Molecular evidence for olfactory neuroblastoma as a tumor of malignant globose basal cells

Abstract

Olfactory neuroblastoma (ONB, esthesioneuroblastoma) is a sinonasal cancer with an underdeveloped diagnostic toolkit and is the subject of many incidents of tumor misclassification throughout the literature. Despite its name, connections between the cancer and cells of the normal olfactory epithelium have not been systematically explored and markers of olfactory epithelial cell types are not deployed in clinical practice. Herein we utilize an integrated human-mouse single cell atlas of the nasal mucosa including the olfactory epithelium to identify transcriptomic programs that link olfactory neuroblastoma to a specific population of stem/progenitor cells known as olfactory epithelial globose basal cells (GBCs). Expression of a GBC transcription factor NEUROD1 distinguishes both low- and high-grade ONB from sinonasal undifferentiated carcinoma (SNUC), a potential histologic mimic with a distinctly unfavorable prognosis. Furthermore, we identify a reproducible subpopulation of highly proliferative ONB cells expressing the GBC stemness marker EZH2 in situ, suggesting EZH2 inhibition may have a role in the targeted treatment of ONB. Finally, we study the cellular states comprising ONB parenchyma using single cell transcriptomics and identify evidence of a divergent differentiation pathway recapitulating aspects of a neuronal versus sustentacular transcriptional regulatory circuit that governs the fate of normal GBCs. These results link ONB to a specific cell type for the first time and identify conserved developmental pathways within ONB that inform diagnostic, prognostic, and mechanistic investigation.

INTRODUCTION

Olfactory neuroblastoma (ONB, also known as esthesioneuroblastoma) is an uncommon sinonasal cancer attributed to malignant transformation of olfactory epithelial tissue (1–4). Classically, the diagnosis is made with immunohistochemical detection of neuroendocrine markers such as CD56/NCAM1, neuron-specific enolase (NSE/ENO4), synaptophysin (SYP), chromogranin A (CHGA), INSM1, and calretinin (CALB2) in the absence of pan-cytokeratin (5–12). This combination can be useful for distinguishing it from many other sinonasal cancers, however these markers become far less reliable (i.e., fail to remain mutually exclusive) in the setting of poorly differentiated, high-grade tumors such as sinonasal undifferentiated carcinoma (SNUC). Further complicating the diagnostic landscape are a population of highly aggressive tumors that display overlapping characteristics with ONB but are distinguished by focal regions of epithelial differentiation. Historically confined to case reports, these have only recently been established in a consolidated fashion under the name “olfactory carcinoma” (13). Their relationship to canonical ONB is unknown, however they appear to represent a more aggressive state of disease.

The overall extent of differentiation within the ONB tumor is useful for ascribing prognosis and is qualitatively assessed by the Hyams histopathologic grading system, which takes into account histoarchitecture, mitoses, nuclear pleomorphism, neurofibrillary matrix, rosette formation, and necrosis within the tumor (14). Well-differentiated ONB (grades I and II) display lobular architecture and neuronal morphology (e.g., Homer-Wright pseudorosettes, fibrillary matrix) whereas poorly differentiated ONB (grades III and IV) generally lack neural morphology and display classic features of aggressive neoplasms such as abundant mitoses, pronounced nuclear pleomorphism, and necrosis. Despite an initially limited sample size, the prognostic utility of the Hyams grade was ultimately established and re-affirmed in recent times by a meta-analysis which identified worsening overall survival with ordinal increases in Hyams grade across 33 studies and 493 patients (15–17).

Unsurprisingly, the diagnostic ambiguity surrounding high-grade ONB and the possibility of ONB variants have confounded interpretations of translational research on the tumor. Multiple independent transcriptomic and epigenomic analyses have identified heterogeneity within cohorts of tumors classified as ONB, however there is controversy regarding whether the findings describe tumor subtypes or are rather the consequence of differences in pre-analytic classifications and inclusion criteria. For example, two transcriptomic studies of ONB heterogeneity identify a subset of “basal-type” ONB expressing epithelial markers (i.e. pancytokeratin), decreased neuroendocrine marker expression, and stem-like transcriptomic signatures (18,19). These tumors also tended to harbor IDH2 hotspot mutations (18). This constellation of immunohistochemical and genetic features has been previously attributed to SNUC (20,21), thus raising questions regarding the molecular boundaries between the two. Similarly an epigenomic study comparing ONB to other sinonasal cancers found that a subset of tumors originally diagnosed as ONB had distinct methylomic profiles resembling undifferentiated carcinomas, adenocarcinomas, squamous cell carcinomas, melanomas, and IDH2-mutant carcinomas (22). In this setting misdiagnosis was considered the most likely source of ONB heterogeneity, rather than ONB subtypes. Fundamentally, interpretations of these studies depend on whether pan-cytokeratin immunostaining undermines a diagnosis of ONB.

At present, a barrier to the accurate diagnosis of ONB (and by extension, contextualization of proposed ONB variants) is the limitation of confirmatory tumor markers. Despite longstanding assumptions of a connection between ONB and the cells of the olfactory epithelium (OE) since its discovery, the standard-of-care diagnostics for this tumor are general neuroendocrine markers which do not directly reflect any particular OE cell type. This is in stark contrast with the state of diagnostics for more common cancers, such as hematopoietic, breast, lung, and gastrointestinal tumors, which are identified with assistance from immunohistochemical markers reflect their precise molecular phenotype or lineage of the analogous normal cell or tissue. To date, the normal cell type that most closely relates to ONB at a molecular level has not been identified. Speculations in the literature invoke vague “progenitor cells” or “basal cells”, but a systematic and thorough molecular comparison of ONB to the normal cells of the OE has yet to occur. To this end, we performed comparative immunohistochemical and transcriptomic analyses of ONB, normal human olfactory mucosal (OM) tissue, and the discrete cell types that reside within the normal mammalian OE. Our results establish a foundational diagnostic reference frame for ONB based on the molecular identity of the normal cell type it most closely relates to.

MATERIALS AND METHODS

Diagnosis of sinonasal tumors

All diagnoses were made by subspecialty head and neck pathologists at the Massachusetts General Hospital and Massachusetts Eye and Ear according to World Health Organization criteria (23). Morphologic characteristics consistent with ONB included cells with round-to-ovoid nuclei containing granular “salt-and-pepper” chromatin, with cells arranged in lobular groups as well as sheets and nests. Depending upon Hyams grade, variable amounts of neurofibrillary stromal elements and rosettes (Homer Wright and/or Flexner-Wintersteiner) were identified supporting the diagnosis. Immunohistochemical profiling for ONB required positivity for neuroendocrine markers, such as chromogranin and synaptophysin, in the absence of immunostaining for CK7, CK20, pan-cytokeratin, p63, and HMB-45. Grading was performed according to Hyams criteria (9,16,17,24,25). Sinonasal undifferentiated carcinoma (SNUC) was called following identification of high-grade carcinoma in the absence of epithelial and glandular differentiation. Morphologic features consistent with the diagnosis included pleiomorphic hyperchromatic or vesicular nuclei with prominent nucleoli forming sheets, lobules, nests, and/or trabeculae. Mitoses were abundant. The immunohistochemical profile of SNUC is positive for cytokeratins (either pan-cytokeratin or AE1/AE3) in the setting of negativity for p63, p40, KRT5, KRT14, CD45, Desmin, S100, and SOX10. Staining for neuroendocrine markers was either absent, focal, or nonspecific. The status of SMARCB1 (INI1), SMARCA4 (BRG1), and IDH1/2 in our archival SNUC samples is unknown.

Olfactory neuroblastoma biopsies

Portions of biopsy-verified olfactory neuroblastoma tissue were obtained from patients during definitive surgical resection of the tumor prior to adjuvant radiation treatment or chemotherapy. The specimens were quickly rinsed in normal saline, divided in half with one portion put into sterile Dulbecco’s Modified Eagle Medium (DMEM, ThermoFisher #11330032) with Hepes Buffer (Gibco/Invitrogen, Carlsbad, CA) on ice and the other portion snap-frozen in a dry centrifuge tube placed on dry ice before further processing. Pieces put in DMEM were fixed in fresh 4% paraformaldehyde in phosphate buffered saline (PBS) (pH 8.0) for 6 days at 4°C and then transferred to fresh PBS for processing in preparation for cryosectioning. Pieces frozen on dry ice were stored at −80°C for later molecular studies.

Human olfactory mucosal biopsies

Specimens of normal human olfactory mucosa (OM) for molecular analysis were obtained from patients with documented olfactory function within the normal range based on smell testing (Smell Identification Test; Sensonics International; Haddon Heights, NJ). Specimens were obtained under endoscopic visualization and local anesthetic using a sickle-knife and forceps (26). Samples were put in DMEM or snap-frozen in a dry centrifuge tube and transported for further processing as above. Biopsies were confirmed to contain OE by immunohistochemical staining for known markers, including anti-olfactory marker protein (OMP).

Bulk transcriptomics

Flash-frozen ONB and human OM biopsies were processed at the same time. RNA was extracted using the Zymo Research Direct-zol RNA Microprep Kit (Catalog #R2061) and library prep were performed using the Illumina TruSeq Stranded mRNA Library Prep Kit (#20020594) following manufacturer’s instructions. Sequencing was performed on the Illumina HiSeq2500 platform using the PE-125 protocol to a depth of 15–20M reads per sample. Alignment to the human genome was performed using STAR (27) and the CRCh37/hg19 assembly. Library normalization and differential expression testing was performed in R using DESeq2 (v1.22.2) (28). Genes were considered differentially expressed if the Benjamini-Hochberg-corrected p < 0.05 with Log2 fold-change > 1. Gene ontology enrichment analysis was performed using Metascape (29).

Cryosectioning, immunofluorescence, and microscopy

Mucosal biopsies and tumor specimens were cryoprotected in 30% sucrose in PBS overnight and snap-frozen with liquid nitrogen in OCT compound. Specimens were cut at 9–12 μm thickness using a Leica CM3050 S cryostat. Tissue sections were collected on Superfrost Plus slides (Fisher Scientific, Catalog #12–550-15) and stored at −20°C until future use. For staining, slides were heated on a plate warmer at 50°C for 20 minutes to promote adherence of the tissue. OCT was removed in PBS for 5 minutes, followed by treatment of the tissue in 3% H₂O₂ in PBS or methanol to quench endogenous peroxidase. Antigen retrieval was performed by immersing tissue in 0.01 M sodium citrate (pH 6.0) in a commercial food steamer for 10 minutes. Slides were incubated with primary antibody in blocking solution (10% v/v normal donkey serum, 5% w/v nonfat dry milk, 4% w/v bovine serum albumin, 0.1% v/v TritonX-100) for 1 hour at 25°C or overnight at 4°C. Secondary antibodies were applied at a concentration of 1:150 in blocking solution and incubations were performed for 1 hour at 25° C. Fluorophore-conjugated streptavidin was applied at 1:150 in biotin-free block for 1 hour at 25° C for tertiary amplification. Tyramide signal amplification (TSA) was performed using a biotinylated secondary at 1:150, HRP-conjugated streptavidin at 1:400 in biotin-free block, and PerkinElmer TSA Plus Cyanine 3 System (Product #NEL744001KT) following manufacturer’s instructions.

Fluorescence microscopy was performed on a Zeiss LSM800 confocal microscope set to one Airy unit in multi-track mode and acquired using Zeiss ZEN microscopy software. Images were processed using Fiji software to set color palette and adjust balance/contrast for optimal signalto-noise representation relative to no-primary staining controls. Adjustments were applied to the entire image and no images were cropped. Table 1 contains a complete description of reagents used in the immunofluorescent profiling of ONB.

Table 1.

Reagents and conditions for immunostaining

Primary antibodies used in clinical immunohistochemistry
Target	Source	Species	Antigen Retrieval	Dilution	Amplification
NEUROD1	R&D Systems #AF2746	Goat polyclonal	EDTA buffer (pH 9.0)	1 to 100	SA-HRP, DAB
EZH2	Cell Signaling Technology #5246	Rabbit monoclonal	Citrate buffer (pH 6.0)	1 to 50	CellSignaling #8114 SignalStain Boost, DAB
KI67	Leica #PA0230 Clone MM1	Mouse monoclonal	Leica #AR9640	Ready To Use	SA-HRP, DAB
Primary antibodies used in CellProfiler analysis
Target	Source	Species	Antigen Retrieval	Dilution	Amplification
NEUROD1 (pos. control)	R&D Systems #AF2746	Goat polyclonal	MeOH, Citrate buffer (pH 6.0)	1 to 100	Tyramide Signal Amplification
TP63 (neg. control)	ATCC clone #4A4	Mouse monoclonal	MeOH, Citrate buffer (pH 6.0)	1 to 100	Tyramide Signal Amplification
EZH2	Cell Signaling Technology #5246	Rabbit monoclonal	MeOH, Citrate buffer (pH 6.0)	1 to 50	Tyramide Signal Amplification
KI67	BD Pharmingen #556003	Mouse monoclonal	MeOH, Citrate buffer (pH 6.0)	1 to 100	Secondary

Case selection and clinical immunohistochemistry

Cases were selected from the surgical pathology archives of the Massachusetts General Hospital, Boston, MA, USA, following approval by the Institutional Review Board, for the years 2006–2017. Representative hematoxylin and eosin-stained slides were reviewed for each case to confirm the clinical diagnoses. Whole-tissue sections were immunostained with 3,3-diaminobenzidine tetrahydrochloride (Table 1) and counterstained with hematoxylin. Normal pancreatic tissue was used as a positive control for NEUROD1; lymphoma tissue was used as a positive control for EZH2. Internal negative controls were used for each slide. Cases were evaluated semi-quantitatively for immunoreactivity by two independent pathologists (ASF, WCF). Extent and quality of immunoreactivity was assessed only in cells where chromogens localized to the nucleus. Table 1 contains a complete description of reagents used in the immunohistochemical profiling of ONB.

Dissociation of murine olfactory mucosal tissue

All animal experimentation was carried out at Tufts University School of Medicine and was approved by the Tufts Institutional Animal Care and Use Committee. The experiment was performed on six 12-week-old CD-1 mice (Charles River, Strain 022), comprised equally of males and females maintained under a 12-hour light-dark cycle with ad libitum access to food and water. Mice were anesthetized with ketamine/xylazine/acepromazine cocktail and euthanized via exsanguination. Olfactory mucosa was separated from contiguous respiratory mucosa and underlying structures and placed in cold sterile HBSS. Tissue was minced into ~1 mm² fragments and resuspended in fresh HBSS containing 250 U/mL hyaluronidase (Worthington #LS002594), 100 U/mL collagenase (Worthington #LS004196), and 75 U/mL DNase I (Worthington #LS002139) for 10 minutes at 37° C to release submucosal cells from the basal surface. Fragments of the OE were pelleted with 80 x g for 1 minute at 4°C and the supernatant was discarded. OE tissue was resuspended in HBSS containing 0.05% trypsin-EDTA (Thermo Fisher Scientific #15400054) and digested for 5 minutes at 37° C. Trypsin inhibitor (Worthington #LS003570) was added to stop digestion along with additional DNase I, and tissue/cells were pelleted at 180 x g for 5 minutes at 4° C. Cells and tissue were further digested by resuspending in HBSS containing DNase I and Dispase II (Stem Cell Technologies #07913) for 20 minutes at 37°C to release cells from the basal lamina. The suspension was passed through a 250-micron mesh to remove large fragments, pelleted at 180 x g for 5 minutes at 4°C, and resuspended in PBS containing 1% BSA for surface marker staining.

Surface marker staining of primary murine olfactory mucosal cells

OSNs and Sus cells were depleted by sorting against anti-CD56 (NCAM1) and monoclonal antibody SUS4 (30) respectively. Cells were stained in PBS containing 1% BSA and SUS4 primary antibody at 1:100 for 15 minutes on ice, then washed by adding three parts PBS containing 1% BSA and pelleted at 300 x g for 3 minutes at 4° C. Cells were resuspended in PBS containing 1% BSA, APC-conjugated anti-Mouse secondary at 1:200, and APC-conjugated anti-NCAM1 (R&D Systems #FAB7820A) at 1:100 for 15 minutes on ice. Cells were again washed with the addition of three parts PBS containing 1% BSA, pelleted at 300 x g for 3 minutes at 4°C, resuspended in fresh PBS containing 1% BSA, and filtered through a 40-micron Flowmi Cell Strainer (#BAH136800040) into a clean tube for sorting. Fluorescence gating was set using unstained controls.

Single cell RNAseq of the murine olfactory mucosa

After recovery from sorting, the cell suspension was divided amongst four lanes of the 10X Single Cell Gene Expression Solution (#1000092) per manufacturer’s instructions. Libraries were combined in equimolar ratios and sequenced on the Illumina HiSeq4000 platform with PE-150 protocol to an average read depth of >150,000 reads per cell (~93% sequencing saturation). Sequences were aligned, counted, filtered, library depth normalized, and aggregated with 10X CellRanger.

Integration of human and mouse scRNAseq

Single cell transcriptomic data from the human nasal mucosa was obtained from the literature (31). For creation of the cell atlas, both respiratory (n = 2) and olfactory (n = 2) mucosal data sets were included. For trajectory analysis of GBCs and iOSNs, only human data sets annotated as originating from olfactory mucosa were included. All data sets were filtered for a maximum mitochondrial gene expression of 10% and minimum number of features greater than 100. Human cells were discarded if total counts were fewer than 400 or larger than 8000. Murine cells were discarded if total counts were fewer than 500 or larger than 5000. Doublets were then removed using Scrublet (32). Human-mouse gene homologs were then identified using the Mouse Genome Database, Mouse Genome Informatics homology resources (http://www.informatics.jax.org/homology.shtml) (33). Data sets were then filtered to only include genes for which homologs have been identified, and features across all data sets were renamed as a human-mouse chimeric pseudogene. Pre-processed Seurat objects were then integrated using the reciprocal PCA method as in the Seurat vignette with the following parameter modifications: FindIntegrationAnchors with k.anchor = 20; FindVariableFeatures with selection.method = “mvp”, num.bin = 30, and nfeatures 5000; RunUMAP with dims = 1:25. Projection of tumor cells onto the nasal mucosal cell atlas was accomplished using the MapQuery function of Seurat onto UMAPs generated using either the top 5000 variable features or using only transcription factors (34). Pseudotemporal trajectory inference and module detection were performed with Monocle3 (35,36).

CellProfiler image analysis

Three sections were randomly selected from each tumor then independently stained (Table 1), imaged, and quantitatively analyzed. Tissue was deemed adequate quality for analysis if adjacent sections labeled positively with NEUROD1 (positive control) and failed to label for TP63 (negative control). Five fields per tissue section were imaged after being chosen by eye on the basis of EZH2 fluorescence diversity. EZH2 was labeled with FITC to facilitate detection by eye whereas KI67 was labeled with AlexaFluor 647 to blind investigators to proliferation status while choosing fields. Antibody cross-reactivity and fluorophore bleed through were assessed using no-primary staining controls on adjacent tissue sections; neither was observed. Acquisition parameters were not standardized between tumor samples due to incompatible differences in staining intensity. Image pre-processing was performed in FIJI: z-stacks were converted to maximum intensity projections to level tissue sections traversing the image plane, brightness was adjusted to comparable ranges across all images, and background fluorescence was subtracted using the rolling-ball algorithm. Replicate image sets were then supplied to CellProfiler. Nuclei were detected based on DAPI labeling using default primary object identification parameters. Objects (nuclei) were discarded if measuring less than 10 pixels in diameter, greater than 30 pixels in diameter, or contacting the border of the image. Objects passing these filters were then eroded by one pixel to separate object clumps prior to fluorescence quantification. Decile bins for EZH2 fluorescence were calculated independently for each replicate. KI67 fluorescence distributions were median centered between replicates on a per-tumor basis and the KI67 positivity threshold was applied uniformly across all tumors and replicates.

Single cell transcriptomics of olfactory neuroblastoma

The biopsy specimen of treatment-naive ONB was obtained via endoscopic resection and transported in DMEM on ice. Tissue was rinsed in sterile, room temperature PBS before transferring to DMEM warmed to 37 degrees. Tissue was mechanically dissociated via mincing before resuspension in fresh DMEM containing Dispase (Stem Cell Technologies #07913). Dispase digestion was performed at 37 degrees with gentle, continuous inversion for 30 minutes. Tissue and cells were then pelleted at 300 x g for 5 minutes at room temperature, then resuspended in Accutase (ThermoFisher #A1110501) with added DNAse (Worthington #LS002139) for one hour. Progression of Accutase dissociation and cell viability was tested at 15 minute intervals with automated cell counting and Trypan blue staining. Tissue and cells were again pelleted at 300 x g for 5 minutes at room temperature and resuspended in DMEM containing 0.05% trypsin-EDTA (Thermo Fisher Scientific #15400054) for 5 minutes with gentle inversion. Trypsin inhibitor (Worthington #LS003570) and DNAse were added after 5 minutes. Final cell yield was 1.01 × 10^7 cells/mL with 42% viability and 2 mL of cell suspension. Cells were then resuspended in 5 mL of 10% DMSO in DMEM and divided into five 1 mL aliquots. For cryopreservation, aliquots were stored at −80C in an isopropanol cryocylinder pre-chilled to 4 degrees. For single cell transcriptomics, cell suspensions were rapidly warmed to 37 degrees and diluted 20X in sterile PBS. Cells were then pelleted and resuspended in 1 mL sterile PBS and passed through a 40 micron Flowmi filter tip. Cell viability after reconstitution was approximately 40%. Single cell transriptomics was then performed using the 10X Single Cell Gene Expression Solution (#1000092) following manufacturer’s protocol. Sequencing was performed to a goal library depth of 50,000 reads per cell.

Statistical analyses

All statistical analyses were performed with R version 4.0.2. A two-sided Welch’s t-test was used for comparing percentages of KI67-positivity between cells in the upper 10^th and lower 90^th percentiles of EZH2 expression. An alpha of 0.05 was considered significant for all hypothesis testing including gene set enrichment, gene ontology enrichment, and differential gene expression analyses. Exact p-values calculated to be less than 0.001 are reported as p < 0.001. Strategies for multiple test correction vary according to method and are delineated in their respective methods sections. Bar plot height equals the mean with error bars displaying standard deviation. Asterisk legend for figures: * indicates p < 0.05, ** indicates p < 0.01, *** indicates P< 0.001.

RESULTS

Olfactory neuroblastoma is a transcriptomic analog of neuronal progenitor GBCs

Our primary research objective was to analyze the molecular phenotype of ONB with respect to normal nasal cell types. To begin, biopsy samples of both ONB and adult human OM were obtained via endoscopic surgical resection (n=3 patient samples for each tissue type) and processed for bulk transcriptomics. Given that our tumor samples were all of the low-grade variety, these samples were then merged with fifteen additional ONB bulk seq from published work (18). We then performed differential gene expression analysis between malignant and normal tissue to define gene sets representative of each, referred to herafter as the “ONB signature” and “OM signature”. Whereas the OM signature was enriched in genes with functions related to neurite extension, transmembrane transport, and cilium biogenesis suggesting a mature neuronal phenotype, analysis showed the ONB signature was greatly enriched in neurodevelopmental gene ontology annotations suggesting a similarity between the overall transcriptomic identity of the tumor and that of neuronal progenitor cells (Figure 1A).

Figure 1. Transcriptomic characterization of ONB vis-à-vis the normal human olfactory mucosa and stages of olfactory epithelial neurogenesis.

A Volcano plot of differential gene expression between ONB and human OM assessed via bulk tissue RNAseq. Cutoffs for Log2 fold change and statistical significance are indicated with solid black lines. To each side, gene ontology networks for the indicated sets of differentially expressed genes. Select terms are annotated. Node size is proportional to the number of genes annotated with each term. Edges indicate the presence of shared genes between terms. Node color indicates communities of similar GO annotations (29). B Top; integrated UMAP embedding of msGBCs, npGBCs, and iOSNs from human and mouse single cell transcriptomic data sets. Number of cells per cell type and species of origin is reported. Lower left; cell clusters correspond to known cell types according to established marker expression. Lower right; pseudotime trajectory and ordering of cells. C Top; expression of two gene expression modules that vary as a function of pseudotime. Rather than coinciding with discrete clusters and annotated cell types, these signatures are maximal at transition points between two cell types along the trajectory; thus, they are regarded as differentiation signatures. The number of genes that comprise each module is indicated. Bottom; gene set enrichment testing for the above-mentioned differentiation modules in ONB versus the human olfactory mucosa.

We then turned our attention to the population of resident neuronal progenitor cells in the olfactory epithelium (OE). Neurogenesis in the OE occurs throughout the mammalian lifespan due to a heterogeneous population of stem and progenitor cells known as globose basal cells (GBCs). During neuronal differentiation, these cells pass through a series of developmental stages that gradually restrict their potency from a multipotent/stem state (the msGBC) to that of a unipotent, committed neuronal progenitor GBC (the npGBC) (37–45). To benchmark the differentiation status of ONB relative to these stages, we turned to single cell transcriptomics. Single cell analyses of the human olfactory epithelium have been previously described, confirming homologous GBC subpopulations in the adult human as in the mouse (31,46,47). In human samples, however, the population of msGBCs, npGBCs, and OSNs (collectively the “olfactory sensory neuron lineage”) constitute a minor fraction of the recovered cells. By integrating previously published human data (31) with a GBC-enriched single cell data set from the murine olfactory epithelium generated by our lab (see Methods), we identified a continuous differentiation trajectory from msGBCs expressing EZH2 and SOX2, to npGBCs expressing NEUROD1, and finally to terminally differentiated immature olfactory sensory neurons (iOSNs) expressing GAP43 (Figure 1B).

We then utilized this integrated data set to define the developmental programs occurring at the transition points between cell states, and tested whether these signatures could be detected in ONB using gene set enrichment analysis (35,36,48,49). We identified two transcriptomic modules that represented two major differentiation events: the transition from msGBC to npGBC (the “GBC differentiation” module), and the transition from npGBC to terminally differentiated OSN (the “OSN differentiation” module). While the GBC differentiation module was statistically enriched in ONB relative to the normal human OM, the OSN differentiation module failed to reach statistical significance (Figure 1C). The lack of concordant expression between two sequential differentiation modules suggested that the differentiation state of ONB may localize between these developmental stages, i.e., in an npGBC-like state.

We next screened for transcriptomic similarity between ONB tumor cells and all cell types present in the mammalian olfactory and respiratory mucosa. To accomplish this, we integrated our GBC-enriched murine data set with the four human data sets (two respiratory mucosae, two olfactory) from Durante et al. (31) to define a referential gene expression space that ONB tumor data could be mapped onto. Investigation of established markers confirmed the presence of many cell types within the atlas: msGBCs, npGBCs, immature OSNs, and mature OSNs; non-neuronal olfactory epithelial cell types such as ionocytes, microvillar cells, sustentacular cells, duct/gland cells, and horizontal basal cells; respiratory epithelial cells including basal, ciliated, and secretory types; submucosal/stromal cells including fibroblasts, pericytes, vascular smooth myocytes, and olfactory ensheathing cells; and leukocytes including both the myeloid and lymphoid lineages (Supplementary Figure 1). In total the atlas contained 51,005 nasal mucosal cells, with 28,830 human cells (31) and 22,175 mouse cells from our GBC-enriched data set.

Next, we obtained a treatment-naive endoscopic biopsy sample of ONB and performed droplet-based single call transcriptomics. Analysis showed the tumor tissue comprised three broad categories of cells: CHGA-positive ONB parenchymal cells, a small PTPRC-positive leukocyte population, and rare VIM-positive stromal cells (Supplementary Figure 2A). When projected into the gene expression space defined by normal respiratory and olfactory mucosal cells (see Methods), we observed that ONB cancer cells almost exclusively localized to the regions defined by GBCs and iOSNs (Figure 2, top). As expected, the projection also revealed a predominance of myeloid over lymphoid cancer-associated leukocytes and heterogeneous stromal cells within the tumor sample. To assess robustness we re-performed the embedding using solely transcription factor expression (34), and found that ONB cells once again projected almost exclusively into the region of GBCs and iOSNs (Figure 2, bottom). Collectively, these transcriptomic data suggest a greater transcriptomic similarity between ONB and npGBCs than any other nasal mucosal cell type.

Figure 2. Projection of ONB single cells onto the nasal mucosal cell atlas.

Top left; uniform manifold approximation projection (UMAP) of 51,005 nasal mucosal cells generated via the standard analytic pipeline. The integrated data set contains 28,830 human cells from Durante et al. (31) and 22,175 mouse cells. Cell type annotations were assigned to clusters according to marker gene expression and tissue of origin. Labels are as follows: msGBC, multipotent stem GBC; npGBC, neuronal progenitor GBC; iOSN, immature olfactory sensory neuron; mOSN, mature olfactory sensory neuron; Olf.HBC, olfactory horizontal basal cell; Sus, sustentacular cell; Olf.DG, olfactory epithelial duct/gland cell; Resp.BC, respiratory basal cell; Cili.REcyte, ciliated respiratory epithelial cell; Secr.REcyte, secretory respiratory epithelial cell; Resp.DG, respiratory epithelial duct/gland cell; Lymph.WBC, lymphoid leukocyte; Myelo.WBC, myeloid leukocyte; OEC, olfactory ensheathing cell. Bottom left; UMAP of nasal mucosal cells generated by limiting the principal component analysis step to transcription factors. Right column; projection of single ONB tumor cells into the gene expression space defined by nasal mucosal cells. ONB cells, leukocytes, and stromal cells are color-coded and labeled according to analysis of the ONB single cell data set (see Supplementary Figure 2A). O; ONB cells. S; stromal cells. L; leukocytes. Axis numbering indicates UMAP coordinates. Dashed contours surround identical regions.

The npGBC marker NEUROD1 differentiates olfactory neuroblastoma from sinonasal undifferentiated carcinoma

ONB has been historically difficult to distinguish from sinonasal undifferentiated carcinoma (SNUC) based on histology alone, however SNUC represents a much more aggressive tumor and thus their discrimination is crucial for clinical risk stratification. Given the ambiguities unresolved from standard diagnostic workflows, we pursued clinical validation of our transcriptomic findings to determine whether the definition of ONB as a tumor of npGBC-like cells held any diagnostic utility in distinguishing ONB from SNUC. Treatment-naïve biopsy samples from sixteen patients with ONB and five patients with SNUC were subject to immunohistochemical detection of the npGBC transcription factor NEUROD1. Slides were then blinded and scored by board-certified subspecialty head and neck pathologists, who determined the percentage of positive tumor cells by quantifying strong nuclear NEUROD1 immunoreactivity. Results showed a clear and statistically significant pattern of distinction between ONB and SNUC tumors, namely, that ONB tissue contained NEUROD1(+) nuclei in varying proportion whereas nuclear NEUROD1 staining was not observed in SNUC tissue (Figure 3A). Despite differences in the homogeneity of NEUROD1 staining across ONB samples (Figure 3B), we observed no statistically significant pattern of difference between high-grade and low-grade tumors (p = 0.43). Importantly, both low-grade and high-grade ONB had significantly more NEUROD1(+) nuclei than SNUC (p = 0.005 for high-grade ONB vs. SNUC; p = 0.001 for low-grade ONB vs. SNUC) (Figure 3C). Semi-quantitative analysis was also performed on KI67 staining as an internal technical control, which revealed a statistically significant elevation in KI67(+) nuclei in SNUC relative to ONB, as expected (Figure 3C). These results suggest that the npGBC transcription factor NEUROD1 may have diagnostic utility for differentiating between ONB and SNUC in a clinical setting.

Figure 3. NEUROD1 immunohistochemistry on patient samples of ONB and sinonasal undifferentiated carcinoma.

A Histopathologic assessment of ONB and SNUC. Slides were stained with hematoxylin and eosin (H&E, top row) or processed for immunoperoxidase labeling with 3,3’-diaminobenzidine. Representative staining patterns for NEUROD1 (middle row) and KI67 (bottom row) are shown. B Higher magnification of the representative images in (A). C Results from semi-quantitative scoring of tissue samples. Tissue categories are as follows: HG-ONB, high-grade ONB; LG-ONB, low-grade ONB; SNUC, sinonasal undifferentiated carcinoma.

The GBC stemness-associated gene EZH2 labels a proliferative compartment of ONB

While the functions of lineage-specifying transcription factors such as NEUROD1 have been well-characterized in the context of GBC differentiation, the roles of epigenetic regulator machinery are less understood. Recent work in the murine olfactory epithelium has shown that EZH2, the catalytic subunit of the Polycomb repressor complex 2 (PRC2), regulates aspects of GBC stemness including lineage specification and proliferation (50,51). To investigate whether this mechanism may be conserved in ONB, we analyzed published ONB bulk RNAseq data (18) which revealed EZH2 as a differentially upregulated gene in high-grade ONB relative to low-grade ONB (Log2 fold change of 1.95, adjusted p-value of 0.002). We also observed a trend in increasing mitotic index with increased expression of EZH2 in this cohort (Figure 4A). Returning to our ONB samples, we performed immunohistochemistry for EZH2 and KI67 on adjacent tumor sections and observed a positive association between the percentage of EZH2(+) cells and the percentage of KI67(+) cells across the cohort (Spearman’s rho = 0.66, p = 0.008) (Figure 4B). Finally, we analyzed the gene expression of the proliferative cells in our single cell transcriptomic dataset of ONB. Differential gene expression analysis indicated the proliferative cell cluster expressed EZH2 at statistically higher levels relative to noncycling cells (average Log2 fold change = 0.55, p.adj < 0.001) (Figure 4C, Supplemental Figure 3B) suggesting that EZH2 may be preferentially expressed by this population.

Figure 4. Expression of EZH2 in the proliferating compartment of ONB.

A Scatterplot of normalized EZH2 expression and mitotic rate for ONB tumors in the Classe et al. cohort (18). B Left; representative images of paired EZH2 and KI67 immunolabeling. Right; semi-quantitative scoring of EZH2 and KI67 immunolabeling across 15 ONB tumors. Linear fit with 95% confidence interval is indicated (red line and grey region, respectively). Spearman’s rho indicates the correlation between the two percentages. Visualization was generated by introducing Gaussian noise to resolve overplotted points. C Expression of NEUROD1, EZH2, and MKI67 in the ONB single cell data set. Colors indicate the proliferating cell cluster versus non-cycling ONB cells. D Representative image of NEUROD1 immunofluorescence in cryopreserved ONB tissue. E Representative image of dual-labeled EZH2 and KI67 immunofluorescence in cryopreserved ONB tissue used as input for the CellProfiler analytic pipeline. F Median-normalized output from CellProfiler for all five ONB tumors under analysis. Total number of cells analyzed from each tumor is noted. Universal threshold for calling KI67 positivity is indicated with a dashed line. G Quantification of the data in (F) by decile binning. Bar height and error bars indicate the mean and standard deviation of n=3 technical replicates. Asterisks indicate statistical significance: * p < 0.05; ** p < 0.01; *** p < 0.001 by the two-sided Student’s t-test.

We then analyzed the relationship between EZH2 expression and cellular proliferation in situ by performing colocalization analyses on ONB tissue. Five additional, cryopreserved ONB tumor samples were first confirmed to be NEUROD1(+) (Figure 4D), followed by two-color immunofluorescence against EZH2 and KI67 simultaneously. Nuclear staining of KI67 and EZH2 was apparent within the tumor parenchyma, and we observed that KI67(+) nuclei were often (but not always) strongly positive for EZH2 immunofluorescence (Figure 4E). We then measured EZH2 and KI67 fluorescence intensity at the per-cellular level using CellProfiler (52) (Figure 4F). To normalize for intertumoral differences in EZH2 intensity, each tumor was analyzed independently by first binning EZH2 fluorescence into deciles. The percentage of KI67(+) cells was then calculated per EZH2 decile according to a universal threshold (Figure 4G). This revealed that most brightly stained EZH2(+) cells were, indeed, much more likely to be positive for KI67 in all five tumors (n = 3 technical replicates per tumor). These results demonstrate that ONB tumor cells expressing EZH2 are more likely to proliferate, and that a proliferative compartment labeled by the GBC stemness gene EZH2 may be a reproducible histologic feature of ONB.

Divergent differentiation within ONB recapitulates a GBC fate-determining transcriptional pathway

ONB occasionally present with S100-positive sustentacular-like cells that can vary in histologic pattern and proportion. This implies that some tumors undergo a process of divergent differentiation, however the mechanisms behind this phenomenon are unknown. In the olfactory epithelium, multipotent GBCs rely on transcriptional interactions between bHLH transcription factors such as HES1 and ASCL1 to promote neuronal differentiation or suppress it in favor of sustentacular cell production (37,43,44,53,54). We turned to our attention to differentiation pathways within ONB to ascertain whether these GBC differentiation mechanisms may be preserved in some form within the tumor. Upon initial inspection of tumor cell clusters within the single cell RNAseq data, it became apparent that some clusters were defined by differences in gene expression that did not originate from differences in core cellular identity. One cluster was defined by the expression of hypoxia-related metabolic genes, another by cell cycle activity, and yet another by elevated expression of ribosomal subunits (Supplementary Figure 2B). To facilitate the detection of true cell types and differentiation pathways within the cancer, we performed multivariate regression to minimize the effects of these and other distorting gene signatures. In total seven variables were regressed out: cellular RNA content, number of unique genes per cell, mitochondrial gene expression, ribosomal subunit expression, cell cycle gene expression, and hypoxia-related gene expression (55) (Supplementary Figure 2C).

Three differentiation states were observed after regressing out the confounding effects of non-identity signatures. These states did not differ significantly in their expression of NEUROD1, however between cell states there were small yet statistically significant differences in the expression of some neuroendocrine markers used in the routine immunohistochemical profiling of ONB (Figure 5A). To further characterize the ONB cell states we performed differential gene expression analysis solely with respect to transcription factors and cytokeratin species (Figure 5B). State 1 was defined by higher levels of the GBC marker HES6 (31,56,57). State 2 expressed elevated levels of transcription factors implicated in olfactory sensory neuron maturation such as ATF5, BCL11B, BPTF, and SOX11 (State 2) (57–60), and State 3 uniquely expressed elevated levels of Notch-associated genes including HES1, a known inhibitor of olfactory neurogenesis (53). Our analysis of cytokeratins found that only KRT8, KRT18, and KRT17 were differentially expressed between the cell states, and that all three were elevated in the HES1-expressing state. Finally, we performed diffusion mapping and analysis of RNA velocity (61–63) to define the flux between these cell states, which showed differentiation proceeding from State 1 to States 2 and 3 in a divergent fashion (Figure 5C).

Figure 5. Single cell analysis of divergent differentiation in olfactory neuroblastoma.

A Left; UMAP embedding and clustering of ONB parenchymal cell types after the influence of non-cell-identity signatures were regressed out (see Supplementary Figure 2B, 2C). Right; violin plots depicting expression of NEUROD1 and diagnostic neuroendocrine markers in each of the ONB cell states. The color of the violin indicates statistical significance (adjusted p-value < 0.05) in a one-versus-all fashion. Red; significantly upregulated in that cluster. Black; significantly downregulated in that cluster. Grey; no statistically significant difference in expression. B Differentially expressed transcription factors and cytokeratin species between ONB cell states. The top ten transcription factors for each cluster are shown. C Diffusion map embedding of the three ONB cell types with and without superimposed RNA velocity streams. D Ranking of normal nasal mucosal cell types according to how strongly they express the ONB cell type signatures. E Expression of ONB cell type signatures in nasal mucosal cells. Top ranking cell types from (C) are indicated across panels.

Our analyses of transcription factor and cytokeratin expression suggested that ONB differentiation states 2 and 3 may reflect attempts at neuronal and sustentacular differentiation, respectively. To test this in an unbiased manner, we compared the gene expression profiles of the ONB differentiation states to the gene expression of normal cell types within the nasal mucosal cell atlas. Using the Seurat module detection algorithm (64), we identified differentially expressed genes between the differentiation states and then screened for their expression on a per-cell basis in normal cells. We then ranked the normal cell types by their average expression of the ONB signatures to identify which normal cell types most resembled each of the ONB cell states at the transcriptomic level. We found the signature for State 1 was weakly detected in the npGBC and olfactory HBC populations, suggesting a poorly differentiated progenitor phenotype. The signature for State 2 was strongly detected in npGBCs and iOSNs, suggesting that this state represents progression down the path of neuronal maturation. Lastly, the gene signature for State 3 was most strongly detected in the olfactory epithelial sustentacular cells and was absent from the neuronal cell types (Figure 5D, 5E). Collectively the expression signatures of, and the relationships between, ONB cellular states are highly suggestive of a HES1-mediated sustentacular-like divergent differentiation occurring in a background of attempted neuronal maturation.

DISCUSSION

The sinonasal cavity gives rise to the largest malignant differential per cubic centimeter of the human body and therefore poses a unique diagnostic challenge. While a small study of six ONB tumors reported consistent detection of the GBC-associated HASH1 (the human homolog for ASCL1) transcript nearly thirty years ago (65), and work performed a decade later suggested that RT-PCR-based detection of ASCL1 RNA could distinguish between ONB and poorly differentiated sinonasal tumors (66), the contemporary immunohistochemical workup for ONB still neglects GBC-specific markers and remains dependent on neuroendocrine stains that do not directly speak to olfactory tissue nor cells of olfactory epithelial origin. This has led to a particularly challenging diagnostic landscape for ONB, which has become increasingly confusing in light of recently proposed ONB subtypes which exhibit aggressive disease progression and are defined by molecular/genetic markers that overlap with other sinonasal malignancies (i.e. pan-keratin expression and IDH2 mutant status, both of which have been previously attributed to SNUC) (18–21,67–69). Naturally, interpretations of isodiagnostic heterogeneity are entirely dependent on pre-analytic classification of the tissue, which in the case of ONB, has yet to require olfactory epithelial cell markers.

For our analysis of ONB, we brought olfactory epithelial cell markers together with transcriptome-wide gene expression profiling to study the cells that comprise both olfactory neuroblastoma and normal olfactory epithelial tissue. Our approach also leveraged a careful strategy of human-murine integration to clarify the transcriptomic identities of underrepresented human cell types. In doing so, we have identified a transcriptomic and immunohistochemical signature that links ONB to a specific cell type within the olfactory epithelium, the neuronal progenitor GBC, and distinguishes it from SNUC; a primitive and aggressive cancer that displays overlapping molecular features with high-grade ONB. While our results don’t comprehensively address the vast differential of undifferentiated, neuroendocrine, and neuroepithelial tumors that can be found in the sinonasal cavity, our findings strongly support the existence of a sinonasal tumor that recapitulates the identity and developmental trajectory of olfactory epithelial neuronal progenitor GBCs. Markers for GBCs and regulators of GBC differentiation discussed herein, including the transcription factor NEUROD1, may represent opportunities for improving diagnostic taxonomies by utilizing a robust normal-tissue reference frame for the identification of ONB instead of distal neuroendocrine markers that are known to be nonspecific.

Beyond static tumor markers, we report evidence for preserved developmental pathways belonging to olfactory epithelial GBCs within the ONB parenchyma; findings that are consistent with the current understanding of tumors that arise from other developing parts of the nervous system (64,70–74). Analyses of ONB tissue, tumor cell subpopulations, and tumor gene expression demonstrate that ONB contains a highly proliferative EZH2(+) cell type that appears to mirror the multipotent/stem GBC population. EZH2 is an emerging therapeutic target in other settings of malignancy (75–77), and our results suggest it may be a positive regulator of proliferation in ONB tumor cells as it is in normal GBCs (51). Additionally, we find transcriptomic evidence that suggests ONB tumor cells can undergo HES1-dependent divergent differentiation down a sustentacular-like pathway. These findings suggest a mechanism that mirrors a crucial fate-determining transcriptional regulatory circuit that governs the fate of normal GBCs by suppressing neuronal differentiation (53), and offers insight into the variable presence of KRT8/18-positive cells documented throughout the ONB literature (46,78–80). In the future, mechanistic studies of ONB stemness, differentiation, and susceptibility to targeted therapeutics (e.g. EZH2 inhibitors) will require in vitro systems. To the best of our knowledge these models are underdeveloped to-date, however our findings of molecular parallelism between ONB and the normal stages of GBC differentiation suggest that tissue culture systems used to study murine GBCs may be adaptable to support ONB cells (51).

While the clinical and histological diversity of ONB is well-documented, controversy and disagreement abound regarding the categorical limits of that heterogeneity and the prognostic and therapeutic implications that follow. By defining ONB as a tumor of GBC-like cells we have established a normal tissue reference frame from which other poorly differentiated sinonasal tumors may be evaluated and classified. Given the rarity of ONB, a multicenter effort would be most appropriate to further explore the diagnostic and prognostic values of GBC markers (e.g. NEUROD1, EZH2, and HES1, and others). Such large-scale studies could explore expanded differentials including SNUC variants (INI1-deficient, BRG1-deficient, and IDH mutant) (81,82), and the newly described olfactory carcinoma (13). Given that NEUROD1 expression has been found in other tumors with neural differentiation (e.g. medulloblastoma) and some neuroendocrine tumors (e.g. a subset of small cell lung cancers), future studies should evaluate whether NEUROD1 improves the immunohistochemical distinction between ONB and sinonasal neuroendocrine carcinoma as well (83–85).

DATA AVAILABILITY

Transcriptomic data generated herein are available from the Gene Expression Omnibus under accession number GSE166612.