NKG2A Is a Therapeutic Vulnerability in Immunotherapy Resistant MHC-I Heterogeneous Triple-Negative Breast Cancer

Immunotherapy, particularly immune-checkpoint inhibitors (ICI), has become a standard treatment for many types of cancer (1). The anti-programmed death-1 (PD-1) monoclonal antibody pembrolizumab is approved for the treatment of early-stage or programmed death ligand-1 (PD-L1)-positive metastatic triple-negative breast cancer (TNBC) in combination with chemotherapy (2–7). However, only about 15% of early-stage TNBC patients experience benefit from adding ICI (3, 8). Current biomarkers such as tumor-infiltrating immune cells and PD-L1 expression have shown limited or no utility in predicting ICI treatment outcomes in the early setting (3, 8–10). Further research is needed to understand why ICI therapy is ineffective for most TNBC patients and to devise strategies to expand its clinical benefit.

CD8⁺ T-cell recognition of major histocompatibility complex-I (MHC-I, comprised of classic HLA-A, -B, and -C in humans) molecules bearing antigens cognate to the unique T-cell receptor is critical for antitumor immune clearance and efficacy of ICIs (11, 12). Loss or dysfunction of MHC-I represents a common mechanism by which tumor cells can evade adaptive immunity and has been associated with resistance to immunotherapy (13–17). There are a variety of mechanisms whereby tumor-specific MHC-I (tsMHC-I) antigen presentation can be disrupted, including immune editing or selection against neoantigens (18, 19), loss or lack of MHC-I allelic diversity (20–22), and development of mutations which limit interferon-sensing pathways and resulting MHC-I upregulation (15). Somatic genetic losses at HLA loci or epigenetic suppression by promoter hypermethylation have also been described in breast cancer (23).

Mouse studies modeling the role of tsMHC-I in antitumor immunity and immunotherapy response frequently rely on complete (clonal) genomic loss of tumor beta-2-microglobulin (B2M), an obligate heterodimeric partner for MHC-I. Genetic deletion of B2M results in an immunologically cold tumor microenvironment and complete lack of response to PD-1/L1-targeted therapy (15, 24, 25), or other therapies targeting adaptive responses requiring antigen recognition (13, 26). However, clonal deletion of tsMHC-I or B2M is relatively uncommon in breast cancer or cancer in general (23, 27, 28), with a majority of tumors exhibiting subclonal lesions, increasing the complexity of tsMHC-I expression within the tumor microenvironment.

Heterogeneous loss or downregulation of tsMHC-I at the single-cell level is anticipated to have numerous consequences. Partial loss could render tumors less responsive or completely unresponsive to immunotherapies. Coordinate reduction in T-cell infiltration and function could also occur. Finally, perhaps more unexpectedly, heterogeneity in the microenvironment could lead to spatial effects on tumor immunity not appreciated in either fully MHC-I-competent or -incompetent tumors (e.g., a neomorphic effect).

In this study, we evaluated tsMHC-I heterogeneity in human breast cancers at the single-cell level and modeled heterogeneity in mice. Human breast cancers demonstrated high variability in tsMHC-I expression, with TNBC displaying the largest percentage of patients with multimodal distributions of tsMHC-I expression. Analyzing the spatial contexture of these tumors, we found that regions of tumors with heterogeneous MHC-I expression contained significantly higher levels of natural killer (NK) cells compared with regions with uniformly high or low MHC-I expression. Murine mammary tumors with enforced MHC-I heterogeneity (MHC-I^HET) mirrored our findings in human TNBC and showed substantially increased NK-cell infiltration compared with MHC-I proficient or B2M-knockout (KO) tumors. Tumor-infiltrating NK cells targeted with combination anti-NKG2A and anti-PD-L1 treatment overcame anti–PD-L1 resistance and extended survival in multiple murine mammary cancer models. Furthermore, we have identified IFNγ as a crucial recruitment factor for tumor-infiltrating NK cells in MHC-I heterogeneous tumors. Collectively, we identified key changes in the tumor microenvironment of heterogeneous MHC-I TNBCs and candidate combination immunotherapy to overcome ICI resistance.

Transcriptional suppression or polymorphic loss of MHC-I is correlated with ICI resistance (29, 30), presumably due to downregulation and reduced diversity of tumor antigen presentation to CD8⁺ T cells. The association of tsMHC-I protein expression on ICI treatment outcomes in patients with breast cancer has not been studied in detail. To test whether loss of MHC-I expression is associated with the efficacy of ICI therapy in breast cancer, multiplexed immunofluorescence (mIF) staining was conducted for DAPI, pan-cytokeratin (panCK, a marker for epithelial cells), and HLA-A/B/C (representing MHC-I), on pretreatment biopsies from 84 patients with metastatic TNBC in a recently completed randomized phase II clinical trial evaluating carboplatin ± atezolizumab (NCT03206203; Fig. 1A) (31). Improved progression-free survival (PFS) was observed in patients with high tsMHC-I expression (top 33% HLA-A/B/C intensity, n = 13) when treated with carboplatin in combination with atezolizumab compared with carboplatin alone. In contrast, there was a substantially reduced and non-significant difference in PFS between the two treatment arms for patients with low tsMHC-I expression (bottom 33% HLA-A/B/C intensity, n = 13; Fig. 1B). Our findings emphasize that adding anti–PD-1/L1 targeting immunotherapy is unlikely to benefit metastatic TNBC patients with low tsMHC-I expression. During this analysis, we observed that many breast cancers demonstrated heterogeneous expression of HLA-A/B/C, comprising regions of high and absent expression (see example in Fig. 1A). Of note, we further validated the specificity of the antibody HLA-A C6 through IHC analysis of various HLA alleles. The C6 antibody recognized the most common types of HLA-A, B, and C alleles, and is therefore representative of a pan-HLA stain (Supplementary Fig. S1A). To more specifically quantify the probability that individual tumor cells with high and low MHC-I expression were present within the same tumor, the tailedness of tsMHC-I expression was measured using the excess kurtosis parameter (Supplementary Fig. S1B). Most TNBC tumors displayed a leptokurtic distribution, demonstrating the coexistence of extreme outliers.

TNBC tumors demonstrate heterogeneous intratumor MHC-I expression, meaning tumor cells with high and low MHC-I expression coexist within the same microenvironment. To investigate whether the heterogeneity of tsMHC-I expression is breast cancer subtype-specific, single-cell-level mIF analysis was conducted for panCK and HLA-A/B/C on tumor microarrays comprising 295 patients with breast cancer (n = 227 ER⁺; 9 HER2⁺/ER⁺; 7 HER2⁺/ER⁻; 52 TNBC; Fig. 2A). Significant variability in tumor HLA-A/B/C expression was observed across breast cancer subtypes, with TNBC exhibiting the highest overall expression (Fig. 2B). Moreover, TNBC tumors at the single-cell level also demonstrated a high coefficient of variation, reaffirming tsMHC-I intrapatient heterogeneity in expression (Fig. 2C). To assess the distribution of tsMHC-I fluorescence intensity, Hartigan's dip test (32) was utilized to measure the probability of unimodal versus multimodal distribution. Many TNBC tumor cells expressed HLA-A/B/C in a multimodal manner (Fig. 2D), indicating the presence of HLA-A/B/C^high and HLA-A/B/C^low tumor cells within the same tumor. HER2⁺/ER⁺ tumors also expressed a high degree of multimodality but were poorly represented in this data set.

Distinct patterns of MHC-I heterogeneity were observed in TNBC patient tumors, but the impact of this observation on immune cell infiltration and response to immunotherapy has not been studied in preclinical models. To assess how the immune microenvironment is shaped in response to tsMHC-I heterogeneity, CRISPR-Cas9 was used to knock out B2M in a BALB/c murine orthotopic mammary cancer model (EMT6), which has homogeneously high MHC-I expression at baseline (Fig. 3A). The EMT6 model displayed variability in response to single-agent anti–PD-L1, ranging from complete response (CR), acquired resistance, and intrinsic resistance (IR; Supplementary Fig. S2A and S2B). As expected, clonal loss of B2M completely abrogated tumor sensitivity to anti–PD-L1 (Fig. 3B; Supplementary Fig. S2A), consistent with prior literature (33–36). To model MHC-I heterogeneity and response to anti–PD-L1, EMT6 MHC-I proficient and deficient isogenic lines were mixed at various ratios prior to injection (Fig. 3C; Supplementary Fig. S2A). As little as a 10% loss of MHC-I resulted in significantly decreased CRs and a 50% loss of MHC-I resulted in no CRs (Supplementary Fig. S2C). Because the 50% loss of MHC-I abrogated CR to anti–PD-L1, this percentage was selected to model MHC-I heterogeneity, termed MHC-I^HET henceforth, to examine changes in the immune microenvironment with respect to MHC-I loss and resistance to anti–PD-L1.

To determine how changes in MHC-I expression might reshape the immune microenvironment in untreated parental (high MHC-I), MHC-I^HET (heterogeneous MHC-I), and B2M-KO (MHC-I null) EMT6 tumors, RNA gene-expression profiling was performed on CD45⁺ infiltrating immune cells for each tumor type. Gene-expression analysis revealed that T-cell markers and interferons were coordinately expressed with increasing MHC-I expression (Fig. 3D; Supplementary Fig. S3A and SB; Supplementary Table S1 and Supplementary Data S1). Interestingly, MHC-I^HET tumors had a uniquely elevated level of NK cell-related genes (Fig. 3E; Supplementary Fig. S3C; Supplementary Table S1 and Supplementary Data S1). Increased NK-cell abundance in MHC-I^HET tumors was further confirmed using flow cytometry through direct measurement of immune cell populations and total tumor populations (Fig. 3F–H; Supplementary Fig. S3D). To describe these differences in NK-cell infiltration among the models at various stages of tumor size, a longitudinal assessment of tumor-infiltrating lymphocytes was performed by fine-needle aspiration. NK cells were present at the beginning of tumor formation, ∼100 mm³, in the parental and B2M-KO tumors, but when tumors reached ∼500 mm³, NK cells were no longer detected. In contrast, in the MHC-I^HET EMT6 tumor model, NK cells infiltrated at significantly higher numbers at ∼100 mm³ and were still present at ∼500 mm³ (Supplementary Fig. S3E), indicating higher NK-cell retention throughout tumor progression. We hypothesized that the lack of immune infiltration in B2M-KO tumors could be due to a paucity of key chemokine and cytokine signals such as Ifnγ/α and CXCL9/10 (Supplementary Fig. S3B and S3F; Supplementary Data S1).

Because the expression of CXCL9/10 was similar between parental and MHC-I^HET EMT6 tumor microenvironments, the level of these chemokines alone could not explain the differences in NK-cell abundance. Therefore, single-cell RNA sequencing (scRNA-seq) and NanoString gene-expression profiling were utilized to characterize tumor-infiltrating NK cells from parental and MHC-I^HET EMT6 tumors to assess factors contributing to NK-cell retention and recruitment. NK cells were characterized with SingleR and further confirmed by their Ncr1 and Klrc1 positivity and lack of Cd3e (Fig. 4A; Supplementary Fig. S4A). Differential gene-expression analysis revealed NK cells from MHC-I^HET EMT6 tumors expressed high levels of Itga2 (CD49b, conventional NK cells), an integrin subunit that forms a receptor for collagen (Fig. 4B; Supplementary Table S2). CD49b has recently been shown to help retain NK cells at the site of infection when binding to collagen, which is highly enriched within EMT6 tumors and surrounding fibroblasts (Supplementary Fig. S4B and S4C; refs. 37–39). In addition, enrichment of IL2 signaling-related genes (Stat5b, Il2rb), and particularly higher Bcl2 expression, an antiapoptotic factor that also serves as an activation or proliferation marker for NK cells (40), was observed in NK cells from MHC-I^HET EMT6 tumors compared with parental EMT6 tumors (Fig. 4B and C). These observations were further confirmed using NanoString gene profiling on CD45+ cells and flow cytometry (Fig. 4D and E; Supplementary Data S1), suggesting that MHC-I^HET tumor-infiltrating NK cells bear specific phenotypes that lead to NK-cell retention and maintenance. Although MHC-I^HET EMT6 tumors are resistant to anti–PD-L1 therapy, there are substantial changes in the amount of tumor-infiltrating NK cells, illustrating an opportunity to therapeutically exploit this phenotype.

Previous studies have established that the presence of conventional type 1 dendritic cells (cDC1) in the tumor microenvironment is reliant on NK cells (41, 42). Flow cytometry analysis revealed that cDC1s (defined here as CD45⁺CD11c⁺MHC-II-high) were significantly more abundant in MHC-I^HET EMT6 tumors compared with parental EMT6 tumors, consistent with the higher NK-cell infiltration (Supplementary Fig. S4D). Transcriptomic analysis of CD8⁺ T cells in untreated MHC-I^HET and parental EMT6 tumors revealed T cells from MHC-I^HET EMT6 tumors displayed markers of naïve CD8⁺ T cells (Tcf7, Ccr7, Sell) and had reduced markers of cytotoxic function and activated CD8⁺ T cells (Gzmb, Nkg7; Supplementary Fig. S4E–S4G; Supplementary Table S2). Thus, despite the increased potential for CD8⁺ T-cell priming by more abundant cDCs, it is possible that the substantial reduction tumor cell–specific antigen presentation in MHC-I^HET tumors dominates the phenotype in basal conditions.

We next hypothesized that therapeutic activation of tumor-infiltrating NK cells in MHC-I^HET EMT6 tumors could overcome resistance to anti–PD-L1 immunotherapy. The inhibitory receptor NKG2A is expressed on NK and T cells, recognizing the nonclassic MHC-I molecule HLA-E in humans and Qa-1b (encoded by the H2-T23 gene) in mice. Expression of Qa-1b was observed in MHC-I^HET EMT6 tumor cells (CD45⁻) in vivo (Fig. 5A). NKG2A expression of NK and T cells in the tumor environment and control tissue (spleen) was examined. NKG2A expression of NK cells was constitutive, whereas it was increased on tumor-infiltrating CD8⁺ T cells as compared with spleen (Fig. 5B). To test if activation of NK cells via NKG2A blockade in MHC-I^HET EMT6 tumor-bearing mice would improve response and overall survival, mice were treated with anti-NKG2A, anti–PD-L1, or a combination of both. Although single-agent anti-NKG2A or anti–PD-L1 showed no improved survival or reduced tumor growth rates, a combination of NKG2A and PD-L1 blockade had a pronounced antitumor effect and restored CRs (CR = 30%) in mice bearing MHC-I^HET EMT6 tumors (Fig. 5C).

Similar results were obtained in the basal-like C57BL/6 murine mammary cancer cell line E0771. When stimulated with IFNγ for 48 hours in vitro, E0771 cells display a naturally occurring heterogeneous expression of MHC-I (Fig. 5D). Additionally, E0771 had similar levels of tumor-infiltrating NK and CD8⁺ T cells to the MHC-I^HET EMT6 model (Fig. 5E). E0771 tumors are known to be resistant to anti–PD-L1 therapy (43). When E0771 tumor-bearing mice were treated with a combination of anti–PD-L1 and anti-NKG2A blockade, survival was significantly extended compared with single-agent anti–PD-L1 and anti-NKG2A (Fig. 5F).

Given that the combination blockade of NKG2A and PD-L1 improved response in MHC-I heterogeneous mammary tumors, it was investigated whether this strategy could also enhance the response in parental or B2M-KO tumors. Although B2M-KO tumors showed no response to combination therapy, parental EMT6 tumors showed significant tumor reduction (CR = 80%) in comparison with single-agent anti–PD-L1 (CR = 33%; Supplementary Fig. S5A). This suggests engagement by CD8⁺ T cells and potential augmentation of limited populations of NK cells, which are enhanced after anti–PD-L1 treatment alone in parental EMT6 tumors (Supplementary Fig. S5B).

To more rigorously test whether PD-L1 therapeutic targeting in the MHC-I^HET tumors is thwarted by the Qa-1–NKG2A axis, Qa-1, the ligand for NKG2A, was deleted in parental EMT6 cells while leaving other MHC-I coding genes intact (Supplementary Fig. S5C). Treatment with anti–PD-L1 resulted in a significant reduction in tumor volume in Qa-1-KO tumors (Supplementary Fig. S5D), similar to the level of response observed in parental EMT6 tumors that received combination treatment with anti-NKG2A. The Qa-1-KO cells were then combined with B2M-KO cells (which cannot express Qa-1 due to lack of B2M) at a 50:50 ratio to create a tumor model that is MHC-I^HET/Qa-1^Neg. Qa-1 knockout partially restored single-agent anti–PD-L1 response (Supplementary Fig. S5E). We hypothesize that the effect was only partial due to the availability of Qa-1 from other stromal cells in the tumor microenvironment but could also hint at alternative ligands for NKG2A. Thus, these data demonstrate the role of the NKG2A/Qa-1 axis in promoting IR to anti–PD-L1 in MHC-I^HET tumors.

NKG2A is expressed on both CD8⁺ T cells and NK cells and is considered a therapeutic target in both cell types. To determine if the antitumor effect of combination therapy depended on NK cells or CD8⁺ T cells, tumor-bearing mice were treated with antibodies for asialo-GM1 (BALB/c and C57BL/6) or NK1.1 (C57BL/6; depleting NK cells) or CD8α (depleting CD8⁺ cytotoxic T lymphocytes). Based on these studies both NK and CD8⁺ T cells are necessary for response to the combination of anti-NKG2A and anti–PD-L1 in MHC-I^HET EMT6 and E0771 tumors (Fig. 5G; Supplementary Fig. S6A–S6C). These data suggest that both therapies are targeting different populations of lymphocytes, each of which may be functionally eliminating different populations of tumor cells in the microenvironment. To test this, tumor MHC-I expression was examined by flow cytometry at the tumor endpoint among the treatment groups. In both the MHC-I^HET EMT6 tumors and E0771 tumors, anti–PD-L1 treatment was selected against MHC-I expressing cells, resulting in tumors that were predominantly MHC-I null, whereas anti-NKG2A selected for MHC-I expressing cells (Fig. 5H; Supplementary Fig. S6D). In EMT6 MHC-I^HET tumors, NK cells were significantly higher in single-agent-treated anti-NKG2A and combination-treated tumors on day 3 after treatment; however, by day 7, only NK cells in combination-treated tumors remained elevated (Fig. 5I). The NK cells in tumors treated with anti-NKG2A exhibited higher degranulation as measured by CD107a (Fig. 5J), whereas T cells in tumors treated with anti–PD-L1 had significantly higher degranulation (Fig. 5K). These data suggest that anti-NKG2A and anti–PD-L1 may be a synergistic combination treatment to overcome immunotherapy resistance in TNBC models with heterogeneous MHC-I expression via cooperative elimination of MHC-I⁺ and MHC-I⁻ cells.

We next hypothesized that the absence of NK cells in B2M-KO tumors was attributed to the lack of IFNγ from antigen-responsive T cells leading to a cold tumor microenvironment. To investigate this, a doxycycline (dox) inducible IFNγ vector was transduced in B2M-KO EMT6 cells, which do not inherently secrete IFNγ (Fig. 6A–C). In vivo dox-mediated induction of IFNγ in B2M-KO EMT6 tumors resulted in a significant increase of tumor-infiltrating NK cells in MHC-I null tumors (Fig. 6D). Subsequent activation of these NK cells with anti-NKG2A significantly reduced tumor burden in the B2M-KO EMT6 tumors, which were previously shown to be unresponsive to multiple types of therapy (Fig. 6E). Our findings suggest that although MHC-I null tumors lack inhibitory signals that could facilitate NK-mediated tumor clearance, the absence of critical stimulatory signals (e.g., IFNγ) results in poor NK-cell recruitment. Consistent with this hypothesis, MHC-I^HET EMT6 tumors implanted into IFNγ^−/− BALB/c mice resulted in a complete loss of NK-cell infiltration into MHC-I^HET EMT6 tumors compared with those established in wild-type mice, whereas T-cell infiltration remained relatively unchanged (Fig. 6F). Depletion of CD8⁺ T cells also abolished NK-cell recruitment in MHC-I^HET EMT6 tumors grown in wild-type mice, suggesting that interferon from antigen-activated CD8⁺ T cells is likely a required signal for NK-cell recruitment (Fig. 6G).

To identify if the NKG2A/HLA-E axis is a relevant therapeutic target in TNBC, expression of tumor HLA-E and NKG2A in infiltrating NK cells was examined in a series of TNBCs. HLA-E was found to be widely expressed in TNBC tumors and was significantly correlated with the presence tumor-infiltrating NK and CD8⁺ T cells (Fig. 7A–B). Furthermore, analysis of previously published scRNA sequencing data from human TNBC (44) revealed that NKG2A (gene name KLRC1) is highly enriched in NK and, to a lesser degree, CD8⁺ T cells, consistent with results in our in vivo model (Supplementary Fig. S7A).

Based on the observed distinct immune cell infiltration pattern changes due to alterations in tsMHC-I expression in mice, we asked whether tsMHC-I expression is associated with similar tumor immune microenvironment features in humans. To test this, mIF assays were performed using panCK and HLA-A/B/C to evaluate tsMHC-I expression, and CD8 or CD56 (NK-cell marker) to spatially assess immune cell infiltration patterns (Fig. 7C–D; Supplementary S7B and S7C). At the bulk level, a positive correlation between tsMHC-I expression and CD8⁺ T-cell infiltration was observed (Supplementary Fig. S7D). However, there was no significant linear correlation between average MHC-I expression and the abundance of NK cells (Supplementary Fig. S7E). Given the tsMHC-I heterogeneity observed in TNBC tumors, the immune cell infiltration patterns around regions of high and low MHC-I expression and regions characterized by heterogeneous MHC-I expression were examined. Density-based spatial clustering of applications with noise (DBSCAN; ref. 45) was used to cluster panCK⁺ cells from each tumor. Based on the average HLA-A/B/C intensity of the tumor clusters, the clusters were classified as HLA-A/B/C^high, HLA-A/B/C^het, and HLA-A/B/C^low. Using the nearest neighbor method, each tumor cluster was paired with its closest CD8⁺ T-cell or NK-cell (Fig. 7C and D). To minimize sampling bias, the number of immune-tumor pairs was normalized by the number of immune cells and tumor clusters. CD8⁺ T cells tended to colocalize with HLA-A/B/C^high tumor clusters. In contrast, regions of HLA-A/B/C^low had the lowest infiltration of CD8⁺ T cells (Fig. 7E). Consistent with our findings in the murine model, HLA-A/B/C^het tumor clusters exhibited the highest levels of infiltrating NK cells and intermediate level of CD8⁺ T cells (Fig. 7E and F). Thus, regions of MHC-I heterogeneity are associated with increased NK-cell localization, consistent with our murine models.

To investigate the impact of heterogeneity in MHC-I expression on the outcome of ICI treatment, the abundance of HLA-A/B/C^het tumor regions in pretherapy biopsy samples from the clinical trial of carboplatin ± atezolizumab in metastatic TNBC was evaluated. Tumor cells were grouped using the DBSCAN algorithm, and the percentage of HLA-A/B/C^het regions out of total tumor clusters was calculated. Patients whose biopsies contained over 85% HLA-A/B/C^het regions were classified as “abundant heterogeneity,” whereas the remaining patients were labeled as “low heterogeneity.” Among chemoimmunotherapy-treated patients, the abundance of HLA-A/B/C^het regions was strongly associated with PFS, but not in patients treated with chemotherapy alone (Fig. 7G). In patients with abundant MHC-I heterogeneity, the NK:CD8 T-cell ratio was higher (Fig. 7H), indicating a relatively higher infiltration of NK cells that could be activated/targeted with anti-NKG2A immunotherapy. These findings highlight the importance of understanding the relationship between tsMHC-I expression and immune cell infiltration in determining patient outcomes and identifying novel immunotherapy targets.

Anti–PD-1/PD-L1–based therapy has the potential to produce durable antitumor immune responses in some patients. However, adaptive and innate immune responses impose selection pressure on tumor cells, which eventually leads to immune escape and immunotherapy resistance. Multiple cancer cell-intrinsic immune-escape mechanisms, such as low neoantigen load (46, 47), loss of antigen presentation (15), and impaired response to interferon (48), have been correlated with immunotherapy outcomes. Among these ICI resistance mechanisms, insufficient antigen presentation to T cells, exemplified by MHC-I downregulation, compromises a large proportion of patients (49, 50). Our analysis of metastatic TNBC patients treated with chemotherapy ± immunotherapy demonstrated that high tsMHC-I heterogeneity is associated with a lack of benefit to PD-L1 therapy. As heterogeneous tsMHC-I expression is common in breast cancer, particularly TNBC, our results highlight the need to overcome aberrant tsMHC-I expression to achieve higher rates of immunotherapy responses and benefits. Although several approaches have been investigated to enhance MHC-I expression and reactivate antitumor immunity (51), our objective was to exploit changes in the immune microenvironment resulting from tsMHC-I heterogeneous expression to overcome immunotherapy resistance. We found that this tumor-specific phenotype drives NK-cell infiltration leading to our investigation of alternative checkpoint blockades, specifically NKG2A.

Our findings reinforce the growing interest in utilizing tumor-infiltrating NK cells as a potential therapeutic strategy for tumors with loss of MHC-I expression (52–54). Other reports have highlighted tumor cells with complete loss of MHC-I and increased sensitivity to NK cells because they fail to engage inhibitory KIR (55–57). Although informative, these studies have not explored the relationship of NK cells and varied MHC-I expression, which is commonly the case we and others have found in breast cancer (23, 27, 28). This study provides evidence that heterogeneity in MHC-I expression, not complete loss of MHC-I, has the strongest correlation with NK-cell infiltration in TNBC patient cohorts and murine models. The observed increase in NK-cell abundance in heterogeneous MHC-I tumors is likely due to several factors: first, as we have shown here, tumor microenvironmental IFNγ is required for NK-cell recruitment, likely coming from antigen-responsive T cells. Second, although not mechanistically shown here, an upregulation of IL2 signaling, crucial for maintaining NK-cell activity, appears to be associated with enhanced NK-cell activity. Finally, we observed an enrichment of CD49b expression, which facilitates NK-cell retention within the tumor microenvironment. Interactions between collagen and CD49b in NK cells can promote the expression of CXCL10 and other chemokines in the presence of low levels of interferon (39). The CD49b^High NK cells that infiltrate MHC-I^HET tumors may actively engage with the collagen-rich environment, leading to an increase in Cxcl9/10 expression. However, further investigation is necessary to explore the impact of NK-dependent chemokine expression in the MHC-I^HET microenvironment on CD4/CD8 infiltration and retention.

Increased infiltration of NK cells has important implications for systemic antitumor immunity, as previous research has demonstrated that NK cells produce chemokines that attract cDC1s (41, 42, 58). We observed higher dendritic cell infiltration in the NK cell-abundant MHC-I^HET tumor models. However, markers of T-cell activation that could be associated with cDC cross-priming were not observed. This phenomenon may be explained by differences in immunogenicity between our syngeneic tumor model and other spontaneous models, or the partial loss of MHC-I on tumor cells. Alternatively, cDC priming of CD4⁺ T cells leading to CD4-mediated cytokine support could also play a more dominant role under basal conditions in the absence of therapeutic checkpoint blockade.

Other studies have highlighted the importance of NK and CD8⁺ T-cell activation through NKG2A blockade therapy (59–61). Monalizumab (anti-NKG2A, AstraZeneca) has become the center of several clinical trials in various cancer types (NCT04307329, NCT04590963, NCT05221840, NCT02671435, NCT05061550, and NCT03833440). Our findings indicate that HLA-E expression is frequently detected in TNBC tumors and is positively correlated with infiltration of NK and CD8⁺ T cells. These findings suggest that combining anti-NKG2A with anti-PD-L1 therapy may represent a promising yet underexplored approach for treating TNBC tumors. Future research could involve spatial profiling to measure the distance between HLA-E–positive tumors and NK/CD8⁺ T cells, with the aim of developing potential biomarkers to predict patient response to NKG2A blockade.

These data reinforce the growing interest on how tumor-specific antigen presentation via MHC-I plays a role in modifying antitumor immunity and ICI response. Together, they endorse the unmet translational and clinical need to address heterogeneity in MHC-I expression as a variable in understanding breast cancer antitumor immunity and response to immunotherapy. Moreover, these data showcase the potential to harness NK-cell function in advancing cancer immunotherapy combinations.

The 314 tissue microarrays used to study tsMHC-I expression heterogeneity were from patients with breast cancer enrolled in the BRE03103 repository trial: NCT00899301, IRB030747, and at the Instituto Nacional de Enfermedades Neoplásicas (Lima, Perú): INEN 10-018. The 84 mTNBC patients who were used to study the correlation between tsMHC-I expression and immunotherapy benefit were enrolled in NCT03206203.

Murine mammary cancer cell lines EMT6 and E0771 were obtained from ATCC. Cell lines were tested at least once quarterly for Mycoplasma contamination. All media components were purchased from commercial vendors and prepared/stored under sterile conditions. Cell lines are utilized within the early passage (<30 passages from acquisition from ATCC) and are DNA fingerprinted through commercial services for validation. EMT6 cells were grown in DMEM (Gibco) supplemented with 10% fetal bovine serum (FBS; Life Technologies). E0771 cells were grown in RPMI (Gibco) and supplemented with 10% FBS. B2M-KO-IFNγ EMT6 cells were maintained in DMEM-F12 (GIBCO) supplemented with 10% TET-System approved FBS (GIBCO). B2M and Qa-1 knockout cell lines were generated using CRISPR plasmids (Santa Cruz) and flow sorted on GFP-positive cells. Cell lines were verified to be complete knockout populations by flow cytometry.

The pSHUTTLE-IFNγ (GeneCopoeia) was recombined into the pINDUCER-20 plasmid (62) using LR Clonase (Invitrogen), resulting in DOX-inducible IFNγ transgene expression. Lentiviral particles were produced by cotransfecting pINDUCER-IFNγ plasmid with pMD2G and psPAX helper plasmids into 293FT cells. Target cells were transduced in the presence of polybrene and selected by neomycin resistance.

Levels of IFNγ secreted from B2M-KO-IFNγ EMT6 cells were evaluated using an ELISA kit according to the manufacturer's instructions (Invitrogen).

All mice were housed at the Vanderbilt University Medical Center vivarium, which is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC). Mouse procedures and studies were approved by the Vanderbilt Division of Animal Care and Institutional Animal Care and Use Committee (IACUC). C57BL/6J and BALB/c mice were purchased from Envigo (Indianapolis, IN) and allowed to acclimatize for at least one week before tumor implantation and experimentation. IFNγ^−/− mice were purchased from The Jackson Laboratory (002287) and genotyping was performed according to distributer protocol. For all experiments, 6- to 8-week-old female mice with 100 mm³ tumors were stratified into specific treatment groups.

For mammary tumor models, 5 × 10⁴ EMT6 or 1 × 10⁵ E0771 cells were orthotopically injected into the fourth right mammary fat pad of female BALB/c (EMT6) or C57BL/6J (E0771) mice. Following the establishment of tumors (∼100–200 mm³) mice were stratified prior to therapy administration. The mice were treated via intraperitoneal (i.p.) injection with isotype IgG2a control (Bio X Cell, clone BE0089), anti-NKG2A (Bio X Cell, clone 20D5), or anti–PD-L1 (Genentech, clone 6E11) dosing at 200 μg on day 7 and 100 μg on days 14, 21, 28, and 35. Mice used for each treatment arm are defined in respective figure legends. For tumor growth analysis, tumors were measured 3 times weekly with calipers, and volume was calculated in mm³ using the formula (length × width × width/2). Mice were humanely euthanized at defined end points or when the tumor volume reached 2 cm³ or tumor ulceration. For the pINDUCER tumor model, mice were primed with doxycycline (2 mg/mL in 5% sucrose, ad libitum) or 5% sucrose (control) for 2 days before injection. Doxycycline was continually administered in drinking water before sacrifice.

Following previously published techniques (57), fine-needle aspiration (FNA) on mouse tumors using a sterile beveled needle attached to a 10 mL syringe with a syringe holder. The needle was then agitated within the tumor and one to three needle passes were collected from each tumor. The FNA-acquired cells were then processed for flow cytometry analysis.

For NK-cell depletion, 100 μL of polyclonal anti–asialo-GM1 (BioLegend, clone Poly21460) or 100 μg of anti-NK1.1 (Bio X Cell, clone PK136) was injected i.p. into mice once weekly prior to tumor implantation. For the depletion of CD8⁺ T cells, 200 μg of depleting anti-CD8α (Bio X Cell, clone 2.43) was administered via i.p injection for the initial treatment and then 100 μg once weekly. A rat IgG2b for anti-CD8α, k (Bio X Cell, clone BE0090) and rabbit polyclonal IgG (BioLegend, clone Poly29108) was used as an isotype control.

Samples were run on an Attune NxT Acoustic Focusing Cytometer (Life Technologies). Analysis was performed in FlowJo. Gating was first done on forward scatter and side scatter to exclude debris. Doublets were excluded by gating on FSC area versus FSC height (Supplementary Fig. S8) Zombie Violet (BioLegend) was used to exclude dead cells from analyses. Antibodies: CD3-AF488 (BioLegend, clone HIT3a), Cd8a-AF488 (BioLegend, clone 53*6.7), CD45-PerCP/Cy5.5 (BioLegend, clone GK1.5), CD45-APC (BioLegend, clone 30-F11), NK1.1-APC (BioLegend, clone PK136), NKp46- APC (BioLegend, clone 29A1.4), CD11b- PE/Cy7 (BioLegend, clone M1/70), CD11c-AF488 (BioLegend, N418), IA/IE-AF700 (BioLegend M5/114.15.2), CD27 (BioLegend, clone LG.3A10), H-2- PE (BioLegend, clone M1/42), NKG2A-PE (Miltenyi Biotec, clone REA1161), H2-Dd-PE (BioLegend, clone 34-2-12), Qa-1b(b)-PE (BD Biosciences, clone 6A8.6F10.1A6), PD-L1–APC (BioLegend, clone 29E.2A3), CD49b-APC/Cy7 (BioLegend, clone DX5), and CD107b-FITC (BioLegend, clone 1D4B), IFNγ-PE (BioLegend, clone XMG1.2).

EMT6 and E0771 tumors were harvested from mice at ∼300–500 mm³ (serum-free RPMI, Gibco) and dissociated using the Tumor Dissociation Kit (Miltenyi Biotec) according to the manufacturer's specifications with the gentleMACS Octo dissociator (Miltenyi Biotec) default tumor protocol (40 minutes at 37°C under constant agitation). The dissociate was then passed through a 40-μm filter and washed with 20–30 mL of PBS and ACK lysed. CD45⁺ cell isolation was performed using CD45 mouse microbeads (Miltenyi Biotec).

After dissociation and CD45⁺ bead isolation, RNA was harvested from mouse tumors using the Maxwell 16 automated workstation (Promega) and the LEV simplyRNA Tissue Kit (Promega). RNA concentration was determined by spectrophotometry (NanoDrop2000, Thermo Fisher Scientific).

Gene-expression analysis on untreated heterogeneous MHC-I EMT6 tumors (CD45⁺ bead sorted) was performed using the NanoString Pan-cancer immunology panel according to the manufacturer's specifications. Tumor dissociates were used for RNA preparation, and 50 ng of total RNA was used for input into nCounter hybridizations. Data were normalized according to endogenous housekeeper controls and transcript counts were log-transformed for statistical analysis. Original data from the NanoString are available at Supplementary Data S1.

Immune cells were harvested from parental EMT6 and MHC-I^HET tumors, and total tumor and CD45⁺ immune populations were collected from parental EMT6 tumors for scRNA-seq. Each sample (targeting 15,000 cells per sample) was processed for single-cell 5′ RNA sequencing utilizing the 10X Chromium system. Libraries were prepared following the manufacturer's protocol. The libraries were sequenced using NovaSeq 6000 with 150 bp paired-end reads. RTA (v.2.4.11; Illumina) was used for base calling, and analysis was completed using 10X Genomics Cell Ranger software. Data were analyzed in R using the filtered h5 gene matrices in the Seurat package. In brief, samples were subset to include cells with >200 but <3,000 unique transcripts to exclude probable noncellular RNA reads and doublets. Cells with >15% of reads coming from mitochondrial transcripts were also excluded as probable dying cells. Batch effect correction was introduced using Seurat's built-in CCA method. SingleR was used to assist with cell-type annotation of clusters. Differential gene-expression analysis was applied calling FindMarkers function with resulting volcano plot produced with EnhancedVolvano. Single-sample gene set enrichment analysis was performed on each individual cell using the Escape package using the Hallmark gene pathway.

Immunofluorescent assays for CK/HLA-A/B/C ± CD8 or CD56 and CK/HLA-E were developed and optimized at Vanderbilt University Medical Center using tyramine signal amplification (TSA) for increased antigen sensitivity. Briefly, formalin-fixed paraffin-embedded tissue was sectioned at 4 mm, deparaffinized, antigen retrieval performed with citrate buffer at pH 6, endogenous peroxidase blocked with hydrogen peroxide, and protein block applied. Sections were then incubated with the first primary antibody overnight at 4°C. Followed by incubation with an HRP-conjugated secondary antibody, and TSA reagent applied according to the manufacturer's recommendations. After washing, antigen retrieval, and protein blockade, incubation steps were repeated for the second and third primary antibodies as described. Primary antibodies: (panCK AE1/AE3 Biocare) at 1:1,600, HLA-A/B/C C6 Santa Cruz at 1:1,300, CD8 144B MM39-10 StatLab, CD56 E7 × 9M Cell Signaling at 1:600, HLA-E MEM-E/02 Abcam at 1:400. Counterstaining was performed with DAPI for nuclei identification. Tonsil and placenta were used as positive and negative control tissues.

The specificity of the C6 antibody was confirmed by IHC analysis of K562 cells (MHC-I null) transduced with individual HLA-A, B, and C alleles. The C6 antibody recognizes the most common types of HLA-A, B, and C and is therefore representative of a pan-MHC-I stain (Supplementary Fig. S1A).

Whole slide images were digitally acquired using an AxioScan Z1 slide scanner (Carl Zeiss) at 20×. Automated quantification was performed via a pathologist-supervised machine learning algorithm using QuPath software. Cell segmentation was determined with the built-in cell detection algorithm. Briefly, nuclei were identified above an optimized threshold on DAPI and defined utilizing watershed segmentation. The cell cytoplasm and membrane were determined by 3 μm cell expansion around the nuclei. Cells were labeled as objects in a hierarchical manner below the tissue annotation or TMA core, each with an xy coordinate. Measurements of intensity and morphology were generated for every cell compartment and channel. Object classifiers were trained on annotated training regions from control tissue and tumor samples to define cellular phenotypes. Tumor cells were defined by panCK expression and subcellular characteristics. T-cell and NK cells were defined by CD8 and CD56 expression, respectively. Once the algorithm was performing at a satisfactory level, it was used for batch analysis. For cases with low, heterogeneous, or null CK expression in which the classifier's performance was not optimal, tumor areas were manually annotated. Out-of-focus areas, tissue folds, necrosis, normal breast, and in situ carcinoma were excluded from the analysis. A pathologist visually assessed each sample for the correct performance of the algorithm. Single-cell data including sample ID, xy coordinate, cell phenotype, and HLA-A/B/C and HLA-E intensity were exported from QuPath for further analysis.

To measure if HLA-A/B/C expression is associated with immunotherapy benefit, a categorical variable of Top and Bottom 33% HLA-A/B/C expression was created. Patients from the carboplatin ± atezolizumab arms were ordered separately from high to low based on their average tumoral HLA-A intensity. For each arm, the top 33% of the rank were selected (for carboplatin + atezolizumab, 41/3 = 13.6 (round to 13 patients); for carboplatin, 39/3 = 13 patients) to compare their PFS. Similarly, 13 patients at the bottom of the rank were selected from each arm to compare their PFS. Single-cell resolution immunofluorescence data, including fluorescence intensity for HLA-A/B/C, CD8, CD56 (NK-cell marker), and pancytokeratin on breast cancer tumors, and spatial cell coordinates were imported from immunofluorescence imaging data. The coefficient of variation (µ/σ) was calculated to measure tumor-specific HLA-A/B/C heterogeneity within the sample. To assess the modality of tumor-specific HLA-A/B/C expression, Hartigen's Dip test and the R package LaplacesDemon were applied to calculate a Bayesian form of test statistic for multimodality.

Tumor cells were categorized as MHC-I^high (mean HLA-A/B/C intensity + standard deviation), MHC-I^low (mean HLA-A/B/C intensity − standard deviation), or MHC-I^mid. Then, the Ripley's K function (R package: spatstat.core, v2.4-4) and the resulting area under the curve were calculated to measure the spatial distribution and clustering pattern. Clustering algorithms DBSCAN (R package: dbscan, v1.1) were applied, taking the tumor cells’ spatial coordinates as input to create tumor clusters. The clusters were categorized as HLA-A/B/C^high (HLA-A/B/C median + IQR), HLA-A/B/C^low (HLA-A/B/C median-IQR), or HLA-A/B/C^het.

NK-cell quantification was performed using the CIBERSORTx deconvolution of Translational Breast Cancer Research Consortium (TBCRC) patient's tumor bulk RNA-seq data. CD8 T-cell quantification was collected from the mIF image result. The NK:CD8 ratio was calculated using the formula log₂NK/CD8.

To quantify the MHC-I heterogeneity within the tumor, dbscan unsupervised clustering to identify individual tumor clusters was performed. Each cluster was then classified as MHC-I high, het, and low based on the cluster's average HLA-A/B/C intensity. The percentage of MHC-I het clusters within the tumor was then calculated. To identify the optimal cutoff points that separate patient's PFS, the R package survminer (version 0.4.9) was used and an 85% MHC-I het cluster was chosen as the optimal cutoff point. Patients with tumors containing ≥85% MHC-I het clusters were labeled as having abundant MHC-I heterogeneity, whereas tumors with <85% were labeled as having low MHC-I heterogeneity.

Statistical analyses were performed as indicated using R or GraphPad Prism (GraphPad Software). For two-group analyses, Student t test was used to compare normally distributed means. Log or square root transformation or nonparametric testing was used if not normally distributed. For comparison within the patient or tumor, paired t test or nonparametric testing was used. For multiple comparisons, ANOVA was accompanied by a posthoc Tukey test or a Kruskal–Wallis test with a posthoc Dunn test. For survival analysis, the log-ranked test is used. To access the data tailedness, the excessive kurtosis parameter was calculated from the base R function kurtosis. In a standard normal distribution, the kurtosis value is 3. Therefore, excess kurtosis is calculated by subtracting 3 from the original kurtosis value. For a data distribution that has very few outliers, like a uniform distribution in extreme cases, a negative excess kurtosis value is expected. A positive excess kurtosis value, such as in the case of a patient's HLA-A intensity, indicates that the data distribution tends to have heavy tails. A heavy-tailed distribution means there is a greater probability of encountering a tumor cell with HLA-A/B/C expression that is a certain standard deviation away from the mean tumoral HLA-A/B/C expression during random sampling of tumor cells.

The data generated in this study are available within the article and its supplementary files. Additional data or resources related to this article are available upon reasonable request from the corresponding authors. Data from the clinical trial, NCT03206203, will be available at the SRA database under extension ID: PRJNA995589.

All studies using human tissues or human subjects were conducted in accordance with the Declaration of Helsinki and were performed after approval by an institutional review board and in accordance with an assurance filed with and approved by the U.S. Department of Health and Human Services, when required. Informed written consent was obtained from each subject in prospective clinical trial subjects. Exceptions to this requirement were made by the institutional review board in cases where human tissues were part of tissue microarrays and were completely deidentified.

L. Van Kaer reports personal fees from Isu Abxis Ltd. (Republic of Korea) outside the submitted work. C. Isaacs reports grants from TBCRC during the conduct of the study, other support from Genentech, PUMA, Seattle Genetics, Astra Zeneca, Novartis, Sanofi, Gilead, Wolters Kluwer, and McGraw Hill, and grants and other support from Pfizer outside the submitted work. T.J. Ballinger reports personal fees from Medscape outside the submitted work. C.A. Santa-Maria reports grants from Pfizer, BMS, AstraZeneca, Merck, Novartis, Genentech, and Celldex outside the submitted work. P.D. Shah reports personal fees from Daiichi Sankyo, AstraZeneca, Gilead Biosciences, and Biotheranostics outside the submitted work. V.G. Abramson reports other support from Astra Zeneca, Macrogenics, Eisai, and Genentech outside the submitted work. J.M. Balko reports grants from NIH/NCI, U.S. Department of Defense, and Incyte during the conduct of the study, grants from Genentech, and personal fees from AstraZeneca and Mallinkrodt outside the submitted work; in addition, J.M. Balko has a patent for MHC-I/II expression to predict immunotherapy outcomes issued, licensed, and with royalties paid from TheraLink. No disclosures were reported by the other authors.

B.C. Taylor: Conceptualization, data curation, formal analysis, visualization, methodology, writing–original draft, project administration, writing–review and editing. X. Sun: Conceptualization, data curation, formal analysis, methodology, writing–original draft, project administration, writing–review and editing. P.I. Gonzalez-Ericsson: Data curation, formal analysis, visualization, methodology. V. Sanchez: Data curation, formal analysis, methodology. M.E. Sanders: Data curation, formal analysis, methodology. E.C. Wescott: Data curation. S.R. Opalenik: Data curation, methodology, writing–review and editing. A. Hanna: Data curation. S.-T. Chou: Data curation. L. Van Kaer: Visualization, methodology. H. Gomez: Data curation. C. Isaacs: Data curation, supervision. T.J. Ballinger: Data curation, supervision. C.A. Santa-Maria: Investigation, project administration. P.D. Shah: Data curation, supervision. E.C. Dees: Supervision, project administration. B.D. Lehmann: Data curation, formal analysis, supervision. V.G. Abramson: Data curation, supervision. J.A. Pietenpol: Data curation, supervision. J.M. Balko: Conceptualization, formal analysis, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

Funding for this work was provided by NIH/NCI SPORE 2P50CA098131–17 (to J.A. Pietenpol), Department of Defense Era of Hope Award BC170037 (to J.M. Balko), and the Vanderbilt-Ingram Cancer Center Support Grant P30 CA68485 (to J.M. Balko). Additional funding was provided by NIH/NCI T32 CA009592 (to B.C. Taylor) and Susan G. Komen ASP231038783 (to J.M. Balko and B.C. Taylor).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.