Immunological biomarkers predict HIV-1 viral rebound after treatment interruption

Treatment of HIV-1 infection with antiretroviral therapy (ART) in the weeks following transmission may induce a state of ‘post-treatment control' (PTC) in some patients, in whom viraemia remains undetectable when ART is stopped. Explaining PTC could help our understanding of the processes that maintain viral persistence. Here we show that immunological biomarkers can predict time to viral rebound after stopping ART by analysing data from a randomized study of primary HIV-1 infection incorporating a treatment interruption (TI) after 48 weeks of ART (the SPARTAC trial). T-cell exhaustion markers PD-1, Tim-3 and Lag-3 measured prior to ART strongly predict time to the return of viraemia. These data indicate that T-cell exhaustion markers may identify those latently infected cells with a higher proclivity to viral transcription. Our results may open new avenues for understanding the mechanisms underlying PTC, and eventually HIV-1 eradication.

A ntiretroviral therapy (ART) for the treatment of Human Immunodeficiency Virus Type 1 (HIV-1) infection has dramatically improved survival 1-3 , but is not a cure-a consequence of a reservoir of latently infected cells 4 . Primary HIV-1 infection (PHI)-the period within weeks to months of seroconversion-may provide insights into how this reservoir is activated and whether it can be suppressed, as stopping ART initiated in this early stage has in certain patients been associated with periods of aviraemic remission 5 -in some cases for over 10 years 6 . Post hoc analyses of other larger PHI cohorts support this phenomenon of sustained post-treatment control (PTC) 7,8 , as well as periods of transient aviraemic remission 5 .
The mechanisms that induce and maintain this remissive state of PTC remain unclear. There is debate over whether the timing of viral rebound after stopping ART is a random, stochastic process based on the spontaneous re-awakening of HIV-1 in latently infected T lymphocytes, or whether the host-for example, through HIV-1-specific or innate immunity-is the determining factor. In the best characterized cohort to date (VISCONTI), PTC was associated with low (but detectable) reservoirs, almost non-existent HIV-specific immune responses, an absence of protective HLA Class I alleles and highly symptomatic seroconversion illnesses 6 -a distinct phenotype from the more widely recognised examples of spontaneous viraemic control shown by 'elite controllers' 9 . The SPARTAC trial 10 was a randomized controlled trial (RCT) comparing ART for either 48 or 12 weeks relative to standard of care in adults with PHI. We reported prolonged periods of aviraemia post-treatment interruption (TI) in participants allocated to receive ART for 48 weeks-14% had undetectable plasma viral loads 1 year after stopping ART 5 . Total HIV-1 DNA level at baseline and week 48 was associated with shorter time to viral rebound 11 .
Here we report an analysis of immunological and virological factors from SPARTAC to explore those that either associate with HIV-1 DNA levels or independently predict viral rebound on stopping ART. We find that pre-therapy measures of T-cell exhaustion (PD-1, Tim-3 and Lag-3) were statistically significant predictors of virological rebound, suggesting that these markers should be considered in future algorithms used to detect PTC.

Results
SPARTAC trial participant characteristics. About 154 participants recruited to the SPARTAC trial 12 were studied based on infection with subtype B HIV-1 and sample availability. SPARTAC incorporated a randomly allocated TI after 48 weeks of ART in a third of patients, providing a unique opportunity to study determinants of rebound. The demographics of the 154 participants are shown in Supplementary Table 1. All 154 patients were sampled for HIV-1 DNA levels at enrolment at the pre-therapy baseline (week 0), a median of 74 days from the estimated date of HIV-1 seroconversion. Of these (due to sample limitations), 78 were also assayed for immunological markers of T-cell exhaustion and activation (Supplementary Table 1). Participants who were randomized to receive 48 weeks of ART, and for whom samples were available at the point of stopping ART, (n ¼ 47; Supplementary Table 1) were studied in analyses of TI.
Selection of biomarkers. We devised a panel of 18 biomarkers to measure parameters of host immunity and markers of the HIV-1 reservoir (Table 1). We wished to determine associations with levels of HIV-1 DNA levels and assess whether the biomarkers might also independently predict rebound at TI. At baseline, biomarker assays were carried out on up to 154 patients (not all assays were conducted on all patients- Table 1 for breakdown). At the point of TI after 48 weeks of ART, 47 participants were available for analysis. Again, and due to sample availability, not all biomarkers were tested on all patients (Table 1).
T-cell immunity correlates with HIV-1 DNA at seroconversion. To determine the relationship between the HIV-1 reservoir and HIV-1-specific T-cell immunity prior to therapy, we studied HIV-1 Gag-directed T-cell responses by ELISpot. Participants who made a CD4 T-cell ELISpot response had significantly lower mean HIV-1 DNA levels compared with those who did not ( Fig. 1a; for total DNA P ¼ 0.018 (3.77 versus 4.01 log 10 copies, respectively) and for Integrated P ¼ 0.046 (3.44 versus 3.72 log 10 copies, respectively; t-test). Total (and to a lesser extent Integrated) HIV-1 DNA was inversely correlated with numbers of HIV-1 Gag-specific CD8 T-cell responses (P ¼ 0.046; b ¼ À 0.064; s.e. 0.032 and P ¼ 0.081, respectively; linear regression; Fig. 1b).
We then looked for associations with HLA Class I, the only consistent genetic association with HIV-1 progression 13 . For each HLA Class I allele in our cohort, the median value of Total HIV-1 DNA was calculated (numbers contributing to each HLA allele are in Supplementary Table 2). In Fig. 1c, HLA alleles were categorized as protective (red) or associated with rapid progression (blue) based on well-characterized associations with HIV-1 progression 14 . While individual favourable (for example, B*27 and B*57) and progressive (for example, B*35 and B*08) HLA Class I alleles were associated significantly with lower and higher Total HIV-1 DNA levels, respectively, we also found that, overall, participants with protective alleles (n ¼ 20) had significantly lower mean Total HIV-1 DNA levels than those with progression-associated alleles (n ¼ 55) (3.46 versus 4.05 log 10 HIV-1 DNA copies per million CD4 T cells; Po0.001 (t-test)) ( Fig. 1c).
Co-variation of biomarkers at PHI. Together, these data support a close interaction between T-cell immunity and the reservoir size in early HIV infection. An exploratory analysis was therefore carried out to illustrate co-variation within 17 immunological and reservoir-associated biomarkers, which were available at the baseline pre-therapy time-point. Results are presented graphically using a 'correlogram' 15 (Fig. 2) in which the order in which the variables are presented was determined by hierarchical clustering with optimal leaf ordering 16 . The heat map (lower triangle of the correlogram) indicates the direction (blue indicates a positive association and red indicates a negative association) and the strength of the correlation (intensity of colour). The same information is also conveyed in the series of pie charts (upper triangle), where the strength of the correlation is indicated by both the size of the pie slice and intensity of colour.
In Fig. 2, the correlogram shows negative associations between CD4 T-cell count and all markers of viral nucleic acid and immune function (except IL-6). PD-1, CD38 and HLA-DR expression on CD4 and CD8 T cells, as represented by the deeper blue shaded region, cluster together with similar positive correlations. The correlogram shows that eight immunological variables: D-dimer, Lag-3 (CD8 only), HLA-DR (CD4, CD8), PD-1 (CD4, CD8), CD38 (CD4, CD8), showed the strongest correlations with Total HIV-1 DNA and were selected for further analysis.
Correlates of Total HIV-1 DNA pre-ART and at TI. As Total HIV-1 DNA was a predictor of time to plasma viral load (VL) rebound in SPARTAC 11 , we wished to explore the joint association of these biomarkers with HIV-1 DNA in a regression analysis. In addition to the eight variables identified in the correlation analysis, we included three clinical markers associated with HIV-1 DNA: plasma VL, CD4 T-cell count and CD4/CD8 T-cell ratio, which were also evident in the correlogram. Plasma VL, CD4 T-cell count and CD4/CD8 ratio were all associated significantly (Po0.001; linear regression) with HIV-1 DNA levels, although in multivariable models (adjusting for just these measures) the CD4/CD8 ratio was no longer significant ( Table 2; Model A). We found strong, significant correlations with biomarkers for CD4 and CD8 T-cell activation and exhaustion, as well as D-dimer (  Table 3); significance (at Po0.05) was lost for HLA-DR). In a further model including plasma VL and CD4 T-cell count, the only biomarker to retain a significant association with HIV-1 DNA was CD8 Lag-3 ( Table 2, Model C).
Many of the immunological associations present pre-ART were not preserved following ART. After 48 weeks of ART (at the point of TI) the only statistical associations with contemporaneously measured Total HIV-1 DNA were unspliced cell-associated HIV-1 RNA and percentage HLA-DR on CD4 T cells (Supplementary Table 4).
Predictors of time to plasma HIV-1 viral load rebound. Would biomarkers associated with HIV-1 DNA also predict time to VL rebound after TI? Although sample limitation restricted the analysis of all biomarkers in all 47 patients, we could find no strong evidence for any biomarker (other than Total HIV-1 DNA) measured at TI predicting viral rebound. CD4 T-cell ELISpots measured at baseline or TI did not predict time to rebound ( Supplementary Fig. 1).
However, pre-therapy levels of T-cell exhaustion were strongly associated with time to VL rebound 4400 copies per ml after TI. In survival analyses dividing the cohort at median values, expression on CD4 and CD8 T cells of PD-1 (P ¼ 0.00017, P ¼ 0.0013), Tim-3 (P ¼ 0.0032, P ¼ 0.08) and Lag-3 (P ¼ 0.016, P ¼ 0.023), respectively, predicted viral rebound to 400 copies per ml (log rank test) ( Fig. 3a-f). These findings were supported by Cox models adjusting for HIV-1 DNA levels (Table 3;  Supplementary Table 5). Adjusting for multiple comparisons (Supplementary Table 6) had no impact on the significance for PD-1, although Lag-3 expression was no longer significant. When using a VL cut-off for rebound of 50 RNA copies per ml, Tim-3 (Po0.001 for CD4; P ¼ 0.0067 for CD8 T cells) and PD-1 (P ¼ 0.023 for CD4 T cells) were predictive of time to rebound (Supplementary Table 7).
No other baseline measures (including markers of activation, IL-6 and D-dimer) were significantly associated with time to rebound. While PD-1 and Lag-3 correlated with HIV-1 DNA levels, Tim-3 did not, suggesting a possible separate mechanism  Table 5). PD-1 and Lag-3 retained significance for expression on CD8 T cells only. Of note, none of PD-1, Tim-3 or Lag-3 predicted time to VL rebound when measured at TI (Supplementary Table 8). Although all badged as 'exhaustion markers' (or 'immune checkpoint markers), it is interesting to note that Tim-3, PD-1 and Lag-3 were predominantly expressed on separate cells. While all three markers were widely expressed on their own on CD4 (range 19.93-33.30%) and CD8 (range 15.00-50.29%) T cells, co-expression of any two markers was limited (Fig. 4).

Discussion
ART in PHI can, in some cases, induce a state of virological remission or PTC. In the search for a cure for HIV-1 infection, it is important to understand the underlying mechanisms that produce PTC and assess the possibility of identifying patients most likely to achieve PTC using an algorithm of biomarkers. The SPARTAC RCT was not designed to identify biomarkers that predict remission, yet the incorporation of a randomly allocated TI after 48 weeks of ART provided a unique opportunity due to the numerous virological and immunological studies undertaken within the protocol. TI studies have been set back by the increased mortality witnessed in the SMART study 17 , yet newer study designs in which ART is recommenced based on viral detection in plasma rather than CD4 decline may provide a safer option. Accordingly, a number of studies are currently being planned to assess the impact of ART on virological remission through intensively monitored TIs, with the assessment of biomarkers being a key endpoint. This analysis of SPARTAC therefore provides preliminary critical data to help inform future larger clinical studies and cohorts.  Although no long-term PTC was seen in this group of patients, we did find a prolonged period of aviraemia post TI-1 year after stopping ART 14% of participants were still undetectable 5 . In SPARTAC, Total (but not Integrated) HIV-1 DNA measurements at baseline and week 48 helped stratify patients according to their viral-rebound kinetics 11 . We now show that-prior to ART-numerous measures of T-cell immunity (both HIV-1 specific and non-specific) measured in    suggesting that these two processes-time to rebound and time to clinical progression-have different mechanisms. Indeed, in SPARTAC, these two time intervals after TI did not correlate ( Supplementary Fig. 3). While there is plenty of evidence to suggest that the progression is immune mediated, it is increasingly proposed-and supported by mathematical modelling 21,22 -that rebound is stochastic, or random 23 , that is, the initiation of viral transcription in a latently infected cell at levels adequate to result in plasma viraemia may not be initially governed by the immune response. These data suggest that the size of the reservoir is determined by T-cell-mediated immunity in early HIV-1 infection, reflected in the number of associations between HIV-1 DNA and immune biomarkers at baseline. Once ART is commenced and viral replication is suppressed, the CD4 T-cell count recovers and the HIV-1 reservoir (in reality, an unknown entity for which we have limited surrogate measures) declines. At the point of TI, two events can occur. (i) viraemia returns; and (ii) absolute CD4 T-cell numbers fall until a point at which ART needs to be reintroduced. We hypothesize that these two processes are determined by different mechanisms: (i) stochastic and dependent on the size of the reservoir size; and (ii) mediated by HIV-1-specific immune responses after viraemia has returned. Accordingly, it is possible that studies describing T-cell-mediated mechanisms driving 'elite controllers' may have little to offer our understanding of how to induce PTC or HIV-1 cure. This model may be supported by our finding that T-cell exhaustion markers at baseline (48 weeks prior to TI) predict time to viral rebound, whereas the same markers measured at the time of TI do not. (A caveat to this finding is that different antibody clones were used for the surface fluorescenceactivated cell sorting (FACS) staining at baseline and at week 48 and, due to the possibility of different affinities, this is a possible confounder which needs to be explored in future studies.) The highly significant impact of the markers PD-1, Tim-3 and Lag-3 is intriguing, especially in view of the lack of association for markers of activation (HLA-DR, CD38), which might be expected to associate with rebound. PD-1 has been widely reported to be upregulated on exhausted T cells in chronic viral infections such as cytomegalovirus, simian immunodeficiency virus (SIV) and HIV, in a process that is reversible through blockade of PD-1 with its ligand. Accordingly, some investigators have suggested that PD-1 is a marker of early exhaustion or even activation-the latter being supported by its clustering with HLA-DR and CD38 in the correlogram. Despite the identification of PD-1, Tim-3 and Lag-3 as potential biomarkers for viral rebound, this does not in itself prove that they are mechanistic. There is much interest in the therapeutic potential of blocking PD-1 (either directly or by antagonizing the PD-1 ligand), and even in the context of targeting the HIV-1 reservoir. Of note, the combination of anti-PD-1 (Nivolumab) and anti-CTLA-4 (Ipilimumab) monoclonal antibodies have shown encouraging results in clinical trials in untreated melanoma 24 , and one might speculate that this enhancement in T-cell function may also be applicable for targeting the HIV-1 reservoir. ARTICLE Tim-3 has been associated with T-cell exhaustion (loss of functionality and decreased proliferation) in the context of HIV-1 infection 25 . Tim-3 expression on both CD4 and CD8 T cells is elevated in HIV þ ve individuals compared with controls, correlates with VL and is associated with disease progression. As for PD-1, blocking the interaction of Tim-3 with its ligand can restore functionality. Interestingly, in the same study it was shown that PD-1 and Tim-3 identify distinct populations of cells, rather than being co-expressed on the same dysfunctional T cells, which we also found in our patients. This might help explain why they both predict time to rebound and yet both are not associated with HIV-1 DNA levels. Increased understanding of which cells express these markers and how they contribute to the HIV-1 reservoir could be important in developing new HIV-1 cure strategies. This will need to be addressed in larger studies carefully designed for this purpose.
The role of Lag-3 (CD223) in HIV-1 infection is less clear, although its upregulation has been associated with disease progression and functional T-cell exhaustion 26 . Lag-3 is structurally similar to the CD4 receptor and is a ligand for MHC Class II, and part of the Ig superfamily on memory T cells 27 . Mouse models have suggested some synergy between PD-1 and Lag-3 (ref. 28), although it is interesting that we only find limited co-expression in our data. However, now that the proof-of-efficacy has been demonstrated in cancer trials for combination immune checkpoint blockade 24 , it will be important to conduct further studies to explore how this applies to HIV latency.
In summary, this analysis of SPARTAC participants undergoing TI after 48 weeks of ART reveals potential new biomarkers that should be considered in larger studies exploring PTC. The correlations between measures of T-cell-mediated immunity and HIV-1 DNA prior to therapy make a case for the reservoir size being determined by T-cell function early in infection. The relationship between immune responses and the duration of PTC at TI is much less clear, but the association with high levels of T-cell exhaustion might help shed light on why some latently infected cells reactivate and others do not. The work should help in future studies to stratify patients for PTC, with the ultimate aim to determine the mechanism of PTC, and identify new targets for HIV-1 cure research.

Methods
Participants and trial design. SPARTAC 10 was an international RCT enroling adults with PHI within 6 months of a last negative, equivocal or incident HIV-1 test. All participants gave written informed consent. Time of seroconversion was estimated as the midpoint of last negative/equivocal and first positive tests, or date of incident test. Participants were randomized to receive ART for 48 weeks (ART-48), 12 weeks (ART-12) or no therapy (standard of care).
Participants for this sub-study were those infected with subtype B HIV-1 and for whom adequate samples were available. For those in the analysis of viral rebound at TI, we only considered participants in the ART-48 arm who had viral load suppression (o50 HIV-1 copies per ml; Chiron bDNA, Bayer) at the point of stopping ART (Table 1). Peripheral blood mononuclear cells (PBMC) were available for analysis in all participants at baseline, regardless of treatment allocation. Participants randomized to the ART-48 arm were sampled at week 48 immediately prior to TI. Viral remission was defined as maintaining a HIV-1 RNA viral load o400 copies per ml after stopping ART.
Ethics statement. The SPARTAC trial, and the immunological and virological analyses performed in this study were approved by the following authorities: the Medicines and Healthcare products Regulatory Agency (UK), the Ministry of Health (Brazil), the Irish Medicines Board (Ireland), the Medicines Control Council (South Africa) and the Uganda National Council for Science and Technology (Uganda). It was also approved by the following ethics committees in the participating countries: Measurement of HIV-1 DNA. CD4 T cells were enriched from frozen PBMC samples by negative selection (Dynabeads) to a purity of 497%. CD4 T-cell DNA was extracted (Qiagen) and used as input DNA for PCR. Cell copy number and Total HIV-1 DNA levels were quantified in triplicate using previously published assays 29,30 . Integrated HIV-1 was measured using an assay based on that previously published 31 but with some minor modifications as described elsewhere 11 .
Unspliced cell-associated HIV-1 RNA Transcript quantitation. RNA was isolated using the Qiagen AllPrep DNA/RNA Mini kit as per the manufacturer's recommendations with the addition of an on-column double DNase digestion. The RNA assay was a modified version of those previously described [32][33][34] . Briefly, complementary DNA was subjected to two rounds of PCR using semi-nested primers. The first reaction was a 15-cycle end-point reaction using MH535 (5 0 -AACTAGGGAACCCACTGCTTAAG-3 0 ) and SL20 (5 0 -TCTCCTTCTAGC CTCCGCTAGTC-3 0 ) primers at a concentration of 400 nM each 35 as previously described [32][33][34] . The second reaction was a 40-cycle quantitative reaction (forward primer 5 0 -TAAAGCTTGCCTTGAGTGCT, SL20 and the probe FAM-5 0 -AGTRGTGTGTGCCCGTCTGTTG-3 0 -BHQ-1), performed using a Roche Lightcycler 480. The HIV standard was generated by in vitro transcribing the plasmid Sp5-NL4.3 (generously provided by the lab of D. Purcell) using the RiboMax Large Scale SP6 RNA production System (Promega). HIV RNA standards were diluted in 10 ng ml À 1 uninfected PBMC RNA. HIV measurements were normalized to input cellular RNA using the 18S gene with the Amplifluor Human/ Mouse 18S rRNA Primer Set (FAM labelled; Millipore, Cat No. SCR593). All samples were negative when run without reverse transcriptase as a control.
CD8 and CD4 T-cell ELISpot assays. Quantification of the HIV-1-specific CD8 response was performed by IFN-g ELISpot analysis 20 . Responses were determined to subtype B consensus HIV-1 Gag overlapping peptides covering the entire Gag protein (123 15mers overlapping by 11 amino acids; AIDS Research and Reference Reagent Program, NIH). CD4 T-cell IFN-g ELISpot assays were carried out in real time in UK subtype B participants on fresh blood samples. Methods are described elsewhere, but essentially, CD4 T-cell frequencies against HIV-1 subtype B p24 were determined using IFN-g ELISpot using fresh CD8 T-cell-depleted PBMCs. CD4 samples were analysed in duplicate and results were expressed as spot forming units per 10 6 CD8-depleted peripheral blood leucocytes. ELISpot data analysis is described elsewhere 20 .
Statistical analysis. Tests between grouped variables and transformed log 10 DNA levels were performed using t-tests and linear regression. Differences between HIV-1 DNA levels for individual HLA Class I alleles and other HLA alleles were examined using Mann-Whitney tests. Comparisons of HIV-1 DNA levels between patients carrying 'protective' and 'progressive' HLA alleles used t-tests. A correlogram 15 was used to display the Pearson correlations between the biomarkers. Variables with similar correlations are grouped together. The direction and size of the correlations are illustrated by heat colour and pie charts. The order of the biomarkers within the correlogram is determined by hierarchical clustering: there are 2 n-1 (2 17-1 ) possible orderings of the heat map, we use the optimal leaf order algorithm 16 to minimize the difference between adjacent biomarkers. The R package correlogram was used to produce the image. Linear regression was carried out between baseline HIV-1 Total DNA and variables selected from the correlogram. The optimized multivariable model in Table 1B is constructed stepwise in both directions to minimize AIC using the R step function. Time to viral rebound was examined with Kaplan-Meier curves, and Cox proportional hazards models. The proportional hazard assumption was tested by examining Schoenfeld residuals using the cox.zph function from the R survival package. Time to VL rebound was defined as the time from TI (at week 48) to the first of more than one consecutive detectable VL4400 copies per ml. For T-cell exhaustion markers, time to rebound was also analysed for 450 copies per ml. All statistics were calculated using R version 3.1.0. Plots were drawn with R and Prism.