The Immune Microenvironment and Neoantigen Landscape of Aggressive Salivary Gland Carcinomas Differ by Subtype

Salivary gland carcinomas (SGC) arise from the major salivary glands (parotid, submandibular, sublingual) or the minor salivary glands of the upper aerodigestive tract. They are a heterogeneous group of malignancies, with 20 subtypes defined by the World Health Organization (1). The standard therapy for salivary gland cancers is surgical resection, with adjuvant radiotherapy or chemoradiotherapy in some cases–an approach that successfully treats a subset of patients. However, disease recurrence or metastasis is common, with distant metastases occurring in 25%–50% of patients (2). Because there are no active systemic therapies for SGCs, recurrent or metastatic disease is usually incurable. Ultimately the 5-year survival rate of recurrent/metastatic SGC is only approximately 20%.

Because of their relative rarity (incidence approximately 1 per 100,000 persons), the molecular features of these tumors are only beginning to be characterized, and features of their immune microenvironment have not been well defined. To improve the effectiveness of treatment, including immunotherapy, a better understanding of how SGCs evade immune surveillance is needed.

In terms of risk of relapse, the three most aggressive histologic subtypes of salivary gland tumors are adenoid cystic carcinoma (ACC), myoepithelial carcinoma (MECA), and salivary duct carcinoma (SDC). Each of these cancers exhibits a high rate of recurrence, distant metastasis, and resistance to conventional therapy.

ACCs account for approximately 10% of salivary gland malignancies (3), have low numbers of mutations, but have oncogenic fusions of MYB-NFIB or MYBL1-NFIB in approximately 60% of cases (4–6). There are no highly active systemic therapies for ACC.

MECAs are rare (∼2%) salivary gland tumors with a highly heterogeneous composition of cells (7). Like ACCs, MECAs are also characterized by low mutational burden and a high prevalence of oncogenic gene fusions, many of which include PLAG1 (8). There are no effective systemic therapies for MECA.

SDCs account for 1%–3% of all salivary gland tumors (9). Morphologically, these tumors resemble invasive ductal carcinoma of the breast, with a molecular profile similar to apocrine breast cancer (10). They often overexpress androgen receptor (AR) or ERBB2 (HER2; ref. 11), both of which appear in early data to be targetable (12), although clinical trials are ongoing. Targetable mutations in the PIK3CA and MAPK pathways have also been identified (10, 13).

In recent years, drugs that inhibit T-cell checkpoints such as CTLA-4 and the PD-1/PD-L1 axis have shown activity and achieved clinically meaningful responses in a subset of patients with different tumor types including melanoma, non–small cell lung, kidney, bladder, and urothelial carcinomas, as well as microsatellite instability-high carcinomas of various sites (14–21). In head and neck squamous cell carcinoma, anti-PD-1 therapies have achieved clinical responses in 15%–20% of patients (22–24).

To date, there is only sparse data regarding the activity of immune checkpoint blockade in SGCs. The KEYNOTE-028 basket trial of PD-L1–positive tumors included a small salivary gland cancer cohort of 28 patients treated with pembrolizumab. Despite enrichment for PD-L1–positive tumors, there was a modest overall response rate of 12% (all in non-ACC adenocarcinoma histologies), with a 6-month progression-free survival of 21% (25).

Recently, molecular features such as tumor mutational load (14, 15, 17, 26), clonal neoantigen load (27), HLA genotype (28, 29), immune infiltration into the tumor microenvironment (30), CD8⁺ T-cell activation (15, 31), and expression of immune checkpoint molecules (31, 32), have been correlated with response to immune checkpoint blockade.

These features have not been well described in SGCs. As clinical research expands beyond checkpoint inhibitor monotherapy, there is a clinical need to better define patients likely to benefit, and to characterize the mechanisms used by different cancer types to evade immune surveillance. These insights can help guide the next steps of clinical investigation. Here, we characterize the immune microenvironment and neoantigen landscape of ACC, MECA, and SDC tumors.

After written informed patient consent under an Institutional Review Board (IRB)-approved salivary tumor biospecimen protocol, tumor samples were prospectively obtained from patients undergoing treatment at Memorial Sloan Kettering Cancer Center (New York, NY): 20 ACC, 16 SDC, and 12 MECA samples immediately frozen in liquid nitrogen at the time of surgery and stored at −80°C. An additional 28 MECA samples were obtained from existing formalin-fixed and embedded in paraffin (FFPE) blocks. Of the ACC cases, 9 were primary tumors collected at the time of surgery, and 11 were biopsies of metastatic sites. Two pairs of ACCs were matched primary and metastatic tumors from two individual patients (ACC_P3 and ACC_M3; ACC_P12 and ACC_M12). All MECA and SDC cases were primary tumors.

Publicly available RNA-sequencing (RNA-seq) data from 10 different cohorts of The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/) dataset were used for a comparison of tumor immune microenvironment with the SGC samples: bladder urothelial carcinoma (BLCA, n = 95), breast cancer (BRCA, n = 1205), colorectal adenocarcinoma (COADREAD, n = 336), glioblastoma multiforme (n = 123), head and neck squamous cell carcinoma (HNSC, n = 293), kidney renal clear cell carcinoma (KIRC, n = 658), lung adenocarcinoma (n = 228), lung squamous cell carcinoma (LUSC, n = 258), ovarian serous cystadenocarcinoma (OV, n = 248), and uterine corpus endometrial carcinoma (UCEC, n = 534).

All studies were conducted in accordance with the Declaration of Helsinki and the International Ethical Guidelines for Biomedical Research Involving Human Subjects (CIOMS) and were approved by an Institutional Review Board (IRB). Written informed consent was obtained from all study participants.

Formalin-fixed paraffin-embedded (FFPE) tissue sections of ACC, SDC, and MECA tumors were sectioned onto glass slides at 4-μm thickness and consecutive tissue sections were stained with antibodies using Discovery XT processor (Ventana Medical Systems). Hematoxylin and eosin (H&E) stains were performed under standard procedures and reviewed by a salivary and head and neck pathologist (N. Katabi) to confirm the histologic diagnosis and evaluate additional immunostaining.

For IHC, tissue sections were deparaffinized with EZPrep buffer (Ventana Medical Systems), antigen retrieval was performed with CC1 buffer (Ventana Medical Systems), and sections were blocked for 30 minutes with Background Buster solution (Innovex), followed by avidin-biotin blocking (Ventana Medical Systems) for 8 minutes. Anti-CD45 (DAKO, catalog no. M0701, 0.5 μg/mL), anti-PD-1 (Ventana, catalog no. 760-4895, 3.1 μg/mL), anti-CD3 (DAKO, catalog no. A0452, 1.2 μg/mL), or anti-PD-L1 (Cell Signaling Technology, catalog no. 13684, 5 μg/mL) antibodies were applied and sections were incubated for 5 hours, followed by a 60-minute incubation with biotinylated horse anti-mouse IgG (Vector Laboratories, catalog no. MKB-2225B) for CD45 and PD-1 antibodies, or goat anti-rabbit IgG (Vector Laboratories, catalog no. PK6101) for CD3 and PD-L1 antibodies at 1:200 dilution in 3% (w/v) BSA-PBS. The detection was performed with a DAB Detection Kit (Ventana Medical Systems) according to the manufacturer's instructions. Slides were counterstained with hematoxylin and enclosed in a coverslip with Permount cover medium (Thermo Fisher Scientific).

Quantification of stained slides was performed manually by a head and neck/salivary pathologist blinded to the immune infiltration score (IIS) and T-cell infiltration score (TIS) values per sample by counting the number of positive leukocytes in the central region of the tumor under 400× magnification. The fields were chosen randomly in the central region of the tumor presupposing a tumor portion of at least 80%. In total, positively stained leukocytes were counted in 10 fields per section and the arithmetic mean was used for statistics.

RNA was extracted with either the RNeasy Plus Mini Kit (Qiagen) or RNeasy FFPE Kit (Qiagen), and RNA-seq libraries were prepared by using KAPA Stranded RNAseq with RiboErase sample preparation kit (KAPA Biosystems) using protocols for salivary cancers (8, 10). Sequencing was performed on HiSeq 2500 (Illumina) with paired-end reads of 50 (PE50) and 137M–234M reads per sample (SDCs), 104M–194M reads per sample (MECA frozen), 92M–400M reads per sample (MECA FFPE), and 170M–321M reads per sample (ACC).

DNA was extracted using the Blood and Tissue Kit (Qiagen). DNA quality and integrity was then analyzed using BioAnalyzer (Agilent Technologies). Whole exome sequencing (WES) libraries were prepared with the SureSelectXT Library Preparation Kit (Agilent) according to the manufacturer's instructions. Extracted DNA was then sheared using an LE220 Focus-ultrasonicator (Covaris), and the fragments were end-repaired, adenylated, ligated to Illumina sequencing adapters, and amplified by PCR. Exome capture was performed using the SureSelectXT v4 51 Mb capture probe set (Agilent), and captured exome libraries were then enriched by PCR. Final libraries were quantified with the KAPA Library Quantification Kit (KAPA Biosystems), Qubit Fluorometer (Life Technologies), and 2100 BioAnalyzer (Agilent) and were sequenced on a HiSeq2500 sequencer (Illumina) using 2 × 125 bp cycles as described previously (8). Tumor mutational burden was calculated as the number of coding mutations per megabase, also incorporating additional exome data from previously published WES samples from our laboratory analyzed using the same mutation calling algorithm (5, 8, 10).

RNA-seq raw read sequences were aligned against human genome assembly hg19 by STAR 2-pass alignment. Gene-level count values were computed by the summarizeOverlaps function from the R package "GenomicAlignments" with hg19 KnownGene as the base gene model. The Union counting mode was used and only mapped paired reads were considered. FPKM (Fragments Per Kilobase Million) values were then computed from gene level counts by using FPKM function from the R package "DESeq2." After principal components analysis to visualize possible batch effects, the ComBat batch correction method was used (33) through implementation in R sva package (34).

A set of previously published, orthogonal tools for assessing immune infiltration and activity in tumors using bulk RNA-seq data were applied. Cell-type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT) uses a reference gene expression signature and performs a linear support vector regression to adaptively select genes from the reference (35). Single-Sample Gene Set Enrichment Analysis (ssGSEA; ref. 36) calculates enrichment scores for a sample and gene set pair, allowing clustering by pathways rather than individual genes, and generates metrics for individual cell populations as well as aggregate measures of IIS and TIS (37). IIS is an aggregate score for innate and adaptive immune scores, while TIS is an aggregate score of nine T-cell subtypes. Estimation of Stromal and Immune Cells in Malignant Tumor Tissues using Expression Data (ESTIMATE) is an ssGSEA-based technique, in which differential gene expression from high and low immune cell–infiltrating tumor samples is used to derive a 141-gene signature estimating the degree of stromal and immune infiltration in a tumor (38). Immune cytolytic activity (“CYT” score) is calculated from geometric means of transcript levels of the two effector genes granzyme A (GZMA) and perforin (PRF1) (39). Batch-corrected normalized data were used for input into TIDE (Tumor Immune Dysfunction and Exclusion; ref. 40).

Five tumor samples with the highest and the lowest IIS per SGC histology were used for differentially expressed gene (DEG) analysis with the DESeq2 package in R. Given the raw count data and gene model used, DESeq2 normalized the expression raw count data by sample specific size factor while testing for genes found with significantly different expression between the high IIS group and the low IIS group samples. Qiagen Ingenuity Pathway Analysis (IPA) was used for canonical pathway analysis of differentially expressed genes, with Benjamini–Hochberg false discovery rate correction.

Mutation-associated neoepitopes were identified and binding to HLA was predicted by NetMHCpan 4.0 (41), with patient-specific HLA type determined using PolySolver (42). Fusion-derived neoantigens and binding to HLA were predicted using INTEGRATE-Neo, which identifies two gene spanning (i.e., gene fusion) transcripts and HLA type based on RNA-seq data and predicts binding using NetMHC (43). HLA zygosity and allele-specific copy number analyses were performed using FACETS (44). Mutational data from exome sequencing was available for all 16 SDC samples, 9 ACCs, and 12 MECAs (total n = 37).

D'Agostino and Pearson omnibus normality test was performed to determine if datasets follow a Gaussian distribution in each comparison. If the data was Gaussian, parametric tests were performed (two-tailed unpaired t tests, one-way ANOVA with Tukey correction for multiple comparisons, or Pearson correlation). If the data was non-Gaussian, nonparametric tests were applied (Mann–Whitney U test, one-way ANOVA using Kruskal–Wallis with Dunn correction for multiple comparisons, or Spearman correlation). The a priori definition of statistical significance for all hypothesis testing was two-tailed α < 0.05. The test used is specified in each figure or table.

RNA-seq and WES data for the included tumors have been uploaded to the NCBI Sequence Read Archive (SRA) under accession numbers SRP109264, SRP096726, and PRJNA601423.

A cohort of 76 salivary gland tumors from patients undergoing treatment at Memorial Sloan Kettering Cancer Center (MSKCC, New York, NY) was analyzed for this study: 20 ACCs, 40 MECAs, and 16 SDCs. Median follow-up time was 41.5 months. The clinical and pathologic features including treatment modalities and past medical history are shown in Table 1 and Supplementary Table S1. Tumors underwent RNA-seq and normalized data were analyzed to quantify and deconvolve immune infiltration using several methods, including CIBERSORT (35), ESTIMATE (38), ssGSEA (37), and cytolytic (CYT) score (ref. 39; Supplementary Table S1).

To place salivary cancer immunity in context, we compared the immune infiltration observed in SGCs to 10 other cancer types. ACCs and MECAs consistently exhibited comparatively immune-depleted microenvironments. On the other hand, SDCs were generally more highly immune infiltrated, comparable with levels observed in clear cell kidney cancer and non–small cell lung carcinoma (Fig. 1A).

Immune cell infiltration and somatic tumor mutational burden (TMB) are complementary measures that together provide information predictive of response to immune checkpoint inhibitor therapy (45), by reflecting aspects of both immunity and neoantigen landscape. Compared with the other cancer types, ACC and MECA tumors generally had low levels of both immune infiltration and TMB. In contrast, SDCs had much higher TMB and immune infiltration levels, although levels were moderate in comparison with other cancer types (Fig. 1B and C). We found no significant correlations between immune infiltration and patient age, sex, or autoimmune disease history.

The 76 SGC samples were then individually compared by performing unsupervised hierarchical clustering of the tumors based on immune deconvolution derived from these 4 methods, revealing an immune-poor cluster and an immune-rich cluster (Fig. 2A). Nearly all SDC tumors (15 of 16, 94%) were in the immune-rich cluster. In contrast, most of the ACCs (16 of 20, 80%) were in the immune-poor cluster, while MECAs were more broadly distributed across both clusters.

Among the three SGC histologies, SDCs generally had a significantly higher level of TIS and T-cell CYT activity (CYT score), compared with ACCs and MECAs (P ≤ 0.001 for both metrics, SDC vs. ACC or MECA, Fig. 2B and C). Identical results were observed for other immunity metrics (IIS, ESTIMATE Immune Score, and CIBERSORT; P ≤ 0.01 for all comparisons, Supplementary Fig. S1A–S1C).

To determine whether these RNA-seq–derived metrics of tumor immunity were consistent with one another, and with histologic assessment of immune cell infiltration in salivary tumors, we systematically examined the intermethod agreement of each of these techniques.

In the analyzed salivary tumors, the RNA-seq–derived immune metrics (CIBERSORT, ESTIMATE, CYT, and ssGSEA) were strongly correlated with one another (Spearman r = 0.73–0.95, P < 0.0001, Supplementary Fig. S2). In 16 representative cases with sufficient tissue available for IHC analysis, IHC staining targeting CD45 and CD3 including respective negative controls (mouse IgG and rabbit IgG) was performed. We found that IHC-derived and RNA-seq–derived measures of leukocyte infiltration (CD45⁺ cells, IIS) were strongly correlated (Spearman r = 0.67, P = 0.005), as were measures of T-cell infiltration (CD3⁺ cells, TIS; Spearman r = 0.63, P = 0.008; Supplementary Fig. S3; Supplementary Table S2) demonstrating a high validity of RNA-seq. These correlations indicate that, in SGC tumors, RNA-seq–derived deconvolution techniques measure aspects of the tumor immune microenvironment in a manner that is consistent and reproducible, and representative of leukocyte infiltration observable with IHC.

RNA-seq data were analyzed to deconvolve the levels of individual immune cell populations, antigen presentation machinery, and T-cell checkpoints. As with the aggregate scores, SDCs tended to have the highest levels, and ACCs the lowest levels, of individual immune cell populations including NK cells, CD8⁺ T cells, Th17 cells, follicular T helper cells, Th2 cells, Treg cells, Th1 cells, activated and inactivated dendritic cells, B cells, mast cells, neutrophils, macrophages, eosinophils, gamma delta T cells, central memory T cells, and effector memory T cells (Fig. 3A). In Supplementary Fig. S4, immune cell subpopulations were analyzed in more detail showing a similar trend with the highest immune cell levels in SDC and the lowest in ACC. Comparable results were observed for measures of the activity of antigen processing/presenting machinery (APM1 and APM2, quantified with ssGSEA, Fig. 3A and B). The APM Score is determined by seven genes: HLA-A/B/C, B2M, TAP1, TAP2, and TAPBP (37). APM score was strongly correlated with T-cell infiltration (Spearman r = 0.84, P < 0.0001; Fig. 3B), consistent with the ability of CD8⁺ T cells to upregulate the expression of APM through secretion of IFNγ, and was highest in SDCs, compared with ACCs and MECAs. Examining the combination of mutational load and antigen presentation, we found that ACCs had the lowest of each measure, indicating that these tumors have profoundly low levels of tumor antigen availability and presentation (Supplementary Fig. S5).

Highly T-cell–infiltrated cancers are often able to escape immune surveillance because the constituent T cells are in a dysfunctional or exhausted state. We therefore examined the expression of T-cell checkpoints PD-1 (and its ligand PD-L1), CTLA-4, LAG3, TIM3, TIGIT, and VISTA (VSIR) in immune-high versus immune-low SGCs, and across histologies. We found that all of these checkpoints were expressed at significantly higher levels in immune-rich salivary tumors, compared with immune-poor tumors: the most significant differences were in PD1, CTLA4, LAG3, TIM3, and TIGIT expression (Supplementary Fig. S6). Delineated for the three different histologies, SDCs had much higher expression of these T-cell checkpoints than ACCs and MECAs (Fig. 3A; Supplementary Fig. S7). We also used an orthogonal measure of T-cell dysfunction, TIDE (Tumor Immune Dysfunction and Exhaustion), which uses a distinct gene signature including TGFB1, SOX10, and other genes independently discovered to modulate the effect of T-cell activity on clinical outcome (40). Again, we found that SDCs had much higher levels of T-cell dysfunction than MECA or ACC (all P < 0.01; Fig. 3D).

More immune-depleted tumors may instead achieve immune evasion by excluding T cells. To examine these possible mechanisms in SGCs, we used TIDE to analyze transcriptome-based gene signatures of T-cell exclusion. We found that immune-poor SGCs had significantly higher T-cell exclusion scores (P = 0.0002), associated with higher levels of myeloid-derived suppressor cells (MDSCs; P < 0.0001) and of M2-polarized tumor-associated macrophages (P < 0.0001; Supplementary Fig. S8A–S8C). This phenotype was seen most clearly in ACCs (Supplementary Fig. S9), which demonstrated higher levels of MDSC and M2 macrophage infiltration (Fig. 3E and F), compared with MECA and SDC. While T-cell exclusion did not significantly differ depending on histology, significant differences were found for MDSC and M2 macrophage infiltration, with the highest levels in ACC (Supplementary Fig. S8D–S8F).

These data reveal several key immune evasion mechanisms at play in salivary tumors. SDCs have high neoantigen load and are highly immune infiltrated, but T cells are characterized by high levels of dysfunction. At the other end of the spectrum, ACCs are characterized by very low neoantigen load, a T-cell exclusion phenotype, and a microenvironment comprised of more immunosuppressive MDSCs and M2 macrophages. MECAs are more heterogeneous, representing the entire spectrum of immunity, with more intermediate mutational load.

To determine whether the level of T-cell infiltration in the tumor microenvironment was associated with clinical outcome, we compared SGCs that ultimately developed recurrence (either locoregional or distant) and those that did not recur. We found that recurrent/metastatic SGCs had a significantly lower level of T-cell infiltration compared with nonrecurrent/metastatic cases (P = 0.02, Fig. 4A). When samples were stratified by histology, the relationship was observed in ACCs (P = 0.008) and MECAs (P = 0.005), but not in SDCs (P = 0.31). Similar results were observed when these comparisons were made with IIS instead (Supplementary Fig. S10).

These data lead us to hypothesize that lack of immune surveillance, as evidenced by an immune-poor microenvironment, may be associated with the ability of ACC and MECA tumors to develop recurrence or distant metastasis after initially definitive therapy. It is important to clarify that in view of the small number of tumors analyzed, these findings are hypothesis-generating. However, a similar association between T-cell infiltration and tumor recurrence was recently observed in ACC by Mosconi and colleagues (46). In our study, this association was not observed in SDCs, perhaps because there were very few SDCs as immune-depleted as the ACCs and MECAs. SDCs that recur or metastasize may utilize other means of immune evasion, although there were no differences in engagement of T-cell checkpoints, T-cell dysfunction, exclusion, or immunosuppressive cell populations in those SDCs developing recurrence.

We also examined the association of immunity with the evolution of carcinomas from benign adenomas. Salivary carcinomas with particularly aggressive behavior can arise from longstanding benign pleomorphic adenomas of the salivary glands, termed “carcinoma ex pleomorphic adenoma” (ex-PA). SDCs and MECAs are the two most common malignant histologies that develop ex-PA.

In this cohort, 6 of 16 SDCs (37.5%) and 24 of 40 MECAs (60%) had arisen from a preexisting pleomorphic adenoma. We observed that SDCs ex-PA had a significantly lower IIS and TIS than SDCs de novo (P = 0.04 and P = 0.03, respectively; Supplementary Fig. S11A and S11B). This difference was mainly driven by lower levels of CD8⁺ T cells in SDCs ex-PA (P = 0.04; Supplementary Fig. S12). In MECA, while IIS and TIS were numerically lower in ex-PA cancers, these differences were not statistically significant (P = 0.09 and P = 0.11, respectively; Supplementary Fig. S11C and S11D).

In these exploratory analyses, we observed that a relative lack of immune surveillance appeared to be associated with more aggressive tumor behavior–the development of tumor recurrence/metastasis in ACCs and MECAs, or the development of a carcinoma from a benign adenoma in SDCs. Further immunophenotyping of salivary cancers and benign tumors will help to elucidate these relationships.

We then examined the transcriptional programs that were differentially activated between immune-rich and the immune-poor tumors (as depicted in Fig. 2A). The transcriptomes from samples with the highest and the lowest IIS values were compared within each of the three SGC histologies. Ingenuity Pathway Analysis (IPA, Qiagen) of significantly differentially expressed genes was used to identify enriched canonical pathways. Among the pathways that were significantly enriched between immune-high and immune-low tumors, 7 pathways were consistently enriched in all 3 SGC histologies (Fig. 4C), including Th1 and Th2 pathways, IL4 signaling, B-cell signaling, and inducible costimulator (iCOS) signaling in Th cells. Although these immune pathways were significantly enriched among immune-rich tumors in all three histologies, the degree of enrichment (as measured by -log₁₀ P) was much higher in SDC and MECA compared with ACC, indicating that the natural range of immunity in ACC tumors is fairly narrow (Fig. 4D). These findings expand upon the phenotypes revealed by RNA-seq deconvolution of enriched immune and T-cell infiltration and activation, and suggest that, in addition, CD4⁺ T-cell activity and Th2 immune responses may have additional relevance to the immune-poor and immune-rich SGC subsets.

In 36 cases (8 ACC, 12 MECA, 16 SDC) with additional exome sequencing available, somatic mutations and mutation-derived neoantigens predicted to bind to patient-specific HLA class I, were identified using netMHCpan (Supplementary Tables S3–S5). Gene fusions can also generate immunogenic neoantigens (47). Gene fusions are particularly prevalent in salivary carcinomas, and represent common oncogenic alterations in ACC and MECA. Analysis of RNA-seq data in SGCs identified fusion-derived neoantigens predicted to bind patient-specific HLA class I in 49 cases (19 ACC, 14 MECA, and 16 SDC; Supplementary Tables S6–S8). SNV- and fusion-associated neoantigen counts are summarized per sample in Supplementary Table S9. As expected, the number of neoantigens generated by missense mutations (single nucleotide variant neoantigens, SNV-NeoAg) was tightly correlated with the number of SNVs in the tumor (Spearman r = 0.82, P < 0.0001; Supplementary Fig. S5). SDCs had the highest, and ACCs the lowest, level of SNV-NeoAg (P = 0.002). Numbers of fusion-derived neoantigens (Fusion-NeoAg) tended to be lower than SNV-Neoag counts. SDCs had the highest, and MECAs the lowest, fusion neoantigen counts (P = 0.02; Fig. 5A).

We analyzed associations between neoantigen landscape and the immune microenvironment. SNV load was correlated with T-cell (Spearman r = 0.48, P = 0.003, Fig. 5B) and immune (Spearman r = 0.52, P = 0.001) infiltration. Predicted SNV-NeoAg abundance had a similar although slightly attenuated correlation with TIS (Spearman r = 0.32, P = 0.052, Fig. 5C) and IIS (Spearman r = 0.34, P = 0.039, Supplementary Fig. S5B). Because both SNV count and immune infiltration vary by histology, we performed multivariable analyses to examine the association between SNV-neoantigen count and immunity metrics, with adjustment for tumor histology. The association between SNV-Neoantigen count and the ESTIMATE immune, CYT, IIS, and TIS scores remained significant, independent of tumor histology (β > 0.35, P < 0.02 for all regression models, Supplementary Table S10).

Similar associations were observed between fusion neoantigen count and immune infiltration (Fig. 5D), with the strongest association seen within the ACC tumors (Supplementary Fig. S5). These associations demonstrate that, among these SGC histologies, which harbor very low (ACC, MECA), or moderate (SDC) mutational load, that the degree of immune infiltration in the tumor microenvironment does appear to correlate with the abundance of neoantigens that derive either from mutations or gene fusions. This suggests that the neoantigen landscape of these tumors impacts immune surveillance, and that for those tumors with very few neoantigens, there is a minimal preexisting antitumor immune response.

For these reasons, salivary cancers with low mutational load and, consequently, low neoantigen load, which also have the least inflamed tumor microenvironments, appear to be the least poised to respond to immune checkpoint inhibitors, at least as monotherapy.

Finally, we examined whether SGCs utilize loss of HLA, which would impair the ability to present a diverse neoantigen repertoire, as a means of immune evasion. Using FACETS (44) to call allele-specific copy number at the HLA class I locus in each case with exome sequencing of normal DNA, we determined which cases harbored either single-copy loss of HLA (hemizygous or copy-neutral loss of heterozygosity) or at least one homozygous HLA class I allele (29). This analysis revealed that HLA LOH/homozygosity was prevalent in ACC (44%) and SDC (31%) tumors, but uncommon in MECAs (8%; Supplementary Fig. S13). We hypothesize that certain SGC histologies, especially ACC, may employ HLA loss as an immune escape mechanism. These findings are preliminary, and further analyses with a larger number of samples will be needed for confirmation. Because HLA status modulates response to immunotherapy in diverse cancer types, these findings may have relevance in future analyses of immunotherapy-treated SGCs.

To further improve clinical benefit from immunotherapy, rational investigation of these drugs should ideally be guided by molecular features that reflect the likelihood of response, and the potential vulnerabilities that may be targetable with different monotherapies or combination therapies.

Here, we describe the immunity landscape of 3 types of salivary gland carcinoma–adenoid cystic (ACC), myoepithelial (MECA), and salivary duct carcinomas (SDC). These are rare, lethal malignancies that are difficult to treat, owing to a lack of active systemic therapies. Data regarding immunotherapy response are limited but thus far have suggested that response rates are low in SGCs. To address this, we performed extensive immunogenomic profiling of 76 clinically annotated salivary cancers.

We find that each of these salivary cancers differs in terms of immunologic microenvironment and likely vulnerabilities. On the one hand, SDCs have the most favorable immunogenomic profile, with moderate TMB and moderate-to-high levels of microenvironment inflammation. In contrast, ACCs have immunogenomic profiles that are less favorable for immunotherapy, with very low TMB, an immune-depleted microenvironment, and possibly a high prevalence of HLA class I loss. MECAs were more intermediate. Understanding these profiles suggests that different immunotherapeutic approaches will be needed for each of these cancer types. Checkpoint inhibitor monotherapy is unlikely to achieve high response rates in any of these cancers.

Our findings from the first comprehensive immune deconvolution of SGCs build upon, and are consistent with, prior studies using IHC or targeted gene expression techniques to describe aspects of immune infiltration in salivary tumors. Mosconi and colleagues analyzed 36 ACC samples by IHC, finding that no cases (0%) were PD-L1 positive, as well as comparably low levels of CD8⁺ and Granzyme B–positive TILs, CD1a⁺ and CD83⁺ cell populations (46), in agreement with our observation that 80% of the ACCs in our study were in the immune-poor cluster, likely driven by a T-cell exclusion phenotype. In a similar study, Chang and colleagues performed IHC staining of several immune markers including CD8 in 70 patients with SGC and found a higher number of CD8⁺ T cells in SDCs and mucoepidermoid carcinomas than in ACCs (48). Sridharan and colleagues performed a targeted gene expression panel in 12 ACCs and again reported a comparably low number of infiltrating immune cells (49).

Our analyses identified possible mechanisms for immune evasion in salivary cancers. In ACCs, T-cell exclusion and comparably higher levels of strongly immunosuppressive MDSC and M2 macrophage populations generate a “cold” immune microenvironment. This immune-depleted microenvironment appeared clinically relevant, because it was associated with a higher likelihood of cancer recurrence in ACCs. Because T-cell checkpoints were universally expressed at low levels in ACCs, and because the mutational load of ACCs is very low (0.3 mutations/megabase), indicating a paucity of tumor antigens, the therapeutic approach of inhibiting single T-cell checkpoints, such as with anti-PD-1 monotherapy, is not likely to be sufficient to achieve responses in ACC. Successful immunotherapy for ACC is likely to require approaches to reverse T-cell exclusion and the immunosuppressive sequelae of MDSC and M2 macrophage infiltration, perhaps leveraging emerging agents such as immune agonists, vaccines, or targeting immunosuppressive cytokines and the peritumoral stroma.

In contrast, SDCs harbor much higher levels of T-cell infiltration that were associated with markedly higher levels of T-cell dysfunction and high expression of T-cell checkpoints. This is in agreement with prior literature reporting rates of PD-L1 positivity in SDCs of 30%–60% (50). In combination with a higher availability of tumor antigens resulting from a moderate mutational burden (1.7 mutations/megabase), these data indicate that SDCs are more poised to respond to T-cell checkpoint inhibitors, although combination therapies that target multiple checkpoints may ultimately be needed.

MECAs were a more heterogeneous group of cancers, with some clustering closer to ACC and some clustering closer to SDC in terms of immune infiltration. Further investigation will be needed to determine if it is clinically helpful to profile these cancers with measures of T-cell infiltration and tumor mutational burden to triage them to therapy. All of these potential approaches for these cancer types will require further investigation.

Despite their lower mutational load, ACCs and MECAs do have another potential source of neoantigens: highly prevalent fusion genes, such as MYB-NFIB in ACC and PLAG1 fusions in MECA. In this study, we found that fusion-derived neoantigens were associated with higher levels of T-cell inflammation. We have previously shown that several patients with metastatic ACC had circulating T cells that recognize a MYB-NFIB fusion–associated peptide (47). Taken together, these data suggest that fusion-derived neoantigens may represent a source of targetable tumor antigens in SGCs.

A major challenge faced by the field of salivary oncology is the rarity of these entities, and the large number of cancer histologies included under the category of “salivary cancer.” These data can help to inform and prioritize immunotherapeutic research strategies for each of these cancer types. The molecular landscape of each salivary cancer histology is unique and informative. It is notable that in recent years, targeted therapies focused on specific molecular alterations in individual SGC histologies have begun to demonstrate activity, including recent trials of lenvatinib (51) and NOTCH1-targeted agents in ACC (52, 53), and androgen receptor and Her2/ERBB2-targeted approaches in SDC (54, 55).

Our study was necessarily limited in that we selected only the most clinically aggressive types of SGC. We did not profile less clinically aggressive SGC types such as mucoepidermoid or polymorphous adenocarcinomas. Clinical data from immune checkpoint blockade in SGCs has been limited to the KEYNOTE-028 study (25), but we await the results of a number of ongoing phase II trials enrolling more patients with SGC (KEYNOTE-158, NCT02628067; NISCAHN, NCT03132038; NCT02538510).

This study aims to provide new insights into the immune microenvironment, potential mechanisms of immune evasion, and possible vulnerabilities needing further study, in these rare, aggressive, often incurable cancers. Our data reveal that these cancer types are vastly different from one another from the standpoint of immunity. Despite the heterogeneity of SGCs, these cancers are often grouped together in the context of clinical trials. These data provide necessary information that may help to guide the next steps of investigation, perhaps helping to prioritize or nominate certain immunotherapeutic approaches for certain SGC histologies.

L.G.T. Morris reports receiving commercial research grants from AstraZeneca and reports receiving other commercial research support from Illumina Inc. D. Chowell holds ownership of intellectual property on using TMB to predict immunotherapy response, with pending patent, which has been licensed to PGDx. N. Riaz is an employee/paid consultant for REPARE Therapeutics and Mirati Technologie, reports receiving commercial research grants from Bristol-Myers Squibb and Pfizer, and reports receiving speakers bureau honoraria from Illumina. A.L. Ho is a paid consultant for Ayala, CureVac, Eisai, Genzyme, Novartis, Kura Oncology, AstraZeneca, Merck, Ignyta, Genentech/Roche, Sun Pharmaceuticals, Regeneron, TRM Oncology, Eisai, Klus Pharma, and reports receiving speakers bureau honoraria from Medscape. T.A. Chan is a paid consultant for Bristol-Myers Squibb and Illumina, reports receiving commercial research grants from AstraZeneca, Illumina, Pfizer, BMS and holds ownership interest (including patents) in Gritstone. No potential conflicts of interest were disclosed by the other authors.

Conception and design: L.G.T. Morris, M. Linxweiler, T.A. Chan

Development of methodology: L.G.T. Morris, M.G. Dalin

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): L.G.T. Morris, S. Dogan, M. Lee

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): L.G.T. Morris, M. Linxweiler, F. Kuo, M.G. Dalin, V. Makarov, D. Chowell, A.A. Hakimi, N. Riaz, T.A. Chan

Writing, review, and/or revision of the manuscript: L.G.T. Morris, M. Linxweiler, F. Kuo, N. Katabi, M.G. Dalin, D. Chowell, S. Dogan, I. Ganly, A.A. Hakimi, R.J. Wong, N. Riaz, A.L. Ho, T.A. Chan

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L.G.T. Morris, Z. Nadeem, S. Dogan, T.A. Chan, M. Lee

Study supervision: L.G.T. Morris, T.A. Chan

We express our deep gratitude to our brave patients and their families who participated in these studies. We thank Jeffrey and Marnie Kaufman and Nicole Spardy Burr from the Adenoid Cystic Cancer Research Foundation for their support and leadership, and members of the Chan Lab for helpful discussions. We acknowledge funding from the Frederick Adler Chair (to L.G.T. Morris), Cycle for Survival (to L.G.T. Morris), the Jayme Flowers Fund (to L.G.T. Morris), the Sebastian Nativo Fund (to L.G.T. Morris), NIH K08 DE024774 and R01 DE027738 (to L.G.T. Morris), the Adenoid Cystic Carcinoma Research Foundation (to T.A. Chan), Pershing Square Sohn Cancer Research grant (to T.A. Chan), the PaineWebber Chair (to T.A. Chan), Stand Up 2 Cancer (to T.A. Chan), NIH R01 CA205426, the STARR Cancer Consortium (to T.A. Chan), NCI R35 CA232097 (to T.A. Chan), and funding to MSKCC through NIH/NCI Cancer Center Support Grant (P30 CA008748).

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