Early immune adaptation in HIV-1 revealed by population-level approaches

Assembling early and chronic infection cohorts matched for size, HLA and HIV-1 diversity

Our study sought to demonstrate that the extent, reproducibility and relative timing (early versus later) of HLA-driven escape in HIV-1 can be inferred via comparative analysis of independent cross-sectional host/virus genotype datasets from different infection stages. This strategy ideally requires cross-sectional datasets that are identically powered with respect to host and viral genetic diversity (i.e. datasets that mimic longitudinal data as closely as possible, in that they differ only with respect to infection stage of the participants). As such, our first step was to assemble early and chronic HIV-1 subtype B cohorts of identical size that were matched as closely as possible for HLA class I allele distribution and HIV-1 diversity. We did so by drawing upon host and viral genotype data from early and chronic infection cohorts in North America, Europe and Australia (methods and [13],[23]-[25]). Our final early and chronic datasets comprised 221 Gag, 203 Pol and 219 Nef HIV-1 subtype B sequences per cohort, for which linked HLA class I types were available. Early cohort patients were recruited a median of 88 [IQR 63–120] days following infection. All early and >75% of chronic patients were antiretroviral naïve; the remainder were untreated at time of sampling.

A total of 59 HLA class I alleles, classified at subtype-level (4-digit) resolution, were observed at a frequency >1% in the early and/or chronic cohorts; these comprised 17 HLA-A, 23 HLA-B and 19 HLA-C alleles (Figure 1). Of these 59 alleles, the frequencies of 56 (94.9%) were comparable between cohorts; only three alleles (HLA-A*02:06, A*30:02 and B*39:01) exhibited significantly different frequencies between cohorts (all p < 0.05; 0.08 < q < 0.33) (Figure 1A-C). As such, our early and chronic cohorts were generally well-matched with respect to host HLA diversity.

HIV-1 Gag, Pol and Nef diversity was also generally comparable between the two cohorts. Mean patristic (pairwise) genetic distances between HIV-1 sequences in early versus chronic datasets, measured in units of substitutions per nucleotide site, were 0.076 (Standard Deviation [SD] ± 0.011) versus 0.071 (SD ± 0.010) respectively for Gag (Figure 1D left and middle panels), 0.057 (SD ± 0.008) versus 0.053 (SD ± 0.008) for Pol, and 0.119 (SD ± 0.018) versus 0.120 (SD ± 0.021) for Nef (not shown). Moreover, no gross inter-cohort segregation was observed in a combined HIV-1 Gag phylogeny (Figure 1D, right), indicating that neither cohort was dominated by large epidemiologically linked clusters nor exhibited evidence of recent descent from distinct ancestors. Together, these data suggest that our early and chronic datasets are similarly powered with respect to host and viral genetic diversity, and thus differ only with respect to infection stage.

Defining the list of HLA-associated polymorphisms for investigation in cohorts of the present size and composition

A total of 453 HLA-associated polymorphisms in Gag/Pol/Nef had previously been identified at q < 0.05 in an independent cohort of N > 1800 individuals with chronic HIV-1 subtype B infection [11], which contained no overlap with the cohorts studied here. These HLA-associated polymorphisms comprise “adapted” associations (HIV-1 amino acids significantly over-represented in the presence of the HLA allele in question) as well as “nonadapted” associations (HIV-1 amino acids significantly under-represented in the presence of the HLA allele). For example, at Gag codon 242 the nonadapted amino acid associated with HLA-B*57:01 is the subtype B consensus “T” whereas the B*57:01 adapted form is “N”, denoted as “B*57:01-Gag-T242N”. The cohort wherein these HLA-associations were originally defined however [11] was more than seven times larger than the cohorts presently studied. Therefore, we do not have sufficient statistical power to interrogate all of them in the present study. As such, our next step was to define, from the published list [11], the subset of known HLA-associated polymorphisms that is appropriate for study in cohorts of the present size and host/viral genetic composition.

Theoretically, if all immune escape mutations, once selected, persisted for the remainder of the host’s lifetime, and if we had achieved perfect genetic matching between our early and chronic cohorts, then we could define an appropriate subset of HLA-associated polymorphisms by interrogating our chronic cohort for the presence of these N = 453 known HLA-associated polymorphisms. Those detectable at the population level in chronic infection, a stage when a majority of within-host adaptation has already occurred, would represent an appropriate subset for study in cohorts of the present size. Thus we first interrogated our chronic cohort for the presence of these 453 published HLA-associated polymorphisms using statistical association with phylogenetic correction (see Methods and [12],[26]), and in doing so identified 157 (35%) “adapted” and “nonadapted” HLA associations at p < 0.01 (corresponding to q < 0.01 in this analysis) (Figure 2 and Additional file 1). These comprised 54, 52 and 51 HLA-associated polymorphisms in Gag, Pol and Nef respectively.

However, if genetic matching of early and chronic cohorts was imperfect, and/or if immune escape mutations occasionally arose transiently during infection, then defining HLA-associated polymorphisms based on the chronic cohort only could potentially lead us to miss some associations. Thus, we interrogated our early cohort for this same list of published HLA-associated polymorphisms using identical methods. In doing so we identified an additional 5 HLA-associated polymorphisms at p < 0.01, two occurring at Gag codon 401 and one each occurring at Nef codons 38, 102 and 133, that had not been identified in the chronic analysis (Figure 2 and Additional file 1).

We therefore took the union of these results, totaling N = 162 HLA-associated polymorphisms, as our definition of known HLA-associated polymorphisms appropriate for cohorts of the present size and composition (Figure 2 and Additional file 1). To further validate this subset, we applied a published phylogenetically-corrected interaction test (see Methods and [12],[27]) to compare the strengths of selection of these individual HLA-associated polymorphisms in early versus chronic cohorts. Given that within-host HIV-1 adaptation increases over the infection course via the selection of immune escape mutations, we would expect higher strengths of association between the restricting HLA and the HIV-1 polymorphism at later versus earlier stages: indeed, of the 162 HLA-associated polymorphisms studied, the strengths of 101 (62%) of them differed significantly (p < 0.01, q ≤ 0.1) between early and chronic infection (Additional file 1). Note also that none of the N = 162 polymorphisms were restricted by the three HLA alleles (A*02:06, A*30:02 and B*39:01) whose frequencies differed significantly between cohorts. All downstream analyses therefore focused on this list of N = 162 HLA-associated polymorphisms.

Fifteen percent of known HLA-associated polymorphisms are already detectable at the population level in early infection

As expected, among these were escape mutations known or previously observed to occur in the first year of infection, including B*57:01 Gag-T242N, B*51:01 RT-I135X and Int-L28I, C*03:04 Nef-V85L, A*11:01 Nef-K92R and A*24:02 Nef-Y135F [2],[13],[16],[18],[19],[21],[28]. Our findings therefore provide proof-of-concept that the most consistently rapid host adaptations in HIV-1 can be identified using cross-sectional methods. It is notable that, by < 3 months post-infection, the magnitude of statistical association between certain HLA alleles and their associated viral polymorphisms is already very high, and in some cases not significantly different from their magnitudes of association in chronic infection. For example, the Odds Ratio [OR] of association between B*57 and Gag-T242N is 33 in early infection (p = 3 × 10^-9) compared to 151 (p = 5 × 10^-16) in chronic infection, which, though stronger during the latter stage, does not represent a statistically significant difference (inter-cohort comparison p = 0.3) (Table 1). This observation underscores the rapid and highly reproducible nature of certain HLA-driven within-host adaptations in HIV-1, where, for certain mutations such as Gag-T242N, escape (and by extension our ability to detect this association via population-level methods) is already near maximal in B*57:01-expressing persons < 3 months post-infection.

Of note, HLA-associated HIV-1 polymorphisms with strong early population-level escape signal also included understudied viral sites. Notable among these were A*31:01-associated polymorphisms at Gag codons 397, 401 and 403, the first of which represented the second strongest p-value detected in the early cohort (Odds Ratio = 6.2, p = 1 × 10^-7, Table 1). These associations are located within the novel A*31:01-restricted CR9 CD8+ epitope originally characterized via detailed longitudinal analysis of a single HIV-1 subtype B-infected individual [18],[19]. By definition, population-level studies identify viral adaptations that occur reproducibly in persons expressing the restricting HLA; as such, the present results extend those of the original individual-level study [18] by indicating that escape within CR9 is both rapid and highly consistent in HLA-A*31:01-expressing persons. By extension, a lack of population-level early escape signal does not necessarily mean that a given site never escapes early: rather, it indicates that a given site does not reproducibly escape early (or at least does not do so to an extent that achieves statistical significance in a dataset of the present size). For example, very rapid (<30 days) escape was previously documented within the A*01-restricted GY9 epitope (Gag codons 71–79) in two HIV-1 subtype C-infected persons [19], but no evidence of reproducible early escape Gag codon 79 in A*01-expressing persons was observed in our early dataset, suggesting that rapid A*01-driven escape at this position is atypical in HIV-1 subtype B.

Taken together, population-level analyses extend those of individual-level studies by identifying escape mutations that are rapidly and reproducibly selected across patients. Our observation that 15% of known HLA-associated polymorphisms, notably those in Nef and Gag, are already detectable < 3 months post-infection, underscores the predictable and rapid nature of HIV-1 adaptation despite each individual’s unique combination of host HLA and transmitted virus genetics. Further, the detection of substantial population-level escape signal within unknown or understudied CD8+ epitopes in HIV-1 (Table 1) argues for continued efforts to map novel epitopes commonly targeted during this critical infection stage.

Can population-level approaches identify transient early escape pathways?

Recent longitudinal studies have revealed that immune escape is often characterized by the initial appearance of transient mutant forms that often retain some ability to be targeted by existing (or de novo) CTL [18],[29], which then drive the selection of more effective escape variants that ultimately become fixed within the host [3],[17],[18]. If such “transient” escape pathways are reproducible across hosts, we wondered whether population-level approaches could theoretically be used as exploratory tools to identify them. If so, we reasoned that such transient escapes would display stronger population-level escape signal in early compared to chronic infection (since, in some persons, the early variant would be subsequently replaced with another, thereby reducing population-level signal in later stages). Although our inter-cohort comparative analysis revealed no HLA-associated polymorphisms that displayed significantly stronger signal in early versus chronic infection (Additional file 1 and data not shown), we were nevertheless intrigued by the five HLA-associated polymorphisms in Gag and Nef that exhibited population-level escape signal of p < 0.01 in our early cohort but p ≥ 0.01 in the chronic cohort (Table 1), suggesting these as possible transient escape pathways.

Escape prevalence in early infection correlates with longitudinal first-year escape rates

Another objective was to investigate to what extent cross-sectional data could be used to infer the extent and time course of immune-driven HIV-1 adaptation. As such, we first wished to demonstrate that early escape frequencies calculated cross-sectionally predict rates of immune escape calculated longitudinally. Published first-year rates of escape were available for 27 optimally-described CD8+ T-cell epitopes [13] which contained one or more HLA-associated polymorphisms investigated in the present study. For example, the estimated first year escape rate for the Gag-TW10 epitope (Gag_240-249) is 38.36% per person-month [13], while the prevalence of the Gag-T242N mutation among B*57-expressing persons in our early dataset is 67% (Figure 3). As expected, longitudinal first-year CD8+ epitope escape rates correlated significantly with HIV-1 polymorphism prevalence among persons expressing the relevant HLA in our early infection dataset (Pearson’s R = 0.68, p = 0.0001; Figure 3). Because ~40% of the patients in the present early infection cohort were included in the published longitudinal study [13], we re-analyzed our data with these overlapping patients removed, and observed that the correlation remained strong (Pearson’s R = 0.55, p = 0.0035, not shown). This supports HLA-associated escape mutation prevalence calculated cross-sectionally at < 3 months post-infection as a reliable surrogate marker of first year escape rates calculated longitudinally.

Inferring the extent of host adaptation via comparative analysis of cross-sectional data from early and chronic infection

We next wished to use our cross-sectional early and chronic cohorts to quantify the extent of HLA-driven escape occurring between these two infection stages. For this analysis, we specifically defined “escape” as the specific adapted viral form associated with a given HLA allele at a given HIV-1 codon - for example, Gag 242 N is the B*57:01-associated adapted form at this position. This adapted list comprised N = 74 HLA-associated polymorphisms (25, 24, and 25 in Gag, Pol and Nef respectively) (Figure 2 and Additional file 1). We calculated the prevalence of each of these polymorphisms in persons expressing the relevant HLA allele in our early versus chronic cohorts, thus allowing us to estimate the extent of within-host HIV-1 adaptation between these two stages. Overall, the median “percentage escaped” (defined as the % of individuals expressing the relevant HLA and harboring the HIV-1 polymorphism of interest) was 23.8% [Interquartile range (IQR) 5.3-44.4%] in early infection versus 55.1% [IQR 28.4-73.0%] in chronic infection (p < 0.0001; Figure 4A). This indicates that, on average, escape prevalence in persons expressing the restricting HLA allele more than doubles between these infection stages. Breaking the analysis down by HIV-1 protein, the median early versus chronic escape prevalence was 23.5% [IQR 15.1-49.7%] vs. 55.6% [IQR 29.7-85.7%] in Gag, 11.3% [IQR 1.2-33.3%] vs. 50.5% [IQR 25.2-69.2%] in Pol, and 31.3% [7.5-63.0%] vs. 54.6% [21.4-73.3%] in Nef (all p ≤ 0.001, not shown). This is consistent with early escape occurring predominantly in Gag and Nef [13],[14],[19], while escape in Pol is generally slower but nevertheless approaches comparable levels by chronic infection.

Though summary statistics are informative, individual polymorphisms differ widely in their timing and extent of selection over the infection course. For this reason, details on polymorphism prevalence in persons expressing vs. not expressing the relevant HLA in early and chronic infection, along with their statistical measures of association, are provided in the Additional file 1. We highlight some examples here. First, for a substantial minority of polymorphisms (notably those in Table 1), escape is rapid, reproducible and largely complete within < 3 months post-infection. For example, 67% of B*57:01-expressing individuals already harbored Gag-242 N in early infection, a proportion that increased to 83% in the chronic phase (Figure 4A). Noting that Gag-242 N frequency was only 5.8% among persons lacking B*57:01 in early infection (Additional file 1), these results are consistent with escape having already occurred in over two-thirds of B*57-expressing persons by < 3 months post-infection [2],[13], with an additional minority escaping somewhat later.

While the prevalence of Gag-242 N is low in the general population (5.8% among B*57:01-negative persons and ~1% among persons lacking an allele belonging to the B58 supertype), other polymorphisms are quite prevalent in circulation, but are nevertheless significantly enriched among HLA-expressing persons in early infection (Figure 4A). In this case, their high early prevalence is attributable to both frequent transmission and reproducible early escape. For example, both the A*31:01-associated Gag-403 K and C*03:04-associated Nef-85 L polymorphisms are observed at >40% prevalence in HIV-1 subtype B sequences, but their prevalence is ~55% and ~72% respectively among persons expressing the relevant HLA < 3 months post-infection (Odds Ratios 1.5 and 2.4 respectively, p < 0.01, Table 1 and Additional file 1). The observation that population-level approaches are capable of detecting strong escape signals despite high polymorphism background frequencies has previously been demonstrated in high-powered chronic infection cohorts [11]; the present study extends this to demonstrate such signals can also be detected very early in infection, in more modestly-powered datasets. Overall, if one uses the original criterion of early population-level statistical signal of p ≤ 0.01 to define HIV-1 sites that predominantly escape early, 15% (11 of 74) of adapted polymorphisms fall into this category; using a more liberal threshold of p < 0.05, this increases to 21.6% (16 of 74) (Figure 4B).

The remaining 78.4% (58 of 74) polymorphisms generally reproducibly escape later than 3 months following infection (Figure 4B). It is important to note that later escape can occur because CTL responses against these regions generally arise later during infection (i.e. there is no immune pressure on these epitopes in early infection), or because CTL responses arise relatively early, but escape does not reproducibly occur rapidly in a significant proportion of individuals expressing the relevant HLA allele. Among these later-escaping polymorphisms are those whose population background (transmission) frequencies are generally low, and those whose background frequencies are generally high. The B*51:01-associated Integrase-32I polymorphism at position 5 of the B*51-restricted LI9 epitope (Integrase_28-36) provides an example of the former. In early infection, its frequency in B*51:01-expressing persons is 5%, not significantly different from background, but this rises to 64% by chronic infection (Figure 4A and Additional file 1). The LI9 epitope is known to be consistently targeted in B*51-expressing persons early after infection [30],[31]. The observation that this epitope ultimately escapes via Int-32I in >60% of B*51:01-expressing persons suggests this epitope is under strong, sustained and reproducible CD8+ T-cell pressure by B*51 in vivo, where delayed escape is likely explained by a combination of mutational/fitness constraints and both intra-individual (“vertical”) and inter-individual (“horizontal”) CD8+ T-cell immunodominance hierarchies [19],[32],[33].

An example of a later-escaping polymorphism with high population background frequency is B*44:02-Gag-312E. Located at position 7 of the B*44:02-restricted AW11 epitope (Gag_306-317), it represents the HIV-1 subtype B consensus residue at this codon. Its >60% frequency in both B*44:02 and non-B*44:02-expressing persons in early infection reflects its high transmission frequency, rather than early selection by B*44:02. Nevertheless, by chronic infection, 83% of B*44:02-expressing persons harbored Gag-312E, consistent with later escape (Figure 4A). A full categorization of HLA-associated “adapted” polymorphisms in terms of “earlier” vs. “later” escaping (defined as early p < 0.05 vs. p ≥ 0.05 respectively) and “lower” vs. “higher” background (estimated transmission) frequency (defined as <30% vs. p ≥ 30% respectively), is provided in the Additional file 1. A graphic depicting the proportion of HLA-associated polymorphisms in each of these categories, broken down by HIV protein, is provided in Figure 4B.

The extent of reversion of HLA-associated polymorphisms over time can be similarly estimated by calculating their prevalence in persons lacking the relevant HLA in the early versus chronic cohorts. The overall median percentage of individuals harboring a given polymorphism in the absence of the restricting HLA allele was comparable in early (13.7% [IQR 4.9-34.2%]) and chronic (14.9% [4.2-34.9%]) infection (p = 0.15; Figure 4C), consistent with slow reversion reported for many polymorphisms [7],[14],[21],[34]. Note that inferred reversion frequencies merit cautious interpretation in cases where polymorphisms are selected by multiple alleles (e.g. the seemingly stable prevalence of Gag-147 L in A*25:01-negative individuals is likely due in part to its selection by B*13:02 and B*57:01 [11], associations that were not investigated in the present study). Nevertheless, results confirmed that HLA-B*57:01-Gag-T242N reverts between early and chronic infection (though RT-245E, Int-122I, Int-124N, or Nef-116N revert slowly or not at all, as reported previously [2],[21],[35]). The reversion analysis additionally revealed novel sites of potential interest. For example, the early escaping A*31:01-Gag 397R polymorphism (Table 1 and [18],[19]) displayed evidence of reversion, suggesting that this mutation may have a high fitness cost.

Host adaptation-related features distinguish protective and non-protective HLA class I alleles

Lastly, we wished to identify adaptation-related features that discriminate protective from non-protective HLA alleles, defined here as their published hazard ratios for progression to AIDS [HR-AIDS] in natural history studies [36]. Although the timecourse of viral escape is influenced by complex factors including epitope immunodominance hierarchies, strength of selection, mutational/fitness constraints and transmitted virus characteristics [17],[19],[34],[37]-[40], we reasoned that HLA alleles that restrict polymorphisms that are already highly prevalent in early infection (due to rapid escape and/or frequent transmission) would be generally unfavorable for HIV-1 control. Thus, for all HLA alleles for which ≥2 adapted polymorphisms were investigated in the present study (N = 17 alleles total), we computed their mean “percentage escaped” in early infection. That is, we took the prevalence of each of these adapted polymorphisms in persons expressing the relevant HLA allele in our early cohort (displayed in Figure 4A), and, for each HLA allele, computed the mean of these values.

Consistent with our hypothesis, we observed a positive correlation between an HLA allele’s average extent of adaptation in early infection, and its HR-AIDS (Pearson’s R = 0.53, p = 0.028; Figure 5A). Of note, HLA-B*57:01 appears as somewhat of an outlier, exhibiting higher than expected escape prevalence in early infection given its protective nature. We hypothesize that the reason B*57 can maintain sustained HIV-1 control despite rapid escape in some epitopes (Table 1 and [2],[13],[16],[18],[28]) is because the early B*57-restricted CD8+ response often simultaneously targets more than one epitope, notably in p24^Gag[31],[41]-[43], where escape is accompanied by fitness costs [44]-[46].

HLA-associated polymorphisms identified at the population level mark viral sites under strong, reproducible in vivo immune pressure by individual HLA alleles [11]. We thus further hypothesized that HLA alleles for which escape was substantial (i.e. selected in a high proportion of persons expressing the relevant HLA) but generally delayed (i.e. selected post early-phase) would tend to be more protective. As such, for the same set of HLA-associated polymorphisms we computed their fold-change in escape between early and chronic infection, and calculated the mean of these values per HLA allele. By this measure, alleles for which the majority of escape had already occurred in early infection would exhibit low subsequent fold-changes, whereas alleles selecting escape mutations that generally occurred later in infection would exhibit fold-changes reflecting the extent of selection pressure on these sites in persons expressing the relevant HLA. Consistent with our hypothesis, we observed an inverse correlation between an HLA allele’s HR-AIDS and its average fold-increase in escape in chronic versus early infection (Pearson’s R = 0.54, p = 0.025; Figure 5B). Of note, B*27 appears as an additional outlier in this analysis, possibly due to escape in the critical Gag-KK10 epitope requiring nearly a decade in some individuals [32],[47],[48] due to its high fitness/mutational barrier [33],[49].

Overall, our findings are consistent with a high early burden of adaptation to host HLA (either via rapid escape or frequent polymorphism transmission) as a correlate of HLA-associated progression risk. Conversely, HLA alleles from which HIV-1 escape is substantial and reproducible yet occurs on a delayed timescale appears to be a correlate of protection. Taken together with observations that protective HLA alleles contribute substantially to the total HIV-specific CD8+ response in early infection [31],[38], that they impose broad selection pressures on HIV-1 [11]; and that some CD8+ epitopes escape slowly despite sustained CD8+ targeting [19],[48], results suggest that the capacity to exert consistent, substantial and sustained pressure, ideally on multiple epitopes, from which the virus can only escape on a relatively delayed timescale, is a correlate of protection.

Some limitations of our study merit mention. Firstly, our early and chronic datasets are relatively modestly powered by association testing standards, so it was not possible to examine all published HLA-associated polymorphisms in HIV-1 subtype B. Secondly, due to the lack of information on duration of infection for chronic patients, it is likely that the chronic cohort comprised patients at a range of infection stages. Inclusion of some chronic patients with less advanced infection could underestimate the extent of escape at this stage. Finally, although care was taken to match our early and chronic datasets as closely as possible for HIV and HLA genetic diversity and distribution, it is essentially impossible to achieve perfectly matched datasets. As such, we cannot rule out small differences in statistical power between cohorts, and therefore should interpret candidate “transient” early escape results with some caution.

Early and chronic infection cohorts

The early cohort was comprised of HIV-1 subtype B infected patients recruited through various observational seroconverter studies including the Acute Infection and Early Disease Research Program (AIEDRP) sites in Boston and New York (USA), Sydney (Australia), a private medical clinic in Berlin (Germany), and observational cohort studies in Montreal and Vancouver (Canada) [13],[23],[24]. Infection dates for the patients in the early cohort were estimated as described in [13],[23]. Briefly, for patients with positive HIV RNA (>5,000 copies/ml) or detectable serum p24 antigen but a negative HIV-1 enzyme immunoassay (EIA), 4 weeks were subtracted from the negative EIA date. For patients with positive EIA but negative/indeterminate Western blot, 6 weeks were subtracted from the positive EIA date. For patients with negative detuned EIA, 4 months were subtracted from this date. For the remainder, infection dates were estimated as the midpoint between the last negative and the first positive HIV test. Clinical histories were incorporated into infection date estimates where available.

To maximize our power to detect HLA-associated polymorphisms in the early infection stage, all available early infection patients were included in the present study. For each of these patients, the sample closest to ~3-months following the estimated date of infection was selected, yielding a median sampling distribution of 88 days [Interquartile Range 63–120 days] post-infection for early samples. In contrast, the chronic cohort was assembled from the baseline (pre-therapy) timepoint from a total of more than 300 HIV-1 subtype B infected individuals initiating antiretroviral therapy in British Columbia, Canada, and untreated HIV-1 subtype B infected individuals in Boston, USA [13],[25]. Time since infection is unknown for individuals in the chronic cohort, however the median CD4 count at sampling was 250 [IQR 147–360] cells/mm³ for this group. To create HIV-1 gene-specific chronic infection datasets of equal size to the early cohort, that were also matched as closely as possible for HLA class I and HIV-1 diversity of the early cohort, chronic patients were selected from the total group using an iterative process to achieve the closest matching of HIV-1 and HLA distributions (Figure 1). All early and >75% of chronic patients were antiretroviral naïve; the remainder were untreated at time of sampling.

HIV-1 and host (HLA class I) genotyping

HIV-1 RNA was extracted from plasma using standard methods. Gag, Pol (including protease, codons 1–400 of Reverse Transcriptase, and Integrase), and Nef were amplified in separate nested RT-PCR reactions using HIV-1 subtype B-specific primers. Amplicons were bulk-sequenced bidirectionally on a 3130xl and/or 3730xl automated DNA sequencer (Applied Biosystems). Chromatograms were analyzed using Sequencher v5.0 (Genecodes) or RECall [51] with nucleotide mixtures called if the height of the secondary peak exceeded 25% of the height of the dominant peak (Sequencher) or 20% of the dominant peak area (RECall). HIV-1 sequences were confirmed as subtype B using the recombinant identification program (RIP; http://www.hiv.lanl.gov/content/sequence/RIP/RIP.html) and aligned to the HIV-1 subtype B reference strain HXB2. Phylogenetic trees were constructed using PhyML [52] and visualized using FigTree (http://tree.bio.ed.ac.uk/software/figtree/). Pairwise genetic distances were computed from newick treefiles using PATRISTIC [53]. HLA class I typing was performed using sequence-based methods [54] and imputed where necessary to high resolution using a machine learning algorithm ([55]; http://research.microsoft.com/en-us/projects/bio/mbt.aspx#HLA-Completion). HIV-1 sequences from cohorts where REBs allow public sequence deposition have been deposited in GenBank: accession numbers are Gag (KJ869442 - KJ869609), Protease-RT (KJ869900 - KJ870015), Integrase (KJ869610 - KJ869735), Nef (KJ869736 - KJ869899). A full summary of polymorphism frequencies, broken down by HLA allele carriage and infection stage is provided as Additional file 1.

Definition and identification of HLA-associated polymorphisms

The published reference list of N = 453 HLA-associated polymorphisms in HIV-1 subtype B Gag, Pol and Nef sequences was defined in an independent international cohort of >1800 individuals chronically infected with HIV-1 subtype-B using phylogenetically-informed methods at q < 0.05 [11]. The cohort used to define these associations had no overlap with the early and chronic cohorts studied here [11]. Briefly, to identify HLA-associated polymorphisms in linked HIV/HLA datasets, maximum likelihood phylogenetic trees (one per HIV-1 gene) are constructed, and a model of conditional adaptation is inferred for each observed HIV-1 amino acid at each codon. The amino acid is assumed to evolve independently along the tree until it reaches the tips, representing the present host. Selection via host HLA-mediated pressures and HIV-1 amino acid covariation is directly modeled using a weighted logistic regression, in which the individual’s HLA repertoire and covarying HIV-1 amino acids are used as predictors, and the bias is determined by the inferred possible transmitted sequences (as inferred via reconstruction of the amino acid frequencies at the penultimate internal nodes in the phylogeny) [12]. Here, the null hypothesis is that the observed amino acids at the tree tips are explained by the phylogeny only, whereas the alternative hypothesis is that they are better explained by the presence of a specific HLA (or covarying HIV-1 amino acid) in the present host. To identify which factors (HLA and/or HIV-1 covariation) contribute to the selection pressure, a forward selection procedure is employed where the most significant association is added to the model in an iterative fashion, with p-values computed using the likelihood ratio test. Statistical significance is reported using q-values [56], the p-value analogue of the false discovery rate (FDR). Q-values denote the expected proportion of false positives among results deemed significant at a given p-value threshold; for example, at q ≤ 0.05, we expect 5% of identified associations to be false positives.

HLA-associated polymorphisms are classified into two categories: (1) “Adapted forms”, amino acids significantly overrepresented in the presence of the HLA allele in question, which represent the putative escape forms associated with that HLA at that codon, and (2) “Nonadapted forms”, amino acids significantly underrepresented in the presence of the HLA allele in question, which represent the immunologically susceptible form associated with that HLA at that codon. In most cases, HLA-associated nonadapted forms represent the subtype consensus amino acid while adapted forms represent polymorphic variants – but exceptions exist.

To identify an appropriate subset of known HLA-associated HIV-1 polymorphisms that are appropriate for study in datasets of the present size (N ~ 200) and host/viral genetic distribution, we interrogated our early and chronic infection cohorts for these N = 453 published polymorphisms [11] using the phylogenetically-corrected methods described above. As described in the results, this yielded a subset of N = 162 HLA-associated polymorphisms detectable in our early and/or chronic datasets (Figure 2 and Additional file 1).

Our analyses also featured comparisons of the strength of selection of HLA-associated polymorphisms between early and chronic cohorts, undertaken using a previously-described phylogenetically-corrected interaction test [12],[22],[27]. Briefly, we took the union of all HLA-associated polymorphisms detectable at the population level in either the early or chronic cohorts (N = 162). For each association on the list, we constructed a phylogenetically-corrected logistic regression model using the restricting HLA as a predictor. Using a likelihood ratio test, we compare this model to a more expressive one that includes an additional interaction term that assigns “1” if the individual expresses the restricting HLA allele and is in the chronic cohort, or “0” otherwise. This allows us to obtain a p-value testing the null hypothesis that HLA-associated selection at that site is not significantly different in early versus chronic cohorts.