Characteristics of study cohorts. In this study, we employed two ATI cohorts: 1) The Philadelphia cohort: a cohort of 24 HIV-infected individuals on suppressive ART who underwent an open-ended ATI.26, 27 This cohort had a wide distribution of viral rebound times (14 to 119 days; median = 28; Supplementary Table 1).26 Importantly, this cohort underwent ATI without concurrent immunomodulatory agents that might confound our signatures at the initial discovery phase.26, 27 2) The AIDS Clinical Trial Group (ACTG) cohort: a cohort combining 74 participants from six ACTG ATI studies (ACTG 371,28 A5024,29 A5068,30 A5170,31 A5187,32 and A519733), tested or not the efficacy of different HIV vaccines and interleukin-2 (IL-2) treatment. These six ATI studies included 567 participants and identified 27 PTCs out of these participants. Our ACTG cohort included all 27 PTCs and 47 matched NCs from the same studies. The definition of post-treatment control was: remaining off ART for ≥ 24 weeks post-ATI with VL ≤ 400 copies for at least 2/3 of time points; had no ART in the plasma; and had no evidence of spontaneous control pre-ART. The remaining 47 were non-controllers (NCs) who rebounded before meeting PTC criteria,34, 35, 36 The PTC and NC groups within the ACTG cohort are matched for gender, age, ethnicity, % treated during early infection, ART duration and pre-ATI CD4 count (Table 1 and Supplementary Fig. 1). Notably, the combined studies within the ACTG cohort reflect six ATI clinical trials where individuals received or not different HIV vaccines and/or immunotherapies.28, 29, 30, 31, 32, 33 This important feature of this cohort allows for identifying/validating markers that predict duration and probability of viral remission independent of potential interventions.
Elevated pre-ATI levels of plasma markers of glutamate and bile acid metabolism associate with delayed viral rebound in the Philadelphia Cohort. We first aimed to examine the utility of plasma metabolites as biomarkers of time-to-HIV-rebound after ART-cessation. Towards this goal, we measured levels of plasma metabolites from the Philadelphia cohort.26, 27 Using an untargeted mass spectrometry (MS)-based metabolomics analysis, we identified a total of 179 metabolites in plasma samples collected immediately before the ATI. Then, we applied the Cox proportional-hazards model to identify metabolomic signatures of time-to-viral-rebound. As shown in Fig. 1A, higher pre-ATI levels of 13 plasma metabolites were significantly associated with a longer time-to-viral-rebound with P < 0.05 and false discovery rate (FDR) < 20%. In contrast, higher pre-ATI levels of 12 plasma metabolites were significantly associated with a shorter time-to-viral-rebound. When participants were separated into low or high groups by the median of each of these 25 metabolic markers, pre-ATI levels of 20 of 25 metabolites significantly indicated hazards of viral-rebound over time using the Mantel-Cox test (Fig. 1B and Supplementary Table 2).
We next sought to determine if the 25 metabolites associated with time-to-viral-rebound shared similar metabolic pathways. Multi-analysis combining KEGG and the STRING Interaction Network (focusing on metabolite-associated enzymatic interactions) revealed that most of the 13 metabolites whose pre-ATI levels associated with a longer time-to-viral-rebound belong to two major metabolic pathways. Specifically, five metabolites lay within the anti-inflammatory glutamate/tricarboxylic acid (TCA) cycle pathway, and three were intermediates within the primary bile acid biosynthesis pathway (Fig. 1C). Confirmatory analysis on these 13 metabolites using the MetaboAnalyst 3.0 pathway feature (http://www.metaboanalyst.ca/) showed enrichment in glutamate metabolism (P = 0.00068) and the bile acid biosynthesis pathway (P = 0.0399) (Fig. 1C and Supplementary Table 3).
Elevated pre-ATI levels of plasma markers of pyruvate and tryptophan metabolism associate with accelerated viral rebound in the Philadelphia Cohort. Multi-analysis of the 12 metabolites whose pre-ATI levels associated with shorter time-to-viral-rebound showed four intermediates in the tryptophan metabolism pathway and three that are central players in the pro-inflammatory pyruvate pathway (Fig. 1D). These observations were confirmed for the 12 metabolites using MetaboAnalyst 3.0, which demonstrated enrichment in pyruvate metabolism (P = 0.0065) (Fig. 1D and Supplementary Table 3). The roles of key discovered metabolites within the glutamate, bile acids, tryptophan, and pyruvate pathways are graphically illustrated in Supplementary Fig. 2. These data reveal a previously undiscovered class of plasma metabolic biomarkers that are associated with time-to-viral rebound post-ATI. They further demonstrate that these biomarkers belong to a specific set of metabolic pathways that may play a previously unrecognized role in HIV control.
L-glutamic acid and pyruvate modulate latent HIV reactivation and/or macrophage inflammation in vitro. Among the top candidate metabolic biomarkers from Fig. 1 are L-glutamic acid (glutamate metabolism) and pyruvic acid (pyruvate metabolism). The higher pre-ATI levels of L-glutamic acid and pyruvic acid associated with longer or shorter time-to-viral-rebound, respectively. These two metabolites can impact inflammation in opposing directions. Glutamate controls the anti-inflammatory TCA cycle through its conversion by glutamate dehydrogenase to α-ketoglutarate,37, 38 whereas pyruvate is centrally positioned within the pro-inflammatory glycolytic pathway.39, 40, 41 We therefore sought to determine if these two metabolites exhibited a direct functional impact on latent HIV transcription and/or myeloid inflammation. We first assessed the impact of these two metabolites on latent HIV reactivation using the established “J-Lat” model of HIV latency. J-Lat cells harbor a latent, transcriptionally competent HIV provirus that encodes green fluorescent protein as an indicator of reactivation (Fig. 2A).42, 43 There are several clones of the J-Lat model with different characteristics, including the type of stimulation to which they respond. For example, the 5A8 is the only J-Lat clone responsive to αCD3/αCD28 stimulation. We examined the impact of L-glutamic acid and pyruvate on two J-Lat clones (5A8 and 10.6). Whereas pyruvate had no observable effect on latent reactivation for either clone (data not shown), L-glutamic acid significantly inhibited the ability of phorbol-12-myristate-13-acetate (PMA)/ionomycin or αCD3/αCD28 to reactivate latent HIV in clone 5A8 without impacting viability compared to stimuli alone controls (Fig. 2B). L-glutamic acid also inhibited the ability of PMA/ionomycin or TNFα to reactivate latent HIV in clone 10.6 without impacting viability compared to stimuli alone controls (Fig. 2C). These data demonstrate that a plasma metabolite, L-glutamic acid, can inhibit latent viral reactivation, consistent with the observation that pre-ATI levels of L-glutamic acid predicted a longer time-to-viral-rebound.
Beyond direct impact on latent viral reactivation, plasma metabolites may exert effects on myeloid inflammation, and such effects may underlie HIV control during ATI. This possibility was tested by examining the effects of L-glutamic acid and pyruvate on lipopolysaccharides (LPS)-mediated secretion of pro-inflammatory cytokines from THP-1 derived macrophage-like cells. These cells characterized by high basal glycolytic activity closely reflect the Warburg-like phenotype observed in HIV infected individuals,44 and exhibit similar inflammatory responses to primary cells under similar in vitro conditions.39 Cells were treated with L-glutamic acid, pyruvate, or appropriate controls for 2 hours before stimulating with LPS and IFNγ for 5 hours (Fig. 2D). L-glutamic acid inhibited LPS/IFNγ-mediated production of pro-inflammatory cytokines such as IL-6 and TNFα (Fig. 2E; other cytokines are shown in Supplementary Fig. 3A). Consistently, L-glutamic acid also increased anti-inflammatory IL-10 secretion (Fig. 2E). Conversely, pyruvate increased IL-6 and TNFα secretion (Fig. 2F; other significantly regulated cytokines are shown in Supplementary Fig. 3B). These data demonstrate that not only do some metabolites associate with time-to-viral-rebound, but also that there is a plausible, functionally significant link between these biomarkers and viral control during and following ATI.
Pre-ATI plasma glycomic and metabolic biomarkers associate with time-to-viral-rebound in the ACTG Cohort. Our recent pilot study showed that pre-ATI levels of a specific set of glycans predicted a longer time-to-viral rebound after ART discontinuation.26 However, this small pilot study did not correct for confounders such as age, gender, and nadir CD4 count on viral rebound. We hypothesized that a set of plasma glycans and metabolites we identified in that pilot study,26 as well as in the results shown in Fig. 1, can predict time-to-viral-rebound and/or probability-of-viral-rebound using plasma samples from a larger validation cohort, even after adjusting for potential demographic and clinical confounders. For this analysis we analyzed samples from the ACTG cohort.
We analyzed the plasma metabolome of samples collected from this cohort before ATI. A total of 226 metabolites were identified using MS-based metabolomics analysis. In addition, we applied two different glycomic technologies to analyze the plasma glycome of the same samples. First, we used capillary electrophoresis to identify the N-linked glycans of total plasma glycoproteins (identified 24 glycan structures, their names and structures are listed in Supplementary Fig. 4) and isolated plasma IgG (identified 22 glycan structures, their names and structures are listed in Supplementary Fig. 5). Second, we used a 45-plex lectin microarray to identify total (N and O linked) glycans on plasma glycoproteins. The lectin microarray enables sensitive identification of multiple glycan structures by employing a panel of 45 immobilized lectins (glycan-binding proteins) with known glycan-binding specificity, resulting in a "glycan signature" for each sample (the 45 lectins and their glycan-binding specificities are listed in Supplementary Table 4).45
We used the Cox proportional-hazards model and a set of highly stringent criteria to identify sets of glycans or metabolites whose pre-ATI levels associated with either time to VL ≥ 1000 (Fig. 3 top panel) or time to two consecutive VL ≥ 1000 (Fig. 3 bottom panel). To ensure high stringency, we only considered markers with a hazard ratio (HR) ≥ 2 or ≤ 0.5. We also included in these sets only those glycomic and metabolic markers with either FDR < 0.1 or markers that emerged from the Philadelphia cohort (Fig. 1 and our previous pilot study26). Importantly, we only included markers that remained significant (P < 0.05) after adjusting for age, gender, ethnicity, ART initiation (during early or chronic HIV infection), ART duration, or pre-ATI CD4 count (Supplementary Table 5). These combined strict criteria identified a signature that predicted shorter time-to-rebound to VL ≥ 1000, comprising four glycan structures and one metabolite (Fig. 3 top panel, red). These five markers include the highly sialylated plasma N-glycan structure (A3G3S3), GalNAc-containing glycans (also known as T-antigen; measured by binding to both MPA and ACA lectins) and the metabolite pyruvic acid. We also identified a signature that associated with a longer time-to-rebound to VL ≥ 1000, comprising seven glycan structures and one metabolite, notably the digalactosylated G2 glycan structure on plasma bulk IgG, fucosylated glycans in plasma (binding to AAL lectin), GlcNac glycans in plasma (binding to DSA, UDA, and STL lectins), and the metabolite L-glutamic acid (Fig. 3 top panel, blue).
Turning to markers that associated with time to two consecutive VL ≥ 1000, and applying the same strict criteria, we identified five glycomic markers and one metabolite whose pre-ATI levels associate with shorter time-to-rebound post-ATI, including A3G3S3 in plasma, T/Tn-antigens (binding to MPA, ACA, and ABA lectins), and the metabolite Nicotinamide (Fig. 3 bottom panel, red). We also identified seven glycan structures and two metabolites whose pre-ATI levels predicted a longer time-to-rebound, including, G2 glycan structure on bulk IgG, core fucosylated glycans (binding to LCA lectin) in plasma, total fucosylated glycans (binding to AAL lectin) in plasma, GlcNac glycans (binding to DSA, UDA, and STL lectins) in plasma, and the metabolites oxoglutaric acid (α-ketoglutaric acid) and L-glutamic acid (Fig. 3 bottom panel, blue). The significance of several of these markers was also confirmed using the Mantel-Cox test in an independent analysis (Fig. 4). In sum, using stringent analysis criteria that also took into account potential confounders, we identified and validated plasma glycomic/metabolomic signatures of time-to-viral-rebound after ART discontinuation in this independent heterogeneous cohort of individuals who underwent ATI and received or not several different interventions before ATI.
Levels of pre-ATI plasma glycomic and metabolic markers that associate with time-to-viral-rebound are linked to levels of cell-associated HIV DNA and RNA. We next examined whether the plasma glycans and metabolites (Fig. 3) that associated with time-to-viral-rebound also reflected levels of virological markers of HIV persistence (levels of peripheral blood mononuclear cell (PBMC)-associated total HIV DNA and HIV RNA) in blood. We found that pre-ATI levels of total fucose (binding to AAL lectin), which predicted delayed viral rebound, showed a significant inverse correlation with pre-ATI levels of cell-associated HIV DNA and RNA (Fig. 5A-B). Similarly, pre-ATI levels of core fucose (binding to LCA lectin), which also predicted delayed viral rebound, also showed an inverse correlation with pre-ATI levels of cell-associated HIV DNA and RNA (Fig. 5C-D). Furthermore, total levels of (GlcNAc)n (binding to UDA lectin), which predicted delayed viral rebound, had an inverse correlation with levels of total HIV DNA (Fig. 5E). Noteworthy, levels of pyruvic acid, whose pre-ATI levels predicted accelerated viral rebound, had a significant positive correlation with pre-ATI levels of cell-associated HIV DNA (Fig. 5F). These data provide more support for a plausible mechanistic connection between our discovered plasma markers and HIV control during ATI.
Multivariable Cox model, using Lasso technique with the cross-validation (CV), selected variables that their combination predicts time-to-viral-rebound. As a single marker would be highly unlikely to strongly predict these complex virological milestones, we next sought to apply a machine learning algorithm to identify a smaller set of plasma biomarkers (from Fig. 3) that together can predict either time to VL ≥ 1000 or time to two consecutive VL ≥ 1000 better than any of these biomarkers individually. The analysis considered biomarkers, both metabolites and/or glycan structures, that emerged as significant from the ACTG cohort (Fig. 3) and using samples with complete data set (n = 70; four samples did not have a complete dataset). The machine learning algorithm, Lasso (least absolute shrinkage and selection operator) regularization, selected seven markers from among the 13 that associated with time to VL ≥ 1000 (Fig. 3 top panel), whose predictive values are independent and combining them together would enhance the predictive ability of the signature compared to each of these markers alone (Supplementary Table 6). Indeed, a multivariable Cox regression model using these seven variables showed a concordance index (C-index) value of 74% (95% confidence interval: 68%-80%), which is significantly higher than the C-index values obtained from Cox models using each variable individually (P < 0.05; Supplementary Table 6). Notably these seven markers included four whose pre-ATI levels associated with accelerated rebound, A3G3S3, T-antigen (MPA and ACA lectins binding), and the metabolite pyruvic acid. The other three markers associated with delayed rebound: total fucose (AAL lectin binding), (GlcNAc)n (STL lectin binding), and the metabolite L-glutamic acid (Supplementary Table 6).
Examining markers associated with time to two consecutive VL ≥ 1000, Lasso selected 12 markers from the 15 identified (Fig. 3 bottom panel) whose predictive values are independent and whose combination enhanced the predictive ability of the signature compared to any single marker alone (Supplementary Table 7). A multivariable Cox regression model using these 12 variables showed a concordance index (C-index) value of 76.4% (95% confidence interval: 70%-84.2%), which is significantly higher than the C-index values obtained from Cox models using each variable individually (P < 0.05; Supplementary Table 7). The 12 markers included some whose pre-ATI levels associated with accelerated rebound, including A3G3S3 glycans and T-antigen (ABA and ACA lectins binding) and some whose pre-ATI levels associated with delayed viral rebound, including G2 glycans, total fucose (AAL lectin binding), (GlcNAc)n (STL lectin binding), and the metabolite L-glutamic acid. (Supplementary Table 7). Together, these data suggest that these multivariable models of combined plasma glycans and metabolites markers warrant further exploration for their capacity to predict time-to-viral rebound in different settings.
Pre-ATI plasma glycomic and metabolic markers distinguish post-treatment controllers (PTC) from non-controllers (NCs). Examining the glycan structures and metabolites obtained from the ACTG cohort, we identified eight glycan structures whose pre-ATI levels were significantly different between PTCs and NCs with FDR < 0.1 (Fig. 6A-H). Among these eight glycans structures, three exhibited lower levels in the plasma of PTCs compared to NCs (FDR < 0.02), including the di-sialylated glycans, A2, in total IgG glycans; the highly-sialylated glycans, A3G3S3, in plasma N-glycans; and T-antigen (binding to ABA lectin) (Fig. 6A-C); and five glycans were higher in PTCs compared to NCs (FDR ≤ 0.035). These included total fucose (binding to AAL lectin), core fucose (binding to LCA and PSA lectins), and (GlcNac)n (binding to STL and UDA lectins (Fig. 6D-H).
Examining metabolites, we found that pre-ATI levels of α-ketoglutaric acid and L-glutamic acid, both of which predicted delayed viral rebound, were higher in the plasma of PTCs compared to NCs (P < 0.01, Fig. 6J-I). Importantly, for both glycans and metabolites, we only selected markers whose levels remained different (P < 0.05) between PTCs and NCs after adjusting for age, gender, ethnicity, ART initiation, ART duration, or pre-ATI CD4 count (Supplementary Table 8). Together, these data suggest that a selective set of plasma glycans and metabolites can distinguish PTCs from NCs and may be used to predict the probability of viral rebound (i.e., the likelihood of PTC phenotype after ATI).
Multivariable logistic model, using CV Lasso technique, selected variables that their combination predicts risk of viral rebound. We next applied the Lasso regularization to select from among the ten markers in Fig. 6 a set of markers whose combined predictive utility is better than the predictive utility of any of these 10 markers individually. The analysis used biomarkers that emerged as significant from the ACTG cohort (Fig. 6) and using samples with complete data set (n = 70). Lasso selected seven markers from the 10 identified as able to distinguish PTCs from NCs (Fig. 6) that their predictive values are independent and combing them together would enhance the predictive ability of the signature compared to each of these markers alone (Supplementary Table 9). Indeed, a multivariable logistic regression model using these seven variables showed an area under the ROC curve (AUC) value of 97.5% (Fig. 7A; 95% confidence interval: 94% -100%), which is significantly higher than the AUC values obtained from logistic models using each variable individually (P < 0.05; Supplementary Table 9). These seven markers included three whose pre-ATI levels are lower in PTCs compared to NCs, namely A2, A3G3S3, and T-antigen (ABA lectin binding), and four whose pre-ATI levels were higher in PTCs compared to NCs, namely total fucose (AAL lectin binding), core fucose (LCA lectin binding), (GlcNAc)n (STL lectin binding), and the metabolite L-glutamic acid (Supplementary Table 9).
Next, a risk score predicting NC was estimated for each individual using the multivariable logistic model. We then examined the ability of these risk scores to classify PTCs and NCs from the ACTG cohort. As shown in Fig. 7B, the model was able to correctly classify 97.7% of NCs (sensitivity), and 85.2% of PTCs (specificity) with overall accuracy of 92.9%. This analysis highlights the potential utility of this risk score estimated from the multivariable model combining six plasma glycans and one metabolite, to predict the risk of NC post-ATI. This prediction can be utilized to select individuals likely to achieve PTC phenotype during HIV cure-focused clinical trials, to be included in ATI studies. In addition, the markers that are included in this model might also serve as windows into the mechanisms that contribute to the PTC phenotype.