Integrated Single-Cell Profiling Reveals Dichotomous NK Cell Populations Associated with Immunosuppression in Solid Tumors

Abstract

Natural killer (NK) cells represent key effectors of antitumor immunity, yet emerging evidence highlights populations with distinct roles in cancer. Despite such expanded diversity within the NK cell repertoire, we lack an understanding of how this heterogeneity impacts immune responses and downstream clinical outcomes. Using single-cell RNA-sequencing (scRNA-seq), we systematically profiled NK cells across cancer and uncovered a dichotomous phenotypic and functional landscape of tumor-infiltrating NK cells shaped by opposing intrinsic signaling programs that drive the expression of IFNG or TGFB1. These divergent programs are associated with distinct transcription factor circuits that integrate cues within the tumor microenvironment and skew NK cells towards pro-inflammatory or suppressive functions. We found that the capacity for NK cells to engage in either functional direction is intrinsically linked to their phenotypic identity. Canonical NK cells recruited from circulation predominantly directed suppressive TGFB1 signals towards effector CD8⁺ T cells in tumors. Of note, these subsets exhibited higher TGFB1 expression than intratumoral myeloid cells across tumor types. In contrast, a tissue-resident adaptive subset exhibited exclusively pro-inflammatory IFNG-driven profiles and was associated with prolonged survival in both primary and metastatic tumor settings. Moreover, these tissue-resident adaptive NK cells, but not other subsets, were linked to response to immune checkpoint blockade. Collectively, our study reveals a previously unrecognized regulatory axis in NK cells that shapes NK cell diversity and augments broader antitumor immune responses.

INTRODUCTION

NK cells are cytotoxic innate lymphocytes that play pivotal roles in tumor immunosurveillance (1–3). Operating independently of antigen specificity, NK cells can rapidly recognize and eliminate transformed targets without prior sensitization. In addition to direct cytolytic activity, NK cells secrete proinflammatory cytokines and chemokines that shape the tumor microenvironment and coordinate broader immune responses (3,4). Collectively, these potent antitumor functions have positioned NK cells as promising immunotherapeutic agents, fueling strategies such as adoptive transfer of engineered NK cells, antibody-mediated engagement, and cytokine-based activation (1,5,6). While early clinical trials have demonstrated encouraging efficacy and safety, durable responses remain limited (7–9), highlighting the need to better understand the regulation of NK cells in tumors.

As the therapeutic landscape of NK cells has evolved, so too has our understanding of basic NK cell biology. Once viewed as a homogenous population in peripheral blood, NK cells are now recognized as a diverse lineage comprising phenotypically and functionally distinct subsets (10,11). While high-dimensional profiling studies have primarily characterized populations conventionally found in circulation (12–15) – namely cytotoxic CD56^dim and pro-inflammatory CD56^bright NK cells – further work understanding the roles of more contemporary tissue-resident (TR) and adaptive, memory-like NK cells is warranted, particularly as these populations may vary across tissues. Adding further complexity, NK cells can acquire regulatory or even pro-tumorigenic phenotypes in certain contexts, challenging the classical view of their exclusively protective function (16–21).

Here, we investigated the phenotypic and functional diversity of tumor-infiltrating NK cells across a spectrum of solid tumors. Leveraging publicly available scRNA-seq datasets, we uncovered a TGFB1–IFNG functional switch that differentiates canonical subsets from tissue-resident and adaptive NK cell populations in tumors. Our findings highlight the dualistic potential of NK cells in cancer and delineate molecular programs that shape their fate and function in cancer, laying a framework to guide the development of more durable NK cell-based immunotherapies.

MATERIALS AND METHODS

scRNA-seq compilation, processing, and cell-type annotation

Publicly available scRNA-seq datasets were compiled from studies listed in table S1. Datasets were selected based on use of the 10x Genomics platform and inclusion of primary, therapy-naïve tumor samples. Sex and age were not considered for analyses. Power analysis was not performed, and sample sizes were based on availability of data. All datasets were collected as count matrices aligned and mapped to the GRCh38 human reference genome. Low-quality cells were excluded based on total unique genes expressed (< 200) and the proportion of mitochondrial RNA expression (>20%). Data were log-normalized using the standard Seurat pipeline (v4.3.0).(22)

Samples from the same cancer type were merged for downstream analysis. The top 2,000 highly variable genes (HVGs) were selected for principal component analysis (PCA), and batch effects were corrected using Harmony (v1.2.0) (23). We applied Uniform Manifold Approximation and Projection (UMAP) using Seurat for dimensionality reduction to assess integration quality and to perform unsupervised clustering.

Cell-type annotations were performed using SingleR (v1.4.1) (24) in conjunction with Blueprint/ENCODE reference transcriptomes, enabling a broad and unbiased classification of immune and non-immune cell populations within each cancer type.

NK cell subset annotation

NK cells were first isolated based on SingleR annotations for each cancer type, which captured a broader NK cell population than marker-based approaches alone. NK cells co-expressing lineage-defining markers for B cells (CD19 or MS4A1) and T cells (CD4 and CD5) were excluded from downstream analyses to enhance purity.

To comprehensively characterize NK cell subsets, we performed two rounds of unsupervised clustering within each cancer type. The first round identified transcriptomically distinct clusters. Differentially expressed genes (DEGs) were computed across clusters to define candidate phenotypic features. Low-quality or ambiguous clusters lacking definitive NK cell subset markers were removed. From the remaining clusters, we performed a second round of DEG analysis to ensure the fidelity of NK cell signatures after removing contaminating populations. The gene expression profile of each cluster was inspected manually to assign high-confidence labels related to NK cell phenotypic markers (Table S2). These manual annotations were independently performed and confirmed by two NK cell biologists (JRL and FC) to ensure accuracy. It was crucial to perform these analyses independently in each cancer type considering the heterogeneity of NK cell markers across different tissues. Final annotations were cross-validated through alignment with consensus gene signatures for NK cell subsets in human blood (12).

Gene regulatory network inference

To elucidate potential transcriptional programs underlying NK cell identity, we utilized pySCENIC (1.3.1) followed by GRNBoost was used to identify transcription factors and regulons based on their co-expression with target genes.(25,26) Putative regulons were filtered based on biological relevance and scaled by enrichment. For subset-specific regulon analysis, we used the six NK cell subsets identified through unsupervised clustering and manual annotations as input. To capture context-dependent regulatory programs, we performed additional analyses using NK cell subsets from three distinct sources: (1) tumor-infiltrating NK cells, (2) peripheral blood NK cells from cancer patients, and (3) peripheral blood NK cells from healthy donors. These analyses allowed us to identify both shared and context-specific regulatory modules.

Signature scoring

To quantify the activation of specific functional programs within NK cells, we employed AUCell (1.24.0) which applies a rank-based scoring method to calculate the level of enrichment of predefined gene sets in single cells (26). Signature gene sets were curated from literature and are outlined in Table S3. These included programs related to cytotoxicity, immunomodulation, stress response, and expression of activating and inhibitory receptors. Additional phenotypic reference signatures were derived from a previous study of NK cell subsets in human blood. Custom IFNG and TGFB1 signature modules were derived from transcription factors associated with each cytokine. These scores were then used to stratify NK cells into IFNG^high and TGFB1^high populations for downstream analyses. For bulk profiles, we employed singscore which allows for stable single sample scoring of molecular signatures (27).

Cell-cell interaction analysis

To characterize intercellular communication patterns within the tumor microenvironment, we applied CellChat (v2.1.2) which predicts ligand-receptor interactions based on the co-expression of signaling molecules across cell types (28). These analyses were done for each cancer type and using SingleR-derived cell type annotations and the manually curated NK cell subset annotations as inputs. This approach allowed for the identification of unique interactions across distinct NK cell subsets. Interactions were also stratified by IFNG^high and TGFB1^high NK cell subsets to further dissect incoming and outgoing signaling patterns.

Clinical association analysis

NK cell subset-specific gene signatures were derived from the top 15 DEGs identified for each annotated subset (Table S4). These signatures were subsequently applied to a cohort of 212,105 patient tumor biopsies that underwent comprehensive clinical-grade molecular profiling. Bulk RNA-sequencing was performed in a Clinical Laboratory Improvement Amendments/College of American Pathologists/ISO15189-certified clinical laboratory (Caris Life Sciences). All analyses were conducted retrospectively using de-identified clinical data derived from insurance claim repositories. This retrospective study was conducted under Caris Life Sciences’ Research Data Banking protocol, which was reviewed and granted IRB exemption by the WCG IRB. Therefore, patient consent was not required. The study adhered to the ethical guidelines of the Declaration of Helsinki, the Belmont Report, and the U.S. Common Rule. Patients remained anonymous and sex and age were not analyzed.

Survival analysis was performed by stratifying patients based on the Z-score normalized gene signature for each NK cell subset. Specifically, hazard ratios were calculated by comparing survival outcomes between patients in the top 25% versus bottom 25% of signature scores in each cancer type. Real-world survival was defined as the time from biopsy collection or treatment initiation to either (1) date of death or (2) date of last recorded contact in the insurance claim database. Patients without contact/claims data for a period of at least 100 days were presumed deceased. Conversely, patients with a documented clinical activity within 100 days prior to the latest data update were censored in the analysis. Cox proportional hazards were used to estimate hazard ratios, and log-rank tests were performed to assess statistical significance using lifelines (v0.29.0) (29). A false discovery rate-adjusted q < 0.05 was used as the threshold for significance.

TGF-β and IFN-γ assays

Peripheral blood products from healthy donors were purchased from Memorial Blood Bank (Minneapolis, MN). All studies were approved by the institutional review committee at the University of Minnesota and work with human cells was performed in accordance with the Declaration of Helsinki. Peripheral blood mononuclear cells (PBMCs) were isolated using Ficoll-Paque (GE Healthcare, cat. 17-1440-03) density gradient centrifugation.

For TGF-β LAP assays, PBMCs were stained with CD3 (BioLegend, cat. 317330), TGF-β LAP (BioLegend, cat. 300010), CD14 (BioLegend, cat. 301830), and CD56 (BioLegend, cat. 304628), assessed using flow cytometry (BD LSR II), and analyzed with FlowJo (version 10).

For TGF-β and IFN-γ immunoassays, NK cells were isolated from PBMCs by negative selection using the EasySep Human NK cell Isolation Kit (STEMCELL Technologies, cat. 100-0960). Enriched NK cells were cultured for 4 hours in the presence or absence of 25 ng/ml PMA and 1 M ionomycin (Sigma, cat. P8139) or 10 μg/ml anti-CD16 Ab (3G8; BD Biosciences; cat. 561248). The levels of TGF-β and IFN-γ in supernatant were measured by enzyme-linked immunosorbent assay (ELISA) using the following kits: Human TGF-beta 1 DuoSet ELISA (Bio-Techne, cat. DY240) and Human IFN-gamma DuoSet ELISA (Bio-Techne, cat. DY285B).

Coculture assay between NK cells and T cells

NK cells were isolated from PBMCs as above, alongside autologous T cells by negative selection using the MojoSort^™ Human CD3 T Cell Isolation Kit (BioLegend, cat. 480022). NK cells were plated with 1×10⁵ autologous T cells in a U-bottom 96 well plate at effector to target (E:T) ratios of 0:1, 1:1, and 2:1. Cells were treated with Dynabeads^™ Human T-Activator CD3/CD28 (Thermo Fisher, cat. 11132D) and supplemented with 1ng/mL IL-15 in RPMI. After 3 days, cells were harvested, stained with live/dead Near-IR dye (Thermo Fisher, cat. L34976), CD3 (BD Biosciences, cat. 562280), CD4 (BioLegend, cat. 317436), and CD8 (Biolegend, cat. 344724) and measured using flow cytometry. Samples were spiked with CountBright^™ Absolute Counting Beads (Thermo Fisher, cat. C36950) to enable measurements of absolute counts.

Data and code availability

This study did not generate any new single-cell RNA-seq data and analyzed existing, publicly available data reported in Table S1. The bulk RNA-sequencing data used for clinical associations can be made available upon reasonable request with the permission of Caris Life Sciences by contacting Andrew Elliott (aelliott@carisls.com) and submitting a letter of intent. The corresponding raw sequencing data is owned by Caris Life Sciences and are subject to controlled access for proprietary and privacy reasons. Processed single-cell analyses of NK cells are available as part of an interactive web tool at https://umnnkatlas.shinyapps.io/shinyapp/ and have been deposited to Zenodo alongside the code used for analyses (https://doi.org/10.5281/zenodo.17634053).

RESULTS

Single-Cell Profiling Reveals Expanded Phenotypic Diversity of NK Cells in Solid Tumors

To comprehensively map the diversity of tumor-infiltrating NK cells across cancer, we analyzed previously published scRNA-seq data from 339 samples across 14 solid tumor types, including peripheral blood from 63 cancer patients and 9 healthy donors (Figure 1A; Table S1). In total, 1,259,383 cells were compiled across datasets and subjected to reference-guided transcriptomic annotations to discern cell type. For each cancer type and sample type, we isolated annotated NK cell populations and performed unbiased clustering. Manual inspection of cluster-specific differentially expressed genes (DEGs) was then conducted to identify transcriptomic markers for canonical (CD56^dim and CD56^bright) and non-canonical NK cell populations, including adaptive and tissue-resident phenotypes (Figure 1B, 1C, and S1; Table S3). These steps were performed in each cancer type and sample type to preserve tissue-specific heterogeneity across NK cell populations.

Figure 1. Pan-cancer phenotypic annotations reveal expanded repertoire of NK cell populations in tumors.

(A) Summary of the discovery cohort composition, including the number of NK cells detected across cancer types and tissue sources. Validation cohort overview for associations with clinical outcomes is shown below. (B) Schematic of the computational workflow for NK cell identification and downstream annotation of distinct tumor-infiltrating NK cell populations from scRNA-seq data. An example for the manual inspection of DEGs to identify transcriptomic markers for NK cell phenotypes followed by annotation of subsets is shown. (C) Dot plot displaying expression of established NK cell phenotypic markers across the six tumor-infiltrating NK cell subsets identified (left). Heatmap shows alignment of these subset annotations with consensus definitions of NK cell populations in human blood (right).

Notably, we uncovered two unique NK cell subsets that have seldom been described: (1) a TR adaptive NK cell population and (2) a TR subset co-expressing features of both CD56^bright and adaptive NK cells. Our annotations were consistent with well-established marker genes (6,12,30) for CD56^dim (SPON2, FCGR3A, and FGFBP2), CD56^bright (XCL1, XCL2, and TNFRSF18), adaptive (CD8B, SH2D1B downregulation, and CD38 downregulation), and tissue-resident (CXCR6, ITGA1, and ITGAE) phenotypes (Figure 1C). These subsets also aligned with consensus definitions of NK1 (CD56^dim), NK2 (CD56^bright), and NK3 (adaptive) populations found in peripheral blood (Figure 1C) (12). Parallel analyses of peripheral blood NK cells from cancer patients and healthy donors identified conventional circulating subsets, alongside a minor population of NK cells exhibiting both CD56^bright and adaptive features (Figure S1; Table S2). Employing SCENIC (25) to infer gene regulatory networks (Figure S2A), we uncovered transcription factors pivotal to the identity of CD56^dim (KLF2, TBX21, and PRDM1), CD56^bright (RXRA, GATA3, and MYC), and adaptive subsets (BCL11B and BATF) (12,13).

Altogether, we compiled 97,826 NK cells across cancer, revealing substantial heterogeneity in human NK cells that extends beyond classical definitions and challenges the notion of a linear developmental trajectory.

Conventional and Tissue-Resident NK Cells Exhibit Distinct Antitumor Properties

To evaluate the functional implications of diverse NK cell subsets, we applied curated gene signatures related to core properties and actions of NK cells. Functional scoring revealed that canonical CD56^dim NK cells exhibited robust cytolytic potential, marked by high expression of PRF1, GZMH, and GZMB (Figure 2A and 2B). While canonical CD56^bright NK cells exhibited the most robust expression of XCL1 and XCL2, their expression of other chemokines and cytokines was relatively limited (Figure 2B). This was particularly evident in comparison to tissue-resident subsets that converged toward immunomodulatory programs with a diverse array of chemokine and cytokine expression. Though each NK cell subset demonstrated distinct sets of functional programs, adaptive and TR adaptive NK cells notably displayed the lowest stress response scores and inhibitory receptor expression whereas canonical CD56^bright and TR CD56^bright NK cells exhibited the highest (Figure 2A and 2B). These differences point to potential distinctions in functional responsiveness and persistence of NK cells related to their identity.

Figure 2. Conventional and tissue-resident NK cell subsets exhibit distinct antitumor capacities across cancer.

(A) Radar plots illustrating distinct functional profiles of conventional (left) and tissue-resident (right) NK cell subsets. (B) Dot plots showing expression levels of genes associated with cytotoxicity, chemokine and cytokine production, activating and inhibitory receptors, and stress response across subsets. (C) Venn diagram showing overlap of differentially expressed genes (DEGs) between tissue-resident subsets and their conventional counterparts, with core genes highlighted based on recurrence (≥3 of 4 comparisons) and functional relevance. (D) Bar graphs showing the proportion so conventional and tissue-resident NK cells cancer types, ranked by frequency of populations. (E) Dot plots comparing Z-score-normalized cytotoxicity (top) and cytokine and chemokine scores (bottom) across NK cell subsets and cancer types. Dots represent average Z-score ± SEM.

Given the historical focus on conventional NK cell populations found in circulation, we sought to delineate features unique to tissue-resident subsets. Through pairwise comparisons of tissue-resident subsets versus their conventional counterparts, we identified a core set of genes that demarcated these groups (Figure 2C, S2B and S2C). These included well-established genes related to tissue-residency and retention (upregulation of RGS1, ITGA1, and CD69 and downregulation of S1PR5), alongside the upregulation of transcription factors related to NK cell maturation and function (ID2 and ID3), cytotoxicity (GZMA and TNF), and proposed immune checkpoints for T cells (SPRY1 and GPR171). Further examining comparisons between TR CD56^bright/adaptive NK cells and their conventional counterparts, we find that they upregulate TMIGD2, a co-stimulatory immune receptor, and CXXC5, a transcription factor positioned at the interface of memory formation in NK cells (31). Additionally, we found differences in the expression of cytokine, chemokine, and stress response genes across subsets, suggesting a unique functional profile for tissue-resident NK cells that has yet to be fully appreciated.

NK Cell Phenotypic and Functional Diversity is Dictated by Cancer Type

Considering the phenotypic annotations of tumor-infiltrating NK cells were performed independently for each cancer type, we next explored how NK cell compositions and functions differed across settings. We found strikingly different proportions of tissue-resident and conventional NK cell subsets between tumors (Figure 2D). For instance, ovarian cancers (OV) predominantly harbored TR CD56^bright NK cells whereas neuroblastoma (NBL) were devoid of any tissue-resident population. Furthermore, while we detected canonical CD56^dim NK cells across all 14 cancer types studied, the representation of the other five NK cell subsets, particularly TR adaptive and TR CD56^bright/adaptive subsets, was highly variable. Nonetheless, we were able to sufficiently parse out distinct subsets within the same cancer type, such as all three tissue-resident subsets in prostate cancer (PRAD). When examining peripheral blood NK cell populations, we were able to identify a minor fraction of CD56^bright and adaptive NK cells displaying prominent tissue-residency scores over circulatory scores (Figure S2D). These potentially tissue-experienced NK cells were more prominent in the peripheral blood of healthy donors than in cancer patients, suggesting that NK cells are likely to be retained in tumors than in healthy tissues.

To assess how NK cell function is influenced by context, we first compared their functional profiles across three compartments: healthy donor peripheral blood, cancer patient peripheral blood, and tumor tissue (Figure S2E). Across subsets, we observed consistent upregulation of chemokine and cytokine expression, alongside stress response genes, after tumor infiltration. In contrast, conventional subsets exhibited lower cytotoxic potential in tumors, indicative of suppressed effector functions in the tumor microenvironment. Comparing these properties across cancer types, we found that canonical CD56^dim NK cells generally exhibited the most robust expression of cytotoxic molecules, with the highest activity in head and neck cancers (HNSC; Figure 2E). However, TR CD56^bright NK cells in HNSC displayed even higher cytotoxic potential, concordant with prior reports of their potent antitumor activity in this setting (32). Of note, when examining the expression of chemokines and cytokines, canonical CD56^bright NK cells exhibited relatively low scores, suggesting drastic functional impairment compared to other subsets. In contrast, most tissue-resident subsets maintained robust immunomodulatory activity, reinforcing a role for tissue-resident NK cells in mediating local immune responses.

Adaptive NK Cells Exhibit T Cell-Like Exhaustion Programs Associated with BATF

Given the context-dependent shifts in NK cell functionality observed across tumors, we next investigated whether stress or exhaustion might further stratify NK cell subsets and impact their antitumor potential. Previous studies have linked NK cell dysfunction to poor clinical outcomes, yet it remains unclear whether such programs are subset-specific or driven predominantly by tumor context. To address this, we first examined the expression of stress-related genes in NK cells across all subsets and settings. While a proportion of tumor-infiltrating NK cells exhibited stress scores comparable to those in peripheral blood, we identified several cancer types – specifically melanoma (SKCM), bile duct (CHOL), esophageal (ESCC), and oral cancers (OSCC) – where NK cells displayed markedly elevated stress-response signatures (Figure 3A). These patterns appeared largely driven by tumor context rather than NK cell identity, suggesting a prevailing nature of immunosuppressive tumor microenvironments. This notion was further supported by the transcription factor regulons of all subsets across cancer types (Figure S3A).

Figure 3. NK cell stress and exhaustion profiles associate with distinct phenotypes and functional patterns.

(A) Box plots comparing stress response scores between NK cells from peripheral blood and tumors. Low-stress and high-stress cancer types are highlighted. Boxes represent median and interquartile range. Whiskers represent 1.5x interquartile range. (B) UMAPs illustrating the distributions of stress response, cytotoxicity, chemokine/cytokine expression, and IFNG levels in NK cells from low-stress (top) and high-stress (bottom) tumor types. Dot plots for associated genes are shown on the right. (C) Ranked gene plot showing transcriptomic associations with IFNG expression across cancers, based on average Spearman correlation values. (D) Heatmaps displaying pairwise Spearman correlations between signatures within adaptive NK cell subsets. (E) Scatter plots showing the relationship between stress and exhaustion scores across NK cell subsets by cancer type, with high-scoring populations highlighted. (F) Spearman correlations between NK cell and CD8⁺ T cell stress scores within and across cancer types. (G) Pan-cancer Spearman correlations between exhaustion scores of adaptive and all NK cells with CD8⁺ T cells. (H) Correlations between exhaustion signatures and expression of BATF, TOX, and TOX2 in adaptive (left) and tissue-resident adaptive (right) NK cells. (I) Heatmap of CXCL13:CXCR5 interaction probabilities across immune populations in esophageal cancer. *p ≤ 0.05; **p ≤ 0.01; ***p ≤ 0.001; ****p ≤ 0.0001.

Despite these elevated stress profiles, NK cells in high-stress tumors retained expression of cytotoxic and immunomodulatory genes (Figure 3B). When comparing NK cell subsets from high- and low-stress cancer types, we found that stress was associated with a conserved upregulation of genes associated with IFN-γ signaling and regulation such as ZC3H12A, NFKBIZ, and AHR (Figure S3B). Indeed, NK cells in high-stress cancer types did not show a deficit in IFNG expression but rather an upregulation (Figure 3B). In line with this, several genes positively correlated with IFNG expression included DNAJB1, HSPA1B, PPP1R15A, and CD69 (Figure 3C). These patterns likely reflect an activated NK cell state, wherein high-stress environments promote an IFNG^high functionally poised program. Conversely, TGFB1 expression was negatively correlated with IFNG, alongside the glucocorticoid-induced tumor necrosis factor receptor (GITR; TNFRSF18) and the interleukin-7 receptor (IL7R).

To better understand the relationships between stress, exhaustion, and functionality, we performed subset specific correlations that revealed divergent trends (Figure 3D and S3C). For instance, in canonical CD56^dim, TR CD56^bright/adaptive, and TR adaptive NK cells, elevated stress was associated with reduced cytotoxic capacity. In contrast, adaptive NK cells demonstrated a paradoxical increase in cytotoxic potential with regard to stress and further showed strong correlations between CD8⁺ T cell-derived exhaustion scores and other functional properties. Altogether, these point to potentially divergent mechanisms driving function or dysfunction in each subset, particularly between adaptive NK cells and their tissue-resident counterpart.

Given the parallels between adaptive NK cells and CD8⁺ T cells (33), we investigated whether exhaustion profiles were conserved between these lineages. Among all NK cell subsets, only adaptive and TR adaptive NK cells in specific tumor types, namely ESCC and OSCC, exhibited transcriptional profiles consistent with T cell-like exhaustion (Figure 3E). These patterns were consistent with our prior comparisons between NK cells in high- and low-stress cancer types where we observed adaptive-specific upregulation of PDCD1, CTLA4, TIGIT, ENTPD1, and CXCL13 (Figure S3B). To determine whether NK cell stress and exhaustion reflect a shared program with CD8⁺ T cells, we compared the transcriptional states between patient and cancer-type matched NK cells and CD8⁺ T cells. We observed strong correlations in stress states between NK cells and CD8⁺ T cells across all tumors, indicating a shared response to environmental stressors (Figure 3F). However, exhaustion profiles were significantly correlated in only adaptive and TR adaptive NK cells (Figure 3G), suggesting that these subsets uniquely recapitulate CD8⁺ T cell features.

To investigate transcriptional regulators of exhaustion in adaptive and TR adaptive NK cells, we examined transcription factors well-established to drive T cell exhaustion. In doing so, BATF emerged as the dominant factor in both adaptive and TR adaptive NK cells, whereas TOX and TOX2 were more prominent in the conventional adaptive subset (Figure 3H and S3D). These suggest divergent paths to exhaustion between these two subsets, reminiscent of naïve CD8⁺ T cells going through early memory or tissue-resident memory trajectories to terminally exhausted states.(34) Looking at the functional consequences of this exhaustion program in adaptive NK cells, we found that adaptive NK cells in ESCC upregulated CXCL13, a chemokine implicated in the recruitment of CXCR5⁺ B cells and CD4⁺ T cells. Notably, these interactions were predicted to be stronger than that of CD8⁺ T cells (Figure 3I), raising the possibility that exhausted adaptive NK cells may contribute to the formation of tertiary lymphoid structures which are linked to favorable responses to immune checkpoint blockade (ICB) (35,36).

An Immunosuppressive Circuit Dichotomizes NK Cell Populations Across Cancer

The inverse correlations observed between TGFB1 and IFNG expression in NK cells prompted us to investigate whether these cytokines define discrete functional states. As TGF-β and IFN-γ orchestrate divergent functions in antitumor immunity, we hypothesized that such an axis may underlie the purported immunoregulatory roles of NK cells. To test this, we stratified tumor-infiltrating NK cells by TGFB1 and IFNG expression and analyzed their transcriptomic profiles. In high-stress cancer types, we found a striking dichotomy between TGFB1 and IFNG expression (Figure 4A) that was further reinforced by paired single-cell expression values which revealed only minor fractions of NK cells co-expressing both cytokines within tumors, cancer patient peripheral blood, and healthy donor peripheral blood (Figure 4B). We found that the vast majority of NK cells in circulation predominantly expressed TGFB1 and lacked IFNG transcripts, shifting towards IFNG expression only in tumors. Stratifying NK cells by expression profiles, we found that IFNG^high NK cells upregulated markers of activation and stress (CD69 and BAG3), in line with our previous results, whereas TGFB1^high NK cells were enriched for members of the TGF-β signaling pathway (SMAD2, SMAD4, and TGFBR2; Figure S4A). Indeed, when further stratifying NK cells by subset and setting, we found concurrent upregulation of these TGF-β signaling pathway genes alongside TGFB1 expression (Figure 4C). Additionally, expression of these genes was more prominent in the peripheral blood of cancer patients compared to healthy donors, suggestive of active immunosuppression in circulation.

Figure 4. NK cells are dominant sources of TGFB1 in circulation and tumors.

(A) UMAPs displaying expression of TGFB1 and IFNG in NK cells from high-stress cancer types. (B) Contour and density plots illustrating the prevalence of TGFB1- and IFNG-expressing populations in tumors, cancer patient peripheral blood, and healthy donor peripheral blood. (C) Dot plots showing the expression of TGF-β pathway genes across NK cell subsets and settings. HD, healthy donor peripheral blood. CP, cancer patient peripheral blood. (D) Lollipop plot depicting the expression of TGFB1 in NK cell subsets and selected cell populations across cancer types.

Given the striking expression of TGFB1 in NK cells, we next aimed to examine how the expression of TGFB1 in NK cells compares to that of other cell populations implicated in TGF-β-mediated immunosuppression. Unexpectedly, NK cells – particularly canonical CD56^dim, canonical CD56^bright, and TR CD56^bright NK cells – emerged as the dominant source of TGFB1 expression in the tumor microenvironment, surpassing macrophages and monocytes in most tumor settings (Figure 4D). Moreover, the expression of TGFB1 in canonical CD56^dim and canonical CD56^bright NK cells strongly correlated to that in macrophages and monocytes (Figure S4B), suggestive of a positive feedback loop that may act to reinforce immunosuppressive signals. When measuring the levels of the TGF-β latency-associated peptide (LAP) in peripheral blood NK cells, we found that a minor fraction expressed LAP. However, LAP⁺ NK cells constituted a distinct subset of circulating NK cells with higher per-cell expression levels relative to monocytes (Figure S4C).

Distinct bZIP transcription factors underlie the TGFB1–IFNG functional switch in NK cells

To understand how such dichotomous expression patterns may be regulated in NK cells, we first employed SCENIC on TGFB1^high and IFNG^high NK cell populations. We identified transcription factor regulons associated with each population, including BATF, MYC, and EOMES regulons in IFNG^high NK cells and MAFF, XBP1, and TBX21 in TGFB1^high populations (Figure 5A), reminiscent of the DEG profiles we observed prior (Figure S4A). Examining transcription factors more granularly, we found that distinct sets of bZIP transcription factors closely associated with the expression of IFNG (JUN, KLF6, and ATF3) or TGFB1 (XBP1, MAFF, CEBPD, and FOSL2) (Figure 5B and S5A). Given that our initial SCENIC inferences linked these signals to transcription factors related to NK cell maturation and identity, we next sought to determine how this TGFB1–IFNG axis maps onto the distinct NK cell subsets we identified. Integrating transcription factor programs into gene signatures for TGFB1 or IFNG expression, we observed a striking separation between canonical and non-canonical subsets in their propensity towards TGFB1 versus IFNG signaling programs. Canonical CD56^bright NK cells were biased towards TGFB1 signaling whereas TR adaptive NK cells, on the other hand, were predominantly shifted towards IFNG programs (Figure 5C and S5B). Within these patterns, however, emerged intermediate populations of NK cells within each subset that strongly exhibited transcriptional programs for both cytokines, suggesting that these sets of bZIP transcription factors may work to coordinate a switch between TGFB1 and IFNG expression.

Figure 5. Distinct transcription factor circuits drive TGFB1 and IFNG signaling patterns in NK cells.

(A) Heatmap featuring predicted transcription factor regulons between TGFB1^high and IFNG^high populations. (B) Scatter plots displaying Spearman correlations between TGFB1 and IFNG expression with putative transcription factors. (C) Contour and density plots illustrating the prevalence of TGFB1 and IFNG signaling networks across NK cell subsets, with quadrants defining populations with distinct signaling patterns. (D) Stacked bar plot displaying the proportions of TGFB1-signature^high, IFNG-signature^high, double-positive, and double-negative canonical CD56^dim (top) and CD56^bright (bottom) NK cells across cancer types. (E) Chord diagram showing the prevalence of TGFB1:TGFBR2 interactions from NK cell subsets to CD8⁺ T cells. Segments indicate proportion of cancer types with significant interactions for that NK cell subset. (F) Dot plots showing incoming and outgoing TGFB1:TGFBR2 interactions between monocytes and macrophages to NK cells and NK cells to CD8⁺ and CD4⁺ T cells, respectively. Myeloid cells were selected as senders due to their well-established roles in immunosuppression, whereas T cells receivers due to their roles in tumor control. Canonical NK cell subsets were divided into IFNG-signature^high and TGFB1-signature^high populations based on their signaling patterns.

We then evaluated the scope of TGFB1 and IFNG transcriptional programs across tumors, finding an enrichment of TGFB1-signature^high canonical CD56^bright and CD56^dim NK cells in pancreatic (PAAD), bone (OS), and adrenal (NBL) tumors. Conversely, these populations were scarce in kidney (KIRC), skin (SKCM), and oral (OSCC) tumors (Figure 5D), patterns reminiscent of stress response signatures. Indeed, the proportion of TGFB1-signature^high NK cells was inversely correlated with stress response signatures whereas the proportion of IFNG-signature^high populations was positively correlated (Figure S5C). The proportion of TGFB1-signature^high NK cells was also related to the expression of genes related to resistance to NK cell cytotoxicity in the epithelial compartment of tumors (Figure S5D). This pattern was largely driven by the expression of P-selectin ligand (SELPLG) and mucin 1 (MUC1) which were highest in PAAD (Figure S5E). Re-examining these distinct cytokine states in peripheral blood, we find that circulating CD56^bright and CD56^dim NK cells are predominantly TGFB1-signature^high in healthy donors, with IFNG-signature^high populations expanding in the periphery of cancer patients and particularly in tumors (Figure S5F), indicative of major reprogramming of NK cells upon tumor entry.

Exploring the functional consequences of these distinct cytokine programs on the broader tumor microenvironment, we conducted cell-cell communication inferences using CellChat (28). We observed that canonical CD56^dim and CD56^bright NK cells represented robust sources of outgoing TGFB1 signals towards CD8⁺ T cells (Figure 5E). Stratifying these subsets further into TGFB1-signature^high and IFNG-signature^high fractions, we found that the former were not only more likely to direct immunosuppressive signals but were also more likely to receive TGFB1 signals from macrophages and monocytes, with these interactions being the most prevalent in ESCC (Figure 5F). In contrast, TR adaptive NK cells exhibited minimal participation in TGF-β networks, reinforcing their independence from the intrinsic suppressive program that is prevalent in canonical NK cells. Notably, co-culture assays between NK cells and T cells derived from the peripheral blood of healthy donors demonstrated that NK cells suppressed the proliferation of autologous CD4⁺ T cells (Figure S5G).

To experimentially test the inducibility of TGF-β and IFN-γ programs in NK cells, we isolated peripheral blood NK cells from healthy donors and quantified cytokine production under basal and stimulated conditions. NK cells secreted robust levels of TGF-β that were unaffected after stimulation with CD16 crosslinking or treatment with phorbol 12-myristate 13-acetate (PMA) and ionomycin (Figure S5H). In contrast, IFN-γ was significantly upregulated after PMA and ionomycin treatment, suggesting that this balance can be skewed through certain activation cues. These data support a model wherein the TGFB1–IFNG axis functions as a tunable switch that can dynamically respond to suppressive and activating cues within the microenvironment.

TR Adaptive NK Cells are Associated with Prolonged Survival Across Cancer Types

While the general abundance of NK cells in tumors has been linked to favorable outcomes in cancer, we have yet to fully understand the clinical impact of distinct NK cell phenotypes. Thus, to investigate the prognostic relevance of each NK cell subset, we derived robust transcriptional signatures for each of the six NK cell subsets based on their top DEGs (Table S4) and projected these signatures onto a comprehensive pan-cancer cohort comprised of 212,105 bulk RNA-sequencing samples from 53 cancer types (Figure 6A). These signatures correlated with the relative frequencies of each subset seen in our single-cell data, supporting their use in deconvoluting NK cell compositions from bulk profiles (Figure S6A).

Figure 6: NK cell identity and the TGFB1-IFNG molecular switch drive divergent outcomes in cancer.

(A) Schematic summarizing clinical cohorts used for survival associations. (B) Dot plot of hazard ratios for each NK cell subset across all cancers and ICB-treated cancers. Hazard ratios were calculated by comparing real-world overall survival in patient cohorts defined by high (top 25%) versus low (bottom 25%) NK cell subset signature score. (C) Scatter plots correlating NK cell subset signature scores and hazard ratios in primary (left) and metastatic tumor samples (right). Cancer types with unfavorable NK cell associations are highlighted. Adjacent heatmap displaying subset-specific Spearman correlations in each setting. PT, primary tumors. MT, metastatic tumors. *p ≤ 0.05; **p ≤ 0.01; ****p ≤ 0.0001. (D) Scatter plot showing subset-specific hazard ratios across cancer types derived from the clinical association bulk RNA-seq cohort (left) alongside corresponding proportions of TGFB1^high cells derived from the single-cell atlas studies (right). (E) Scatter plots correlating the frequency of TGFB1^high and IFNG^high NK cells (scRNA-seq-derived) with hazard ratios across cancer types (bulk RNA-seq-derived).

Strikingly, we found that the abundance of TR adaptive NK cells in tumors was strongly associated with improved overall survival across cancer types (hazard ratio [HR] 0.68, 95% confidence interval [CI] 0.66–0.69, q<0.0001; Figure 6B). This prognostic effect was more pronounced than any other subset, particularly canonical CD56^bright NK cells which only exhibited a modest association with prolonged survival (HR 0.94, 95% CI 0.93–0.96, q<0.0001). These findings suggest that not all NK cells contribute equally to antitumor immunity and may point to subset-specific functional states dictating their clinical relevance.

We further examined whether these associations hold true in the context of ICB and found that the prognostic separation between subsets narrowed (Figure 6B), suggesting that ICB treatment may broadly influence the NK cell compartment in tumors. Nonetheless, TR adaptive NK cells remained the most favorable subset. To validate these observations, we analyzed two independent datasets with pre- and post-treatment biopsies: a scRNA-seq cohort of OSCC (37) and a bulk RNA-seq cohort of SKCM (38) (Figure S6B and S6C). In OSCC, we observed a marked expansion of the TR adaptive NK cell compartment in responders, alongside a contraction of the canonical CD56^bright population (Figure S6B and S6D). Similar patterns were found in SKCM where we found significant increases in all subset signatures in responders that were not observed in non-responders (Figure S6E). Notably, this enrichment was restricted to patients naïve to prior ICB, highlighting the potential effects of treatment history in modulating NK cell dynamics.

Though NK cells are well recognized for their role in controlling metastasis, their contributions within primary tumors remain controversial (39). To explore how distinct NK cell subsets may contribute to immune surveillance at different stages of disease, we stratified our pan-cancer bulk RNA-seq cohort by biopsy site, comparing primary versus metastatic tumors from cancer types included in our single-cell analyses (Figure 6A and 6C). From these stratifications, we found that, in primary tumors, only the signatures for canonical CD56^dim and TR adaptive NK cells were associated with more favorable hazard ratios across cancer types. In metastatic tumors, on the other hand, the associations shifted exclusively toward adaptive and tissue-resident subsets, suggesting their potential roles in controlling disseminated disease (Figure 6C).

Importantly, TR adaptive NK cells were the only subset that demonstrated consistent, robust associations with improved survival across primary and metastatic settings and in response to ICB treatment. Meanwhile, canonical CD56^bright NK cells remained only weakly prognostic across settings, aligning with their potentially dysfunctional profiles we observed in our earlier analyses. Intriguingly, in a minority of tumor types, such as metastatic liver cancers (LIHC) and primary thyroid cancers (THCA), higher NK cell abundance correlated with worse survival (Figure 6C). These findings reinforce the need for subset- and context-resolved investigations to fully comprehend the contributions of NK cells to antitumor immune responses.

TGFB1–IFNG Balance Underlies the Prognostic Value of Tumor-Infiltrating NK Cells

Having established the clinical relevance of each NK cell subset, we next investigated whether the intrinsic programs governing NK cell functionality might further explain the differential associations with patient outcomes. Given our prior observations of a dichotomous TGFB1–IFNG functional switch demarcating suppressive and inflammatory NK cell states, we hypothesized the relative balance of these pathways may shape the specific prognostic value of NK cell subsets across cancer.

To test this, we linked the hazard ratios derived from our pan-cancer bulk RNA-seq analysis with TGFB1 and IFNG transcriptional signature scores derived from single-cell profiles of tumor-infiltrating NK cell populations. Specifically, we calculated the proportion of TGFB1-signature^high and IFNG-signature^high NK cells in each subset across cancer types and correlated those frequencies with the corresponding hazard ratios from the bulk RNA-seq cohort (Figure 6D and 6E). This approach revealed strong associations between the balance of these programs with the frequency of TGFB1-signature^high NK cells correlating with worse overall survival (r = 0.59517, p = 0.00005) and IFNG-signature^high frequencies with prolonged survival (r = −0.42533, p = 0.00533) (Figure 6E). These associations were particularly notable in cancer types where NK cell activity is blunted. For instance, pancreatic tumors (PAAD) that were previously found to upregulate markers of resistance to NK cell cytotoxicity (Figure S5D), were among the cancer types in which NK cell abundance was least prognostic (Figure 6D). To further probe whether therapeutic intervention could modulate these functional axes, we investigated the changes in TGFB1 and IFNG signatures following ICB treatment. In OSCC, we observed a marked upregulation of IFNG signatures within the TR adaptive NK cell compartment of responders after ICB, but not in non-responders (Figure S6F). A similar pattern emerged in SKCM, where IFNG signatures increased significantly in responders to ICB treatment, specifically in those without prior exposure (Figure S6G). Taken together, these findings support the notion that the TGFB1–IFNG axis represents a microenvironmentally-tuned balance that governs the downstream activity of NK cells within tumors.

DISCUSSION

Our comprehensive characterization of NK cell heterogeneity across human cancers reveals a previously unrecognized functional switch that challenges the traditional view of NK cells in promoting antitumor immunity. Here, we identify a striking dichotomy between TGFB1 and IFNG expression that not only delineates distinct NK cell subsets but is also reflective of a dynamic, tunable regulatory switch. We find that this switch is associated with distinct bZIP and AP-1 transcription factor circuits that may serve to integrate microenvironmental signals and shift NK cells between a homeostatic, regulatory state and an activation-induced, pro-inflammatory state. Our findings suggest that manipulating this axis may offer an avenue for reprogramming both NK cell functionality and the immunosuppressive tumor microenvironment.

Importantly, our work builds upon prior studies characterizing the landscape of NK cells in cancer by identifying non-conventional phenotypic and functional states that differentially contribute to survival outcomes in patients. Our more nuanced, manual annotations of NK cell populations unraveled an expanded repertoire of subsets in tumors that span multiple differentiation and maturation states. We further examined the biological and clinical consequences of distinct adaptive subsets, namely identifying a tissue-resident adaptive subset exhibiting exhaustion patterns reminiscent of CD8⁺ T cells. We identified the specific expansion of this subset in responders to ICB and report on their prognostic value across solid tumors. In addition, we were able to discover discrete populations of tissue-infiltrating and tissue-resident NK cells that exhibited distinct functional properties within the same tumor tissues.

The biology of tissue-resident NK cells has only recently come into focus, revealing substantial functional diversity driven by tissue-specific cues. For example, liver-resident NK cells can exert both antiviral and regulatory functions (40) whereas decidual NK cells act to support pregnancy by guiding vascular remodeling and trophoblast differentiation (41). While prior studies have predominantly described tissue-resident NK cells as CD56^bright, our data expand this paradigm by identifying three transcriptionally distinct tissue-resident subsets, including those with adaptive features. Moreover, our identification of adaptive NK cells with transcriptomic profiles resembling CD56^bright phenotypes challenges the traditional linear model of NK cell differentiation that currently prevails (42). Notably, this aligns with contemporary studies demonstrating that CD56^bright NK cells are capable of adopting both tissue-resident and memory-like programs in settings of infection and in cancer (43–45). Taken together with lineage tracing studies demonstrating the acquisition of NKG2C expression in naïve NK cells (30), our findings suggest a previously underappreciated degree of plasticity in NK cell development, wherein distinct differentiation programs can converge across multiple maturation states. One may posit that strenuous inflammatory conditions throughout the human lifespan, including infection and cancer, may skew NK cells towards different developmental paths depending on microenvironmental cues and intrinsic molecular circuits that differ between subsets. Additionally, this expanded paradigm may reflect natural diversification of the NK cell pool over time that relies on multiple inflammatory events rather than a singular disease state, though human cytomegalovirus infection representing the primary driver of adaptive NK cell formation argues against this notion. Indeed, further work is warranted to understand the full breadth of NK cell diversity and differentiation across physiological and disease states.

Further expanding upon this paradigm, emerging evidence has shown that NK cells are capable of suppressing CD8⁺ T cell responses in infection and cancer (16–18,20,21). This immunoregulatory function has been linked to the secretion of suppressive cytokines, such as IL-6 (18), IL-10 (46), and IL-22 (17), as well as the direct elimination of activated T cells. NK cells have also been implicated in promoting the development of myeloid-derived suppressor cells (MDSCs), thereby contributing to further immunosuppression (18). However, the role of NK cells as robust producers of TGF-β has received relatively little attention. The earliest report of TGF-β secretion by NK cells dates back to 1998, when CD2 ligation was demonstrated to induce TGF-β production (47). This observation remained largely unexplored until more recent studies implicating TGF-β in the establishment of tissue-resident NK cell populations (48) and TGF-β⁺ NK cells having protective roles in acute graft-versus-host disease (GVHD) (49).

While exogenous TGF-β is well-established to impair NK cell cytotoxicity (50), paradoxical findings have emerged showing that TGF-β imprinting during NK cell activation can enhance antitumor functions (51). Specifically, TGF-β exposure has been shown to generate NK cells with elevated IFN-γ and TNF-α secretion and reduced sensitivity to downstream TGF-β-mediated suppression (51). Our data extend these observations by demonstrating that NK cells can endogenously produce TGF-β and, in doing so, feed an immune suppressive circuit – sending TGF-β signals to neighboring immune cells while also receiving reinforcing input from other suppressive populations. These findings underscore the context-dependent effects of TGF-β in NK cells and highlight the need for further investigation into how intrinsic and extrinsic signals interplay to shape NK cell fate and function.

Importantly, we find that the TGFB1–IFNG switch is not merely a functional feature but is also tightly aligned with NK cell identity. Canonical CD56^bright and CD56^dim NK cells exhibited strong TGFB1 signatures whereas adaptive subsets were marked by robust IFNG profiles. This was particularly evident in TR adaptive NK cells that were associated with favorable clinical outcomes across cancer types and tumor settings. Intriguingly, TR adaptive NK cells also displayed an exhaustion-like signature reminiscent of CD8⁺ T cells and associated with BATF transcriptional programs. These functional parallels likely underlie the preferential expansion of TR adaptive NK cells after ICB treatment. Nonetheless, the molecular mechanisms governing memory-like formation and long-term persistence in NK cells remain incompletely understood. While adaptive NK cells have traditionally been studied in the context of viral infection, their enrichment in tumors and ability to adopt tissue-resident phenotypes point to broader roles in tumor immunosurveillance.

Finally, our data suggest that the TGFB1–IFNG functional switch is shaped by both homeostatic cues and stress- and activation-induced signals. Stress and activation, likely modulated by tumor-derived factors appear to drive NK cells towards IFNG signaling states that promote antitumor immunity. While Tang et al. linked NK cell stress signatures to dysfunction and poor clinical outcomes (14), our data suggest a more nuanced view wherein a certain threshold of stress may be required to appropriately engage and activate NK cells. Our findings are reminiscent of Netskar et al. whose findings demonstrated that stressed CD56^bright NK cells are not consistently linked with poorer survival outcomes. Additionally, they detected the expression of stress response markers in activated CD56^dim NK cells, reflecting different thresholds for stress and dysfunction across subsets. In this context, the observed dysfunction of canonical CD56^bright NK cells may reflect exceedingly overt stress signals, whereas other subsets may be better equipped to integrate stress signals and mount productive immune responses. This more refined view is supported by our subset-resolved analyses which extend beyond traditional classifications. Importantly, by linking these subsets to over 200,000 tumor samples, we were able to validate the clinical relevance of our findings at an unprecedented scale.

While we have systematically characterized the phenotypic and functional landscape of NK cell diversity in tumors, several limitations should be noted. Our study predominantly utilizes single-cell transcriptomic data from publicly available datasets of primary, therapy-naïve solid tumors. Correlates across different clinical settings, particularly in heavily pre-treated, metastatic cancers, were unexplored. Additionally, the functional validation of distinct natural killer subsets and their pro-inflammatory versus suppressive properties remains limited, namely by the sensitivity and scarcity of NK cells within tumor tissues. Importantly, our annotations of NK cell populations for each tumor type are limited by the quality of data in each setting, with the scarcity of NK cells in some minimizing the detection of more diverse subsets. Nonetheless, our data aligns with contemporary studies redefining the landscape of NK cell diversity in tissues and highlighting their dualistic roles in promoting versus suppressing immune responses.

Collectively, our study builds upon a growing body of work redefining the role of NK cells in cancer. By uncovering an intrinsic the TGFB1–IFNG functional switch that shapes opposing NK cell functions, we reveal a dynamic and therapeutically actionable axis with broad implications. This intrinsic balance between TGF-β and IFN-γ signaling not only shapes NK cell behavior in cancer but may also operate in other disease settings, including autoimmunity and infection, where NK cells are increasingly recognized as key immune effectors. We envisage our findings will provide a conceptual and translational framework for harnessing NK cell plasticity, ultimately guiding the development of next-generation NK cell-based immunotherapies.

SYNOPSIS.

While natural killer cells are important for antitumor immunity, their heterogeneity is poorly understood. The authors identified natural killer cell populations that differentially contribute to survival outcomes and exhibit distinct functional properties that can be exploited for therapy.