Natural Killer Cell Regulation of Breast Cancer Stem Cells Mediates Metastatic Dormancy Open Access

Metastatic dormancy is a significant clinical problem for the 3.8 million women in the United States with breast cancer (1). Most women will have early-stage diagnosis and receive therapy that is curative on the scale of 5 to 10 years. However, many of these patients will recur with metastatic disease 10 years or more after initial diagnosis (2). The risk factors for recurrence are unclear, but evidence suggests a role for immune system function (3). Most mouse models of metastatic dormancy have utilized human tumor cells inoculated into immunocompromised mice (3), limiting their ability to uncover key immune components and thus their clinical relevance.

Natural killer (NK) cells are a major component of the immune system that has been implicated in controlling metastasis (4) and metastatic dormancy (5, 6). NK cells are innate immune cells that can distinguish normal cells from virus-infected and transformed cancer cells without requiring antigen specificity like T cells or B cells. NK cells receive signals from activating and inhibitory receptors on target cells and integrate these signals in the context of the tissue microenvironment to regulate cellular killing. Activating receptors such as NKG2D, natural cytotoxicity receptors, DNAM, and CD16 (7), recognize “induced self” signals on target cells upregulated during viral infection, inflammation, or cellular stress. Inhibitory receptors such as killer immunoglobulin receptors (KIR), NKG2A, and others recognize MHC class I molecules present on the surface of target cells. Tumor cells and virally infected cells frequently downregulate MHC class I, allowing them to escape recognition of mutated or viral antigens by T cells. As a result, a compensatory mechanism for NK cell regulation of these cells with “missing self” signals evolved (7). Similarly, a clinically relevant tumor in a human or a mouse has escaped this regulation through modulation of these receptors or inhibition of NK cell function in the tumor or metastatic microenvironment (8). In other words, evasion of the immune system is required for the development of clinically relevant disease.

The paucity of immunocompetent mouse models has hindered research into the role of the immune system in regulating tumor dormancy. To address this, we evaluated murine breast cancer cell lines for their short- and long-term dormancy in fully immunologically competent mouse models and found that each is regulated by different components of the immune system. We focused on the D2.0R model that exhibits NK cell-dependent metastatic dormancy via quiescent D2.0R cancer stem cells (CSC) evading NK cell-mediated immunity through downregulation of activating NK ligands, STING, and STING targets and upregulation of MHC partially controlled by Bach1 and Sox2 expression.

Met1 cells were a gift from the laboratory of Alexander D Borowsky (9). D2.0R and D2A1 cells were derived by the Jeffrey Green lab and were a gift from the laboratory of Hunter Kent (10). PyMT cells were derived from a spontaneous MMTV-PyMT tumor in the laboratory of M.S. Wicha. E0771 cells were a gift from the laboratory of Qiao Li.

Murine mammary carcinoma cell lines were cultured at 37°C and 5% CO₂ in a humidified incubator. PyMT and Met1 cells were maintained in RPMI 1640 (Gibco) with 10% FBS. D2.0R, D2A1, and E0771 were maintained in DMEM (Gibco) with 10% FBS and 1× GlutaMAX (Gibco). Cells were passed with TrypLE (Gibco) and subcultured at a 1:20 ratio upon reaching 90% confluency. Cells were confirmed Mycoplasma free every 2 months via MycoAlert Mycoplasma Detection Kit (Lonza).

Murine mammary carcinoma cell phenotypes were tested in two-dimensional (2D) and sphere culture conditions. 2D culture was defined as attachment culture with low-serum 2% FBS that supported slow growth over a period of 1 week. Sphere culture was defined as ultralow attachment culture with mammosphere media that supported the formation of tumor spheres over a period of 1 week. Mammosphere media was prepared with 475 mL of MEBM (Lonza) with 1× GlutaMAX (Gibco), 1× B27 (Gibco), 1× Pen/Strep (Gibco), 5 μg/mL insulin (Sigma), 20 ng/mL EGF (BD Biosciences), bFGF (Gibco), 0.5 μg/mL hydrocortisone (Sigma), 0.4 IU/mL heparin (STEMCELL), and 1% wt/vol of 400 cP methylcellulose (Sigma). Mammosphere media was prepared and placed on a rocker overnight at 4°C to dissolve the methylcellulose. After the methylcellulose was fully dissolved, media was sterile filtered and stored at 4°C for up to 2 months.

PKH67 Green Fluorescent Cell Linker Midi Kit (Sigma) was used according to manufacturer instructions to label murine mammary carcinoma cell lines. CellTrace Far Red (CTFR) Cell Proliferation Kit (Thermo Fisher Scientific) was used according to manufacturer instructions to label murine mammary carcinoma cell lines. For short-term dormancy and immunogenicity studies in vivo, tumor cells were labeled with fluorescent dye and immediately injected into mice. For studies that investigated the differential phenotype of proliferative versus quiescent tumor cells, cells were labeled with fluorescent dye and cultured for 1 week in 2D conditions described above to generate a label-retaining population that was quiescent and a nonlabel-retaining population that was proliferative.

For sphere formation assays, cells were sorted, and counts were taken from the sorter as described above. Cells were seeded at 100 cells/well in mammosphere media as described above and allowed to form spheres for 1 week. After 1 week, spheres larger than 70 μm were manually counted via light microscopy and recorded.

For colony assay, tumor cells were cocultured with primary NK cells for 24 hours. Primary NK cells were isolated from spleens of naïve BALB/c mice at 6 to 10 weeks of age. Splenocytes were isolated via mashing against a 70-μm cell strainer, and red blood cells were removed via 5 seconds of hypotonic lysis in water. NK cells were purified from total splenocytes using the NK Cell Isolation Kit, mouse (Miltenyi Biotec), and magnetic-activated cell sorting was performed using the MACS system (Miltenyi Biotec). This kit uses a depletion strategy, and the negative fraction after sorting is NK cells. NK cells were cultured in RPMI 1640 (Gibco) with 10% FBS (Gibco), 0.375% sodium bicarbonate (Gibco), 1 mmol/L sodium pyruvate (Gibco), 1% Pen/Strep (Gibco), 1,000 U/mL murine recombinant IL2 (PeproTech), and 0.001% 2-mercaptoethanol (Sigma) and used immediately for coculture assays. Target cells were first passaged from original culture conditions, counted using the Luna slide counter (Logos Biosystems) with acridine orange and propidium iodide to identify live and dead cells, respectively, and resuspended at 1e6 cells/mL in PBS. Cells were then diluted to 1e5 cells/mL in NK media, and 100 μL of cells was plated with 250 μL of NK cells in a six-well plate. After 24 hours, culture media was aspirated, and tumor cells were washed with PBS two times to remove any nonadherent NK cells. Adherent tumor cells were trypsinized, resuspended in tumor cell media, and counted by Luna slide counter. Briefly, tumor cells survived by NK cell cytotoxicity and untreated control tumor cells were inoculated in six-well plates (500 cells/well) and incubated for 8 days. The colony formation of cells was determined using crystal violet staining (Beijing Solarbio Science & Technology Co., Ltd.). Cells were then fixed with 100% methanol for 10 minutes at room temperature. Next, cells were stained by 0.1% crystal violet for 5 minutes at room temperature. Finally, the cell colonies were observed under an inverted microscope (IX73; Olympus Corporation; magnification, ×2.5) and manually counted.

For measurement of proliferation over time, cells were sorted and seeded at 100 cells/well in full serum media and allowed to grow for 2 weeks in an IncuCyte microscope incubator. Images were taken every 6 hours, and the confluency mask was used to calculate the confluency percentage over time. Data were exported to csv format from the IncuCyte software and analyzed using R.

All animal studies were approved by the University of Michigan Institutional Animal Care and Use Committee and were performed in accordance with approved protocols.

Tumor cells were prepared in sterile PBS such that the number of cells desired to be delivered to each mouse was in a volume of 100 mL. Mice were anesthetized with isoflurane, and the chest wall was depilated and cleaned with alcohol. A bright light was used to localize the heartbeat through the chest wall, and the bottom right area of this region was taken to be the left ventricle. An insulin syringe was prepared with a 100-μL air gap, and 150 μL of cell suspension was drawn into the syringe, and bubbles were removed. The needle was slowly inserted into the chest at the site of the heartbeat, and correct placement in the left ventricle was confirmed by a bright flash of pulsing blood into the syringe. Following proper placement, 100 mL was expelled from the syringe, and then slight negative pressure was applied to prevent tumor cells from spilling into the chest cavity. Pressure was applied to the chest wall, and mice were returned to a clean cage and monitored until they recovered from anesthesia.

Mice that received tumor cells were monitored at least twice weekly for the duration of the experiment. Mice were euthanized once they reached moribund conditioning (>20% decrease in body weight, limited physical activity, labored breathing, significant hunching behavior, or limping).

Tumor cells were prepared in sterile PBS with 50% Matrigel (Corning) such that the number of cells desired to be delivered to each injection was in a volume of 50 μL. Mice were anesthetized with isoflurane, and the skin above the fourth mammary fat pads was depilated and cleaned with alcohol. A bright light was used to localize the fourth nipple, and the mammary fat pad was grabbed using Adson skin grabbers near the nipple. An insulin syringe was used to deliver 50 μL of cell suspension to the fat near the nipple through the skin. A successful injection was confirmed by a ballooning of the fat at the injection site. Mice were returned to a clean cage and monitored until they recovered from anesthesia.

Mice were euthanized with CO₂, and euthanasia was confirmed via pneumothorax. The femur and tibia were removed together from the hip joint and cleaned of skin and muscle. The metaphysis at the knee was removed for the bottom of the femur and top of the tibia to expose the bone marrow cavity. Bones were placed marrow side down in a 0.5-mL tube with a small hole in the bottom. This tube was placed in a larger 1.5-mL tube and centrifuged at 10,000 × g for 15 seconds to remove the bone marrow from the bones. Successful centrifugation was confirmed by a large red pellet at the bottom of the larger tube and a white appearance of the bones in the top tube. Bones were discarded, and marrow was resuspended in sterile PBS for further staining or directly proceeding to flow cytometry analysis of LRCs.

For in vivo monitoring of bioluminescence, mice were injected intraperitoneally with d-luciferin (Regis Technologies) at a final dose of 150 mg/kg. Mice were then anesthetized using isoflurane and arranged on a warmed imaging stage in the IVIS Lumina imaging system (Perkin Elmer). After 10 minutes, mice were imaged using “Auto” settings with an open emission filter and the excitation filter blocked. Regions of interest (ROI) were drawn using the Aura software (Spectral Instruments Imaging), and the same ROIs were applied to all subsequent time points. Data were exported as flux (photons/sec) for each ROI and further analyzed using R.

For depletion of NK cells, mice were injected intraperitoneally with 300 μg of Ultra-LEAF purified antiasialo GM1 rabbit polyclonal antibody (clone Poly21460, BioLegend) every 4 days for the duration of the experiment. Control mice were injected with 300 μg of normal rabbit serum (Abcam) every 4 days for the duration of the experiment. For the depletion of macrophages, mice were injected intraperitoneally with 2 mg of Clodrosome (Encapsula), whereas control mice were injected with 2 mg of Encapsome (Encapsula) every 4 days for the duration of the experiment.

MSA-2 was prepared by dissolving one 10-mg vial of MSA-2 (MedChemExpress) in 0.2 mL of DMSO for a concentration of 50 mg/mL and then taking 100 μL of this solution and adding 400 μL of PEG300 (MedChemExpress), 40 μL of Tween80 (MedChemExpress), and 450 μL of dPBS. This created a final concentration of 5 mg/mL MSA-2. Vehicle control solution was prepared the same except without MSA-2. Approximately 200 μL of MSA-2 solution (1-mg MSA-2 per injection into 20-g mice for approximately 50 mg/kg dose) was injected intraperitoneally one time, 7 days after intracardiac inoculation of tumor cells. Vehicle control solutions were prepared exactly as described without MSA-2 and injected 200 μL per mouse, 7 days after intracardiac inoculation of tumor cells. Survival monitoring was performed as previously described.

For immunofluorescence studies, cells were cultured on sterile glass coverslips in a six-well plate. For label retention studies combined with Ki67 labeling (Anti-Ki67 clone D3B5, Cell Signaling Technology), cells were cultured in low serum for 1 week. For estrogen receptor staining (Antiestrogen Receptor alpha clone 6F11, Abcam), cells were cultured until 80% confluent. For staining, coverslips were washed with PBS, fixed with 10% neutral buffered formalin, and permeabilized with 0.5% Triton-X100 in PBS for 3 minutes. Following permeabilization, coverslips were washed with PBS twice. Primary and no primary control staining solutions were prepared with a 1:200 dilution of antibody in PBS containing 1% bovine serum albumin, and a 50-μL droplet of staining solution was added to a fresh piece of parafilm, and then the coverslip was inverted on the parafilm surface to incubate with the antibody solution for 20 minutes in the dark. The coverslips were washed twice with PBS, and the process was repeated with secondary antibody (1:200) containing DAPI for nuclear staining. Coverslips were washed twice with PBS and mounted using Fluoromount-G (SouthernBiotech) onto glass slides and imaged on Nikon A1Si confocal microscope managed by the Biomedical Research Core Facilities microscopy core (University of Michigan).

The EBImage package in R analyzes cell confluence and eccentricity using brightfield microscopy images. This package defines eccentricity as sqrt(1-minor.axis2/major.axis2), which results in a circle having an eccentricity value of zero and a straight line an eccentricity value of 1.

For analysis of LRCs, fixed unstained cells and fixed freshly labeled cells were used as negative and positive controls, respectively. Bone marrow was isolated using previously described method and analyzed on Coulter CytoFLEX or Bio-Rad ZE5 managed by the Flow Cytometry Core (University of Michigan) using the 488 laser and 525/35 emission filter for PKH67 and the 640 laser and 670/30 emission filter for CTFR. For all experiments, the presence of label-retaining cells (LRC) was reported as a fold change from a naïve control to account for any autofluorescent background present in bone marrow.

For analysis of cancer stem cell markers, tumor cell lines were cultured in 2D or sphere according to the experiment and plated in 96-well plates in triplicate for each single color control, full minus one (FMO), and master mix sample including unstained, DAPI only (Thermo Fisher), DEAB (STEMCELL) as a negative control for ALDEFLUOR, ALDEFLUOR (STEMCELL) only, PE-isotype control for CD90 (Abcam), PE-CD90 (clone G7, Abcam), APC-isotype control for s (BioLegend), APC-Sca1 (clone D7, BioLegend), FMO-ALDH, FMO-CD90, FMO-Sca1, and master mix (DAPI, ALDEFLUOR, CD90, Sca1). Samples were analyzed on Bio-Rad ZE5 managed by the Flow Cytometry Core (University of Michigan). FCS files were analyzed using FlowJo (Treestar Inc.) and/or R using the flowCore package. Single color controls were used to create a compensation matrix that was applied to all samples. FMOs were used to set gates for each stain, and these gates were applied to the master mix sample.

For analysis of epithelial-to-mesenchymal transition markers, tumor cell lines were cultured in 2D or sphere according to the experiment and plated in triplicate in 96-well plates as described above in combination with the CSC markers. All EMT markers were used in the AF647 channel, and the same Sca1 clone described above was used and conjugated to PECy7. The following antibodies were used in addition to those described above: AF647 E-cadherin (clone DECMA1, BioLegend), AF647-EpCAM (clone G8.8, BioLegend), and AF647-Vimentin (clone OD1D3, BioLegend). Samples were analyzed as described above.

For fluorescence-activated cell sorting, cells were stained as described above and sorted using the MoFlo Astrios sorter. Cells were sorted into pure FBS to maintain viability, and counts were recorded from the sorting instrument, which was used to resuspend the cells for downstream in vitro or in vivo applications.

For all NK cytotoxicity assays, primary NK cells were isolated from spleens of naïve BALB/c mice at 6 to 10 weeks of age. Splenocytes were isolated via mashing against a 70-μm cell strainer, and red blood cells were removed via 5 seconds of hypotonic lysis in water. NK cells were purified from total splenocytes using the NK Cell Isolation Kit, mouse (Miltenyi Biotec) with magnetic-activated cell sorting using the MACS system (Miltenyi Biotec). This kit uses a depletion strategy, and the negative fraction after sorting is NK cells. NK cells were cultured in RPMI 1640 (Gibco) with 10% FBS (Gibco), 0.375% sodium bicarbonate (Gibco), 1 mmol/L sodium pyruvate (Gibco), 1% Pen/Strep (Gibco), 1,000 U/mL murine recombinant IL2 (PeproTech), and 0.001% 2-mercaptoethanol (Sigma) and either used immediately for cytotoxicity assays or cultured overnight before adding target cells. Target cells were first passaged from original culture conditions, counted using the Luna slide counter (Logos Biosystems) with acridine orange and propidium iodide to identify live and dead cells, respectively, and resuspended at 1e6 cells/mL in PBS. If required, cells were sorted via FACS as previously described. Cells were then diluted to 1e5 cells/mL in NK media, and 50 μL of cells was plated with 50 μL of NK cells in each well of a 96-well plate. Target and effector cells were cultured 4 hours or overnight, and cytotoxicity was measured using the CytoTox96 LDH assay (Promega) or using click beetle green luciferase-expressing cells and read on a BioTek Synergy H1 plate reader (BioTek Instruments). The cytotoxicity percentage was calculated via manufacturer instructions as (experimental–effector spontaneous–target spontaneous)/(target maximum–target spontaneous) × 100 for LDH assay or for click beetle green (CBG)-expressing cells by using the fold change in luminescence from the tumor cell only control.

For NK cytotoxicity assays with siRNA knockdown, target cells were plated at 4e3 cells/well in 96-well plates with full serum media overnight. siRNA lipid complexes were then prepared using Opti-MEM media (Gibco), Lipofectamine RNAiMAX (Invitrogen), and Silencer siRNAs (Negative Control No. 1 or pan-Raet1 siRNA, Thermo Fisher), and complexes were added to cells according to manufacturer instructions. Cells were cultured for 1 to 3 days and analyzed for knockdown efficiency via flow cytometry or NK cytotoxicity assay.

For NK cytotoxicity assays with antibody treatment, target tumor cells and effector NK cells were prepared as described above, and 1 μg of Anti-NKG2D (clone C7, BioLegend) or Armenian Hamster IgG Isotype Control (clone HTK888, BioLegend) was added per well. Incubation and LDH assay were performed as described above.

Cells were harvested using TrypLE (Life Technologies), counted, and suspended at 5 × 10⁶ cells/mL prior to FACS isolation of LRCs (CTFR positive) and non-LRCs (CTFR negative); at least 500,000 cells were sorted for each population into pure FBS. Following FACS, cells were pelleted via centrifugation at 500 × g for 5 minutes, and the supernatant was aspirated. RNA from cell pellets was subsequently isolated using RNeasy Micro Kit (Qiagen) according to manufacturer instructions. Isolated RNA concentrations were quantified using NanoDrop 2000 (Thermo Fisher) and resuspended at 8 ng/μL. RNA samples were kept at −80°C until further use. Three biologic replicates were generated by labeling three different passage numbers of D2.0R cells with CTFR, culturing in 2% FBS for 7 days, isolating cells via FACS, and finally isolating RNA as previously described on three independent days. After all samples were collected, RNA was submitted to the University of Michigan Advanced Genomics Core for quality control, library preparation, and sequencing. RNA quality was assessed using an Agilent Bioanalyzer, and RNA integrity number (RIN) values were measured. All samples had an RIN higher than nine. Library construction was performed with the Illumina platform with poly(A) enrichment. Sequencing was performed using a NovaSeq shared flow cell run with paired-end 150 bp (PE150) on a 300-cycle S4 flow cell targeting ∼400 million total reads for all nine sample runs (44.4 million reads per sample).

BCL Convert Conversion Software v3.9.3 (Illumina) was used to generate demultiplexed Fastq files. The reads were trimmed using Cutadapt v2.3. FastQC v0.11.8 was used to ensure the quality of data. Fastq Screen was used to screen for various types of contamination. Reads were mapped to the reference genome GRCm38 (ENSEMBL), using STAR v2.7.8a and assigned count estimates to genes with RSEM v1.3.3. Alignment options followed ENCODE standards for RNA-seq. MultiQC v1.7 compiled the results from several of these tools and provided a detailed and comprehensive quality control report. The expected gene counts from alignment were then processed to determine the differential gene expression using the DESeq2 package in R. First, genes with counts less than 50 were removed. Then, differential gene expression analysis was performed between LRC and non-LRC, with the reference condition being non-LRC. Normalized counts were then calculated. Batch effects were investigated via PCA and removed using the limma package in R. Exploratory data analysis was performed via generation of MA plots and volcano plots. GSEA was performed using the ClusterProfiler package in R using KEGG gene sets.

For ChEA3 analysis, we used the ChIP-Seq Literature library of transcription factors, which includes transcription factor/gene interactions mined from the literature rather than those that are predicted across all possible libraries. This approach included 164 transcription factors, in which we then received a rank of each for each gene set we queried.

Statistical analysis and hypothesis testing was performed in R. All reported values represent the mean and standard error of the mean from three to nine biologic replicates, with at least technical triplicates within each experiment. Whenever possible, all data points are included in graphical representations of data. The method used for hypothesis testing is included in each figure caption. For analysis of the Kaplan–Meier survival curves, initial statistical testing was performed via the Wilcoxon rank sum test. Post hoc testing was also performed via the Wilcoxon rank sum test. Reported P values on panels are the overall P value for the initial test. Post hoc tests comparing individual curves of P values and corrected for multiple hypothesis testing using Bonferroni are reported in each figure caption.

The data generated in this study are publicly available in Gene Expression Omnibus at GSE267583. All other raw data generated in this study are available upon request from the corresponding author.

We chose five murine mammary adenocarcinoma cell lines (PyMT, Met1, D2.0R, D2A1, and E0771) to investigate for long-term dormancy following intracardiac inoculation. In this manuscript, we define long-term dormancy as the survival of mice after intracardiac inoculation of 100,000 tumor cells longer than 3 months without signs of metastasis. We injected 100,000 PyMT, Met1, D2.0R, D2A1, or E0771 murine mammary carcinoma cells into the left ventricle of syngeneic mice and monitored mouse health over time (Fig. 1A). Our experiments revealed that these cell models fit two phenotypes: cell lines that form metastasis very quickly and cell lines that exhibit an extended period of metastatic dormancy (>100 days; Fig. 1B). D2A1 and E0771 belong to the first group with mice exhibiting extensive metastases within 30 days. PyMT, Met1, and D2.0R belong to the second group with mice surviving past 100 days.

We next developed a model of short-term dormancy of breast cancer cells in bone marrow following intracardiac injection in immunocompetent syngeneic mice. We utilized fluorescent label retention (PKH67 or CTFR; ref. 11) to label nonproliferating tumor cells in vitro and in vivo (Fig. 1C). Cells that retain the label are quiescent LRCs, whereas cells that do not retain the label (non-LRC) are proliferative, as confirmed by Ki67 staining (Fig. 1D). Another advantage of this labeling technique is that unlike genetically introduced tags such as RFP, which elicit an immune response (Supplementary Fig. S1A), these dyes are nonimmunogenic (Supplementary Fig. S1B and S1C). We used these dyes to label PyMT, Met1, and D2.0R murine mammary carcinoma cell lines and injected these cells into syngeneic mice (FVB/N for PyMT and Met1 and BALB/c for D2.0R) via intracardiac injection to allow the cells to distribute throughout the circulation (Fig. 1E). In this manuscript, we define short-term dormancy as the presence of nonproliferating (label-retaining) tumor cells in bone marrow 1 to 4 weeks after intracardiac inoculation of tumor cells. LRCs were detectable in the femurs of mice injected with PyMT, Met1, and D2.0R cells and tibias of mice injected with PyMT cells (Fig. 1F). This short-term model of dormancy indicates that nonproliferative tumor cells lodge in bone marrow and can be detected after 1 week. Altogether, these data indicate that PyMT, Met1, and D2.0R exhibit short- and long-term dormancies, whereas D2A1 and E0771 do not. As a result, we chose to further investigate the PyMT, Met1, and D2.0R cell lines.

After identifying PyMT, Met1, and D2.0R as murine mammary carcinoma cell lines that display long-term metastatic dormancy in vivo, we investigated the CSC populations in these cell lines (D2.0R; Fig. 2, PyMT, Met1; Supplementary Fig. S2) by staining for ALDH, Sca1, and CD90 (see Supplementary Fig. S3 for gating scheme), which are well-established markers of CSC in other models of human and murine breast cancer (12). We compared 2D (2% FBS and tissue culture treated flasks) and sphere formation conditions for each cell line (Fig. 2A; Supplementary Fig. S2A). Mammosphere culture enriches for a proliferative stem-like phenotype (13), whereas low-serum 2D culture enriches for a quiescent stem-like population (14). The enrichment of ALDH⁺ and Sca1⁺CD90⁻ and the depletion of Sca1⁻CD90⁺ in a more proliferative sphere culture relative to quiescent 2D cultures suggest that ALDH⁺ and Sca1⁺CD90⁻ CSCs are more proliferative, whereas Sca1⁻CD90⁺ CSCs are more quiescent. To test this hypothesis, we performed combined staining of the CSC panel with CTFR labeling (Fig. 2B; Supplementary Fig. S2B). We compared the expression of CSC markers in 2D culture for all live cells, LRCs, or non-LRCs to determine whether these populations are quiescent (LRC) or proliferative (non-LRC). The ALDH⁺ and Sca1⁺CD90⁻ populations are significantly reduced in LRC relative to live, confirming that these populations are proliferative, whereas the Sca1⁻CD90⁺ population was significantly enriched in the LRC relative to live, indicating that this population is quiescent.

In human breast cancer cell lines, proliferative CSC is more epithelial, whereas quiescent CSC is more mesenchymal (15). We compared the percentage of cells expressing E-cadherin and vimentin for each CSC marker relative to the total live population of cells (Fig. 2C; Supplementary Fig. S2C). E-cadherin expression was increased in the ALDH⁺ and Sca1⁺CD90⁻ cells and decreased in Sca1⁻CD90⁺ relative to total live cells, whereas vimentin was increased in Sca1⁻CD90⁺. We also treated PyMT, Met1, and D2.0R cell lines with 3 ng/mL TGFβ1 for 1 week and performed flow cytometry to measure CSC and EMT markers (Supplementary Fig. S4). We found TGFβ1 treatment increases the %Sca1⁻CD90⁺ of live cells and decreases the %Sca1⁺CD90⁻ of live cells for all cell lines (Supplementary Fig. S4E and S4F). Additionally, TGFβ1 treatment increases the overall proportion of cells that are N-cadherin⁺ (Supplementary Fig. S4G), increases the proportion of N-cadherin⁺ (Supplementary Fig. S4H) and vimentin⁺ (Supplementary Fig. S4I), and decreases EpCAM⁺ cells (Supplementary Fig. S4J) in the Sca1⁻CD90⁺ population relative to all live cells. Altogether, these results suggest that ALDH⁺ and Sca1⁺CD90⁻ cells are more epithelial and that Sca1⁻CD90⁺ cells are more mesenchymal. We next compared the sphere formation efficiency and proliferation of each population after FACS sorting. Sphere forming efficiency was higher for all populations relative to ALDH⁻CD90⁻Sca1⁻ cells (Fig. 2C; Supplementary Fig S2D), indicating that CSC properties are enriched across all CSC populations tested. Additionally, we found that LRC also demonstrated increased sphere formation relative to non-LRC (Fig. 2E). We also monitored the proliferation of each sorted population in normal growth media conditions (10% FBS) by measuring the percent confluence overtime with the IncuCyte imaging system (Supplementary Fig. S2E). Consistent with other results, ALDH⁺ cells had the fastest proliferation, whereas Sca1⁻CD90⁺ cells had the slowest proliferation. Finally, we performed limiting dilution, tumor initiation assays for ALDH⁺/⁻ and bulk cells as the gold standard for CSC activity. We found the TIC frequency was highly dependent on the mouse strain used, though ALDH⁺ had the highest TIC frequency in NOD scid and NSG mice (Supplementary Fig. S5A–S5Dc). We were unable to perform these experiments with CD90⁺ cells as we found the antibodies labeling the cells enhanced NK cell killing of tumor cells via antibody-dependent cellular cytotoxicity in vitro and in vivo.

After determining that PyMT, Met1, and D2.0R cell lines display metastatic dormancy in vivo after intracardiac inoculation, we next investigated the importance of the immune system in this phenotype. We performed intracardiac inoculation of each cell line in hosts differing in immunocompetence: fully immunocompetent (FVB/N for PyMT and Met1, BALB/c for D2.0R) as well as immunodeficient NOD scid or NSG mice (Fig. 3A). NOD scid mice lack T and B cells but have NK cells and macrophages, which may be slightly less functional than in the syngeneic host (16). NSG mice lack T, B, and NK cells. Relative to NOD scid mice, the primary alteration is the loss of NK cells because of the added mutation of the IL2 receptor gamma chain (17). For PyMT, we found a significant decrease in survival between FVB/N and NOD scid mice and NOD scid compared with NSG (Fig. 3B). This suggests that metastatic dormancy in PyMT is controlled partly by the adaptive immune system and partly by an IL2 receptor-dependent population, likely NK cells. Interestingly, some fully immunocompetent FVB/N mice inoculated intracardiac with 100,000 PyMT tumor cells never succumbed to metastatic disease. We hypothesize that this is a result of a T-cell–specific response against PyMT as a neoantigen, which has been reported by DeVette and colleagues for autochthonous into a naïve FVB/N mouse (2). For Met1, we found a significant decrease in survival between FVB/N and NOD scid and NSG mice but no difference between NSG and NOD scid (Fig. 3C), suggesting that only the adaptive immune system plays a role in Met1 dormancy. For D2.0R, we found a significant decrease in survival between BALB/c and NSG but no significant difference between NOD scid and BALB/c (Fig. 3D), suggesting that the adaptive immune system is not required for metastatic dormancy of D2.0R, but rather a population of immune cells that is IL2 receptor-dependent controls dormancy. Altogether, these findings indicate that the immune system plays a role in metastatic dormancy in all three syngenetic mouse models, but the specific immune cell populations that control dormancy differ between models.

We focused on the D2.0R model in which all mice succumbed to metastasis after a period of dormancy. To confirm that survival times differ between NOD scid and NSG for mice injected with D2.0R cells, we used CBG luciferase D2.0R cells. We injected 100,000 luciferase-expressing D2.0R tumor cells into the left ventricle of NOD scid or NSG mice and measured bioluminescence over time using IVIS (Fig. 3E). The bioluminescent signal was initially the same but diverged between NOD scid and NSG starting at 14 days and maintained separation through 56 days following injection (Fig. 3F; Supplementary Fig. S6). This indicates that initial clearing of tumor cells does not change in NOD scid compared with NSG, but rather their ability to grow after metastatic seeding is impaired in NOD scid mice. We also confirmed that there was a significantly shorter survival time for NSG mice relative to NOD scid for mice injected with D2.0R-CBG (Fig. 3G). These results confirm that an IL2 receptor-dependent population of immune cells controls metastatic dormancy of D2.0R cells.

To determine which immune cell population was responsible for these differences in dormancy, we determined the effect of depletion of specific immune cell populations. We hypothesized that NK cells were the IL2 receptor-dependent population because they are present in NOD scid but absent in NSG mice. However, others have reported that macrophage function may also be altered by the loss of IL2 receptor in NSG mice (18). We tested the importance of NK cells and macrophages via depletion with anti-ASGM1 (Fig. 4A) or Clodrosome, respectively. Our depletion strategy was successful in both cases (Supplementary Fig. S7), and we measured tumor growth by bioluminescent signal over time in depleted and control mice. Signal was significantly increased in mice depleted of NK cells (Fig. 4B), and their survival time was significantly shorter (Fig. 4C). However, for mice with macrophage depletion (Fig. 4D, Clodrosome ), there was no significant difference in bioluminescent signal over time relative to control (Fig. 4E) nor in survival times (Fig. 4F). These results indicate that NK cells are primarily responsible for the metastatic dormancy phenotype in the D2.0R model.

We next conducted NK cytotoxicity assays to examine the sensitivity of CSC populations to NK cells in vitro. We first validated that all cell lines are sensitive to NK cell killing via comparing cytotoxicity among all cell lines to the model NK-sensitive cell line Yac1 (Supplementary Fig. S8A). We also tested the impact of NK cell coculture on colony formation and found that D2A1, D2.0R, and PyMT cell lines had reduced colony formation after treatment with NK cells but Met1 did not (Supplementary Fig. S8B). This result is consistent with the fact that there was no significant difference in survival time for NOD scid versus NSG mice inoculated with Met1 cells. Next, we performed indirect cytotoxicity assays (Fig. 5A), in which we cocultured D2.0R cells and NK cells and compared the populations relative to monoculture. In the direct assays, we sorted specific CSC populations via FACS and then cocultured them with NK cells. We first stained untreated (target only) or NK treated (target + NK) for CSC markers (Fig. 5B) Sca1⁺CD90⁻ cells were significantly decreased, suggesting that this population is sensitive to NK cells. We also stained for EMT markers and found EpCAM-expressing cells decreased, whereas vimentin-expressing cells increased upon treatment with NK cells (Fig. 5C). Finally, LRC percentage was significantly enriched when D2.0R cells were cultured with NK cells (Fig. 5D; see Supplementary Fig. S9 for gating scheme). Altogether, the indirect assays suggest that Sca1⁺CD90⁻, epithelial, proliferative D2.0R cells are more sensitive to NK cells than their Sca1⁻CD90⁺, mesenchymal, quiescent counterparts.

To confirm these results, we performed assays directly measuring the sensitivity of D2.0R cells in 2D or sphere culture (3D; Fig. 5E). We previously identified that low-serum 2D culture enriches Sca1⁻CD90⁺ cells relative to sphere culture, due to the low-serum conditions enriching for quiescent cells (Fig. 2A). We found that cells cultured in 2D were less sensitive to NK lysis than those cultured as spheres (Fig. 5F). We next sorted ALDH⁺ or ALDH⁻ D2.0R cells from 2D culture conditions and observed a significant reduction in cytotoxicity for ALDH⁺ cells relative to ALDH⁻ (Fig. 5G). We also sorted bulk (mixed population), LRCs, and non-LRCs and found that non-LRCs were significantly more sensitive to NK lysis (Fig. 5H). Finally, we sorted non-LRC ALDH⁻, non-LRC ALDH⁺, LRC ALDH⁻, and LRC ALDH⁺ D2.0R cells to investigate the intersection of stemness and quiescence phenotypes (Fig. 5I). We found non-LRC ALDH⁻ cells were more sensitive to NK cell-driven cytotoxicity than all other populations. This indicates that stemness and quiescence result in an NK cell-resistant phenotype. In summary, these results support the indirect assays and suggest that Sca1⁺CD90⁻, epithelial, proliferative non-LRC D2.0R cells are more sensitive to NK cells than their Sca1⁻CD90⁺, mesenchymal, quiescent LRC counterparts, though stemness and quiescence are important for NK resistance.

We next determined whether the findings from our in vitro experiments also apply in the in vivo setting. We injected CTFR-labeled D2.0R cells into the left ventricle of BALB/c mice and isolated bone marrow after 7 days (Fig. 5J). Using CD90 staining, we analyzed the overlap between CD90 expression and label retention in vivo. We compared the CD90 expression in live tumor cells to non-LRC and LRC populations and found a significant increase in CD90 expression in the LRC population relative to live and non-LRC population (Fig. 5K). This result supports our in vitro findings, indicating that the CD90⁺ population is predominately label-retaining in vivo. We also investigated the resistance of LRC D2.0R cells to NK cells in vivo. We sorted non-LRCs and LRCs using FACS and inoculated them into the mammary fat pads of either NOD scid or NSG mice and monitored tumor growth using IVIS (Fig. 5L). We found no significant difference in the growth rate of non-LRC versus LRC in NSG mice (Fig. 5M). However, in NOD scid mice, which have NK cells, LRCs demonstrate a significant growth advantage relative to non-LRCs (Fig. 5N). This finding suggests that LRCs are more resistant to NK cell-mediated lysis in vivo compared with non-LRCs.

To identify the mechanisms linking dormancy to NK cytotoxicity in a nonbiased manner, we performed RNA-seq of LRC and non-LRC. A total of 4,199 genes were differentially expressed (adjusted P value < 0.05) between the two cell populations (Fig. 6A), 279 of which had an absolute log₂ fold change > 2. Upon hierarchical clustering, data were clustered by condition and not by batch (Fig. 6B). GSEA demonstrated that pathways related to inflammation, including neutrophil extracellular trap formation, cytokine–cytokine receptor interaction, complement and coagulation cascades, allograft rejection, and others, were enriched in the LRC population, whereas in the non-LRC population, enriched pathways were related to cell cycle, including cell cycle and DNA replication (Supplementary Fig. S10). Activating NK ligands were more highly expressed in non-LRCs (Fig. 6C), whereas MHC genes were decreased in non-LRCs (Fig. 6D). The decrease in NK activating ligands and the increase in MHC gene expression in LRC are consistent with the relative LRC resistance to NK cell cytotoxicity. We validated these results via flow cytometry and found the NKG2D ligand RAE1 to be partially associated with the sensitivity of proliferative D2.0R tumor cells to NK cell killing in vitro and in vivo (Supplementary Fig. S11). Interestingly, STING and STING target expression were increased in non-LRCs (Fig. 6E). These results are consistent with the phenotype we observed in vitro and in vivo that non-LRCs are more sensitive to NK cell cytotoxicity than LRCs.

We next investigated transcription factor drivers of these gene expression programs using ChEA3 analysis (19). We compared the predicted transcription factors for the genes that were upregulated in LRCs to the predicted transcription factors for MHC genes, complement genes, and genes associated with EMT and murine mammary stem cells. We reasoned that the overlap of these gene sets should identify transcription factors that are responsible for general differences in gene expression, as well as for the increase in MHC gene expression and EMT/stem-like phenotype. Bach1 was the top ranked transcription factor within the intersection of these gene sets (Fig. 6F). Additionally, Sox2 was the fifth ranked transcription factor, and as Bach1 has been shown to regulate Sox2 via increasing its stability (20), we chose to investigate these genes further.

To confirm these transcription factors as drivers of NK resistance and sensitivity, respectively, we knocked these genes down using a Tet-On inducible shRNA system in D2.0R-CBG cells. We found knockdown of Bach1 or Sox2 did not alter proliferation or cell eccentricity (Supplementary Fig. S12). We confirmed that knockdown of Bach1 reduces Sox2 protein levels (Fig. 7A and B), as reported by Wei and colleagues (20). We also tested whether knockdown of Bach1 or Sox2 altered the sensitivity of D2.0R cells to NK cell cytotoxicity. We found a small but statistically significant increase in cytotoxicity after knockdown of Bach1 and a trend toward increased NK cytotoxicity after knockdown of Sox2 (Fig. 7C). Finally, we inoculated these cells into the mammary fat pads of NOD scid or NSG mice and treated mice with either control water or doxycycline containing water to investigate the difference in tumor growth. We found knockdown of Bach1 and Sox2 only decreased tumor growth in NOD scid mice with NK cells and not in NSG mice without NK cells (Fig. 7D–I). This finding confirms that Bach1 and Sox2 expressions contribute to the NK-resistant phenotype of D2.0R cells.

Finally, as we found low STING and STING target expression in LRCs compared with non-LRCs, we investigated the STING agonist MSA-2 as a method to reverse the NK-resistant phenotype. MSA-2 was recently reported by the Massagué group to prevent reactivation from dormancy (21). MSA-2 significantly enhanced NK cell killing in vitro at 16 to 24 hours upon overnight cotreatment of tumor cells and NK cells (Fig. 8A), or by overnight pretreatment of NK cells alone (Fig. 8B), but not by overnight pretreatment of tumor cells alone (Fig. 8C). We reasoned that there were two possibilities for why we did not observe an effect of the STING agonist pretreatment of tumor cells. First, either overnight treatment with MSA-2 was not sufficient to alter tumor cell phenotype or treating the bulk population (LRC and non-LRC) was not sufficient to upregulate STING to a level in which it would impact NK cell killing because non-LRCs already have high STING activity. To test these hypotheses, we first performed overnight pretreatment of tumor cells alone but then followed the coculture of tumor cells and NK cells over time (Fig. 8D). We found no significant difference in NK cytotoxicity after 24 hours of coculture between untreated and MSA-2–treated tumor cells; however, after 48 hours, we could see significantly more cytotoxicity for MSA-2–treated tumor cells relative to control. This indicates that the STING agonist treatment has a lag phase before tumor cell phenotype is sufficiently shifted to impact sensitivity to NK cells. With this finding in mind, we then tested the second hypothesis that non-LRCs and LRCs may have differential sensitivity to STING agonist treatment. We sorted non-LRCs and LRCs via FACS and then cultured them for 1 week in varying concentrations of MSA-2 (Fig. 8E). Interestingly, we observed that even after 1 week of culture after FACS and before coculture with NK cells, LRCs still maintained the NK-resistant phenotype relative to non-LRCs. Furthermore, we observed a significant increase in NK cell cytotoxicity for 10 μmol/L and 33 μmol/L MSA-2 treatment for both populations. Thus, we conclude that the STING agonist MSA-2 can increase tumor cell sensitivity to NK cells via a tumor cell-intrinsic effect and by impacting NK cell phenotype, which has been reported previously (22). We subsequently found that a single dose of MSA-2 given 7 days after inoculation significantly increased mouse survival time (Fig. 8F). This dramatic effect of a single dose is consistent with the dual effect of MSA-2 on tumor cell sensitivity and NK cell activity.

Understanding breast cancer dormancy has emerged as a major research focus in the last decade (23). Although much of this research has focused on identifying tumor-intrinsic mechanisms of breast cancer dormancy, some studies have explored the role of the immune system. The lack of immunocompetent models poses a major challenge in this research effort. Human tumor cells inoculated into immunocompromised hosts have been used in many studies, but the species mismatch between the host and the tumor cells is a major limitation. Even so, the tissue microenvironment is a well-established regulator of metastatic breast cancer dormancy including the immune microenvironment and the perivascular niche in bone marrow. Although interactions with innate immune cells are still present in this model and may be crucial for the dormancy phenotype, it does not reflect the clinical setting. In this study, we evaluated the metastatic dormancy phenotype of five murine mammary carcinoma cell lines and identified three lines that exhibited long-term metastatic dormancy, with survival exceeding 100 days following the intracardiac inoculation of 100,000 tumor cells.

CSCs have been extensively studied in human breast cancer cell lines because their initial identification by Al Hajj and colleagues (24) in 2003. However, CSCs have been less well studied in murine models. One reason for this is that the CSC hypothesis was called into question after a report from Sean Morrison’s group demonstrated that, in melanoma, the tumor initiation ability of a cell population was highly dependent on the immune background used (25). Although only a subpopulation of melanoma cells was capable of tumor initiation in NOD scid mice, NSG mice without NK cells were able to form tumors from single melanoma cells, calling into question the relevance of the cancer stem cell hypothesis. However, our present study and others have now shown that immune evasion is an important characteristic of CSCs and stress the importance of using immunocompetent models to generate clinically relevant data. Two CSC populations have been reported in PyMT: a tumor-initiating (ALDH⁺) and a metastasis-initiating (CD24⁺CD90⁺) population (26). We chose CD90 and ALDH as markers to investigate for all cell lines because of this report. We also chose Sca1 as it was reported to be a CSC marker in TUBO, PyMT, 4T1, and other murine cell lines (12). Our findings indicate that the ALDH⁺ and Sca1⁺CD90⁻ populations were more proliferative and epithelial, whereas the Sca1⁻CD90⁺ population was more quiescent and mesenchymal, in agreement with the report in PyMT (26) and the dichotomy seen in human breast cancer cells (15).

We tested three different mouse breast cancer cell lines for their dependence on the immune system (innate or adaptive) and found three different phenotypes. The fact that we observed different immune dependencies for each cell line suggests that there are many ways to get to the same metastatic dormancy phenotype. Further, for women with long latency periods and then late recurrence, the same immune cell population may not have kept the tumor cells in check. This implies that to prevent metastatic recurrence, we must identify the vulnerabilities of tumor cells to the immune system and maintain these relationships.

NK cells play a crucial role in determining the outcome of metastasis (27). The earliest studies on metastasis by Hanna and colleagues (4) in 1981 showed that NK cells cleared tumor cells after intracardiac inoculation. This classic study has recently been revisited using modern techniques such as bioluminescence and intravital imaging (28). The importance of NK cells in metastatic dormancy has also been established. Malladi and colleagues (5) found that human dormancy-competent cell lines required NK cells for the dormancy phenotype. However, one disadvantage of using human tumor cells in mouse models is that there is a species mismatch between human tumor cells and mouse NK cells. This mismatch can alter the strength of activating and inhibitory receptor interactions and may result in reduced activation or inhibition (29). The interactions between dormancy-competent cell lines and NK cells were validated using human NK cells and tumor cells in vitro (5), but this approach does not replicate the in vivo metastasis environment. In this report, we used syngeneic interactions between D2.0R tumor cells and NK cells in vitro and in vivo to validate the importance of NK cells in D2.0R metastatic dormancy.

Goddard and colleagues (30) have recently very elegantly described a concept they term “relative scarcity,” in which disseminated tumor cells in the lung are not controlled by the inability of tumor cells to be recognized by T cells but rather by the lack of interactions between DTCs and T cells. This work also uses the D2.0R cell line but has a few important differences to the report here including delivery of tumor cells via tail vein instead of intracardiac and the use of GFP and luciferase-expressing tumor cells in a fully immune competent mouse. Further work in this system is required to investigate if relative scarcity is also at play in the bone marrow microenvironment after intracardiac inoculation of D2.0R cells that do not express GFP and/or luciferase.

We found that the quiescent, mesenchymal D2.0R CSC population is resistant to NK lysis, whereas the more epithelial, proliferative D2.0R non-CSC population is sensitive to NK lysis. Interestingly, stem-like ALDH⁺ proliferative D2.0R cells are similarly resistant to NK lysis to quiescent D2.0R cells. This indicates that stemness and quiescence can result in an NK cell-resistant phenotype and provides a mechanism to explain the emergence of metastasis in this model. This is the first report to investigate the sensitivity of D2.0R cells to NK cells and is consistent with previous literature that showed that TUBO spheres are sensitive to NK lysis, whereas cells grown in 2D are resistant (13). In contrast, there are multiple reports that studied the sensitivity of human breast cancer CSC to NK lysis. Human BC has a quiescent, mesenchymal CSC population (CD24⁻CD44⁺) and a proliferative, epithelial CSC population (ALDH⁺; ref. 15). Three reports tested the sensitivity of the CD24⁻CD44⁺ population to NK cells, and two of the three found they were more resistant to NK cells than their differentiated counterparts (31–33). Similarly, three reports tested the sensitivity of the ALDH⁺ population to NK cells, and two of the three found they were more sensitive to NK cells than their differentiated counterparts (13, 34, 35). Moreover, coculture of NK cells and breast cancer cells increases the mesenchymal stem-like population (36). Interestingly, EMT and stemness have been shown to decrease the sensitivity of tumor cells to cytotoxic T cell and NK lysis (37). This phenomenon may hold across mouse and human breast cancers, but for other cancers, there is some evidence that EMT enhances sensitivity to NK cell killing and that stem-like cells are more sensitive to NK lysis than their differentiated counterparts (38). Thus, it is not yet clear how widely applicable this model may be.

One limitation of this work is the focus on short-and long-term dormancy rather than investigation of the entire time course. Future work will require more established connections between the short-/long-term dormancy phenotypes and will investigate the emergence of macrometastasis in various organs over time to determine differential tissue regulation of metastasis.

In this report, we have identified three immunocompetent models of metastatic dormancy with varying immune dependencies. We have conducted a detailed investigation of the D2.0R model and found that metastatic dormancy in this model is dependent on NK cells, with different sensitivities of quiescent mesenchymal CSCs compared with proliferative epithelial CSCs. To prevent disease recurrence, it may be possible to target the mechanisms that lead to quiescent CSC resistance to NK cells. In addition, we demonstrate that this quiescent CSC population is sensitive to STING agonists. The utilization of immunocompetent mouse models may facilitate the development of strategies to effectively target dormant CSC reducing metastatic recurrence, a major cause of breast cancer mortality.

G.G. Bushnell reports grants from the National Institutes of Health during the conduct of the study. M. Burness reports grants from Pfizer and GSK during the conduct of the study and grants from Cyteir outside the submitted work. M.S. Wicha reports grants from NCI and Breast Cancer Research Foundation during the conduct of the study. No disclosures were reported by the other authors.

G.G. Bushnell: Conceptualization, formal analysis, funding acquisition, investigation, visualization, methodology, writing–original draft, writing–review and editing. D. Sharma: Formal analysis, validation, investigation, writing–review and editing. H.C. Wilmot: Validation, investigation, writing–review and editing. M. Zheng: Validation, investigation, writing–review and editing. T.D. Fashina: Validation, investigation, writing–review and editing. C.M. Hutchens: Formal analysis, validation, investigation, visualization, writing–review and editing. S. Osipov: Investigation, writing–review and editing. M. Burness: Project administration, writing–review and editing. M.S. Wicha: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.

We acknowledge support from the National Institutes of Health (NIH R35 CA197585) to M.S. Wicha, Breast Cancer Research Foundation (BCRF-18-173) to M.S. Wicha, and NIH K99 CA267261 to G.G. Bushnell.