Comprehensive Molecular Characterization of Salivary Duct Carcinoma Reveals Actionable Targets and Similarity to Apocrine Breast Cancer

This article is featured in Highlights of This Issue, p. 4537

Salivary duct carcinoma (SDC) is one of the most aggressive head and neck tumors, accounting for 2% of salivary gland cancers. The majority of cases originate in the parotid or submandibular gland, with lymph node metastasis and facial nerve palsy commonly present at diagnosis (1, 2). Standard treatment includes surgical resection with or without adjuvant radiotherapy, but local disease recurrence and distant metastasis are common and rarely responsive to chemotherapy. Therefore, the overall prognosis remains poor with a long-term survival of 30% to 55% (1–3).

SDCs may arise de novo or develop as the malignant component of a carcinoma ex-pleomorphic adenoma (denoted SDC ex-PA), which is a cancer that has transformed from a longstanding pleomorphic adenoma. Histologic findings include a hyalinized, fibrous stroma infiltrated by neoplastic ducts, which resembles the morphology of invasive ductal carcinoma of the breast (4). SDC and breast cancer also share several IHC characteristics, including positive staining for the androgen receptor (AR), which is detected in a majority of both tumor types (3, 5–7). Several case series have reported clinical benefit from androgen deprivation therapy (ADT), which has been suggested as potential first-line therapy in patients with recurrent AR-positive SDC (8, 9).

Amplification of ERBB2 (also known as HER2) has been detected in approximately one third of SDCs and is associated with poor prognosis (3, 5, 10, 11). The ERBB2 inhibitor trastuzumab has been tested with reported cases of clinical response (12–14), although no controlled study has been performed. It has been proposed that SDC, analogous to breast cancer, can be divided into different subtypes based on receptor status: AR-positive, ERBB2-positive, and basal-like (AR- and ERBB2-negative) phenotypes (6). However, this hypothesis is based on immunohistochemical data only, and it is unclear whether SDC resembles breast cancer on a broader molecular level.

Previous genetic studies of SDC have focused on a limited number of genes and have found TP53 alterations in the majority of patients. Other recurrently mutated genes include KIT, PIK3CA, and HRAS (5, 15). To date, no unbiased comprehensive molecular characterization of SDC with whole-exome or transcriptome sequencing has been reported.

In this study, we performed whole-exome sequencing of tumor and matched normal DNA, as well as tumor RNA sequencing (RNA-Seq), and analyzed the spectrum of genetic alterations in 16 SDC patients. To classify SDC in comparison with breast cancer, global gene expression was analyzed and compared with published datasets of breast carcinoma. For additional comparison, we also analyzed targeted sequencing data in a second cohort of 15 recurrent/metastatic SDCs that were profiled as part of an institutional precision medicine platform. We report a high frequency of potentially actionable genetic alterations in this rare, often lethal cancer. Taken together, this is the first comprehensive genetic study of SDC, providing a molecular foundation for future research and clinical trials.

With written informed consent and following Institutional Review Board (IRB) approval, tumor and matched normal (peripheral blood or nonneoplastic normal tissue) specimens were obtained from SDC patients treated at Memorial Sloan Kettering Cancer Center (New York, NY) during the years 2000 to 2015. For cohort 1 (denoted SDC1–SDC16), tissue samples were snap frozen in liquid nitrogen at the time of surgery and stored at −80°C. Hematoxylin and eosin–stained tumor sections were independently reevaluated by a head and neck pathologist (N. Katabi) confirming the SDC diagnosis. Sixteen cases with sufficient tumor material were subjected to whole-exome sequencing of tumor and normal DNA, as well as tumor RNA-Seq.

For cohort 2 (denoted SDC17–SDC31), tumors were sequenced using a clinical next-generation sequencing platform, Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT; ref. 16). These were patients with advanced (all recurrent or metastatic) cancers who were offered targeted tumor sequencing in the context of an IRB-approved study (NCT01775072). For these cases, tumor DNA and matched normal blood DNA were subjected to massively parallel sequencing, targeting the exons of 410 selected genes and selected intronic and regulatory regions (listed in Supplementary Table S1), with technical details of this assay described previously (16).

DNA was extracted from fresh-frozen tissue or whole blood using the DNeasy Blood and Tissue Kit (Qiagen) and quantified with the PicoGreen assay (Thermo Fisher Scientific). RNA was extracted using the RNeasy Mini kit (Qiagen) and quantified with the RiboGreen assay (Thermo Fisher Scientific). DNA and RNA quality and integrity were analyzed by BioAnalyzer (Agilent Technologies) and Fragment Analyzer (Advanced Analytics).

Whole-exome sequencing libraries were prepared using the SureSelectXT Library Preparation Kit (Agilent). DNA was 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 enriched by PCR. Final libraries were quantified using 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.

RNA-Seq libraries were prepared using the KAPA Stranded RNA-Seq with RiboErase sample preparation kit (KAPA Biosystems). Total RNA (100 ng) was ribosomal RNA-depleted and fragmented, followed by first and second strand synthesis, A-tailing, adapter ligation, and PCR amplification (using 11 cycles). Final libraries were quantified using the KAPA Library Quantification Kit (KAPA Biosystems), Qubit Fluorometer (Life Technologies), and 2100 BioAnalyzer (Agilent) and were sequenced on a HiSeq2500 v4 chemistry sequencer (Illumina) using 2 × 125 bp cycles.

The match between tumor and normal samples of each patient was confirmed with fingerprinting analysis using an in-house panel of 118 SNPs and with VerifyBamID (17). Raw sequencing data were aligned to the hg19 genome build using the Burrows–Wheeler Aligner version 0.7.10 (18). Indel realignment, base quality score recalibration, and removal of duplicate reads were performed using the Genome Analysis Toolkit version 3.2.2 (broadinstitute.org/gatk), following guidelines for raw read alignment (19).

Single nucleotide variations (SNV) were independently detected by 4 callers; MuTect (broadinstitute.org/cancer/cga/mutect), Somatic Sniper v1.0.4.2 (gmt.genome.wustl.edu/packages/somatic-sniper), Strelka v1 (sites.google.com/site/strelkasomaticvariantcaller), and Varscan v2.3.8 (varscan.sourceforge.net). SNVs that were identified by at least 2 different callers, with >10% variant allelic fraction and >7× coverage in tumor and ≥15× coverage with >97% normal allelic fraction in normal, were considered high-confidence variants. SNVs that did not meet these criteria, but had ≥4× coverage, >6% variant allelic fraction in tumor, ≥4× normal coverage with >97% normal allelic fraction in normal, and passed manual review via Integrative Genomics Viewer (IGV) v2.3 (broadinstitute.org/igv), were considered low-confidence variants. Insertions and deletions (indels) were detected by Strelka and VarScan. Variants that passed manual review in IGV with ≥4× tumor allelic coverage, >10% tumor allelic fraction, ≥4× normal DNA coverage, and >97% normal allelic fraction were considered as potential indels.

All potential indels and low-confidence SNVs, a random selection (15%) of the high-confidence SNVs, and the telomerase reverse transcriptase (TERT) gene promoter were subjected to orthogonal validation using NimbleGen SeqCap EZ target enrichment (Roche), using 500× and 250× sequencing depth for tumor and normal DNA, respectively. Variants with >100× of tumor and normal DNA coverage, >15% variant allelic fraction in tumor, and <3% variant allelic fraction in normal DNA, were considered validated. The overall validation rate in high-confidence SNVs was 96.2%. Validated indels and low confidence SNVs, as well as all high-confidence SNVs, were included in subsequent analyses.

Allele-specific copy number data were acquired by analysis of whole-exome sequencing bam files using OncoSNP-SEQ (sites.google.com/site/oncosnpseq). Chromosome arm-level alterations and statistically significant amplifications and deletions were determined and visualized using GISTIC 2.0 (broadinstitute.org/cancer/cga/gistic). To determine tumor purity, we analyzed RNA-Seq data using the ESTIMATE algorithm (20).

For SDC samples, raw FASTQ files were aligned to the hg19 genome using STAR aligner with default parameters (21). Aligned fragments were counted with Rsamtools v3.2 and annotated using the TxDb.Hsapiens.UCSC.hg19.knownGene version 3.2 transcript database. The raw count matrix for 591 PAM50-annotated breast cancers published by The Cancer Genome Atlas (TCGA; ref. 22) was downloaded from the TCGA Genome Data Analysis Center firehose (doi:10.7908/C11G0KM9) and merged with the SDC count matrix. Regularized logarithm transformation of the matrix was obtained with the rlog function of DeSeq2 v1.10.1 after removing the batch effect using the svaseq function of the sva package v3.18.0. FPKM were obtained with DESeq2 using the robust method. The 100 genes with the highest variance of rlog-transformed data across both breast and SDC samples were selected and clustered together using a Euclidean distance to define molecular subtypes.

Unsupervised consensus clustering of SDCs and 109 basal-like breast cancer samples was performed using ConsensusClusterPlus package v1.22.0. The number of clusters was determined empirically according to the consensus index and delta provided in the package. Gene set enrichment analysis (GSEA) of the different clusters was performed using the Farmer and colleagues breast cancer signatures 1 through 7 (23).

Sections of formalin-fixed paraffin-embedded (FFPE) tumor tissue were stained with AR antibody clone AR-441 (monoclonal mouse, dilution 1:100; Dako) and were considered positive if immunoreactivity was detected in ≥5% of tumor cell nuclei. ERBB2 staining was done using the anti-HER-2/neu antibody, clone 4B5 (monoclonal rabbit; Roche), and was visualized using the UltraView Universal DAB Detection Kit (Ventana Medical Systems). S100 was detected using antibody GA504 (polyclonal rabbit, dilution 1:8,000; Dako).

FFPE tissue samples were tested for NTRK3 rearrangements by a break-apart FISH assay using a commercial NTRK3 Break Apart FISH Probe (Empire Genomics). Four-micron (4 μm) FFPE tissue sections generated from FFPE blocks of tumor specimens were pretreated by deparaffinizing in xylene and dehydrating in ethanol. Dual-color FISH assay was conducted according to the protocol for FFPE sections from Vysis/Abbott Molecular with a few modifications. FISH analysis and signal capture were conducted on a fluorescence microscope (Zeiss) coupled with ISIS FISH Imaging System (Metasystems). We analyzed 100 interphase nuclei from each tumor specimen.

The AR locus of the RNA-Seq data was displayed with Sashimi plots in IGV 2.3. A sample was reported as AR splice variant 7 (AR-V7) positive if at least two uniquely aligned junction reads spanning exon 3 and the downstream cryptic exon (expressed intronic region CE3) were detected, with a minimum of 5 nucleotide overhang on either side without mismatches. The number of reads spanning AR-V7 variant splice junctions was normalized to the total number of reads aligned to the reference transcriptome.

For validation of fusion genes and expression of the AR-V7 splice variant, cDNA was synthesized from tumor RNA using SuperScript III Reverse Transcriptase (Thermo Fisher Scientific), amplified by PCR, and visualized on a 1.5% agarose gel. For validation of fusion genes with intronic breakpoints, PCR was performed using tumor genomic DNA. All primer sequences are listed in Supplementary Table S2.

To generate an AR signaling index, we calculated the average z-score of 46 canonical target genes that are upregulated by the AR (Human Androgen Receptor Signaling Targets PCR Array, Qiagen), calculated from FPKM data.

To investigate the spectrum of somatic genetic alterations in SDC, we performed whole-exome sequencing of snap-frozen tumor and matched normal DNA from 16 patients (see Table 1 for clinical information). The mean coverage was 143× and 74× for tumor and normal DNA, respectively, with 98% of the target sequence covered to at least 20× depth (Supplementary Fig. S1). We also performed RNA-Seq, with an average of 170M total reads, of which 88% were aligned to the mapping sequence with high quality (Supplementary Fig. S2). Orthogonal resequencing showed a high validation rate (>96%) of the mutations detected, and only high-confidence and/or validated mutations were included in subsequent analyses (see Materials and Methods). The average estimated tumor purity was 68% (range, 52%–93%), which is higher than reported for most tumor types, including breast cancer (20).

We detected a median of 50 nonsilent somatic mutations per tumor, corresponding to approximately 1.7 mutations per megabase (Fig. 1; Supplementary Table S3). This places the mutational load of SDC in the lower one third of solid tumors, close to breast cancer as well as pancreas, prostate, and kidney cancers (24). This mutational burden is higher than that of adenoid cystic carcinoma of the salivary gland, which has 0.3 mutations per megabase (25). In total, 86% were missense mutations, 12% truncating mutations, 2% splice site mutations, and 0.6% in-frame indels. RNA-Seq detected 50% of all mutations but 94% of mutations in the COSMIC database occurring in cancer genes (Supplementary Fig. S3A). The latter group mostly contained mutations with evidence supporting likely importance as drivers of oncogenesis, suggesting that such alterations may be commonly expressed and detected by RNA-Seq. The variant allele frequencies in this group of mutations were higher in RNA than in DNA (Supplementary Fig. S3B), which is in line with a previous study (26), and supports the hypothesis that many transforming mutations are highly expressed in tumor cells.

We noted a relatively low rate of copy number alterations (CNA), with amplifications or deletions on chromosome arm-level occurring in 5 patients. The only recurrent arm-level CNA was 8q amplification, present in 3 of 16 (19%) patients (Fig. 1), which is in line with previous findings in estrogen receptor–negative breast cancer and head and neck squamous cell carcinoma (22, 27). The 8q arm contains oncogenes such as MYC, which has shown recurrent amplifications associated with poor prognosis in several tumor types, including prostate cancer (28), and PLAG1, which is rearranged and overexpressed in a majority of pleomorphic adenomas (29). Significant focal amplifications included loci 17q12 (including ERBB2) and 10q11.21 (including RET) in 4 and 1 of 11 patients, respectively (Supplementary Fig. S4).

Unique fusion genes were detected in 5 of 16 patients using RNA-Seq and were all confirmed by PCR analysis (Fig. 1 and Supplementary Fig. S5). One patient had an ETV6–NTRK3 fusion, which is found in a majority of mammary analogue secretory carcinoma (MASC) tumors (30). The fusion gene included ETV6 exons 1–4 and NTRK3 exons 14–20, which is different from the most common fusion variant (where the breakpoints are located in ETV6 exon 5 and NTRK3 exon 15), but has been previously reported in MASC (31). FISH analysis showed rearrangement of the NTRK3 locus in 92% of the cells (Supplementary Fig. S6D). To confirm the SDC diagnosis, we re-reviewed additional tissue sections and noted typical SDC morphology (Supplementary Fig. S6A–S6C), which is distinct from that of MASC (32). Furthermore, the tumor was positive for AR and negative for S100, which have been previously detected in 0 of 9 and 15 of 15 cases of MASC, respectively (33). Although the ETV6–NTRK3 fusion has been reported in breast secretory carcinoma and sporadically in other diseases, such as congenital mesoblastic nephroma, congenital fibrosarcoma, and acute myeloid leukemia (30), it has not been detected in salivary gland tumors other than MASC. This finding may suggest that the ETV6–NTRK3 fusion is not pathognomonic for MASC. An alternative explanation could be that this case was initially a MASC that underwent a transformation with the emergence of a high-grade SDC-like tumor. However, this appears unlikely considering the absence of MASC histology or IHC findings in multiple regions in the tumor and that no previously reported cases of MASC with high-grade transformation have shown SDC morphology. This case also demonstrates the potential clinical value of tumor sequencing, because the ETV6–NTRK3 fusion is targetable. Our group recently reported a near-complete response to TRK inhibition in a patient whose salivary carcinoma harbored a similar ETV6–NTRK3 fusion gene (34).

Two patients with SDC ex-PA had CTNNB1–PLAG1 or LIFR–PLAG1 fusions, which have been previously described in pleomorphic adenoma (30). Another patient had a BCL6–TRADD fusion gene. BCL6 is commonly translocated in diffuse large B-cell lymphoma (35) but has never been reported with TRADD as a fusion partner. Finally, one tumor harbored an ABL1–PPP2R2C fusion. Rearrangements including the 3′ portion of ABL1 with several different 5′ fusion partners are found in hematologic malignancies (36). However, the reciprocal PPP2R2C–ABL1 fusion was not detected by RT-PCR.

Recurrent mutations and CNAs are depicted in Fig. 1A. We interrogated selected genes in a second cohort, consisting of 15 patients with recurrent or metastatic SDC, whose tumors were analyzed using MSK-IMPACT, a clinical sequencing assay designed to detect mutations and CNAs in 410 cancer-related genes (Supplementary Fig. S7A and Supplementary Table S4 for results; Supplementary Table S5 for clinical information) (16). Overall, SDCs showed a relatively low number of mutations and CNAs compared with other tumor types that were analyzed using the same platform (Supplementary Fig. S7B), which is in line with our findings in cohort 1.

Across both cohorts (n = 31), TP53 alterations were most prevalent and were detected in 17 (55%) cases (Fig. 1B). The frequency of TP53 alterations was higher in cohort 2 than in cohort 1 (80% vs. 31%, OR = 8.8; P = 0.011; Fisher exact test), suggesting that this mutation may be enriched in tumors that behave aggressively, as all cohort 2 tumors were recurrent/metastatic.

Amplification of ERBB2 was detected in 10 of 31 (32%) cases, which is similar to previous SDC studies (3, 5, 10, 11). The ERBB2 inhibitor trastuzumab is used as standard treatment of ERBB2-amplified breast cancer (37) and has been tested with promising results in SDC (12–14). In our study, 3 patients with ERBB2 amplification were treated with trastuzumab in combination with chemotherapy. One patient received trastuzumab for 1 year as part of adjuvant therapy and had no sign of disease at follow-up 6 years later, one was alive with stable bone metastases after receiving trastuzumab as maintenance treatment for 4 years, and one was enrolled on a clinical trial with an anti-PD1 antibody after experiencing progression of disease after 1 year of trastuzumab treatment. Alterations leading to activation of the PI3K/AKT/mTOR pathway may cause ERBB2 inhibitor resistance in breast cancer (38), and AKT1 amplifications, RPTOR amplification/mutations, or truncating PTEN mutations were detected in 4 of 10 ERBB2-amplified tumors (Fig. 1A and Supplementary Fig. S7A). Our findings warrant further studies to investigate the effect of such alterations on response to ERBB2 inhibition in SDCs harboring ERBB2 gene amplification.

Of note, most tumors (74%) had mutations in either the MAPK pathway (BRAF, HRAS, and NF1) or in ERBB2 (P for mutual exclusivity = 0.057, Fisher exact test). This indicates that ERBB2 amplification and MAPK pathway activation are the predominating drivers of oncogenesis in SDC and may act through independent or redundant mechanisms.

Seven of 31 (23%) patients harbored activating HRAS mutations. Interestingly, 3 of 3 patients with an HRAS G13R mutation also had a cooccurring PIK3CA H1047R mutation. On the other hand, no PIK3CA mutation was detected in the 4 patients with HRAS Q61R (P = 0.029; Supplementary Table S6). These data suggest that HRAS Q61R might be a more potently activating mutation, whereas HRAS G13R cooperates with PIK3CA H1047R to promote oncogenesis in SDC. Similar findings have been observed in other cancer types, where PIK3CA has been shown to cooperate with specific RAS mutations (39), in a manner that can generate dependency on PI3K signaling and resistance to ERBB2 inhibition. These data may have relevance for the targeting of PIK3CA or ERBB2 in SDC.

Mutations in the promoter region of the human TERT gene occur in several cancer types (40). We sequenced the TERT promoter and found no mutations in any of the 31 tumors, suggesting that these alterations are rare in SDC.

In total, 21 of 31 (70%) patients had a history of tobacco smoking. There were no associations between smoking status and any genetic alterations detected in the study. There were also no significant associations between the above-described genetic alterations and overall or recurrence-free survival (data not shown).

To investigate the clinical relevance of molecular characterization in SDC patients, we categorized alterations according to levels of evidence supporting standard or investigative therapies. There were no alterations that have standard-of-care therapeutic implications, as there are no FDA-approved drugs or biomarkers in SDC. However, 19 tumors (61%) had alterations that were potentially actionable, because the specific alteration corresponded to either an FDA-approved treatment or to a treatment with published clinical evidence in another cancer type. An additional 3 tumors (10%) contained alterations in which preclinical studies suggest potential activity of such drugs (Fig. 1C; Supplementary Table S7).

Taken together, a high percentage of SDCs have actionable alterations with supportive clinical evidence. These findings are in stark contrast to the two other salivary cancer types that have been genetically profiled, adenoid cystic carcinomas and polymorphous low-grade adenocarcinomas, which both have a relative paucity of actionable targets (25, 41). Therefore, routine clinical tumor sequencing in SDCs may yield potentially actionable information.

IHC showed positive staining for AR in 12 of 16 (75%) cases (Fig. 1A). As ADT has been suggested as potential treatment for AR-positive SDC (8, 9), we investigated potential molecular mechanisms of ADT resistance in these tumors. AR-V7, including AR exons 1–3 and a cryptic exon 3 (CE3) instead of 1–8 (Fig. 2A), has been detected in prostate cancer and is associated with ADT resistance (42). AR-V7 lacks the ligand-binding domain but is transcriptionally active. Analysis of the RNA-Seq data showed evidence of AR-V7 expression in 8 of 16 cases. When compared with full-length AR, the ratios of AR-V7 (mean 5.1%, range 0.5%–15.1% of total AR reads) were comparable with those detected in prostate cancer (Fig. 2B and C; ref. 43). Expression of AR-V7 was confirmed by RT-PCR in 7 of the 8 cases (Fig. 2C). This confirms data from a prior study identifying AR-V7 in SDC (44) and additionally demonstrates that in many cases, the ratios of AR-V7 are quite low in comparison with full-length AR. These results suggest that future clinical trials investigating ADT in SDC should incorporate assessment of the presence, and ratios of AR-V7, in tumors before and during/after treatment and that RNA-Seq is a more sensitive measure than RT-PCR.

Forkhead box protein A1 (FOXA1) is a pioneer transcription factor that acts by binding to and exposing chromatin, thereby enabling AR and other hormone receptors to activate transcription of their target genes. In prostate cancer, AR and FOXA1 protein levels correlate with each other, and high FOXA1 levels are associated with poor prognosis and the development of metastasis (45). In SDC, we also noted a strong correlation between FOXA1 and AR mRNA levels (Fig. 2D).

Although functional studies are lacking, clinical data have suggested that FOXA1 mutations may be associated with ADT resistance (46). Three SDC patients had FOXA1 mutations, which did not appear to affect AR expression levels (Fig. 2D). Like most FOXA1 mutations found in ADT-resistant prostate cancer (46), these were all located in the DNA-binding domain (Fig. 2E). To investigate the effect of the FOXA1 mutations on AR activity, we compared AR mRNA levels with a composite estimate of AR signaling, derived from the average expression of 46 canonical genes upregulated by AR (see Materials and Methods). AR signaling correlated with AR expression at both protein and mRNA levels. Interestingly, the tumor with a Q263-C268 FOXA1 deletion showed low AR signaling despite the highest AR expression of all cases (Fig. 2F and G). This suggests that the Q263-C268 deletion represses AR signaling, although these data are preliminary and the hypothesis requires additional confirmation. Nevertheless, the presence of FOXA1 mutations in SDC is likely to have significant implications as ADT is investigated in SDC.

We also detected two missense mutations, one frameshift insertion, and one amplification of the fatty acid synthase (FASN) gene. Overexpression of FASN is prevalent in several tumor types and been proposed as a potential ADT resistance mechanism in prostate cancer (47), although mechanistic data remain lacking.

In our study, 4 patients with AR-positive tumors and recurrent disease received ADT treatment with or without chemotherapy. Of these patients, 3 developed progression of disease and 1 currently remains on therapy with stable disease after 17 months.

Morphologically, SDC resembles invasive ductal carcinoma of the breast, and the two diseases share some immunohistochemical features, such as AR and ERBB2 expression, which are seen in 58% to 87% and 20% of breast cancers, respectively (6, 7, 48). Thus, it is possible that SDC and breast cancer may be molecularly similar. However, this has not been previously investigated at the scale of global gene expression.

We performed unsupervised hierarchical clustering based on gene expression data from RNA-Seq in the 16 cases of SDC and 591 breast cancers sequenced by TCGA (22). Despite arising from a different organ, the SDC tumors did not cluster separately from the breast cancers. Instead, they clustered together with basal-like and HER2-enriched breast tumors and separately from Luminal A, Luminal B, and normal-like breast tumors (Fig. 3A). As has been suggested based on IHC data (4, 6), we found that SDCs could be divided into three groups based on AR and ERBB2 gene expression. Four tumors had low AR and ERBB2 levels, 8 tumors had high AR and low ERBB2, and 4 tumors had high AR and ERBB2 (Fig. 3B). Interestingly, 3 of 4 SDCs with low AR and ERBB2 expression clustered with basal-like breast cancers, whereas AR-high cases, regardless of ERBB2 status, predominately clustered with ERBB2-enriched breast cancers (Fig. 3C). These findings indicate that SDC shows a distinct similarity to breast cancer not only by histology and IHC, but also by global gene expression patterns.

To determine which breast cancer subtype SDC is closest to based on molecular profile, we then performed consensus clustering of the SDCs and 109 basal-like breast tumors from TCGA according to methods described by Lehmann and colleagues (49). Similar to that study, we detected an optimal number of 7 clusters (Supplementary Fig. S8). All SDC cases were classified together with a subset of the breast tumors forming cluster 1 (Fig. 4A). To explore the nature of this cluster, we performed GSEA, using gene sets specific for different subtypes of breast cancer as defined by Farmer and colleagues (23). The differentially expressed genes associated with tumors of cluster 1 showed a significant enrichment for Farmer and colleagues' gene cluster 7 (P < 0.0001), which defines the molecular apocrine subtype of breast cancer (Fig. 4B). There was no significant overlap with any other of Farmer and colleagues' gene clusters (Fig. 4C).

Of note, all SDCs were classified in the cluster of molecular apocrine-like tumors, regardless of AR status. This raises the possibility that AR-negative tumors have alternative mechanisms for the activation of AR target genes. Indeed, 3 of 4 tumors negative for AR gene expression had a high to average AR signaling index (Fig. 2G). GSEA analysis of genes differentially expressed in the 12 AR-positive compared with 4 AR-negative SDCs revealed a strong enrichment of genes dysregulated by androgen stimulation of the molecular apocrine breast cancer cell line MDA-MB-453 (Supplementary Fig. S9; ref. 50), further indicating that the two cancer types respond to androgens in a similar fashion. A clinical trial evaluating ADT in AR-positive breast cancer is ongoing (NCT01889238) and may give further insights to molecular markers of response to this treatment.

To our knowledge, this is the first study to comprehensively analyze genetic alterations and gene expression features in SDC, providing a rationale for future targeted therapy trials. Our results illuminate the unique biology of SDC, identify a high prevalence of actionable molecular alterations, and have direct implications for clinical trials.

SDCs are aggressive salivary cancers, with no highly active systemic therapy, and no predictive biomarkers. The majority of tumors are ultimately treatment resistant and lethal. Our results show that most (61%) SDCs have alterations with published clinical evidence supporting specific targeted therapies. In the emerging era of precision oncology and biomarker-defined basket studies, these data suggest that routine clinical tumor sequencing for this disease is likely to be of high clinical value.

Our results suggest that ADT could be investigated in the majority of SDCs, and ERBB2 inhibition in a subset of the patients. We found evidence supporting several potential mechanisms of resistance to these drugs, such as AR-V7 expression and FOXA1 mutations in AR-positive tumors, and PI3K–AKT pathway alterations in cases with ERBB2 amplification. It will be important to assess these genetic and expression-based biomarkers in correlative analyses for patients enrolled on ADT or ERBB2 inhibitor trials.

These data also define the diverse molecular landscape of SDC. For example, we describe the first occurrence of the ETV6–NTRK3 fusion gene in a tumor with histologic features and IHC profile typical of SDC.

Finally, we show that the gene expression pattern of SDC is highly similar to that of molecular apocrine breast carcinoma. This has long been speculated on the basis of similarities in IHC but has not heretofore been confirmed at the level of genome-wide data. It is likely that parallel clinical investigation of ADT and other treatments in apocrine breast cancer will generate useful data and hypotheses for SDC trials, which are currently constrained by the rarity of this cancer.

L. Wang is a consultant/advisory board member for Roche Diagnostics. No potential conflicts of interest were disclosed by the other authors.

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

Development of methodology: M.G. Dalin, A. Desrichard, L.A. Walsh, K.-W. Lee, L. Wang, D.B. Solit, T.A. Chan, L.G.T. Morris

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M.G. Dalin, A. Desrichard, L. West, S. Dogan, L. Wang, D.B. Solit, N.D. Schultz, T.A. Chan, L.G.T. Morris

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.G. Dalin, A. Desrichard, N. Katabi, V. Makarov, L.A. Walsh, Q. Wang, J. Armenia, A.L. Ho, D.B. Solit, M.F. Berger, J.S. Reis-Filho, T.A. Chan, L.G.T. Morris

Writing, review, and/or revision of the manuscript: M.G. Dalin, A. Desrichard, N. Katabi, L.A. Walsh, Q. Wang, L. Wang, A.L. Ho, I. Ganly, D.B. Solit, J.S. Reis-Filho, T.A. Chan, L.G.T. Morris

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L.A. Walsh, D. Ramaswami, M.F. Berger, T.A. Chan, L.G.T. Morris

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

The authors thank R. Ghossein for expert head and neck pathology tumor review; T. Nielsen for graphic design; and G. Stenman, M. Persson, PA. Watson, E. Adams, and H. Hieronymus for discussions and technical assistance.

This work was supported by NIH P30 CA008748. M.G. Dalin was supported by Sahlgrenska University Hospital, The Gothenburg Medical Society, The Swedish Society of Medicine, and Svensson's Fund for Medical Research. L.G.T. Morris was supported by the Damon Runyon Cancer Research Foundation, NIH K08 DE024774, the Society of MSK, and the MSK Translational and Integrative Medicine Research Fund T.A. Chan was supported by grants from the Adenoid Cystic Carcinoma Research Foundation and the Geoffrey Beene Cancer Center.

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.