CD103⁺CD56⁺ ILCs Are Associated with an Altered CD8⁺ T-cell Profile within the Tumor Microenvironment

Immune regulation is composed of a complex multilayered network of cellular and molecular pathways that regulate the immune system to prevent tissue damage and autoimmunity. Innate lymphoid cells (ILC) are a heterogeneous population of cells that primarily reside in peripheral tissues and are capable of rapidly orchestrating innate and adaptive immunity in response to pathogenic and physiologic signals (1). The ILC family includes NK cells, ILC1s, ILC2s, ILC3s, and lymphoid tissue inducer cells (1). ILC1s are defined by the expression of CD127, a lack of surface markers that define other immune subsets, and the absence of CRTH2 or cKit expression (2, 3)_. However, distinguishing ILC1s from NK cells has been challenging as various lineage-defining markers can vary depending on the tissue microenvironment and activation status (4). Moreover, ILC1s have been shown to express surface markers similar to NK cells (e.g., C-X-C chemokine receptor type 6; ref. 5). ILC2s secrete type 2 cytokines including IL-4, IL-5, IL-9, IL-13, and amphiregulin, express the transcription factor (TF) GATA3, and play a role in clearing parasite infections (1). ILC3s secrete IL-17, IL-22, and GM-CSF and contribute to intestinal homeostasis and response to bacterial infections (1).

Human NK cells in the periphery have been historically subdivided based on the level of expression of CD56 and CD16 (6). CD56^dimCD16^+/− cells were initially described as potent cytolytic effector cells that express high levels of IFN-γ and cytolytic molecules (including perforins and granzymes), whereas CD56^brightCD16⁻ cells have reduced lytic abilities and increased expression of cytokines. However, recent studies have found that CD56 brightness does not necessarily correlate with NK cell function (7, 8), challenging the current nomenclature of NK cells. For example, NK cells treated with either IL-15 (7, 9) or IL-12 in the presence of the OVCAR carcinoma cell line (10), or IL-2 and membrane-bound IL-21 on feeder cells (8), led to the induction of CD56^bright NK cells with enhanced degranulation and cytotoxicity in vitro. Moreover, IL-21–induced CD56^bright NK cells had an increased ability to clear autologous tumors in a patient-derived xenograft murine model. CD56^brightCD16⁺ NK cells derived from melanoma metastasis in the lymph node that were activated with IL-15 had enhanced cytotoxic function (11). Although several studies have characterized CD56^bright cells in human cancers (12, 13), the functional role of these cells within the tumor microenvironment (TME) is unclear.

Beyond the conventional roles of ILCs, there is growing evidence that they also have potent immunosuppressive roles with similar mechanisms that mirror regulatory T cells (Tregs; ref. 14), adding to the complexity of immune regulatory networks. Notably, NK-like regulatory cells have been shown to directly suppress T cells in colitis (15), autoimmunity (16–22), GVHD (refs. 23–25), and chronic viral infections (14, 26–30). In addition, ILC2s and ILC3s have also demonstrated diverse mechanisms of suppression in a variety of contexts (14, 30).

ILCs with inhibitory properties have been recently identified in the context of cancer (14, 31–34). For example, NKp46⁻ ILC3s from a mouse model of hepatocellular carcinoma suppressed the proliferation of CD8⁺ T cells and increased apoptosis in vitro (33). Moreover, human NK cells that have been previously cocultured with primary sarcoma tumor cells were able to subsequently suppress CD4⁺ T-cell proliferation and IFN-γ production (34). In our previous work, we identified CD56⁺ immunoregulatory ILCs from in vitro–expanded tumor-infiltrating lymphocyte (TIL) cultures derived from patients with epithelial ovarian carcinoma (EOC; ref. 32). CD56⁺ ILCs could suppress autologous CD4⁺ T-cell and CD8⁺ T-cell expansion in vitro, in part through an NKp46-dependent mechanism. Because these immunoregulatory CD56⁺ ILCs were found in slow growing TIL cultures (supplemented with high-dose IL-2), our next goal was to examine whether we could identify their counterparts in primary EOC tumors. In this study, we found that CD103 is a marker for a distinct population of CD56⁺ ILCs and explored their properties and the corresponding TME from primary EOC tumors.

This research study was conducted in accordance with the guidelines of the Declaration of Helsinki and was approved by the University Health Network (UHN) Institutional Research Ethics Board (10-0335). Informed consent was obtained from all study participants. Surgical specimens and fluids were obtained through UHN Biospecimen Services. The tonsil, spleen, and appendix were retrieved from the Mount Sinai Services Surgical Pathology Department under Research Ethics Board 20-0178-E. Peripheral blood mononuclear cells were retrieved from consented healthy donors through Miltenyi Biotec (150-000-450).

Tumor specimens were retrieved from the UHN Biospecimen Program through surgical resection of 54 patients diagnosed with EOC. Patient characteristics are summarized in Supplementary Table S1. Fresh tumor specimens were enzymatically digested [1 mg/mL collagenase I (catalog no. C5138, Sigma-Aldrich), 10 μg/mL pulmozyme (DIN: 02046733, Hoffmann-La Roche), 2 mmol/L L-glutamine (catalog no. 25030-081, Gibco), and 100 μg/mL amphotericin B (catalog no. 15290-018, Gibco) in RPMI-1640 (catalog no. 11875093, Thermo Fisher Scientific)] using a Gentle MACs Dissociator (Miltenyi Biotec). Tumor digests were cryopreserved in human serum (catalog no. 106-512, GeminiBio) containing 10% DMSO (catalog no. 472301, Sigma-Aldrich) in liquid nitrogen tanks and were thawed prior to analysis.

Cells were washed twice in FACS buffer [PBS, 2% FBS (catalog no. A15-701, PAA Laboratories), and 0.05% sodium azide (catalog no. 7144.8-16, RICCA)] or PBS, followed by staining with fixable viability dye eFluor 780 (catalog no. 65-0865-14, 1:1,000, eBioscience) or eFluor 506 (catalog no. 65-0866-14, 1:800, eBioscience) and FcR blocking (catalog no. 564220, 1:200, BD Biosciences, RRID: AB_2869554) in PBS at 4°C for 20 minutes. The cells were washed and subsequently stained for surface marker antibody cocktails at 4°C for 30 minutes (see Supplementary Table S2 for the list of antibodies used). The cells were washed and fixed with eBioscience Foxp3 Fixation/Permeabilization buffer (catalog no. 00-5523-00, Thermo Fisher Scientific) for 30 minutes and washed with 1× permeabilization buffer. The cells were stained for 60 minutes at 4°C with an intracellular antibody cocktail in 1× permeabilization buffer (see Supplementary Table S2 for the list of antibodies used). Data were acquired on a BD LSRFortessa (BD Biosciences) at the Princess Margaret Flow Cytometry Core Facility using FACSDiva v.9.0.1 (BD Biosciences) and analyzed using FlowJo software v10.8.1 (FlowJo LLC, RRID: SCR008520). Cell sorting was performed on a Beckman MoFlo Astrios or BD FACSAria Fusion (BD Biosciences) by the Princess Margaret Flow Cytometry Core Facility. Intratumoral Tregs were gated as previously described (35). CD56⁺ ILCs were negatively gated for lineage markers and were defined as either CD3⁻TCRαβ⁻TCRγδ⁻CD14⁻CD19⁻CD20⁻CD34⁻CD123⁻CD303⁻FcεRIα⁻ or CD3⁻CD14⁻CD19⁻.

Gene signatures were developed from bulk RNA sequencing (RNA-seq) analysis of CD56⁺ inhibitory ILCs derived from TIL cultures (32) using the top 25 most negative and positive genes determined by log fold change values. Genes that contributed positively to the signature score included DNAH2, COL18A1, HSPA2, COL11A2, RAB26, FAM43A, IER5L, CTSL, TTN, JUNB, JUN, SIGLEC17P, COL9A2, KIAA1683, MST1, RHOB, HIST2H2BD, XKR6, KANK3, FOSB, ADAMTS17, MAFB, SNAI1, DUSP1, and MYO15B. Genes that contributed negatively to the signature score included PARP9, ARHGAP11A, WDR76, OAS1, MMS22L, ASPM, TENM1, MX1, ATAD5, A2M, KIR2DL3, KIR3DL1, KIR3DX1, BRIP1, CHMP4C, KIAA1324L, ZNF280B, IFI44L, KIR2DL1, CDKN3, NOSTRIN, NUTM2F, DNAH11, CHRNA7, and ISY1-RAB43. The genes contributing positively and negatively were assigned a weight of +1 and −1, respectively. The full list of genes and their weights is provided in Supplementary Table S1.

Bulk RNA-seq datasets of tumors from patients with high-grade serous ovarian carcinoma (HGSC) were curated from GSE9891, GSE17260, GSE26193, GSE30161, GSE49997, and TCGAOVARIAN (36–41). These datasets were chosen because they included at least 50% of the genes in our signature. Patients within these datasets that were in the top 50th percentile in expression of CD56 were included. Patients were scored based on the expression of the CD56⁺ inhibitory ILC gene signature and were split into two groups (i.e., high vs. low using the mean of signature scores to determine the 50th percentile threshold) for downstream survival analysis using recurrence-free survival data (RFS). Data were censored at 10 years. A random effects model was used to account for variability across different dataset, and D index was used as the statistical measure to assess the prognostic ability of our gene signatures.

RNA was extracted from sorted CD103⁺CD56⁺Lin⁻ ILCs (a total of 250,900 cells) and CD103⁻CD56⁺Lin⁻ ILCs (a total of 282,800 cells) matched from four patient samples using RNeasy Mini Kit (catalog no. 74104, Qiagen) according to the manufacturer’s instructions. After quantity and quality assessments using Agilent Bioanalyzer RNA 6000 Nano Kit (catalog no. 5067-1511, Agilent) and KAPA SYBR FAST qPCR kit (catalog no. KK44602, Roche), RNA libraries were prepared, and rRNA were removed using SMARTer Stranded Total RNA-seq Kit V3 – Pico Input Mammalian (catalog no. 634485, Takara Bio). Libraries were normalized, pooled, and loaded onto Illumina NovaSeq V1.5 cartridges for cluster generation, and sequencing was performed on a NovaSeq 6000 instrument (Illumina). Using the paired-end 101 bp protocol, approximately 40 million reads per sample were produced. RNA-seq was performed by the Princess Margaret Genomics Center.

RNA-seq FASTQ files were mapped using STAR aligner v2.5.2b (RRID: SCR_004463; ref. 42) with human transcript coordinates from Human Release 29 GENCODE (GRCh37; RRID: SCR_014966; ref. 43). Reads were summarized per gene using RSEM v1.3.0 (RRID: SCR_00262; ref. 44). Bioconducter (RRID: SCR006442; ref. 45) BioMart (RRID: SCR_019214; ref. 46) was used to map the Ensemble transcript IDs to NCBI gene IDs and UniGene symbols. Immunoglobulin genes were removed before further analysis. Differential gene expression analysis was performed using DESeq2 (RRID: SCR_015687; ref. 47), and P values were adjusted for multiple testing using Benjamini and Hochberg FDR correction (48).

GSEA v4.2.3 (RRID: SCR_003199; ref. 49) was used to perform gene set enrichment, with FDR-adjusted P values < 0.15 considered significantly enriched. The Hallmark, Canonical Pathways, Transcription Factor Targets, and Gene Ontology gene set collections from MSigDB v2022.1 (RRID: SCR_016863) were tested for enrichment. Next, Cytoscape v3.9.1 (RRID: SCR_003032; ref. 50) EnrichmentMap v.3.3.4 (RRID: SCR_016052; ref. 51) was used to find clusters of enriched gene sets with large numbers of shared genes. Thus, gene sets in each cluster were likely enriched because of the differential expression of shared member genes. Only clusters with five or more gene sets (nodes) were included in the figure, and cluster names were manually curated for clarity based on the names of gene sets in each cluster.

Three separate single-cell RNA sequencing (scRNA-seq) experiments were performed and cells were stained, sorted, and pooled depending on the specific experiments listed below.

• (i)

  CD45⁺ cells were mixed with CD45⁻ cells in a 1:1 ratio for each sample (OV348, OV744, and OV749);

• (ii)

  CD45⁺CD56⁺CD127^+/−Lin⁻ ILCs were mixed with CD45⁺CD56^–CD127⁺Lin⁻ ILCs in a 1:1 ratio for each sample (OV699, OV702, and OV710)

• (iii)

  CD45⁺Lin⁺ immune cells (OV699, OV702, and OV710).

Staining was performed as per the manufacturer’s guidelines and in consultation with Princess Margaret Genomics Center. For dataset (i), cells were stained with FcR blocking antibody (catalog no. 564220, 1:200, BD Biosciences, RRID: AB_2869554) in PBS at 4°C for 20 minutes. The cells were then stained with fixable viability dye eFluor 780 (catalog no. 65-0865-14, 1:1000, eBioscience) and surface antibody cocktails before FACS. For datasets (ii) and (iii), cells were also hashtagged and pooled with four other tissue/tumor types and cellular indexing of transcriptomes and epitopes sequencing was performed. Briefly, cells were stained with FcR blocking antibody (catalog no. 564220, 1:200, BD Biosciences, RRID: AB_2869554) and DNA-barcoded tetramer specific to MR1 (catalog no. 405261, BioLegend) and CD1d (catalog no. 405271, BioLegend) to identify mucosal-associated invariant T cells and NK T cells, respectively, at 4°C for 40 minutes. Human MR1 5-OP-RU and CD1d PBS-57 tetramers and their control counterparts (i.e., MR1 6-FP and CD1d unloaded) were provided by NIH and incubated with biotinylated monomers with TotalSeq-C oligo-barcoded streptavidin (i.e., catalog no. 405261, 405271, 405273, and 405275, respectively, BioLegend) as per the manufacturer’s instructions. Cells were stained with fluorescence surface antibody cocktail for FACS at 4°C for 30 minutes. Finally, the cells were stained with a cocktail of Hashtag DNA-barcoded antibodies targeting β2M and CD298 (catalog no. 394661, 394663, 394665, 394667, and 394669; RRIDs: AB_2801031, AB_2801032, AB_2801033, AB_2801034, and AB_2801035; BioLegend) and DNA-barcoded antibodies targeting unique surface antigens. For the latter, TotalSeq-C Human Universal Cocktail targeting 130 unique antigens (catalog no. 399905, RRID: AB_2876728, BioLegend), along with antibodies targeting cKit (catalog no. 313243, RRID: AB_2810474, BioLegend), CRTH2 (catalog no. 350131, RRID: AB_2814279, BioLegend), and CD307a (catalog no. 374413, RRID: AB_2876703, BioLegend), was used. Live cells were then sorted and pooled before sequencing. For (ii) and (iii), lineage markers were defined as CD3⁻TCRαβ⁻TCRγδ⁻CD14⁻CD19⁻CD20⁻CD34⁻CD123⁻CD303⁻FcεRIα⁻.

Samples were prepared at the Princess Margaret Genomics Center using Chromium Single Cell 5′ v2 Reagent Kit (catalog no. PN-1000263, 10× Genomics) according to the manufacturer’s instructions. Library size was determined using Agilent High Sensitivity DNA Kit (5067-4626, Agilent). Samples were pooled and quantified using Qubit dsDNA HS Assay Kit (Q32851, Qubit) on the Qubit 2.0 Fluorometer. Pools were sequenced with an S2 flow cell on a NovaSeq 6000 platform (Illumina) to a targeted depth of 50,000 reads per cell. FASTQ files were aligned to GRCh38, and counts data were assembled using CellRanger v6.1.2 (10× Genomics, SCR_023221).

Quality control and preprocessing included the removal of cells with low features (≤200), high features (≥5,000), and/or a high fraction of mitochondrial genes. Moreover, genes expressed in less than 10 cells were excluded. For datasets (i), (ii), and (iii), 15,742 cells, 1,374 cells, and 23,289 cells passed the threshold, respectively, and were included in the analysis. After log-normalization of the data, samples were batch corrected using Seurat’s integration function (52), which used reciprocal principal component analysis (PCA; ref. 53). After scaling the data and PCA, cells were clustered using Seurat’s built-in clustering function (resolutions for datasets i, ii, and iii were 1.0, 0.5, and 0.5, respectively). Uniform Manifold Approximation and Projection (UMAP) for dimension reduction was then used to obtain pairs of values (UMAP1 and UMAP2) summarizing the expression profile of each cell for visualization purposes. These coordinates were plotted with colors determined by the Seurat cluster assignment of each cell. Cell clusters were annotated following examination of differentially expressed markers obtained by Seurat FindAllMarkers in the initial round of clustering, and its cell sub-clustered to identify subpopulations. For dataset (i), CD56⁺ILCs underwent unsupervised clustering (resolution = 0.2) and resulted in subclusters CD103⁺CD56⁺ ILCs (90 cells) and CD103⁻CD56⁺ ILCs (344 cells). For dataset (ii), CD103⁺CD56⁺ ILCs underwent unsupervised clustering (resolution = 0.5) and resulted in two subclusters: sc0 (256 cells) and sc1 (106 cells). Plots were generated using ggplot2 v3.3.6. CD56⁺CD3⁻ cells were gated by protein expression using CITEViz (54).

Single-cell regulatory network inference and clustering (SCENIC) analysis was performed on CD56⁺ ILC clusters from scRNA-seq (dataset i) to reconstruct gene regulatory networks by identifying regulons (or TFs and their target genes; ref. 55). pySCENIC (v.0.12.0, Python 3.10.6, RRID: SCR_025802) identifies potential gene targets for each TF based on co-expression (GENIE3/GRNBoost2), selects potential direct binding targets based on DNA motif analysis (RcisTarget), and analyzes network activity in each individual cell. Only regulons that reached P adjusted value of <0.01 were included in output figures.

Methods for TIL expansion are described by Nguyen and colleagues (56). Single-cell suspensions of TILs were plated at 1 × 10⁶ cells/well in a Nunc Cell-Culture Treated 24 well plate (catalog no. 142475, Thermo Fisher Scientific) at 37°C and 5% CO₂. The cells were cultured in 2 mL of complete media comprising Iscove’s modified Dulbecco’s medium (catalog no. 12440-046, Gibco), 10% human serum (catalog no. 106-512, GeminiBio), 25 mmol/L HEPES (catalog no. 15630-080, Gibco), 100 μg/mL penicillin/streptomycin (catalog no. 17-602E, Lonza), 2 mmol/L L-glutamine (catalog no. 25030-081, Gibco), 10 μg/mL gentamicin (catalog no. 15290-018, Gibco), 0.05 mmol/L 2-Mercaptoethanol (catalog no. 21985-023, Gibco), and 6,000 IU/mL human recombinant IL-2 (i.e., Proleukin, Novartis) for 4 weeks. Media were replenished three times a week. TILs were split when confluent and were combined, mixed, and replated. “Rapidly expanding TILs” refer to TIL cultures that yielded ≥4 × 10⁶ cells by 4 weeks, and “Slow or not expanding TILs” refer to TIL cultures that yielded <4 × 10⁶ cells by 4 weeks. The cells were processed for flow cytometry analysis.

Intratumoral CD103⁺CD56⁺Lin⁻ (CD3⁻CD14⁻CD19⁻) ILCs and CD45⁺ immune cells (excluding CD56⁺Lin⁻ ILCs) were sorted using FACS from viably frozen samples of primary HGSC tumors (n = 4). CD45⁺ immune cells were stained with 10 μmol/L cell proliferation dye eFluor 450 (catalog no. 65-0842-90, eBioscience) in PBS for 15 minutes at 4°C. CD103⁺CD56⁺Lin⁻ ILCs were co-cultured with CD45⁺ immune cells (5 × 10⁴ cells/well) in a 1:3 or 1:5 (suppressor:responder) ratio with soluble anti-CD3 (1 μg/mL, catalog no. 555337, BD Biosciences) for 4 days. All cell cultures were completed in RPMI-1640 complete media (catalog no. 11875119, Invitrogen) containing 10% FCS (catalog no. A15-701, PAA Laboratories), 2 mmol/L L-glutamine (catalog no. 25030-81, Gibco), 100 μg/mL penicillin–streptomycin (catalog no. 15140-122, Gibco), and 0.05 mmol/L 2-Mercaptoethanol (catalog no. M7522, Sigma-Aldrich). In a separate assay, CD103⁺CD49a⁺CD56⁺Lin⁻ ILCs were cocultured with 10,000 or 15,000 autologous CD8⁺ T cells (1:1 or 1:13 suppressor to responder ratio) with soluble anti-CD3 (2 μg/mL, catalog no. 555337, BD Biosciences) and anti-CD28 (2 μg/mL, catalog no. 555726, BD Biosciences) for 3 to 4 days. CD49a was used in these assays to increase purification of these cells, although majority of CD103⁺CD56⁺Lin⁻ ILCs were CD49a⁺. To increase survival of intratumoral T cells, IL-2 (100 IU) or CD25⁻CD4⁺ autologous T cells (1:1 ratio of CD8⁺ T cells and CD25⁻CD4⁺ T cells) were added into the cultures. The cells were incubated in Nunc MicroWell 96-well polystyrene microplates (catalog no. 12-565-66, Thermo Fisher Scientific) at 37°C and 5% CO₂. Precision Count Beads (catalog no. 424902, BioLegend; 10,000 beads/well) were added to retrieve absolute cell numbers from wells. The cells were analyzed on flow cytometry to identify changes in the cytolytic marker granzyme B (GZMB; AF700, QA16A02, 1:50, BioLegend), the activation marker CD25 (PE, M-A251, 1:50, BioLegend), and the proliferation marker Ki-67 (FITC, 20Raj1, 1:25, BioLegend).

A cohort of eight HGSC samples (Supplementary Table S1) was included in this analysis. No statistical method was used to predetermine sample size, and all samples were stained simultaneously. Tissue samples were sectioned at 4 μmol/L thickness and formalin-fixed, paraffin-embedded at University Health Network with the Immune Profiling Team in Toronto. Tissue sections were baked at 60°C for 1 hour and deparaffinized in xylene for 30 minutes, followed by rehydration in a graded series of alcohol (ethanol:deionized water) at 100:0, 96:4, 90:10, 80:20, and 70:30 for 5 minutes each. Slides were placed in 1× TBS for 10 minutes following the last step in the series. In a decloaking chamber at 95°C, heat-induced epitope retrieval was performed in Tris-EDTA buffer (pH 9.2) for 30 minutes After complete cooling, the slides were incubated in blocking buffer [3% BSA and 5% horse serum (H0146, Sigma-Adlrich) in 1× TBS] at room temperature for 1 hour. Tissue on the slides was circled with a hydrophobic pen and samples were incubated in blocking buffer containing primary antibodies at appropriate concentrations at 4°C in a wet chamber. Tissue samples were washed in 1× TBS three times for 5 minutes each followed by a 5-minute incubation step in iridium at a dilution of 1:1,000 to counterstain slides. The slides were then washed in 1× TBS three times for 5 minutes again, briefly rinsed in ddH₂O to remove any salt, and dried before imaging mass cytometry (IMC) measurements. The antibody panel described in Supplementary Table S2 was used to stain the tissue sections.

Antibodies were optimized on different control tissues including tonsil, spleen, appendix, and HGSC (see Supplementary Table S2 for a list of antibodies used). Initial immunofluorescence staining was performed to ensure antibody specificity and co-localization. Slide preparation followed the same steps as outlined in the previous section, and all samples were stained simultaneously. The slides were incubated in blocking buffer containing primary antibodies at 4°C in a wet chamber overnight followed by washing in 1× TBS three times for 5 minutes each. Tissues were then incubated in blocking buffer containing fluorescent secondary antibody [with either Goat anti-Rabbit IgG Alexa Fluor 488 (catalog no. A27034, Invitrogen), Goat anti-Mouse IgG Alexa Fluor 488 (catalog no. A11029, Invitrogen), Goat anti-Mouse IgG Alexa Fluor 555 (catalog no. A21424, Invitrogen), or Goat anti-Rabbit IgG Alexa Fluor 555 (catalog no. A27039, Invitrogen)] at room temperature for 1 hour. The slides were then washed in 1× TBS three times for 5 minutes again and then counterstained with DAPI (2 μg/mL) for 5 minutes and washed again in 1× TBS three times for 5 minutes Coverslips were mounted using Prolong gold anti-fade (Thermo Fisher Scientific) and imaged. All antibody conjugations were performed at the Lunenfeld-Tanenbaum Research Institute according to the manufacturer’s instructions (Maxpar X8 Multi-Metal Antibody Labeling Kit, Fluidigm). Metal-tagged antibodies were stored in a Candor Antibody Stabilizer (CANDOR Bioscience) at 4°C.

Image acquisition and single-cell quantification and clustering were blinded to patient identifiers and clinical metadata. Acquisition order was performed randomly, independently of patient. IMC images were acquired at a resolution of roughly 1 μm. Cores were laser-ablated at a frequency of 200 Hz in a rastered pattern using the Hyperion Imaging System (Standard BioTools). Ablated tissue aerosol was ionized using inductively coupled plasma, and resulting isotopic ion reporters were quantified using time-of-flight mass spectrometry to infer protein abundance. Raw data were compiled using the Standard BioTools commerical acquisition software. All image acquisitions were successful and processed, whereby nearly each entire whole section was acquired in a number of regions of interest per patient sample.

Due to the nature of IMC application, spillover of signal from one channel into another results in contamination between markers. As such, each individual metal-tagged antibody was mixed with trypan blue at 1:1 and spotted onto an agarose-coated slide and dried. Quantification by IMC of all metal isotopes was performed. A spillover matrix measuring the contamination between channels was generated and signal spillover between channels was corrected using functions from the CATALYST R package (RRID: SCR_017127; ref. 57).

Images were viewed to qualitatively assess spatial profiles in histoCAT (58). Data were converted into TIFFs and segmented into single cells using Mesmer (Deepcell, RRID: SCR_022197; ref. 59), which is available through our analysis pipeline at https://github.com/JacksonGroupLTRI/ImcPQ (58). Single-cell masks were generated using DNA, CD45, pan-cytokeratin, E-cadherin, and SMA to identify nuclear and membrane boundaries of cells. Single-cell expression data for 37 channels and DNA (Ir191 and Ir193) were outputted. The data were not transformed, and all analysis was performed on raw measurements. To remove “noise”, the data were censored at the 99.9th percentile to remove outliers. Z-score of the mean across samples for each marker was applied to visualize using heatmaps. For UMAP generation (PhenoGraph clustering), data were normalized to the 99th percentile. Image analysis and segmentation were performed in Jupyter Notebook (RRID: SCR_018315) using Python.

The single-cell IMC analysis pipeline was implemented in R. After segmentation with Deepcell into single-cell data, cell types were labeled using gating strategies. Each individual marker used for cell-type gating was visualized by histograms and by Cytomapper (60), and the cells selected by each cutoff were compared against original images to confirm gating accuracy. Combinatorial gates were employed to identify each population of interest (i.e., PanCK^–Ecadherin^–SMA^–Vimentin⁻CD45⁺CD3⁺CD8⁺). CD103⁺CD56⁺ILCs were gated as lineage negative cells (PanCK⁻SMA⁻Vim⁻CD3⁻CD20⁻CD14⁻CD15⁻CD16⁻CD123⁻CD127⁻CD117⁻). A contact-based network representation of the segmentation mask was created with the Griottes package (61), an algorithm designed to generate a network representation and pairwise interactions. First-degree neighbors of each cell were extracted. Paired t tests were performed to find significant contact-based interactions between cell types. Dimensionality reduction of the dataset was initially carried out using PCA, followed by cell clustering with PhenoGraph (RRID: SCR_016919) and visualization using UMAP. Populations -1 and 38 were removed because of the small cell numbers which may represent debris. All statistical tests were conducted using R packages, and graphical representations were created using GraphPad (Prism 10, RRID: SCR_000306).

Statistical significance was determined by a two-tailed Student t test, one-way ANOVA, or two-way ANOVAs using appropriate multiple corrections depending on experimental conditions. For Kaplan–Meier curves, significance was determined by the log-rank (Mantel–Cox) test. The n values used to calculate statistics are defined and indicated in figure legends. Significance is indicated within figures or figure legends.

Datasets generated during the current study are available online with Gene Expression Omnibus number GSE276966 and GSE276563. Code generated during data analysis will be available on Zenodo (RRID: SCR_004129) via the following link: https://zenodo.org/records/13743867. All other data generated in the study are available in the article and its supplementary files or from the corresponding author upon reasonable request.

To identify potential markers of inhibitory ILCs, we examined CD56⁺Lin⁻ cells from slow-growing TIL cultures, compared with those in rapidly expanding cultures, and found an increased proportion of cells expressing CD103 (Fig. 1A and B). We then further characterized CD103-expressing CD56⁺ ILCs using multiparametric flow cytometry using cells directly isolated from surgically resected tumors from patients with EOC. First, we found that intratumoral Lin-negative CD56⁺ ILCs expressed distinct surface markers and TFs compared with ILC1s, ILC2s, NKp46⁺ ILC3s, and NKp46⁻ ILC3s/ILC progenitors (Supplementary Fig. S1A and S1B). In addition, the CD103-expressing CD56⁺ ILCs were preferentially found in primary ovarian tumors, and to some extent patient ascites, but were absent in peripheral blood from healthy donors (Fig. 1C; Supplementary Fig. S1C). This is consistent with previous studies that have identified CD103 as a transmembrane heterodimer involved in cell adhesion and tissue retention and as a common marker of tissue-resident lymphocytes (62, 63).

Although CD56 has been used to identify NK cells with canonical antitumor and antiviral cytotoxic functions (64), recent studies have demonstrated the extensive heterogeneity within the CD56⁺ ILCs including subsets with noncanonical functions (i.e., angiogenic, cytokine-producing, and immune regulation; refs. 13, 65–68). Flow cytometry analysis of CD103⁺CD56⁺ ILCs showed that GZMB expression was not detectable, along with very low to undetectable expression of the degranulation marker CD107a and the antibody-dependent cellular cytotoxicity protein CD16 compared with CD103⁻CD56⁺ ILCs in the tumor (Fig. 1D and E). The CD103⁺CD56⁺ ILC population looked homogeneous in terms of the expression of these markers that were evaluated, supporting the idea that they represent a defined subset of cells. The reduction in cytotoxicity-related proteins in CD103⁺CD56⁺ ILCs suggests a noncytolytic role within the TME, although this does not exclude the possibility that cytotoxic functions may be upregulated upon further activation or polarization of this population.

To determine whether CD103⁺CD56⁺ ILCs are associated with clinical prognosis, we analyzed viably frozen samples from 32 patients for the proportion of CD103 expression on CD56⁺ ILCs by flow cytometry. We found that tumors with a high abundance of CD103⁺CD56⁺ ILCs (≥60% of ILCs expressing CD103) had a trend, although not statistically different, toward reduced RFS compared with tumors with low abundance of CD103⁺CD56⁺ ILCs (≤40% of ILCs expressing CD103; Fig. 1F). Tissue-resident CD8⁺ T cells have been shown to correlate with improved prognosis in the high-grade serous subtype of EOC (69) and play an important role in antitumor immunity (69, 70). To further refine this analysis, we examined the impact of the CD103⁺CD56⁺ cells in tumors with high numbers of tissue resident CD8⁺ T cells. We found an association, although not statistically significant, between CD103⁺CD56⁺ ILCs and a trend toward reduced RFS to occur in patients with high proportion of CD103⁺CD8⁺ T cells (Fig. 1G). We also evaluated whether the presence of Tregs correlated with patient prognosis in our patient cohort and we found that they were associated with an opposite trend toward an increased RFS (Fig. 1H). To validate our findings with a larger cohort of patients with ovarian carcinomas, we developed a gene expression signature score based on bulk RNA-seq data of CD56⁺ inhibitory ILCs derived from slow-growing TILs from Crome and colleagues (Supplementary Table S1; ref. 32). We then applied this signature score to a curated list of transcriptomic data of tumors from patients with ovarian carcinoma (36–41). Survival analysis showed a statistically significant difference, with patients having a high CD56⁺ inhibitory ILCs signature (n = 112) being associated with a lower RFS compared with those with a low score for this signature (n = 140; Fig. 1I). Notably, the survival analysis from this larger cohort (Fig. 1I) showed similar kinetics with our flow cytometry cohort (Fig. 1F), with the split between the CD56⁺-high and -low signature occurring at a similar time frame. Overall, these findings suggest that CD103⁺CD56⁺ ILCs may play an influential role within the TME and are associated with a trend toward poor clinical outcomes.

To further understand the biology of the CD103⁺CD56⁺ ILCs, we performed bulk RNA-seq on sorted, patient-matched populations of CD103⁺CD56⁺ ILCs and CD103⁻CD56⁺ ILCs (N = 4). CD103⁺CD56⁺ ILCs showed a distinct transcriptomic profile and grouped separately in a PCA from CD103⁻CD56⁺ ILCs (Fig. 2A and B; Supplementary Fig. S2A; Supplementary Table S3). Gene set enrichment analysis (ref. 49), followed by clustering of gene sets based on shared genes, revealed an enrichment of gene networks involved in cytokine signaling (i.e., IL-2, STAT5, IL-9, IL-37, and TPO), cell-cycle processes, and metabolic functions (i.e., glycogen synthesis and degradation and retinoic acid signaling) in CD103⁺CD56⁺ ILCs (Fig. 2C and D; Supplementary Fig. S2B; Supplementary Table S4). Correspondingly, CD103⁺CD56⁺ ILCs also had an enriched expression of genes encoding for IL-15 (i.e., IL15RA), IL-2 (i.e., IL2RB and IL2RG), and IL-27 (i.e., IL27RA) receptors (Supplementary Fig. S2C), suggesting a potential role for these cytokines in modulating the functions of these cells. Moreover, CD103⁺CD56⁺ ILCs were enriched in the gene expression of chemokine CCL5 and the cytokine IL-32. Although CD103⁺CD56⁺ ILCs had a reduced expression of cytolytic molecules (Fig. 1D and E), there was an increased protein and gene expression of ENTPD1 (ATP ectonucleotidases) and TIGIT, both often associated with exhaustion or inhibitory mechanisms (Fig. 3A and B; Supplementary Fig. S2D) and the proliferation marker Ki67 (Supplementary Fig. S2D). In addition, there was an enrichment in co-stimulatory molecules TMIGD2 (encoding for CD28H) and KIR2DL4 (Fig. 3A). CD28H binds to B7 family member HHLA-2, inducing T-cell proliferation and cytokine production through AKT-dependent signaling cascade (71). KIR2DL4 is an activation receptor for nonclassical MHC class I molecule HLA-G (72). These findings suggest that CD103⁺CD56⁺ ILCs have been activated and may have distinct functions within the TME.

In contrast, CD103⁻CD56⁺ ILCs were more enriched in gene networks associated with canonical NK cell functions (i.e., allograft rejection, leukocyte migration, and leukocyte degranulation; Fig. 2C and D; Supplementary Fig. S2E; Supplementary Table S4). For example, CD103⁻CD56⁺ ILCs were enriched in expression of genes encoding Toll-like receptors (i.e., TLR1 and TLR10), activating receptors (i.e., LILRA2 and KLRF1), and cytolysis-related molecules (i.e., FCGR3A and GZMB; Fig. 3A). Furthermore, CD103⁻CD56⁺ ILCs exhibited increased expression of class II HLA (i.e., HLA-DMA, HLA-DMB, HLA-DOA, HLA-DQB1, HLA-DPB1, HLA-DRB1, HLA-DPA1, HLA-DRA, and HLA-DQA1), suggesting a potential role in antigen presentation or an activated phenotype (Supplementary Fig. S2F; refs. 73–77). These findings suggest that CD103⁺CD56⁺ ILCs are transcriptomically distinct from CD103^–CD56⁺ ILCs.

As expected, CD103⁺CD56⁺ ILCs expressed increased levels of genes associated with tissue-homing and retention (i.e., CD69, CXCR6, ITGA7, ITGA1, and ITGAE; ref. 62) but a decrease in genes associated with tissue egress (i.e., CD62L, CCR7, ITGA5, ITGAX, KLF2, KLF4, S1PR1, and SELL; Supplementary Fig. S2C). Zhou and colleagues (78) previously reported that CD49a-expressing tissue-resident group 1 ILCs suppressed antiviral T cells in the liver of LCMV-infected mice. Moreover, tissue-resident group 1 ILCs in the decidua are involved in tissue remodeling and protection against fetal loss during pregnancy (79). Consistent with a tissue retention–like phenotype, intratumoral CD103⁺CD56⁺ ILCs had reduced protein expression of the tissue-egress marker CD49e (encoded by ITGA5) and an increased expression of the tissue retention marker CD69 compared with CD103⁻CD56⁺ ILCs (Fig. 3B and C; Supplementary Fig. S2D). Moreover, CD103⁺CD56⁺ ILCs almost all uniformly expressed CD49a by flow cytometry (Fig. 3B and C), consistent with a tissue retention profile that has been previously described to have inhibitory roles.

CD103⁺CD56⁺ ILCs also had an increased expression of genes previously reported to be upregulated in inhibitory ILCs (i.e., CD7 and CD244; refs. 32, 80). The increased expression of CD7 and other markers upregulated in inhibitory CD56⁺ ILCs from TIL cultures (i.e., GITR, NKp30, and NKp46; ref. 32) were validated by flow cytometry (Fig. 3B; Supplementary Fig. S2D). We previously found that CD56⁺ ILCs from TIL cultures inhibited T cells using an NKp46-dependent mechanism (32). In summary, intratumoral CD103⁺CD56⁺ ILCs are transcriptomically distinct from CD103⁻CD56⁺ ILCs, expressing increased features of tissue residency and molecules such as CD7, GITR, NKp30, and NKp46, which have been reported in inhibitory NK/ILC populations associated with the TME.

A marker that we had not expected to find to be enriched in CD103⁺CD56⁺ ILCs was CD101 (Fig. 3A). CD101 is a member of the immunoglobulin superfamily and has been shown to be expressed on immunoregulatory cells [i.e., Tregs, CD8αα⁺ regulatory T cells, and IL-10⁺ dendritic cells (DC); Supplementary Fig. S3; refs. 81–85]. CD101^high Tregs have a higher suppressive capacity than CD101^dim/− Tregs (81). Moreover, engagement of CD101 through agonist antibodies promoted human cutaneous DCs to suppress T cells through IL-10 (84). The increased gene expression of CD101 in CD103⁺CD56⁺ ILCs (Fig. 3A), which we validated by flow cytometry (Fig. 3D and E), is consistent with an inhibitory role. Moreover, CD101, as well as GITR, distinguishes CD103⁺CD56⁺ ILCs from other NK cell subsets [including CD49a⁺ NK cells (86), CD56^bright NK cells, and cytolytic CD56^dim NK cells; Fig. 3F and G; Supplementary Fig. S4A]. Although there has been limited evidence that CD101 is a marker for terminally differentiated cells with an exhausted state in LCMV models (87), these cells are likely not exhausted as CD101⁺ cells within CD103⁺CD56⁺ ILCs had an increased expression of co-stimulatory molecule GITR and proliferation marker Ki67 compared with CD101⁻ cells (Supplementary Fig. S4B and S4C). Overall, intratumoral CD103⁺CD56⁺ ILCs, but not other CD56⁺ NK cell subsets, expressed CD101. Moreover, CD101⁺CD103⁺CD56⁺ ILCs displayed an activated and proliferative state.

We performed scRNA-seq to further characterize the CD103⁺CD56⁺ ILCs in the TME. First, we generated scRNA-seq data from a mixture of intratumoral CD45⁺ and CD45⁻ cells (1:1), resulting in 23 populations from unsupervised clustering (Fig. 4A; Supplementary Fig. S5, Supplementary Table S5 and S6). The CD56⁺ (NCAM1) clusters (including clusters 10a and 10b) had higher expression of killer Ig–like receptors (i.e., KIR2DL1, KIR2DL3, KIR2DL4, KIR3DL1, and KIR3DL2), natural cytotoxicity receptors (i.e., NCR1 and NCR3), and killer cell lectin–like receptors (i.e., KLRF1, KLRD1, KLRC2, KLRK1, KLRC1, KLRC3, KLRG1, and KLRB1), suggesting that they are CD56⁺ ILCs (Supplementary Fig. S5D). When we applied unsupervised clustering on the CD56⁺ ILC cluster, two populations were found and distinguished by the presence or absence of CD103 (ITGAE; Fig. 4A; Supplementary Fig. S5B).

SCENIC reconstructs gene regulatory networks through cis-regulatory motif analyses and infers TFs and cellular states at a single-cell level in silico (55). CD103⁺CD56⁺ ILCs showed enrichment of TF regulatory networks associated with development, differentiation, and function of group 1 ILCs (i.e., NFIL3, RUNX3, GATA3, and BACH2; Fig. 4B; Supplementary Fig. S6; Supplementary Table S7; refs. 88, 89). For example, BACH2 acts as a negative regulator of NK cell maturation and inhibits cytolytic programming (88). RUNX3 is essential for normal development and survival of ILC1s (90) and programs CD8⁺ T-cell residency in non-lymphoid tissues and tumors (91). Regulons involved in RUNX3-dependent transcriptional circuitry including interferon regulatory factor 4 (IRF4) and PRDM1 (92) were also enriched in CD103⁺CD56⁺ ILCs. The increased protein expression of IRF4 was validated with flow cytometry (Fig. 4C). It was also observed that Retinoid X receptor alpha (RXRA) regulon, a nuclear receptor activated by retinoic acid, was enriched in CD103⁺CD56⁺ ILCs. Retinoic acid has been shown to induce IL-10–expressing inhibitory ILC2s (93) and modulate Treg differentiation from naïve CD4⁺ T cells (94). These findings suggest a distinct TF circuitry in intratumoral CD103⁺CD56⁺ ILCs compared with CD103^–CD56⁺ NK cells.

To complement the SCENIC analysis, we investigated differences in TFs that are often defined in ILCs in the intratumoral CD103⁺CD56⁺ ILCs in contrast to CD103^–CD56⁺ NK cells (Supplementary Fig. S7). ILC subsets are often defined by expression of the TFs TBET, EOMES, and RORC (Supplementary Fig. S1A and S1B; ref. 95). Although there was a trend toward a reduced level of TBET, CD103⁺CD56⁺ ILCs did not show a significant difference in protein expression of EOMES, compared with CD103⁻CD56⁺ ILCs. Although SCENIC analysis showed an enrichment of GATA3, RORC, and FOXP3 regulons, we did not observe expression of these proteins by flow cytometry (Supplementary Fig. S7). However, CD103⁺CD56⁺ ILCs upregulated gene expression of TFs related to FOXP3-associated complexes (i.e., ETS1 and IKZF3) and ILC development (i.e., TOX and ZBTB16; Fig. 3A). ETS1 (ETS proto-oncogene 1) interacts with FOXP3 intronic enhancer, which is required for Treg development and suppressive functions (96). Enrichment of TOX and PLZF (ZBTB16) were further confirmed by flow cytometry (Supplementary Fig. S7A and S7B). PLZF identifies CD8αα⁺ T cells that control T cell–mediated autoimmunity through perforin (97). Although intratumoral CD103⁺CD56⁺ ILCs did not express detectable levels of FOXP3 (Supplementary Fig. S7D and S7E), they expressed a subset of genes that overlap with those expressed by Tregs (Fig. 4D; refs. 35, 81, 84, 88, 96, 98–111). In contrast, CD103⁻CD56⁺ NK cells expressed gene signatures more similar to γδ T cells, which may be because of shared innate cytotoxicity mechanisms. Overall, these findings suggest that intratumoral CD103⁺CD56⁺ ILCs express unique gene programs.

Several groups have identified other ILC subsets within the CD56⁺ populations, including CD56⁺ ILC3s and CD56⁺ intraepithelial ILC1s within the tonsil (66, 67, 112). To determine whether CD103⁺CD56⁺ ILCs shared features with other ILC subtypes, we performed cellular indexing of transcriptomes and epitopes sequencing on sorted CD45⁺Lin⁻CD56⁺ ILCs mixed with CD45⁺Lin⁻CD127⁺ ILCs in a 1:1 ratio for three additional ovarian tumors to enrich for ILCs and ensure sufficient cell numbers for analysis. After batch-correction and preprocessing in silico, 4,703 genes and 134 protein markers were utilized for unsupervised clustering, which resulted in five clusters identifying ILCs (Fig. 4E–G; Supplementary Tables S5 and S6). Gating on the protein expression of CD56⁺CD3⁻ cells, we found that they clustered mostly with NK cells (39.11%), CD103⁺ILCs (35.02%), ILC3s (19.46%), ILC1s (4.28%), and ILC2s (2.14%; Fig. 4F; Supplementary Fig. S8A and S8B). The CD56⁺CD3⁻ cells that express protein levels of both CD103 and CD49a were mostly specific to cluster 0 (Supplementary Fig. S8C and S8D). Although there were some ILC3s (cluster 3) and Ki67⁺ ILC1s (cluster 5) that expressed CD56, these cells did not express CD103. CD103⁺CD56⁺ ILCs did not express markers related to conventional NK cells (i.e., CD16, CD57, CD45RA, KIR2DL2, KIR3DL1, and KLRF4) or ILC3-related markers (i.e., cKit, ICOS, and CD25; Supplementary Fig. S9A and S9B). Similarly, markers associated with tonsillar CD56⁺ ILC3s characterized by Cella and colleagues (66) were not expressed by CD103⁺CD56⁺ ILCs (i.e., IL1R1, IL23R, IL7R, CCL20, CD33, CSF2, FXYD5, NCR2, NFKBIA, PRR5, RORC, TNFSF13B, and TOX2; Supplementary Fig. S9C). Therefore, there is no evidence to support that these CD103⁺CD56⁺ ILCs are related to the CD56⁺ ILC3 subset. Currently, there are no clear markers that can distinguish tissue-resident NK cells and CD56⁺ ieILC1s in human tissues (4). Although CD103⁺CD56⁺ ILCs also expressed various markers associated with group 1 ILCs (i.e., KLRB1, KLRD1, and KLRC3), CD103⁻CD56⁺ NK cells and Ki67⁺ ILC1s also expressed these genes (Supplementary Fig. S9C). In contrast to CD103⁻CD56⁺ NK cells, both CD103⁺CD56⁺ ILCs and Ki67⁺ ILC1s expressed KLRC1 and ZNF683 (encoding protein Hobit; Fig. 4H). Hobit is a key TF that regulates tissue residency and differentiation in murine tissue–resident NK cells and ILC1s, but the requirement of these cells for Hobit varies in different tissues (113, 114). CD103⁺CD56⁺ ILCs were distinct from Ki67⁺ ILC1s because of their expression of TF IKZF3 (encoding protein Aiolos; Fig. 4H) and the absence of CD27 and CD71 (Supplementary Fig. S9D). Finally, we did not find evidence that CD103⁺CD56⁺ ILCs were composed of heterogeneous populations. Subclustering of these cells in silico resulted in two subclusters that did not have many genes that were differentially expressed (Supplementary Fig. S9D and S9E; Supplementary Table S6). It is possible that these subclusters represent differences in the cellular state, such as activation status, rather than representing different ILC subsets. Overall, CD103⁺CD56⁺ ILCs isolated from ovarian cancer represent a distinct population of cells that have unique transcriptomic programs.

To determine whether CD103⁺CD56⁺ ILCs affect CD8⁺ T cells, we cocultured sorted CD45⁺ immune cells from surgically resected tumors (n = 4) with soluble anti-CD3s with or without intratumoral CD103⁺CD56⁺ ILCs in a 1:3 or 1:5 ratio (suppressor: all other CD45⁺ immune cells; Fig. 5A). These assays were done using samples from four patients across two experiments in which the analyses were kept separately. CD8⁺ TILs cultured with intratumoral CD103⁺CD56⁺ ILCs showed reduced expression of cytolytic molecule GZMB (Fig. 5B and C). Moreover, when intratumoral CD103⁺CD49a⁺CD56⁺ ILCs were cultured with T cells and soluble anti-CD3 and anti-CD28 (Fig. 5D), CD8⁺ T cells had a reduced expression of activation marker CD25 and reduced proportion of CD25⁺GZMB⁺CD8⁺ T cells (Fig. 5E and F). The proliferation marker Ki67 was also reduced in CD8⁺ T cells when cultured with CD103⁺CD49a⁺CD56⁺ ILCs (Fig. 5G). These findings suggest that intratumoral CD103⁺CD49a⁺CD56⁺ ILCs may inhibit CD8⁺ T-cell effector function.

To examine whether CD103⁺CD56⁺ ILCs were associated with a similar CD8⁺GZMB^lo phenotype ex vivo, we characterized intratumoral T cells from patients with varying proportions of CD103⁺CD56⁺ ILCs (Fig. 6A). It is important to note that there was interpatient heterogeneity in the proportion of CD103⁺CD56⁺ ILCs present within the TME (Supplementary Fig. S10A). We found that CD103⁺CD56⁺ ILC^high tumors correlated with a reduction in GZMB-expressing CD8⁺ T cells (Fig. 6A), consistent with the ILC coculture assays above. Similar findings were observed when we evaluated GZMB expression in CD8⁺ T cells from samples with high proportions of CD103⁺CD56⁺ ILCs (ILC^high; >60% of CD56⁺ ILCs positive for CD103) compared with low proportions of CD103⁺CD56⁺ ILCs (ILC^low; <40% of CD56⁺ ILCs positive for CD103; Fig. 6B). Similarly, linear regression analysis showed that as tumors had an increasing proportion of CD103⁺CD56⁺ ILCs, there was a decrease in GZMB-expressing CD8⁺ T cells (Fig. 6C). Duhen and colleagues (115) and others have previously found that CD39⁺CD103⁺ cells are enriched for tumor-reactive T cells in human solid tumors (116). Therefore, we examined whether antigen-specific CD8⁺ T cells were present and whether they correlated with the proportion of CD103⁺CD56⁺ cells (Supplementary Fig. S10B–S10D). The presence of CD103⁺CD56⁺ ILCs was associated with high abundance of potentially antigen-specific CD39⁺CD103⁺CD8⁺ T cells (Supplementary Fig. S10C) despite our findings that there were reduced proportions of GZMB-expressing CD8⁺ T cells (Fig. 6A–C). Similarly, patients with the CD103⁺CD56⁺ ILC^high group had a trend toward increased proportion of CD39⁺CD103⁺CD8⁺ T cells (Supplementary Fig. S10B). Linear regression analysis showed that as tumors had an increasing proportion of CD103⁺CD56⁺ ILCs, there was also an increase in CD39⁺CD103⁺ CD8⁺ T cells (Supplementary Fig. S10D). Therefore, our data are consistent with the interpretation that the presence of CD103⁺CD56⁺ ILCs correlate with the downregulation of GZMB on potentially antigen-specific cytotoxic T cells found in the same TME.

To determine whether these CD103⁺ inhibitory ILCs were physically interacting with T cells within the tumor, we performed IMC on formalin-fixed, paraffin-embedded slides from HGSC tumors (Fig. 6D). NKp46 was used to identify CD103⁺ ILCs as CD56 expression is highly abundant in ovarian tumor cells and identifying these ILCs from background through histology approaches with CD56 is not feasible. Moreover, NKp46 is expressed on the majority of CD103⁺CD56⁺ ILCs (Fig. 3E). We found that the CD8⁺ T cells in contact with CD103⁺NKp46⁺Lin⁻ ILCs had a lower expression of GZMB and Ki-67 compared with CD8⁺ T cells not in contact with these ILCs (Fig. 6E and F). To determine other lymphocytes that interacted with CD103⁺NKp46⁺Lin⁻ ILCs, we plotted CD45⁺ immune cells that were in contact with CD103⁺NKp46⁺Lin^– ILCs on an UMAP and identified cell clusters after unsupervised clustering (Fig. 6G and H; Supplementary Fig. S10E and S10F). Among all the cells that came into contact with CD103⁺NKp46⁺Lin⁻ ILCs, we found that they were preferentially in contact with GZMB^lowCD8⁺ T cells compared with other cell subsets (Fig. 6I). Overall, our findings suggest that CD103⁺ ILCs are interacting with T cells with a lower cytolytic potential in the TME, providing a direct association between ex vivo findings and our in vitro cultures.

To further investigate the association between CD103⁺CD56⁺ ILCs with altered an CD8⁺ T-cell profile, we performed scRNA-seq of intratumoral CD45⁺Lin⁺ cells. First, we characterized three subsets of intratumoral CD8⁺ T cells (Fig. 7A; Supplementary Table S5 and S6) and performed differentially expressed gene analysis for each subset (c1, c2, and c9) comparing CD8⁺ T cells from tumors with high proportions of CD103⁺CD56⁺ ILCs [i.e., OV710 (80.4%) and OV348 (93.4%)] compared with tumors with lower proportions of CD103⁺CD56⁺ ILCs [i.e., OV749 (4.6%); OV744 (9.4%); and OV702 (40.9%); Fig. 7B; Supplementary Table S8]. CD8⁺ T cells from CD103⁺CD56⁺ ILC–high tumors had a lower expression of genes within T-cell receptor (TCR) signaling pathway (e.g., CD3G, CD8B, CD247, IKBKB, LAT, MAPK3, NFATC4, PAK4, PPP3CB, RASA1, SHC, SOS1, VAV3, and ZAP70) compared with cells from CD103⁺CD56⁺ ILC–low tumors (Fig. 7C). To identify genes associated with T-cell activation, we performed scRNA-seq of CD8⁺ T cells from healthy donor peripheral blood mononuclear cells and stimulated them with anti-CD3 and anti-CD28 for 18 hours in vitro and identified a T-cell activation signature. Surprisingly, intratumoral CD8⁺ T cells in CD103⁺CD56⁺ ILC–high tumors had a lower expression of genes that were involved in T-cell activation (e.g., CD38, CD47, CDK4, CISH, FASLG, IL2RA, IL2RB, IKZF4, PTPN6, TIGIT, TNFRSF4, STAT1, and STAT2) compared with cells from CD103⁺CD56⁺ ILC–low tumors (Fig. 7D). Overall, CD103⁺CD56⁺ ILCs are associated with CD8⁺ T cells with less cytotoxic capacity and reduced activation status.

In this study, we have examined the cellular composition of the ovarian TME with the goal to identify intratumoral ILCs that have given rise to the inhibitory population originally identified in TIL cultures (32). We have found that the expression of CD103 delineated CD56⁺ ILCs that were associated with poor proliferative capacity of TILs in culture. We characterized CD103⁺CD56⁺ ILCs directly from the TME and found that they expressed distinct features and transcriptional regulons. CD103⁺CD56⁺ ILCs were able to reduce the expression of the cytolytic molecule GZMB on autologous CD8⁺ T cells in vitro. Moreover, CD8⁺ T cells found in the same TME with CD103⁺CD56⁺ ILCs also showed a reduced GZMB expression and had a transcriptional signature consistent with a less activated phenotype. Using IMC, we further showed evidence for a direct interaction of these cells in situ. Collectively, the presence of CD103⁺CD56⁺ ILCs correlated with GZMB^lowCD8⁺ T cells, suggesting that tumors with a prominent population of CD103⁺CD56⁺ ILCs would show limited immunosurveillance and poor prognosis.

Group 1 ILCs expressing tissue retention molecules (i.e., CD103 and CD49a) have been shown to suppress T cells in chronic viral infection models (78, 117, 118) and may have tumor-promoting functions in cancer models (119, 120). Moreover, human CD103⁺ group 1 ILCs induced in vitro have been previously demonstrated to suppress T cells. For example, Neo and colleagues (34) found that CD73⁺CD56⁺ ILCs from peripheral blood cocultured with tumor cells could subsequently suppress T-cell proliferation and reduce IFN-γ expression via IL-10. Although they did not characterize the expression of tissue retention proteins from ILCs directly from tumors, their supplementary table showed an increase in ITGAE (CD103) gene expression in CD73⁺ peripheral blood cells cocultured with primary sarcoma tumors. Recently, our group has found that peripheral blood NK cells treated with TGF-β1 and IL-15 had an increased expression of CD103, CD49a, GITR, and CD101 (121) which have a similar expression profile as intratumoral CD103⁺CD56⁺ ILCs from ovarian cancer as described here. Moreover, TGF-β1/IL-15–induced CD103⁺CD56⁺ NK–like cells inhibited proliferation, reduced absolute numbers, and reduced CD25 expression of autologous T cells in vitro (121). In this current study, we found that CD103⁺CD56⁺ ILCs directly isolated from ovarian tumors downregulated the expression of activation marker CD25 and proliferation marker Ki67 on autologous CD8⁺ T cells in vitro. We also found that CD103⁺CD56⁺ ILCs downregulated GZMB expression on CD8⁺ T cells in vitro and were associated with GZMB^lowCD8⁺ T cells within the TME. Similarly, Tregs have also been shown to reduce effector functions and cytolytic activity in T cells (122–124). Although we were unable to clearly demonstrate that these CD103⁺CD56⁺ ILCs isolated directly from the tumors could inhibit proliferation of CD4⁺ T cells or CD8⁺ T cells because of the extremely low numbers of these cells within the TME, our analysis here, in combination with findings from our previous work with in vitro–generated CD103⁺CD56⁺ cells (121), is consistent with the possibility that these cells may acquire a regulatory function if they encountered appropriate signals in the TME and further develop into inhibitory cells. Collectively, these findings suggest that CD103⁺CD56⁺ ILCs from ovarian tumors have inhibitory properties.

It is important to note that the regulation of CD8⁺ T cells by ILC1-like cells occurs in very defined conditions. Xu and colleagues (125, 126) and Crouse and colleagues (126) found that type I IFNs were required to render CD8⁺ T cells susceptible to NK cell–mediated killing in LCMV-infected mice through a Qa-1/NKG2A-dependent mechanism (127). This suggests that the potential inhibitory role of CD103⁺CD56⁺ ILCs may differ according to the TME context and the ligands present on CD8⁺ T cells. Moreno-Nieves and colleagues (128) described a population of CD103⁺CD49a⁺CD56⁺ NK cells with an ieILC1-like signature in head and neck squamous cell carcinoma (HNSCC). Although they found that in vitro–induced CD49a⁺ NK cells, derived from healthy peripheral blood NK cells stimulated with IL-15 and the HNSCC cell line PCI-33, showed potent antitumor activity against xenograft tumors, it is unclear whether intratumoral CD103⁺CD56⁺ NK cells directly isolated from HNSCC have an immune-inhibitory function. It is possible that NK cells and ILCs have various roles within the TME, and further studies are needed to investigate mechanisms that promote antitumor immunity or trigger an inhibitory function of ILCs.

Although previous studies have correlated CD8⁺ T-cell infiltration with prognosis or survival, several studies have evaluated GZMB levels in the CD8⁺ T cells as well. For example, IHC approaches have found that tumors with both low GZMB expression and low CD8⁺ T-cell densities were associated with poor clinical outcomes in several cancers (129–131). Zhu and colleagues (132) found that low densities of GZMB⁺CD8⁺ T cells by IMC, indicating lower proportion of cytotoxic CD8⁺ T cells, were associated with poor overall survival in patients with high-grade serous carcinomas. Patients with metastatic melanoma or triple-negative breast cancer that responded to anti-PD1/PDL1 single agent or combination therapy were associated with increased GZMB-expressing CD8⁺ T cells (133–138). Recent studies have attempted to categorize tumors with functional markers expressed by CD8⁺ T cells and found a correlation with a distinct TME subset (136, 139, 140). Using scRNA-seq, Tietscher and colleagues (140) identified a subset of patients with breast tumors that did not have CD8⁺ T cells with a PD-1^highCTLA-4^highCD38^high–activated/exhausted phenotype and had a TME with innate and adaptive lymphocytes that expressed lower levels of cytolytic markers GZMB, GNLY, and FASLG. Moreover, these tumors were associated with myeloid cells with a reduced inflammatory and phagocytic phenotype. Gruosso and colleagues (141) also found that GZMB⁺CD8⁺ T cells identified using immunofluorescence were associated with more immunoreactive microenvironment with higher intratumoral type I IFN signature in patients with triple-negative breast cancer. These studies underscore the importance of understanding correlates and mechanisms that dictate different functional and dysfunctional profiles of CD8⁺ T cells within the TME. One of the key findings in this current study is that CD103⁺CD56⁺ ILCs were associated with a reduction in GZMB-expressing CD8⁺ T cells by flow cytometry and IMC. Moreover, scRNA-seq revealed that CD103⁺CD56⁺ ILCs were associated with CD8⁺ T cells with a lower activation status. These findings suggest that CD8⁺ T cells in tumors with a high abundance of CD103⁺CD56⁺ ILCs have a limited ability for immunosurveillance, possibly contributing to resistance to immunotherapy in patients with EOC (142).

In summary, we have identified intratumoral CD103⁺CD56⁺ ILCs that are associated with CD8⁺ T cells with limited immune activity. Future work characterizing CD103⁺CD56⁺ ILCs, their mechanisms of activation, and whether they are associated with clinical outcomes across different cancers may lead to better understanding of the barriers for immunosurveillance, define new targets, and lead to improved treatment strategies.

H.W. Jackson reports grants, personal fees, and nonfinancial support from Standard BioTools, personal fees and nonfinancial support from Abcam, and nonfinancial support from Somalogic outside the submitted work. P.S Ohashi reports grants from Canadian Institutes of Health Research, Canadian Cancer Society, TRANSCAN3, and Princess Margaret Cancer Foundation Wolfond Immunotherapy Fund during the conduct of the study, as well as other support from Providence Therapeutics, Treadwell Therapeutics, Tikvo Allocell, and Rondo Therapeutics outside the submitted work. S.Q. Crome, P.S. Ohashi, and L.T. Nguyen filed a patent serial no. 16/325,923. No disclosures were reported by the other authors.

D.C. Chung: Conceptualization, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. N. Shakfa: Data curation, software, formal analysis, methodology, writing–review and editing. J. Vakharia: Software, formal analysis, methodology. K. Warner: Data curation, writing–review and editing. N. Jacquelot: Data curation, writing–review and editing. A. Sayad: Software, formal analysis, writing–review and editing. S. Han: Data curation, writing–review and editing. M. Ghaedi: Data curation, writing–review and editing. C.R. Garcia-Batres: Data curation, writing–review and editing. J. Sotty: Data curation, writing–review and editing. A. Azarmina: Software, formal analysis. F. Nowlan: Software. E.L.Y. Chen: Methodology. M. Zon: Formal analysis. A.R. Elford: Resources, project administration. B.X. Wang: Resources, project administration, writing–review and editing. L.T. Nguyen: Resources, project administration, writing–review and editing. M. Mrkonjic: Resources. B.A. Clarke: Resources. M.Q. Bernardini: Resources. B. Haibe-Kains: Formal analysis. S.E. Ferguson: Resources. S.Q. Crome: Formal analysis, writing–review and editing. H.W. Jackson: Resources, project administration, writing–review and editing. P.S. Ohashi: Conceptualization, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing.

We would like to thank the past and present members of the Ohashi Laboratory for technical help and scientific discussions, including Michael St. Paul, Meghan Kates, Ning An, Aras Toker, Khalid Dhuban, and Stephanie Wong. We would also like to thank Somi Afuni from the Jackson lab for her computational help. Furthermore, we thank the Princess Margaret Cancer Center Flow Facility and Princess Margaret Genomics Center for their technical support. Special acknowledgements are made to Drs. Arthur Mortha and Matthew Buechler for their suggestions and insights for this project. We would like to note that the Canadian Cancer Society grant was generously funded in memory of Jim JoTak, and the Wolfond Immunotherapy Fund at the Princess Margaret Cancer Foundation was generously donated by the Wolfond Family. We would also like to thank NIH for providing tetramers to MR1 and CD1d. This work was supported by the following grants: CIHR foundation award (CIHR FDN #143220), TRANSCAN Canadian Institutes of Health Research grant (CIHR – TRN #184713), Canadian Cancer Society (CCSRI #706152), Canadian Cancer Society grant (CHRYSALIDS TRANSCAN 3), and the Wolfond Immunotherapy Fund.