptide-like-3 (APOBEC3) innate cellular cytidine deaminase family, particularly APOBEC3F and APOBEC3G, can cause extensive and lethal G-to-A mutations in HIV-1 plus-strand DNA (termed hypermutation). It is unclear if APOBEC3-induced mutations in vivo are always lethal or can occur at sublethal levels that increase HIV-1 diversification and viral adaptation to the host. The viral accessory protein Vif counteracts APOBEC3 activity by binding to APOBEC3 and promoting proteasome degradation; however, the efficiency of this interaction varies, since a range of hypermutation frequencies are observed in HIV-1 patient DNA. Therefore, we examined “footprints” of APOBEC3G and APOBEC3F activity in longitudinal HIV-1 RNA pol sequences from approximately 3,000 chronically infected patients by determining whether G-to-A mutations occurred in motifs that were favored or disfavored by these deaminases. G-to-A mutations were more frequent in APOBEC3G-disfavored than in APOBEC3G-favored contexts. In contrast, mutations in APOBEC3F-disfavored contexts were relatively rare, whereas mutations in contexts favoring APOBEC3F (and possibly other deaminases) occurred 16% more often than average G-to-A mutations. These results were supported by analyses of >500 HIV-1 env sequences from acute/early infection.

IMPORTANCE Collectively, our results suggest that APOBEC3G-induced mutagenesis is lethal to HIV-1, whereas mutagenesis caused by APOBEC3F and/or other deaminases may result in sublethal mutations that might facilitate viral diversification. Therefore, Vif-specific cytotoxic T lymphocyte (CTL) responses and drugs that manipulate the interplay between Vif and APOBEC3 may have beneficial or detrimental clinical effects depending on how they affect the binding of Vif to various members of the APOBEC3 family.

INTRODUCTION

Human immunodeficiency virus type 1 (HIV-1) continuously adapts to selection pressures in infected individuals. Because of its small genome, high mutation rate, large population size, and rapid turnover rate, nearly all HIV-1 single-mutation variants are produced in most untreated hosts each day, and any variant with the slightest survival advantage rapidly prevails (1). The high mutation rate of HIV-1 has been linked to the high error rate of HIV-1 reverse transcriptase (RT) and RNA polymerase II (RNA Pol II) (2) and to innate cellular cytidine deaminases, such as apolipoprotein B mRNA-editing enzyme-catalytic polypeptide-like 3 (APOBEC3) (3).

Several members of the APOBEC3 family (APOBEC3A to APOBEC3H) have antiviral effects; however, APOBEC3F and APOBEC3G exert the strongest levels of antiviral activity and are the primary restriction factors inhibiting HIV-1 replication in human primary CD4⁺ T cells, the key HIV-1 target cells (4). During reverse transcription, APOBEC3F and APOBEC3G induce cytidine-to-uridine (C-to-U) deamination of single-stranded minus-strand HIV-1 DNA, which generates guanosine-to-adenosine (G-to-A) mutations in plus-strand DNA (3, 5–11). These G-to-A mutations occur in APOBEC3F- and APOBEC3G-favored di-, tri- and tetranucleotide contexts (the GGnn-to-AGnn and GAnn-to-AAnn contexts, respectively [12]) (the mutated nucleotide residue is underlined). The remaining APOBEC3 proteins favor the same dinucleotide context as APOBEC3F. Because G-to-A transitions far exceed all other mutations, the mutated provirus is referred to as “hypermutated.”

This intrinsic cellular defense mechanism is thwarted by the HIV-1-encoded protein Vif, which mediates APOBEC3 degradation. Vif simultaneously binds to APOBEC3 and to a cellular ubiquitin ligase (cullin5 [Cul5]-elongin [Elo]B/C-Rbx)) through distinct amino acids motifs (13, 14). This binding induces the polyubiquitination of both APOBEC3 and Vif, which are then targeted for proteasomal degradation (3, 15–20). This process depletes the pool of cytosolic APOBEC3 that is available for incorporation into assembling viral particles and reduces the risk of HIV-1 hypermutation in the next infected cell.

However, the efficiency of the Vif-APOBEC3 interaction clearly varies among hosts, because a range of hypermutation frequencies has been observed in HIV-1 DNA from infected patients. Whereas the lowest observed levels of APOBEC3G-induced hypermutation in vitro are highly likely to introduce multiple lethal stop codons within HIV-1 (21), it is unclear whether some or all of the APOBEC3 proteins induce mutations at sublethal levels that increase viral diversification in vivo. Such increased diversification may facilitate viral adaptation to selective pressures in the host and thereby enhance HIV-1 propagation.

To determine if APOBEC3-induced mutagenesis can facilitate HIV-1 evolution, we focused on APOBEC3F and APOBEC3G. We examined longitudinal HIV-1 sequences from ∼3,000 chronically infected patients to detect polymorphisms that were indicative of APOBEC3F or APOBEC3G activity (i.e., APOBEC3F or APOBEC3G “footprints”) by determining whether G-to-A mutations occurred in motifs that were favored or disfavored by the deaminases. Because the remaining APOBEC3 proteins favor the APOBEC3F-favored dinucleotide context, the APOBEC3F footprints could be generated by APOBEC3F and/or other deaminases. We hypothesized that deleterious mutations, even if nonlethal, would be quickly eliminated from the population and that the footprints would indicate remaining deaminase-induced mutations that were selectively neutral or possibly advantageous to the viruses.

MATERIALS AND METHODS

HIV-1 RNA sequences from patients without hypermutated viruses during chronic infection.

We obtained a curated set of 5,614 longitudinal pairs of partial pol sequences (average length = 1,084 bp) from 3,190 patients who were previously analyzed by Deforche et al. (22). The sequences were generated using population sequencing and originated with the Stanford Drug Resistance Database in Palo Alto, CA (23), the University Hospitals in Leuven, Belgium, and Hospital Egas Monis in Lisbon, Portugal. Briefly, HIV-1 was amplified from plasma samples and sequenced using various clinically validated genotypic in-house (24) or commercial tests (e.g., Bayer Health Diagnostics' HIV-1 TrueGene and Celera Diagnostics/Abbott Laboratories' ViroSeq) (Fig. 1A). These assays produce a consensus sequence representative of the circulating and actively replicating HIV population in the patient, and these sequences were used to decide the patients' antiretroviral treatments (24, 25).

FIG 1.

Outline of analyses. (A) Longitudinal blood samples were obtained from 3,190 HIV-infected patients to test for drug resistance mutation as part of their treatment program. HIV-1 was amplified using PCR and sequenced using population sequencing. The sequences were deposited in either the Stanford Drug Resistance Database in Palo Alto, CA, or at the EuResist database. (B) The sequences were curated and organized into sequence pairs (n = 5,614). (C) The observed substitutions provided information of the underlying misincorporation rates by considering the set of nucleotide substitutions that were equivalent at the amino acid level. In these examples, two alternative substitutions could have generated the same synonymous and nonsynonymous substitutions, respectively. Because we assumed that selection acts mostly at the amino acid level, the observed nucleotide substitutions could be considered to be the outcome of a probabilistic experiment, where the probability of observing any of the equivalent nucleotide substitutions would be proportional to the misincorporation rate.

The data were curated by creating phylogenetic trees using the software program PAUP (26). Codons at positions that were associated with drug resistance mutations in protease (positions 30, 46, 48, 50, 54, 56, 82, 84, 88, and 90) and reverse transcriptase (41, 44, 65, 67, 70, 74, 75, 77, 100, 103, 106, 108, 115, 116, 151, 181, 184, 188, 190, 210, 215, and 219) (27) were excluded to avoid the confounding effects of convergent evolution on the tree reconstructions (28). Sequence pairs from the same patient that did not cluster together in the tree (suggesting either contamination or HIV-1 superinfection) were excluded (a total of 450 sequence pairs). Each of the remaining 5,614 sequence pairs contained two sequences (Fig. 1B). Because 2 to 4 sequences were available from each patient, we obtained 1 to 3 pairs per patient. However, 1 or 2 pairs were most commonly observed. All of the sequence pairs represented the total number of available sequences per patient. The genetic distance was estimated using PAUP, and the sequences were subtyped using the subtyping tool in the HIV database (www.hiv.lanl.gov). If dual HIV-1 infections were present in all samples from a given patient, it would have added unbiased noise to the population sequencing results, limiting the resolution of our analyses. However, the quality of all sequence reads was carefully evaluated at the time patient treatment was decided, and signs of dual infections were not observed either during proofreading of the sequences or during the phylogenetic clustering analysis. No CD4 counts or viral load data were available.

Maximum-likelihood-based nucleotide misincorporation rate estimator.

A modification of the method developed by Deforche et al. (22) was used to estimate the misincorporation rates of the observed substitutions in the serially sampled sequences. The misincorporation rates of individual mutations will depend on their mutation rates, the other mutations with which they cooccur, and the effects of natural selection on the protein sequence. Because we are interested in the first two factors, we corrected for changes to the protein sequence by considering the relative rates of changes whose effects on the amino acid sequence were equivalent.

Briefly, nucleotide substitutions were considered equivalent when they had the same effect at the amino acid level (i.e., they either changed the motif to encode the same amino acid or were silent mutations), and the observed nucleotide substitution was considered against the alternative equivalent substitution alone. The observed nonsynonymous substitutions, which could have been achieved by alternative equivalent nucleotide substitutions (e.g., a Phe-to-Leu mutation may be caused by a TTC-to-CTC (T-to-C) mutation rather than the alternative TTC-to-TTA (C-to-A) or TTC-to-TTG (C-to-G) mutations, and synonymous substitutions were used to generate the maximum-likelihood estimation (MLE) of the misincorporation rates (Fig. 1C). In our analyses, we implemented two novel models. To assess the influence of the two downstream nucleotides on observed G-to-A misincorporation rates, the first model included parameters that estimated the GGn-to-AGn (APOBEC3G-type) misincorporation rates, and the second model included parameters that estimated the GAn-to-AAn (APOBEC3F-type) misincorporation rates. Both models integrated the following parameters: (i) 11 parameters that corresponded to the relative rates of misincorporation other than those for the A-to-C change, which was used as a reference; (ii) a parameter that corrected for a bias toward G-to-A mutations that remove CpG (i.e., CG-to-CA); (iii) a parameter that corrected for a bias toward other mutations that remove CpG; and (iv) a parameter that corrected for a bias against mutations that create CpG. Therefore, all of the mutation rates should reflect the context-dependent bias in the generation of the mutations (e.g., mediated by the APOBEC3 proteins or RT and/or RNA Pol II), or unknown selection factors not associated with the amino acid sequences or selection against CpG motifs. This analysis is not based on the frequency of observed mutations but on estimating a preference for a certain observed mutation over possible alternative mutations, which have the same influence on the amino acid sequence. Consequently, the analysis looks at relative rates of mutation between trinucleotide contexts, rather than simply plotting the proportion that are mutated. The set of cross-sectional acute and early HIV-1 sequences (29) was analyzed based on APOBECF and APOBEC3G dinucleotide preferences.

Ethics statement.

The research was conducted according to the Declaration of Helsinki. Informed consent was obtained from all of the patients according to the regulations of each country regarding experimentation using human tissue. The protocol and this consent procedure were approved by the Ethical Committee UZ Leuven (reference ML-8627, approval B322201316521 S52637).

The sequence accession numbers from GenBank and the Stanford Drug Resistance Database can be found in Table S1 in the supplemental material, whereas sequence identifiers (IDs) from the EuResist database can be made available to researchers after submission of a study protocol to EUresist (http://www.euresist.org) and study approval by the local ethical committee. The EUresist sequence IDs cannot be made available without study approval because the IDs are linked to patient treatment and the patients' informed consents are time limited.

RESULTS

Analyses of longitudinal HIV-1 pol sequences from patients without hypermutated viruses during chronic infection.

To investigate if sublethal APOBEC3-induced mutagenesis occurs in natural HIV-1 infection, we analyzed the curated set of longitudinal pairs of HIV-1 pol RNA sequences from 3,190 chronically infected patients for APOBEC3F and APOBEC3G footprints. The pairs comprised 2 to 4 sequences per patient. For each patient, we first compared the sequences from the first and second time points (equal to one pair), followed by the sequences from the second and third time points (equal to a second pair) and those from the third and fourth time points (equal to a third pair) when appropriate. The sequences were sampled over 1 to 3 years. The median time from baseline to follow-up was 469 days (25 to 75% quartile range, 231 to 894). The sequences were all derived from HIV-1 RNA isolated from plasma, and none were hypermutated. The mean genetic distance was 0.017 nucleotide substitutions per site. Approximately 90% of the sequences belonged to subtype B, 4% to subtype G, and the remainder to other subtypes, circulating recombinant forms (CRF), or unique recombinant forms (URF). Dual infections were not observed in the cohort.

A previous study used this data set and APOBEC3G- and APOBEC3F-dinucleotide contexts to examine the relative importances of RT and APOBEC3 for the G-to-A misincorporation rates in HIV-1 (22). Although that study confirmed the previously described bias toward G-to-A mutations in HIV-1 (30–32), APOBEC3G or APOBEC3F footprints could not be reliably detected. We previously determined the tri- and tetranucleotide target preferences for APOBEC3F and APOBEC3G (12), which enabled us to study the influence of the downstream nucleotide on G-to-A mutation rates and clearly distinguish the APOBEC3F and APOBEC3G target sites.

Possible footprints of APOBEC3F and/or other deaminase—but not APOBEC3G—activity on HIV-1 RNA sequences from chronic infection.

To evaluate whether G-to-A mutations in vivo occurred in motifs that were known to be favored or disfavored by APOBEC3F or APOBEC3G, we used the modified version of the maximum-likelihood-based nucleotide misincorporation rate estimator that incorporates the specific trinucleotide target motifs for APOBEC3F (GAn-to-AAn) and APOBEC3G (GGn-to-AGn) (see Materials and Methods) (12). This program included correction factors for mutations that resulted in nonsynonymous (NS) changes and a factor that modified the rate at which CpG dinucleotides were added or removed by a mutation; the latter factor was necessary because selection against CpG motifs occurs only at the nucleic acid level, possibly due to the role of cytosine methylation in decreasing viral gene expression (22, 33). Therefore, the estimated misincorporation rates likely reflect the context-dependent bias in the generation of the mutations (e.g., the motifs that were favored by APOBEC3F [and possibly other deaminases], APOBEC3G, or RT/RNA Pol II) or unknown selection factors that were not associated with the amino acid sequences (Tables 1 and 2).

TABLE 1.

Misincorporation rate estimates generated by analysis of 5,614 pairs of longitudinal nonhypermutated patient-derived HIV-1 pol sequences with separate model parameters for GGn-to-AGn^a

Statistic	Value for misincorporation rate
	A→C	A→G	A→T	C→A	C→G	C→T	G→A	G→C	G→T	T→A	T→C	T→G	fAddCpG	fDelCpG	fGGAtoAGA	fGGCtoAGC	fGGGtoAGG	fGGTtoAGT	fDelCpG.GA
Estimate	0.425	2.090	0.232	1.472	0.211	3.356	5.156	0.198	0.461	0.629	2.075	0.659	0.274	3.375	1.206	1.675	0.607	1.831	5.542
Lower 95% CI	0.384	1.948	0.208	1.302	0.167	3.106	4.754	0.159	0.391	0.559	1.929	0.563	0.250	2.919	1.069	1.266	0.517	1.476	4.627
Upper 95% CI	0.472	2.241	0.258	1.663	0.269	3.625	5.591	0.247	0.544	0.708	2.232	0.771	0.300	3.898	1.361	2.215	0.712	2.269	6.625

^aThe model estimates both the rates of each individual single nucleotide misincorporation (demonstrating that G to A is the most frequent change), the parameter that corrected for the bias toward G-to-A mutations that remove CpG [factor (f)AddCpG], a parameter that corrected for the bias against mutations that create CpG (fDelCpG), the context-specific rates in the context of interest (i.e., GGA to AGA, GGC to AGC, GGG to AGG, and GGT to AGT), and the overall G-to-A rate after factoring in corrections for CpG (fDelCpG.GA). CI, confidence interval.

TABLE 2.

Misincorporation rate estimates generated by analysis of 5,614 pairs of longitudinal nonhypermutated patient-derived HIV-1 pol sequences with separate model parameters for GAn-to-AAn^a

Statistic	Value for misincorporation rate
	A→C	A→G	A→T	C→A	C→G	C→T	G→A	G→C	G→T	T→A	T→C	T→G	fAddCpG	fDelCpG	fGAAtoAAA	fGACtoAAC	fGAGtoAAG	fGATtoAAT	fDelCpG.GA
Estimate	0.433	2.125	0.235	1.496	0.215	3.412	4.831	0.205	0.474	0.640	2.110	0.670	0.274	3.372	1.416	0.736	1.297	1.309	6.058
Lower 95% CI	0.390	1.983	0.212	1.325	0.170	3.161	4.430	0.165	0.403	0.569	1.964	0.573	0.250	2.917	1.274	0.597	1.080	1.154	5.070
Upper 95% CI	0.480	2.276	0.262	1.690	0.273	3.683	5.266	0.255	0.559	0.720	2.267	0.784	0.300	3.896	1.574	0.908	1.557	1.484	7.224

^aSee footnote a of Table 1.

We estimated two models for the HIV-1 in vivo relative misincorporation rates using the patient-derived sequence pairs: a GAn-to-AAn model (reflecting APOBEC3F-favored trinucleotides) and a GGn-to-AGn model (reflecting APOBEC3G-favored trinucleotides) (Fig. 2A and B). The results for the parameters that were included in both models were similar, and large variations in the misincorporation rate were observed (Fig. 2A). The most frequent mutation (G to A) was approximately 25 to 30 times more prevalent than the least frequent mutations (C to G and G to C). Furthermore, transitions were more frequent than transversions; the most common transitions were G to A and C to T, and the most frequent transversions were C to A and T to A. The parameters that modeled a bias in G-to-A misincorporation in the GGn-to-AGn and GAn-to-AAn contexts revealed that the bias was dependent upon the downstream nucleotide (Fig. 2B).

FIG 2.

The examination of APOBEC3G and APOBEC3F mutational footprints within natural HIV-1 sequences. (A and B) A total of 5,164 pairs of partial pol sequences from 3,190 HIV-1-positive patients were analyzed using the modified maximum-likelihood-based nucleotide misincorporation rate estimation method. We estimated the nucleotide misincorporation rates in a GGn-to-AGn model and a GAn-to-AAn model across the set of sequence pairs (A), and we assessed the influence of the two downstream nucleotides on the observed G-to-A misincorporation rates (B). The GGn-to-AGn model included parameters that estimated APOBEC3G-type misincorporation rates, and the GAn-to-AAn model included parameters that estimated APOBEC3F-type misincorporation rates. In panel A, the relative misincorporation rates are scaled to obtain a geometric mean of “1” for the estimated rates to allow for objective comparisons, whereas the misincorporation rates are expressed relative to the average G-to-A misincorporation rate in panel B. The relative frequencies of mutation in the trinucleotides that contained the known preferred APOBEC3G and APOBEC3F target dinucleotides are shown (APOBEC3G, left “GG”; APOBEC3F, right “GA”), and the trinucleotides that were most and least favored by APOBEC3G and APOBEC3F in vitro (see parts C and D, respectively) are indicated with arrows. We previously demonstrated that the preferred APOBEC3G and APOBEC3F tetranucleotide motifs correlated strongly in vitro and in vivo (12). (C and D) We previously determined APOBEC3G and APOBEC3F tetranucleotide preferences using nearly full-length in vitro-hypermutated HIV-1 genomes (12). These sequences were reanalyzed in this study to determine the odds ratios (95% confidence intervals in parentheses) by comparing the mutation frequencies between pairs of trinucleotides that contained the preferred dinucleotide motif targets for APOBEC3G “GGn” (C) and APOBEC3F “GAn” (D). The ratios represent the relative frequencies of the column over row category. In the set of sequences that were generated in the presence of APOBEC3G (C), GGA-to-AGA mutations occurred approximately 14 times more frequently than GGC-to-AGC mutations.

We next examined the relative preferences of APOBEC3F and APOBEC3G for trinucleotide motifs, including the favored dinucleotide target of each deaminase (i.e., GAn for APOBEC3F and GGn for APOBEC3G) (Fig. 2C and D), by reanalyzing the hypermutated sequences from our previous study (12). We previously demonstrated that APOBEC3G and APOBEC3F display highly conserved target motif preferences in vitro and in vivo. In agreement with other studies (34, 35), we determined here that APOBEC3G targeted GGG motifs at least 30 times more frequently than GGC motifs (Fig. 2C), whereas APOBEC3F targeted GAA motifs nearly 15 times more frequently than GAC ones (Fig. 2D). Thus, mutagenesis induced by either deaminase was inhibited when a plus-strand C occurred 2 bp downstream from the target G (Fig. 2C and D, second columns).

G-to-A mutations were the most frequently observed mutations but did not generally occur in nucleotide contexts favored by APOBEC3G (Fig. 2C). GG-to-AG mutations in contexts that were disfavored by APOBEC3G (e.g., GGC-to-AGC) were more frequent than those in APOBEC3G-favored contexts (Fig. 2B and C). Because G-to-A mutations were observed less frequently in the optimal APOBEC3G context GGG (12, 35) than in other contexts (0.36 < 1 [95% likelihood interval = 0.25 to 0.51]), it is highly likely that HIV-1 viruses with mutations in this specific context were incapable of propagation (i.e., they constituted inactivated, “dead-end” viruses).

In contrast, GAA-to-AAA mutations did occur 16% more often than average GA-to-AA mutations and may therefore have been induced by APOBEC3F or other deaminases (e.g., APOBEC3A [monocytes/macrophages only], APOBEC3B, APOBEC3C, APOBEC3DE, and/or APOBEC3H haplotypes II, V, and VII) (34, 36–42) or a combination of several deaminases, with contributions from RT and/or RNA Pol II (34, 43–45) (Fig. 2C and D). The relatively lower likelihood of these mutations occurring in contexts that were disfavored by APOBEC3F and/or possibly other deaminases suggests that a substantial proportion of the G-to-A mutations may have been induced by these deaminases. Since APOBEC3F, APOBEC3A, APOBEC3B, APOBEC3C, APOBEC3DE, and APOBEC3H preferentially target 5′GA motifs, stop codons are less likely to be induced because they would occur only when a mutation occurred in an in-frame “TGG Ann” codon pair (i.e., “TGG Ann” to “TGA Ann, " where “n” signifies any base).

Because we analyzed longitudinal sequences from patients and counted mutations that occurred in APOBEC3F and APOBEC3G contexts over time, a mutation that occurs at one time point and is carried over to the next time point is likely sublethal. None of the observed mutations generated stop or start codons, and the majority of changes were synonymous. Our method is robust to the selective pressures that operate at the amino acid level because it focuses only on observed preferences between mutations that are equivalent on the amino acid level (i.e., mutations that are either silent substitutions or changed the amino acid to the same amino acid) (22). Therefore, our results are not confounded by “hitchhiking effects” in which, for example, a substitution is fixed by positive selection of a mutation in a different position in the sequence or by sequential evolution at multiple codons. If no mutations occurred in a given dinucleotide context in the longitudinal data set from a patient, then this specific context was not included in our analyses.

Taken together, our results suggest that APOBEC3G is highly unlikely to increase HIV-1 diversity in vivo, whereas APOBEC3F and possibly other deaminases (e.g., APOBEC3A [monocytes/macrophages only], APOBEC3B, APOBEC3C, APOBEC3DE, and/or APOBEC3H haplotypes II, V, and VII) may increase HIV-1 diversification and facilitate viral adaptation and propagation.

Analyses of HIV-1 env sequences from patients without hypermutated viruses during acute and early infection.

We expanded our analyses of APOBEC3F- and APOBEC3G-induced HIV-1 mutagenesis in chronic infection to include sequences from acute and early infection. Using previously published sequences (29), we analyzed 514 HIV-1 RNA envelope sequences from these disease stages. The previous analysis reported that several RNA sequences had greater numbers of G-to-A mutations in APOBEC3-favored contexts than expected (29); however, these analyses did not differentiate between APOBEC3F and APOBEC3G contexts. Our analyses included the dinucleotide preferences of APOBEC3F and APOBEC3G and demonstrated that the majority of these G-to-A mutations occurred in APOBEC3F-favored motifs rather than APOBEC3G-favored motifs (70% in the APOBEC3F context versus 12% in the APOBEC3G context for all sequences with G-to-A mutations and 68% in the APOBEC3F context versus 12% in the APOBEC3G context for the nonhypermutated sequences only) (Fig. 3), in agreement with the results of the chronic infection analysis. No nonhypermutated sequences with GG-to-AG mutations carried more than a single GG-to-AG mutation, whereas 25% of the nonhypermutated sequences with GA-to-AA mutations carried multiple GA-to-AA mutations.

FIG 3.

Analysis of mutation characteristics in 514 HIV-1 env sequences from acute and early infection. We divided the sequences from reference 29 into those with (n = 120) and without (n = 394) G-to-A mutations. First, we asked whether sequences with G-to-A mutations were hypermutated (1); second, we examined the hypermutation bias (2) and found that 90% carried GA-to-AA hypermutation; GA is a motif preferentially mutated by APOBEC3F and other deaminases (e.g., APOBEC3A, APOBEC3B, APOBEC3C, APOBEC3DE, and/or APOBEC3H haplotypes II, V, and VII) but not APOBEC3G. For simplicity, we characterize this motif as an “APOBEC3F” context in the figure. Third, we examined the Gn-to-An mutation characteristics in the nonhypermutated sequences with G-to-A mutations (3) and found that in 68%, mutations were in the GA (“APOBEC3F”) context, whereas 12% were found in the GG (APOBEC3G) context. Two sequences (1.8%) carried equal numbers of GG-to-AG and GA-to-AA mutations (signified by “*”), and 18% carried G-to-A mutations in non-APOBEC3-preferred contexts. Fourth, we examined whether there were multiple GG-to-AG mutations in the nonhypermutated sequences (4) and found only evidence of single GG-to-AG mutations. Fifth, we likewise examined if there were multiple GA-to-AA mutations in the nonhypermutated sequences (5) and found that 75% of the sequences carried a single GA-to-AA mutation, whereas 25% carried multiple GA-to-AA mutations, in contrast to the GG-to-AG pattern.

Thus, we determined that G-to-A mutations occurred more often than expected in APOBEC3F-favored contexts in sequences from acute, early, and chronic infection, consistent with a role of APOBEC3F (and possibly other deaminases) in the evolution of HIV-1 during the course of the disease.

Collectively, our data provide compelling evidence that G-to-A mutations accumulate over time in natural HIV-1 infection and that these mutations are unlikely to be caused by APOBEC3G but may be induced at least partly by other cytidine deaminases at all stages of disease.

DISCUSSION

In this study, we investigated the footprint of APOBEC3F and APOBEC3G activity in HIV-1 evolution in infected patients. We analyzed thousands of longitudinal, plasma-derived, nonhypermutated pol sequence pairs from more than 3,000 chronically HIV-1-infected patients (22) to examine whether it is likely that APOBEC3-induced mutations in vivo are always lethal for HIV-1 or whether they might occur at sublethal levels that increase viral diversification and aid in adaptation to the host.

As expected, G-to-A transition mutations were the most frequently occurring mutations over time (22, 32); however, our analyses revealed that their frequency varied greatly depending on the sequence contexts. Specifically, we observed that G-to-A mutations occurred more frequently in APOBEC3G-disfavored contexts than in APOBEC3G-favored contexts, strongly suggesting that APOBEC3G activity does not contribute to beneficial HIV-1 diversification in vivo. This result is consistent with our previous in vitro and in silico analyses of APOBEC3G, which indicated that the activity of even a single APOBEC3G unit is highly likely to be lethal for HIV-1 (21).

In contrast, we observed a 16% overrepresentation of G-to-A mutations in APOBEC3F-favored contexts and an underrepresentation of G-to-A mutations in APOBEC3F-disfavored contexts, which suggest that APOBEC3F contributes to HIV-1 evolution. However, these results do not rule out a minor role for other APOBEC3 proteins, such as, e.g., APOBEC3A (monocytes/macrophages only), APOBEC3B, APOBEC3C, APOBEC3DE, and/or APOBEC3H haplotypes II, V, and VII (34, 36–42) and/or a contribution from RT and/or RNA Pol II (34, 43–45). The activity of APOBEC3A differs from that of other deaminases because it exerts its antiviral effect in monocyte/macrophage producer cells (46) and may be targeted by Vpx (47) but does not restrict HIV-1 replication in the target cell (48). Homogeneous deletion of APOBEC3B increases the risk of HIV acquisition and AIDS progression in some cohorts (49) but not all (50), although it is not normally expressed in lymphoid cells (51). The antiviral effect of APOBEC3DE has been demonstrated in a single-cycle experimental system and in T-cell lines (52, 53) but not in human primary cells. The impact of APOBEC3H in natural infection is unclear. However, overexpression of APOBEC3H haplotype II restricts the replication of some lab-adapted HIV-1 strains (37, 54), and haplotypes II, V, and VII are incorporated into virions and restrict HIV-1 replication in 293T cells (54).

An alternative explanation for the deficit of G-to-A substitutions in APOBEC3G contexts is that most of the other observed changes were caused by previously occurring unknown positive selection and that the population adaptation to this selection pressure is complete. This would account for a surprising “lack” of events in some contexts because the expected “background” neutral rates would be overestimated from a set of changes that were, in reality, positively selected. However, we suggest that this explanation is unlikely, particularly because most of the observed changes were synonymous.

A potential limitation of our analyses is the assumption that alternative equivalent mutations are silent. Although natural selection occurs at the amino acid level and mutations that result in incorporation of the same amino acid therefore can be considered silent, some mutations could theoretically facilitate other and/or further mutations at the next time point. However, if this should occur, we estimate that the effect would be so weak that it would have no effect on our overall results.

The results of the chronic HIV-1 infection analysis are strongly supported by the analyses of the envelope sequences from acute and early infection (29). G-to-A mutations occurred in APOBEC3F-favored contexts more often than expected in these sequences, which suggests that APOBEC3F (and possibly other deaminases) may play a role in HIV-1 evolution throughout the course of the disease, e.g., by facilitating immune escape (55). Because our previous analyses of almost-full-length genomes of in vitro-mutated HIV_IIIB and in vivo patient-derived hypermutated sequences demonstrated similar patterns and levels of APOBEC3-induced mutagenesis (12, 21), the limitation that we are not able to compare the mutational context in the same gene is unlikely to affect the results.

Studies of the consequences of APOBEC3G activity for patient-derived HIV sequences are an area of active research. Previous studies either support our finding that APOBEC3G is unlikely to contribute to viral diversification (22, 56, 57) or offer no evidence that contradicts our findings (55, 58). The impact of APOBEC3G in HIV-1 evolution has also been studied extensively in vitro using a wide range of assays, lab-adapted HIV-1 strains, reporter genes, and cell lines, and experiments often include the nucleoside analogue RT inhibitor 2′,3′-dideoxy-3′-thiacytidine (3TC or lamivudine) (7, 21, 59–61). The use of 3TC is potentially problematic because it accumulates to different degrees in different cell lines (62–65) and increases intracellular dATP levels (66), possibly affecting RT misincorporation rates (67). The relevance of these studies for APOBEC3G's role in HIV evolution in vivo is unclear and requires further examination.

The hypothesis that APOBEC3F-induced mutations in natural infection may occasionally be tolerable during viral propagation is supported by in vitro studies demonstrating that APOBEC3F has minimal or no antiviral activity in primary cells or when APOBEC3F is expressed in cell lines at levels similar to those in primary cells (68, 69). To some extent, these findings conflict with the antiviral activity of APOBEC3F observed in other in vitro studies, although the observed antiviral effect was consistently less than that due to APOBEC3G (4), and Vif-deficient lab-adapted HIV-1 produced in an APOBEC3F-null/APOBEC3G cell line did not exhibit a significant change in the fraction of hypermutations that occurred within 5′GG or 5′GA dinucleotides (41). However, the latter study did not consider any contribution from RT and reported that mutations in the 5′GA context were nearly absent only in HIV-1 produced in APOBEC3F-null/APOBEC3DE knockdown cells; this finding contrasts with other reports that APOBEC3DE has no or little antiviral activity (34, 57, 70).

The conflicting results of in vitro studies highlight the need to examine patient-derived HIV-1 sequences for footprints of APOBEC3 activity to understand the consequences of this activity for HIV-1 evolution in natural infection. The results of the present study are largely in agreement with those of a cross-sectional study of approximately 2,000 full-length HIV-1 sequences from the HIV database that found no evidence of an evolutionary footprint of APOBEC3G (56). However, in contrast to the results presented here, the previous study identified only a nonsignificant trend that suggested footprints of APOBEC3F activity. The latter result could be due to the higher noise levels of the previous study, which may have been the result of the cross-sectional nature of the study and the use of HIV B sequences from patients in different disease stages. Analyzing longitudinal sequences allowed us to distinguish between founder effects and mutations that develop within a patient and potentially facilitated the identification of nonlethal, APOBEC3-induced mutations, which are likely to accumulate over time, by comparing within-patient sequences from different time points.

Both overlapping and distinct Vif domains are involved in the binding of APOBEC3F and APOBEC3G and other deaminases, such as, e.g., APOBEC3C and APOBEC3DE (10, 40, 56, 71–76); however, it is likely that a single Vif molecule binds to only one APOBEC3 molecule. Because the number of Vif molecules in HIV-infected cells at any given time point is finite, we hypothesize that viral selection may favor the preferential binding of Vif to APOBEC3G because it would confer a greater survival advantage to HIV-1 than binding to other deaminases. We acknowledge that APOBEC3F and other deaminases can restrict HIV-1 proliferation in vitro (53, 77–79) and possibly in vivo, and thus we hypothesize only that the lethal antiviral effects of these deaminases are less than that of APOBEC3G.

The greater viral survival advantage of APOBEC3G-Vif neutralization is suggested by the following: (i) APOBEC3G is more likely to generate stop codons than APOBEC3F and other deaminases (3, 12), (ii) the processive single-stranded DNA scanning behavior of APOBEC3G (“jumping” and “sliding”) increases its mutagenic potential relative to that of APOBEC3F (predominately “jumping”) (80), (iii) most (81–86) although not all (87) studies demonstrate that APOBEC3G expression levels correlate inversely with viral load and disease progression in adults, (iv) APOBEC3G expression levels are higher in long-term nonprogressors than in uninfected controls and lowest in rapid progressors (81), (v) APOBEC3G mRNA is expressed at higher levels than that of APOBEC3F in lymphocytes and lymphoid tissue (88, 89), and APOBEC3F is not expressed in monocytes/macrophages and dendritic cells, whereas APOBEC3G is (71), (vi) APOBEC3F and other deaminases have little or no antiviral effect in primary cells or when expressed at similar levels in cell lines (68, 69), although a cumulative antiviral effect of APOBEC3F, albeit less than that of APOBEC3G, was observed in a long-term (27 days) culture of lab-adapted virus in CD4⁺ T cells and monocyte/macrophages (4), (vii) APOBEC3F has no significant effect on virus infectivity when stably expressed in HeLa cells, even though it was packaged into nascent virions (68), (viii) the protein expression levels of APOBEC3G are higher than those of APOBEC3F in peripheral blood mononuclear cells (PBMCs) from different donors, although the use of different antibodies precluded an accurate quantification of protein expression (69), and (viiii) APOBEC3G had a stronger antiviral effect than APOBEC3F when they were expressed as fusions with the FLAG-epitope at the same levels in cell lines (90). We acknowledge that there may be a discordance between mRNA and protein expression levels and antiviral effects and that these studies do not analyze the degree to which each APOBEC3 inhibits reverse transcription, stimulates cDNA degradation, or inhibits integration; however, these limitations do not challenge the conclusion that APOBEC3G activity is more deleterious to HIV survival than the activities of other deaminases (4).

Based on the results of the analyses performed in this study, we hypothesize that Vif selection for preferential binding to APOBEC3G results in the efficient elimination of APOBEC3G from assembling virions while allowing some incorporation of APOBEC3F (and/or other deaminases) into virions. In addition, it is possible that selection pressures may favor both the preferential binding of APOBEC3G and the less efficient binding of APOBEC3F and possibly other deaminases. This hypothesis is supported by reports that APOBEC3G and APOBEC3F rarely comutate the same HIV-1 genome (91) and by a study of plasma-derived HIV-1 RNA env sequences in which 10/3,449 sequences were hypermutated and 9/10 predominantly carried APOBEC3F-type GA-to-AA mutations (57).

However, other factors beyond viral selection pressures may affect the interaction of Vif with APOBEC3. The pleiotropic effects of mutations caused by other selective pressures could affect Vif binding to APOBEC proteins and/or Elongin C, such as the selective pressure of Vif-specific cytotoxic T cell (CTL) responses. CTLs recognize fragments of viral proteins (CD8 epitopes) that are presented by HLA on the infected cell surface and subsequently kill the infected cell. Numerous studies have observed that CTLs frequently target CD8 epitopes in Vif (92–98). The pleiotropic effects of CTL-induced escape mutations that prevent viral recognition and the killing of infected cells depend on the affected Vif region and may destabilize the interactions between Vif and APOBEC3F, APOBEC3G, and/or other deaminases or the interactions between Vif and Elongin C. The importance of the specific Vif sequence in determining the biological outcome of the interaction between Vif and a range of APOBEC3 proteins is emphasized by reports demonstrating that, e.g., APOBEC3C and APOBEC3B are neutralized by some Vif variants but not others (40) and that Vif subtype variability affects the efficiency of the interaction with APOBEC3G (99).

We hypothesize that suboptimal Vif-APOBEC3G interactions are the result of a trade-off by the virus between competing fitness costs. For HIV-1, the disadvantage of producing defective viral progeny as a result of hypermutation is outweighed by the possible advantages of either nonlethal APOBEC3F-induced mutagenesis (and/or nonlethal mutagenesis by other deaminases) or CTL escape to viral progeny that avoid hypermutation (illustrated in Fig. 4). Because defective viruses generated by APOBEC3G have been demonstrated to potently activate HIV-specific CTL responses (100), hypermutation may increase the magnitude of CTL responses and the killing of infected cells, which would subject the HIV-1 population to greater fitness costs and increase the selective pressure to escape these CTL responses. However, Vif-specific CTL responses may be harmful to the host when they result in mutations that increase nonlethal APOBEC3F-induced mutagenesis (or nonlethal mutagenesis by other deaminases) and thereby facilitate viral adaptation. This could potentially lead to an increase in the HIV load in plasma, as suggested by a report of significantly higher viremia in patients with Vif-specific CTL responses (95).

FIG 4.

Model of proposed pleiotropic fitness effects of a CTL escape mutation. In this model, wild-type Vif can exclude all APOBEC3F and APOBEC3G units but encodes a CTL epitope that can be recognized by the host's CTL response. A CTL escape mutation within this epitope results in a hypothetical 20% decrease in the ability of the mutated Vif to bind to APOBEC3F and APOBEC3G. (A) Incorporation of APOBEC3G. We assumed that incorporation of APOBEC3G into HIV-1 resulted in viral inactivation in 100% of cases (based on the findings in reference 21). Thus, the incorporation of APOBEC3G in 20% of nascent virions results in inactivation of 20% of the viral population. (B) Incorporation of APOBEC3F. Based on the results of the analysis of sequences from chronic infection (Fig. 3), we estimated that about 10% of the sequences with APOBEC3F-induced G-to-A mutations would be hypermutated. Thus, if APOBEC3F were incorporated into 20% of nascent viruses, only 10% of these would be hypermutated, resulting in inactivation of 2% of the viral population, whereas 18% would carry nonlethal APOBEC3F-induced mutations that might facilitate viral adaptation to the host.

In conclusion, we demonstrated that APOBEC3G activity is unlikely to increase HIV-1 evolution in vivo, whereas the activity of other members of the APOBEC3 family, specifically APOBEC3F, may increase HIV-1 diversification through sublethal mutagenesis and thereby facilitate viral adaptation to the host. Therefore, vaccine-induced Vif-specific CTL responses and drugs that manipulate the interplay between Vif and APOBEC3 should be tested with caution because they may have either beneficial or detrimental clinical effects on the patient depending on how they affect the binding of Vif to various members of the APOBEC3 family.