Olfactory neuroblastoma mimics molecular heterogeneity and lineage trajectories of small cell lung cancer

SUMMARY

The olfactory epithelium undergoes neuronal regeneration from basal stem cells and is susceptible to olfactory neuroblastoma (ONB), a rare tumor of unclear origins. Employing alterations in Rb1/Trp53/Myc (RPM), we establish a genetically-engineered mouse model of high-grade metastatic ONB exhibiting a NEUROD1⁺ immature neuronal phenotype. We demonstrate that globose basal cells (GBCs) are a permissive cell of origin for ONB, and that ONBs exhibit cell fate heterogeneity that mimics normal GBC developmental trajectories. ASCL1 loss in RPM ONB leads to emergence of non-neuronal histopathologies, including a POU2F3⁺ microvillar-like state. Similar to small cell lung cancer (SCLC), mouse and human ONB exhibit: mutually exclusive NEUROD1 and POU2F3-like states, an immune-cold tumor microenvironment, intratumoral cell fate heterogeneity comprising neuronal and non-neuronal lineages, and cell fate plasticity—evidenced by barcode-based lineage tracing and single-cell transcriptomics. Collectively, our findings highlight conserved similarities between ONB and neuroendocrine tumors with significant implications for ONB classification and treatment.

In brief

Finlay et al generate mouse models of olfactory neuroblastoma (ONB) that demonstrate striking similarities to human ONB and other neuroendocrine tumors, including expression of lineage-related oncogenes ASCL1, NEUROD1 and POU2F3, an immune-cold tumor microenvironment, intratumoral cell fate heterogeneity and plasticity, and shared expression of therapeutic targets.

Graphical Abstract

INTRODUCTION

Olfactory neuroblastoma (ONB, esthesioneuroblastoma) is a rare malignant tumor with an unknown cellular origin that is predicted to arise from the olfactory epithelium (OE)^1–4. ONB primarily develops in the superior nasal cavity at the anterior skull base, and its intracranial or orbital extension can lead to pain, loss of smell or vision, headache, epistaxis, and potential facial disfiguration. The current standard-of-care is surgery and post-operative radiation, and occasionally adjuvant chemotherapy, including cisplatin and etoposide. ONBs are classified using the Hyam’s grading scale where high-grade ONB is considered Hyam’s grade III-IV, which is associated with increased metastases and poor outcomes^5,6. Pathologically, ONB resembles a “small round blue cell” tumor that is difficult to distinguish from other classifications such as small cell neuroendocrine carcinoma (SNEC), sinonasal undifferentiated carcinoma (SNUC), and large cell neuroendocrine carcinoma (LCNEC)—an issue that is exacerbated by limited tissue for molecular analyses. Although ONBs often express neuroendocrine markers, some of which overlap with SNEC, they typically lack widespread expression of keratins, distinguishing ONB from other sinonasal malignancies.

In the olfactory epithelium (OE), neurogenesis is a dynamic process that continues throughout adulthood⁷. The OE comprises various cell types including horizontal basal cells (HBCs), globose basal cells (GBCs), immediate neuronal precursors (INPs), immature and mature olfactory sensory neurons (iOSNs and mOSNs), non-neuronal microvillar (MV) cells (i.e., POU2F3⁺ brush/tuft-like MV1 and FOXI1⁺ ionocyte-like MV2), sustentacular (Sus) cells, and ducts of the submucosal Bowman’s glands (BG) (Figure 1A). The multipotent HBC and GBC populations can differentiate into all OE cell types, and maintain neuroepithelial homeostasis⁸. HBCs are relatively quiescent but can respond to severe injury by entering the cell cycle and regenerating non-neural Sus cells and the GBC/neuronal lineages^9,10. GBCs are mitotic progenitors that undergo progressive neuronal differentiation marked by the expression of bHLH transcription factors; they transiently express ASCL1 followed by NEUROD1 as they commit to an INP state^11,12, but GBCs can also produce non-neuronal cells including MV cells^10,13–16. POU2F3⁺ MV cells in the OE highly resemble POU2F3⁺ brush/tuft cells in the lung, which are chemosensory cells with brush-like apical microvilli^15,17. The OE appears to be highly conserved between mouse and human, as scRNA-seq studies have identified all OE cells in both species, including the GBC (HES6⁺, ASCL1⁺), INP (NEUROD1⁺) and immature neuronal states¹⁴. ONB transcriptionally resembles GBCs⁴, although the ability to directly investigate GBCs as a cell of origin has been hindered by a lack of model systems.

Figure 1. Rb1/Trp53/Myc (RPM) mice develop high-grade metastatic olfactory neuroblastoma (ONB).

A) Normal cell types in the olfactory epithelium. Sus = sustentacular; MV1/2 = microvillar-1 or −2; HBCs = horizontal basal cells; BG = Bowman’s gland; OSN = olfactory sensory neuron; GBC/INP = globose basal cell / immediate neuronal precursor; OEC = olfactory ensheathing cell. B) Pan-cancer UMAP of 604 bulk RNA-seq samples from 39 tumor types. Dashed circles indicate clustering of neuroendocrine and neural tumor types corresponding with Leiden clusters in Figure S1B. C) MicroCT images of uninfected control (top) or Ad5-Cgrp-Cre-infected RPM mouse (bottom) 10 weeks post-infection in multiple planes. Orange arrows indicate tumor occluding black air space. D) Representative H&E of RPM mouse brain (top left; scale bar, 5 mm) with high magnification inset (top right). Middle panel: Arrowheads indicate rosettes. Bottom panels: RPM olfactory tumor with arrows indicating organized necrosis. Scale bars, 50 μm, unless indicated. E) Representative immunohistochemistry (IHC) of RPM lung (bottom) vs olfactory (top) tumor for indicated antibodies with H-score quantification. Scale bar, 25 μm. Mean ± standard deviation (SD); Student’s unpaired t test, **** p<0.0001, * p<0.03, not significant (ns). F) Representative IHC or IF for indicated markers (IHC scale bar, 25 μm; IF scale bar, 50 μm). Positive control (+) insets; pan-CK = kidney; P63 = Sox2^LSL/LSL;Nkx2–1^fl/fl;Lkb1^fl/fl squamous tumor. G) Cervical lymph node (LN) metastases (mets) in an RPM GEMM (left, yellow arrows indicate tumor) and representative H&E of ONB mets (middle, top scale bar,1 mm; bottom scale bar, 50 μm). Bar graph indicates percentage of mice with ONB that have mets +/− SCLC. H) Representative IHC of primary tumors (1°) or mets in mice with SCLC (left) vs mice with ONB (right) for indicated antibodies with H-score quantification of n=5–14 tumors per category. Mean ± SD; Student’s unpaired t test, **** p<0.0001, *** p<0.0002, ** p<0.004, * p<0.03, not significant (ns).

See also Figure S1 and Table S1.

Genomic studies of ONB are limited by scarce available tissue. Across numerous ONB studies, point mutations in TP53 are the most commonly-reported genetic mutations^18–21, and gain/loss in chr 17p could involve gain of mutant-TP53 or loss of wildtype-TP53, respectively. RB1 mutations and chromosomal loss are reported in ONB^19,21, as well as alterations in other cell cycle and proliferation regulators including CDKN2A, CCND1, PIK3CA/PTEN, KIT, FGFR, and APC^18,20,22. Some ONB studies have reported MYC amplification and gains in chr 8q^19,23, as well as MYC mutations of unknown function²⁴. Notably, alterations in these genes are among the top recurring mutations in small cell lung cancer (SCLC)^25–27, but the ability of these genes to promote ONB has not been tested.

SCLC, like ONB, is considered a “small round blue cell” tumor that shares the expression of lineage-specific transcription factors important to olfactory neurogenesis. While historically treated as a single disease, recent studies demonstrate that SCLC can be molecularly stratified based on expression of lineage-related drivers ASCL1, NEUROD1, POU2F3, and more controversially, YAP1^28–32. These molecular subtypes are characterized by neuroendocrine (ASCL1, “SCLC-A”), neuronal (NEUROD1, “SCLC-N”), tuft (POU2F3, “SCLC-P”), or mesenchymal cell (YAP1) phenotypes, respectively. Recent studies have shown that SCLC subtype states are highly plastic^33–40. The MYC oncogene can dynamically drive SCLC cells from an ASCL1⁺ neuroendocrine to a NEUROD1⁺ neuronal to a YAP1⁺ mesenchymal state^33,40–42. Mutually exclusive ASCL1, NEUROD1 and POU2F3 states are also found in human SCLC and neuroendocrine prostate cancer^{30,33,36,43–46}, suggesting conservation of these lineage relationships across tissues. A recent study using bulk RNA-sequencing of ONB tumors categorized distinct neural and basal states²¹, but bulk profiling did not enable assessment of intratumoral heterogeneity. Another preliminary single-cell RNA-sequencing (scRNA-seq) study involving human ONBs validated the existence of neural and basal subtypes while suggesting additional molecular subtypes⁴⁷. However, unlike SCLC, the ability to address the role of lineage-related transcription factors and cell fate plasticity in ONB has been prevented by a lack of models for functional studies.

Given the scarcity of tissue and models for ONB, new models are urgently needed for molecular and functional studies. Very few mouse models of olfactory tumors have been described beyond virally-induced models, such as SV40^48,49 or damage-associated rat and hamster models. These models are limited by a lack of precise genetics, long latencies, and superficial characterization. Human cell lines and culture models for ONB are currently unavailable, together presenting significant challenges for understanding the disease. Here, we seek to develop models of ONB to interrogate genetic alterations, cell of origin, lineage-related transcription factors, and cell fate plasticity.

RESULTS

Rb1/Trp53/Myc (RPM) mice develop high-grade metastatic olfactory neuroblastoma

We first asked how closely ONB resembles neuroendocrine versus neural tumors and performed a pan-cancer analysis integrating human ONB with 39 different tumor types across 604 patient samples using bulk RNA-sequencing datasets^{21,26,50–52}. Hierarchical (Figure S1A) and UMAP (Figure 1B) clustering across all samples revealed that tumors clustered by histopathology as opposed to tissue of origin (Figures 1B and S1A). The majority of ONB samples clustered in proximity to neuroendocrine tumors—in one group containing SCLC-A, SCLC-N, and neuroendocrine prostate cancer (NEPC), and another containing a subset of NEPC and pancreatic tumors (Figure S1A, right inset). Unbiased Leiden clustering in UMAP space similarly grouped ONB with SCLC (Cluster 1, Figure S1B). Notably, ONB did not cluster with neuroblastomas or other neural tumors including high and low-grade glioma, pheochromocytoma, and medulloblastoma (Figures 1B and S1B). Interestingly, two ONB samples clustered in a small, distant branch containing all four SCLC-P samples (Figure S1A, left inset). Thus, the majority of human ONBs closely resemble the transcriptional profiles of lung and prostate neuroendocrine tumors.

Rb1^fl/fl;Trp53^fl/fl;LSL-Myc^T58A/T58A (RPM) mice have been used to model neuroendocrine lung tumors^33,41,53, so we sought to determine whether they can develop ONB. Intranasal adenovirus delivery can target murine olfactory epithelium⁵⁴. Of the cell-type-specific viruses available to us, we chose Adeno-Cgrp promoter-Cre (Ad-Cgrp-Cre) to potentially target neuronal progenitor cells. Using a modified intranasal method of Ad-Cgrp-Cre administration in RPM mice (n = 18), some animals exhibited symptoms of sinonasal or anterior skull base tumors in the absence of lung tumor burden, including asymmetrical facial swelling, head tilting, difficulty breathing, and weight loss. Upon skull dissection, dense and opaque tissue was observed in the olfactory area consistent with tumor. Abnormal lesions were detected in the olfactory space by microCT imaging in 12 of 18 RPM animals starting ~7–8 weeks post-infection (Figure 1C), similar to SCLC latency in this model⁴¹. As some virus can travel to the lungs, 6 RPM animals developed SCLC that required sacrifice prior to the onset of olfactory tumors, precluding proper analysis of tumor penetrance; still, 12 of 12 animals that did not require sacrifice from lung tumor burden harbored olfactory tumors. Pathological assessment of hematoxylin & eosin (H&E)-stained tissue identified large olfactory tumors with “small round blue cell” morphology (Figure 1D).

Olfactory tumors were robustly proliferative, indicated by MKI67 immunohistochemistry (IHC) (Figure 1E). RPM animals infected with Ad-Cgrp-Cre develop SCLC but not lung adenocarcinoma^41,53, both of which express NKX2–1 (also known as TTF1). Olfactory tumors were uniformly negative for NKX2–1 but expressed the olfactory lineage marker LHX2 (Figure 1E), suggesting olfactory tumors were unlikely SCLC metastases and more likely arising in the olfactory area. A board-certified pathologist with expertise in olfactory tumors (Diana Bell, MD)^1,6 evaluated H&E-stained RPM tumors without knowledge of genotype or biomarker expression. All nine tumors were classified as high-grade malignancies favoring neuroectodermal derivation, specifically SNEC or Hyam’s grade III/IV ONB (Table S1). RPM tumor cells exhibited high nuclear/cytoplasmic ratio, speckled (salt-n-pepper) nuclear chromatin, visible nucleoli, occasional neurofibrillary stroma, and occasional rosettes (Flexner-Wintersteiner and Homer Wright) (Figure 1D). Most tumors demonstrated high mitotic index and focal areas of apoptotic bodies and necrosis, consistent with high-grade malignancies (Figure 1D).

To better characterize the molecular phenotype of these tumors, we stained tumors for pan-cytokeratin (pan-CK), basal epithelial marker TRP63 (P63), and neuronal/neuroendocrine markers, chromogranin A (CHGA), synaptophysin (SYP), NCAM1/CD56, and UCHL1. RPM tumors were uniformly negative for pan-CK and P63 and expressed abundant neuronal/neuroendocrine markers (Figures 1F and S1C). Cells expressing S100B were sparsely detected in RPM tumors adjacent to tumor regions expressing neuronal antigens, suggesting ingrowth of S100B⁺ olfactory ensheathing cells (OECs) (Figure 1F). Recent studies suggest NEUROD1 expression can distinguish human ONB from SNUC⁴. Mouse RPM olfactory tumors expressed abundant NEUROD1 (Figures 1F and S1C), at levels higher than that observed in SCLC from RPM mice. Given the neuronal/neuroendocrine NEUROD1⁺ identity of RPM olfactory tumors with a lack of keratin expression, the consensus differential diagnosis for these tumors by the pathologist is high-grade ONB. MYC appears to be important for ONB tumorigenesis because intranasal Ad-Cgrp-Cre infections in Rb1/Trp53/Rbl2 (RPR2) mice, another frequently used SCLC model^37,55, did not lead to nasal tumors by microCT imaging as late as 10 months post-infection (Figure S1D).

Cervical lymph nodes (LNs) are the most common site of metastases for human ONB⁵⁶. Approximately 29% of RPM mice (9 of 31 with ONB) had cervical LN metastases (Figure 1G). Given that 4 of 9 animals also harbored SCLC at the time of sacrifice, we sought to determine whether we could predict the origin of the cervical LN metastases, or whether ONB metastasized to the lung. Metastases from animals with only SCLC were predominantly in local lymph nodes and the liver, and were NKX2–1⁺/LHX2⁻ (Figure 1H). In contrast, metastases from animals with only ONB were found in cervical LNs and were NKX2–1⁻/LHX2⁺ (Figures 1G–H). LHX2 was not detected in lung tumors of animals with ONB, suggesting that ONB likely does not travel to the lung but rather to cervical LNs. These data support that genetic alterations in Rb1/Trp53/Myc in mice promote highly-penetrant, metastatic ONB that recapitulates the biology of high-grade human ONB.

Olfactory neuroblastoma can arise from GBC progenitors

The potential cell(s) of origin of ONB are unknown, but evidence suggests that ONB transcriptionally resembles GBC/INPs^4,21,57,58. We sought to determine the origins of ONB in the RPM model. Since Ad-Cgrp-Cre uses the Cgrp promoter, we assessed olfactory mucosa for potential Cgrp⁺ target cell populations. ScRNA-seq data of the normal OE suggest that Cgrp (also known as Calca) is lowly and sparsely detected in captured olfactory cell populations (Figure 2A). In agreement, CGRP protein was only detected in trigeminal nerve fibers terminating at the mucosal surface (Figure 2B), consistent with prior studies⁵⁹. Trigeminal neuron cell bodies reside intracranially in the trigeminal ganglion, caudal to the olfactory bulbs; RPM tumors did not arise in the trigeminal ganglion but did arise in the basal layer of the OE, invading into the lamina propria (Figure 2C). Tumors were not detected in other regions supplied by trigeminal branches V1–3, such as the oral cavity, forehead, etc. As such, Cre expression by intranasal Ad-Cgrp-Cre likely occurs within the OE.

Figure 2. Olfactory neuroblastoma can arise from GBC progenitors.

A) UMAP of scRNA-seq data from normal olfactory epithelium (OE) cells (left) and corresponding Cgrp/Calca expression (right). B) Immunofluorescence (IF) staining of CGRP (green: trigeminal nerves, white arrowheads), SOX2 (pink: HBCs, Sus cells), TUBB3 (yellow: INPs, OSNs) and Hoescht (blue: nuclei) in the normal OE of uninfected RPM mice. C) In situ tumors of RPM GEMMs emerging from the basal layer in the OE. Arrowheads indicate TUBB3⁺/KI67⁺ tumor cells invading the lamina propria. AS=nasal airspace, LP=lamina propria. D) Experimental timeline indicating Ad-Cgrp-Cre delivery to reporter mice at day 0 and collection of tissue at indicated days post-infection (dpi) (top) with IF of the normal OE for tdTomato (tdTom) (pink: Cre-mediated recombination) and Hoescht (blue: nuclei). E) IF on normal OE for tdTom (pink: Cre-mediated recombination), TUBB3 (yellow: INPs, OSNs), Hoescht (blue: nuclei) and left: ASCL1 (green: GBCs), middle: NEUROD1 (green: GBCs, INPs), or right: SOX2 (green: HBCs, Sus cells). White indicates overlap of ASCL1, NEUROD1, or SOX2 with tdTom. Arrowheads in left and middle panel indicate ASCL1⁺/tdTom⁺ or NEUROD1⁺/tdTom⁺ GBCs. Arrowheads in right panel indicate TUBB3⁺/tdTom⁺ OSNs while asterisks indicate SOX2⁺/tdTom⁺ Sus cells. F) Quantification of tdTom⁺ GBC vs non-GBC-derived cells in Ai9 reporter mice at Day 5–7 post-Cre (n = 3). Mean ± standard deviation (SD), Student’s unpaired t-test, *** p<0.001. G) Experimental schematic depicting 1–2) MMZ treatment of uninfected RPM mice and isolation of normal RPM GBCs, 3) ex vivo Cre-mediated transformation validated via PCR, and 4) implanting transformed, GBC-derived organoids as allografts. H) Representative H&E of primary RPM tumor (top) vs GBC-derived RPM allograft (bottom) and IHC staining of indicated antibodies with H-score quantification. Mean ± SD; Student’s unpaired t test, * p<0.02, not significant (ns). All scale bars, 50 μm. All dashed white lines indicate basal layer of OE.

See also Figure S2 and Table S1.

To address this issue directly, we infected Cre reporter “Ai9” mice (Rosa26-LSL-tdTomato) with intranasal Ad-Cgrp-Cre and harvested nasal tissue after 3, 5 and 7 days (Figure 2D). Rare, single tdTomato⁺-fate-mapped cells were detected at Day 3. At Day 5, rare solitary labeled cells and small clusters (< 4 cells per cluster) were detected, with substantially increased labeled cells by Day 7. Quantification of cell type-specific markers demonstrated that the vast majority of labeled cells are GBCs or their derivatives (Figures 2E–F and S2A), including INPs, OSNs, and MV cells; in other words, the vast majority of cells harboring recombination events at Day 5–7 were not lone cells, but found in clusters, mirroring the results of GBC-driven lineage tracing studies¹⁶. Importantly, MV cells, a post-mitotic population, were only labeled when found in clusters of cells with other cell identities, and not as single cells (Figure S2A), suggesting that MV cells are not likely the initial cell of origin. Given some similarities of ONB to respiratory-derived tumors such as SNEC and SNUC, we quantified the number of tdTomato⁺ respiratory cells. TdTomato⁺ respiratory cells were rare across all timepoints, and when present, consisted of a solitary, isolated cell that had not divided (Figures S2B–C). We did not detect recombination in trigeminal neurons (Figure S1D), OECs (Figure S1E), or cells in the olfactory bulb (Figure S2F). These results, combined with the lack of tumors in the respiratory portion of the nasal cavity, suggest that the virus has high specificity for the OE. Although GBC/INP cells appear to be CGRP-negative, Cgrp is driven by bHLH factors⁶⁰, and the proliferative GBC/INPs strongly express several bHLH transcription factors including ASCL1, NEUROG1, and NEUROD1. Together, our results indicate that GBCs/INPs are infected and express Cre, suggesting GBCs/INPs are a potential cell of origin in the ONB model.

To specifically address if GBCs can transform into ONB, we treated uninfected RPM mice with methimazole (MMZ) to eliminate differentiated OE cells and promote GBC expansion⁶¹ (Figure 2G). GBCs were immuno-magnetically selected by surface KIT expression to purify highly neurogenic progenitors consistent only with GBC states^14,61. GBCs were treated with Ad-Cre, expanded in culture to induce RPM genetic alterations (Figure 2G), and recombined alleles were PCR verified after transformation (Figure S2G). Transformed GBC-derived organoids were transplanted into flanks of SCID/beige immunodeficient mice and palpable tumors developed in ~6–7 weeks. RPM allografts histologically resembled primary RPM ONB (Figure 2H and Table S1), but with more ASCL1 and slightly less NEUROD1. Together, these data demonstrate that GBCs can serve as a permissive ONB cell of origin.

Loss of ASCL1 enhances non-neuronal histopathologies

GBCs are marked by ASCL1 expression^13,14,16,58, which is important for OE neuronal lineages^13,14,58,62. ASCL1 is expressed in human ONB but is reportedly reduced in higher-grade tumors^6,63,64. ASCL1 is critical for SCLC tumorigenesis in the lung³², and depending on the genetic context, ASCL1 loss can either obliterate tumorigenesis³² or delay tumor latency and block neuroendocrine fate, including the emergence of NEUROD1⁺ tumors^53,62. Also in the lung, MYC can drive ASCL1⁺ neuroendocrine tumors toward a NEUROD1⁺ state via plasticity³³. Thus, we sought to determine whether loss of ASCL1 would abolish tumorigenesis or the NEUROD1⁺ state in ONB, as it does in lung cancer^32,53. To address this, we infected RPM-Ascl1^fl/fl (RPMA) mice with intranasal Ad-Cgrp-Cre. Compared to RPM mice, RPMA olfactory tumors arose with delayed latency of 10–21 weeks with an average latency of 16 weeks post-infection (Figure 3A). The delayed latency of lung tumorigenesis in RPMA compared to RPM mice⁵³ allowed more time for olfactory tumors to develop, such that we detected olfactory tumors in 100% of RPMA animals (14 of 14 mice) by microCT imaging. We verified by IHC that RPMA tumors lack ASCL1 (Figure S3A) and NKX2–1 (Figure S3B). Additionally, ~30% of RPMA mice with ONB developed NKX2–1⁻/LHX2⁺ metastases (7/27 with cervical LN metastases, 1/27 with liver metastases) (Figures S3C–D). Only 1 of 8 mice with ONB plus metastases had a lung tumor, consistent with ONB, not SCLC, being the origin of these metastases (Figure S3C).

Figure 3. Loss of ASCL1 enhances non-neuronal histopathologies.

A) Tumor latency post-Ad-Cgrp-Cre infection in RPM vs RPMA GEMMs. Triangles indicate first signs of ONB by microCT imaging in indicated numbers of RPM (blue) or RPMA (orange) GEMMs. B) Representative H&E of RPMA ONBs with glandular growth and necrosis (left) and hyperplastic seromucinous glands overlying ciliated pseudostratified epithelium (right). C) Representative IHC for indicated antibodies in RPM and RPMA ONB tumors with H-score quantification. **** p< 0.0001, * p<0.04. D) Representative IHC in RPM and RPMA ONBs with H-score quantification. ** p<0.003, * p<0.03; not significant (ns). E) Representative POU2F3 IHC in RPM and RPMA ONBs with H-score quantification by genotype (left), and by histotype in n=16 RPMA tumors (right), ** p< 0.002, * p< 0.03. F) Representative IF for NEUROD1 (green), POU2F3 (pink), TUBB3 (yellow) and Hoescht (blue: nuclei) in an RPMA ONB. G) IF of Hoescht (blue: nuclei), POU2F3 (green), and TUBB3 (pink: INPs, OSNs) on normal human OE (left) or two distinct human ONBs (middle, TUBB3-high; right, TUBB3-low). Dashed white line indicates basal lamina. H) IF of Hoescht (blue: nuclei), HES1 (green), and TUBB3 (pink: INPs, OSNs) on human ONB-1 and −2 from 3G. For all IF/IHC, scale bars are 50 μm. For all bar graphs, error bars are mean ± standard deviation (SD) with Student’s unpaired t-tests.

See also Figure S3 and Table S1.

Pathological evaluation (Dr. Bell) of 17 distinct RPMA tumors (one per animal) revealed increased histopathological heterogeneity compared to RPM tumors. While 9 of 17 tumors (~53%) resembled SNEC and/or high-grade ONB, 8 RPMA tumors (~47%) were diagnosed as SNUC or olfactory carcinoma with adenocarcinoma-like features (Figures 3B, S3E, and Table S1). In contrast to results in the RPMA lung model, RPMA olfactory tumors maintain NEUROD1 in sparse and heterogeneous patterns, but significantly less frequently than in RPM ONBs (Figure 3C). Consistent with reduced NEUROD1, we observed decreased TUBB3 and CHGA expression in RPMA ONB, suggesting reduced neuronal fates upon ASCL1 loss (Figure 3C). SYP and UCHL1 staining were not significantly decreased in RPMA tumors when analyzed by H-score (a function of intensity and percentage of positive cells) (Figure 3D). However, RPM ONB tumors have uniformly diffuse expression of SYP and UCHL1, whereas RPMA tumors have emergence of cells that are starkly negative for these neuronal markers (Figure 3D). HES1 is implicated as a suppressor of OE and ONB neuronal differentiation^4,65, and Notch signaling and ASCL1 are often mutually antagonistic in SCLC^37,53. HES1 was significantly increased in RPMA compared to RPM tumors (Figure 3D), consistent with a gain of non-neuronal fates. Pan-CK was low and focal but enriched in RPMA compared to RPM tumors (Figure 3D), also suggestive of increased non-neuronal fates. RPMA tumors were devoid of the basal marker, P63 (Figure S3F). We asked if differences in IHC markers could stratify RPMA tumors by histopathology. Indeed, RPMA tumors classified as SNUC versus ONB/SNEC had decreased neuronal and increased non-neuronal markers (Figure S3G).

Given that some human ONBs clustered transcriptionally with POU2F3⁺ SCLC (Figure S1A), we assessed POU2F3 expression. Strikingly, in the absence of ASCL1, precocious and robust POU2F3 was detected in ~50% of ONB tumors (Figure 3E). The emergence of POU2F3⁺ tumors was surprising given that it differs from prior reports in RPMA mouse lungs^40,53, but is consistent with multiple neuroendocrine tumors where ASCL1 and POU2F3 are expressed in a mutually exclusive pattern. Single-cell studies from NEPC, SCLC, and other extrapulmonary neuroendocrine tumors have shown that ASCL1 and POU2F3 are expressed in distinct tumor cell populations^{30,43–46,66,67}. Co-staining of POU2F3 and NEUROD1 in RPMA tissue revealed that these transcription factors are expressed in non-overlapping tumor cell populations (Figure 3F). Importantly, POU2F3 was specifically expressed in RPMA tumors classified as SNEC/ONB, not SNUC (Figure 3E), supporting the presence of a non-neuronal, MV-like subtype of ONB that is distinct from SNUC. Consistent with the non-neuronal fates present in murine ONB, we detected POU2F3⁺ and HES1⁺ human ONB tumor cells enriched in non-neuronal, TUBB3⁻ populations (Figures 3G–H). Clonal resolution multi-color fate mapping revealed that KIT⁺ GBCs can give rise to both neurons and non-neuronal MV populations, both MV1 and MV2 subsets¹⁶. Thus, the ability of ASCL1 to regulate a MV versus neuronal cell fate switch in our ONB GEMMs is consistent with developmental GBC lineage potential. Further, a reduced neuronal fate upon loss of ASCL1, which is expressed in GBCs, is consistent with GBCs as a cell of origin in this model. Together, these data demonstrate that ASCL1 represses non-neuronal fates during olfactory tumorigenesis.

Mouse ONB transcriptionally resembles human ONB

To better understand the transcriptional heterogeneity in mouse ONB and its relationship to human tumors, we performed scRNA-seq on independent RPM and RPMA tumors. We harvested tumor cells from a RPM-Rosa26-LSL-Cas9-Ires-Gfp (RPM-GFP) mouse in which tumor cells are GFP⁺ (Figure S4A) and 3 independent RPMA tumors that span the histological spectrum in this model including ONB and SNUC (Table S1). To identify tumor cells and determine whether RPM and RPMA ONB cells resemble normal OE cell types, we integrated normal mouse OE^68,69 with RPM and RPMA ONB tumor scRNA-seq data (Figures 4A and S4B–D). RPMA and RPM tumor cells clustered distinctly from each other, and from normal OE (Figures 4A, S4B and S4C)—allowing us to distinguish tumor versus contaminating normal cells from ONB samples. Identified tumor clusters were appropriately enriched for Myc, Cas9, and Luciferase (Figure S4D). Cell cycle analysis showed that tumor cells, GBCs, and early INPs are proliferative compared to more differentiated normal OE cells (Figure S4E). Optimal transport distance between cell type clusters revealed that tumor cells are most transcriptionally similar to GBCs and immature neuronal states versus other normal OE cell types (Figure 4B), consistent with recent observations from human ONB bulk RNA-seq data^4,21. Hierarchical clustering and differential gene expression analysis supported similarity between RPM and RPMA ONB tumors and GBCs (Figure S4F and Table S2).

Figure 4. Mouse ONB transcriptionally resembles human ONB.

A) UMAP of tumor and non-tumor scRNA-seq data from n=1 RPM (n=3,350 cells) and n=3 RPMA (n=1,633 cells) ONBs integrated with normal mouse olfactory epithelium (OE). Inset is tumor cells only. B) Optimal transport (OT) distance matrix for 100 subsampled cells per cluster. Darker colors represent closer transcriptional distance between cell types. Dendrogram is hierarchical clustering by OT between cell types. C) Dot plot of expression of OE cell type markers in RPM vs RPMA tumor cells by scRNA-seq. D) Volcano plot of enriched genes in RPM (purple) vs RPMA (orange) ONB. E) Module scores of indicated human tumor types applied to ONB tumors and normal OE cells in 4A UMAP. F) Published module scores derived from scRNA-seq data of subtyped, human SCLC tumors³⁰ applied to 4A UMAP. Bar graphs indicate average score per cell type. Error bars are 95% confidence intervals. Mann-Whitney with Bonferroni correction, ***p<0.0001. G) UMAP of scRNA-seq data from two human ONB tumors, ONB-A⁴ (red) and ONB-B (ONB-2 in 3G-H, blue) integrated with a published human OE scRNA-seq atlas^14,70,71 (left). Heatmap of average gene expression per ONB tumor (right). H) RPM and RPMA ONB scores applied to human ONB-A vs -B from 4G. Mann-Whitney with Bonferroni correction, **** p<0.0001.

See also Figure S4 and Tables S1 and S2.

In comparison to cell types of the normal OE, RPM ONB tumors are enriched for markers of GBC/INP cells, whereas RPMA ONB tumors are enriched for non-neuronal states, including MV cells (Figures 4C, S4D, and S4G). Consistently, ENRICHR pathway analysis of DEGs reveals increased neuronal/neuroendocrine fates in RPM ONB, and enrichment of non-neuronal and dedifferentiated embryonic-stem-like states in RPMA (Table S2). Olfactory neuron lineage-specific transcription factors are among the most enriched genes in RPM, including Lhx2, Ebf1, Neurod1, Nhlh2, and Myt1l (Figure 4D), while non-neuronal lineage markers including Sox9 (BG, MV), Foxi1 (MV2-ionocyte), and keratins Krt8 and Krt18 were highly enriched in RPMA tumors (Figures 4C–D, S4D, and S4G).

To address the relationship between mouse and human ONB directly, we generated signatures from the pan-cancer bulk RNA-seq analysis in Figure 1B (Table S2). ONB GEMM tumors were highly enriched for human ONB signatures, more so than human neuroblastoma or lung adenocarcinoma (Figure 4E). Consistent with the transcriptional similarity between human SCLC and ONB (Figures 1B and S1B), mouse ONB was also enriched for the human SCLC gene signature (Figure 4E). Similar to SCLC, Mycl and Myc were expressed in RPM ONB tumors, while Mycl was decreased upon ASCL1 loss (Figure S4G), consistent with Mycl being an ASCL1 target gene³².

Given the similarities to SCLC, we utilized gene sets enriched in human SCLC-A, N, and P subtypes and applied module scores to our integrated dataset. Compared to RPMA, RPM ONB had significantly higher scores for the SCLC-A and SCLC-N modules (Figure 4F). Conversely, RPMA ONB had a significantly higher SCLC-P score compared to RPM (Figure 4F), consistent with Ascl1-knockout inducing a shift away from neuronal and toward a microvillar fate (Figures 3E and 4C). Known NEUROD1 targets conserved in mouse and human SCLC^32,53 were also significantly enriched in RPM versus RPMA ONB (Figure S4H). While mouse ONB and SCLC were transcriptionally highly concordant (Figure S4I), 3–5% of genes were specifically enriched in either ONB or SCLC. Importantly, Lhx2 was a top DEG in RPM ONB, while Nkx2–1 was a top DEG in RPM SCLC. These findings are consistent with the enrichment of NKX2–1 in RPM SCLC vs ONB primary tumors (Figure 1E) and the ability of these markers to discriminate the origin of metastases in these models (Figures 1H and S3D).

While single cell transcriptomic data from human ONB is scarce, we sought to take advantage of the one publicly available specimen (“ONB-A”)⁴ and combined it with a new tumor collected for this study (“ONB-B”). The human ONBs were integrated with our previously published single cell atlas of normal human OE^14,70,71. This allowed us to distinguish tumor cells from surrounding normal tissue and infiltrating cells, which defined two distinct tumor clusters—one comprising cells primarily from ONB-A and the other from ONB-B (Figure 4G). ONB-A exhibited higher expression of neuronal genes including NEUROD1, GNG8, CHGA, SYP, and UCHL1, while ONB-B was enriched for non-neuronal genes including KRT8/18, CFTR, WWTR1 (Figure 4G). Comparing mouse ONB signatures (Table S2) to the human dataset, the RPM ONB signature was enriched in neuronal, ONB-A, whereas the RPMA ONB signature was enriched in the less neuronal, ONB-B (Figure 4H). These data reveal that RPM and RPMA genetic alterations recapitulate transcriptional signatures of human tumors.

GBC-derived ONB exhibits plasticity between neuronal and non-neuronal states

To assess the capacity for ONB plasticity directly, we performed single-cell lineage-tracing on GBC-derived ONB allografts to track their ability to become neuronal or non-neuronal tumor populations and determine whether individual cells undergo phenotypic switching. During neurogenesis, GBCs predominantly progress toward neuronal fates as the bulk of the OE comprises neurons, and concordantly, RPM ONB is predominantly NEUROD1⁺. As ASCL1 loss promotes ONB tumors harboring both neuronal and non-neuronal fates, we expected that RPMA GBC-derived tumors may have increased lineage plasticity. To test this, we treated un-infected RPM and RPMA mice with MMZ to eliminate differentiated OE cells and promote GBC expansion (Figure 5A)⁶¹. RPM and RPMA GBCs were immuno-magnetically selected based on surface KIT expression, treated with Cre, grown ex vivo as organoids (Figure 5A), and allele recombination was PCR verified following transformation (Figures S2G and S5A). GBC-derived organoids were infected with a lentiviral CellTag library that expresses unique barcodes in the 3’UTR of Gfp to enable clonal analysis via downstream scRNA-seq^72,73. After in vitro growth to allow clonal expansion, CellTagged, GBC-derived organoids were transplanted into the flanks of SCID/beige immunodeficient mice (Figure 5A). Palpable tumors developed in ~6–7 weeks and were harvested for histology and scRNA-seq. As observed for RPM allografts (Figure 2H), RPMA allografts histologically resembled primary RPMA ONB with mixed neuroendocrine differentiation and glandular features (Figures 5B, S5B, and Table S1). RPMA allografts harbored heterogeneous expression of neuronal and non-neuronal markers, including NEUROD1 and POU2F3 (Figure S5B). Importantly, merging scRNA-seq data from the RPM and RPMA allografts with normal OE and primary ONB tumors (from Figure 4A) demonstrates that RPM and RPMA allografts are highly similar to primary tumors and GBCs (Figures 5C–D). RPM allografts appeared slightly less differentiated compared to their primary counterparts, with increased levels of ASCL1, HES1 and SOX2, and reduced NEUROD1 (Figure S5B–S5C), which may be due to organoid culture prior to transformation and transplant. Altogether, GBC-derived organoid-based tumors resemble their primary tumor counterparts by protein and transcriptomic analyses.

Figure 5. GBC-derived ONB exhibits plasticity between neuronal and non-neuronal states.

A) Experimental design with representative brightfield and GFP organoid images. Scale bar, 275 μm. B) Representative H&E of RPMA primary ONB vs GBC-derived allograft tumors. Scale bars, 50 μm. C) UMAP of scRNA-seq on RPM (n=4,161) and RPMA (n=19,302) GBC-derived allograft, primary RPM and RPMA ONB, and normal OE cells. D) Optimal transport (OT) distance matrix of 100 subsampled cells per cluster. Darker colors indicate closer transcriptional distances between cell types. Dendrogram shows hierarchical clustering based on average OT distances. E) Force atlas (FA) projection of transcriptionally distinct Leiden clusters (top) from indicated samples (bottom). F) Feature plots of indicated genes in 5E FA space. G) Frequency of cells in each CellTagged clone (x-axis) or in the whole RPM or RPMA allograft sample per Leiden cluster from 5E. H) Cell state assignments based on expression of normal OE markers per Leiden cluster in 5E in UMAP (top) and frequency of cell state per clone (bottom) or per whole sample. I) Diffusion pseudotime of RPM and RPMA cells in 5E FA map (top), and predicted differentiation trajectories through cell states (bottom). J) Representative FA maps of clonally-linked cells in multiple states for indicated clones. Colors correspond with state trajectories from 5I. Clone IDs match x-axis labels in 5G-H.

See also Figures S5–S6 and Tables S1 and S3.

To examine GBC-derived allografts and primary ONB at higher resolution, we subset and reclustered the tumor cells and examined transcriptional states via Leiden clustering (Figures 5E, S6A–B, and Table S3). As expected, RPM primary tumors clustered in populations highly expressing neuronal fate markers and RPMA primary tumors in clusters enriched for non-neuronal markers (Figures 5E–F). RPMA allograft cells spanned the transcriptional heterogeneity of RPM and RPMA primary tumors, residing in neuronal and non-neuronal states and occupying a greater diversity of Leiden clusters than RPM samples (Figures 5E–F and S6A–B). Analysis of CellTagged clones in the RPM and RPMA allografts reveals that cells from 18 of 19 clones span multiple transcriptionally-distinct states (Figures 5G and S6A–D). These states correspond with cell fate identities defined by expression of OE marker genes (Figures 5F–H, S6B–C and Table S3) including: GBC-like, neuronal/INP, non-neuronal (epithelial/MV-like), non-neuronal (mesenchymal/glandular), and stem-like (Figure 5H). The majority of cells in each RPM clone were in a GBC-like state corresponding to cluster 7 (Figures 5G–H, S6B and S6D). The majority of cells in each RPMA clone were in Leiden clusters 0 and 2 (Figure 5G), corresponding to states expressing immature neuronal/INP and non-neuronal (epithelial/MV-like) genes (Figures 5F, 5H, S6B and S6D). The abundance of GBC or INP-like cells in RPM and RPMA clones are consistent with self-renewal capacity of those cell fates and their ability to progress to multiple transcriptionally-distinct states. While RPM tumor cells exhibited fate plasticity that spanned GBC, neuronal, and some non-neuronal (mesenchymal) states, RPMA clones reached additional non-neuronal and dedifferentiated stem-like states with greater frequencies (Figures 5E–H and S6D). These clonal dynamics are consistent with our findings in primary tumors that ASCL1 loss promotes greater intratumoral heterogeneity and access to non-neuronal fates. We performed diffusion pseudotime analysis on RPM and RPMA allograft cells, with the starting state defined as the GBC-like cluster (the cell of origin for the allografts). Pseudotime predicted multiple distinct lineage trajectories—each corresponding with movement towards one of the four major cell states (Figures 5I and S6E)—and that RPM and RPMA clones progress to different extents in multiple trajectories (Figure 5J). These analyses support that clonally-linked cells can adopt multiple, unique fates from a transcriptionally-similar starting point. Altogether, lineage-tracing data provide strong evidence that GBCs give rise to ONB tumors comprising cells with multi-lineage plasticity potential, which may explain the cell fate heterogeneity observed in human ONB.

Human ONB exhibits SCLC similarities in intratumoral heterogeneity, immune-cold tumor microenvironment, and therapeutic targets

Human ONB may have the capacity for intratumoral cell fate plasticity. To examine this, we obtained 12 human ONB specimens from surgical resections for spatial transcriptomic profiling (Figure 6A and Table S4). Eight specimens were diagnosed as low-grade, and four as high-grade, with the majority of samples from treatment-naïve patients or following late (>1 year) recurrences (Table S4). Tissue was stained with H&E and proliferation (KI-67), neuronal (TUBB3), immune (CD45) and nuclei (SYTO 13) markers (Figure 6B). From 2–3 regions of interest (ROIs) per tumor (Figure S7A), whole transcriptome readouts were collected and defined as either KI-67⁺, TUBB3⁺, CD45⁺, or pan-negative. Unbiased clustering of ROIs in UMAP space segregated stromal/immune, low-grade, and high-grade regions, such that CD45 was specific to stromal/immune regions, and KI-67 and TUBB3 were confined to tumor regions (Figure 6C). DEGs were identified in TUBB3⁺, KI-67⁺, or pan-negative ROIs that were elevated in high- vs low-grade ONB (Table S4). GBC and non-neuronal signatures were enriched in high-grade ONB, whereas neuronal signatures were enriched in low-grade ONB—suggesting a shift away from neuronal cell fate in more aggressive tumors (Figure 6D).

Figure 6. Human ONB exhibits SCLC similarities in intratumoral heterogeneity, immune-cold tumor microenvironment, and therapeutic targets.

A) Schematic indicating collection of human specimens by surgical resection, formalin-fixation and paraffin embedding (FFPE), and Nanostring GeoMx profiling. B) Representative regions of interest (ROIs) in a sample including H&E staining (left), immunostains for KI-67 (pink: proliferation), TUBB3 (yellow: OSNs, tumor cells), and CD45 (green: immune cells) (middle). Representative ROI segmentation (right). Scale bar, 50 μm. C) UMAP clustering of transcriptomic profiles from all 48 ROI segments across samples. Marker expression and tumor grade per ROI indicated according to legend. Pan-negative ROI segments are negative for all three markers. D) Violin plots showing cell-type signature scores per ROI in low- (blue) vs high- (orange) grade human ONB. Each dot represents one ROI; Mann-Whitney with Bonferroni correction; *p<0.05; **p<0.01; ***p<0.001; ns = not significant. E) Leiden clustering of ROIs (excluding stromal/immune segments). Bar graphs depict proportion of ROIs per Leiden cluster 1 and 2 based on tumor grade (left), immunostain (middle), or patient ID (right). F, G) Violin plots of indicated F) non-neuronal vs neuronal genes or G) RPM vs RPMA ONB gene modules per Leiden cluster. H) ScRNA-seq expression of indicated genes in mouse RPM and RPMA tumors in 4A UMAP. I) Violin plots of indicated genes per Leiden cluster. J) Normalized expression of indicated genes in human ONB-A vs -B tumors from scRNA-seq data in 4G. For violin plots each dot represents one ROI, and edgeR quasi-likelihood F-test with Benjamini Hochberg correction was performed; *p<0.05; **p<0.01, ***p<0.001, ****p<0.0001, ns = not significant.

See also Figure S7 and Table S4.

Next, we excluded immune ROIs and used unbiased Leiden clustering to identify transcriptional clusters that discriminate tumor populations. This analysis identified two distinct Leiden clusters; cluster 1 was highly enriched for high-grade ONB that spanned KI67⁺, TUBB3⁺ and pan-negative marker ROIs, whereas cluster 2 was enriched for low-grade TUBB3⁺ regions (Figure 6E). Importantly, most human ONBs had ROIs spanning both Leiden clusters (Figures 6E and S7A), suggesting frequent intratumoral transcriptional heterogeneity. MYC and other non-neuronal markers (KRT8, GRHL1, KI-67) were increased in cluster 1, which had a greater proportion of high-grade ONB; in contrast, NEUROD1 and other neuronal markers were enriched in cluster 2 (Figure 6F). Likewise, the RPM ONB gene signature was significantly enriched in cluster 2, while the RPMA ONB gene signature was significantly enriched in cluster 1 (Figure 6G). ENRICHR pathway analyses (Table S4) supported the similarities between neuronal RPM ONB and Leiden cluster 2, and non-neuronal RPMA ONB and Leiden cluster 1. SOX2, TEAD2, and FOXM1 target genes and Notch signaling were enriched in cluster 1, while REST and NEUROD1/2 target genes and neuronal/neuroendocrine pathways were enriched in cluster 2 (Table S4). These results suggest that human ONB, which presumably arises from a single cell of origin, harbors intratumoral heterogeneity in cell state with both neuronal and non-neuronal populations (Figures 6B, 6E, and S7A), similar to the RPM and RPMA ONB GEMMs.

Given the rise of FDA-approved immune checkpoint therapies, we sought to take advantage of the spatial transcriptomics platform to profile tumor infiltrating immune populations. H&E stains and CD45 IHC revealed a low degree of immune infiltration across ONB tumors (Figure S7B). Assessment of established immune markers revealed that infiltrating immune cells were primarily of the myeloid lineage, specifically macrophages (Figures 6D and S7C), consistent with prior studies⁴. Analysis of inflammatory cytokines and effectors, as well as immune checkpoints showed little to no expression within low- and high-grade tumors (Figure S7C). We observed a lack of CD3⁺ T cell infiltration in RPM and RPMA ONBs and some myeloid populations (Figures S7D–E). This suggests that mouse ONB recapitulates the immune-cold tumor microenvironment of human ONB, consistent with the immune-cold nature of SCLC and other neuroendocrine tumors.

Lastly, we examined expression of SCLC therapeutic targets potentially relevant to human ONB, including DLL3, SEZ6, BCL2, SSTR2, and UCHL1, which are enriched in neuroendocrine-high SCLC^{19,32,74–83}. Importantly, all of these genes are enriched in more neuronal RPM versus RPMA ONB tumors (Figure 6H), and are enriched in more neuronal/low-grade human ONBs—cluster 2 of the human spatial data (Figure 6I) and ONB-A of the human scRNA-seq data (Figure 6J). This analysis across multiple mouse and human tumors with distinct platforms reveals that ONB mouse models recapitulate the cell fate heterogeneity of human ONB, and that ONB has remarkable similarities to SCLC with implications for diagnosis and treatment.

DISCUSSION

ONB research has been hindered by a lack of model systems for understanding tumor development and testing therapeutic strategies. We present a mouse model of NEUROD1⁺ ONB with short latency, high penetrance, and frequent metastases, which is tractable for genetic manipulation and drug studies. While this model relies on Rb1/p53/Myc alterations and underscores the importance of Myc in tumorigenesis, Mycl is also expressed in normal OE⁸⁴ and GEMM tumors. MYCL is a SCLC oncogene expressed in neuroendocrine-high tumors^{25–27,85–87}, and although MYCL amplifications and fusions have not been reported in ONB, chr1p34 amplifications, where MYCL is located, have been noted^23,88,89. These GEM models offer a platform to investigate how genetic alterations impact ONB tumorigenesis, cell of origin, fate, and lineage plasticity.

The elusive cell of origin for ONB is speculated to be GBCs based on their transcriptional similarities⁴. GBCs, as mitotic progenitors, are vulnerable to DNA damage and genetic alterations upon infection, injury, and/or inflammation. While our experiments show that GBCs are a permissive cell of origin for ONB, it is clear from studies in SCLC that genetic alterations can influence tumor permissivity—a concept termed “oncogenic competence”⁹⁰. For instance, MYCL-associated SCLC arises mainly in lung neuroendocrine cells, while club cells are relatively refractory⁹¹. However, in MYC-driven SCLC, club cells become as susceptible as neuroendocrine cells^40,53—suggesting context-dependent oncogenic competence. Future studies with cell type-specific promoters can clarify the impact of genetic alterations on ONB cell-of-origin.

In contrast to SCLC mouse models^32,53, ASCL1 loss in ONB promotes a POU2F3⁺ microvillar-like state, suggesting a lineage relationship between ASCL1⁺ GBCs and POU2F3⁺ microvillar cells, as observed in normal OE renewal^13,16. Despite similar states (ASCL1⁺, NEUROD1⁺, POU2F3⁺) in SCLC, a comparable relationship between ASCL1 and POU2F3 is yet to be observed in lung tumors. We speculate that an equivalent cell of origin to the olfactory GBC has not been tested in the context of SCLC, i.e. multipotent basal cells^92,93. Lung basal cells were not targeted in Olsen et al.⁵³ or Chen et al.⁴⁰, so future studies should address whether basal cells can adopt a tuft cell fate during SCLC tumorigenesis. A preliminary study suggests that tuft cells can transdifferentiate to neuroendocrine cells in pancreatic cancer models⁹⁴, but the mechanisms are unknown. Our POU2F3⁺ ONB model could shed light on POU2F3-driven microvillar/tuft-like tumors, potentially applicable to multiple cancer types.

Our data suggest ONB can evolve from neuronal to non-neuronal phenotypes, which may influence tumor progression or resistance to NE-targeted therapies. SCLC subtypes exhibit plasticity, transitioning from ASCL1⁺ to NEUROD1⁺ or other non-NE states, impacting therapeutic resistance^{29,33–36,95–98}. Our barcode-based lineage analysis indicates similar plasticity in ONB GEMMs, consistent with a multipotent ASCL1⁺ state that precedes NEUROD1 expression during OE neurogenesis. This plasticity may explain a proportion of histological heterogeneity in human sinonasal tumors, challenging classification, and suggesting that tumor types may exist on a spectrum of differentiation. Insights from prostate, lung, and olfactory systems could unveil conserved plasticity mechanisms for therapeutic exploitation.

Although SCLC and ONB are considered immune-cold tumors^{4,20,39,99–102}, recent studies demonstrate that SCLC can exhibit a more inflamed phenotype with improved immunotherapy responses^29,103–106. Future studies on ONB’s immune landscape, in both mouse models and human samples, are warranted to elucidate the immune similarities and differences between ONB and other neuroendocrine tumors, and how these may impact immunotherapy approaches.

ONB has historically been classified based on pathology rather than molecular information. Our findings reveal significant similarities between ONB and prostate and lung neuroendocrine tumors, prompting consideration of whether the name ONB accurately reflects the molecular identity of these tumors. Comparative studies with adrenal neuroblastoma (NB) cell lines show shared transcriptional and proteomic similarities with SCLC¹⁰⁷, though lacking POU2F3 expression found in subsets of human SCLC, NEPC, and ONB. While both SCLC and NBs often rely on MYC family oncogenes, NB tends to express MYCN, while SCLC and ONB more commonly utilize MYCL or MYC. ALK-activating mutations are common in NB, but so far have been extremely rare in ONB and SCLC^20,108. Both NB and SCLC have been linked to neural crest lineages^{3,53,109–112}, and MYCN is sufficient to drive NB in neural crest stem cells in zebrafish and mouse^113,114. In contrast, targeting non-neural-crest cells in the lung epithelium leads to SCLC, and in the OE leads to ONB. Our collective findings suggest ONB’s closer resemblance to SCLC than other NBs. This highlights the relevance of testing SCLC targets like DLL3 or SEZ6^115–119 in ONB and considering ONB in basket trials with other neuroendocrine malignancies.

Limitations of the study

A significant challenge in ONB research stems from the dearth of human samples for molecular analysis, hindering our understanding of the frequency and co-occurrence of genetic alterations such as TP53, RB1, and MYC. Our study describing ONB GEMMs offers a platform to systematically interrogate the impact of specific genetic alterations on ONB phenotype.

While our spatial transcriptomic dataset lacks single cell resolution, it, along with human scRNA-seq samples and IHC analyses, supports the presence of molecular heterogeneity in both mouse and human ONB; given that human tumors presumably arise from a single cell, this suggests that human ONB exhibits lineage-related plasticity.

While evidence from reporter mice, GBC isolation, and ASCL1-knockout suggests GBCs as a permissive cell of origin for ONB, Ad-Cgrp-Cre occasionally recombines in other cell types. Other OE cell types, akin to SCLC^53,91, might initiate ONB. Development of precise adeno-Cre tools or sorting non-GBCs is essential for dissecting their contribution to ONB initiation, progression, and phenotype. Additionally, OE damage with MMZ and/or ex vivo culture may influence GBC fate trajectories.

Our barcode-based lineage tracing reveals GBC-derived tumor cell plasticity between neuronal and non-neuronal states. It remains unclear whether a highly-plastic, self-renewing GBC/INP-like pool continuously generates differentiated tumor cells, or if more-differentiated cells retain plasticity and can de- or trans-differentiate to other cell states. Future studies with additional rounds of CellTagging or alternative lineage-tracing methods may provide insights into the dynamics of ONB plasticity.

STAR METHODS

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Trudy G. Oliver (tgo@duke.edu).

Materials availability

There are limitations to the availability of RPM/A GBC-derived organoid lines generated in this study due to their derivation from primary cells in the olfactory epithelium. Distribution of organoid cultures requires payment for processing and shipment, and completion of a material transfer agreement. Human ONB tissue used in this study is not available due to sample scarcity.

Data and code availability

Mouse and human single-cell RNA-seq and de-identified human spatial transcriptomics data have been deposited at GEO under Superseries GSE244123 and will be publicly available on the date of publication. GEO Subseries accession numbers matching each data type deposit are listed in the key resources table. All original code has been deposited on GitHub and at Zenodo (DOI: 10.5281/zenodo.10829939) and is publicly available as of the date of publication. DOIs and Github links are additionally listed in the key resources table. Any other information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
APC anti-KIT	eBioscience	Cat#17-1171-82 RRID:AB_469430
ASCL1 [EPR19840]	Abcam	Cat#ab211327 RRID:AB_2924270
Beta III Tubulin (TUBB3)/TUJ1	Abcam	Cat#ab18207 RRID:AB_444319
Beta III Tubulin (TUBB3)/TUJ1: for IF costains	Biolegend	Cat#801201 RRID:AB_2728521
CD11B	Abcam	Cat# ab133357 RRID:AB_2650514
CD3	Abcam	Cat#ab5690 RRID:AB_305055
CGRP	Sigma	Cat#C8198 RRID:AB_259091
Chromogranin A (CHGA)	Novus Biologicals	Cat#NB120-15160 RRID:AB_789299
TRP63 (P63)	R&D Systems	Cat#599-AF1916 RRID:AB_2207174
KIT	Cell Signaling Technology	Cat#3074 RRID:AB_1147633
HES1	Cell Signaling Technology	Cat#11988 RRID:AB_2728766
LHX2	Sigma	Cat#ABE1402 RRID:AB_2722523
mCherry (tdTomato)	Sigma	Cat#AB356481 RRID:AB_2861426
MKI67: for mouse IHC	BD Biosciences	Cat#556003 RRID:AB_396287
NCAM1	R&D Systems	Cat#AF2408 RRID:AB_442152
NEUROD1 [EPR4008]: for mouse IHC	Abcam	Cat#ab109224 RRID:AB_10861489
NEUROD1: for IF	Cell Signaling Technology	Cat#4373 RRID:AB_10549071
NEUROD1: for IF costains	R&D Systems	Cat#AF2746 RRID:AB_2149217
NKX2-1 (TTF1)	Abcam	Cat#ab76013 RRID:AB_1310784
Pan-Cytokeratin (Pan-CK)	Thermo Fisher Scientific	Cat#BS1712R RRID:AB_10855057
POU2F3	Sigma	Cat#HPA019652 RRID:AB_1855585
S100B	Abcam	Cat#ab4066 RRID:AB_304258
SOX2	Cell Signaling Technology	Cat#3728 RRID:AB_2194037
SOX9: for IF	Cell Signaling Technology	Cat#82630 RRID:AB_2665492
SOX9: for IHC	Abcam	Cat#ab185966 RRID:AB_2728660
Synaptophysin (SYP)	Cell Signaling Technology	Cat#36406S RRID:AB_2799098
UCHL1	Sigma	Cat#HPA00593
YAP1	Cell Signaling Technology	Cat#14074 RRID:AB_2650491
Peroxidase AffiniPure Donkey Anti-Rabbit IgG (H+L)	Jackson ImmunoResearch	Cat#711-035-152; RRID: AB_10015282
Peroxidase AffiniPure Goat Anti-Mouse IgG1, Fcg Subclass 2b Specific	Jackson ImmunoResearch	Cat#115-035-205; RRID: AB_2338513
CD45-532: for Spatial Transcriptomics	Novus Biologicals	Cat#NBP2-34528AF532 RRID:AB_2864384
TUBB3-594: for Spatial Transcriptomics	Biolegend	Cat#657408 RRID:AB_2565285
KI67-647: for Spatial Transcriptomics	Cell Signaling Technology	Cat#12075 RRID:AB_2728830
Bacterial and Virus Strains
Ad5-CGRP-Cre	University of Iowa Viral Vector Core Facility	Cat# VVC-Berns-1160
Ad5-CMV-Cre	University of Iowa Viral Vector Core Facility	Cat# VVC-U of Iowa-5
CellTag Barcode Library V1	Addgene	RRID:Addgene_115643
Biological Samples
Human ONB tissue	Duke University Hospital	N/A
Mouse tissues	This paper	N/A
Chemicals, Peptides, and Recombinant Proteins
Normal Goat Serum	Jackson Immunoresearch	Cat#005-000-121
Normal Donkey Serum	Jackson Immunoresearch	Cat#017-000-001
10% Neutral Buffered Formalin	Fisher Scientific	Cat#22-110-869
Ethanol (200 Proof)	VWR	Cat#TX89125172DUK
Hydrogen Peroxide	Fisher Scientific	Cat#H325-500
Tween-20	Fisher Scientific	Cat#BP337-500
ACK Lysing Buffer	Thermo Fisher Scientific	Cat#A10492
Collagenase Type 1A	Sigma	Cat#C9891
Collagenase, Type 4	Worthington Biochemical	Cat#LS004186
Dispase	Worthington Biochemical	Cat#LS02104
DMSO	Fisher Scientific	Cat#BP231-100
CaCl2	Sigma	Cat#C5670
MEM	Thermo Fisher Scientific	Cat#11095080
Formaldehyde (37% by weight)	Thermo Fisher Scientific	Cat#BP531-500
HBSS-free media	Thermo Fisher Scientific	Cat#14175
HBSS without calcium or magnesium	Thermo Fisher Scientific	Cat#14025
Trypsin-EDTA	Thermo Fisher Scientific	Cat#25200-072
Leibovitz’s L15 media	Thermo Fisher Scientific	Cat#11415-064
Fetal Bovine Serum (FBS)	Sigma	Cat#12303C
DNase	Sigma	Cat#D4527
Sucrose	VWR	Cat#97061-428
SYTO 13 nuclear dye	Thermo Fisher Scientific	Cat#S7575
Methimazole (MMZ)	Sigma	Cat#M8506
DNase I	Stem Cell Tech	Cat#07900
Papain	Stem Cell Tech	Cat#07465
Neurocult media	Stem Cell Tech	Cat#05700
4% paraformaldehyde	Sigma Aldrich	Cat#158127
OCT	VWR	Cat#25608-930
Antigen Unmasking Solution, Citrate Based	Vector Laboratories	Cat# H-3300-250
BSA	VWR	Cat#97061-422
Triton X-100	Chem Impex	Cat#1279
Hoescht stain	Thermo Fisher	Cat#62249
Vectashield	Vector Laboratories	Cat#H-1000
TAT-Cre Recombinase	EMD Millipore	Cat#SCR508
SPRIselect	Beckman Coulter Life Sciences	Cat#B23317
Advanced DMEM/F-12	Fisher Scientific	Cat#12-634-028
L-glutamine	Thermo Fisher	Cat#35050079
Penicillin-Streptomycin (Pen-strep)	Thermo Fisher	Cat# 15140163
Y27632 (Rho-kinase inhibitor)	MedChemExpress	Cat# HY-10071
TrypLE	Thermo Fisher	Cat#12604013
DNeasy DNA Isolation Kit	Qiagen	Cat#69506
GoTaq	Promega	Cat#M7123
Polybrene	Santa Cruz	Cat#sc-134220
PRR Matrigel, Phenol Red-Free	University of Utah	N/A
L-WRN conditioned organoid media	University of Utah	N/A
Critical Commercial Assays
DAB Peroxidase (HRP) Substrate Kit	Fisher Sci	Cat#NC9276270
VECTASTAIN ABC Kit (Rabbit IgG)	Vector Laboratories	Cat#PK-4001
SignalStain Boost IHC Detection Reagent (HRP, Rabbit)	Cell Signaling Technology	Cat#8114
SignalStain Antibody Diluent	Cell Signaling Technology	Cat#8112
Mouse on Mouse (M.O.M) Basic Kit	Vector Laboratories	Cat#BMK-2202
TSA Fluorescein System	Akoya Biosciences	Cat#NEL701A001KT
Chromium Single Cell 3’ Library & Gel Bead Kit v3.1	10X Genomics	Cat#PN-1000268
Chromium Single Cell Controller	10X Genomics	Cat#PN-120263
Chromium Next GEM Chip G Single Cell Kit	10X Genomics	Cat# 1000120
GeoMx Human Whole Transcriptome Atlas Probe Mix	NanoString	Cat#GMX-RNA-NGS-HuWTA-4
APC selection kit	Stem Cell Tech	Cat#17667
10X Magnetic Separator	Fisher Scientific	Cat#NC1469069
Plasmid Plus Mega Kit	Qiagen	Cat#12981
Deposited Data
Single cell RNA-Seq normal unlesioned mouse OE	Horgue et al. 2022	GEO: GSE185168
Single cell RNA-seq lesioned mouse OE dataset	Ko et al, 2023	GEO: GSE224894
ASCL1 conserved targets by ChIP-seq	Borromeo et al. 2016	GEO: GSE69398
NEUROD1 conserved targets by ChIP-seq	Borromeo et al. 2016	GEO: GSE69398
NEUROD1 ChIP-seq data from RPM mouse tumors	Olsen et al. 2021	GEO: GSE155692
Single cell RNA-seq RPM ONB	This study	GEO Superseries: GSE244123 (Subseries GSE244122)
Single cell RNA-seq RPMA ONB	This study	GEO Superseries: GSE244123 (Subseries GSE244122)
Single cell RNA-seq RPM and RPMA GBC-derived allografts	This study	GEO Superseries: GSE244123 (Subseries GSE244119)
Spatial transcriptomics human ONB	This study	GEO Superseries: GSE244123 (Subseries: GSE244117)
Single cell RNA-seq human ONB-B	This study	GEO Superseries: GSE244123 (Subseries: GSE248746)
RNA-seq from 19 human ONBs	Classe et al., 2018	GEO: GSE118995
RNA-seq from 21 prostate neuroendocrine carcinomas	Labrecque et al., 2019	GEO: GSE126078
RNA-seq from 16 small cell lung cancers	Rudin et al., 2012	EGAS00001000334
RNA-seq from 20 medulloblastomas	Northcott et al., 2017	N/A
RNA-seq from 3 normal human OEs	Olender et al., 2016	GEO: GSE80249
Single cell RNA-seq from a human ONB (ONB-A)	Zunitch et al., 2023	GEO: GSE166612
Single cell RNA-seq atlas from 7 human patients	Finlay et al., 2022; Durante et al., 2020; Oliva et al., 2022	GEO: GSE139522; GEO: GSE184117
GitHub and Zenodo deposit of all source code	This study	GitHub: https://github.com/Goldstein-Lab/Finlay_and_Ireland_et_al_mouse_ONB Zenodo: DOI: 10.5281/zenodo.10829939
Experimental Models: Organisms/Strains
Mouse: Rb1fl/fl;Trp53fl/fl;MycT58ALSL/LSL (RPM)	Trudy G. Oliver, Duke University, Mollaoglu et al, 2017	The Jackson Lab RRID:IMSR_JAX:029971
Mouse: Rb1fl/fl;Trp53fl/fl; MycT58ALSL/LSL;Ascl1fl/fl (RPMA)	Trudy G. Oliver, Duke University, Olsen et al, 2021	N/A
Mouse: Rb1fl/fl;Trp53fl/fl;MycT58ALSL/LSL-Cas9 (RPM-Cas9 or RPM-GFP)	Trudy G. Oliver, Duke University, Ireland et al, 2020	N/A
Mouse: C.B-Igh-1b/GbmsTac-Prkdcscid-Lystbg N7 (Scid/Beige)	Gift of Zachary C. Hartmann, Duke University	Taconic RRID:IMSR_TAC:CBSCBG
Mouse: B6.Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze/J (R26R-Ai9 tdTomato Reporter)	The Jackson Laboratories	The Jackson Lab RRID:IMSR_JAX:007909
Mouse: Rb1fl/fl;Trp53fl/fl;Rbl2fl/fl	Schaffer et al., 2010	RRID: MMRRC 043692-UCD
Cell Line: HEK 293T/17	ATCC	Cat#CRL-11268 RRID:CVCL_1926
Tumor organoids: KIT+ GBCs from Rb1fl/fl;Trp53fl/fl; MycT58ALSL/LSL;Ascl1fl/fl (RPMA) (RPMA) and Rb1fl/fl;Trp53fl/fl; MycT58ALSL/LSL (RPM) olfactory epithelium	This study	N/A
Software and Algorithms
Graphpad Prism	Graphpad Software	www.graphpad.com/scientific-software/prism/
Quantum GX2 mCT Software	PerkinElmer	N/A
Analyze 11.0	AnalyzeDirect	https://analyzedirect.com/
Python (v3.8.8)	Python Software	https://www.python.org/
Scanpy (v1.9.1)	Wolf et al., 2018	https://scanpy.readthedocs.io/en/stable/
scvi-tools (v0.17.4)	Gayoso et al., 2022	scvi-tools.org
otscomics (v0.1.0)	Huizing et al., 2022	https://ot-scomics.readthedocs.io/en/latest/
Osn package (v0.0.1)	Tsukahara et al., 2021	https://github.com/dattalab/Tsukahara_Brann_OSN/tree/main/osn
10X Genomics Cell Ranger v7.0.0	Zheng, et al. 2017	https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/using/tutorial_ov
Gene Set Enrichment Analysis (GSEA)	Broad Institute, and the Regents of the University of California	http://software.broadinstitute.org/gsea/index.jsphttp://software.broadinstitute.org/gsea/index.jsp
Enrichr	Kuleshov et al., 2016	http://amp.pharm.mssm.edu/Enrichr/http://amp.pharm.mssm.edu/Enrichr/
Seurat	Butler et al., 2018; Stuart and Butler et al., 2019	https://satijalab.org/seurat/https://satijalab.org/seurat/
R Statistical Programming	The R Foundation	www.r-project.orgwww.r-project.org
edgeR (v3.40.2)	Robinson et al., 2010	https://bioconductor.org/packages/release/bioc/html/edgeR.html
BioRender	BioRender.com	https://app.biorender.com
Velocyto (v0.17.17)	La Manno et al. 2018	https://velocyto.org/
UCSC Genome Browser	Kent et al. 2002	https://genome.ucsc.edu/
scVelo (v0.2.5)	Bergen et al. 2020	https://pypi.org/project/scvelo/
NanoStringNCTools (v1.6.1)	Aboyoun et al., 2021	https://bioconductor.org/packages/release/bioc/html/NanoStringNCTools.html
GeomxTools (v3.2.0)	Ortogero et al. 2023	https://www.bioconductor.org/packages/release/bioc/html/GeomxTools.html
GeoMxWorkflows (v1.4.0)	Reeves et al. 2023	https://www.bioconductor.org/packages/release/workflows/html/GeoMxWorkflows.html
CellTagR	Kong et al. 2020	https://github.com/morris-lab/CellTagR
TCGAbiolinks (v2.25.3)	Colaprico et al., 2015	https://www.bioconductor.org/packages/release/bioc/html/TCGAbiolinks.html
ImageJ (v2.3.0)	Reuden et al. 2017	https://imagej.net/software/imagej2/
NCI GDC Data Portal	National Cancer Institute (NCI)	https://gdc.cancer.gov/access-data/gdc-data-portal

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Mice

Rb1^fl/fl;Trp53^fl/fl;H11b-LSL-Myc^T58A/^T58A-Ires-Luciferase (RPM) (JAX #029971) mice⁴¹ and Rb1^fl/fl;Trp53^fl/fl;Rbl2 ^fl/fl (RPR2) (RRID:MMRRC #043692-UCD) mice⁵⁵ have been previously described. We crossed RPM mice with Rosa26-LSL-Cas9-Ires-Gfp mice¹²⁰ (JAX # 024857) to generate RPM-Cas9 mice with Cre-dependent expression of Cas9 and Egfp³³. RPMA mice were previously described⁵³, and the tdTomato Cre-reporter strain R26R-Ai9 is commercially available (JAX#007909, RRID:IMSR_JAX:007909). SCID/beige mice (CBSCGB) are available from Taconic.

All mice were housed and treated according to regulations set by the Institutional Animal Care and Use Committee of Duke University. Viral infections were performed in a Biosafety Level 2+ room following guidelines from Duke University Institutional Biosafety Committee. Male and female mice were distributed equally for all experiments.

GBC-derived organoid cultures

GBC-derived organoid cultures from RPM and RPMA mice were obtained and transformed ex vivo (see METHOD DETAILS). Organoid lines were determined to be free of pathogens by IDEXX 18-panel mouse pathogen testing and confirmed mycoplasma-negative before implantation to SCID/beige hosts.

Patient tissue for spatial transcriptomic analyses, immunostaining, and scRNA-seq

Archival tissue from patients undergoing surgical resection of ONB between 2010 and 2023 at the Duke University Hospital was considered for profiling via the NanoString GeoMx spatial transcriptomic platform and/or scRNA-seq. Candidate FFPE tissue samples were sectioned into 5 μm sections and stained with H&E for review by a pathologist (J.N.). For spatial transcriptomics and immunostaining, tissue sections with minimal/no tumor or with significant artifact were excluded from analysis leaving 12 ONB tissue specimens. For single-nucleus transcriptomic analyses, patient consent was obtained and ONB tissue was collected in the operating room, processed by the Duke Biorepository and Precision Pathology Center, and flash frozen for downstream processing. Patient demographics, including age, sex, tumor location, tumor grade, and tumor stage, were collected for all human samples. All tissue samples and patient data were collected and stored in accordance with an approved Institutional Review Board protocol for this project.

METHOD DETAILS

Intranasal infections

Anesthetized mice at 6–8 weeks of age were infected by intranasal instillation with 1×10⁸ plaque-forming units of Ad5-Cgrp-Cre adenovirus (University of Iowa) using methods based on¹²¹, but modified to promote formation of lesions in the olfactory epithelium. Briefly, mice were anesthetized with isoflurane at a flow rate of 20–25ml/hr, depending on the size and sex of the mouse, with the assumption that increased isoflurane may reduce the rate of breathing and virus administered to the lungs. The optimal breathing rate was approximately one breath every 2–3 sec. Mice were held in a supine position and 20 μl of viral cocktail (0.3 μL CaCl₂ (Sigma cat# C5670, resuspended to 2M in sterile water) + 1×10⁸ pfu adenovirus + MEM (Thermo cat# 11095080) up to 20 μL total volume per mouse) was administered via a P20 pipette, alternating between the left and right naris for each drop. Mice were maintained in supine position for ~30 sec post-infection and monitored until they regained full motility and recovered from anesthesia.

MicroCT imaging

To monitor tumor development, mice were imaged beginning four weeks after Ad-Cgrp-Cre administration for RPM mice, or after 8 weeks for RPMA mice, and every two weeks thereafter. Mice were anesthetized with isoflurane and imaged using a small animal Quantum GX2 microCT (Perkin Elmer). Quantum GX2 images were acquired with 18 sec to 2 min scans at 45 μm resolution, 90 kV, with 88 mA of current. Mice were sacrificed when tumor burden resulted in any difficulty breathing or significant weight loss.

Immunohistochemistry

Tissues were fixed in 10% neutral buffered formalin for 24 hr at room temperature (RT), washed in PBS and transferred to 70% ethanol. Formalin-fixed paraffin embedded (FFPE) sections at 4–5 μm were dewaxed, rehydrated, and subjected to high-temperature antigen retrieval by boiling 20 min in a pressure cooker in 0.01 M citrate buffer at pH 6.0 (Tris-EDTA buffer at pH 9.0 for pan-CK staining). Slides were quenched of endogenous peroxide in 3% H2O2 for 15 min, then blocked in 5% goat serum in PBS/0.1% Tween-20 (PBS-T) for 1 hr, and then stained overnight with primary antibodies in blocking buffer (5% goat serum or SignalStain antibody diluent, Cell Signaling Technology (CST) cat# 8112). For non-CST primary antibodies, an HRP-conjugated secondary antibody (Vector Laboratories) was used at 1:200 dilution in PBS-T, incubated for 45 min at RT followed by DAB staining (Vector Laboratories). All samples were stained simultaneously with equivalent DAB conditions. Alternatively, CST primary antibodies were detected using 150 uL of SignalStain Boost IHC Detection Reagent (CST cat# 8114). All staining was performed with Sequenza cover plate technology. The primary antibodies include: ASCL1 (Abcam cat# 211327) 1:300; CD11B (Abcam cat# ab133357) 1:2000; CD3 (Abcam cat# ab5690) 1:100; Chromogranin A (CHGA, Novus Biologicals, cat# NB120–15160) 1:500; HES1 (CST cat# 11988) 1:300; LHX2 (Sigma cat# 1402) 1:250; MKi67 (BD cat# 556003) 1:300; NEUROD1 (Abcam cat# 109224) 1:300; NCAM1 (R&D cat# AF2408) 1:150; NKX2–1/TTF1 (Abcam cat# ab76013) 1:250; P63 (R&D cat# AF1916) 1:400; pan-Cytokeratin/pan-CK (Thermo cat# BS1712R) 1:150; POU2F3 (Sigma cat# HPA019652) 1:300; SYP (CST cat# 36406S) 1:200; TUBB3/TUJ1 (Abcam cat# ab18207) 1:1000; UCHL1 (Sigma cat# HPA005993) 1:100; YAP1 (CST cat# 14074) 1:300. For manual H-score quantification, images were acquired on a Nikon Ci-L LED Microscope with DS-Fi3 Camera. H-score was quantified on a scale of 0–300 taking into consideration percent positive cells and staining intensity as described¹²², where H Score = % of positive cells multiplied by intensity score of 0–3. For example, a tumor with 80% positive cells with high intensity of 3 = 240 H-Score.

Tissue immunofluorescence

Mouse noses were dissected as previously described⁶¹ and fixed in 4% paraformaldehyde (Sigma Aldrich, St. Louis, MO) in PBS overnight. Noses were washed twice in PBS for 5 min each, then cryopreserved in 30% sucrose solution, 250 mM EDTA, and PBS for approximately 7 days at 4C. Noses were embedded in Optimal Cutting Temperature compound (VWR, Radnor, PA). 10 μm sections were cut and placed on charged, superfrost slides (Thermo Fisher Scientific) using a cryostat (CryoStar NX50, Thermo Fisher Scientific) and stored at −20C. Prior to staining, slides were rehydrated in PBS for 5 min, followed by 1 min in 70% ethanol, 1 min in 95% ethanol, 1 min in 100% ethanol, 1 min in 95% ethanol, 1 min in 70% ethanol, and 5 min in PBS. Antigen retrieval was performed using citrate-based antigen unmasking solution (Vector Laboratories, Newark, CA). Tissue sections were blocked for 45 min in a moist chamber at RT in blocking buffer (5% nonfat dry milk, 4% BSA, 10% normal donkey serum). Primary antibody was diluted in blocking buffer and incubated overnight at 4C in a moist chamber. Tissue sections were washed 3 × 5 min each in PBS+0.1%Triton X-100. Sections were incubated in secondary antibodies (Jackson Immunoresearch) diluted 1:100 in 5% normal donkey serum in a moist chamber at RT for 45 min. Slides were washed 3x for 5 min each in PBS+0.1% Triton X-100. Hoescht stain (Thermo Fisher) diluted 1:1000 in Millipore water was added to tissue sections for 3 min. Slides were mounted with Vectashield (Vector Laboratories) and coverslips (VWR). For antibodies that required TSA at the secondary antibody incubation step, the appropriate biotinylated antibody (Vector Laboratories) was diluted 1:250 in TSA blocking buffer (TSA Fluorescein System, Akoya Biosciences, Marlborough, MA) and added to tissue sections for 45 min at RT. Slides were washed 3 × 5 min each in PBS+0.1% Triton X-100. ABC reagent (Vectastain Elite ABC Kit cat# PK-4001) was added to slides for 10 min at RT. Slides were washed 3 × 5 min each in PBS+0.1% Triton X-100. TSA Fluorescein Reagent diluted 1:50 in amplification diluent (Akoya Biosciences) was added to slides for 5 min at RT. Slides were washed 3 × 5 min each in PBS+0.1% Triton X-100. Hoescht stain was added and slides were mounted as described above. Primary antibodies included: ASCL1 (Abcam cat# 211327) 1:300 TSA; CGRP (Sigma cat# C8198) 1:1000 TSA; CGRP (CST cat# 14959) 1:100; HES1 (CST cat# 11988) 1:200 TSA; LHX2 (Millipore cat# ABE1402) 1:400; mCherry (tdTomato, Sigma cat# AB356481) 1:500; NEUROD1 (CST cat# 4373) 1:500 TSA; NEUROD1 (R&D cat# AF2646) 1:100 TSA; POU2F3 (Sigma cat# HPA019652) 1:50; SOX2 (eBioscience cat# 14–9811-82) 1:50; SOX9 (CST cat# 82630) 1:50 TSA; S100B (Abcam cat# ab4066) 1:100; TUBB3 (Biolegend cat# 801201) 1:500. Images were acquired using a Leica DMi8 microscope system and Leica Application Suite X software (v3.7.5). Exported .lif image files were opened and processed in ImageJ (v2.3.0) , and scale bars were added to final images using metadata from .lif files.

Analysis of Cgrp-Cre-based reporter activity

R26R-Ai9 mice were infected with 1×10⁸ plaque-forming units of Ad5-Cgrp-Cre adenovirus as described above, and harvested at 0, 3, 5, or 7 days post-infection. Tissue was dissected and fixed for immunostaining as described. Olfactory epithelial cells were defined based on marker expression, morphology, and location. GBCs = ASCL1+; INP = NEUROD1+; INP/iOSN/mOSN = TUBB3+; Sus = apical SOX2+ nuclei; HBCs = basal SOX2+ nuclei; BG = SOX9+ lamina propria and morphology; MV1 = SOX9+ apical nucleus and morphology; MV2 = FOXI1+ and morphology; OECs = S100B+; trigeminal nerve fibers = CGRP+. GBC-derived cells were defined as GBCs or clusters of cell types known to be derived from GBCs and similar in pattern to our prior GBC-based lineage tracing analyses¹⁶. Non-GBC-derived cells were defined as lone cells harboring recombination events that were not GBCs. Full sections of olfactory bulb immediately adjacent to OE were fixed and stained from each mouse, which confirmed an absence of virus-induced recombination in the bulb.

10X Genomics Single Cell RNA Seq

Sample and library preparation:

One RPM-GFP ONB tumor was micro-dissected using a fluorescent dissection microscope and processed into a single-cell suspension in preparation for scRNA-seq. Three RPMA ONB tumors were micro-dissected from the olfactory region and processed for scRNA-seq. Specifically, the turbinate and septal regions were micro-dissected and abnormal-appearing tumor mass identified by microCT imaging was carefully isolated. Representative portions of tumors with enough material were set aside for FFPE and IHC staining, and remaining material was subject to dissociation for scRNA-seq. We note that scRNA-seq and/or FFPE on portions of each tumor may not capture the full spectrum of intratumoral heterogeneity; however, we capture multiple samples and/or regions and use multiplexed approaches for analysis (IHC, IF, scRNA-seq) to minimize this confounding factor. Each tumor was mechanically dissociated into small clumps using scissors, then digested to a single-cell suspension using 1 mL of an enzymatic digestion cocktail per sample for 10 min at 37°C. The digestion cocktail consisted of 4200 μL of HBSS-free media (Thermo Fisher cat# 14175), 600 μL of trypsin-EDTA (0.25%) (Thermo Fisher cat# 25200–072), 600 μL of collagenase type 4 (Worthington Biochemical cat# LS004186) from a 10 mg/mL stock prepared in HBSS without calcium and magnesium (Thermo Fisher cat# 14025), and 600 μL of dispase (Worthington Biochemical cat# LS02104). Enzymatic digestion was quenched on ice with 500 μL of quench media containing 7.2 mL of Leibovitz’s L15 media (Thermo Fisher cat# 11415–064), 800 μL of FBS (Sigma cat# 12303C), and 30 μL of DNase (Sigma cat# D4527) at 5 mg/mL in HBSS-free media. The tissue suspension was passed through a 100-μm cell strainer. Cells were spun at 2000 rpm for 5 min. Supernatant was removed and replaced with 500 μL of ACK (ammonium-chloride-potassium) lysis buffer per sample to remove red blood cell contamination (3 min incubation at 37°C; Thermo Fisher cat# A10492). Reaction was quenched with 10 mL of cold 1X PBS. Cells were spun at 1500 rpm for 5 min and resuspended in cold 0.04% BSA in 1X PBS at a concentration of 1 × 10^6 cells/mL and loaded onto a 10X chromium controller. The RPM-GFP ONB tumor was processed targeting 10,000 cells, while the RPMA ONB tumors were processed targeting 10,000 (n=1) or 5,000 (n=2) cells, as cells from these samples were more limited.

One human ONB specimen was collected immediately upon resection in the operating room in Hibernate E media (Gibco cat# A1247601). Following processing by the Duke BioRepository and Precision Pathology Core to confirm diagnosis and ensure that adequate tumor tissue remained for clinical needs, remaining portions of tumor were snap frozen and stored in liquid nitrogen for a total processing time of less than 1 hr. Single nuclei were isolated with the Chromium Nuclei Isolation Kit (10X Genomics cat# CG000505) according to defined conditions, with a nuclear lysis time of 2 min. Nuclear integrity and counts were assessed with a Cellometer Auto 2000 (Nexcelom), and nuclei were loaded onto a 10X chromium controller, targeting 1000 nuclei.

Library preparation following standard workflow with 10X 3′ chemistry, version 3.1 (10X Genomics cat# 1000268) was performed. Completed libraries were sequenced on an Illumina NovaSeq 6000 at a depth of 300 million paired reads when 10,000 cells were targeted (n=1 RPMC9, n=1 RPMA), 150 M paired reads when 5,000 cells were targeted (n=2 RPMA), or 30 M paired reads when 1,000 nuclei were targeted (n=1 human ONB) with the 10X-recommended paired end sequencing mode for dual indexed samples.

Demultiplexing and data alignment:

scRNA-seq data from the RPM-GFP and RPMA tumor sample was demultiplexed and processed into FASTQ files via the 10X Cell Ranger v7.0.0 pipeline. RPM-GFP reads were aligned to a custom mouse genome (mm10–2020-A) with eGFP, Cas9, and firefly Luciferase (fLuc) transcripts included, and RPMA tumors were aligned to a custom mouse genome (mm10–2020-A) with Venus and fLuc transcripts included. Count barcodes and UMIs were generated using cellranger count.

Quality control:

Quality control and downstream analysis were performed in Python (v3.8.8) utilizing Scanpy (v1.9.1), according to current expert recommendations for single cell best practices¹²³. Anndata objects were created from filtered feature matrices with sc.read_10x_mtx(). Initially, cells with >80,000 total counts, <2500 total counts, >8000 genes by counts, or >30% mitochondrial content were removed. Normalized count layers were created with sc.pp.normalize_total() with a target sum of 10,000. For the creation of larger integrated objects containing multiple scRNA-seq datasets, adata.concatenate() was run with join=‘outer’ to combine all anndata objects into one prior to clustering. For the creation of the integrated atlas containing normal mouse olfactory epithelium and RPM and RPMA tumors, we utilized normal unlesioned mouse OE datasets from GSE185168 and lesioned mouse OE from GSE224894⁶⁸. For the creation of the normal human olfactory epithelium integrated atlas, we utilized datasets from four patients from GSE139522¹⁴ and from three normosmic patients in GSE184117⁷¹. Transcriptomic data from a previously published ONB sample was also included (ONB-A in 4G; GSE166612)⁴.

Integration and Clustering:

For clustering, we utilized scvi-tools(v0.17.4), which has passed benchmarking standards for minimizing batch effects while maintaining true biological variability, particularly across large integrated objects¹²⁴. First, highly variable genes were determined with sc.pp.highly_variable genes() with 5000–10000 top genes, depending on the size of the dataset, flavor set to “seurat_v3”, and batch key set to the name of the specific scRNA-seq batch. Poisson gene selection was then calculated with scvi.data.poisson_gene_selection() with the same n_top_genes and batch_key utilized for the highly variable gene selection process. The probabilistic deep learning model was set up with scvi.model.SCVI.setup_anndata() with an anndata object containing only the top genes from the highly variable Poisson gene selection. Categorical covariate keys included mouse genotype and tumor, continuous covariate keys included percent mitochondrial counts, and the batch key was the same as specified above. The model was trained with default parameters, an early stopping patience of 20, and 500 max epochs, with the model.train() function. The latent representation of the model was obtained with model.get_latent_representation() and added to the .obsm of the full anndata object (including all genes). Neighbors were then calculated with sc.pp.neighbors() with use_rep set to the .obsm category added from the latent representation. sc.tl.umap() was then run with min_dist=0.5 to generate a UMAP. Finally, Leiden clusters were generated with sc.tl.leiden() with resolution set to 2.0. As is required throughout the scVI pipeline, we utilized raw count layers for all steps above.

We then performed another round of quality control by assessing n_genes_by_counts, total counts, and percent mitochondrial counts per cluster. In general, the model tends to cluster low quality and doublet cells together, so any particular clusters with exceptionally high or low average genes_by_counts, total counts, or mitochondrial content are subject to filtering out of the dataset. Prior to removing a particular low-quality cluster, we assessed gene expression based on known markers of cells in the olfactory epithelium (see Figure S4D), and ran sc.tl.rank_genes_groups() on the normalized layer with method=‘wilcoxon’ to assess marker genes for each cluster. This helped to ensure that biological cells that normally have higher or lower n_genes_by_counts, total_counts, or percent mitochondrial content (i.e. immune cells) were not aberrantly filtered out. Each time a cluster was removed, we ran the scVI pipeline on the new anndata object (starting from sc.pp.highly_variable_genes) iteratively through this quality control step until there were no longer any low-quality cell clusters in the anndata object. Additionally, each time clusters were subset out from a larger anndata object, the pipeline was re-run for optimal clustering.

Plot Generation:

UMAP plots and featureplots showing expression of specific genes were generated with sc.pl.umap(), with vmin=0, vmax=‘p99.5’, and the normalized count layer as input. Dotplots were generated with sc.pl.dotplot() using the normalized count layer and standard_scale=‘var’. In dotplots, color indicates expression level and dot size indicates abundance of expression in cells per sample. For matrix plots, the normalized count layer was scaled between 0 and 1 with sc.pp.scale() followed by .clip(0,1), which allowed for direct comparison of genes expressed at different magnitudes on a single plot. Matrix plots were then generated with sc.pl.matrixplot().

Cell Cluster Distance Analysis:

In order to determine transcriptional distance between cell types in the olfactory epithelium and mouse ONB tumors, optimal transport (OT) analysis calculating Wasserstein distance was performed, as recently described¹²⁵, using otscomics (v0.1.0) and sklearn (v1.2.2). First, an anndata object containing 3000 highly variable genes was subset from the full dataset by running scvi.data.poisson_gene_selection(), with the batch_key set to each mouse sample. Due to computational limitation, 100 cells per cell-type cluster were then randomly subsampled. Data were logarithmized with sc.pp.log1p(), and then normalized per-cell, as described¹²⁵. A cost matrix was calculated using otscomics.cost_matrix() with standard parameters. Then, the OT distance matrix was calculated with otscomics.OT_distance_matrix() with cost set to the cost matrix calculated above, an entropic regularization parameter of 0.1, and a batch size of 128. Matrix plots were generated with Matplotlib (v3.7.1) and Seaborn (v0.11.2). For hierarchical clustering of average OT distance values across each cell type, sns.clustermap() was run with default parameters.

We also utilized PCA-based clustering analysis to confirm OT transcriptional distance results. 3000 highly variable genes were identified and subset from the global anndata object using scvi tools, as described above. 30 principle components for the dataset were then calculated with sc.pp.pca() with n_comps=30, followed by PCA-based dendrogram clustering using sc.tl.dendrogram(). To generate the cluster markers heatmap shown in Figure S4F, sc.tl.rank_genes_groups() with method=‘wilcoxon’ was run to determine markers of each cell type marker. For heatmap plotting purposes, the normalized count layer was logarithmized and scaled with sc.pp.log1p() and sc.pp.scale(). 100 cells from each cell type cluster were then randomly subsampled, to allow for equal column width in the heatmap for each cell type. sc.pl.rank_genes_groups_heatmap() with n_genes=30, was then used to plot the heatmap.

Differential gene expression analysis:

We utilized edgeR (v3.40.2) to conduct differential expression analysis between RPM and RPMA tumors. From the original anndata object in Python, 3 pseudobulk samples per tumor were created by randomly subsampling 20 unique cells from each cell-type cluster, as described¹²³. The new pseudobulk anndata object contained raw counts in the .X matrix. From here, a pseudobulk samples by gene name count matrix was exported from Python and loaded into R (v4.2.2). Lowly expressed genes with <1 counts per million were filtered out. Raw pseudobulk counts were then normalized with calcNormFactors(), and dispersion was estimated with a genewise negative binomial generalized linear model with quasi-likelihood tests, by running glmQLFit() followed by glmQLFTest(), with default parameters. Differentially-expressed genes (DEGs) were then calculated with topTags() using a Benjamini-Hochberg correction. DEGs were plotted as a volcano plot using ggplot2 (v3.4.2).

ScRNA-seq gene signatures and gene set enrichment analysis:

Gene sets in Figure 4 include human ONB, SCLC, NB, and LUAD module scores derived from the top 500 marker genes of those tumor types in the pan-cancer bulk RNA-seq analysis from Figure 1 (Table S2) and SCLC subtype A, N, and P signatures obtained from Chan et al³⁰. Human genes were converted to mouse orthologs with the gorth() function in R from gprofiler2 (v0.2.2). Module scores for gene sets were then calculated with sc.tl.score_genes() with default parameters, and were projected onto the global UMAP plot and bar graphs. RPM and RPMA ONB scores were generated based on the top 100 DEGs between primary tumors by scRNA-seq analysis (Table S2). Human orthologs were taken from these lists to apply to human scRNA-seq and transcriptomic data using sc.tl.score_genes(). Statistical differences between RPM and RPMA tumor module scores were calculated with Mann-Whitney with two-sided Bonferroni correction across all cells within the RPM and RPMA tumor clusters.

Gene sets for “NEUROD1 targets by ChIP-seq” represent conserved transcriptional targets identified by ChIP-seq. The NEUROD1 conserved target gene list was derived from published ChIP-seq data from human SCLC cell lines (GEO: GSE69398)³² and NEUROD1 ChIP-seq data from RPM mouse tumors (GEO: GSE155692)⁵³. As described above, gene lists were then fed into sc.tl.score_genes() to calculate module scores and were similarly projected on UMAPs and bar graphs.

Similarly, cell cycle profiling was performed in Scanpy. Specifically, published human G1, S, and G2/M gene sets¹²⁶ were converted to mouse orthologs, as described above. Then sc.tl.score_genes_cell_cycle() was run with default parameters. UMAP plots were generated in Scanpy as described above.

Gene set enrichment analysis to compare RPM and RPMA tumors was performed with Enrichr via gseapy (v1.0.4) in Python. From the set of marker genes for RPM and RPMA tumors (from sc.tl.rank_genes, described above), the top 300 genes from each cluster were subset out and run through Enrichr using gp.enrichr() with organism=‘mouse’ and the following gene sets: ‘KEGG_2019_Mouse’, ‘Reactome_2022’, ‘GO_Biological_Process_2021’, ‘MSigDB_Hallmark_2020’, ‘NCI-Nature_2016’, ‘Tabula_Muris’¹²⁷. To generate dotplots showing the top 5 terms from each dataset, dotplot(enr.results) was run. Differentially expressed genes in the analysis of transcriptional overlap between RPM SCLC and ONB in Figure S4I are included in Table S2.

Spatial Transcriptomics:

Sample processing and data collection:

We performed multiplexed and spatially resolved transcriptomic profiling using the NanoString GeoMx digital spatial profiling (DSP) platform, with mostly default parameters¹²⁸. Briefly, 5 μm sections from high quality archival ONB tissue blocks were mounted on glass slides for profiling. Two adjacent sections were cut and positioned on separate slides to allow for side-by-side H&E comparison. Following deparaffinization, antigen retrieval was performed with Tris-EDTA buffer (pH 9.0) in a steamer for 20 min. Slides were then incubated for 15 min at 37°C in PBS buffer with 0.1 μg/mL proteinase K for RNA target exposure. In situ hybridization was conducted by incubating slides with GeoMx Human Whole Transcriptome Atlas Probe Mix (NanoString) overnight at 37°C. Slides were then stained with fluorescence-conjugated morphology markers for two hrs at RT with the following labels: SYTO 13 nuclear dye (NanoString) 250nM, CD45–532 (NovusBio cat# NBP2–34528AF532) 7.50 μg/mL, TUBB3–594 (Biolegend cat# 657408) 1:50, KI67–647 (CST cat# 12075) 1:100. Slides were then loaded on the NanoString GeoMx DSP instrument, and 2–3 representative ROIs in spatially distinct locations were selected based on tumor morphology and antibody staining markers. In general, ROIs were 200–600 μm in diameter, and contained 100–300 nuclei. Note, some tumor specimens also contained normal overlying OE, which was captured for controls. ROIs were then segmented based on morphology markers, and RNA probe oligos were photocleaved and collected in the following order: CD45, KI-67, TUBB3, pan-negative. Oligos were then processed for library preparation according to the standard Nanostring protocol, with the exception that beads were dried for 1 min instead of 5. DNA libraries were sequenced on an Illumina NovaSeq 6000 S-Prime flow cell with a read length of 2×27bp and a 5% PhiX spike-in by volume. Output fastq files from sequencing runs were converted to .dcc using the GeoMx NGS Pipeline with default parameters.

Quality control:

Downstream analyses were performed in R (v4.2.2) and Python. First, filtering of low-quality ROIs was performed utilizing NanoStringNCTools (v1.6.1), GeomxTools (v3.2.0) and GeoMxWorkflows (v1.4.0), with a standard pipeline and parameters. A data object containing all ROI segments was created from .dcc files with the readNanostringGeoMxSet() function. Expression counts with a value of 0 were shifted to 1 with shiftCountsOne(), allowing for downstream transformation. Filtering of low quality probes was performed with setBioProbeQCFlags() with default settings (minProbeRatio=0.1, percetFailGrubbs=20, removeLocalOutliers=True). Next, the limit of quantification for each segment and gene was determined, with a cutoff of 1.25 and a minLOQ of 1.25. Segments with a gene detection rate <15% were removed, leaving 82 total segments across all samples. Following filtering of low quality ROIs, background correction on raw data was performed with CountCorrect (v0.1) in Python¹²⁹.

ROI Segment Analyses:

Background corrected raw data from ROI segments were quantile normalized, as has been shown to be the superior normalization method for Nanostring GeoMx whole transcriptome data¹³⁰. To assess unsupervised clustering of all ROIs, a UMAP was generated using umap(t(log2(assayDataElement()))) and plotted with ggplot2 (v3.4.1). For differential gene expression analysis, edgeR (v3.40.2) was utilized, treating each ROI segment as a bulk RNA-seq sample. Additional filtering based on CPM was not performed due to the lower counts observed in Nanostring GeoMx data. A quasi-likelihood F-test generalized linear model was run, with Benjamini Hochberg corrections for each comparison between low- and high-grade tumors (ie. for TUBB3⁺, KI-67⁺, and pan-negative segments separately). Output results tables were generated and written to .csv files, and volcano plots were produced with ggplot2. In order to query gene set scores and plot specific genes, an anndata object in Python was created using the quintile normalized, background corrected counts matrix. To calculate gene signature scores for specific cell types, the following genes were used based on published markers derived from scRNA-seq data^13,14,58,70: GBC (HES6, KIT, CXCR4, ASCL1, SOX2); INP (LHX2, EBF1, SOX11, NEUROD1); neuronal (OLIG2, GNG8, EBF4, TUBB3, CHGA, SYP, INSM1); non-neuronal (KRT8, KRT18, SOX9, POU2F3, FOXI1); myeloid (CD68, C1QA, C1QB, C1QC); and lymphoid (CD3E, CD3G, CD4, CD8A, MS4A1). Module scores based on these gene sets were calculated with sc.tl.score_genes(), and plotted with Seaborn (v0.11.2) violinplots. Statistical differences between gene signature score per ROI segment were calculated with Mann-Whitney test with two-sided Bonferroni correction. Seaborn violin plots were also used to plot normalized expression of specific genes across ROI segment.

To generate unbiased clusters of tumor ROIs specifically, CD45+ ROIs were first filtered out. Next, to regress out batch effects associated with individual ROIs for downstream clustering, an scvi model was trained using scvi-tools (v0.17.4). Background corrected raw count matrices were filtered down to 3,000 highly variable genes, identified with scvi.data.poisson_gene_selection(). For model setup, the slide on which an ROI was located was used as a categorical covariate, and the area and number of cells within an ROI were used as continuous covariate keys. The batch key was set to the tumor ID. Following model training with 500 maximum epochs and an early stopping patience of 20, the latent representation of the model was transferred to the .obsm metadata of the full anndata object containing all genes. This latent representation was then used as the indicated representation for calculating neighbors with sc.pp.neighbors() with n_neighbors=10. Leiden clusters were generated with sc.tl.leiden() with a resolution of 0.6. Clusters with an average total counts of less than 25,000 were excluded due to poor quality, leaving 48 tumor-specific ROIs across two clusters. Cluster markers for each Leiden cluster were generated with sc.tl.rank_genes_groups() using the background corrected, quantile normalized counts and method=Wilcoxon. Individual gene expression across Leiden clusters was analyzed as described above with Mann-Whitney tests and Seaborn violin plots. To analyze gene set scores from mouse RPM and RPMA ONB, the top 50 genes identified in RPM vs RPMA tumors by edgeR differential expression analyses (see above), were first converted to human orthologs using gprofiler2. Human ortholog gene lists were used as input to calculate gene set scores across each Leiden cluster with sc.tl.score_genes() using the quintile normalized counts. Statistics and plotting were performed similarly to analysis of individual genes, described above. This approach was also applied to calculate RPM and RPMA gene set scores in human scRNA-seq datasets comparing ONB-A and ONB-B.

Bulk RNA-Seq Atlas:

TCGA and TARGET bulk RNA-seq datasets were downloaded from the NCI GDC Data Portal into R utilizing TCGAbiolinks (v2.25.3)¹³¹. Due to dataset size constraints, 15 samples per cancer type were randomly selected for download from the database. Specifically, raw counts from bulk RNA-seq count matrices were downloaded with GDCquery(). TCGA and TARGET datasets were then concatenated with bulk RNA-seq samples from 19 ONB (GSE118995), 21 prostate neuroendocrine carcinoma (GSE126078), 16 SCLC (EGAS00001000334), 20 medulloblastoma⁵², and 3 normal human OE (GSE80249). From here, edgeR was used to process data. Genes with less than 100 cpm were removed, and counts were normalized with calcNormFactors(). A generalized linear model was used to estimate dispersion by running estimateGLMCommonDisp() followed by estimateGLMTrendedDisp(), with method=bin.spline. Using dispersion values from the model, 1000 highly variable genes were selected. Next, a dissimilarity matrix between bulk RNA-seq samples was calculated with dist() and method=euclidean. Hierarchical clustering was then performed on this dissimilarity matrix by running hclust() with method=complete. The hierarchical clustering was converted to a dendrogram for plotting with as.dendrogram(). A circus plot was generated using circlize (v0.4.15) and dendextend (v1.17.1) with circlize_dendrogram(). To generate UMAPs and Leiden clusters, normalized count matrices and metadata were exported from R, and used to create an anndata object in Python. The counts matrix was then filtered down to the same 1000 highly variable genes used to plot the dendrogram above. sc.pp.pca() was run with 60 principle components, followed by sc.pp.neighbors() with n_neighbors=10 and all 60 pcs. Finally, Leiden clusters were calculated with sc.tl.leiden() with a resolution of 0.8, and UMAP coordinates generated with sc.tl.umap(). Genes marking neural and neuroendocrine tumors were identified with sc.tl.rank_genes_groups() with method=Wilcoxon.

Organoids:

GBC isolation from mice:

Live GBCs were isolated from mice as described previously⁶¹. Briefly, n=4–5 RPM or RPMA mice were intraperitoneally injected with 50 μg/g body weight methimazole (Sigma cat# M8506, dissolved in PBS at 5 mg/mL) approximately 10 days prior to isolation. Mice were sacrificed by CO₂ asphyxiation and decapitated. Olfactory mucosa was carefully dissected off of turbinate blocks and septum, and placed in cold enzyme dissociation cocktail, consisting of 5 mL dispase, 1 mg/mL collagenase type 1A (Sigma cat# C9891), 100 uL DNase I (Stem Cell Tech cat# D4527), and 10 mg papain (Stem Cell Tech cat# 07465). Once all dissections were complete, cells were incubated in enzyme cocktail for 20 min at 37°C with intermittent trituration. Next, 2 mL of prewarmed Trypsin (Stem Cell Tech) were added to the 37°C incubation for 2 min with continuous trituration. 4 mL of FBS was added to the mixture to stop enzyme reactions, and the solution was filtered through a 250 μm cell strainer, centrifuged at 400g for 5 min, and resuspended in 10 mL of HBSS. The solution was then filtered through a 70 μm cell strainer and centrifuged again. The cell pellet was resuspended in 300 μL of sort buffer, consisting of 2% FBS and 1mM EDTA in HBSS.

For selection of KIT⁺ cells, the cell suspension was incubated with 3 μL FCR block (Stem Cell Tech APC selection kit) and 12 μL APC anti-KIT (eBioscience cat# 17–1171-82) at RT for 15 min. 30 μL of APC selection cocktail was added to the suspension and incubated for 15 min at RT. Finally, 22 μL of Rapidsphere Nanosphere particles were added to and pipette-mixed in the suspension and incubated for 10 min at RT. Next, 2.2 mL of sort buffer was added to the suspension, and cells were placed in a sorting magnet (Stem Cell Tech cat# 18000), and incubated for 5 min. The negative fraction of cells was then removed by carefully decanting the solution. Cells were washed two more times with 2.5 mL of sort buffer for a total of 3 washes. The positive fraction of cells was washed with 7.5 mL of sort buffer and centrifuged at 400g for 5 min. The KIT⁺ cell pellet was then re-suspended in Neurocult media (Stem Cell Tech cat# 05700) in approximately 500–1000μL, depending on the size of the cell pellet.

GBC-derived organoid culture and Cre administration:

Isolated KIT⁺ GBCs from RPMA mice were immediately subject to recombination via cell-permeable TAT-Cre recombinase (Sigma cat# SCR508) to avoid differentiation of GBCs before transformation. ~50–100k cells were pelleted for 5 min at 500 x g at 4C, then resuspended in 500 uL of “organoid culture media, OCM” defined as 50% L-WRN conditioned organoid media and 50% Advanced DMEM/F-12 (Fisher Sci cat# 12–634-028) supplemented with 1% L-glutamine, 1% Pen-strep, and 10% FBS. 10 uL (100 Units) of TAT-Cre recombinase was added to the OCM, pre-filtered through a 0.22 uM syringe and warmed to 37C before adding to cells. Cells in one well of a 24-well plate were incubated 4 hr at 37C in TAT-Cre media, then collected, pelleted, and replated in 30 uL Matrigel. After Matrigel solidified, cultures were fed 500 uL/well of pre-filtered OCM supplemented with 10 uM Y27632 (Rho-kinase inhibitor; MedChemExpress cat# HY-10071) to support organoid formation and 5 uL (50 units) additional TAT-Cre. After incubation for an additional 6–12 hr, TAT-Cre media was removed and replaced by 500 uL of OCM supplemented with 10 uM Y27632. Every 3–5 days, OCM was replaced, and/or cultures were dissociated and expanded using TrypLE (Thermo Fisher cat# 12604013) for 10–30 min followed by quenching with Advanced DMEM/F-12 complete media described above, pelleting at 500 x g for 5 min at 4C, then resuspending in Matrigel and plating in OCM. Validation of recombination in RPMA organoids occurred ~4 weeks post-Cre treatment by PCR (details in following section).

Isolated KIT+ GBCs from RPM mice were similarly subject immediately to recombination following methods described above but were resistant to transformation and displayed lack of recombination at timepoints when RPMA GBC organoids were fully recombined. RPM-derived GBCs were subsequently subject to transformation via adenoviral-CMV-Cre after ~10 weeks in culture and ~6 weeks post-CellTagging. Organoids were dissociated using TrypLE (Thermo Fisher cat# 12604013) for 10–30 min. After quenching, ~500k cells were resuspended in OCM + 5×10⁷ total pfu Ad5-CMV-Cre (VVC-U of Iowa-5) + 0.5 ug/mL polybrene (Santa Cruz cat#sc-134220A) and plated into one well of a 24-well plate. Cells were spinfected for 30 min at RT, 300 x g. Cells were then incubated at 37C for 4 hr, pelleted, and replated in 40 uL of Matrigel + 2.5×10⁷ total pfu Ad5-CMV-Cre. Organoids were incubated 48 hr, then replated in fresh OCM+Matrigel. Every 3–5 days, OCM was replaced, and/or cultures were dissociated and expanded using TrypLE for 10–30 min followed by quenching with Advanced DMEM/F-12 complete media described above, pelleting at 500 x g for 5 min at 4C, then resuspending in Matrigel and plating in OCM. Validation of recombination in RPM organoids occurred 3–4 weeks post-Cre treatment by PCR (details in following section).

PCR validation of recombination efficiency:

Qiagen DNeasy kit (cat# 69506) was used to isolate genomic DNA from RPM and RPMA GBC-derived organoids following exposure to Cre recombinase. Fully recombined RPM-tumor-derived cell lines were used for positive recombination controls for Rb1, Trp53, and Myc^T58A alleles. A fully recombined RPMA-tumor-derived organoid culture was used for a positive recombination control for the Ascl1 allele. DNA concentrations were measured on a BioTek Synergy HT plate reader. Equal quantities of tumor genomic DNA (100 ng) were amplified by PCR with GoTaq (Promega cat# M7123) using primers to detect Rb1 recombination: D1 5’-GCA GGA GGC AAA AAT CCA CAT AAC-3’, 1lox 5’ 5’-CTC TAG ATC CTC TCA TTC TTC CC-3’, and 3’ lox 5’-CCT TGA CCA TAG CCC AGC AC-3’. PCR conditions were 94 deg 3 min, 30 cycles of (94 deg 30 s, 55 deg 1 min, 72 deg 1.5 min), 72 deg 5 min, hold at 4 deg. Expected band sizes were ~500 bp for the recombined Rb1 allele, and 310 bp for the floxed allele. Primers to detect Trp53 recombination include the following: A 5’-CAC AAA AAC AGG TTA AAC CCA G-3’, B 5’-AGC ACA TAG GAG GCA GAG AC-3’, and D 5’-GAA GAC AGA AAA GGG GAG GG-3’. PCR conditions were 94 deg 2 min, 30 cycles of (94 deg 30s, 58 deg 30s, 72 deg 50s), 72 deg 5 min, hold at 4 deg. Expected band sizes were 612 bp for the Trp53 recombined allele, and 370 bp for the floxed allele. Primers to detect Myc^T58A recombination include the following: CAG-F2 5’-CTG GTT ATT GTG CTG TCT CAT CAT-3’, MycT-R 5’-GCA GCT CGA ATT TCT TCC AGA-3’. PCR conditions used were 94 deg 2 min, 35 cycles of (95 deg 30 sec, 60 deg 30 sec, 72 deg 1.5 min), 72 deg 7 min, hold at 4 deg. Expected band sizes were ~350 bp for the recombined allele, and ~1239 bp for the floxed allele. Primers to detect Ascl1 recombination include the following: Sense Ascl1 5’UTR: 5’-AAC TTT CCT CCG GGG CTC GTT TC-3’ (for Cre recombined fwd), VR2: 5’-TAG ACG TTG TGG CTG TTG TAG T-3’ (for Cre recombined rev), MF1 5’-CTA CTG TCC AAA CGC AAA GTG G-3’ (for floxed fwd), and VR2 5’-TAG ACG TTG TGG CTG TTG TAG T-3’ (for floxed rev). PCR conditions used were 94 deg 5 min, 30 cycles of (94 deg 1 min, 64 deg 1.5 min, 72 deg 1 min), 72 deg 10 min, hold at 4 deg. Expected band sizes were ~700–850 bp for the Ascl1 recombined allele, and ~857 bp for the floxed allele. All degrees are in Celsius. PCR products were run on 1.2% agarose/TAE gels containing ethidium bromide and images were acquired using a BioRad Gel Dox XR imaging system.

Lentiviral transduction of organoids with CellTag Library V1:

The CellTag V1 plasmid library was purchased from Addgene (cat# 124591) and amplified according to the published protocol for this technology⁷³. Briefly, the plasmid library was transformed using Stellar competent cells at an efficiency of ~220 cfus per unique CellTag in the V1 library. The library was isolated from E. coli culture via the Plasmid Plus Mega Kit (Qiagen, cat#12981), and assessed for complexity via high-throughput DNA sequencing with the Illumina MiSeq (75-cycle paired-end sequencing v3). Generation of the CellTag Whitelist from sequencing data resulted in 13,836 unique CellTags in the 90^th percentile of detection frequency. High titer lentivirus (1.5E7 transducing units (TU)/mL) was generated from the CellTag V1 library following published protocols^33,132, and titered based on GFP fluorescence with 293T cells (ATCC, cat# CRL11–268).

RPM and RPMA GBC-derived organoids were expanded for 2–3 weeks post-isolation, then ~1 × 10⁶ cells were transduced with the CellTag V1 lentiviral library. Organoids were dissociated into single cells with TrypLE for 30 min and subject to mechanical dissociation every 10 min of TrypLE incubation. TrypLE was quenched and cells were pelleted and resuspended in 500 uL of 50% conditioned media plus 8 ug/mL polybrene (Santa Cruz, cat# sc-134220) and 25 uL of CellTag V1 lentivirus, then plated in one well of a 24-well plate. Cells were spinoculated at 300 x g for 30 min at RT to increase transduction efficiency, incubated immediately following spinoculation for 3–6 hr at 37C, then pelleted and replated in 50 uL of Matrigel and 500 uL of fresh viral supernatant. Organoids were incubated for 24 hr, then viral media was replaced with normal 50% conditioned media for organoid expansion. GFP was visible in >50% of cells as soon as 24 hr after viral transduction.

RPMA organoids were expanded for ~1 week after CellTagging to allow limited clonal expansion, then subject to dissociation and 10X Genomics scRNA-seq library preparation (following methods described above) or sorting for GFP to enrich for CellTagged populations. GFP-sorted cells were allowed to re-form organoids, then 1 week after sorting were implanted into flanks of SCID/beige mice (Taconic) with ~1–2 × 10⁶ cells per flank in 50 uL of 50:50 Matrigel:50% conditioned media. In total, RPMA organoids grew in culture for ~4 weeks following transformation by Cre before being allografted. Due to difficulty with recombination, RPM organoids grew for ~9 weeks after initial CellTagging then were subject to scRNA-seq library preparation and implantation to SCID/beige mice as described above. RPM or RPMA GBC-derived organoid allografts were collected ~6–7 weeks following implantation, both at a size of ~1 cm³, and then subject to FFPE and dissociation for scRNA-seq. Tumor tissue was dissociated into single cells with an enzymatic cocktail described in previous methods³³, then prepared as described above for 10X Genomics scRNA-seq. All libraries were sequenced on an Illumina NovaSeq 6000 aiming for a read depth of at least 50k paired-end reads per cell for dual indexed samples.

CellTag analysis and clone calling:

ScRNA-seq data from CellTagged, GBC-derived allografts were demultiplexed and processed into FASTQ files via the 10X Cell Ranger v7.0.0 pipeline. Reads were aligned to a custom mouse genome (mm10–2020-A) with the CellTag.UTR and GFP.CDS transcripts included as described in the CellTagR computational pipeline⁷³. Count barcodes and UMIs were generated using cellranger count. CellTags were extracted from the RPM and RPMA allograft samples’ processed BAM files following methods described in the CellTagR pipeline documentation (https://github.com/morris-lab/CellTagR). In short, BAM files for the allograft samples were filtered to exclude unmapped reads and include reads that align to the GFP.CDS transgene or CellTag.UTR. CellTag objects were created in R, and CellTags were extracted from the filtered BAM files to generate matrices of cell barcodes, 10X UMIs, and CellTags. The matrix was further filtered to include only barcodes identified as cells by the CellRanger pipeline, and then subject to error correction via Starcode. CellTags not detected in our Whitelist generated from assessment of our lentiviral library complexity (described above) were also removed. Clones were assigned as cells expressing >2 but <20 CellTags with similar combinations of CellTags (Jaccard similarity better than 0.7 for RPM and 0.75 for RPMA). For scRNA-seq analysis and CellTag visualization, we utilized Scanpy (v1.9.1) for initial QC and scvi-tools(v0.17.4) for integration and clustering in Python (v3.8.8), following methods described for other data in this study. Data include RPMA allograft clones with >10 cells per clone, to ensure robust sampling of each clone (n=16), and all detected RPM allograft clones (n=3). CellTag-based clonal information was added as metadata by 10X-assigned barcode to our analysis to visualize clone distribution and cell identity per clone using standard visualization functions in Python and R. Optimal transport and pseudotime analysis on GBC allograft data were performed according to methods described in the “Pseudotime” and “Cell Cluster Distance Analysis” methods sections.

Pseudotime:

Pseudotime analysis was performed with Scanpy PAGA pseudotime. PCAs were calculated with sc.pp.pca() and nearest neighbors were computed with sc.pp.neighbors() with use_rep set to the final scvi model embedding from .obsm. Leiden clusters were calculated with sc.tl.leiden(), followed by sc.tl.paga() with groups=“leiden”, and sc.pl.paga(). Plots were generated with sc.tl.draw_graph() with init_pos=“paga”, followed by sc.pl.draw_graph() and sc.pl.paga_compare(), with threshold=0.01, and edge_width_scale=0.5. Arrows on Figure 5I were manually annotated based on predicted differentiation trajectories. Likewise, for gene expression over pseudotime graphs, paths and markers for specific cell types were specified based on the literature and enriched marker genes per cell state. Sc.pl.paga_path() was used to generate plots.

QUANTIFICATION AND STATISTICAL ANALYSIS

Any remaining statistical analysis was performed using GraphPad Prism. Error bars show mean ± SD unless otherwise specified. Significance was determined by Student’s two-tailed unpaired t tests with 95% confidence intervals and p values <0.05 considered statistically significant, unless otherwise indicated. All statistical details are further described in respective figure legends. Additional statistical methods related to quantification of immunohistochemistry, and bioinformatic analyses of scRNA-seq and spatial transcriptomics can be found in METHOD DETAILS. No statistical methods were used to predetermine sample sizes.

Supplementary Material

2

Table S1. Histological assessment of RPM and RPMA ONB primary tumors and allografts, related to Figures 1–5.

3

Table S2. Mouse olfactory neuroblastoma transcriptionally resembles human ONB, related to Figure 4.

4

Table S3. GBC-derived ONB exhibits plasticity between neuronal and non-neuronal states, related to Figure 5.

5

Table S4. Human ONB exhibits similarities to SCLC in intratumoral heterogeneity, immune-cold tumor microenvironment, and therapeutic targets, related to Figure 6.

Highlights.

• Rb1/Trp53/Myc alterations promote high-grade metastatic NEUROD1⁺ ONB

• Mouse and human ONB exhibit transcriptional similarity to neuroendocrine tumors

• ONB harbors intratumoral cell fate heterogeneity and lineage plasticity

• ASCL1 suppresses non-neuronal lineages including POU2F3⁺ microvillar-like states