We used an unbiased approach that integrates datasets from peripheral blood transcriptome, high density flow cytometry and plasma cytokine measures to identify cellular and molecular drivers of HIV persistence and lack of CD4 T cell recovery (Supplementary Fig. 1a), in two independent cohorts of HIV-infected subjects (see Study Participants in Material and Methods; Table S1, Table S2). All subjects were under cART for at least three years and maintained consistently low CD4 T cell counts up to four years prior to sample collection (Supplementary Fig. 1b). The Cleveland Immune Failure cohort (CLIF) was comprised of 61 subjects (17 immune responders - IRs - >500 CD4 T-cells/mm3 and 44 immune non-responders - INRs - <350 CD4 T-cells/mm3); whereas, the Study of the Consequences of the Protease Inhibitor Era cohort (SCOPE) included 41 subjects (20 IRs with >500 CD4 T-cells/mm3 and 21 INRs with <350 CD4 T-cells/mm3).
Transcriptional profiling reveals systemic senescence as a driver of poor immune reconstitution
Exploratory analyses of whole blood transcriptomic data (CLIF cohort subjects; using unsupervised clustering described in the methods) (Fig. 1a and Supplementary Fig. 1c) identified three groups of subjects with unique transcriptional profiles: IRs and two distinct INR groups, “INR-A” and “INR-B”. INR-As exhibited the highest transcriptomic variation from IRs with approximately 3000 differentially expressed genes (DEGs). In contrast, INR-Bs were proximal to IRs with <400 DEGs (Table S3-5). Age, years on cART, CD4 T cell counts, CD4 T cell nadir, markers of gut barrier dysfunction (sCD14), and inflammation (IL-6, IP10) - known predictors of morbidity and mortality in INRs27 - failed to distinguish the two INR groups (Supplementary Fig. 1d-p). Interestingly, DEGs specific to INR-As were mostly down-regulated (72% and 66% when contrasted against IRs and INR-Bs, respectively), suggesting a quiescent transcriptional state in these subjects (Supplementary Fig. 2a). Comprehensive pathway analyses using MSigDb’s Hallmark module28-30 (Fig. 1b, Table S6) revealed that a majority of the pathways, likely driven by significantly elevated CD4 T cell numbers, were upregulated in IRs when compared to both INR-As and INR-Bs. Notably, when contrasting INR-As against INR-Bs (where CD4 T cell numbers are comparable), INR-As were observed to have significantly decreased expression of inflammation, cell cycling, apoptosis, metabolism genesets31,32. This transcriptional profile was in line with reduced expression of genes regulated by Myc (a primordial transcription factor (TF) that regulates proliferation and metabolism of activated T cells)33 and was indicative of a senescent state31,32.
Cell subset deconvolution (using gene signatures from Nakaya et al34) showed that genesets specific to all major innate and adaptive immune cell subsets were reduced in INR-As (vs INR-Bs) (Fig. 1c, Table S7). Similarly, down-regulation of gene signatures for specific CD4 and CD8 T-cell subsets (naive, memory and effector cell, extracted from Novershtern et al35 - see methods for detail) were also observed (Fig. 1d, Table S8). The down-regulated signatures in T-cells, myeloid dendritic cells (mDCs) and monocytes mapped to genes that drive apoptosis, transcription/cell migration, and cellular/lipid metabolism (lipid storage, GTP metabolic process), respectively (Supplementary Fig. 2b). These data provide evidence for decreased global transcription and systemic senescence in immune cells from INR-As, when compared to INR-Bs or IRs.
Transcriptional profiles downstream of FOXO3 and SMAD2/3 signaling are enriched in senescent INRs
To identify mechanisms underlying poor immune reconstitution observed in INR-As, we identified TFs (p < 0.05, Supplementary Fig. 2c, Table S9) and mapped the pathways that were associated with genes upregulated (Reactome pathway database, ClueGO plugin in Cytoscape, FDR <0.05)36,37 in the INR-As (vs IRs and INR-Bs; FDR < 0.05, Table S10). The upregulated TFs included IRF3 (driver of type I IFN production38), FOXO3 (transcriptional repressor39), SMAD2 (TGF-β signaling40) and CCNT2 (Negative regulator of HIV Tat protein; i.e. driver of HIV latency induction41) (Supplementary Fig. 2c). Interestingly, enrichment of cellular processes downstream of these TFs - including heme metabolism, TGF-β signaling, IRF3 activation, reactive oxygen species (ROS) production and inhibition of NF-kB activation (Fig. 1e) - were observed in these subjects. Specifically, INR-As showed increased expression of features of senescence that included FOXO3 regulated genes like SOD/CAT42-44 (driver of ROS production) (Fig. 1e); FOXO4/TP5345 driven anti-apoptotic genes (BCL2L1; FDR < 0.01)46-48 (Table S5); and targets of SMAD2/3 including ‘Inducer of Promyelocytic Leukemia’ (PML; conductor of TGF-β signaling via SMAD2/3, and regulator of HIV latency49,50), WEE151 (cell cycle regulator) and GLUL52,53 (metabolic regulator). The downregulation MYC target genes (known to control ribosomal biosynthesis and translation)54, genes of the electron transport chain, and genes regulating major metabolic pathways (including glycolysis, oxidative phosphorylation and fatty acid metabolism) further confirmed the senescent nature of these subjects (Supplementary Fig. 2d, Fig. 1b). The senescent/anti-inflammatory profile of INR-As contrasted the pro-inflammatory profile of INR-Bs where the enrichment of pathways downstream of TLR/IL1, cell cycling, and apoptosis/pyroptosis was observed. Increased expression of several members of the pro-inflammatory NF-kB family of TFs (NFKB1, NFKB2 and RELB), upregulation of NF-kB target genes including chemokines (e.g. CCL2 and CCL17)55, genes of the inflammasome complex (NLRP3, IL1B and IL18)56 and effector genes driving apoptosis/pyroptosis (e.g. CASP4, DIABLO, CASP2 and CASP3)57 were characteristics of this group (Supplementary Fig. 2d).
Overall, our data suggest that the INR-As have increased expression of a signaling cascade downstream of TGF-b (SMAD2/3; PML) which culminates in the upregulation of FOXO3/4 driven senescence - characterized by impaired cell metabolism and cell cycle arrest58. INR-As will henceforth be referred to as “Senescent-INRs”.
A gene-based classifier segregates the two INR groups in whole blood and in sorted central memory CD4 T cells
Next, we built a gene-based nearest shrunken centroid classifier (see methods) to identify Senescent-INRs in other cohorts of HIV infected cART treated subjects. The 352 features (genes, Table S11) classifier was trained on the CLIF cohort and had a misclassification error rate of 0.26 (Fig. 1f). To validate the capacity of the classifier to differentiate immune cell subsets from Senescent INRs (vs INR-Bs), we tested the enrichment of the classifier geneset in sorted CD4 TCM, CD4 TEM (effector memory T cells) and innate immune (HLA-DR+CD19-) cells. We observed that unlike the CD4 TEM and innate subsets, the classifier geneset was enriched in the CD4 TCMs from the senescent INRs (Fig. 1g). In line with the whole blood signatures (Fig. 1b), metabolic, cell cycling and apoptosis pathways were significantly reduced in these CD4 TCMs. Whereas, pathways that define quiescence (WNT/b-catenin signaling pathway) and senescence (cellular/oxidative stress induced) were enriched in CD4 TCMs from senescent INRs (Fig. 1h). Overall, these observations suggest that CD4 TCMs from the senescent INRs are metabolically impaired, have poor cell cycling capacity and could serve as hubs that maintain high levels of HIV reservoir.
Classifier confirms the generalizability of the Senescent-INR phenotype in HIV-infected subjects
Using an unsupervised approach (initially applied in the CLIF cohort), we confirmed two distinct INR groups in the SCOPE cohort (Supplementary Fig. 3a, Supplementary Fig. 3b). The 352-gene classifier segregated these two INR groups in the SCOPE cohort with an accuracy of 81% (Fig. 2a) - validating the generalizability and reproducibility of two INR groups (Fig. 2b: PCA representation of the 352-gene classifier across the SCOPE cohort) and highlighting the potential use of this classifier to distinguish Senescent-INRs in clinical settings. Our classifier (Table S11) included genes that drive a senescent biology (FOXO3, FOXO4, TGFBR2, RIOK3, IRF3 and BCL2L1)46,59,60,61 and regulators of mitochondrial activity (NDUFS3, CYB5R3, ATP5G2)62. Several transporters of macromolecules: SLC6A8 (Sodium- and chloride-dependent creatine transporter 163), SLC48A1 (Heme transporter64), SLC4A1 (Band 3 anion transporter65), SLC25A23 (Mitochondrial Calcium carrier66) and SLC38A5 (glucose67) upregulated in Senescent-INRs and were also features of the classifier (Table S11). Overall, we observed that molecular pathways specific to cellular senescence68-71 discriminate the two INR groups and counter-intuitively suggest that failure to increase CD4 T-cell numbers in senescent could be due to senescence and not to previously described pro-inflammatory cascades10.
The classifier gene set highlights the role of senescence in driving the magnitude of the inducible HIV reservoir
Although poor immune reconstitution has been associated with HIV persistence72, the cellular effectors and molecular mechanisms driving this association remain unidentified. To assess the role of senescence in HIV persistence, we measured frequencies of CD4 T-cells with inducible multi-spliced HIV RNA (“inducible HIV”) - in all subjects from the SCOPE cohort - using the Tat/rev Induced Limiting Dilution Assay (TILDA)73. Our data show that Senescent-INRs have significantly higher inducible HIV when compared to IRs (~4.3-fold increase in median values, p<0.021) and INR-Bs (~5.5-fold increase in median values, p = 0.046) (Fig. 2c). And, these Senescent-INR subjects with the highest levels of inducible HIV were significantly enrichmed in (i.e. could be predicted by) the classifier geneset (Fig. 2d). Interestingly, inducible HIV levels in Senescent-INRs were negatively correlated with CD4 counts (rho = -0.52; p < 0.05; while this correlation was not significant in inflammatory INRs and in IRs). Comprehensive pathway analyses of the whole transcriptome further confirmed that genesets characteristic of Senescent-INRs (i.e. cell cycle arrest and ROS production) were associated with higher inducible HIV (Table S12).
Increased frequencies of TGF-b producing Tregs and PD1+ TCMs with impaired metabolism drive lower CD4 counts and HIV persistence
Given that transcriptional profiling identified the TGF-β pathway (PML, SMAD2/3; Fig. 1e, Supplementary Fig. 2 c,d) as a driver of cellular senescence in our cohort and that Forkhead box P3 (FOXP3) expressing T-regulatory cells (Tregs) are a primary source of TGF-β74,75, we sought to ascertain the role of Tregs in driving senescence and HIV persistence. To this end, we developed a high-dimensional flow cytometry panel (Table S13) to quantify master regulators of Treg function (SATB1 and FOXP376), to discriminate differentiated Tregs (CD45RA, CD49B and CD3977), to assess proliferation (Ki67) and TGF-β activation potential of a CD4 T cell (expression of latent (LAP) or activated (GARP) forms of TGF-β78,79). UMAP dimension reduction (https://arxiv.org/abs/1802.03426) (Fig. 3a). followed by unbiased clustering80 (Fig. 3a) led to the identification of three clusters of CD4 T cells (Fig. 3b, Fig. 3c, Table S14) that were enriched in senescent INRs (vs IRs) and expressed markers characteristic of TGF-b producing Tregs (FOXP3, CD25, lowCD127, LAP and GARP) (Fig. 3c). The most abundant of these clusters (i.e. Cluster 7; 0.81-5.34% of CD4 T cells) also showed an effector Treg phenotype (high CD39, low CD45RA) and low levels of markers that define IL-10 production capacity (Tr1) (i.e. CD49B, LAG381) (Fig. 3c).
We assessed the systemic impact of these Treg clusters by studying their association with clusters of cytokines that define the host plasma milieu. We applied an unsupervised clustering using independent methods (k-means, and hierarchical clustering82) to identify four clusters of plasma cytokines (from 43 cytokines testes) across all subjects of the SCOPE cohort (Supplementary Fig. 4a,b,c). Of these, the overall centroid score of “Cytokine Cluster 3” was significantly associated with the 352-gene classifier in Senescent-INRs (NES = 2.9, FDR = 0, Supplementary Fig. 4d, Table S15). No significant correlation between “Cytokine Cluster 3” centroid score and the classifier genes was observed in INR-Bs or IRs (Table S15). This cytokine cluster included several anti-inflammatory cytokines: TGF-β1, TGF-β2 (triggers of the TGF-β pathway83), IL13 (known to inhibit inflammatory cytokine production84), KC/GRO (CXCL1) (another anti-inflammatory cytokine85), VEGF (the inhibitor of apoptosis86) and homeostatic cytokines (IL3, IL7)87,88 (Supplementary Fig. 4c). Importantly, our data showed that frequencies of GARP+ Tregs (Treg cluster 7) were univariately correlated with members of “Cytokine Cluster 3” (Fig. 3d, including TGF-b2 and IL-7; Table S16), and were associated with an enrichment in the classifier genes (significantly overlapped leading edge genes represented in Fig. 3e, Table S17).
The downstream impact of heightened frequencies of TGF-b secreting Tregs on HIV latency was further confirmed by the observation that ChIP-Seq validated SMAD2/3 targets89 and HDAC1/2 targets (induced after siRNA knockdown)90 were more abundant in subjects with the highest inducible HIV and frequencies of GARP+ Tregs (Fig. 3e). These genesets included genes that activate latent TGF-β (i.e. FURIN91), restrict T cell differentiation (LGAL3; inducer of TGF-β driven activation of b-catenin92,93), reduce cell cycling (i.e. GADD45A94) and restrict chromatin accessibility (co-operation between HDAC1/2 targets and SMAD2/3) can to induce cell quiescence (Fig. 3e).
Increased frequencies of senescent PD1+ TCMs with impaired mitochondrial function drive lower CD4 counts in the senescent INRs
Given that signaling via TGF-β/FOXO3 axis drives surface PD1 expression95, cell cycle arrest and impaired metabolism96 - we used high-dimensional flow cytometry, transcriptomics and cytokine data to identify the cellular drivers of immune reconstitution or lack thereof; we hypothesized that the upregulation of surface PD-1 and reactive oxygen species (ROS; measured by intracellular CellROX staining) in CD4 T-cells could be associated with lower CD4 counts97. Using the analytical strategy described in the section above, we identified a cluster of PD1hi ROShi CD4 central memory T-cells (TCM, Cluster 9; Fig. 4a,b,c, Table S18) that was induced in Senescent-INRs when compared with IRs (but not in INR-Bs vs IRs). Conversely, a cluster of PD1hi ROSlo CD4 effector memory T-cells (TEM, Cluster 17; Fig. 4a,b,c, Table S18) was uniquely upregulated INR-Bs. Frequencies of PD1+ CD4/CD8 TCM cells were directly correlated with the frequencies of GARP+ Tregs (Treg cluster 7) and with cytokine cluster 3 (Fig. 4d). In line with previous studies24,98, these data confirm the negative impact of PD-1 expression on T-cell homeostasis and immune reconstitution.
Evidence of impaired metabolism and cellular senescence in these PD1hi ROShi TCMs was obtained by identifying metabolic and cell cycling genesets that associate with this cell subset (in senescent INRs vs IRs). Individuals with higher levels of PD1hiROShi TCMs had poor mitochondrial metabolism profiles (higher ROS and lower oxidative phosphorylation), reduced cMyc activity and increased expression of genes that drive cellular senescence (Fig. 4f, Table S19). Specifically, expression of catalase (CAT), peroxidins (PRDXs) and cell cycle inhibitory genes (i.e. CDKN2D) was increased in senescent-INRs; while genes that regulate oxidative phosphorylation (NDUFs and COXs) and MYC target genes were expressed at lower levels. Concurrently, frequencies of PD1hiROShi TCMs were also associated with cellular senescence pathways downstream of impaired mitochondrial activity (i.e: oxidative stress induced senescence, SASP), and were observed to be the highest in subjects with higher inducible HIV (Fig. 4e, Table S19). Altogether, these results indicate that PD1 expressing TCMs with impaired mitochondrial metabolism are senescent cells lacking the capacity to cycle and to differentiate into effector cells, thereby driving low CD4 cellular counts.
Bile and short chain fatty acid profiles drives Treg differentiation and TGF-β production in Senescent-INRs
The role of Tregs and impaired T cell homeostasis has been well characterized in prominent metabolic diseases (from diabetes, cancer and cardio vascular diseases)99-103. Subjects presenting such metabolic aberrations have impaired host/microbe driven metabolic profiles that impair T cell function. To understand the metabolic milieu that could regulate Treg function in senescent INRs, we performed mass spectrometry (MS) to identified plasma metabolites that would be associated (ex vivo) with or could trigger (in vitro) Treg differentiation, TGF-β production or senescence. An unbiased assessment of the plasma metabolite profile of SCOPE cohort participants revealed that one of the primary components of variation (Principal Component 1 and Principal Component 2; derived from PCA analyses of ~750 detectable metabolites) was associated with inducible HIV (PC2 vs TILDA p-value <0.05; Fig. 5a). Specifically, the metabolite that were univariately correlated (p-value<0.05) with inducible HIV levels included members of the liver-biliary axis (i.e. primary liver/bile metabolites like bilirubin, biliverdin, cholate and glycol-beta-muricholate; secondary liver/bile metabolites like ursodeoxycholate), carnitine derivatives and members of the hydroxybutyrate family (i.e. a-ketobutyrate and hydroxybutyryl carnitine) (Fig. 5b). Several, but not all, of these metabolites were also associated with SMAD2/3 and HDAC1/2 target genesets and frequencies of GARP+ Tregs (Fig. 5b). Importantly, this analysis showed that a-ketobutyrate - a correlate of inducible HIV - was also correlated with C7 GARP+ Treg frequencies (p = 0.089), emphasizing the association between Tregs, metabolome and inducible HIV. Together these data indicate that metabolites could constitute integral components of the mechanisms that fuel HIV persistence.
TGF-β production resulting from alpha-ketobutyrate driven Treg differentiation causes increased HIV latency in vitro
To establish a causal link between butyrate metabolites on GARP+ Treg differentiation, we used the approach described by Ohno and Rudensky17,104. Briefly, increasing concentrations of a-ketobutyrate were added to healthy sorted naïve CD4 T-cells stimulated with IL-2, anti-CD3/28 and/or TGF-β (Fig. 5 c,d) changes in frequencies of TGF-b secreting FOXP3+ cells were monitored105. We observed that, stimulation of naïve T cells with high concentrations of a-ketobutyrate (in the absence of TGF-β) preferentially led to the differentiation of naïve T-cells into GARP+FOXP3+ cells (Fig. 5 c,d) that secreted significant amounts of TGF-β1 (Fig. 5e). Of note, a further increase in GARP+ Treg differentiation when both TGF-β and alpha-ketobutyrate were added to the culture (Supplementary Fig. 5a). Aside from Tregs, Increasing concentrations of a-ketobutyrate also led to significantly increase in PD1 expressing CD4 T-cells (Fig. 5 c,d, Supplementary Fig. 5b) that were associated with loss of effector function - shown by reduced secretion of T-helper cytokines in the culture media (i.e. IL17A, IFN-γ, IL9) (Fig. 5f). Altogether, these data indicate that in addition to enhancing Treg differentiation, the abundance of b-hydroxybutyrates could drive the upregulation of TGF-β associated suppressive activity of Tregs. The latter, as shown in Fig. 3, are critical for the maintenance of the HIV reservoir.
To provide evidence for a direct role of TGF-β in the induction of HIV latency, we developed an in vitro culture model106 where TGF-β was added to HIV-infected CD4 T-cells. Increased numbers of CD4 T-cells with integrated proviral DNA were observed after 14 days of culture with increasing concentrations (0 to 50 ng/mL) of TGF-β (r= 0.7; p = 0.009) (Fig. 6a), demonstrating that TGF-β contributes to heightened levels of non-replicative forms of HIV DNA and establishment of HIV latency. In addition, we observed (in line with ex vivo observations shown in Fig. 4b) an increase in surface PD-1 levels at the two highest doses of TGF-b in the CD4 TCM subset (Fig. 6b). We then reversed latency by stimulating these CD4 T cells with immobilized CD3 and soluble CD28 specific antibodies. Our data show that frequencies of HIV p24+ cells were significantly higher in conditions where the highest concentrations of TGF-β (0.2-20ng/mL) were added (Fig. 6c). These frequencies of HIV p24+ cells were correlated with surface PD-1 levels on CD4 TCM (Fig. 6d). Altogether, these in vitro/ex vivo observations validate the critical role of TGF-β producing Tregs in driving the mechanistic establishment and/or maintenance of the HIV reservoir.
Integrated multi-omic model highlights cellular and molecular effectors of senescence as drivers of HIV persistence and lack of immune reconstitution
Using unbiased and holistic approaches, we have highlighted gene expression profiles, cytokines and T-cell subsets that are independently associated with HIV persistence and lack of CD4 reconstitution in the Senescent INRs. To investigative the interplay of pathways driving these cellular phenotypes and HIV persistence, we integrated multi-omic signatures above (Figures 2, 3 and 4, Table S20) across all subjects (n=41). Our data show that the levels of inducible HIV were positively correlated to senescence - characterized by impaired metabolism97 (increased ROS, decreased OXPHOS), reduced transcription/translation (RPSs and EIFs) downstream of poor cMYC activity and TGF-β signaling (SMAD2/3 target genes) (Fig. 6e, Supplementary Fig. 5c). The induction of these cascades was consistent with higher expression of the transcriptional repressors that induced senescence (i.e. FOXO3, FOXO4)45, and with lower expression of master regulators of innate antiviral activity (i.e. IRF7 and restriction factors of viral replication) (Supplementary Fig. 6). Lack of CD4 reconstitution and HIV persistence was also driven by higher frequencies of GARP+ Tregs, increased expression of SMAD2/3 targets (including FURIN, SMOX) and heightened levels of IRF3 induced genes (i.e. AFF, DARC, GUK1) (Supplementary Fig. 6). Our integrated analyses indicate that most pathways that drive HIV persistence are negatively associated with the recovery of CD4 numbers upon cART initiation in senescent subjects; and provided further evidence for the direct interplay between Treg frequencies, TGF-β production, heightened FOXO3 expression, interferon signaling, establishment of cellular senescence, impaired T cell homeostasis, and quiescent cellular and metabolic state as conditions that favor the maintenance of HIV reservoir in the unique senescent INR population described here.