Reversible downregulation of HLA class I in adenoid cystic carcinoma

Abstract

Background

Adenoid cystic carcinoma (ACC) is a rare, but lethal cancer with low response rates to systemic therapies, such as cytotoxic chemotherapy and immune-checkpoint inhibitors (ICIs). Despite extensive clinical trials, no effective treatments for patients with recurrent or metastatic ACC are available, and ACC mortality rates remain poor.

Methods

We employed automated multiplex immunofluorescence (mIF), single-cell RNA sequencing (scRNA-seq) Gene Expression analysis, RNA in-situ hybridization, and spatial transcriptomics analysis to characterize the immune landscape of ACC tumors, ACC metastasis, and normal tissues from regions where ACCs arise. Based on results from these studies, we treated freshly resected ACCs with interferon-γ or a stimulator of the interferon genes (STING) agonist in vitro. Additionally, we included one patient with ACC in a phase 1 clinical study of a novel STING agonist (dazostinag) plus pembrolizumab.

Results

The mIF analysis revealed that ACC tumors are immunologically “cold”, with few tumor-infiltrating T-lymphocytes and low programmed death-ligand 1 (PD-L1) expression. The most striking finding was a very low beta-2-microglobulin (B2M) expression in nearly all ACCs, with only focal expression found in some ACC metastases. mIF and RNA sequencing analyses of normal salivary gland and breast tissues revealed a p63+, NFIB+, basal duct cell population, with similarly low B2M/human leukocyte antigen (HLA) class I expression. Spatial transcriptomics analysis of the focally B2M-positive ACC metastases uncovered the genetic pathway driving upregulation of B2M, an interferon-γ program mediating the reintroduction of HLA-I/B2M; the significantly upregulated genes included IRF1, GBP1, and TAP1. On short-term treatment of primary ACC tissues in vitro with interferon-γ or a STING agonist, we observed strongly upregulated HLA class I/B2M expression. Moreover, treatment of a patient with recurrent, metastatic breast ACC with a STING agonist and pembrolizumab led to a partial response with a 70% tumor reduction.

Conclusions

Low B2M/HLA class I expression may explain why ACCs are immunologically cold and the lack of response to ICIs. Our findings suggest that the normal cell of ACC origin exists in a B2M/HLA-class I low state, and that pharmacologic manipulation with immune activators, such as STING agonists, can restore HLA/B2M in ACCs, as supported by the promising response observed in a patient with metastatic ACC. These findings indicate a potential path to urgently needed immunotherapies.

Correspondence to Professor A. John Iafrate; [email protected]

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Adenoid cystic carcinoma (ACC) is a rare cancer, with poor response rates to systemic therapies, including immune checkpoint inhibitors. Currently, no standardized therapy is available for recurrent or metastatic disease. We characterized the ACC immune landscape to understand the mechanisms underlying these tumors’ poor response rates to immune-checkpoint inhibitors.

WHAT THIS STUDY ADDS

• We examined the previously described “cold” immune landscape of ACCs and identified a very low beta-2-microglobulin (B2M)/human leukocyte antigen class I expression that most likely contributes to the low number of infiltrating T lymphocytes observed in ACC tumors. However, focally B2M-positive ACC metastasis revealed a possible reversibility of this downregulation, which we were able to achieve through pharmacological intervention with interferon-γ or a stimulator of the interferon genes (STING) agonist. Treatment with a novel systemic STING agonist (dazostinag) and pembrolizumab yielded a partial response with a 70% tumor reduction in a patient with ACC with metastatic disease.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The findings from this study provide novel insights into the immune landscape of ACCs, their likely cell of origin, and address the pressing need for systemic treatments for patients with advanced disease.

Introduction

Adenoid cystic carcinoma (ACC), a rare epithelial cancer of the salivary glands, accounts for approximately 1% of all head and neck cancers.¹ It less frequently arises in the secretory glands of other organs, including the lacrimal glands, tracheobronchial tree, and breast.^2–4 ACC most often exhibits a slow, but progressive disease course, due to its ability to spread through perineural invasion^{5 6} and hematogenous dissemination^{7 8} to distant organs. Its three distinct histological subtypes—tubular, cribriform, and solid—represent increasing aggressiveness and decreasing overall survival.^{9 10} Even with appropriate adjuvant therapy following surgery, local recurrence is common, with increased risk seen with the solid histological subtype, perineural invasion, positive surgical margins, or advanced stage at diagnosis.¹¹ Studies have found that 40–70% of patients with ACC develop distant metastases, most often to the lungs, but also to liver, bones, and brain, with clinical behavior ranging from indolent to highly aggressive.^{7 12} The standard of care for treatment of patients with ACC is surgery and adjuvant radiotherapy.¹³ Systemic therapies in the adjuvant setting are currently recommended only in the context of clinical trials or in cases of metastatic disease.^{13 14}

Chromosomal translocations, found in approximately 60% of patients with ACC, include the translocation t(6;9), leading to the MYB-NFIB fusion, or rarer t(8;9) translocations, leading to MYBL1-NFIB fusions.^15–17 As these translocations are thought to drive ACC pathogenesis, extensive molecular profiling studies have sought to clarify the mechanisms by which they and other oncogenes, such as NOTCH1 mutations, drive progression.^{16 18 19} This work, however, has yet to impact clinical decision-making for patients with targeted therapies for ACC. Approaches based on cytotoxic agents and VEGFR inhibitors, as well as immune checkpoint inhibitors (ICI), have also shown low response rates and are associated with significant toxicities, as well as a major impact on quality of life.^20–22

Such challenges are highlighted in a 2022 clinical review that examined the outcomes of 55 clinical trials focused on molecular targets for locally recurrent or metastatic ACC.²³ The approaches to the most common mutational targets and targeting therapies had been designed to: (1) interrupt tumor cell proliferation and survival by targeting c-KIT, EGFR, FGFR, or NOTCH1; (2) prevent neoangiogenesis by targeting VEGF; (3) disrupt immune checkpoint evasion by targeting PD-1; and (4) exploit other targets, such as PRMT5 and ATRA, to develop novel, effective strategies against ACC. None of the included 55 studies yielded a complete response in any patient, and a partial response was found only in a low fraction of patients in a few studies. For example, a partial response rate of 16% (5 of 32 patients) was observed in a trial of the multikinase inhibitor lenvatinib; still, this meager response rate, as the most promising observed, led to a category 2B recommendation in the National Comprehensive Cancer Network (NCCN) Head and Neck Cancer guidelines.^{14 21 23} These data illustrate that, despite robust, ongoing efforts, access to effective treatment for patients with unresectable or recurrent tumors largely remains out of reach.

Several challenges have limited the study of ACC biology. Chief among them is a lack of validated cell lines, rendering many studies reliant on human tumor tissue biopsies and animal models. Additionally, the rarity of ACC makes large cohort studies unfeasible, leading to unpredictability in diagnosis and treatment. However, computational tools now assist discoveries relevant to ACC. A recent proteogenomic study of salivary gland ACCs revealed two molecular ACC subtypes potentially useful for clinical prognostication.¹⁹ This study showed that ACC-I, with a 37% prevalence, is characterized by a strong upregulation of MYC target genes, enrichment of NOTCH-activating mutations, and a more aggressive disease course. By contrast, ACC-II, with a 63% prevalence and upregulation of both TP63 and the receptor tyrosine kinases (AXL, MET, EGFR), is associated with a more indolent disease course. Not only were MYC and TP63 sufficient, in this study, to distinguish the two groups, but overlapping, actionable protein/pathway alterations were identified for each subtype.¹⁹

Similarly, a transcriptomics study of salivary gland ACC evaluated paired normal and tumor tissues from 15 patients.²⁴ This study identified a new gene fusion, TVP23C-CDRT4, upregulation of TVP23C and CDRT4 in solid ACC tumors, and the well-known MYB-NFIB and MYBL1-NFIB fusions. In addition, infiltration of T cells, B cells, and natural killer cells in ACC tissues was observed, as was significant upregulation of the nuclear receptor transcription regulator PRAME, alongside downregulation of antigen-presenting human leukocyte antigen (HLA) genes.²⁴

Our ability to analyze the previously described “cold” immune environment of a rare tumor, such as ACC,^25–27 that responds poorly to immune therapies, has been greatly improved by automated multiplex immunofluorescence (mIF) platforms, which permit multiparametric studies of single tissue sections. The COMET automated mIF platform (Lunaphore Technologies SA), which supports staining up to 40 markers on the same tissue slide, confers unique advantages for examining the spatial relationships between different protein markers and their context within different cellular compartments (eg, normal tissue, tumor and tumor-adjacent stromal tissue, and tumor-infiltrating lymphatic structures). This spatial context is critical for investigating mechanisms that may support the absence of T-lymphocyte infiltrates observed in tumors thereby deemed “cold”.²⁸ Additionally, a spatial transcriptomics platform with single-cell resolution, the 10x Genomics Visium HD, now permits additional, in-depth analyses of gene expression to address critical questions of spatial heterogeneity and tumor-stromal interactions.

Our study leveraged mIF and spatial transcriptomics to explore the immune landscape of ACC tissues in their spatial context. We sought to better understand why ACCs are resistant to ICIs by identifying pathways potentially supporting these tumors’ cold immune environment, as well as molecular characteristics and novel prognostic biomarkers able to differentiate populations of patients with ACC and to advance effective, personalized therapies.

Results

ACCs are immunologically “cold”

To understand the immune landscape of ACCs, we performed a detailed analysis of a cohort of ACC tumors from our pathology archive, utilizing the COMET mIF system, with a panel of antibodies (including T cell, B cell, and macrophage markers, as well as pan-cytokeratin to identify tumor cells, B2M, and Ki67) (figure 1A–G). This cohort included a total of 24 ACC cases from 23 patients: 15 head and neck ACCs, 2 lung (ACC19 and ACC20), 2 breast (ACC17 and ACC18), 4 metastases from head and neck ACCs (ACC21, ACC22, ACC23, and ACC24), and 1 ACC from an unknown primary site (ACC7) (online supplemental table 1). Additionally, we analyzed a set of five comparator tumors, including two human papillomavirus (HPV)-negative head and neck squamous cell carcinomas (HNSCC) of the oral cavity, two HPV-positive oropharynx HNSCC, and one basal-like carcinoma of the breast. The HNSCCs were chosen as control tumors because they, like ACCs, arise in the immune environment of the oral mucosa. The basal-like carcinoma of the breast served as a control tumor for ACCs arising from other organ sites as it is an ACC mimic, showing some similar histological features. 17 of 23 patients with ACC underwent RNA-based fusion detection, finding 13 fusion-positive cases (online supplemental table 2). All tumors were reviewed by a head and neck subspecialty pathologist (WCF), who verified the diagnoses and ACC histological subtypes. Using a custom image analysis pipeline, we segmented and classified single cells into different cell types based on manually-thresholded values to define individual cell types.

In the ACC tumor cohort, we observed a complex and highly variable tumor immune landscape, with the proportion of immune cells (B cells, T cells, and macrophages) to total cells in the tumor area ranging from 1.5% to 34.4%, with a mean of 7.55% (95% CI (4.59%, 10.52%)). This proportion was lower than that of the five non-ACC comparison cohort cases, whose immune cell to total cell ratios were 12.5–22.1%, with a mean of 15.57% (95% CI (10.8%, 20.35%), p value=0.02) (figure 1H). When examining the composition of the immune cells, CD8+ cytotoxic T cells showed a significantly lower abundance, at 14.54% (95% CI (11.0%, 18.08%)), in the ACC cases versus 23.55% (95% CI (10.32%, 36.78%)) in the control tumor cohort (p=0.046) (figure 1I). A non-significant trend of lower percentages in ACCs versus control tumors could be seen for T-helper cells, at 25.61% (95% CI (17.61%, 33.61%)) versus 31.21% (95% CI (13.87%, 48.55%), p=0.54) and for B cells, at 2.72% (95% CI (1.17%, 4.27%)) versus 5.5% (95% CI (−4.93%, 15.93%), p=0.13). CD68+ macrophages represented the only individual cell type showing a trend of greater abundance in ACCs compared with control tumors, comprising 57.12% of immune cells seen in ACC cases (95% CI (47.27%, 66.91%)) versus 39.73% of immune cells in control tumors (95% CI (5.2%, 74.27%), p=0.15) (figure 1I).

These data must be considered in the context that most ACCs arise from the major and minor salivary glands, which are adjacent to the mucosa of the oral cavity and can display an abundance of tissue-resident mucosal immune cells (representative image online supplemental figure 1A). We were concerned that this proximity had led to overestimation of the tumor-associated immune infiltrate, and thus examined and quantified immune cells within the malignant epithelium itself, limiting our analysis to regions of interest (ROI) defined by cytokeratin-positive tumor glands (100 glands per case). When examining the amount of total immune cells in the malignant epithelium (figure 1J), almost all ACC cases showed lower immune cell infiltration than did the control tumors. Strikingly, and in contrast to the analysis of total T cells in the general tumor sample area, the number of T cells within the malignant epithelium was significantly lower in the ACCs than in the control tumors. For ACCs, the percentage of T cells to epithelial cells within the malignant epithelium was 0–1.72%, with a mean of 0.48% (95% CI (0.27%, 0.69%)); for control tumors, that percentage was 2.33–12.9%, with a mean of 5.81% (95% CI (0.56%, 11.06%), p<0.001), indicating that ACCs are immunologically cold (figure 1K).

We also observed a difference between ACCs and control tumors in the ratio of intraepithelial B cells (0.036% vs 0.72%, p=0.003) and of intraepithelial macrophages (1.65% vs 6.11%, p=0.007). For ACCs, a low percentage of Ki67+ proliferating CD8+ cytotoxic T cells to total T cells was observed in the total tumor area, with a mean of 5.87% (95% CI (3.77%, 7.97%)), while for the control tumors, the percentage was 27.94% (95% CI (0.81%, 55.07%), p<0.001) (online supplemental figure 1B). These findings indicate a more active, expanding subset of CD8+ T cells in the control tumors than in the ACCs.

Loss of major histocompatibility complex class I/beta-2-microglobulin may underlie the lack of ACC immune infiltrate

The lack of tumor-infiltrating T cells led us to examine several common mechanisms of immune avoidance, specifically: (1) programmed death-ligand 1 (PD-L1)/programmed death-ligand 2 (PD-L2) expression on tumor cells, (2) the presence of regulatory T cells in the tumor environment, and (3) major histocompatibility complex (MHC) class I/beta-2-microglobulin (B2M) expression on tumor cells. The antibody panel on the COMET mIF platform includes PD-L1, whose expression correlates with immune evasion and provides a principal biomarker of immunotherapy response. We used tumor cell fluorescence signals to calculate the combined positivity score (CPS) and tumor proportion score (TPS) for PD-L1. 23 of 24 ACC cases were negative for PD-L1 TPS, with scores of <1% and an average score of 0.24%, 95% CI (0.12%, 0.35%), while the control tumors showed TPS scores ranging from 15.89% to 56% (40.28%, 95% CI (20.73%, 59.83%)) (p<0.0001) (figure 1L). For the PD-L1 CPS score, six ACC cases scored just above the >1 threshold and had an average score of 0.57, while for the control tumors, the CPS scores ranged from 19.76 to 69.8 (49.6, 95% CI (24.83, 74.37)) (p<0.0001) (figure 1L). These findings suggest that PD-L1 expression likely does not play a major role in ACC immune avoidance, consistent with prior publications.^29–31

PD-L2 is also a ligand of programmed cell death protein 1 (PD-1), but its role in tumor development and progression is still not fully understood. We added PD-L2 to our panel to analyze a subset of ACC cases and our control cohort. We first considered only membranous staining and observed 4 of 11 ACC cases with PD-L2 positivity (CPS>1). However, we observed that including cytoplasmic staining revealed that 9 of 11 ACC cases exhibited PD-L2 positivity. When using previously published cut-offs for PD-L2 in ACC,^{26 27} these cases showed TPS>10% and CPS>10, as did all five control tumor cases (online supplemental figure 1C). High PD-L2 expression in ACC (largely cytoplasmic) has been previously described^{26 27 32}; however, in our cohort, the positivity of PD-L2 in ACC cases was highly variable and did not show a correlation with T lymphocyte infiltration (R=0.11, p=0.75) (online supplemental figure 1D). Since our control HNSCCs were also positive, we consider it unlikely that PD-L2 expression, which was predominantly expressed in the cytoplasm, is the major driving mechanism behind the cold immune environment of ACCs.

Since abundant regulatory T cells (Tregs) are known to create an immunosuppressive microenvironment, we quantified the percentage of FoxP3+regulatory T lymphocytes to total cells. Tregs were sparse in the ACCs (0.45%), but averaged 2.09% for the non-ACC cases (p<0.001) (online supplemental figure 1E). Similar percentages were seen for the number of Tregs compared with the total number of immune cells: in ACCs, Tregs made up 5.88% of the immune cells (95% CI (3.4%, 8.36%)), whereas HNSCCs and the basal-like carcinoma showed an average ratio of 12.36% (95% CI (5.17%, 19.56%)) (p=0.041) (online supplemental figure 1F). This trend continued for percentages of Tregs to T-helper cells, at 20.84% for ACCs, but significantly higher, at 43.39%, for control tumors (p=<0.001) (online supplemental figure 1G). Taken together, these findings suggest that Tregs are unlikely to play a major role in regulating the cancer immune environment of ACCs.

As the loss of MHC-I/B2M is a well-described mechanism of tumor immune evasion, we included B2M in our mIF panel. A near-complete absence of B2M expression was observed in 20 of 24 ACCs (figure 2A,B). Additionally, staining the ACCs with an anti-HLA-A/B/C antibody confirmed that loss of B2M correlated with decreased MHC class I expression (figure 2C, online supplemental figure 2A–F). This contrasted with clear B2M positivity in four of the five non-ACC control tumors (figure 2B), with only one HNSCC lacking B2M (online supplemental figure 3A). In some ACC cases, the HLA-A/B/C expression showed more positivity and heterogeneous expression levels than were seen for B2M, but the levels were still substantially lower than those of control tumors and adjacent normal tissues (figure 2C, online supplemental figure 4).

Only four ACC cases were identified as having very minimal focal B2M expression at the edge of the tumor (ACC21, ACC22, ACC23, and ACC24). These four cases were metastases, two to cervical lymph nodes and two to the lung, suggesting a unique interaction of the metastatic site’s microenvironment with tumor cells at the edge of the tumor mass. The four metastases with focal B2M expression further suggested that downregulation might be reversible. The B2M-positive foci show increased infiltration of T lymphocytes in the tumor glands (online supplemental figure 3B,C and online supplemental figure 5A,B), as well as the highest numbers of T cells in the malignant epithelium of the ACC cohort (1.52%, 1.72%, 1.29%, and 1.03%) (figure 1K).

Mechanism of downregulation of MHC class I and B2M in ACCs

B2M can be downregulated in certain tumors through nonsense mutations, increased protein degradation, or decreased transcription.^{33 34} However, examination of published ACC exome sequencing data sets in the COSMIC database of cancer variants revealed no B2M mutations,³⁵ suggesting that the lack of B2M expression in the cases we examined was unlikely to have occurred through acquired mutations. However, the samples analyzed in this study did not have whole exome sequencing data available. As many ACCs can have large amounts of lymphoid tissue and stroma, possibly contributing to B2M levels in bulk RNA extracts, we analyzed messenger RNA (mRNA) expression levels in situ at a single cell level using RNA in-situ hybridization (RNA-ISH) with a B2M-specific probe. In the cancer cells, we observed very low levels of B2M RNA, whereas a clear maintenance of B2M expression levels was seen in the adjacent immune cells (figure 2A,D). This finding strongly suggested that the lack of B2M protein was a result of decreased gene transcription, rather than changes in protein translation or protein stability.

We interrogated the impact of the ACC-associated MYB-NFIB fusion on B2M transcription. MYB is known to be highly expressed in ACCs because of this fusion, but nuclear factor I B (NFIB) expression is not as well studied. We examined 21 of 24 ACC cases in our cohort for NFIB expression levels through immunohistochemistry (IHC), and found NFIB to be highly overexpressed in all 21 ACC cases (figure 2E). Since both genes are transcription factors, their high expression suggested one or both might have transcriptional suppressor activity with respect to the B2M gene. We searched for potential MYB or NFIB binding sites at the B2M promoter, identifying two potential NFIB binding sites between the transcriptional start site and the well-annotated interferon (IFN) response element site (online supplemental figure 6A).

To test whether NFIB was suppressing B2M expression, we knocked down NFIB using lentiviral shRNA transduction in five cell lines. Since no confirmed representative ACC cell line models exist, we performed lentiviral shRNA-induced knockdown of NFIB in lines with varying levels of NFIB. Three of these cell lines (MCF7, NCI-H446, and S6) showed similar patterns of expression to the ACC cases, with high NFIB and low B2M levels (online supplemental figure 7A). Knockdown of NFIB did not impact B2M protein expression in any line (online supplemental figure 7B, online supplemental figure 8A,B), nor did overexpression of NFIB lead to change in B2M expression (online supplemental figure 7C, online supplemental figure 8C,D). We also used CRISPR-Cas9 to knock out the two predicted NFIB binding sites at the B2M promoter. We performed single-cell cloning of cell lines with edits to these sites in three cell lines, MCF7, 293T, and NCI-H446. No increase in B2M protein expression was seen in any of the clones with successful editing at these sites (online supplemental figure 6B–D, online supplemental figure 8E). In fact, editing of the predicted site closest to the transcriptional start site (guide RNA (gRNA) a) resulted in decreased expression of B2M (online supplemental figure 6B–D, online supplemental figure 8E).

We also investigated the influence of MYB expression on B2M expression using siRNA-mediated MYB knockdown in MCF7. Western blot analysis revealed no effect of the MYB knockdown on B2M expression (online supplemental figure 7D, online supplemental figure 8F,G). MYB only has two low-confidence binding sites in the B2M promoter (JASPAR score 243 and 249), so we did not proceed with CRISPR/Cas9 editing of that site (online supplemental figure 7E). Taken together, these data indicate that B2M is not downregulated because of direct transcriptional suppression by either MYB or NFIB. B2M (14 kDa), MYB (80 kDa), and NFIB (47–60 kDa) displayed the expected molecular weights on western blot.

We explored a second hypothesis, that the ACC cell of origin may have been a B2M-low precursor cell. Reanalysis of a published ACC single-cell RNA-seq data set confirmed that ACC tumor cells are HLA/B2M low.^{36 37} Inspection of the Uniform Manifold Approximation and Projection (UMAP) analysis revealed four clear tumor subclusters, consistent with the known phenotypic variability of ACC tumor cells. One showed myoepithelial/basal markers (TP63+, NFIB+, ACTA2+) (cluster C2), while another two showed luminal markers (KIT+) (C0, C7) (figure 3A–D). A fourth subcluster (C3), intermediate between these three clusters, was defined as an NFIB-intermediate and HLA/B2M-very low state (figure 3C,D). Reanalysis of normal salivary gland single-cell RNA sequencing (RNA-seq) data from the Human Protein Atlas (HPA)^{38 39} (figure 3E), and specifically of salivary gland duct cells (figure 3F), also revealed two TP63-high subclusters (C0 and C5), both of which showed significantly lower B2M than the KIT-high subcluster (p<0.001) (figure 3G,H).

We asked whether we could validate these observations in the glandular ducts of a normal salivary gland sample and a normal breast sample by using an expanded 19-plex Lunaphore mIF panel (adding MYB, SOX2, cytokeratin 5, and p63 to include additional stem cell/basal cell markers), and were able to identify these as p63-high and NFIB-high basal cells that showed very low expression of HLA (representative breast duct in figure 4A–C). In the salivary gland, we identified a corresponding NFIB-high and HLA-low population of basal cells in the salivary gland ducts (figure 4D–G). Prior reports provide evidence that the basal cells of the intercalated ducts are the cells of origin of ACCs.^{40 41} A recent report also found p63 to be upregulated in a subtype of ACC with a less aggressive clinical course than other ACC subtypes.¹⁹ Detailed analysis of breast samples revealed a similarly NFIB-high and HLA-low basal cell population in the ducts; in breast, these cells are also smooth muscle actin-positive and of myoepithelial differentiation (figure 4H–K). Earlier reports have also described such a subpopulation of basal cells in human skin, with low/negative expression of MHC class I.⁴² These data support a model in which HLA/B2M expression is low in ACCs because the stem cells that transform into ACC naturally show low expression: our data thus indicate that the MYB-NFIB fusion might not actively downregulate HLA/B2M expression.

Genetic pathways regulating focal HLA class I/B2M upregulation in ACC metastasis

The four ACC metastases in our cohort showed focal upregulation of B2M, suggesting that its expression could be reinduced in the cancer cells, and that the tumor’s interaction with the local microenvironment at the metastatic site was likely driving a reinduction of B2M. Due to the spatial heterogeneity of B2M expression in the tissue, we applied a spatial transcriptomics platform with single-cell resolution, 10x Genomics Visium HD, to investigate the gene expression programs driving the upregulation of B2M and HLA class I. Two of the four ACC metastases with regions of focal B2M upregulation (ACC23 and ACC24) were run on the Visium HD platform using 8 µm binning. For ACC24, a lung metastasis, analysis of the spatially-encoded Visium HD data identified 10 clusters (figure 5A,B), including stromal and immune cell clusters, a cluster of type II pneumocytes, and 3 tumor clusters. B2M expression was high in the immune cells, endothelial cells, fibroblasts, and type II pneumocytes (figure 5C,D), as well as in one tumor cluster (C5). The C5 cluster could be mapped back to the region of upregulated B2M/HLA-class I in our mIF analysis by comparison with the scout histology image.

We used differential gene expression to identify genes significantly upregulated in cluster C5 relative to the other two B2M-low tumor clusters (figure 5E,F), followed by Gene Set Enrichment Analysis (GSEA) on the logFC-ranked gene list (figure 5G,H). GSEA revealed 143 significantly positively enriched Gene Ontology (GO) biological process gene sets, the overwhelming majority of which were related to general immune activation, notably “antigen processing and presentation” and both “response to type I interferon” and “response to type II interferon”. Terms were clustered based on semantic similarity to reduce redundancy (online supplemental figure 9), revealing several more subtly enriched processes related to cell adhesion, migration, and proliferation. Analysis of ACC23, a lymph node metastasis, yielded highly concordant results, including a similar group of upregulated genes enriched for “response to type II interferon” and other immune-related biological processes (online supplemental figure 10).

We selected a subset of the shared, differentially upregulated genes - TAP1, IRF1, and GBP1 - based on their functional relevance in immune activation, meaningful change in expression of log2FC>2 and availability of antibodies. We validated these genes with immunohistochemistry and found an upregulation in expression in the HLA-class I positive region compared with the HLA-class I negative areas in the ACC metastasis (figure 5I–L, online supplemental figure 11A–D). GBP1 and IRF1 are directly inducible by IFN-γ; they regulate cellular immune response and cytokine production and can lead to a higher inflammatory state. TAP1 plays an important role in immune response, as well as antigen processing and presentation via MHC class I. Therefore, these results suggest that increased IFN-γ signaling in the metastatic microenvironment of ACCs can reinduce the expression of HLA class-I and B2M.

Pharmacologic rescue of HLA and B2M in ACC short-term cultures

Our mIF and spatial transcriptomics data for the four focally B2M-positive ACC metastases suggested that B2M low expression could be modulated; therefore, we sought to test several pharmacologic approaches known to stimulate the expression of immune genes, such as IFN-γ, a synthetic systemic stimulator of the interferon genes (STING) agonist, and proteasome inhibitors. As no ACC cell line models exist for in vitro drug tests, we employed a short-term culture model to evaluate the effects of in vitro drug treatments on HLA and B2M expression in four freshly resected ACC tissue specimens (online supplemental figure 12A).

Immediately following surgery, the ACC tissue was manually dissected, divided into tissue culture plate wells, and incubated for 48 hours in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS) with either dimethyl sulfoxide (DMSO), IFN-γ (5 ng/mL), the STING agonist ADU-S100 (15 µM), pembrolizumab (10 µg/mL), or bortezomib (50 nM). IFN-γ and ADU-S100 both resulted in a fourfold upregulation of HLA and B2M protein (figure 6A,B, online supplemental figure 12B,C, online supplemental figure 13A,B). The reintroduction of protein expression was primarily detected at the tumor edges, likely due to limited drug diffusion deeper into the tissue, and the drug therefore acting at the tumor edges. RNA-ISH analysis confirmed extremely strong upregulation of B2M transcript levels in the tumor tissue (figure 6C, online supplemental figure 12D, online supplemental figure 13A,B). IHC analysis of IFN-γ and ADU-S100-treated tumors revealed a weak upregulation of the immune checkpoint protein PD-L1 (figure 6D, online supplemental figure 12E, online supplemental figures 13A,B). These data from short-term cultures indicate that treating ACCs with IFN-γ or STING agonists could restore HLA/B2M expression and possibly result in an antitumor immune response. PD-L1 checkpoint induction suggests that any such pharmacologic therapy may need to be combined with anti-PD-1 or anti-PD-L1 antibodies; otherwise, any antitumor effect may be self-limiting. By contrast, treatments with pembrolizumab or bortezomib led to no significant change in B2M, HLA class I expression, or PD-L1 expression (online supplemental figure 12G–I). These four cases treated in vitro included three NFIB-MYB fusion cases and one case with a NOTCH1 mutation (figure 6E, online supplemental figure 12F,J).

STING agonist treatment of a patient with ACC led to a partial response

Based on these findings, we were able to enroll and treat a patient with recurrent/metastatic ACC in the dose escalation cohort of a phase 1 clinical study of a novel STING agonist (dazostinag, TAK-676) plus pembrolizumab (NCT04420884). This patient, in their late 50s, presented with a long-standing, neglected, right breast mass first noticed in the 1990s. In October 2022, the patient had an episode of aphasia and was diagnosed with ischemic stroke. During the patient’s evaluation, a large, fungating, bleeding right breast mass was found, measuring 14.9×16.2×20 cm. Biopsy revealed ACC with positive MYB expression by immunohistochemistry. Next-generation sequencing revealed an MYB:NFIB gene fusion, and based on our findings from the public sequencing data set, we assumed the absence of mutations in B2M and HLA class I genes. CT revealed subcentimeter pulmonary nodules up to 8 mm in size, suspicious for distant metastases, in addition to the right breast mass.

The patient received chemoradiation to the right breast, for a total of 70 Gy, and weekly, concurrent cisplatin at a dose of 40 mg/m². A right mastectomy and latissimus dorsi flap reconstruction was subsequently performed. The primary tumor, analyzed using the mIF platform, revealed very low expression of HLA class I/B2M and PD-L1, limited immune cell infiltrate, and positive MYB, NFIB, p63, and SOX2 expression (figure 7A), corresponding to the expression pattern typical of the ACC tumor tissue cohort. In December 2023, the lung nodules increased, with the largest nodule, measuring 25×22 mm (right lower lobe), pathologically confirmed as ACC.

Palliative systemic therapy options were reviewed. Given the recent cerebrovascular accident, vascular endothelial growth factor tyrosine kinase inhibitors, commonly used in advanced ACC, were contraindicated. The patient enrolled in the phase 1 study of a synthetic systemic STING agonist, dazostinag, dosed at 14 mg administered weekly by intravenous infusions, in combination with intravenous pembrolizumab dosed at 200 mg every three weeks (NCT04420884).⁴³ After the initial two doses, the STING agonist dose was reduced to 10.5 mg weekly for a non-study drug-related adverse event (the patient had a prolonged hospitalization for seizures after missing a dose of levetiracetam and prior stroke, as well as a mild fever from witnessed aspiration pneumonia). The pulmonary nodules shrank, and a confirmed partial response was achieved. The patient continued to receive protocol treatment for 9 months. The most recent 9-month restaging scan showed a 70% reduction of the sum of target lesions from baseline (figure 7B). Due to the reduced lesion size after treatment, there was insufficient tumor to permit an on-treatment biopsy.

Discussion

ACC is typically a slow-growing, yet progressive malignancy of the salivary glands and, more rarely, other anatomic sites. In patients with recurrent or metastatic disease following surgery and radiotherapy, treatment options are severely limited.¹⁴ Despite a number of systemic therapies evaluated to date in clinical trials, activity has been modest at best. The NCCN guidelines recommend clinical trial participation, when available, as a preferred approach for recurrent or metastatic ACC. Other off-label treatment options that may be considered for select patients with recurrent or metastatic ACC include cytotoxic chemotherapy, lenvatinib, or axitinib in combination with avelumab.^{21 44} Therefore, we explored the ACC immune landscape using mIF to understand possible reasons that could explain why these tumors are largely unresponsive to immunotherapy. We found that ACCs are immunologically cold and have low expression of PD-L1, confirming several prior studies,^{26 45} and that this cold phenotype may result from downregulation of B2M, concealing the tumor from T cell recognition. Using short-term tumor cultures, we further showed this downregulation to be reversible with STING agonist or IFN-γ treatment, potentially revealing new therapeutic approaches to this challenging cancer.

Our study is not the first to describe ACCs as cold tumors. A 2021 study used RNA-seq to evaluate ACCs and found a universal low expression of T cell checkpoint proteins, a scarcity of tumor antigens, and a clinically relevant association between the cold ACC tumor microenvironment and higher likelihood of recurrence.⁴⁵ Additionally, they reported an immune-excluded environment for ACCs, M2-polarized macrophages and myeloid suppressor cells, as well as a low mutational load.⁴⁵ Underscoring the challenges of analyzing rare tumors through scarce biopsy resources and the lack of validated models, another recent study observed a predominantly cold environment with low expression of PD-L1 and an elevation of M2 macrophages.³⁰ RNA-seq of 62 patient samples found most ACC tumors to be immunologically cold, but about 30% to be “hot”; however, spatial analysis revealed a restriction of immune cells to the stroma.⁴⁶ Combined with our data, these studies provide mechanistic insight into why ACCs do not respond to ICI therapy.

The downregulation of B2M and HLA is the predominant immune signal we observed in ACCs; however, we also observed a trend of higher percentages in CD68+ macrophages as compared with control tumors. Since macrophages can have a role in creating an immunosuppressive microenvironment, this is certainly an area worthy of additional investigation. One important observation was that the total number of immune cells in the ACC tumor differed only slightly from the control HNSCC tumors; only when measuring T cells within the epithelium itself did we see a striking difference. A likely explanation is that normal salivary glands can have adjacent lymphoid structures, and that the tumor grows into this stromal environment without invoking an antitumor response. This explanation is consistent with histologic observations of many tumor glands infiltrating directly into lymphoid follicles. It is unclear why these close approximations do not induce B2M in the ACC cells.

MHC class I downregulation/B2M loss as a mechanism of immune evasion has been described in a number of viral infections and in cancer.^{33 47 48} Multiple virally encoded proteins have been shown to directly interfere with the expression of HLA/B2M, or to interfere with other components of the antigen presentation machinery.⁴⁹ Recent studies have focused on acquired mutations in B2M in tumors resistant to immunotherapies, and on loss of heterozygosity of the HLA locus in tumors resistant to immunotherapies.⁵⁰ Mutations in HLA/B2M have not been observed in the large number of ACCs studied with exome sequencing.⁵¹ Our analysis strongly supports a mechanism in which B2M is downregulated via transcription silencing. We observed that B2M downregulation is not linked to the transcriptional activity of the MYB-NFIB gene fusion, as neither knockdown of NFIB and MYB nor CRISPR/Cas9 editing of putative NFIB binding sites in the B2M promoter had any impact on B2M expression. We extrapolated from other NFIB and MYB-expressing tumor lines, since no validated ACC cell lines currently exist. For in vitro drug treatments, the short-term tissue culture provided a sufficient representation of tumor response, similar to that of other models, such as tumor organoids and xenograft models, which are similarly unable to fully recapitulate the human immune response.⁵²

Ultimately, ACCs likely do not involve active downregulation of HLA/B2M. Instead, they show a lack of HLA/B2M expression, as do the cells of origin for ACCs—the basal cells of the salivary gland ducts—also known as the duct reserve cells.^{40 41} At baseline, these basal cells exist in an HLA/B2M-low state. Studies have shown that the intercalated duct cells potentially act as transient amplifying progenitor cells and can differentiate into acinar and granular duct cells.^{53 54} Single-cell RNA sequencing (scRNA-seq) and mIF analysis support these basal cells as stem-like cells, able to give rise to both the myoepithelial and luminal cells that comprise ACCs. These basal duct cells have a neural crest-like expression pattern (NFIB+, SOX2+, p63+), interestingly shared by basal cells of the normal breast ducts. Other cells share this staining pattern, suggesting that other tumors may also have this natural HLA/B2M-low state. We do not know why these NFIB/SOX2/p63 expressing cells have low expression of HLA/B2M, but it likely reflects a general silencing of immune genes in the tumor cells, possibly because of critical signaling pathways, such as nuclear factor kappa B (NF-κB) signaling.

The clinical response of patients with ACCs to ICIs has been underwhelming. In fact, the most recent clinical trials of immunotherapies in this disease have shown very low response rates (2 of 32 patients with ACC), even when using an aggressive combination of ipilimumab and nivolumab.⁵⁵ The study authors showed lower immune cell infiltration in the ACC tumors, and responding salivary gland tumors demonstrated the preferential loss of mutational variants with stronger HLA-binding affinity, and a pre-existing, clonally skewed intratumoral T cell receptor (TCR) repertoire⁵⁵; however, the authors did not appear to consider the potential role of low MHC class I in their tumors.

Spatial transcriptomics analysis of ACC metastasis revealed upregulation of IFN-γ induced pathways in regions of focally high HLA/B2M as compared with regions that were HLA/B2M-low. These findings suggest that: (1) the B2M-low state is reversible, and (2) the IFN pathway plays an important role in mediating upregulation of B2M. Our data indicate that such a restoration of HLA/B2M is required in tumor cells before checkpoint inhibition can have a reasonable chance of success.

In our study, pharmacologic manipulation of short-term ACC cultures with IFN-γ or a STING agonist each resulted in robust HLA/B2M gene expression and protein upregulation. The B2M promoter has a consensus IFN response element, providing evidence for a likely direct transcriptional upregulation of the immune activators via IRF3 or NF-κB. While systemic IFN-γ would be very challenging to implement clinically, due to side effects, several STING agonists now in clinical trials target various advanced solid tumors. These studies have largely focused on the ability of STING agonists to modulate the cells of the immune system. Our data suggest that perhaps of equal importance is the impact of immune activators on tumor cells. A similar effect of STING agonists on HLA-negative small cell carcinoma of the lung has been observed.⁵⁶ Earlier clinical trials involving the STING agonist focused on intratumoral injections and modulating immune responses to the tumor. However, more recent STING agonist trials administer the drug systemically, as seems necessary if the drug must modulate ACC tumor cells themselves, perhaps at metastatic sites. Given that STING agonists induce PD-L1 in tumor cells (and likely PD-1 on T cells), checkpoint inhibition will likely remain an important approach, in combination with immune activators. This hypothesis is supported by the robust clinical response of an index patient with recurrent and metastatic MYB:NFIB fusion-positive ACC originating from the breast gland, who achieved a partial response with a 70% confirmed tumor size reduction from baseline after a 9 months treatment course of STING agonist with pembrolizumab. The patient’s primary tumor showed the typical HLA/B2M negative phenotype with positive MYB and NFIB expression. Due to the limited size of the tumor after treatment, we were not able to obtain an on-treatment biopsy to examine the upregulation of HLA/B2M in the patient tumor. Therefore, other mechanisms of STING agonist and pembrolizumab treatment could have contributed to the partial response by modulating the immune environment.

In summary, we report that low/absent B2M and HLA class I expression explains why ACCs are immunologically cold, potentially explaining their lack of a systemic response to ICIs. Our findings suggest that the ability to restore B2M could lead to therapies for treating this particularly difficult-to-treat tumor, which essentially continues to have no effective systemic treatments despite extensive clinical trials. Future studies should explore the unique concept reported here—that the normal cell of ACC origin exists in a B2M/HLA-low state—as well as the possibility that other intractable tumor types may derive from similar B2M/HLA-low precursor cells. Finally, STING agonist treatment with dazostinag in combination with pembrolizumab has demonstrated a promising response in a patient with metastatic ACC. Expanded studies investigating the activity of immune activators, including STING agonists, in patients treated for recurrent and metastatic ACC are warranted.

Materials and methods

Patient cohort and study design

We selected 29 cases from 28 patients who had undergone biopsy or surgery and were diagnosed with ACC, HPV+ or HPV− HNSCC, or basal-like carcinoma at the Pathology Department of Massachusetts General Hospital (MGH) between January 2004 and March 2024. Formalin-fixed paraffin-embedded (FFPE) tissue slides, with their associated pathology reports, were obtained from the pathology archives. The informed consent for the archived tissue was waived, due to the minimal risk to patient privacy. The study was conducted in accordance with the Declaration of Helsinki. The patient cohort included 24 ACC cases (15 head and neck, 4 metastases from head and neck ACCs, 2 lung; 2 breast; 1 unknown site), four HNSCCs, and one basal-like carcinoma of the breast. The two HPV+ and two HPV− HNSCCs were selected to act as control tumors, because these heterogeneous tumors arise from a similar immune environment of the oral mucosa as ACCs. The basal-like carcinoma of the breast showed histological features similar to ACCs, and therefore served as a control tumor for ACCs arising from other organ sites, such as the breast and lung.

Multiplex immunofluorescence staining and imaging

mIF staining was performed on FFPE tissue slides on the COMET platform (Lunaphore Technologies SA). Prior to staining, antigen retrieval and deparaffinization were conducted simultaneously with the PT module (Epredia) by applying the Tris-EDTA-based Dewax and HIER Buffer H (TA-999-DHBH, Epredia) at a 1:15 dilution in distilled water for 1 hour at 102°C. Slides were then washed in Multistaining Buffer (BU07, Lunaphore Technologies) before staining.

Slides were first incubated in Quenching Buffer (BU08, Lunaphore Technologies) for 30 s at 37°C, next incubated in 4’,6-diamidino-2-phenylindole (DAPI) for 1 min at 37°C (1 µg/mL, 62248, Thermo Fisher Scientific), and then immediately washed in Imaging Buffer (BU09, Lunaphore Technologies). Three rounds of sequential imaging were then performed in DAPI, Cy5, and TRITC channels, with images captured using a 20× objective. After imaging, the slides were washed with a Multistaining Buffer and blocked in 1% bovine serum albumin (BSA) (9048-46-8, Research Products International) for 2 min at 37°C, followed by incubation in two primary antibodies for 8 min at 37°C. After this 8 min incubation, a 4 min incubation in secondary antibodies Mouse Alexa Fluor Plus 555 (1:200, A32727, Thermo Fisher Scientific), Rabbit Alexa Fluor Plus 647 (1:400, A32733, Thermo Fisher Scientific), and DAPI was conducted at 37°C. After a wash in Imaging Buffer, sequential images were again captured. Following imaging, the slides were washed in Multistaining Buffer and the antibodies were eluted by a 2 min incubation in Elution Buffer (Lunaphore Technologies, BU07) at 37°C. Autofluorescence was then quenched with Quenching Buffer at 37°C for 30 s. These steps were repeated for the subsequent staining cycles.

All primary antibodies are listed in online supplemental table 3. The primary commercial platform antibodies were selected based on the quality of the vendor’s validation and literature, and on receipt, each antibody was validated by manual immunohistochemistry on normal tissue and/or cancer tissue. The accuracy of the staining was validated by a pathologist, based on prior descriptions of protein expression localization and intensity. Depending on antibody quality, one to three antibodies were tested per protein target. After manual validation, we validated the antibody for mIF at various dilutions, and its staining consistency across multiple rounds of staining and elution.

The PD-L2 antibody was evaluated using two clones, previously described in ACC studies: 176611 (MAB1224, Biotechne) and 366C.9E5 (MABC1120, MilliporeSigma). In addition, the RNAscope Probe Hs-PDCD1LG2 (551891, Advanced Cell Diagnostics) was used to confirm correct staining of antibodies. Both the antibodies and RNAscope probe stained similarly in tonsil (online supplemental figure 14). PD-L2 clone 176611 was used for further staining on the mIF COMET platform, due to the antibody staining quality being optimal at the pH9 epitope retrieval used for our mIF staining.

Multiplex immunofluorescence image analysis

Image preprocessing

Leveraging autofluorescence measurements created by the COMET device for the TRITC and Cy5 channels, pixel-based autofluorescence background subtraction was done using Horizon software (Lunaphore Technologies SA). Subtracting background intensities from foreground staining intensities permitted us to more accurately extract the staining intensity for each antibody stain during subsequent analysis steps. These background-subtracted whole-slide images (WSI) were exported as OME-TIFF files for downstream analysis. Additionally, these WSIs were also converted to the Digital Imaging and Communications in Medicine (DICOM) format as VL Whole Slide Microscopy Images. We followed the data model for a DICOM digital pathology workflow,⁵⁷ inserting anonymized clinical metadata as FHIR resources throughout the conversion process and encoding staining procedures, based on TIFF metadata. The DICOM instances were then uploaded to our DICOM store for retrieval according to open DICOMweb Application programming interface (API) specifications. We leveraged the Slim Viewer,⁵⁷ connected to our DICOM store to visually inspect our data sets and to annotate ROI for downstream cohort analysis, as well as highlight artifacts to exclude.

Cell segmentation

Cell segmentation was implemented iteratively, leveraging an open-source, pretrained generalist segmentation model, “cyto3”^{58 59} via the cellpose-qupath-extension V.0.9.0⁶⁰ in QuPath V.0.5.1.⁶¹ For each WSI data set, nuclei detection objects were identified via segmentation analysis that leveraged DAPI as the primary channel, with no secondary channel specified. Subsequently, we calculated cell detection objects for each of the cytoplasmic and membrane markers independently by specifying a single marker as the primary channel, in conjunction with DAPI as the secondary channel. This process resulted in independent segmentation masks for DAPI, as well as each of the cytoplasmic and membrane markers present in each of the WSI data sets. Next, by intersecting all the centroid coordinates for each independently segmented marker within 130 px (~30 µm) of each DAPI segmentation, these data were combined on a cell-by-cell basis to generate complete cell segmentations that included nuclei and cellular contours for all valid cell detections.

Cell fluorescence measurement extraction

Intensity summary statistics and cell-specific measurements for each valid cell detection were collected post-segmentation and cell object reconstruction. Using annotation objects in QuPath, we extracted nuclei and cytoplasmic measurements across optical channels for each antibody staining channel present in each of the WSI data sets. To assess the morphological features of each cell detection, we measured the area, compactness, convexity, and solidity for each nucleus and cytoplasm. To obtain optical measurements within the areas of identified cell detections for each nucleus and cytoplasm, we measured the pixel intensity as a mean, median, 90th percentile, and sum using QuPath and exported these data as a singular csv for each WSI data set for further analysis in R.

Cell classification

For each sample, a pathologist marked ROI, such as total tumor area, malignant epithelium, and normal ducts to be independently analyzed. For each antibody staining channel, the positive expression threshold was manually set based on visual inspection by at least two pathologists. During processing, cells were omitted from analysis according to the following criteria: DAPI below threshold, non-nuclear DAPI (ratio nucleus/cytoplasm <1), and intensity outliers caused by background subtraction (zeros or extreme values >10 SDs above the mean).

The retained cells were subsequently classified as either positive or negative for each marker by using the manually determined thresholds, and then classified according to the positive marker combination in online supplemental table 4. These cell classifications and the spatial orientations of their nuclei provided further insight into the imaging of each individual WSI.

PD-L1 and PD-L2 scoring

Cells were classified as PD-L1 or PD-L2 positive or negative based on a digital threshold of the immunofluorescence images, manually set for each sample. The TPS was generated by dividing PD-L1/PD-L2 positive tumor cells by the total number of tumor cells and multiplying by 100. The CPS was generated by dividing the number of PD-L1/PD-L2 positive cells of any cell type by the total number of tumor cells and multiplying by 100. For PD-L1, only membranous staining was considered positive, with a cut-off score of TPS>1% and CPS>1. For PD-L2, both membranous and cytoplasmic staining were considered as positive, with a cut-off score of TPS>10% and CPS>10, consistent with prior studies on PD-L2 in ACC.^{26 27}

Immunohistochemistry

FFPE slides were deparaffinized and rehydrated via serial incubations in xylene and ethanol. Heat-induced antigen retrieval was performed in a hot water bath in pH 9 Tris/EDTA buffer or pH 6 citrate buffer, and stained with primary antibodies anti-B2M antibody (1:1000, HPA006361 Sigma-Aldrich), anti-B2M antibody (1:100, MA5-36022, Thermo Fisher Scientific), anti-HLA-I-ABC antibody (1:1000, ab70328, Abcam), anti-NFIB (1:100, ab186738, Abcam), anti-PD-L1 antibody (1:100, IHC411-1, GenomeMe), anti-IRF1 (1:100, 8478S, Cell Signaling), anti-GBP1 antibody (1:150, ab119236, Abcam), or anti-TAP1 antibody (1:200, 11114–1-AP, Proteintech) for 1 hour at room temperature. The corresponding secondary antibody, RTU ImmPress-HRP Goat IgG polymer reagent mouse (MP-7452, Vector Laboratories) or RTU ImmPress-HRP Goat IgG polymer reagent rabbit (MP-7451, Vector Laboratories) was applied for 30 min at room temperature. The sections were then developed using a DAB Substrate Kit (ab64238, Abcam). The slides were subsequently stained with hematoxylin, dehydrated via serial incubations in ethanol and xylenes, mounted, and cover slipped.

Immunohistochemical stains were quantified by two methodologies. First, the whole slide was evaluated by a head and neck subspecialty pathologist (WCF) by using a sum of the combined stain intensity (0–3) and the percentage of positive cells (0=0–9%, 1=10–29%, 2=30–59%, 3=60–100%). PD-L1 staining was evaluated by the TPS score. Additionally, 20× images of the stained tissues were analyzed using FIJI. The color channels of the images were deconvoluted using the Color Deconvolution 1.7 Plugin. The H DAB vector was used for IHC images, and the FastRed FastBlue DAB was used for RNA-ISH images. ROI containing only cancer cells were marked for each image and analyzed using the reciprocal staining intensities (RSI=255 – mean gray value).⁶²

RNA in-situ hybridization

RNA-ISH was performed using the RNAscope 2.5 assay Reagent Kit-RED (322350, Advanced Cell Diagnostics) according to the manufacturer’s instructions. Sections were deparaffinized and dehydrated with xylene and 100% ethanol before being air dried at room temperature. RNAscope Hydrogen Peroxide was next applied to each section for 10 min and then subjected to manual target retrieval in RNAscope Target retrieval solution for 30 min. Afterwards, the sections were washed with distilled water, dipped in ethanol, and dried at room temperature. A hydrophobic barrier was drawn around the tissue sections using a barrier pen. Dried slides were covered in Protease Plus treatment and placed into the HybEZ oven (Advanced Cell Diagnostics), set at 40°C for 30 min. The slides were then washed in distilled water, incubated with an anti-B2M probe (442201, Advanced Cell Diagnostics) or at 40°C for 2 hours, washed again with the 1X Wash Buffer, and incubated in amplification solutions and FastRed detection reagents. Finally, the slides were washed in tap water, stained in hematoxylin and bluing solution, and fully dried at 60°C before mounting.

Drug treatment of tissue slice cultures

Surgically resected ACC tissue specimens were obtained in cold DMEM (11965, Thermo Fisher Scientific) containing 10% FBS (A52568, Thermo Fisher Scientific) and 1% Penicillin/Streptomycin (15140–122 Thermo Fisher Scientific), and immediately prepared for in vitro drug treatments. Based on the specimen size, they were manually cut into thin slices with a scalpel and plated into a 12-well cell culture dish. The tissue slices were then incubated in DMEM media, supplemented with 10% FBS and 1% Pen/Strep containing different therapeutic agents in a standard 37°C humidified incubator with 5% CO₂. Tissue slices were subjected either to DMSO (1:1000, D2650, Sigma-Aldrich) as a control or to one of four different therapeutic agents: 50 nM of bortezomib (S1013 Selleck Chemicals), 15 µM STING agonist ADU-S100 ammonium salt (HY-12885B MedChemExpress), 10 µg/mL of pembrolizumab (HY-P9902 MedChemExpress), or 5 ng/mL of IFN-γ (PHC4031, Fisher Scientific). After 48 hours, slices were fixed in 10% neutral buffered formalin for 1 hour/mm at room temperature. After fixation, they were embedded in paraffin and cut into 5 µm sections for further analysis.

Cell culture

All cell lines were obtained from the American Type Culture Collection (ATCC) (Manassas, Virginia, USA), except for S6, an MGH patient-derived glioblastoma multiforme cell line. HEK293T cells, U2OS cells, MCF-7 cells, and S6 cells were cultured in DMEM (11965, Thermo Fisher Scientific), supplemented with 10% FBS (A5256801, Thermo Fisher Scientific) and 1% Penicillin/Streptomycin (15140–122, Thermo Fisher Scientific). NCI-H446 cells were cultured in Roswell Park Memorial Institute (RPMI)-1640 medium (11675, Thermo Fisher Scientific) supplemented with 10% FBS and 1% Penicillin/Streptomycin. The cells were kept at standard conditions in 37 °C with 5% CO₂. The cell lines were tested for Mycoplasma infection every 2 months through PCR testing.

Total nucleic acid extraction from FFPE tissue

ROIs were circled by the pathologist on an H&E-stained slide from each tissue block, and a corresponding number of 5 µm slides per block needed for extraction were also marked by the pathologist. Total nucleic acid was extracted using the Zymo Quick-DNA/RNA in FFPE MiniPrep Kit (R1009, Zymo Research). FFPE slides were deparaffinized in xylene for 5 min, followed by a wash in 100% ethanol for 5 min. Using a scalpel blade, the ROI was then manually scraped directly into an Eppendorf tube containing 95 µL of DNase/Rnase free water, 95 µL 2X Digestion Buffer, and 10 µL Proteinase K. Each Eppendorf tube was incubated for 1 hour at 55°C, and then for an additional 20 min at 94°C. 600 µL of DNA/RNA Lysis Buffer was next added to the tissue and centrifuged at 10,000 rpm for 60 s. Following this, the lysate was washed with 100% ethanol and transferred into a Zymo-Spin IICR Column and collection tube. Nucleic acid was extracted from the lysate according to the manufacturer’s instructions, and DNA/RNA concentrations were quantified using the Qubit RNA HS Assay kit (Q32855, Thermo Scientific) or Qubit ssDNA HS Assay kit (Q32854, Thermo Fisher Scientific).

NFIB and MYB knockdown

NFIB shRNA lentiviral particles (sc-43565-V, Santa Cruz Technologies) or Control shRNA Lentiviral Particles (sc-108080, Santa Cruz Technologies) were used for lentiviral transduction in cell lines on 12-well plates at 60% confluency, according to the manufacturer’s protocol. After 24 hours of transduction, the cells were selected with puromycin (A1113803, Thermo Fisher) for 48 hours at the following concentrations: MCF7 2 µg/mL, S6 2 µg/mL, 293T 3 µg/mL, U2OS 1.5 µg/mL, and NCI-H446 0.75 µg/mL. For the c-MYB shRNA knockdown, the cells were seeded in 6-well plates and transfected either with the ON-TARGETplus human MYB siRNA SMARTPool (L-003910-00-0005, Horizon Discovery) or the Control ON-TARGET non-targeting negative Pool (D-001810-10-05, Horizon Discovery) at a concentration of 90 nM for 48 hours and 72 hours per manufacturers protocol using the DharmaFECT 1 Transfection Reagent (T-2001–01, Horizon Discovery).

Plasmid construction

Two gRNAs targeting the NFIB binding site on the B2M promoter region gRNAa: AAACGCGTGCCCAGCCAATC, gRNAb: CTGGCACTGCGTCGCTGGCT were cloned into the plasmid lentiCRISPR V.2 (plasmid #52961, Addgene). NFIB overexpression plasmid used the NFIB full-length sequence in the pLV plasmid backbone.

Lentivirus production

Lentivirus was produced in 293T Lenti-X cells (632180, Takara) with Roche Xtreme-Gene Transfection reagent (XTG9-RO, Millipore Sigma), the packaging plasmid psPAX2, and the VSV-G expressing plasmid pMD2G (#12260, #12259, Addgene). The viral supernatants were collected after 48 hours of transfection. The target cells were transduced in media containing 5 µg/mL of polybrene (SC-134220 Santa Cruz technologies) for 24 hours and afterwards selected with puromycin for 48 hours.

Single cell cloning

The cell lines MCF7, 293T, and NCI-H446 were transfected with lentivirus containing gRNA a, gRNA b, and the combination of gRNA a+b. After puromycin selection, the cells were flourescence-actived cell sorting (FACS)-sorted into 96-well plates to establish single cell clones and expanded. DNA was extracted from the single cell clones using the QIAamp DNA Mini Kit (56304, Qiagen), and the target site at the B2M promoter was amplified using PCR and subsequent PCR purification with the QIAquick PCR Purification Kit (28106, Qiagen). The purified PCR product underwent Sanger sequencing, and the successful modification of the targeted sequence was confirmed using Synthego ICE (Interference of CRISPR editing) web-based tool. Single cell clones with the successful modifications were chosen for further analysis.

Western blot analysis

Western blotting was performed by applying standard protocols. The primary antibodies were diluted in 5% milk at the following dilutions: anti-NFIB (1:1,000, ab186738, Abcam; the two bands observed for NFIB are most likely due to the multiple NFIB isoforms previously reported,⁶³ anti-c-MYB (1:500, 12319, Cell Signaling), anti-B2M (1:500, MA5-36022, Invitrogen), and anti-cyclophilin B (1:5,000, 43603S, Cell Signaling) and then incubated for 1 hour. The secondary antibodies—anti-rabbit IgG, horseradish-peroxidase-linked (1:2,000, 7074S, Cell Signaling), anti-mouse IgG, horseradish-peroxidase-linked (1:2,000, 7076S, Cell Signaling)—were applied for 1 hour. The membranes were then developed with SuperSignal West Pico PLUS Chemiluminescent Substrate (34580, Thermo Scientific) and imaged using autoradiographic film (XAR ALF 2025, LabScientific).

The relative protein expression of the western blots was quantified using FIJI image analysis software, using the Band/Peak Quantification Macro. The images were inverted before determining ROIs for each band. The Band/Peak Quantification macro estimated the background using the median intensity of the top and bottom of the ROI. The signal values were then normalized to the cyclophilin B housekeeping protein.

scRNA-seq gene expression analysis

The normal salivary gland scRNA-seq raw count table and cell type annotations were downloaded from HPA (data available from V.23.proteinatlas.org).^{38 39} Cells with fewer than 500 genes or 1,000 reads, greater than 40% ribosomal reads, or greater than 30% mitochondrial reads were removed. scRNA-seq read counts for a primary salivary gland ACC tumor were downloaded (data accessible at National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) database accession GSE217084)³⁷ and filtered in the same manner. The filtered counts were processed and analyzed independently for both data sets using Seurat V.5.0.2. Briefly, the counts were log-normalized and scaled. Using the first 20 principal components of the top 2,000 most variable genes, the cells were then clustered using Louvain modularity optimization of the shared nearest neighbors graph and visualized using UMAP dimensionality reduction.

The subset of cells labeled as salivary duct cells in the HPA normal salivary gland data set was also re-clustered and visualized separately by using the same parameters. For the ACC primary tumor data set, cell types were labeled using CHETAH V.1.18.0, with a custom reference generated from the HPA data processed as described above. All expression-valued UMAPs were generated using the “FeaturePlot” function from Seurat, with minimum and maximum cut-offs set to the 0.01 and 0.99 quantile, respectively, for each gene shown. Box plots with significance annotations were generated using ggplot2 V.3.5.0 and ggsignif V.0.6.4.

Fusion detection

To assess MYB-NFIB or MYBL1-NFIB fusion positivity in ACC cases, approximately 200 ng of total RNA were extracted from FFPE tissue identified as tumor by a molecular pathologist, and then processed according to vendor specifications with the ArcherDX FusionPlex Pan Solid Tumor V.2 kit (AB0137, Archer—Integrated DNA Technologies). Final sequencing libraries of samples were multiplexed and sequenced on the Illumina NextSeq 550 platform with the High Output V.2.5 (300 Cycle) kits (20024908, Illumina). Raw sequencing data were demultiplexed and converted to FASTQ files with BCL Convert V.4.2.1, which were subsequently uploaded to the Archer Analysis V.7.1.0–14 (Archer—Integrated DNA Technologies). Additionally, we leveraged a custom fusion analytics pipeline, implemented in Bash, to quality control raw FASTQ reads with TrimGalore,⁶⁴ align reads to the GRCh38 reference assembly^{65 66} with STAR⁶⁷ and identify fusions using Arriba. Fusion results for each sample were verified on a case-by-case basis by expert molecular pathologist review (AJI).

Visium HD

The single-cell resolution spatial transcriptomics data sets of ACCs were generated using the Visium HD, Human Transcriptome, 6.5 mm reagent kit (1000675, 10x Genomics) following the manufacturer’s protocol (CG000684 Rev A, CG000685 Rev B, 10x Genomics). In brief, total RNA from the FFPE slides was purified using the RNeasy FFPE Kit (73504, Qiagen), and RNA quality was measured using the High Sensitivity RNA ScreenTape (5067–5579, Agilent) and the TapeStation (Agilent). Tissue slides with DV200>50% were selected. The FFPE tissue slides were deparaffinized and H&E stained before coverslipping and imaging. Whole slide images were acquired at 20× magnification (0.52 µm/pixel resolution) using a MoticEasyScan Infinity digital pathology scanner. Subsequently, the slides were destained and decrosslinked, followed by probe hybridization, ligation, CytAssist-enabled probe release and capture onto the Visium HD slide (1000670, 10x Genomics), probe extension, and elution. Visium HD FFPE libraries were constructed with the Visium Human Transcriptome V.2 Probe kit (1000466, 10x Genomics) and Dual Index Kit TS Set A (1000251, 10x Genomics) and sequenced on the Illumina NextSeq 2000 platform with the P3 (300 cycles) Kit (20040561, Illumina).

Raw sequencing data were demultiplexed using BCL Convert V.4.1.7 and processed using the count command from Space Ranger V.3.1.1. The resulting gene-barcode matrices of 8 µm × 8 µm binned count data along with spatial coordinates aligning the barcodes to the microscope image were analyzed in R using Seurat V.5.1.0. For ACC47, bins containing <150 genes or >25% mitochondrial reads were excluded. As ACC23 contained tissue fold artifacts, boundaries were hand-drawn using the Loupe Browser V.8.0.0, and bins falling within these regions were excluded, as were bins containing <100 genes or >25% mitochondrial reads. For both samples, counts were log-normalized and scaled, then the first 20 PCs of the top 2000 most variable genes were used for Louvain clustering and UMAP visualization. One cluster was discarded from the analysis of ACC23 as it contained only three barcodes and was deemed a product of outliers. The remaining clusters were manually annotated by reviewing the top positive marker genes found using the “FindAllMarkers” function from Seurat, along with custom cluster statistics (eg, percent keratins, collagens, and known ACC-specific genes), and visual inspection of the clusters overlaid on the H&E image.

Differential gene expression analysis between the tumor clusters was performed using “FindMarkers” from Seurat. GSEA of GO terms was conducted on ranked lists of log fold change values using clusterProfiler V.4.12.6, with a minimum gene set size of 20 and a p value cut-off of 0.05. To create a non-redundant summary of the enrichment results, simplifyEnrichment V.1.14.1 was used to calculate and cluster a pairwise matrix of GO semantic similarities via the functions “GO_similarity” and “simplifyGO”. The lowest p value term in each cluster was selected as representative for visualization purposes.

Statistical analysis

Differences of means between continuous variables for ACC versus control tumors were compared using Student’s t-test (“t.test” in R). Fisher’s exact test (“fisher.test” in R) was used to assess differences in proportions between categorical metadata features. For scRNA-seq, Visium HD, and COMET data, Mann-Whitney U tests were performed to assess the locational shift of expression distributions between cell subsets of interest via the “wilcox.test” function in R. Prior to statistical analysis and box plot visualization, all COMET cell mean intensities were transformed using the following sigmoid function to obtain values ranging from 0 to 1, with the threshold mapped to 0.5:

f(x)=11+e2(t−x)

where x is the observed cell mean intensity and t is the intensity threshold for a given marker. This transformation allowed us to minimize the effects of outliers and to better visualize differences between values in proximity to the threshold. All statistical calculations were performed using base R V.4.3.1.

Statistical analysis of IHC, RNA-ISH, and western blot quantifications was performed using GraphPad Prism V.10 (San Diego, California, USA). The results are shown as mean±SD. One-way analysis of variance test with a Dunnett’s post hoc were performed.

We thank our colleagues in the clinical teams at Massachusetts Eye and Ear Hospital and Massachusetts General Hospital for the assistance provided with tumor collection. We thank the Mass General Krantz Family Cancer Center Translational Cartography Core, which provided support for whole slide imaging. Figure 1 (Created in BioRender. Li, A. (2025) https://BioRender.com/d48c325), 3 (Created in BioRender. Li, A. (2025); https://BioRender.com/s68e936), 4 (Created in BioRender. Li, A. (2025) https://BioRender.com/r08d378), 5 (Created in BioRender. Li, A. (2025) https://BioRender.com/n01b941), and online supplemental figures 1 (Created in BioRender. Li, A. (2025) https://BioRender.com/a00g248), 10 (Created in BioRender. Li, A. (2025) https://BioRender.com/z87z482) were partially created with BioRender.com.

Data availability statement

Data are available in a public, open access repository. Data are available upon reasonable request. All code used to perform statistical analysis, as well as relevant figures, is available at: https://github.com/IafrateLab/li_et_al-2024. The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Ethics statements

Patient consent for publication

Not applicable.

Ethics approval

This study involves human participants and was approved by Mass General Brigham Institutional Review Board: 2023P001964 and 2022P003132 and Dana-Farber Cancer Institute Internal Review Board: Protocol 13-416.

References

¹ Spiro RH, Huvos AG, Strong EW. Adenoid cystic carcinoma of salivary origin. A clinicopathologic study of 242 cases. Am J Surg 1974; 128: 512–20. doi:10.1016/0002-9610(74)90265-7

² Lawrence JB, Mazur MT. Adenoid cystic carcinoma: a comparative pathologic study of tumors in salivary gland, breast, lung, and cervix. Hum Pathol 1982; 13: 916–24. doi:10.1016/s0046-8177(82)80052-x

³ Ghabach B, Anderson WF, Curtis RE, et al. Adenoid cystic carcinoma of the breast in the United States (1977 to 2006): a population-based cohort study. Breast Cancer Res 2010; 12: R54. doi:10.1186/bcr2613

⁴ Benali K, Benmessaoud H, Aarab J, et al. Lacrimal gland adenoid cystic carcinoma: report of an unusual case with literature review. Radiat Oncol J 2021; 39: 152–8. doi:10.3857/roj.2021.00122

⁵ Wang J, Yang Z, Liu Y, et al. The GAL/GALR2 axis promotes the perineural invasion of salivary adenoid cystic carcinoma via epithelial-to-mesenchymal transition. Cancer Med 2023; 12: 4496–509. doi:10.1002/cam4.5181

⁶ Huang Z, Pan J, Chen J, et al. Multicentre clinicopathological study of adenoid cystic carcinoma: A report of 296 cases. Cancer Med 2021; 10: 1120–7. doi:10.1002/cam4.3707

⁷ Spiro RH. Distant metastasis in adenoid cystic carcinoma of salivary origin. Am J Surg 1997; 174: 495–8. doi:10.1016/s0002-9610(97)00153-0

⁸ Gao M, Hao Y, Huang MX, et al. Clinicopathological study of distant metastases of salivary adenoid cystic carcinoma. Int J Oral Maxillofac Surg 2013; 42: 923–8. doi:10.1016/j.ijom.2013.04.006

⁹ Fordice J, Kershaw C, El-Naggar A, et al. Adenoid cystic carcinoma of the head and neck: predictors of morbidity and mortality. Arch Otolaryngol Head Neck Surg 1999; 125: 149–52. doi:10.1001/archotol.125.2.149

¹⁰ Spiro RH, Huvos AG. Stage means more than grade in adenoid cystic carcinoma. Am J Surg 1992; 164: 623–8. doi:10.1016/s0002-9610(05)80721-4

¹¹ He S, Li P, Zhong Q, et al. Clinicopathologic and prognostic factors in adenoid cystic carcinoma of head and neck minor salivary glands: A clinical analysis of 130 cases. Am J Otolaryngol 2017; 38: 157–62. doi:10.1016/j.amjoto.2016.11.014

¹² Jang S, Patel PN, Kimple RJ, et al. Clinical Outcomes and Prognostic Factors of Adenoid Cystic Carcinoma of the Head and Neck. Anticancer Res 2017; 37: 3045–52. doi:10.21873/anticanres.11659

¹³ Geiger JL, Ismaila N, Beadle B, et al. Management of Salivary Gland Malignancy: ASCO Guideline. J Clin Oncol 2021; 39: 1909–41. doi:10.1200/JCO.21.00449

¹⁴ NCCN. National comprehensive cancer network clinical practice guidelines in oncology: head and neck tumors version 3.2024. 2024. Available: https://www.nccn.org/professionals/physician_gls/pdf/head-and-neck.pdf [Accessed 5 Apr 2024 ].

¹⁵ Mitani Y, Liu B, Rao PH, et al. Novel MYBL1 Gene Rearrangements with Recurrent MYBL1-NFIB Fusions in Salivary Adenoid Cystic Carcinomas Lacking t(6;9) Translocations. Clin Cancer Res 2016; 22: 725–33. doi:10.1158/1078-0432.CCR-15-2867-T

¹⁶ Persson M, Andersson MK, Mitani Y, et al. Rearrangements, Expression, and Clinical Significance of MYB and MYBL1 in Adenoid Cystic Carcinoma: A Multi-Institutional Study. Cancers (Basel) 2022; 14: 3691. doi:10.3390/cancers14153691

¹⁷ Persson M, Andrén Y, Mark J, et al. Recurrent fusion of MYB and NFIB transcription factor genes in carcinomas of the breast and head and neck. Proc Natl Acad Sci U S A 2009; 106: 18740–4. doi:10.1073/pnas.0909114106

¹⁸ Thierauf J, Ramamurthy N, Jo VY, et al. Clinically Integrated Molecular Diagnostics in Adenoid Cystic Carcinoma. Oncologist 2019; 24: 1356–67. doi:10.1634/theoncologist.2018-0515

¹⁹ Ferrarotto R, Mitani Y, McGrail DJ, et al. Proteogenomic Analysis of Salivary Adenoid Cystic Carcinomas Defines Molecular Subtypes and Identifies Therapeutic Targets. Clin Cancer Res 2021; 27: 852–64. doi:10.1158/1078-0432.CCR-20-1192

²⁰ Ho AL, Dunn L, Sherman EJ, et al. A phase II study of axitinib (AG-013736) in patients with incurable adenoid cystic carcinoma. Ann Oncol 2016; 27: 1902–8. doi:10.1093/annonc/mdw287

²¹ Tchekmedyian V, Sherman EJ, Dunn L, et al. Phase II Study of Lenvatinib in Patients With Progressive, Recurrent or Metastatic Adenoid Cystic Carcinoma. J Clin Oncol 2019; 37: 1529–37. doi:10.1200/JCO.18.01859

²² Mahmood U, Bang A, Chen Y-H, et al. A Randomized Phase 2 Study of Pembrolizumab With or Without Radiation in Patients With Recurrent or Metastatic Adenoid Cystic Carcinoma. Int J Radiat Oncol Biol Phys 2021; 109: 134–44. doi:10.1016/j.ijrobp.2020.08.018

²³ Miller LE, Au V, Mokhtari TE, et al. A Contemporary Review of Molecular Therapeutic Targets for Adenoid Cystic Carcinoma. Cancers (Basel) 2022; 14: 992. doi:10.3390/cancers14040992

²⁴ Tang Y-F, An P-G, Gu B-X, et al. Transcriptomic insights into adenoid cystic carcinoma via RNA sequencing. Front Genet 2023; 14: 1144945. doi:10.3389/fgene.2023.1144945

²⁵ Seifi S, Azizi Z, Ghanghermeh A, et al. Eosinophil density in common benign and malignant salivary gland tumors with Congo red staining. Indian J Med Paediatr Oncol 2019; 40: 212–6. doi:10.4103/ijmpo.ijmpo_217_17

²⁶ Mosconi C, de Arruda JAA, de Farias ACR, et al. Immune microenvironment and evasion mechanisms in adenoid cystic carcinomas of salivary glands. Oral Oncol 2019; 88: 95–101. doi:10.1016/j.oraloncology.2018.11.028

²⁷ Sridharan V, Gjini E, Liao X, et al. Immune Profiling of Adenoid Cystic Carcinoma: PD-L2 Expression and Associations with Tumor-Infiltrating Lymphocytes. Cancer Immunol Res 2016; 4: 679–87. doi:10.1158/2326-6066.CIR-16-0031

²⁸ Bonaventura P, Shekarian T, Alcazer V, et al. Cold Tumors: A Therapeutic Challenge for Immunotherapy. Front Immunol 2019; 10: 168. doi:10.3389/fimmu.2019.00168

²⁹ Wolkow N, Jakobiec FA, Afrogheh AH, et al. PD-L1 and PD-L2 Expression Levels Are Low in Primary and Secondary Adenoid Cystic Carcinomas of the Orbit: Therapeutic Implications. Ophthalmic Plast Reconstr Surg 2020; 36: 444–50. doi:10.1097/IOP.0000000000001585

³⁰ Michaelides I, Künzel J, Ettl T, et al. Adenoid cystic carcinoma of the salivary glands: a pilot study of potential therapeutic targets and characterization of the immunological tumor environment and angiogenesis. Eur Arch Otorhinolaryngol 2023; 280: 2937–44. doi:10.1007/s00405-023-07884-3

³¹ Vital D, Ikenberg K, Moch H, et al. The expression of PD-L1 in salivary gland carcinomas. Sci Rep 2019; 9: 12724. doi:10.1038/s41598-019-49215-9

³² Chang H, Kim JS, Choi YJ, et al. Overexpression of PD-L2 is associated with shorter relapse-free survival in patients with malignant salivary gland tumors. Onco Targets Ther 2017; 10: 2983–92. doi:10.2147/OTT.S134589

³³ Wang H, Liu B, Wei J. Beta2-microglobulin(B2M) in cancer immunotherapies: Biological function, resistance and remedy. Cancer Lett 2021; 517: 96–104. doi:10.1016/j.canlet.2021.06.008

³⁴ Bernal M, Ruiz-Cabello F, Concha A, et al. Implication of the β2-microglobulin gene in the generation of tumor escape phenotypes. Cancer Immunol Immunother 2012; 61: 1359–71. doi:10.1007/s00262-012-1321-6

³⁵ Sondka Z, Dhir NB, Carvalho-Silva D, et al. Data from: cosmic: a curated database of somatic variants and clinical data for cancer cosmic. 2024. Available: https://cancer.sanger.ac.uk/census

³⁶ Zhou M-J, Yang J-J, Ma T-Y, et al. Increased retinoic acid signaling decreases lung metastasis in salivary adenoid cystic carcinoma by inhibiting the noncanonical Notch1 pathway. Exp Mol Med 2023; 55: 597–611. doi:10.1038/s12276-023-00957-7

³⁷ Zhou MJ, Yang JJ, Ma TY, et al. Data from: Increased retinoic acid signaling decreases lung metastasis in salivary adenoid cystic carcinoma by inhibiting the noncanonical Notch1 pathway NCBI GEO Gene Expression Omnibus. 2023. Available: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE217084

³⁸ Karlsson M, Zhang C, Méar L, et al. A single-cell type transcriptomics map of human tissues. Sci Adv 2021; 7: eabh2169. doi:10.1126/sciadv.abh2169

³⁹ Tabula Sapiens C, Jones RC, Karkanias J, et al. Data from: The Tabula Sapiens: a multiple-organ, single-cell transcriptomic atlas of humans Human Protein Atlas - Single cell type data. 2022 Available: https://www.proteinatlas.org/humanproteome/single+cell/single+cell+type/data#datasets

⁴⁰ Batsakis JG, Regezi JA, Luna MA, et al. Histogenesis of salivary gland neoplasms: a postulate with prognostic implications. J Laryngol Otol 1989; 103: 939–44. doi:10.1017/s0022215100110552

⁴¹ Akrish S, Peled M, Ben-Izhak O, et al. Malignant salivary gland tumors and cyclo-oxygenase-2: a histopathological and immunohistochemical analysis with implications on histogenesis. Oral Oncol 2009; 45: 1044–50. doi:10.1016/j.oraloncology.2009.07.016

⁴² Matic M. A subpopulation of human basal keratinocytes has a low/negative MHC class I expression. Hum Immunol 2005; 66: 962–8. doi:10.1016/j.humimm.2005.07.007

⁴³ Luke JJ, Gao X, Olszanski AJ, et al. 996MO Dazostinag (TAK-676) alone and in combination with pembrolizumab (pembro) in patients (pts) with advanced/metastatic solid tumors: data from phase I dose escalation. Ann Oncol 2024; 35: S678–9. doi:10.1016/j.annonc.2024.08.1055

⁴⁴ Ferrarotto R, Sousa LG, Feng L, et al. Phase II Clinical Trial of Axitinib and Avelumab in Patients With Recurrent/Metastatic Adenoid Cystic Carcinoma. J Clin Oncol 2023; 41: 2843–51. doi:10.1200/JCO.22.02221

⁴⁵ Linxweiler M, Kuo F, Katabi N, et al. The Immune Microenvironment and Neoantigen Landscape of Aggressive Salivary Gland Carcinomas Differ by Subtype. Clin Cancer Res 2020; 26: 2859–70. doi:10.1158/1078-0432.CCR-19-3758

⁴⁶ Sousa LG, McGrail DJ, Lazar Neto F, et al. Spatial Immunoprofiling of Adenoid Cystic Carcinoma Reveals B7-H4 Is a Therapeutic Target for Aggressive Tumors. Clin Cancer Res 2023; 29: 3162–71. doi:10.1158/1078-0432.CCR-23-0514

⁴⁷ Greenberg ME, Iafrate AJ, Skowronski J. The SH3 domain-binding surface and an acidic motif in HIV-1 Nef regulate trafficking of class I MHC complexes. EMBO J 1998; 17: 2777–89. doi:10.1093/emboj/17.10.2777

⁴⁸ Montesion M, Murugesan K, Jin DX, et al. Somatic HLA Class I Loss Is a Widespread Mechanism of Immune Evasion Which Refines the Use of Tumor Mutational Burden as a Biomarker of Checkpoint Inhibitor Response. Cancer Discov 2021; 11: 282–92. doi:10.1158/2159-8290.CD-20-0672

⁴⁹ Røder G, Geironson L, Bressendorff I, et al. Viral proteins interfering with antigen presentation target the major histocompatibility complex class I peptide-loading complex. J Virol 2008; 82: 8246–52. doi:10.1128/JVI.00207-08

⁵⁰ Sade-Feldman M, Jiao YJ, Chen JH, et al. Resistance to checkpoint blockade therapy through inactivation of antigen presentation. Nat Commun 2017; 8: 1136. doi:10.1038/s41467-017-01062-w

⁵¹ Stephens PJ, Davies HR, Mitani Y, et al. Whole exome sequencing of adenoid cystic carcinoma. J Clin Invest 2013; 123: 2965–8: 67201. doi:10.1172/JCI67201

⁵² Huang Y, Huang Z, Tang Z, et al. Research Progress, Challenges, and Breakthroughs of Organoids as Disease Models. Front Cell Dev Biol 2021; 9: 740574. doi:10.3389/fcell.2021.740574

⁵³ Rocchi C, Barazzuol L, Coppes RP. The evolving definition of salivary gland stem cells. NPJ Regen Med 2021; 6: 4. doi:10.1038/s41536-020-00115-x

⁵⁴ Man YG, Ball WD, Marchetti L, et al. Contributions of intercalated duct cells to the normal parenchyma of submandibular glands of adult rats. Anat Rec 2001; 263: 202–14. doi:10.1002/ar.1098

⁵⁵ Vos JL, Burman B, Jain S, et al. Nivolumab plus ipilimumab in advanced salivary gland cancer: a phase 2 trial. Nat Med 2023; 29: 3077–89. doi:10.1038/s41591-023-02518-x

⁵⁶ Mahadevan NR, Knelson EH, Wolff JO, et al. Intrinsic Immunogenicity of Small Cell Lung Carcinoma Revealed by Its Cellular Plasticity. Cancer Discov 2021; 11: 1952–69. doi:10.1158/2159-8290.CD-20-0913

⁵⁷ Gorman C, Punzo D, Octaviano I, et al. Interoperable slide microscopy viewer and annotation tool for imaging data science and computational pathology. Nat Commun 2023; 14: 1572. doi:10.1038/s41467-023-37224-2

⁵⁸ Stringer C, Wang T, Michaelos M, et al. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods 2021; 18: 100–6. doi:10.1038/s41592-020-01018-x

⁵⁹ Stringer C, Pachitariu M. Cellpose3: one-click image restoration for improved cellular segmentation. Bioinformatics [ Preprint ]. doi:10.1101/2024.02.10.579780

⁶⁰ Burri OS, Fehlmann T. BIOP/qupath-extension-cellpose: parallel object reader, normalization bugfix gpu disable option and more (v0.9.3). 2024.

⁶¹ Bankhead P, Loughrey MB, Fernández JA, et al. QuPath: Open source software for digital pathology image analysis. Sci Rep 2017; 7: 16878. doi:10.1038/s41598-017-17204-5

⁶² Nguyen D. Quantifying chromogen intensity in immunohistochemistry via reciprocal intensity. 2013. Available: https://assets-eu.researchsquare.com/files/nprot-2931/v1/ff8f61e8-c9d9-44ea-94b2-c04c8f25b243.pdf?c=1631825057

⁶³ Nanda JS, Awadallah WN, Kohrt SE, et al. Increased nuclear factor I/B expression in prostate cancer correlates with AR expression. Prostate 2020; 80: 1058–70. doi:10.1002/pros.24019

⁶⁴ Krueger F. TrimGalore: a wrapper tool around cutadapt and fastQC to consistently apply quality and adapter trimming to FastQ files. Babraham Bioinformatics 2015.

⁶⁵ Navarro Gonzalez J, Zweig AS, Speir ML, et al. The UCSC Genome Browser database: 2021 update. Nucleic Acids Res 2021; 49: D1046–57. doi:10.1093/nar/gkaa1070

⁶⁶ Lander ES, Linton LM, Birren B, et al. Initial sequencing and analysis of the human genome. Nature New Biol 2001; 409: 860–921. doi:10.1038/35057062

⁶⁷ Dobin A, Davis CA, Schlesinger F, et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 2013; 29: 15–21. doi:10.1093/bioinformatics/bts635