MicroRNA-92 Expression in CD133⁺ Melanoma Stem Cells Regulates Immunosuppression in the Tumor Microenvironment via Integrin-Dependent Activation of TGFβ

Primary melanomas have been reported to harbor subpopulations of tumor cells with intrinsic self-renewal and proliferative capacity termed as cancer stem cells (CSC; ref. 1). The CSC theory may help explain the plastic, chemoresistant, and invasive nature of refractory melanomas. Several biomarkers have been utilized in the identification and isolation of melanoma CSCs including CD20 (1), aldehyde dehydrogenase (2), CD133 (3), and ABCB5 (4). The murine melanoma cell line B16-F10 was recently shown to contain a distinct subset of cells expressing CD133 that had long-term tumorigenic potential and highly expressed the stem cell markers Oct4, Nanog, and Sox10 (5). In this study, we utilized CD133⁺ B16-F10 cells to explore the intricate interactions between immune cells and CSCs to determine how specific subpopulations may drive immunosuppression in the tumor microenvironment (TME).

Immunosuppression can be mediated through several immune cell phenotypes including regulatory T cells (Treg), myeloid-derived suppressor cells (MDSC), and alternative macrophages (M2; ref. 6). Many cancers, including melanoma, exploit these immune cell phenotypes to secrete cytokines and growth factors that create a permissive environment for cancers to proliferate and eventually metastasize. TGFβ is a pleiotropic cytokine with robust immunosuppressive activity including the ability to repress T-cell activation and proliferation (7). Part of this immunosuppressive effect can be carried out by Tregs, which produce abundant TGFβ to modulate immune response to self and foreign antigens (extensively reviewed in ref. 8). However, TGFβ is secreted in an inactive form and must undergo activation to stimulate downstream signaling cascades through binding of the TGFβ receptor (TGFBR; ref. 9). One mechanism for converting latent TGFβ to its active form is through interactions with RGD-recognizing integrins (i.e., integrin α_v), which associates with latent TGFβ-binding proteins (LTBP) and the TGFβ prodomain to free TGFβ via mechanical shearing (9). TGFβ activation and subsequent signaling through its receptor has been associated with immune evasion (10), epithelial-to-mesenchymal transition (EMT; ref. 11), and tumor cell invasion (12); thus, targeting of TGFβ signaling in cancer remains a priority (13, 14).

miRs are small (20–30 nucleotide) noncoding RNAs that generally function to suppress gene expression by targeting the 3′ UTR of mRNAs, several of which have been demonstrated to regulate cellular functions pertinent to oncogenesis and tumor progression (15). miR-92, a member of the miR-17-92 cluster, has been reported as both an oncomiR (16, 17) as well as a tumor suppressor (18, 19) depending on the cancer model. Importantly, miR-92 was shown to regulate expression of integrin α₅ in an ovarian cancer model (20). Remarkably, the role of miR-92 in melanoma has yet to be explored.

Our study reveals, for the first time, that miR-92 may regulate an integrin-mediated axis driving TGFβ-induced immunosuppression in the TME. Furthermore, this axis may confer a selective survival advantage to CSCs present within the heterogeneous tumor population by modulating immunosuppression and exploiting immunosuppressive cell phenotypes such as Tregs and M2 macrophage populations present within the TME. These studies shed light on the biological function of CSCs in the context of immune surveillance and also provide a potential therapeutic target in refractory melanomas in which CSCs may contribute to patient relapse.

The B16-F10 cell line was obtained from ATCC. All cell lines were grown in DMEM supplemented with 10% heat-inactivated FBS (Atlanta Biologicals), penicillin (100 U/mL, Gibco), and streptomycin (100 μg/mL, Gibco). Cells were incubated at 37°C at 5% CO₂ and subcultured every 72 hours. Routine monitoring for Mycoplasma contamination was performed using the MycoAlert Detection Kit (Lonza #LT07-218). Cells recovered from frozen aliquots were allowed one passage to reach exponential growth phase following recovery before being used in this study. Cells at passages greater than ten were not used in the experiments performed in this study. CD133⁻ and CD133⁺ cells were isolated by FACS and were grown in DMEM/F-12 serum-free media (SFM) containing 1 × N-2 Supplement (Gibco #17502-048) 10 ng/mL basic fibroblast growth factor (PeproTech #450-33), and 10 ng/mL EGF (PeproTech #315-09) in low-cluster 6-well plates (Corning #3471).

B16-F10 cells were grown as nonadherent oncospheres in SFM as described previously (21). After 7–10 days of culture in low-cluster plates, oncospheres were dissociated into single-cell suspensions and labeled using a PE-conjugated CD133 antibody [(BioLegend #141204) in 100 μL of staining buffer (2% FBS/2 mmol/L EDTA in PBS)] at a dilution of 1:100. The appropriate isotype control (BioLegend #400508) was used to gate the CD133⁻ and CD133⁺ populations. Cells were sorted using a BD FACSAria II into 15-mL conical collection tubes containing approximately 10 mL of ice-cold PBS at 4°C. Representative histograms demonstrating our gating strategy and postsort purity have been provided (Supplementary Fig. S1). After sorting, cells were centrifuged at 300 × g for 10 minutes, resuspended in an appropriate amount of PBS, and counted by trypan blue exclusion assay on a BioRad TC20 Automated Cell Counter before use in subsequent assays.

Primary tumors generated from subcutaneous injection of B16 cells or lungs from metastasis-bearing mice intravenously were dissociated using a Tumor Dissociation Kit (Miltenyi Biotec #130-096-730) to dissociate whole-tumor tissues into single-cell suspensions following the manufacturer's recommended protocol. The resulting cell suspensions were washed with and resuspended in PBS prior to initiating labeling with antibodies. Single-cell labeling with fluorophore-conjugated primary antibodies against CD45 (BioLegend #103116), CD3 (BioLegend #100306), CD4 (BioLegend #100453), CD8 (BioLegend #100708), NK1.1 (BioLegend #108748), FOXP3 (BioLegend #126419), IL10 (BioLegend #505031), TGFβ (BioLegend #141410), IL17 (BD #564168), IFNγ (BD #563854), CD11b (BioLegend #101222), F4/80 (BioLegend #123110), CD11c (BioLegend #117334), GR1 (BioLegend #108457), Ly6C (BD #560595), Ly6G (BD #560603), and CD206 (BioLegend #141723) was performed for at least 30 minutes on ice, washed with staining buffer, and subsequently analyzed on a BD FACSCelesta Flow Cytometer equipped with BD DIVA software in conjunction with FlowJo software. For intracellular labeling against transcription factors and cytokines (i.e., FOXP3), cells were fixed and permeabilized using the True-Nuclear Transcription Factor Buffer Set Kit (BioLegend #444201) following the manufacturer's recommendations. Data were compensated using BD CompBeads (anti-mouse #552843, anti-rat/hamster #552845), labeled with single antibodies or isotype controls, and analyzed using FlowJo. Spanning-tree Progression Analysis of Density-normalized Events (SPADE) V3.0 (22) was used to down-sample and cluster similarly labeled populations of cells following compensation and gating in FlowJo. Compensated FCS 3.0 files were exported and analyzed using the standalone version of SPADE 3.0 using the following parameters: Arcsinh transformation = 150, maximum allowable cells in pooled data = 200,000, outlier density = 1, fixed number of remaining cells = 100,000, clustering parameter = K-means, and the desired number of clusters = 50. Determinations for phenotyping each node/cluster was carried out based on single-color controls and a representative figure is provided in Supplementary Fig. S2.

Briefly, CD133⁺ and CD133⁻ populations were isolated via FACS from the B16-F10 murine melanoma as described above. Total RNA was extracted (Qiagen, miRNeasy #74106) from B16-F10 cells sorted from three independent experiments. Each sample was individually analyzed for quantity (NanoDrop 2000, Thermo Fisher Scientific) and quality (BioAnalyzer 2100, Agilent). For miRNA microarray, aliquots from individual samples were pooled for each group (n = 3/CD133^+/−). All samples used for downstream analysis had an RNA integrity number of at least 8. RNA profiling from samples was performed using the FlashTag Biotin HSR RNA Labeling Kit for GeneChip miRNA Arrays for the Affymetrix GeneChip miRNA 4.0 array platform. Labeled and hybridized chips were scanned on a GeneChip Scanner (Affymetrix) and microarray image data were analyzed using Affymetrix Power Tools. Data analysis and generation of representative figures (i.e., scatter plot) were performed using the Transcriptome Analysis Console (TAC, Affymetrix). MiRNAs with a fold change greater than 1.5 or less than −1.5 were considered for further validation and analysis. Predicted targets and alignment scores for specific miRNAs were generated using online software including TargetScan Mouse 6.2 and miRDB. Ingenuity Pathway Analysis (IPA, Qiagen) in combination with MetaCore pathway analysis tools (Thomson Reuters) were used to generate potential gene networks associated with significantly altered miRNAs and generate miR-gene interactome pathway maps.

CD133⁺ and CD133⁻ B16-F10 cells were isolated by FACS, and total RNA was isolated using miRNeasy Kit (Qiagen), following the manufacturer's protocol. The expression of indicated mRNA and miRNA levels was determined by qRT-PCR. Total RNA was quantitated using a Nanodrop 2000 (Thermo Fisher Scientific). For miRNA expression analysis, cDNA was generated from total RNA using miScript II cDNA Synthesis Kit (Qiagen # 218161). Two-step miRNA qRT-PCR were carried out using SsoAdvanced SYBR Green Mix (Bio-Rad #1725270) with mouse primers for Snord96a (Qiagen #MS00033733), miR-669a-5p (Qiagen #MS0026222), miR-669l-5p (Qiagen #MS00043337), miR-466h-5p (Qiagen #MS00012201), and miR-92a-3p (Qiagen #MS00005971). Expression levels for miRNAs were normalized to Snord96a. For mRNA expression analysis, cDNA was made from total RNA using miScript II cDNA synthesis kit. A two-step amplification with a 60°C annealing temperature for qRT-PCR was carried out using SsoAdvanced SYBR Green Supermix from Bio-Rad with mouse primers for IL10, TGFβ1, TGFβ2, TGFβ3, Smad2, ITGB1, ITGB3, ITGA5, and ITGAV customized and ordered from IDT. All PCR experiments used a CFX96 Touch Real-Time PCR Detection System (Bio-Rad), and expression levels were normalized to β-actin mRNA levels. Fold changes were calculated using the 2^−ΔΔC_t method. Specific primers sequences are provided in Supplementary Table S1.

Cells were harvested and resuspended in RIPA (150 mmol/L NaCl, 1.0% IGEPAL CA-630, 0.5% sodium deoxycholate, 0.1% SDS, 50 mmol/L Tris, pH 8.0) buffer (Sigma #20-188) containing a protease inhibitor cocktail (Sigma #P8340) and PhosStop Phosphatase Inhibitor (Roche # 04906845001). Protein concentrations of cell lysates were determined by a Bicinchoninic Acid Assay (Thermo Fisher Scientific #23225) and 40–60 μg of total protein was loaded per lane on 10% Tris-Gly Gels (Bio-Rad #4561033), subjected to SDS-PAGE, and transferred to a nitrocellulose membrane using the iBlot System (Invitrogen). Lysates were probed with antibodies that recognize phosphorylated SMAD2 (Cell Signaling Technology, #8828S), total SMAD2 (Cell Signaling Technology, #5678S), β-Actin (Cell Signaling Technology, #4970S), Integrin β₁ (Cell Signaling Technology, #4749T), Integrin β₃ (Cell Signaling Technology, #4749T), Integrin α_v (Cell Signaling Technology, #4749T), and Integrin α₅ (Cell Signaling Technology, #4749T), and GAPDH (Cell Signaling Technology, #5174S). Densitometry and image analysis were performed using a ChemiDoc station equipped with ImageLab Software (Bio-Rad). Densitometry analysis of bands of interest from immunoblots was performed using ImageJ software.

B16-F10–sorted populations were isolated on the basis of CD133 positivity as described previously. Sorted cells were cultured in low-adherent 6-well plates (Corning) in SFM at a density of 1 × 10³ cells/mL. Cultures were grown for up to 10 days and amended with fresh SFM media twice per week. Oncospheres (>100 μm) were counted and imaged using an EVOS Light Microscope (Life Technologies) and images were analyzed using ImageJ Software (NIH, Bethesda, MD).

Female C57Bl/6 mice (Jackson #000644) were used at 6–8 weeks of age. All mice were handled in accordance with the American Association for Laboratory Animal Science guidelines with the approval of the appropriate Institutional Animal Care and Use Committees at the University of South Carolina (Columbia, SC; protocol no. 2371). Mice were injected subcutaneously with 1 × 10⁵ B16-F10 cells in PBS (100 μL). Tumor size was monitored three times weekly until animals were sacrificed because of tumor burden. Tumor volume [V = L × W² × (π/6)] was determined by measuring the greatest linear dimensions in length (L) and width (W).

For our experimental metastasis models, 2 × 10⁵ B16-F10 cells suspended in 100 μL PBS were injected intravenously into 6- to 8-week-old, female C57Bl/6 mice via the lateral tail vein. After approximately 14–16 days, mice were sacrificed. Upon sacrificing the mice, lungs were resected, imaged, dissociated, and labeled with antibodies for subsequent flow cytometry analysis.

In experiments involving in vivo growth of CD133⁺-transfected cells, mice were injected with CD133⁺ cells transfected with miR-92a mock (HiPerfect reagent only) or mimic (as described below), and tumor volume was measured. On day 15, mice were sacrificed, tumors were dissociated, and labeled with antibody panels for various immune phenotypes using flow cytometry.

In brief, CD133⁺ cells (1.5 × 10⁵/well in 0.5 mL) postsorting were cultured in 24-well plates at 37°C, 5% CO₂. The following day (24 hours postseeding), transfection was performed following the manufacturer's protocol. Seventy-five ng of miR-92a mimic, miR-92a inhibitor, or miR-92a mimic + inhibitor (to a final concentration of 10 nmol/L) were diluted in 100 μL of culture medium without serum. HiPerFect Reagent (4.5 μL; Qiagen, #301705) was added to the diluted miR-92a mimic, miR-92a inhibitor, or miR-92a mimic + inhibitor. The reagents were incubated for 10 minutes at room temperature to allow for the formation of transfection complexes. The complexes were added to their respective wells and subsequently mixed by pipetting to ensure uniform dilution of the transfection complexes. The culture medium was changed after 12–15 hours. Following the change in medium, cells were incubated for 72 hours at 37°C, 5% CO₂. The cells were collected 72 hours posttransfection and used for miRNA assays or gene expression. Primer assays and gene expression were determined by RT-PCR and are described in Material and Methods. Snord96a (#3150530, Qiagen) was used as an internal control for miR-92a expression and Actin (primer sequences provided previously) was used to normalize gene expression.

Sorted tumor populations were cultured alone or with freshly isolated whole-splenic cells at a 1:1 ratio (1 × 10⁶ total cells) in 100 μL of serum-free media for 24 hours in cell culture–treated 96-well plates (Corning #3595). The resulting supernatants were centrifuged at 400 × g to remove cells and debris, and frozen at −80°C until analysis. A free TGFβ precoated ELISA kit was used (BioLegend #437707) to determine the concentration of active TGFβ in each sample following the manufacturer's recommended protocol. Splenic cells from each well were isolated by centrifugation at 400 × g for 10 minutes and labeled using the fluorophore-conjugated antibodies described previously. Following labeling, cells were washed and resuspended in 500 μL of staining buffer for flow cytometric analysis.

Expression data from cutaneous melanoma samples contained in The Cancer Genome Atlas (TCGA) database were assessed and visualized using cBioPortal (23, 24) and UCSC Xena (http://xena.ucsc.edu/). Queried genes included ITGAV, ITGA5, and TGFB1. Correlations based on the Pearson coefficient are represented in each scatter plot, and represent mRNA expression data for each of the genes provided. Mutational status for each queried gene is also provided above each mRNA expression heatmap and explained in detail in the appropriate figure legend. Correlation analysis between miR-92a expression and the genes described above were also conducted using Stata 14 (StataCorp, 2015) from cutaneous melanoma datasets containing both miRNA and mRNA expression data available from the TGCA study. Simple linear regression analysis was performed to predict expression of integrin alpha subunits as a function of miR-92 expression using Stata. Figures reflecting these analyses are visualized by a scatter plot of the original data, the linear regression line, and the 95% confidence intervals of the regression line.

GraphPad Prism 5.0 Software (GraphPad Prism Software, Inc.) was used for all statistical analyses. For all in vitro studies, two-group comparisons between control and test samples were done by two-tailed Student t test and representative data from three independent experiments were presented. A one-way ANOVA was performed on in vitro experiments containing more than one group, and significance was determined and denoted for each group accordingly. Subcutaneous and experimental metastasis in vivo data were analyzed for significance using two-way ANOVA and a two-tailed Student t tests, respectively. For all tests, statistical significance was assumed when P < 0.05. P values were reported in each figure or in their respective figure legends.

To study the differential characteristics of CD133⁺ and CD133⁻ B16-F10 cells, we injected 2 × 10⁵ B16-F10 cells from each phenotype subcutaneously into C57Bl/6 mice. We observed a 58% increase in mean tumor volume and a 52% increase in mean wet tumor weight in the CD133⁺ group compared with CD133⁻ group. Not only did CD133⁺ cells form palpable tumors quicker than CD133⁻ cells, they were also more tumorigenic (Fig. 1A and B). CD133⁺ cells formed tumors in 6/6 mice, while CD133⁻ cells only formed tumors in 4/6 mice (Fig. 1B). Using in vitro functional assays, we observed that CD133⁺ cells had a higher propensity to proliferate, form colonies, and generate anchorage-independent oncospheres when compared with CD133⁻ cells (Fig. 1B–D). In colony-forming assays, CD133⁺ cells generated an average of 42.8 ± 8.4 colonies, while CD133⁻ cells only generated an average of 18.2 ± 3.0 colonies (Fig. 1C). We also observed a significant increase in the ability to form anchorage-independent oncospheres in CD133⁺ populations compared with CD133⁻ cells (Fig. 1D). Not only were CD133⁺ cells capable of generating significantly more floating spheres (106.8 ± 11.6) compared with CD133⁻ cells (41.8 ±10.7), but oncospheres from CD133⁺ cells were observed to be much larger in diameter (Fig. 1D), suggesting that anchorage-independent survival and proliferation was enhanced in the CD133⁺ population.

Following syngeneic transplantation of CD133⁺ or CD133⁻ B16 tumor cells into C57Bl/6 mice, we allowed palpable tumors to grow to approximately 1 cm³ before resecting the tumors. Following resection, tumors were enzymatically dissociated and labeled to determine the infiltrating immune cell phenotypes in CD133⁺ and CD133⁻-initiated tumors. The data from a representative experiment are shown in Fig. 1E–G and from multiple experiments are summarized in Fig. 1H. CD133⁺-initiated tumors were observed to not only grow faster and larger than those initiated by CD133⁻ cells, but were also observed to have higher abundance of tumor-infiltrating Tregs, granulocytic MDSCs (gMDSC; CD45⁺CD11b⁺GR1⁺Ly6G⁺), and M2 macrophages (CD45⁺CD11b⁺F4/80⁺CD206⁺; Fig. 1E–I). Infiltrating macrophages, identified as CD45⁺F4/80⁺CD11b⁺ were reduced within the TME of CD133⁺-generated tumors (0.53% of total CD45⁺ cells) compared with CD133⁻ tumors (1% of total CD45⁺ cells). No changes in pan-T cell (CD45⁺CD3⁺) or pan-MDSC (CD45⁺CD11b⁺GR1⁺) populations were observed. Concurrently, we observed a significant increase in T cells staining positive for TGFβ as well as IL17a in tumors initiated by CD133⁺ cells (Fig. 1E). TGFβ⁺ T cells increased from 13.6% in CD133⁻ tumors to 25.2% in tumors generated by CD133⁺ cells. We also observed that Helios⁺ Tregs were significantly increased in CD133⁺ tumor samples increasing from 4.9% in CD133⁻ tumor samples to nearly 10% in CD133⁺ tumors. Although we did not observe a difference in macrophage or MDSC populations, we did find that tumors initiated by CD133⁺ cells had significantly increased proportions of infiltrating gMDSC (10.1% of total MDSCs in CD133⁺ group vs. 2.9% of total MDSCs in CD133⁻ group) and alternate macrophages (13.4% of total macrophages in CD133⁺ group vs. 8.9% of total macrophages in CD133⁻ group). To highlight the changes associated with tumor initiation between CD133⁺ and CD133⁻ cells, we have included a representative SPADE tree from pooled tumor samples (Fig. 1I). SPADE analysis shows a significant increase in immunosuppressive phenotypes including Tregs and cells staining positive for TGFβ or IL10. Taken together, these data suggest that more suppressive immune cells infiltrated CD133⁺ tumors allowing for superior tumor growth when compared with CD133⁻ cells.

We used an Affymetrix microarray to screen expression profiles of several thousand miRNAs in CD133⁺ and CD133⁻ B16-F10 cells. Analysis of data showed that of the 3,195 miRs screened, 2,995 miRs were common to these two cell types while 144 miRs were downregulated and 56 upregulated in CD133⁺ cells when compared with CD133⁻ cells (Fig. 2A). A comprehensive list of all miRNAs with greater than 2-fold change difference between samples is provided (Supplementary Table S2). The microarray also identified miRNAs of the miR-297-669 cluster to be downregulated in CD133⁺ cells including miR-669a-5p, miR-669l-5p, and miR-446h-5p, which was validated by qRT-PCR (Fig. 2A). The data confirmed that these miRNAs were, in fact, downregulated in CD133⁺ cells with relative expression levels (normalized to CD133⁻) of 0.49, 0.51, 0.57, and 0.49 for miR-466h, miR-669a, miR-669l, and miR-92a, respectively (Fig. 2A). Because the miR-297-669 cluster is not present in humans (but is conserved in rodents), further analysis of miRs in this cluster were not selected for further characterization. Assessment of miR-92 using Metacore (Fig. 2B) and Ingenuity (Fig. 2C) pathway analysis tools identified a network of genes associated with melanoma progression including CDC42, PTEN, and MAP2K (Fig. 2C) and have been identified for clarity by blue circles. With the recent discovery that miR-92 could regulate integrin subunit expression (20), we used predictive sequence alignment software to explore potential integrins, which may be targeted by miR-92. Integrin α_v and α₅ were highly predicted targets of miR-92 with weighted context scores of −0.28 and −0.34, and P_CT values of 0.92 and 0.90, respectively (Fig. 2D). With this highly predictive targeting of integrin subunits coupled with the mechanistic relationship between integrin activation of TGFβ, we further explored the relationship between miR-92a and TGFβ in an in vivo model.

The increase in TGFβ from dissociated primary tumors formed by CSCs, along with mechanical integrin-dependent TGFβ activation, lead us to initially look for miRNAs targeting integrins. Microarray analysis in combination with queried miR databases led to the identification of integrins as potential targets of miR-92a (Fig. 2D); thus, we next utilized qRT-PCR and immunoblot analysis to determine whether or not integrin expression was significantly increased in CD133⁺ cells when compared with CD133⁻ cells (Fig. 3). Normalized integrin mRNA expression was increased in CD133⁺ cells compared with CD133⁻ cells by 1.5-, 2.5-, 1.3-, 1.6-fold for integrin β₁, β₃, α_v, and α₅ subunits, respectively (Fig. 3A). Protein levels of integrin α subunits were also increased as assessed by immunoblot from two independent experiments in which B16 cells were isolated via FACS based on CD133 expression (Fig. 3B). qRT-PCR analysis of several isoforms of TGFβ identified TGFβ1 as the primary isoform responsible for the disparities in protein expression. No significant difference was observed in TGFβ3 between the two groups (Fig. 3A), and TGFβ2 was too lowly expressed to amplify using our parameters. All qRT-PCR experiments validating mRNA expression utilized β-actin as a housekeeping gene. Although qRT-PCR analysis determined a significant difference in SMAD2 mRNA expression, we did not observe any difference in protein level expression in our Western blot analysis; however, phosphorylated SMAD2, an indicator of TGFβ signal activation, was significantly induced in CD133⁺ tumor cells when compared with CD133⁻ cells (Fig. 3B). Full images of the exposed membranes have been provided (Supplementary Fig. S3).

When transplanted via tail vein injection, CD133⁺ cells created larger and more abundant lesions when compared with CD133⁻ cells (Fig. 4A). CD133⁺ cells generated an average of 41.0 ± 4.5 metastatic nodules compared with 22.6 ± 3.6 nodules from mice receiving intravenous injection of CD133⁻ cells. When dissociated and labeled for infiltrating immune cells, both phenotypes were able to induce immune cell infiltration of the pulmonary tissues; however, when compared with CD133⁻-initiated metastases, metastatic outgrowths initiated by CD133⁺ cells demonstrated a significantly higher proportion of immunosuppressive cell phenotypes including gMDSCs, and TGFβ⁺ Tregs (Fig. 4B and C). In myeloid panels, we observed a significant increase in gMDSCs with a marginal decrease in mMDSCs; macrophage populations did not exhibit any significant change in lung tissues from either tumor cell type (Fig. 4C). Spleens from each group showed a significant shift in mMDSCs as well as M1 macrophage populations (Supplementary Fig. S4). In the lymphocyte panel, we observed a significant increase in total T cell (CD45⁺CD3⁺) and NK cells (CD45⁺NK1.1⁺) infiltration along with a concurrent decrease in CD8⁺ T cells in the lungs of CD133⁺ cell–initiated mice when compared with mice in the CD133⁻ group (Fig. 4D). A significant increase in the percentage of TGFβ⁺ Tregs (CD45⁺CD3⁺CD4⁺FOXP3⁺) as well as IL17a⁺CD4⁺ T cells was observed along with a decrease in IFNγ⁺CD4⁺ T-cell populations in CD133⁺ when compared with CD133⁻ samples (Fig. 4D). The spleens of CD133⁺-transplanted mice also showed similar results with a significant decrease in IFNγ-producing cells as well as a concurrent increase in Treg populations (Supplementary Fig. S4). SPADE analysis (Fig. 4E) highlighted the differences in immune cell tumor infiltration between CD133⁺-initiated lesions (right) and CD133⁻-initiated lesions (left) in both myeloid (bottom) and lymphocyte (top) panels. Grayscale legends are provided for each panel to discriminate populations identified for each SPADE tree. These data suggested that tumors initiated by CD133⁺ tumor cells not only generated more immunosuppressive phenotypes within the TME, but also potentially resulted in less cytotoxic T-cell response as well.

After isolating CD133⁺ and CD133⁻ cell populations, we used a coculture model with splenic cells to determine whether either phenotype was able to readily polarize splenocytes to the immunosuppressive phenotypes observed in our dissociated tumor samples. Concurrently, we measured the concentrations of active TGFβ in coculture system as well as in splenocytes and tumors cells alone. Active TGFβ was significantly enhanced in cocultures using CD133⁺ and splenic cells (4.9 ± 0.3 and 6.6 ± 1.6 pg/mL at 8 and 24 hours, respectively) when compared with cocultures using CD133⁻ cells (2.3 ± 0.5 and 3.9 ± 0.9 pg/mL at 8 and 24 hours, respectively; Fig. 5A). Interestingly, we did not observe any differences in active TGFβ when CD133⁺ and CD133⁻ cells were cultured alone at either time point; although, samples from both tumor phenotypes contained significantly more free TGFβ than splenocytes cultured alone (Supplementary Fig. S5). When cells isolated from our cocultures were analyzed by flow cytometry, we observed a significant increase in CD11b⁺TGFβ⁺ cells in cocultures of CD133⁺ cells when compared with CD133⁻ cocultures (Fig. 5B–D). No significant changes were observed in T-cell populations (Fig. 5C and D). SPADE analysis of flow cytometric data is provided to further represent the shift in TGFβ-producing myeloid cells we observed in our samples.

To directly test whether miR-92a targeted the integrins, B16-F10 cells sorted on positive CD133 expression were transfected with miR-92a mimic, inhibitor, and a combination of mimic and inhibitor for 48 hours. Gene expression was measured by qRT-PCR for miR-92a, ITGAV, ITGA5, TGFB1, and SMAD2 following the transfection and isolation of total RNA; all groups were normalized to CD133⁺ mock-transfected cells as a reference. When transfected with a miR-92 mimic, we observed an 8.2-fold increase in miR-92a expression. Relative expression levels of ITGAV, ITGA5, TGFB1, and SMAD2 mRNA were 0.20, 0.17, 0.27, and 0.19, respectively (Fig. 6A). Conversely, transfection with an inhibitor of miR-92a decreased miR-92a expression, while significantly increasing mRNA levels for ITGAV, ITGA5, TGFB1, and SMAD2 (2.7, 4.7, 2.8, and 2.4-fold increases, respectively). This phenotype was partially rescued when both mimic and inhibitor were transfected into CD133⁺ B16 cells returning to baseline mRNA levels of ITGA5, TGFB1, and SMAD2. These results indicated that miR-92a was involved in the expression of genes that regulate TGFβ signaling and activation (Fig. 6A).

To test whether alterations in the expression of miR-92a in CD133⁺ cells would change tumor growth and immune response in vivo, CD133⁺ cells isolated from the B16 cell line were transfected with mock or miR-92 mimic. Transfected cells were next injected subcutaneously into C57Bl/6 mice and allowed to form tumors over 14 days (Fig. 6B). Upon endpoint of the experiment, tumors initiated by cells transfected with miR-92 mimic showed a significantly decreased tumor burden with an average tumor volume of 114.4 mm³. This represents a 58% decrease in tumor volume from the mean tumor volume of the mock-transfected CD133⁺ cells of 272.8 mm³. Although all mice in both groups formed tumors, the significant reduction in tumor volume suggested that miR-92a expression had an inverse association with tumor growth. It was also observed that cells transfected with miR-92 mimic showed a significant delay in tumor formation compared with CD133⁺ mock-transfected cells.

Tumor-bearing mice were sacrificed upon endpoint and tumors were dissociated, labeled with panels of antibodies against phenotypic markers for lymphocytes and monocytes, and analyzed by flow cytometry (Fig. 6C–F). In tumors initiated by miR-92/mimic, we observed a significant shift in tumor-infiltrating cells with an overall increase in total T cells (4.9% vs. 1.8%), but a decrease in pan-MDSC (3.0% vs. 1.8%) and pan-macrophage (1.7% and 0.5%) populations when compared with tumors initiated by the mock-transfected cells. When we analyzed subsets of Th cells (CD4⁺), we found that immunosuppressive phenotypes of FOXP3⁺ Tregs, IL10, and TGFβ-producing CD4⁺ cells were significantly decreased following miR-92 mimic transfection (17.8% vs. 4.3%; 34.5% vs. 11.0%; 28.3% vs. 5.7%, respectively). Antitumor proinflammatory phenotypes producing IFNγ were significantly increased from 0.5% to 2.3% when compared with tumor tissues from mock-transfected group.

To associate our findings in preclinical murine models of melanoma to clinical data, we used the publicly available TCGA database to authenticate our results in human samples. We specifically identified samples taken from patients with melanoma and observed a moderate positive correlation between integrin α₅ and TGFβ1 expression levels (Pearson coefficient = 0.66; Fig. 7A and B). Further analysis identified positive correlations between integrin α₅ and LTBP1 (Pearson coefficient = 0.67) as well as NRP-1 (Pearson coefficient = 0.69; Fig. 7B). These associations between groups of genes involved in TGFβ activation and signaling in human samples validated our preclinical studies in murine models and indicate that CSCs may, in fact, use these signaling molecules and their connected signaling networks to evade immune-mediated tumor ablation. We next identified datasets consisting of both miRNA and mRNA expression to explore the relationship between miR-92a and the genes involved in TGFβ activation (i.e., integrins). Using these publicly available data, we observed a moderate inverse association between integrin alpha 5 and alpha V subunits, and miR-92a using Spearman (−0.33 and −0.38, respectively) and Pearson correlation (−0.30 and −0.37, respectively) coefficients. Linear regression analysis and the resulting graphics are provided to help visualize the relationship described previously (Fig. 7); as miR-92a expression increased, we saw a subsequent reduction in integrin α_V and α₅ expression (Fig. 7C). Along with the scatter plot of the original data, the regression line from simple linear regression analysis, and 95% confidence intervals of the regression line were also provided.

The main goal of this study was to investigate whether CD133 expression on CSCs would alter the expression of miRNA, which would target genes involved in the regulation of immune response and consequently control tumor growth. On the basis of the expression of CD133, this study identified a potential role for miR-92 in regulating immunosuppression by mediating integrin-dependent TGFβ activation. We showed that CSCs based on the CD133 biomarker are functionally distinct from the bulk tumor population and demonstrate superior tumorigenicity and growth in vitro and in vivo. These results are in line with previous studies describing CD133 as a biomarker for CSCs present in melanomas (3, 25), and that these cell populations have intrinsic chemoresistant (26), angiogenic (27), and metastatic properties (28). CD133⁺ B16-F10 melanoma cells had enhanced tumor growth in subcutaneous tumor growth models and established larger and more abundant pulmonary lesions in our experimental metastasis model compared with CD133⁻ cells. Moreover, tumors and metastatic lesions initiated by CD133⁺ cells were significantly more immunosuppressed than CD133⁻-initiated tumors as assessed by intratumoral abundance of tumor-associated macrophages and Tregs. Indeed, CSCs in several cancer models have been reported to exploit immune cells to create a tumor-tolerant niche during tumorigenesis and metastasis (29–31). Infiltration of the TME by these immune cell phenotypes resulted in significantly higher production of TGFβ, shifts in MDSC populations toward a granulocytic phenotype, and lower IFNγ-producing cells in tumors initiated by CD133⁺ cells when compared with tumors initiated by CD133⁻ cells. Secretion and activation of TGFβ is a well-studied mechanism of immunosuppression and has been described in melanomas (32) as well as other models (33). Interestingly, blockade of TGFβ signaling in a B16-F10 melanoma model led to a T-cell–mediated eradication of tumors (34). Other studies have found that TGFβ blockade was sufficient for significantly increasing antitumor immune responses (35), and helps promote response to immune checkpoint inhibitors in different mouse models of melanoma (36, 37). In our model, we observed a significant increase in Tregs, a phenotype that can be induced by TGFβ (38, 39), in CSC-initiated tumors when compared with tumors generated from non-CSCs. These data indicate that tumor growth and metastasis by B16-F10 cells may be driven by Treg-mediated immunosuppression.

Using microarray technology, we identified several miRs that were downregulated in the CD133⁺ CSC population, which targeted TGFβ and its associated gene networks. Preliminary data indicated that several miRNAs from the miR-297-669 cluster were downregulated in CSCs compared with non-CSCs. Interestingly, it was recently reported that members of this cluster directly regulated TGFβ2 (40). Another miRNA identified by our microarray screen, miR-92a, is characterized as an oncomiR in various cancers (16, 41, 42) and is being employed as a potential serum biomarker for certain malignancies (43); however, miR-92 can also act as a tumor suppressor in other cancers (18, 44, 45). Downregulation of miR-92 in human breast cancers was associated with poor prognosis and correlated with stage and disease-free survival. Interestingly, the researchers also observed a significant increase in macrophage infiltration; however, the phenotype (M1/M2) of these tumor-infiltrating macrophages was not reported (46). Modulation of macrophage populations by B16-F10 cells has been linked to disease progression and metastasis (47), thus miR-92a may have far-reaching effects outside of regulating TGFβ signaling mechanisms. In fact, it was recently shown that miR-92 was acknowledged to alter miRNA profiles of induced pluripotent stem cells, an effect that was suggested to be p53 mediated (48). Additional studies have advocated a role for miR-92 in neuroblast self-renewal and maintenance (49). Interestingly, a recent report by Huber and colleagues described several miRNAs that were found in melanoma exosomes and mediated MDSC expansion and differentiation from CD14⁺ monocytes (50). Additional evidence from a study of glioma exosomes, which reported that miR-92a can stimulate the proliferation and function of MDSCs (51), indicates that exosome-mediated transfer of miRNAs may, in part, function to induce immunosuppression within the TME. The functional effects of miR-92 seem to be somewhat ubiquitous as well as tissue- and context-dependent; thus, more work exploring the functions of miR-92 in melanoma is justified.

CD133⁺ cells expressed higher integrin α₅ and α_v (RGD-recognizing subunits) on both a protein and mRNA level when compared with CD133⁻ cells. Integrin α₅ has been reported to be regulated by miR-92 in an ovarian cancer model (20); however, whether miR-92 regulates the α_v subunit has yet to be clarified. The integrin α_v subunit has been explicitly characterized to heterodimerize with integrin β subunits to form integrins α_vβ₃, α_vβ₅, α_vβ₆, and α_vβ₈, all of which have been reported to modulate TGFβ activation (52). Integrin α₅β₁ is the major receptor for fibronectin (53); fibronectin is required for TGFβ activation (54) and fibronectin matrix assembly (55), suggesting the unique possibility that integrin α₅ may also play a role in the liberation and activation of TGFβ. In endothelial cells, it was shown that fibronectin and its receptor (e.g., integrin α₅β₁) mediated SMAD phosphorylation following exogenous application of TGFβ1 and BMP-9 (56). Conversely, it was reported that TGFβ1 may regulate integrin α₅β₁ and integrin signal transduction (57). Further interactions between TGFβ1 and integrin α₅β₁ were reported in T cells where TGFβ-activated cells were protected from apoptosis by an integrin-dependent mechanism (58). qRT-PCR analysis of the integrin β₁ and β₃ subunits reflected higher mRNA expression in CD133⁺ cells compared with CD133⁻. Interestingly, the integrin β₃ was shown to regulate senescence through induction of the TGFβ signaling pathway (59). In addition, we observed an increase in TGFβ1 and activating phosphorylation of the downstream signaling molecule SMAD2. Our data demonstrate that CD133⁺ B16-F10 cells highly express components of integrin and TGFβ-associated signaling cascades, which may, in turn, provide a selective survival advantage in the context of immunosuppression within the TME. Mechanistically, this axis may be regulated through miR-92 modulation of integrin-dependent TGFβ activation.

Cocultures of splenic and tumor cells showed that while no significant change in active TGFβ concentrations was observed between CSCs and non-CSCs cultured alone, cocultures of CD133⁺ tumor and splenic cells resulted in significantly higher free TGFβ when compared with cocultures using CD133⁻ cells. Membrane-bound TGFβ was significantly increased in tumors resulting from CD133⁺ cells when compared with CD133⁻ cells in subcutaneous and intravenous models of melanoma; however, liberated TGFβ from each cell type remained unchanged in vitro. These data indicated that there are significant interactions between immune cells and tumor cells that collaborate to produce immunosuppression within the TME. In addition, they indicate that TGFβ secretion and activation may involve multiple mechanisms involving several cell phenotypes that are not recapitulated in the culture of cancer cell lines. After 24 hours, splenic cells from cocultures were labeled and analyzed by flow cytometry to determine whether any changes in myeloid cell or lymphocyte populations were stimulated by either tumor cell phenotype. Although we did not see a significant change in regulatory T-cell populations (as in our in vivo models), we did observe a significant increase in TGFβ⁺ myeloid cell populations in splenic cells cocultured with CSCs compared with non-CSCs. It was recently described that B16-F10 tumors undergo significant changes in immune cell infiltration depending upon the stage of disease (60). These studies suggest the intriguing concept that palpable tumor formation and initial immunosuppression may be driven by immunosuppressive myeloid cell types (i.e., alternate macrophages), while late-stage tumor growth and metastasis is controlled by Treg populations. In fact, an increased production of immature myeloid cells was observed in patients with cancer (61). In these studies, immature myeloid cells inhibited antigen-stimulated T-cell responses, which may help explain the significant decrease in CD8⁺ cytotoxic T lymphocytes in our CSC-induced metastasis models, and potentially clarify the disparities in tumor-associated immunophenotypes between our in vivo models and those characterized in our in vitro coculture models.

Finally, we characterized miR-92 to functionally modulate expression of integrin subunits as well as mediators of TGFβ signaling, and correlate these in vitro data with clinical data derived from the TGCA database. When mimics and inhibitors of miR-92a were transfected into B16 CD133⁺ cells, the resulting expression profiles supported our original hypothesis that miR-92 controls TGFβ-induced immunosuppression. Expression of ITGA5, ITGAV, TGFB1, and SMAD2 were all significantly affected when expression of miR-92a was altered. Mice receiving B16 cells with transfected miR-92a mimic showed significantly reduced tumor growth compared with mock-transfected B16 cells.

We wanted to correlate our studies with clinical data from the TCGA database in which thousands of human tumor samples have been characterized. Exploring melanoma samples within the database, we identified a positive correlation between integrin α₅ and TGFβ1, LTBP1, and neuropilin-1 (NRP-1). LTBP1 targets latent TGFβ complexes to the ECM where it is subsequently activated (62). NRP-1 was recently shown to modulate TGFβ signaling in glioblastomas (63). Interestingly, NRP-1 is a biomarker to distinguish natural and inducible Tregs (64), enforcing the idea that immunosuppression in melanomas might be T-cell–mediated. We also analyzed datasets containing both expression of mRNA and miRNAs. When association analyses were performed, we identified a moderate inverse correlation between integrin subunits and miR-92a as predicted by several databases, which aligned the sequences of miR-92a and integrin α_v and α₅. We further performed a simple linear regression model to predict integrin mRNA expression using miR-92a expression as an independent variable based on the available miRNA and mRNA expression data. These data clearly showed a reduction in ITGAV and ITGA5 expression with increased expression of miR-92a.

Our study categorized miR-92 as a potential tumor suppressor in melanoma by modifying TGFβ-induced immunosuppression. Disparities in miRNA expression between stem cell–like populations and bulk tumor cells may confer the properties attributed to CSCs such as chemoresistance and metastatic outgrowth. While integrins are clearly involved in adhesion, we demonstrated that CSCs express significantly higher levels of RGD-recognizing alpha subunits than non-CSCs. Furthermore, these disparities in expression may have resulted in altered TGFβ activation and subsequent immunosuppression. Given our initial data, further studies exploring the miR-92/integrin/TGFβ axis are warranted.

While our studies demonstrate a link between miR expression in tumor cells and their ability to grow in vivo and modulate immune cell phenotype in TME, it has some limitations and opportunities. Use of the B16-F10 cell line, although commonly used and easily accessible, may not mimic human disease as accurately as newer model systems including the Yale University Mouse Model (YUMM; ref. 65). The YUMM systems have fully characterized driver mutations, are genomically stable, and syngeneic to the C57Bl/6 mouse strain providing a much more comprehensive understanding of the underlying genetic interactions in progressing disease states. Our microarray analysis was only performed using a sample size of one (i.e., one chip/experimental group) lending limited interpretability based on our significantly altered miRNAs. Nonetheless, we used microarray only as a screening tool and validated the expression of miRs of interest using RT-PCR. Finally, when reviewing TCGA data, only a small proportion of cases (limited to cutaneous melanomas) had both RNA and miRNA sequencing data available. Because of the small sample size, a more rigorous study involving a larger sample size would be needed to understand the translational impact.

Nonetheless, our study also forms the basis for further exploring the role of miRs in CSCs and may be of particular interest to researchers and clinicians exploring new therapeutics modalities targeting miRs thereby modulating the complex interactions between tumor and immune cells in the TME. Future studies elucidating miR expression in tumor cells and patient's response to immunotherapy or characterizing the degree of immunosuppression in patient prognosis and treatment, would also be relevant clinically. In summary, our study found that melanoma CSCs modulate the TME by regulating a miRNA-gene network consisting of miR-92a, integrin α₅/α_v, and TGFβ.

No potential conflicts of interest were disclosed.

Conception and design: C. Shidal, P. Nagarkatti, M. Nagarkatti

Development of methodology: C. Shidal, P. Nagarkatti

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Shidal, P. Nagarkatti, M. Nagarkatti

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Shidal, N.P. Singh, P. Nagarkatti

Writing, review, and/or revision of the manuscript: C. Shidal, N.P. Singh, P. Nagarkatti, M. Nagarkatti

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Shidal, N.P. Singh, M. Nagarkatti

Study supervision: P. Nagarkatti, M. Nagarkatti

This work was supported, in part, by NIH grants P01AT003961, R01AT006888, R01AI123947, R01AI129788, R01MH094755, and P20GM103641 (to P. Nagarkatti and M. Nagarkatti).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.