Identification of NK Cell Subpopulations That Differentiate HIV-Infected Subject Cohorts with Diverse Levels of Virus Control

HIV infection results in the establishment of a stable reservoir of latently infected cells; ART is usually required to keep viral replication under control and disease progression at bay, though a small subset of HIV-infected subjects can control HIV infection without ART through immunological mechanisms. In this study, we sought to identify subpopulations of NK cells that may be involved in the natural immunological control of HIV infection. We used mass cytometry to measure surface marker expression on peripheral NK cells. Using two distinct semisupervised machine learning approaches, we identified a CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells that differentiates HIV controllers from noncontrollers. These cells can be sorted out for future functional studies to assess their potential role in the immunological control of HIV infection.

ABSTRACT

HIV infection is controlled immunologically in a small subset of infected individuals without antiretroviral therapy (ART), though the mechanism of control is unclear. CD8⁺ T cells are a critical component of HIV control in many immunological controllers. NK cells are also believed to have a role in controlling HIV infection, though their role is less well characterized. We used mass cytometry to simultaneously measure the levels of expression of 24 surface markers on peripheral NK cells from HIV-infected subjects with various degrees of HIV natural control; we then used machine learning to identify NK cell subpopulations that differentiate HIV controllers from noncontrollers. Using CITRUS (cluster identification, characterization, and regression), we identified 3 NK cell subpopulations that differentiated subjects with chronic HIV viremia (viremic noncontrollers [VNC]) from individuals with undetectable HIV viremia without ART (elite controllers [EC]). In a parallel approach, we identified 11 NK cell subpopulations that differentiated HIV-infected subject groups using k-means clustering after dimensionality reduction by t-neighbor stochastic neighbor embedding (tSNE) or linear discriminant analysis (LDA). Among these additional 11 subpopulations, the frequencies of 5 correlated with HIV DNA levels; importantly, significance was retained in 2 subpopulations in analyses that included only cohorts without detectable viremia. By comparing the surface marker expression patterns of all identified subpopulations, we revealed that the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells are more abundant in EC and HIV-negative controls than in VNC and that the frequency of these cells correlated with HIV DNA levels. We hypothesize that this population may have a role in immunological control of HIV infection.

IMPORTANCE HIV infection results in the establishment of a stable reservoir of latently infected cells; ART is usually required to keep viral replication under control and disease progression at bay, though a small subset of HIV-infected subjects can control HIV infection without ART through immunological mechanisms. In this study, we sought to identify subpopulations of NK cells that may be involved in the natural immunological control of HIV infection. We used mass cytometry to measure surface marker expression on peripheral NK cells. Using two distinct semisupervised machine learning approaches, we identified a CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells that differentiates HIV controllers from noncontrollers. These cells can be sorted out for future functional studies to assess their potential role in the immunological control of HIV infection.

INTRODUCTION

HIV is a retrovirus that establishes a life-long latent infection and remains a global health burden despite highly effective therapies. HIV-infected individuals treated with suppressive antiretroviral therapy maintain undetectable HIV plasma viral loads, but this treatment is not curative due to the slow decay of a reservoir of latently infected cells (1, 2). Interruption of antiretroviral therapy is typically followed by a robust rebound of viremia; thus, therapy must be continued for the life of the individual. Intriguingly, a subset of HIV-infected individuals can control HIV load in the absence of antiretroviral therapy (3–6). Viremic controllers (VC) maintain a low HIV load (40 to 2,000 copies of HIV RNA/ml of plasma), and elite controllers (EC) maintain an undetectable HIV load (≤40 copies of HIV RNA/ml of plasma). The mechanism of immunological control in these subjects is not fully understood.

NK cells and cytotoxic T lymphocytes (CTLs) have both been implicated in immunological control of HIV infection (7). Certain major histocompatibility complex class I (MHC-I) molecules are overrepresented in controller cohorts, most notably, HLA-B*57 and HLA-B*27 (8, 9), and several studies have demonstrated a quantitatively and qualitatively superior CTL response in EC and VC who possess these protective MHC alleles (10–15). Nevertheless, a substantial proportion of controllers lack protective MHC alleles and may control HIV infection through a non-CTL mechanism (16). NK cells may contribute to the controller phenotype, both in the absence of (17) and in conjunction with (18–20) protective MHC-I alleles. Recently, in a case of early immunological control of HIV by a subject lacking protective MHC-I alleles, researchers found an effective cytotoxic NK effect without robust CTL activity (21).

In peripheral blood, classical NK cells are historically identified as CD56^bright CD16⁻ (cytokine-producing) and CD56^dim CD16⁺ (cytotoxic) (22). A nonclassical NK cell population (CD56⁻ CD16⁺ CD7⁺) with reduced functional capacity is expanded in HIV-infected subjects and may comprise recently activated cells that have become anergic (23–27). This subset has been described in other chronic viral infections, such as hepatitis C virus (HCV), but is largely absent in HIV-negative (HN) donors (28, 29).

NK cell activity is regulated by a balance of signals downstream of multiple activating and inhibitory receptors (30), including the killer cell immunoglobulin-like receptor (KIR), C-type lectin receptor (e.g., CD94), and natural cytotoxic receptor (NCR) families. These receptors interact with MHC-I, adhesion molecules, and costimulatory molecules, among others.

Interactions between HLA-B Bw4-80Ile alleles (including HLA-B*57) and KIR3DL1 as well as homozygosity for KIR3DS1 have been shown to be protective against HIV acquisition (31, 32) and progression (18–20), and viral evolution to escape pressure applied by KIR molecules has been described previously (33). A recent study demonstrated the importance of immunological control of HIV by NK cells that express NKG2A (34). In addition, upon activation with alpha interferon (IFN-α), NK cells are able to kill HIV-infected CD4⁺ T cells via activating NKG2D and NKp46 receptors (35), though the HIV Nef protein has been shown to be able to downregulate ligands for both NKG2D and KIRs (36, 37). Additionally, NCR molecules are downregulated during chronic HIV infection, possibly contributing to the dysfunction of NK cells (38).

Prior studies have characterized NK cells from HIV controllers (39) or compared NK cell functionality to clinical parameters (40). In this study, we used mass cytometry to simultaneously profile 24 NK cell surface markers on the single-cell level in 93 subjects with differing degrees of HIV infection control, including both VC and EC. We used semisupervised machine learning approaches to identify and quantify NK cell subpopulations associated with the ability to control HIV infection.

RESULTS

Subject characterization.

Subjects were categorized into 5 groups based on HIV load and antiretroviral therapy status (see Materials and Methods). Clinical characteristics of the participating subjects are presented in Table S1 in the supplemental material. The distributions of subject CD4⁺ T cell counts and viral loads are shown in Fig. 1A.

FIG 1.

Study subject characterization. A total of 93 subjects were recruited for participation in this study (EC, n = 13, red; VC, n = 27, blue; VNC, n = 12, yellow; cART, n = 21, green; HN, n = 20, gray). (A) Subject peripheral blood CD4⁺ T cell count (left) and HIV load (right). All subjects had a CD4⁺ T cell count above 200 cells/μl of blood (left, dotted line). The limit of detection for viral load was 40 copies of HIV RNA/ml of plasma (right, dotted line). (B) HIV DNA measurement for each subject. The limit of quantitation was 10 copies/10⁶ PBMCs (dotted line). (C) Events were first gated on intact cells, followed by dead cell discrimination gating. NK cells were identified as CD3⁻ CD14⁻ CD19⁻ CD33⁻ CD45⁺ CD66b⁻ CD235⁻ and then differentiated by CD56, CD16, and CD7 expression. (D) Frequency of NK cell subsets within lineage-negative (CD3⁻ CD14⁻ CD19⁻ CD33⁻ CD45⁺ CD66b⁻ CD235⁻) PBMCs. (E) Expansion of CD56⁻ CD16⁺ CD7⁺ NK cells as measured by the ratio of CD56⁻ CD16⁺ CD7⁺ to CD56^dim CD16⁺ NK cells. Each color-filled point represents an individual subject, and horizontal lines represent the group medians. Significance was calculated using a Kruskal Wallis test with a post hoc Dunn’s test. Each subject group was compared to VNC. P values were adjusted by multiplying by the number of comparisons made (Bonferroni correction). *, P < 0.05; **, P < 0.005; ***, P < 0.0005.

We estimated HIV reservoir size for HIV-infected subjects with a previously published PCR-based approach (41, 42). Although PCR-based total HIV DNA measurements in peripheral blood mononuclear cells (PBMCs) detect defective provirus and unintegrated DNA and may not accurately measure reservoir size (43, 44), these measurements can distinguish groups of HIV-infected subjects and can be used as a proxy for reservoir size. Our data are consistent with other reports of HIV DNA reservoir measurements in the subject groups representing combination antiretroviral therapy (cART) and viremic noncontrollers (VNC) (41) (Fig. 1B). HIV DNA levels were higher in most VNC subjects than in cART subjects, likely due to detection of actively infected cells, though this difference was not statistically significant. VC and EC subjects had significantly lower HIV DNA levels than VNC subjects (P = 0.0051 and P = 0.0001, respectively).

Mass cytometry NK cell phenotyping.

To fully explore the repertoire of NK cell subsets by simultaneously interrogating the expression levels of a wide variety of activating and inhibitory receptors, we utilized a comprehensive mass cytometry approach. We designed and qualified an NK cell-focused panel (Table S2) of 38 antibodies (Abs). A total of 34 antibodies passed our quality control, as 4 antibodies had either minimal staining or high background levels. All 10 antibodies chosen for cell lineage identification passed quality control. All 93 subjects were subjected to immune phenotyping by mass cytometry in 6 independent runs. Run-to-run variability was minimal, as determined by t-neighbor stochastic neighbor embedding (tSNE) (data not shown). We interrogated mass cytometry results in two ways: first, we measured correlations between individual surface marker expression frequencies and the size of the HIV DNA reservoir; second, we used semisupervised machine learning approaches to identify NK cell subpopulations that differed in abundance among the enrolled HIV-infected subject groups.

Expansion of CD56⁻ NK cells occurs in viremic subjects.

Because NK cell subset imbalance has been observed in subjects chronically infected with HIV (23–27), we measured the frequencies of the two major NK cell subsets (CD56^bright CD16⁻ and CD56^dim CD16⁺) and of the more recently described CD56⁻ CD16⁺ CD7⁺subset in all HIV-infected subject groups (23) to verify that the NK cell subset distribution in our subject cohort reflected previous findings (23–27). The gating strategy used is shown in Fig. 1C. There were no significant differences in the frequencies of either of the two major NK cell subsets across HIV-infected subject groups, compared to HN, after adjusting for multiple comparisons (Fig. 1D). We observed a higher frequency of CD56⁻ CD16⁺ CD7⁺ NK cells within lineage-negative cells (CD3⁻ CD14⁻ CD19⁻ CD33⁻ CD45⁺ CD66b⁻ CD235⁻) in VNC than in HN (P = 0.014, false discovery rate [FDR] adjusted). Additionally, we saw a significantly higher ratio of CD56⁻ CD16⁺ CD7⁺ cells to CD56^dim CD16⁺ cells in VNC than in HN (P < 0.0005, FDR adjusted), suggesting that CD56⁻ CD16⁺ CD7⁺ cells may have expanded in VNC subjects (Fig. 1E).

Frequencies of individual markers on NK cell subsets correlate with HIV DNA.

To identify the NK cell subpopulations whose frequencies correlated with immunological control, we calculated correlations between the frequencies of NK cell subsets expressing individual markers and viral load, CD4⁺ T cell count, or HIV DNA levels. We observed a number of individual NK cell markers that correlated with viral load or CD4⁺ T cell count, some of which have been previously reported as differentially expressed in HIV controllers, including LIR1 (45), though none of these correlations withstood Bonferroni correction for multiple comparisons (P > 0.05). The only significant correlations remaining after correction for multiple comparisons were those between HIV DNA and expression of either CD62L or Siglec-7 on CD56^dim CD16⁺ NK cells (P = 0.001 and P = 0.001, respectively) or of either CD27 or Siglec-7 on CD56⁻ CD16⁺ CD7⁺ NK cells (P= 0.011 and P = 0.026, respectively) (data not shown). These results corroborate findings from multiple previous studies (26, 46–48).

CD56^dim CD16⁺ NK cell subpopulations that differentiate EC from VNC.

In order to discover novel NK cell subpopulations defined on the basis of multiple markers, we analyzed the entire mass cytometry panel concurrently. We first used CITRUS (49) to identify NK cell subpopulations that differentiated groups of HIV-infected subjects. In order to minimize noise, our analysis included only NK cells and NK cell-related markers (i.e., lineage markers were used to gate NK cells but were excluded from downstream analyses). HIV-infected subject groups were compared in a pairwise fashion. Using CITRUS, we could not distinguish any HIV-infected subject groups from one another using the CD56^bright CD16⁻ NK cell subset (≤15% model cross-validation error rate). However, we were able to differentiate EC from VNC using the CD56^dim CD16⁺ NK cell subset. Three clusters of cells identified by CITRUS within CD56^dim CD16⁺ NK cells (clusters 24427, 24429, and 24432) differentiated EC from VNC (Fig. 2A). The expression levels of selected markers that helped differentiate these clusters of cells are shown in Fig. 2B; the expression levels of all NK markers measured are shown in Fig. S1 in the supplemental material.

FIG 2.

CITRUS analysis of NK cell subsets revealed three clusters within CD56^dim CD16⁺ NK cells that differentiate EC from VNC. (A) Frequency of cells within clusters identified by CITRUS to differentiate EC (red) and VNC (yellow) subject groups. Each color-filled point represents an individual subject, and horizontal lines represent the group median. Significance was calculated using the Kruskal Wallis test. P values were further adjusted by multiplying by the number of comparisons made (Bonferroni correction). **, P < 0.005; ***, P < 0.0005. (B) Expression intensity of markers that distinguish bulk CD56^dim CD16⁺ NK cells (black line, shaded background) from the cells identified in cluster 24427, 24429, or 24432 (lines of gradating purple color). Histograms representing the expression intensities of all NK markers measured are shown in Fig. S1. x-axis data represent marker expression intensity; y-axis data represent frequency.

Dimensionality reduction followed by k-means clustering revealed NK cell subpopulations that differentiated multiple HIV-infected subject groups.

As a distinct approach, we performed supervised and unsupervised dimensionality reduction prior to cluster identification and pairwise comparisons of cluster frequencies. We concatenated cells from subjects across all HIV-infected subject groups and ran principal-component analysis (PCA), linear discriminant analysis (LDA), or tSNE. PCA and tSNE are both unsupervised dimensionality reduction algorithms that preserve the information in a data set while allowing identification of new clusters. In contrast, LDA is a supervised dimensionality reduction algorithm that separates data points using defined response variables as a guide to maximize separation. Using ACCENSE (50), we used the first two dimensions from each dimensionality reduction algorithm for k-means clustering to define NK cell subpopulations. To our knowledge, using a supervised dimensionality reduction algorithm prior to clustering is a novel approach to identify cellular populations that differentiate subject groups from high-parameter mass cytometry data sets.

Dimensionality reduction by PCA performed poorly, as the information contained in the first principal component was explained almost entirely by expression of CD57. From two-dimensional coordinates generated by tSNE, we identified one subpopulation within the CD56⁻ CD16⁺ CD7⁺ NK cells (cluster 1) that differentiated viremic subjects (VC, VNC) from aviremic subjects (EC, cART, HN) (Fig. 3A and B). This NK cell subpopulation expressed higher levels of KIR3DL1 (Fig. 3C), suggesting a loss of this inhibitory marker in response to viremia.

FIG 3.

Clusters identified after dimensionality reduction and k-means clustering differentiate viremic subjects from aviremic subjects. (A and B) Frequency of cells within cluster identified after running tSNE analysis on CD56⁻ CD16⁺ CD7⁺ NK cells followed by k-means clustering. Each color-filled point represents an individual subject (orange, EC, cART, HN; purple, VC, VNC), and horizontal lines represent the group median. Significance testing was done by the Kruskal Wallis test with a post hoc Dunn’s test. ***, P < 0.0005. (C) Expression intensity of markers that distinguish bulk CD56⁻ CD16⁺ CD7⁺ NK cells (black line, shaded background) from the cells identified in cluster 1 (green line). Histograms representing the expression intensities of all NK markers measured are shown in Fig. S2. x-axis data represent marker expression intensity. (D and E) Frequency of cells within clusters identified after running LDA on CD56^bright CD16⁻ NK cells followed by k-means clustering. Each color-filled point represents an individual subject (orange, EC, cART, HN; purple, VC, VNC), and horizontal lines represent the group median. Significance testing was done by Kruskal Wallis test with a post hoc Dunn’s test. P values were further adjusted by multiplying by the number of comparisons made (Bonferroni correction). **, P < 0.005; ***, P < 0.0005. (F) Expression intensity of markers that distinguish bulk CD56^bright CD16⁻ NK cells (black line, shaded background) from the cells identified in clusters 2 to 6 (lines in gradations of red). Histograms representing the expression intensities of all NK markers measured are shown in Fig. S3. x-axis data represent marker expression intensity.

From two-dimensional coordinates generated by LDA, k-means clustering identified five subpopulations within CD56^bright CD16⁻ NK cells (clusters 2 to 6) that differentiated viremic subjects (VC, VNC) from aviremic subjects (EC, cART, HN) (Fig. 3D and E). The relevant markers for identification of these NK cell subpopulations are shown in Fig. 3F. The cells in clusters 2 to 4, which were less abundant in viremic subjects, showed high levels of CD11b and CD161 expression. Cells expressing these markers are matured and play a proinflammatory role (51, 52); CD161-expressing NK cells have been shown previously to be depleted in viremic HIV-infected subjects (52). The cells in clusters 5 to 6, which were more abundant in viremic subjects, had higher expression of DNAM-1 and NKp46, suggesting that they were matured and active NK cells (53, 54).

Also, using LDA followed by k-means clustering, we identified an additional five subpopulations within the CD56^dim CD16⁺ NK cells (clusters 7 to 11) that differentiated VNC from other groups of HIV-infected subjects (Fig. 4A). Similarly to the cells in clusters 2 to 4, the cells in clusters 7 to 11 expressed elevated levels of CD11b and CD161, but they also showed increased expression of Siglec-7 and lacked expression of CD57 (Fig. 4B). Clusters 7 to 11 were differentially abundant between VNC and EC, with intermediate frequencies observed in the VC and cART groups. The cells identified by cluster 10 were also more frequent within the EC group than within the HN group (P = 0.001, FDR adjusted).

FIG 4.

Several clusters identified after dimensionality reduction differentiate subject groups. (A) Frequency of cells within clusters identified after running LDA on CD56^dim CD16⁺ NK cells. Each color-filled point represents an individual subject, and horizontal lines represent the group median. Significance was tested using the Kruskal Wallis test with a post hoc Dunn’s test. Each subject group was compared to VNC. Additionally, EC and HN groups were tested for normal distribution (D’Agostino & Pearson, Shapiro-Wilk, and KS normality tests); cluster frequencies with distributions that did not significantly differ from the normal distribution were compared by unpaired t test, and cluster frequencies with distributions that did significantly differ from the normal distribution were compared by Mann-Whitney test. P values were all further adjusted by multiplying by the number of comparisons made (Bonferroni correction). *, P < 0.05; **, P < 0.005; ***, P < 0.0005. (B) Expression intensity of markers that distinguish bulk CD56^dim CD16⁺ NK cells (black line, shaded background) from the cells identified in clusters 7 to 11 (lines of gradating blue color). Histograms representing the expression intensities of all NK markers measured are shown in Fig. S4. x-axis data represent marker expression intensity.

We next performed a linear regression analysis to determine whether the frequencies of the cells in clusters 1 to 11 were associated with HIV DNA levels. HIV DNA levels correlated significantly with the frequencies of cells identified by clusters 7 to 11 (Fig. 5A); excluding the viremic subjects from the analysis (Fig. 5B), clusters 8 and 10 maintained significant correlations with HIV DNA levels. Importantly, this suggests that these clusters were correlated with integrated HIV DNA as an approximate indicator of the frequency of cells containing proviral DNA and representing the overall infection burden in T cells rather than a latent replication-competent virus reservoir. This conclusion was unaffected by converting the HIV DNA copy number per 10⁶ PBMCs to the HIV DNA copy number per 10⁶ CD4⁺ T cells for correlative analysis (data not shown).

FIG 5.

Frequencies of clusters 7 to 11 correlate with HIV DNA levels. The frequency of cells within clusters 1 to 11 was correlated with the log value of HIV DNA measurements using an ordinary least-squares regression. (A) All HIV-infected subject groups were included in the analysis. Only correlations with slopes that differ significantly from 0 are shown. (B) Only aviremic HIV-infected subjects (EC, cART) were included in the analysis. Each color-filled point represents an individual subject, and each solid line represents the line corresponding to the best fit. P values were adjusted by multiplying by the number of comparisons made (Bonferroni correction) and are displayed in the upper right corner of each individual graph. EC, red; VC, blue; VNC, yellow; cART, green.

The CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells differentiated HIV-infected subject groups.

The expression levels of CD11b, CD161, CD57, and Siglec-7 on cells identified by clusters 7 to 11 followed broadly similar distribution patterns (Fig. 4B). When clusters 7 to 11 were combined, we observed that the distribution of these markers on these cells showed a profile similar to that seen with the cells in cluster 24429 identified by CITRUS. These relationships are shown in Fig. 6A. This suggests that two disparate machine learning algorithms independently converged on identification of a similar NK cell subpopulation that showed a high level of expression of CD11b, CD161, and Siglec-7 and a low level of expression of CD57.

FIG 6.

Computational approaches identify a novel NK cell population that differentiates HIV-infected subject groups. (A) Histograms representing the expression intensities of markers that distinguish all CD56^dim CD16⁺ NK cells (black line, shaded background) from the cells identified in clusters 7 to 11 combined (LDA, blue line) or in cluster 24429 (CITRUS, gold line). x-axis data represent marker expression intensity. (B) Frequency of CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ cells within the CD56^dim CD16⁺ NK cell subset (left). The frequency of the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ CD56^dim CD16⁺ NK cell subpopulation was correlated with the log value of HIV DNA measurements using an ordinary least-squares regression (right). Each color-filled point represents an individual subject, and horizontal lines represent the group median. Significance was calculated using the Kruskal Wallis test with a post hoc Dunn’s test. Each subject group was compared to VNC. *, P < 0.05.

Using manual gating approaches, we compared the frequencies of CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ cells within the CD56^dim CD16⁺ NK cell population across the HIV-infected subject groups and the HN group (Fig. 6B, left). This subpopulation was more frequently observed in EC and HN than in VNC (P= 0.048 and P = 0.010, respectively). Though CD161⁺ cells within the CD56^dim CD16⁺ NK cell population have been previously shown to differentiate cART and VNC subjects (52), all 4 markers were necessary to identify a subpopulation that differentiated the HIV-infected subject groups in this study. We confirmed the expression of CD57 and CD161 from these mass cytometry findings in an independent experiment performed using traditional flow cytometry to verify that these markers alone were not sufficient to identify the NK cell subpopulation (data not shown). The frequency of this NK cell subpopulation, either per 10⁶ PBMCs or per 10⁶ CD4⁺ T cells, correlated with HIV DNA measurements (Fig. 6B, right).

Increased functionality in the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells.

In order to determine if the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells had enhanced functional capacity, we used standard NK cell stimulation conditions to measure the levels of cytokine production and degranulation. Enriched NK cells showed low production of IFN-γ without stimulation. Under conditions of stimulation with interleukin-12 (IL-12)/IL-18, CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ cells showed significantly more IFN-γ production than CD11b⁻, CD161⁻, or Siglec-7⁻ subpopulations (Fig. 7A). We next cocultured enriched NK cells with K562 cells; K562 cells are MHC class I deficient and act as NK cell targets. NK cells cultured without K562 cells had a low background level of degranulation. We observed a significantly higher frequency of degranulation in the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cells than in CD11b⁻, CD161⁻, or Siglec-7⁻ subpopulations (Fig. 7B). The CD57⁺ subpopulation, which is believed to be “antigen-experienced” (55), showed high levels of cytokine production and degranulation; there was no significant difference between CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ cells and CD57⁺ cells under either of the sets of stimulation conditions.

FIG 7.

Functional activity of the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation of CD56^dim CD16⁺ NK cell subset. (A) Frequency of IFN-γ producing NK cells with or without stimulation with IL-12/IL-18. + and − above the horizontal line indicate whether stimulation was added; + and − below the horizontal line indicate the expression of indicated marker on gated CD56^dim CD16⁺ NK cell subpopulation. (B) Frequency of degranulating NK cells as measured by CD107a analysis during stimulation with or without coculture of K562 target cells. + and − above the horizontal line indicate whether stimulation was added; + and − below the horizontal line indicate the expression of the indicated marker on the gated CD56^dim CD16⁺ NK cell subpopulation. Each color-filled point represents an individual subject, and horizontal lines represent the group medians. Significance was calculated using a Friedman’s test (paired) with a post hoc Dunn’s test. P values were adjusted by multiplying by the number of comparisons made (Bonferroni correction). n.s., no significant difference; **, P < 0.005; ***, P < 0.0005.

DISCUSSION

We used a multifaceted approach to identify novel NK cell phenotypes and subpopulations associated with immunological control of HIV. A key advantage of mass cytometry over traditional flow cytometry is the ability to generate high-parameter data sets to interrogate many cellular phenotypic differences across subject cohorts simultaneously, considering a comprehensive receptor repertoire rather than individual receptors in isolation. High-parameter data sets can be interrogated computationally in an unbiased fashion to uncover novel subsets of cells that may not be identified by more-limited flow cytometry panels.

We identified several clusters that differentiated subjects with or without active HIV viremia, a finding that suggests that NK cell receptor expression may be impacted by the presence of HIV antigens and/or active viral replication. More intriguingly, after LDA dimensionality reduction, the frequencies of cells identified in clusters 7 to 11 correlated with HIV DNA measurements. Importantly, the frequencies of cells identified in cluster 8 and cluster 10 remained significantly correlated with HIV DNA in the analyses that included only aviremic subjects (EC and cART; Fig. 5B). The cells in clusters 7 to 11 were more abundant in EC than in VNC and when combined showed a pattern of receptor expression similar to that seen with cluster 24429 (Fig. 6A).

The NK cell clusters identified from the semisupervised machine learning approaches used in this study are largely distinguished from other CD56^dim CD16⁺ NK cells by their reduced expression of CD57 and increased expression of CD11b, CD161, and Siglec-7.

Some of the markers that define this cell subset have been implicated in previous studies of NK cell function. Siglec-7 is a sialic acid-binding protein that has an immune tyrosine-based inhibition motif (ITIM). Siglec-7 is expressed on highly active NK cells (56), and loss of expression is believed to correspond to dysfunction of NK cells, which occurs in chronic HIV and HCV infection (46, 57). CD11b is a marker of maturation of NK cells in humans and mice (58, 59). CD57 is a marker of terminally matured NK cells (60) and has been suggested to identify antigen-experienced NK cells; consistent with this hypothesis, the frequency of CD57-expressing NK cells is higher in HIV-infected subjects (61). CD161 is a C-type lectin expressed on some T cells and on the majority of NK cells early in maturation (62). CD161 marks cytotoxic NK cells, but interaction with its cognate ligand LLT1 has been shown to repress NK cell function (63, 64). CD161⁺ NK cells, like CD161⁺ T cells, have an enhanced ability to respond to cytokines, including IL-12 and IL-18 (52, 65), and CD161⁺ CD57⁻ NK cells have enhanced proliferative capabilities (52). On the basis of the identified marker expression, the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ CD56^dim CD16⁺ NK cells are likely highly active, partially matured, cytotoxic NK cells (51, 52, 56, 60) and thus might contribute to the spontaneous control of HIV replication in the absence of antiretroviral therapy. Indeed, our in vitro data support this hypothesis, as the CD11b⁺ CD57⁻ CD161⁺ Siglec-7⁺ subpopulation showed a higher response to stimulation than other CD56^dim CD16⁺ NK cell subsets.

NK cells represent a heterogeneous population, and the complexity of NK cell maturation and activation is an area of ongoing investigation. The identification of NK cell clusters that correlate with HIV DNA levels could be a fingerprint of antigen exposure level during the course of infection or, more intriguingly, could reveal NK phenotypes that contribute to the control of viral replication in early stages of infection. The interactions between the identified differentially expressed receptors on antigen-inexperienced CD56^dim CD16⁺ NK cells, including CD11b, CD161, and Siglec-7, and the functional impact of their modulation on the NK cell activity could contribute to better understanding of various distinct mechanisms of HIV control and thus warrant further exploration. Future studies to understand the effect of modulating this NK cell subpopulation would benefit from knowing if a similar subpopulation exists in animal species.

MATERIALS AND METHODS

HIV-infected subject recruitment.

Prior to enrollment in the study, all subjects provided informed consent as approved by the Committee on Human Research of the University of California, San Francisco. A total of 93 subjects were recruited from the SCOPE Study cohort as follows: 21 chronically infected, antiretroviral therapy-treated subjects with a viral load of <40 copies/ml of plasma (cART); 12 chronically infected, antiretroviral therapy-naive subjects with a viral load of >5,000 copies/ml of plasma (viremic noncontrollers [VNC]); 27 antiretroviral therapy-naive subjects with a viral load of between 40 and 2,000 copies/ml of plasma (viremic controllers [VC]); 13 antiretroviral therapy-naive subjects with a viral load of <40 copies/ml of plasma (elite controllers [EC]); and 20 HIV-seronegative subjects (HIV-negative [HN]). Viral load measurements were confirmed in-house using the COBAS TaqMan test v2.0. HN were verified as HIV antibody negative by Western blotting. All subjects were screened to have a CD4⁺ T cell count of >200 cells/μl of blood. All subjects were screened to be HCV antibody negative and/or HCV load negative at the time of donation; HCV antibody positivity was not an exclusion criterion if the subject had undetectable HCV load at the time of sample collection. Additional details are available in Table S1 in the supplemental material.

Sample preparation and storage.

Whole blood was collected in acid citrate dextrose (ACD)-containing Vacutainers. Blood was processed on the day of collection. PBMCs were isolated by density centrifugation over Ficoll-Paque media. PBMCs were counted and resuspended at 10 M cells/ml of freezing media (90% fetal bovine serum [FBS], 10% dimethyl sulfoxide [DMSO]). Cells were frozen to −80°C in a controlled-rate freezing chamber. The next morning, samples were transferred to liquid nitrogen storage and maintained at −160°C for long-term storage.

Genomic DNA isolation and cell-associated HIV quantification.

Genomic DNA was isolated from thawed PBMCs, and HIV was quantified as previously reported. In brief, genomic DNA was isolated by sonicating cells, treating them with chaotropic salts, and precipitating them with alcohol (41). Genomic DNA isolated from U1 cells (containing 2 copies of integrated HIV/cell) was used to make a standard curve for human genome copy number as well as HIV copy number determinations. CCR5 was used as a standard control for human genome copy number determinations (66). HIV was detected using primers and probes within the HIV integrase gene as previously optimized for the integrase single-copy assay (67). Quantitative PCR (qPCR) was performed using TaqMan Fast Advanced master mix (Thermo Fisher) in a final volume of 25 μl (42).

Mass cytometry Ab validation.

Monoclonal antibodies were conjugated to metals isotopes (Table S2) using a MaxPAR antibody conjugation kit (Fluidigm) and were diluted to 0.2 to 0.4 mg/ml in antibody stabilizer phosphate-buffered saline (PBS) (Candor Bioscience). Optimal staining concentrations for each antibody were determined by titrating antibody concentrations prior to staining cell lines and primary cells as positive and negative controls.

Sample processing and data acquisition.

Frozen PBMCs were thawed rapidly at 37°C. Prewarmed RPMI 1640 supplemented with 10% FBS and 0.1 mg/ml DNase I was added dropwise in a swirling motion to maximize viability and prevent clumping. Each sample was treated with cisplatin for the exclusion of dead cells (68) and fixed with 1.5% paraformaldehyde. A total of 1 × 10⁶ cells from each sample were aliquoted, and pre-Perm palladium barcoding was performed as previously described (69, 70). Cells were then pooled, treated with Fc receptor blocking solution (BioLegend), and stained with metal-labeled antibodies and DNA intercalators as previously reported (69, 71–73). Cells were washed and resuspended with normalization beads (74) prior to being loaded into a CyTOF 1 instrument for acquisition (Fluidigm). All samples were run in a total of 6 batches. PBMCs from the same 3 subjects (1 EC, 1 cART, and 1 HN) were included in each run for all 6 batches to verify consistency from run to run.

After acquisition, individual markers were checked to verify that the positive population had been successfully resolved from the negative population. A total of 3 antibodies (2B4, NKp30, and NKp44) showed minimal staining. CD137-positive cells could not be resolved from CD137-negative cells due to high background levels. Those four markers were excluded from downstream analysis. Additionally, t-distributed stochastic neighbor embedding (tSNE) was run to rule out batch effects (data not shown).

Linear regressions and statistical tests.

Ordinary least-squares linear regressions were calculated using Python (statsmodels). Pairwise comparisons were performed using a Kruskal-Wallis (KW) test with a post hoc Dunn’s test (https://gist.github.com/alimuldal/fbb19b73fa25423f02e8) or Friedman’s paired test (GraphPad Prism). To correct for false discovery, P values corresponding to the KW-Dunn results were further adjusted by multiplying by the number of comparisons made (Bonferroni correction). All reported P values were adjusted for multiple testing.

CITRUS analysis.

CITRUS (cluster identification, characterization, and regression) (49) was run using a Cytobank platform. CD56^bright CD16⁻, CD56^dim CD16⁺, and CD56⁻ CD16⁺ CD7⁺ NK cells were downsampled to equal sample sizes for each subject (CD56^bright CD16⁻, 206 cells; CD56^dim CD16⁺, 1,111 cells; CD56⁻ CD16⁺ CD7⁺, 265 cells). Subjects with too few cells per NK cell subtype were excluded from analysis. Cluster sizes were limited to a minimum of 5% of total cells. HIV-infected subject groups were compared in a pairwise fashion. Cells were concatenated from subjects within the two subject groups being compared. After unsupervised agglomerative hierarchical clustering, cluster abundances were inputted into a LASSO-regularized logistic regression to build a logistic classifier; hyperparameters were optimized using k-fold cross-validation.

Dimensionality reduction and cluster identification.

The downsampled data used for CITRUS were used also for principal-component analysis (PCA), linear discriminant analysis (LDA), and tSNE using Python (scikit-learn). PCA prioritizes describing data sets as a combination of attributes (e.g., in this study, individual marker intensities), representing their variation through principal components (PCs), which preserve as much variance in the data set as possible. The focus of tSNE is to deconvolute high-parameter data by identifying similarities between data points (e.g., in this study, individual cells) and then projecting the relationships in two dimensions. Both of these approaches preserve the information in a data set in an attempt to portray the data in a way that allows the user to identify new clusters in an agnostic fashion (e.g., in this study, to identify NK cell subpopulations without taking HIV-infected subject groups into consideration). LDA is a supervised dimensionality reduction algorithm that identifies combinations of attributes (e.g., in this study, individual marker intensities) that maximize the separation of a given response variable (e.g., in this study, HIV subject groups); LDA does not maintain the overall variance of the data set but separates the data points based on response variables more effectively (e.g., in this study, identified the NK cell subpopulations that differentiated HIV-infected subject groups). tSNE was run with a random seed, a maximum of 5,000 iterations, and perplexity set to 50, using the Barnes-Hut approximation (angle = 0.5). LDA was run using eigenvalue decomposition and was limited to two dimensions. Clusters were subsequently identified by k-means clustering using the ACCENSE package (50).

NK cell functional assays.

Frozen PBMCs were thawed rapidly at 37°C and washed in RPMI 1640 supplemented with 10% FBS. Cells were rested at 37°C and 5% CO₂ for 3 h prior to enrichment of NK cells (Stemcell Technologies). After being rested overnight at 37°C and 5% CO₂, 2 × 10⁵ NK cells were plated in individual wells of a 96-well flat bottom plate and stimulated with either cytokines (10 ng/ml IL-12 [BioLegend] plus 100 ng/ml IL-18 [BioLegend]) or K562 cells (10:1 effector/target ratio) for 5 h at 37°C and 5% CO₂. After the first hour of stimulation, brefeldin A (BioLegend), monensin (BioLegend), and CD107a (clone H4A3; BioLegend) were added. After stimulation, cells were stained for viability (Zombie dyes; BioLegend), CD3 (clone UCHT1; BioLegend), CD11b (clone ICRF44; BioLegend), CD16 (clone 3G8; BioLegend), CD56 (clone HDC56; BioLegend), CD57 (clone H CD57; BioLegend), CD161 (clone 191B8; Miltenyi Biotec), and Siglec-7 (clone 6 to 434; BioLegend). Cells were fixed and permeabilized with CytoFix/CytoPerm reagent (BD Biosciences) prior to staining for intracellular IFN-γ (clone B27; BioLegend). Stained cells were analyzed on an LSRFortessa X-20 instrument (BD Biosciences), and data were analyzed using FlowJo (TreeStar).