Salivary Gland Cancer Patient-Derived Xenografts Enable Characterization of Cancer Stem Cells and New Gene Events Associated with Tumor Progression

Salivary gland cancer (SGC) afflicts approximately 4,000 adults in the United States annually and includes several distinct histotypes, of which mucoepidermoid carcinoma (MEC) and adenoid cystic carcinoma (ACC) are the most prevalent (1, 2). Preclinical models of SGC are scarce due to the difficulty in establishing models (3). These already limited models are further diluted by disease histotype (4) and cell line misidentification (5). Research using patient-derived xenografts (PDX) from more common malignancies has led to enhanced preclinical and translational research (6). However, to date, only ACC PDX models have been reported (7, 8), with no reports characterizing PDX models from the most frequent subtype, MEC.

Late recurrences (longer than 5 years after diagnosis) manifesting primarily as distant metastases occur in upwards of 26% and 17% of ACC and MEC patients, respectively (9). This is particularly striking in ACC patients with a promising 5-year survival rate of 75% to 80%, which plummets to 35% at 10 years and to 10% by 20 years (10, 11). Recurrences after a decade or more suggest the persistence of a tumor cell population that is senescent, can self-renew, and upon migration to distant sites can replicate the morphology and heterogeneity of the originating tumor; these characteristics are consistent with the accepted definition of a cancer stem cell (CSC) subpopulation (12).

Earlier studies inferred the presence of SGC CSCs and tested surface markers (CD44, CD133, and ABCG2) and CSC-related factors (SOX2, OCT4, and NANOG; refs. 13, 14); however, these did not prospectively define CSCs by sphere formation or tumor initiation (15). Aldefluor was the sole marker used to determine that the ALDH⁺ population was tumorigenic in ACC PDX models (16), and ALDH⁺CD44^high cells from MEC cells lines formed spheres and initiated tumors (10, 17). Using ACC PDX models and primary cultures, Panaccione and colleagues (18, 19) determined that CD133⁺ CSCs were maintained through SOX10 and Notch signaling and high expression of CD44.

Here, we report the successful engraftment of the first three known PDX models of a high grade MEC (derived from the same patient after successive relapses), as well as one acinic cell carcinoma (AciCC), and four ACC cases. We have characterized the somatic mutational landscape, confirmed the presence of previously reported gene fusion events, and identified a novel gene fusion. We determined that the ALDH⁺CD44^high phenotype defines the CSC fraction across ACC and MEC tumors using sphere formation and in vivo tumor initiation. The detailed analysis of MEC models generated from successive relapses identified an increase in mutation load, associated with increased expression of tumor promoting genes, and decreased transcription of tumor suppressor genes. Finally, in a landmark finding in solid tumor CSC biology, we found a time and relapse-dependent increase in both CSC frequency, as well as individual CSC sphere forming potential in vitro and tumorigenicity in vivo. The accumulation of CSCs over time contributing to disease progression or cancer relapse is considered a key principle in CSC biology, but this concept has been elusive and difficult to demonstrate in solid tumors.

The protocol for studies involving human subjects was approved by the Colorado Multiple Institutional Review Board (COMIRB #08-0552) in accordance with the Belmont Report and U.S. Common Rule. Informed written consent was obtained from all patients whose tissues were used for this study. The University of Colorado Institutional Animal Care and Use Committee (IACUC) approved all experiments involving mice. PDX generation and characterization was previously reported (20).

Processing of tumor tissue for sorting, analysis, and in vivo implantation of CSCs was previously reported (21).

Immunohistochemistry (IHC) analyses were performed as previously described (3). Primary antibodies and dilutions; 1:750 ALDH1A1 (#61195 BD Biosciences), 1:50 CK5 (#CM353C BioCare), 1:200 Cleaved caspase-3 (#9664 Cell Signaling Technology), 1:100 EGFR (#4267 Cell Signaling Technology), 1:100 phospho-EGFR (#3777 Cell Signaling Technology), 1:100 Ki67 (#RM-9106-S1 Thermo), and 1:2000 phospho-SMAD2 (#ab188334 Abcam).

Protein analyses were performed as previously described (21). Primary antibodies and dilutions; 1:2,000 Actin (pan; Cell Signaling Technology), 1:500 ALDH1A1 (Sigma-Aldrich), 1:750 p4EBP1 (Cell Signaling Technology), 1:1,000 SMAD2/3 (Cell Signaling Technology), 1:1,000 pSMAD2 (Cell Signaling Technology), 1:1,000 pSMAD3 (Cell Signaling Technology), 1:500 SOX2 (Cell Signaling Technology).

DNA was isolated from blood and tumor with a DNeasy Blood & Tissue kit (Qiagen), whereas total RNA was isolated from tumor tissue using an RNeasy kit (Qiagen). Isolated DNA and RNA were sent to the University of Colorado Cancer Center Genomics and Microarray Core, which performed library preparation, sequencing, and FASTQ generation. Single-end sequencing of mRNA (Poly A) was performed on an Illumina HiSEQ instrument with read lengths of 100bp (Illumina). DNA underwent paired-end sequencing using Agilent Technologies SureSelect Human All Exon Versions 5 and 6 with read lengths of 150bp (Agilent, Santa Clara).

DNA sequencing was conducted as previously described (22), and new primer sets were developed for the fusions listed below.

FISH analysis of gene fusions was previously described (6) and is further described in detail in the Supplementary Materials.

Experiments were compared with a two-group t test. Calculations were done using GraphPad Prism version 7.0 and SPSS version 11. Data are represented graphically as mean ±SEM.

Materials will be shared per the University of Colorado's Office for Technology Transfer policies and Institutional Review Board.

After securing patient informed consent, excess tumor tissue samples from 12 SGC surgeries were implanted subcutaneously on the flanks of athymic nude mice, including 5 ACC, 4 MEC, 1 salivary duct carcinoma (SDC), 1 AciCC, and 1 mammary analogue secretory carcinoma (MASC). Engraftment rates varied between histotypes resulting in PDX models of 3 MEC (75%), 3 ACC (60%), and 1 AciCC (100%; Supplementary Table S1; Supplementary Table S2). The engraftment rate of relapsed tumors was higher than that of primary cases and were 63% and 20%, respectively (Supplementary Table S1), whereas time to initial engraftment and subsequent passaging for MEC (3), ACC (3) and a single AciCC required on average 170 ± 46, 401 ± 7, and 343 days to generate tumors (>1,500mm³) respectively (Supplementary Fig. S1A). Morphology (hematoxylin and eosin; H&E staining) and cytokeratin 5 (CK5) staining were conserved in PDX tumors when compared with the originating patient tissue (Fig. 1A).

The most unique PDX were three models (CUSG006, −007, −012) generated from the same patient with a high-grade MEC from three successive surgeries of relapses occurring over 3 years (Fig. 1B). The time between surgeries was similar (∼9-10 months), but by the third operation the tumor had migrated outside of the primary parotid bed site, invading into the deep neck space and around the carotid artery. The growth rate of CUSG012 was nearly identical to CUSG006 and CUSG007 once stably passaged in mice (Supplementary Fig. S1B).

Whole-exome sequencing (WES) of nine tumors from 7 patients with paired normal samples determined the mutation burden for each case (Fig. 1C). Possibly damaging somatic mutations present in more than one sample are displayed in a co-mutation (coMut) plot (Fig. 1D), whereas all mutations are reported in Supplementary Table S3. We observed substantial overlap in the three MEC cases and an approximately 50% increased mutation burden in the two later samples, CUSG007 (62 mutations) and CUSG012 (58 mutations), compared with CUSG006 (41 mutations). We found that CUSG012 had acquired an inactivating mutation in the SHPRH gene. This gene has been shown to suppress genomic instability (23), which may allow for rapid accumulation of mutations and tumor progression. When filtered by known somatic cancer-related mutations (COSMIC), we observed few oncogenic mutations per tumor (0 mutations in CUSG002, CUSG003, CUSG006, CUSG007; 1 mutation in CUSG004, CUSG005, CUSG012; 2 mutations in CUSG013 [ATM, KDM6A]), except for the SDC case (CUSG008; 5 COSMIC mutations), which included missense mutations in ERBB2, CDH1, and BAP1 (Supplementary Fig. S1B). We also identified inactivating mutations in genes with putative roles as tumor suppressors and promoting genomic stability, including missense mutations of RB1CC1 in multiple histotypes (MEC, SDC, and MASC), whereas SHPRH mutations were found in two cases (MEC and SDC). We found that two ACC tumors (CUSG004, −005) had frameshift mutations in SPEN and KMT2D.

The analysis of SGC tumor RNA-seq data using the STAR-fusion pipeline identified multiple high-confidence gene fusion events in patient tumors. Because single-end sequencing was performed, junction read counts, or the number of reads that split alignment between two genes, were used as fusion evidence (Supplementary Table S4). We identified a novel fusion event, NFIB-MTFR2, in the hypomutated ACC model CUSG004. This molecular fusion event was confirmed by Sanger sequencing (Fig. 1E) and subsequently visualized by FISH using green 5'NFIB and the red 3'MTFR2 DNA probes in the CUSG004 patient tumor, but not in a matched normal tissue sample (Fig. 1F) or other controls (Supplementary Fig. S2; Supplementary Table S5). We identified several previously reported fusions, including NTRK3-ETV6 (CUSG002, MASC) and MYB-NFIB (CUSG005, ACC; Supplementary Table S4). Coinciding with its high mutation burden (Fig. 1C), CUSG008 (SDC) had numerous genomic rearrangements, including ADNP2-KCNG2, YTHDF1-NKAIN4, and LAMA5-FTCD (Supplementary Table S4). Finally, the NFIB–MTFR2 (CUSG004) and MYB–NFIB (CUSG005) fusions observed in patient tumor was conserved following engraftment on mice.

We first assessed unsupervised gene clustering and the drift in gene expression following PDX engraftment on mice and subsequent passaging by analyzing patient tissue and multiple PDX passages for three cases by RNA-seq. We found that the two ACC cases, CUSG004, −005, clustered more closely than the MEC case, CUSG006 (Supplementary Fig. S3A). We also found that 85% of genes that had significant changes in expression following engraftment on nude mice overlapped across all three cases (Supplementary Fig. S3B–S3E) and were related to immune response (Supplementary Fig. S4A–S4D). Numerous differentially expressed genes were identified when comparing ACC or MEC samples versus all others (Supplementary Table S6). Gene set enrichment analysis (GSEA) identified significant upregulation of the Hallmark Pathways, “MYC Targets,” “Mitotic Spindle,” and “E2F Targets” in ACC tumors (Fig. 2A).

GSEA comparing the PDX tumors engrafted from subsequent MEC relapses identified the upregulation of Hallmark pathways over time, including “E2F Targets,” “Myc Targets,” “DNA Repair,” and “TGF-beta” and downregulation of “Epithelial to Mesenchymal Transition” and “Inflammatory Response” (Fig. 2B). When comparing tissue collected from the first two surgeries (CUSG006, CUSG007), we observed progressively increased expression of growth promoting genes (CR1 [97-fold], MAGEC2 [21-fold], MMP1 [3.1-fold], and HEY1 [2.1-fold]). We next compared tissue from the second and third surgeries and found expression of genes related to migration (MT1E [1,445-fold]), survival (EN1 [6.9-fold]) and CSCs (LGR5 [28-fold], LEF1 [19-fold]) to be dramatically enriched in the relapsed third tumor (Fig. 2C). Just as striking, expression of key tumor suppressors (CDKN2B [-1,628-fold], TP53 [-2.3-fold], SIK1 [-1,709-fold]) was also inhibited in this same case.

Next, we assessed gene-expression changes throughout disease progression by examining protein levels in patient tissue by IHC. Tumor cell proliferation, Ki67 expression, did not change between cases CUSG006, −007, and −012 (Fig. 2D). However, we did observe incremental increases in pSMAD2 staining and ALDH1A1-high cells by IHC (Fig. 2D). Changes in gene expression and IHC were confirmed by western blot for key pathways, which were replicated three times with different PDX tumors. Levels of pSMAD2, pSMAD3, NOTCH1, HES1, SOX2, ALDH1A1, and MYC increased over disease progression in the three PDX cases, whereas EGFR signaling (EGFR, pEGFR, and pMAPK) decreased (Fig. 2E).

We determined Aldefluor activity and CD44 expression in two ACC and three MEC PDX models and found the ALDH⁺CD44^high population consistently fell between 0.15% and 4.5%, except for the ACC PDX model CUSG005 that had a dual positive population of 17% (Fig. 3A; Supplementary Table S7). Cells from PDX tumors were sorted by Aldefluor activity and CD44 expression and seeded as spheres (5 × 10³ cells/well ALDH⁺CD44^high, ALDH⁺CD44^low, ALDH⁻CD44^high or 5 × 10⁴ cells/well ALDH⁻CD44^low) in 96-well low-bind plates. The ALDH⁺CD44^high population generated the most tumor spheres for both ACC (CUSG004 P = 0.032, CUSG005 P < 0.001) and MEC (CUSG007 P < 0.001, CUSG012 P = 0.018) when sorted from PDX. The initial MEC tumor of three successive surgeries (CUSG006) was the sole exception, where no sorted subpopulation generated more than 3 spheres per well (Fig. 3B; Supplementary Fig. S5). Out of all subpopulations sorted across cases, the CUSG012 ALDH⁺CD44^high subpopulation consistently formed the largest spheres across all cases and histotypes (Fig. 3B).

The ALDH⁺CD44^high subpopulation from the three MEC relapses increased from 0.2% (CUSG006) to 0.3% (CUSG007) and then to 4.5% (CUSG012; ref. Fig. 3C; Supplementary Table S7). As noted above, subpopulations sorted from CUSG006 tumors had difficulty generating spheres, whereas CUSG007 ALDH⁺CD44^high cells readily formed spheres (Fig. 3C). Of note, all subpopulations sorted from CUSG012 tumors were able to generate spheres (Fig. 3C and D); however, the ALDH⁺CD44^high cell subpopulation had the greatest sphere-generating potential, forming the most and the largest spheres across all three cases (Fig. 3C). Comparing sorted ALDH⁺CD44^high and bulk tumor cells, we found that HES1, SOX2, and pSMAD2 were enriched in sphere-forming ALDH⁺CD44^high cells (Fig. 3D).

Tumor formation in mice using cell dilution assays remains the gold standard to identify CSCs (24, 25). All four tumor subpopulations (10² and 10³ cells for ALDH⁺CD44^high, ALDH⁺CD44^low, ALDH⁻CD44^high; 10⁵ cells for ALDH⁻CD44^low) and bulk tumor cells (10⁴ and 10⁵ cells) from one ACC (CUSG004) and three MEC (CUSG006, −007, −012) PDX were implanted into the flanks of nude mice and monitored for up to 500 days for tumor formation. The ALDH⁺CD44^high subpopulation was the most tumorigenic when ≤10³ cells were injected (Supplementary Table S8), but tumor initiation/growth took over a year when 10-fold fewer (≤10²) ALDH⁺CD44^high cells were implanted (Fig. 4A), suggesting that CSCs can remain senescent for long periods before actively proliferating. Importantly, 10³ ALDH⁺CD44^high cells were as tumorigenic as 10⁵ bulk tumor cells supporting that it is the approximately 1% CSC fraction within bulk cells that bears tumorigenicity; indeed, 10⁵ ALDH⁻CD44^low cells rarely formed tumors (Supplementary Table S8). Tumors resulting from CSC implantation harbored CSC fractions similar to the originating PDX (Supplementary Table S7), and subsequent passaging of tumor subpopulations sorted from CSC-generated tumors (CUSG004, CUSG007) gave similar results (Supplementary Table S8 and Fig. 4B).

No subpopulations sorted from CUSG006 tumors generated tumors whereas ALDH⁺CD44^high, and to a lesser extent ALDH⁺CD44^low cells, from CUSG007 and CUSG012 cells readily formed tumors in cell dilution studies with inoculates as low as 10² cells (Supplementary Table S8). It is important to note again that the ALDH⁺CD44^high fraction increased from 0.3% to 4.5% during the progression of CUSG007 to −012, possibly accounting for the subtly lower single marker (ALDH⁻CD44^high and ALDH⁺CD44^low) subpopulation tumorigenicity in CUSG012.

Orphan diseases like SGC are often understudied due to a scarcity of laboratory models such as cell lines or PDX for functional, mechanistic, and/or therapeutic experiments. Currently, research into the molecular drivers responsible for SGC initiation and progression is limited, and this is perhaps reflected in the lack of targeted therapies for this disease. There is also incomplete evidence that CSCs promote proliferation and relapse of SGC tumors across histotypes (16, 17). Therefore, establishing PDX models will greatly aid in the identification of key genomic and expression alterations in engrafted cases and in the definition of the CSC subpopulations across SGC histotypes. This will provide additional and valuable tools in determining the processes that lead to late recurrences and metastasis (9) and developing hypothesis-based novel therapeutics (26). In this report, we show the first successful establishment of high-grade MEC PDX models reported to date, the fact that these originated from the same individual at different disease times, and the observation of mutation accumulation and enhanced growth-promoting gene dysregulation over time. In addition, a novel NFIB--MTFR2 gene fusion was identified in an ACC case. Importantly, we demonstrate that both CSC frequency and individual CSC “aggressiveness” (sphere forming potential, tumorigenicity) increase in subsequent relapses.

The characterization of the genomic and expression changes of successive relapses of an increasingly aggressive tumor over time revealed that during the approximately 10 months between the first (CUSG006) and second (CUSG007) relapses there were dramatic increases in CR1 and MAGEC2 expression, which regulate the CSC compartment of colorectal tumors (27) and can promote metastases through STAT3 (28), respectively. In comparing CUSG007 and CUSG012, we found that CUSG012 had acquired an inactivating mutation in SHPRH, a gene that suppresses genomic instability (23), which could account for the accompanying genomic and gene-expression changes. Along with downregulation of tumor suppressors (CDKN2B, SIK1, TP53), we noted increased expression of genes associated with survival (EN1; ref. 29), Wnt signaling (LEF1), and colon stem cells and colorectal CSCs (LGR5). These gene-expression changes were accompanied by increased TGF-β signaling and expression of known protein effectors of the CSC phenotype (SOX2, MYC, ALDH1A1), suggesting a transition to a CSC-enriched state.

To formally characterize CSCs, we used both sphere formation and tumor initiation in mice using cell dilution assays, the in vitro and in vivo “gold standards” when prospectively defining CSCs (15). Similar to previous CSC studies in solid tumors (15) the ALDH⁺CD44^high fraction ranged between approximately 0.1% and 3% in four out of five cases. However, the ALDH⁺CD44^high population in the CUSG005 tumor was >15%, which may be due to the homogeneity of this MYB-NFIB fusion-driven tumor (30). The ALDH⁺CD44^high population consistently generated the most, and the largest, spheres compared with other subpopulations across ACC and MEC subtypes. As few as 10² ALDH⁺CD44^high cells initiated tumors in vivo, and resulting tumors recapitulated the morphology, tumor cell populations, and CSC fraction as that of the originating tumor. Possibly the most striking data arose from the inherently controlled assessment of CSC profiles and properties from PDX models arising from three subsequent relapses from the same subject. Increased mutation load and ensuing gene dysregulation ultimately led to both increased CSC fraction and individual CSC aggressiveness (sphere formation, tumorigenicity), in what we propose is a relevant observation highlighting the translational value of PDX models. This finding supports a core tenet in CSC biology. Similarly, it has been reported that CSC directly extracted from pre- and post-treatment samples of breast cancers showed increases in breast CSC numbers and sphere formation potential in response to cytotoxic therapy (31),emphasizing that originating patient samples are critical to study delicate biological processes such as the characterization of the CSC fraction.

The timing of CSC initiated tumor formation closely followed what is observed in the clinic, as 100 highly tumorigenic CSCs took nearly 250 days to give rise to measurable (∼100 mm³) tumors on mice, a size still several-fold smaller than what is clinically or radiologically detectable in patients. The ability of a few cells to survive undetected for many months in a hostile environment before proliferating are characteristics consistent with residual disease leading to late recurrence and metastases in SGC patients (9). The observation that PDX-derived CSCs can model the clinical behavior of SGC is critical to support the utilization of these models to study the process of metastasis and to identify relevant therapies.

We successfully engrafted PDX models for three MEC, three ACC, and one AciCC cases that generally preserved the originator tumor's morphology. The time required for initial SGC PDX engraftment following patient tumor implantation exceeded a year in some cases, which was much longer than what we have previously observed in our extensive experience with head and neck squamous cell carcinomas (HNSCC). HNSCC commonly grow within 3 to 4 months (22); notably, this follows the usual timing of HNSCCs relapses that commonly occur in the first year after diagnosis (32). The growth dynamic of SGC PDX is also consistent with the protracted natural history where SGC distant relapses decades after diagnosis are common (9, 10) and previous SGC PDX reports (7, 8). To our knowledge, we have generated the first functional PDX models of MEC, which retain the morphology of their respective histotypes. An interesting observation is that the majority of successfully engrafted cases were relapses, consistent with pancreatic and breast cancers (33, 34).

SGCs are hypomutated compared with other neoplasms (35). Through WES analysis, we determined that ACC tumors (CUSG004, CUSG005, CUSG013) had between 19 and 25 somatic mutations, which is consistent with a previous study that reported 2-35 mutations per ACC case (36). We did not observe high levels of TP53 mutations or alterations in PI3K signaling that were reported in studies using targeted NextGen Sequencing of 182-315 cancer genes (37–39). This may be due in part to the lack of sequenced paired normal tissue to remove germline SNPs/mutations in these other reports. However, we did observe mutations in cancer associated genes that have been reported using multiple platforms. We found that two of three ACC tumors harbored frameshift mutations in SPEN, a negative regulator of Notch signaling, which is mutated in up to 20% of ACC cases (36, 38). The third ACC case had mutations in both KDM6A and ATM that are mutated in approximately 15% and approximately 5% of ACCs respectively (38, 40). The BRCA1-Associated Protein 1 (BAP1) was mutated in the single SDC case (CUSG008) analyzed, which is a gene consistently mutated in 5% to 10% of SGCs across disease histotypes (37, 38, 40). We also observed inactivating mutations across histotypes in putative tumor-suppressors RB1CC1, RARRES1 and TP53BP1 (41–43), as well as in SHPRH, which suppresses genomic instability (23) by promoting error free replication (44). Loss of SHPRH function may account for rapid changes in gene expression and progression of the CUSG012 tumor. Also, two ACC cases with low mutation burdens shared damaging rearrangements in the tumor-suppressor KMT2D (45). Finally, we have confirmed the presence of common SGC genomic alterations and novel mutations that may allow for tumorigenesis and disease progression.

Given this low frequency of oncogenic mutations, we tested gene fusions known to occur in several SGC histotypes, including ETV6-NTRK3 fusions in MASCs and MYB–NFIB fusions in ACCs (46, 47), and these genetic events were identified in cases CUSG002 (MASC) and −005 (ACC), respectively. We identified a novel gene fusion between the NFIB and MTFR2 genes in an ACC PDX model (CUSG004). This is likely a driver event because the MYB domain of the MYB–NFIB fusion promotes carcinogenesis through its role as a transcription factor (30). Notably, NFIB has been identified as an oncogene in small-cell lung cancer where it appears to regulate cell proliferation and viability through transcriptional regulation (48). Furthermore, the NFIB–MTFR2 fusion results in the loss of the transcriptional regulatory domain of NFIB and leads to a truncated NFIB protein. Further studies examining the consequences of this truncation are merited as NFIB is thought to be important in the regulation of the TGF-beta and Sonic Hedgehog pathways (49).

In summary, here we report detailed genomic and CSC characterization from a comprehensive panel of SGC PDX, which includes three sequential models from the same MEC patient that have enabled us to dissect the molecular evolution and functional changes in CSCs over time. This is the first report of successful MEC PDX generation. We observed a doubling in the mutation burden in subsequent relapses and found an enrichment of pro-survival gene expression and “stemness.” The size of the CSC subpopulation increased over time, and in addition the CSCs had a significant rise in sphere initiation capacity and in vivo tumorigenicity. Overall these findings suggest that CSCs may be associated with tumor progression in MEC, as well as highlighting the importance of SGC PDX as a valid tool in expanding our understanding of the molecular progression of cancer and the impact of disease progression on tumorigenic cell subpopulations, such as CSCs.

No potential conflicts of interest were disclosed.

Conception and design: S.B. Keysar, M.E. Reyland, A. Jimeno

Development of methodology: S.B. Keysar, J.R. Eagles, A.-C. Tan, A. Jimeno

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.B. Keysar, J.R. Eagles, B. Miller, F.N. Chowdhury, J. Reisinger, T.-S. Chimed, J.J. Morton, M. Varella-Garcia, J.I. Song, D.W. Bowles, A. Jimeno

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.B. Keysar, J.R. Eagles, B. Miller, B.C. Jackson, H.L. Somerset, M. Varella-Garcia, A.-C. Tan, D.W. Bowles, A. Jimeno

Writing, review, and/or revision of the manuscript: S.B. Keysar, J.R. Eagles, B. Miller, F.N. Chowdhury, J. Reisinger, P.N. Le, J.J. Morton, H.L. Somerset, M. Varella-Garcia, J.I. Song, D.W. Bowles, M.E. Reyland, A. Jimeno

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Jimeno

Study supervision: A. Jimeno

This work was supported by NIH grants R01-CA149456 (to A. Jimeno), R21-DE019712 (to A. Jimeno), R01-DE024371 (to A. Jimeno), P30-CA046934 (University of Colorado Cancer Center Support Grant), R01-DE015648 (to M.E. Reyland), R56-DE023245 (to M.E. Reyland and A. Jimeno), the Adenoid Cystic Carcinoma Research Foundation (to A. Jimeno and M.E. Reyland), University of Colorado Adenoid Cystic Cancer Research Fund (to D.W. Bowles), the Daniel and Janet Mordecai Foundation (to A. Jimeno), and the Peter and Rhonda Grant Foundation (to A. Jimeno). The authors wish to thank the patients who donated their tissue, blood and time, and to the clinical teams who facilitated patient informed consent, as well as sample and data acquisition.

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.