YAP-Driven Malignant Reprogramming of Epithelial Stem Cells at Single Cell Resolution

Tumor initiation represents the first step in tumorigenesis during which normal progenitor cells undergo cell fate transition to cancer. Capturing this process as it occurs in vivo, however, remains elusive. Here we employ cell tracing approaches with spatiotemporally controlled oncogene activation and tumor suppressor inhibition to unveil the processes underlying oral epithelial progenitor cell reprogramming into cancer stem cells (CSCs) at single cell resolution. This revealed the rapid emergence of a distinct stem-like cell state, defined by aberrant proliferative, hypoxic, squamous differentiation, and partial epithelial to mesenchymal (pEMT) invasive gene programs. Interestingly, CSCs harbor limited cell autonomous invasive capacity, but instead recruit myeloid cells to remodel the basement membrane and ultimately initiate tumor invasion. CSC transcriptional programs are conserved in human carcinomas and associated with poor patient survival. These findings illuminate the process of cancer initiation at single cell resolution, thus identifying candidate targets for early cancer detection and prevention.


INTRODUCTION
Adult stem cells play a central role in tissue homeostasis by balancing self-renewal and differentiation. 1 However, cells with self-renewal capacity may also accumulate and propagate oncogenic genomic alterations, ultimately leading to carcinogenesis. 2,3Stem cells thus possess intrinsic tumor suppressive mechanisms to guard against inappropriate oncogene activation, including terminal differentiation, 4 oncogene-induced senescence, 5 and apoptosis. 6Current models of carcinogenesis posit that tumor initiation requires the inactivation of intrinsic tumor suppressive mechanisms concomitant with oncogene activation. 7These insights have been supported by recent genome-wide cancer sequencing efforts that have cataloged candidate genomic alterations underlying most human malignancies. 8However, these studies in established, often advanced tumors are confounded by cellular and mutational heterogeneity and thus cannot directly identify cancer-initiating cells, often referred to as cancer stem cells (CSCs), or discriminate between alterations driving tumor initiation from those promoting tumor progression.As such, the underlying molecular mechanisms mediating malignant reprogramming of normal progenitor cells into CSCs remains poorly understood.
Stem cells in the oral mucosa reside in the basal layer of the strati ed squamous epithelium, and consist of a single pool of self-renewing oral epithelial progenitor cells (OEPCs). 9This single progenitor cell population renders the oral epithelium an ideal system to elucidate mechanisms underlying malignant reprogramming, 10 unlike other tissue systems such as the skin or gut, in which elegant studies have revealed multiple unique stem cell pools that play distinct roles in cancer initiation. 11,12Head and neck squamous cell carcinoma (HNSC) represents the most common malignancy arising from the upper aerodigestive epithelia. 13Extensive molecular characterization of HNSC has revealed that while each individual tumor exhibits alterations in a large number of genes, these alterations converge to impact a nite set of oncogenic molecular pathways. 14HNSC is characterized by near universal loss-of-function of TP53 and CDKN2A tumor suppressors by genomic alteration or human papillomavirus (HPV)-mediated inhibition. 14Notably, alterations in FAT1, observed in nearly 30% of HNSC, 14 disrupt Hippo pathway signaling and result in unrestrained activation of the transcriptional co-activator YAP. 15 We now combine knowledge of the landscape of oncogenic pathway alterations in HNSC with genetically engineered animal models, lineage tracing, and single cell transcriptomics to unveil the underpinnings of cancer initiation in vivo.Here, we found that in the context of TP53 and CDKN2A suppression, YAP activation rapidly reprograms OEPCs into CSCs, providing an opportunity to illuminate the process of cancer initiation at single cell resolution.

YAP activation and E6-E7 expression is su cient to induce rapid tumor initiation
Tumor initiation represents the crucial rst step in tumorigenesis during which normal progenitor cells undergo cell fate transition to cancer.To investigate this process, we developed genetically engineered murine systems focusing on prevalent and co-occuring genomic alterations in HNSC (Fig. 1a).We employed conditional expression of the HPV16 E6-E7 oncogenes to concomitantly inhibit the TP53 and CDKN2A tumor suppressors. 16Given the frequency of FAT1 mutations and widespread YAP nuclear expression in HNSC, 17 we investigated the effects of YAP activation in conjunction with E6-E7 by expression of the YAP1 S127A allele.Keratin 14 (KRT14) is expressed in the basal layer of oral epithelia and marks OEPCs, which may represent the cell of origin for HNSC. 18We employed a tamoxifen-inducible Cre-recombinase (CreERT) driven by the Krt14 promoter to target genomic alterations to KRT14 + OEPCs. 19We bred mice bearing E6-E7 ("E"), YAP1 S127A ("Y"), or both transgenes ("EY").Littermates lacking these transgenes but possessing Krt14-CreERT, LSL-rtTA regulatory transgenes, and the H2B-GFP reporter were used as normal controls ("N").Local administration of tamoxifen activated CreERT-mediated recombination of a oxed STOP cassette (LSL) and enabled transcription of the reverse tetracycline-controlled transactivator (rtTA) in KRT14 + OEPCs (Fig. 1b).Administration of doxycycline chow then induced expression of the tetracycline response element-regulated HPV16 E6 − E7 , YAP1 S127A , and H2B-GFP transgenes (Extended Data Fig. 1a-d). 16,20,21rial examination of mouse tongues identi ed macroscopic lesions as early as 8 days after transgene induction.Median lesion-free survival was dramatically different across genotypes (Fig. 1c-d).By 20 days, the majority (65%) of EY mice bore at least two lesions, while few Y and no E or N mice bore any lesions (Extended Data Fig. 1e).Evaluation of hematoxylin and eosin (H&E) staining and pan-cytokeratin immunohistochemistry (IHC, Extended Data Fig. 1f) showed at least one invasive carcinoma in 81% of EY mice, compared to 18% of Y mice and no E or N mice (Fig. 1e,f).Most carcinoma-bearing EY mice had multiple independent carcinomas, which were also larger and more deeply invasive than those of Y mice (Fig. 1g-i).We next investigated tumor initiation at higher temporal resolution using a pulse-chase strategy.At day 10, we observed a marked increase in EY epithelial thickness compared to other epithelia (Fig. 1j,k).Invasive carcinoma occurred in 44% of EY mice by 10 days, while no carcinoma in Y epithelia were observed until day 20 (Fig. 1l).These ndings show that in KRT14 + OEPCs, unrestrained YAP activation in the context of E6-E7 expression is su cient to induce oral carcinoma with high penetrance and rapid kinetics.
Invasive carcinoma is preceded by the expansion of a stem-like cell population We investigated the early processes of tumor initiation employing H2B-GFP for lineage tracing of transgene-activated cells.In epithelial whole mounts 36 hours after a single dose of intralingual tamoxifen, we observed a similar distribution of H2B-GFP + basal cells across transgenic conditions (Fig. 2a,b).By 6 days, H2B-GFP + cells had expanded throughout the full thickness of EY epithelia (Fig. 2c,d).While isolated mitotically active cells (KI67 + staining) remained restricted to the basal layer in normal (N) epithelia, the expression of E, Y, or EY transgenes resulted in extension of KI67 + cells beyond the basal layer (Fig. 2e,f, Extended Data Fig. 2a,c).We next evaluated the expression of P63 and SOX2, transcription factors that regulate epithelial morphogenesis and OEPC maintenance, 22,23 and the transmembrane protein ITGA6, a component of the hemidesmosome. 24,25In normal epithelia, P63 and SOX2 expression was restricted to the basal layer, and ITGA6 to basal cells in contact with the basement membrane.Expression of E, Y, or EY led to an increase in P63 + epithelial cells.YAP activation resulted in ectopic extension of P63 + cells into suprabasal epithelial layers.Concomitant YAP and E6-E7 expression resulted in expansion of the P63 + compartment to occupy the full thickness of the oral epithelia (Fig. 2g,h, Extended Data Fig. 2b,d).Similarly, EY expression drove the expansion of SOX2 + cells into all suprabasal layers (Fig. 2i,j).While E6-E7 expression alone did not alter ITGA6 compartmentalization, Y and EY expression diminished ITGA6 basement membrane localization and resulted in diffuse low (Y) to high (EY) ITGA6 expression throughout suprabasal strata (Extended Data Fig. 2b,e).
Strikingly, in EY tumors, most cells at the invasive front were KI67 + /TP63 + /SOX2 + (Fig. 2k-m).Taken together, these ndings suggest that YAP activation, together with E6-E7-mediated tumor suppressor inhibition, results in the rapid expansion of a proliferating stem cell-like population at the pre-invasive stage.

Transcriptional programs underlying OEPC reprogramming: YAP promotes mTOR signaling
We performed RNA sequencing (RNAseq) of microdissected tongue epithelia to evaluate transcriptional changes occurring upon tumor initiation (Extended Data Fig. 3a).We selected 15 days post-transgene induction as the normal to carcinoma transition point since half of EY mice bore invasive carcinoma at this time point.Principal component analysis showed clear distinctions between EY, Y, E, and N samples (Extended Data Fig. 3b).As expected, upregulation of YAP target genes 26 and YAP signatures by gene set enrichment analysis (GSEA) 27 were observed in YAP-expressing conditions (Fig. 3a,b, Extended Data Fig. 3c,d, Supplementary Table 1), including cell cycle, cell identity, differentiation, and stemness programs (see below Fig. 4g).
Unexpectedly, among multiple transcriptional signatures, Y and EY mice also showed enrichment for gene sets for mTOR pathway activation, a commonly activated signaling mechanism in HNSC 28 (Fig. 3b,c and Extended Data Fig. 3e).Consistently, IHC of epithelia from EY and Y mice showed a pronounced increase in phospho-S6 (pS6) ribosomal protein levels, a downstream marker of mTOR activity 29 (Fig. 3d,e).These ndings raised the possibility that YAP may drive mTOR activity.We noted that Axl, a YAP target gene, is upregulated in epithelia of Y and EY mice (Fig. 3f), potentially representing a mechanistic candidate linking YAP activation to mTOR signaling.We performed siRNA-mediated knockdown of YAP and/or its paralog TAZ (WWTR1) in human HNSC cells (Cal27 and Cal33) to test if YAP indeed regulates mTOR activity.Consistent with reports that YAP and TAZ serve mutually compensatory functions, 30 combined knockdown of YAP and TAZ was required to diminish expression of the YAP target CYR61 and mTOR activity, based on pS6 abundance (Fig. 3g,h).Intriguingly, combined knockdown also resulted in diminished pEGFR but not total EGFR, potentially related to decreased expression of YAP-regulated EGFR ligands, including Epgn, Nrg1, and Nrg4 (Fig. 3i), suggesting that YAP initiates multiple transcriptional mechanisms resulting in mTOR activation (Fig. 3j).

Oncogenic transcriptional reprogramming de nes YAP and E6-E7 activated epithelia
We next extended our global pathway analysis to OEPC gene programs.Recent characterization of normal murine oral epithelia at single cell resolution showed that the basal layer consists of stem, cycling progenitor, and differentiating cells (Fig. 4a). 9Using signatures of these physiologic cell states, we evaluated the effects of transgene expression in bulk oral epithelia.We observed pronounced enrichment for the cycling progenitor signature and a dramatic depletion of the differentiation signature, suggesting that EY expression promotes oral epithelial cell cycle progression concomitant with differentiation arrest and depletion of the normal stem cell program (Fig. 4b,d and Extended Data Fig. 4a).At the gene level, differentially expressed gene (DEG) analysis showed EY-mediated upregulation of OEPC stemness factors (Ybx1, Procr, Ezh2, Suz12, Eed) and downregulation of differentiation (Grhl1, Ovol1, Krt4, Krt13, Ivl, Lor, Krtdap, late corni ed envelope genes) and apicobasal polarity factors (Pard3, Pard3b, Vangl2, Camsap3).Notably, several basal progenitor state factors (Hoxc13, Pax9, Gli2, Krt15) displayed paradoxical downregulation (Extended Data Fig. 4b), indicating that EY-driven transcriptional changes do not re ect a physiologic OEPC cell state, but rather constitute a unique cellular state related to tumor initiation.
To gain a global view of transcriptional programs, DEGs in each condition were compared to normal.From E to Y to EY, a progressively larger number of DEGs was observed (Supplementary Table 2).Venn analysis revealed 2318 genes differentially expressed solely in EY mice (Fig. 4e, Supplementary Table 3).Gene ontology (GO) 31,32 and GSEA of Molecular Signatures Database (MSigDB) Hallmark pathways 33,34 identi ed enrichment among these 'EY-unique' DEGs for processes underlying cell proliferation, epithelial cell development and identity, and acute phase in ammatory responses (Fig. 4f and Extended Data Fig. 4c).
Further analysis revealed that distinguishing features of EY transcriptomes included activation of cell cycle and stemness programs, inhibition of differentiation, and activation of epithelial to mesenchymal transition (EMT, Fig. 4g).These ndings suggest that beyond cell cycle activation, the OEPC to CSC transition involves differentiation arrest, activation of stemness, and EMT-induced epithelial plasticity.

Activation of CSC programs in nascent carcinoma at single cell resolution
Our transcriptional analyses of bulk epithelia shed insight into the processes occurring in cells undergoing malignant conversion.However, the cellular heterogeneity of the pre-malignant microenvironment precludes elucidation of tumor initiating transcriptional programs speci cally in CSCs.We thus performed single cell RNAseq (scRNAseq) of oral epithelia at the same time point as bulk RNAseq (Fig. 5a).In total, 12,771 epithelial cells were identi ed across 8 clusters, which could broadly be divided into de ned physiologic cell states 9 (see above, Fig. 4a) and transgene-associated cell states (Fig. 5b).
To test tumor initiating capacity, we implanted cell suspensions generated from E, Y, and EY transgene induced epithelia into the tongues of NOD-SCID-gamma (NSG) immune compromised mice (Fig. 5h).Large tumors formed in all mice implanted with EY-induced cells, compared to only small tumors in fewer mice implanted with Y-induced cells, and no tumors in mice implanted with E-induced cells (Fig. 5i-k).Collectively, these ndings demonstrate that combined YAP activation and E6-E7 expression endows OEPCs with an oncogenic transcriptional phenotype, and results in e cient OEPC reprogramming into CSCs.

CSCs co-opt collagenase-expressing G-MDSCs to facilitate tumor invasion
The basement membrane represents an anatomic barrier against invasive carcinoma.Second harmonic generation microscopy 46 identi ed a dramatic reduction in brillar collagen abundance at the invasive front of nascent in ltrative carcinoma (Fig. 6a,b).However, CSCs did not express collagenases (Fig. 6c), and hence may lack the intrinsic ability to initiate tumor invasion.In search for an alternative mechanism, we hypothesized that in ltrating immune cells (CD45 + ) observed within EY lesions (Fig. 6d) may contribute to basement membrane extracellular matrix (ECM) remodeling.To explore this possibility, we analyzed the scRNAseq data of 11,286 immune cells present in our scRNAseq data, which distributed across 13 cluster identities (Fig. 6e and Extended Data Fig. 6a,b, Supplementary Table 5).Remarkably, myeloid derived suppressor cells (MDSCs) comprised 65% of immune cells (31% of granulocytic, G-MDSCs, and 34% of monocytic, M-MDSCs) in EY-expressing epithelia (Fig. 6f,g).Ligand-receptor analysis revealed that CSCs express chemokine ligands (Cxcl1, Cxcl2, S100a8, S100a9, Csf3) whose corresponding receptors (Cxcr2, Csf3r) are speci cally expressed by G-MDSCs (Fig. 6h,i).Analysis of bulk RNAseq from epithelia also demonstrated granulocyte genes including Csf3, Retnlg, Cd177, and Ly6g ranked among the most highly upregulated genes in EY epithelia (Extended Data Fig. 6c).In oral epithelia, TNFα, G-CSF, IL-23, and IL-17 initiate a cytokine-chemokine cross talk, which induces sustained granulocyte recruitment during in ammation. 47In line with this model, EY-expressing epithelia showed increased abundance of Il23a and Il17f transcripts and TNFα, G-CSF, and IL-17 proteins, as well as the granulocyte-speci c chemokines CXCL1 and CXCL2 (Extended Data Fig. 6d).
These ndings suggest that promoting G-MDSC in ltration constitutes a distinctive feature of CSCs.We thus asked if G-MDSCs contribute to the pre-invasive to invasive transition of CSCs.We observed increased abundance of LY6G + G-MDSCs at tumor invasive fronts (Fig. 6j).Furthermore, in contrast to the absence of collagenase expression in EY epithelial cells, G-MDSCs expressed collagenases Mmp8 and Mmp9 (Fig. 6k), and displayed elevated levels of pro-MMP9 protein (Extended Data Fig. 6e).Consistent with a causal role of G-MDSCs in tumor invasion, treatment of transgene-induced EY mice with anti-LY6G depleting antibody signi cantly reduced the overall multiplicity of carcinoma (Fig. 6l,m and Extended Data Fig. 6f).These ndings suggest that cytokines and chemokines secreted by EY-transformed CSCs recruit collagenase-producing G-MDSCs, which cleave basement membrane collagen brils and facilitate tumor invasion.

CSC programs are enriched in HNSC and associated with poor prognosis
To test the translational signi cance of the CSC tumor initiation transcriptional program in human disease, we performed high dimensional weighted gene coexpression network analysis (hdWGCNA) 48,49 .We decomposed the tumor initiating cluster into 12 co-expressed transcriptional modules (Fig. 7a, Supplementary Table 6), then subjected each module to ssGSEA analysis in matched TCGA-HNSC tumor and normal tissue samples.We noted enrichment of 8 of 12 modules exclusively in HNSC, unveiling RNA metabolism and processing, intracellular tra cking, hypoxia response, G1/S and G2/M cell cycle progression, interferon response, motility and migration, and cytoskeleton and cell polarity as distinguishing features of the HNSC transcriptome (Fig. 7b).Among all modules, enrichment for CSC G1/S cell cycle, motility and migration, and cytoskeleton and polarity modules were associated with worse diseasefree and overall survival (Fig. 7c-h).These ndings suggest that our transgene-induced oral CSCs display coherent transcriptional programs enriched in aggressive HNSC.

DISCUSSION
The overwhelming majority of work investigating cancer-driving mechanisms has relied on established tumors, which requires retrospective inference of tumor initiating events and limits distinction between processes governing tumor initiation from progression.In HNSC, despite this mutational and cellular heterogeneity, the diverse genomic alterations observed in HNSC converge to activate a limited number of oncogenic signaling pathways.Through the use of a spatiotemporally controlled murine system targeting genomic alterations to a single pool of epithelial progenitor cells, we now show that unrestrained YAP activation in the context of TP53 and CDKN2A inhibition by HPV E6-E7 oncogenes induces carcinoma with rapid kinetics and nearly complete penetrance, enabling in-depth investigation of tumor initiation.The ability to combine this genetically-de ned system with lineage tracing and single cell transcriptomics provided a unique opportunity to investigate the conversion of epithelial progenitor cells into CSC as it occurs in vivo.
The identi cation of a rapidly emerging CSC population in vivo offered the opportunity to directly examine the biological processes giving rise to cancer initiating cells.In this regard a widely accepted framework for carcinogenesis posits that accumulating genomic alterations lead to stepwise transcriptional changes driving normal tissue to invasive carcinoma via a series of intermediate states.Our ndings indicate that once a minimal complement of pathway alterations is achieved -loss of TP53 and CDKN2A function and YAP activation -OEPCs are rapidly reprogrammed into CSCs endowed with hallmarks of invasive carcinoma, including cell cycle progression, hypoxia, squamous differentiation defects, interferon responses, and a pEMT cell identity.While the role of pEMT in tumor initiation has remained elusive, our interrogation of the CSC state now reveals that pEMT represents an early and distinguishing phenotype of CSCs.By examining coordinately expressed programs in CSCs, we further dissected the elements of pEMT relevant to transgene-mediated tumor initiation into distinct transcriptional programs driving cellular motility and cell-ECM interactions.Remarkably, our new CSC-speci c transcriptional programs, including pEMT, were enriched in HNSC and correlated with worse prognosis, thus supporting their likely direct relevance to human squamous malignancies.
One unexpected nding was that tumor initiation could be achieved in the absence of genomic alterations in the PI3K/AKT/mTOR signaling axis, given the extensive evidence implicating widespread mTOR activation in HNSC 28, [50][51][52] .However, we observed robust mTOR activation in pre-invasive lesions and throughout SCC in YAP-expressing epithelia, suggesting that YAP can activate mTOR signaling.Indeed, mTOR program enrichment represented a distinguishing feature of the EY-induced CSC state.Mechanistically, we found that YAP drives the transcriptional upregulation of Axl and multiple EGFR ligands, which may explain EY-mediated mTOR activation.The fact that > 70% of HNSCs do not harbor genomic alterations in the PI3K-mTOR pathway but exhibit a widespread activation of YAP 17,53 and mTOR 52 is aligned with a potential signaling cross-talk in which YAP can activate mTOR.Speci cally, our ndings support a model in which YAP:TEAD-mediated transcription activates an AXL-and EGFR-initiated signaling cascade resulting in the activation of mTOR in epithelial cells, thus representing an actionable target to prevent tumor initiation.
Remarkably, transgene-induced carcinoma displayed tumor invasive fronts with extensive modi cation of basement membrane collagen brils, concordant with the overall perspective that ECM modi cation and invasion are de ning features of premalignancy-to-cancer transition.However, our single cell analysis of CSCs revealed that these cells do not express collagenases.Instead, our ndings indicated that CSCs express multiple cytokines and chemokines that promote G-MDSC recruitment, and that recruitment of G-MDSCs expressing Mmp8 and Mmp9 collagenases to the invasive front is required for CSC invasion.These observations provide a model in which CSCs lack cell autonomous intrinsic invasive capacity and alternatively polarize the tumor microenvironment (TME) and exploit innate immune cells to disrupt the basement membrane, ultimately enabling tumor invasion.
In summary, we demonstrate that a genetically-de ned, traceable system simultaneously activating oncogenic pathways and disabling tumor suppressive mechanisms in normal oral epithelial progenitor cells induces the emergence of a distinct cancer initiating stem-like cell state.Through multimodal analysis of nascent CSCs at the single cell level in vivo, we de ne tumor-autonomous transcriptional programs and CSC-TME cross-talk as tumor initiating events during invasive carcinoma formation (Fig. 8).This conceptual framework of cancer initiation has the potential to open multiple novel avenues for early intervention, including precision targeting of tumor cell-autonomous cancer initiating signaling pathways, and disrupting CSC-TME networks mediating the development of invasive carcinoma.Declarations LMC, K.Sato; Writing, Original Draft: FF, SIR, JSG; Writing, Review & Editing: FF, SIR, JSG, ODK, QS, KSato, PT, ZM; Visualization: FF; Supervision: JSG, PT, QS, ODK, JAC; Project Administration: FF, JSG; Funding Acquisition: FF, JSG.

DECLARATION OF INTERESTS
J.S.G. has received other commercial research support from Kura Oncology, Mavupharma, Dracen, Verastem, and SpringWorks Therapeutics, and is a consultant/advisory board member for Domain Therapeutics, Pangea Therapeutics, and io9, and founder of Kadima Pharmaceuticals.The remaining authors declare no competing interests.

Lead contact
Further information and requests for resources and reagents should be directed to the Lead Contact, J. Silvio Gutkind (sgutkind@health.ucsd.edu).

Materials availability
Transgenic mice and cell lines generated in this study are available from the Lead Contact upon reasonable request and completion of Material Transfer Agreements (MTA).There may restrictions to the availability of these reagents due to cost or limited quantities.Tg(KRT14-cre/ERT) 20Efu and Tg(tetO-HIST1H2BJ/GFP) 47Efu . 19,20The Col1a1tm1(tetO-Yap1*) Lrsn mouse was kindly provided by Dr. Fernando Camargo (Harvard University). 21The B6.Cg-Gt(ROSA)26Sor tm1(rtTA,EGFP)Nagy/J mouse was obtained from The Jackson Laboratory. 57The Tg(tetO-HPV16-E6E7) SGu mouse was designed by the Gutkind laboratory and generated in house. 16All transgenic mouse experiments were performed in age-and sex-balanced groups of 8-16 week old littermates.NSG™ mice (NOD.Cg-Prkdc scid Il2rg tm1Wjl/SzJ ) mice were originally obtained from The Jackson Laboratory and propagated at the Moores Cancer Center.Implantation of transgene epithelial cell suspensions were performed in 8-week-old female NSG mice.

Genotyping
Human HNSC cell lines CAL27 and CAL33 cell lines were obtained from the NIDCR Oral and Pharyngeal Cancer Branch cell collection. 58Cell identity was con rmed by STR pro ling.CAL27 (CVCL_1107) was derived from a 56 year old male with tongue adenosquamous carcinoma.CAL33 (CVCL_1108) was derived from a 69 year old male with tongue squamous cell carcinoma.Both cell lines were cultured in DMEM (D-6429, Sigma-Aldrich, St. Louis, MO), 10% fetal bovine serum, 5% CO 2 , at 37 °C, and both tested free of Mycoplasma infection directly prior to experimentation.

Transgene induction
Mice were anesthetized with iso urane and 100µL of tamoxifen solution (20 mg/mL in miglyol) was administered into the tongue under stereomicroscopic visualization.One dose of tamoxifen was administered every other day for a total of 3 doses.

Epithelia
After in situ in ltration with 500uL collagenase+dispase solution (1mg/mL, 2.5 mg/mL) (Millipore Sigma), the tongues of euthanized mice were dissected free and incubated for 30 minutes at 37°C.The tongue epithelium was then dissected free from the underlying muscle under stereomicroscopic visualization.

LoxP-STOP-LoxP excision assay
For oxed stop excision assay, high-quality genomic DNA was isolated from whole epithelia using the DNeasy Blood and Tissue Kit per manufacturer protocol (Qiagen).PCR products were generated using REDTaq® polymerase and LSL excision primers, and were subjected to electrophoresis on 2% agarose gel in Tris acetate EDTA buffer.

RT-qPCR
RNA was prepared by homogenization of whole tongue epithelia in TRIzol® (Invitrogen) followed by phenol:chloroform extraction and RNeasy Mini Kt based column puri cation with on-column DNase treatment (Qiagen).For quantitative PCR (qPCR), cDNA library preparation was performed using Bio-Rad iScript™ reverse transcriptase and qPCR was performed using Applied Biosystems Fast SYBR® Green Master Mix per manufacturer's instructions.

Evaluation of gross tongue
Following transgene induction, mice were examined under anesthesia using a stereomicroscope every 3-7 days for the appearance of tongue epithelial lesions.Lesion free survival in days was de ned as the time from rst tamoxifen treatment to the appearance of the rst gross lesion.

Fluorescence microscopy
For whole mount uorescence imaging, epithelial sheets were isolated from mice euthanized 36 hours after a single dose of intralingual tamoxifen, washed in HBSS, stained with NucBlue for 3 hours at room temperature, washed again, mounted immersed in HBSS between two cover glasses, and Z-stacks were acquired with a confocal microscope.
For cross-sectional uorescence imaging, immediately after euthanasia, mice underwent intracardiac perfusion rst with 2mM EDTA in PBS followed by 1.6% paraformaldehyde in PBS.Perfusion xed tongues were dissected and incubated in 1.6% paraformaldehyde at room temperature overnight, then transferred to 30% sucrose in PBS for 2-3 days at 4°C, then washed in PBS, then embedded in OCT media and snap frozen in cryomolds for frozen section slide preparation.
For uorescent analyses, slides were thawed in the dark, blocked, incubated overnight at 4°C with primary antibodies, and then incubated with uorophoreconjugated secondary antibodies for 2 hours at room temperature.Nuclei were then stained with Hoechst 33342 in PBS for 15 minutes and slides were mounted with ProLong Diamond mounting medium.
Second harmonic generation for collagen imaging The generation imaging was done on an upright Leica SP8 microscope with a resonant scanner and hybrid non-descanned detectors.Ti-Sapphire femtosecond pulsed Chameleon Ultra II (Coherent Inc.) laser was tuned to 855 nm and the beam was focused on the sample with an HC PL APO CS 10x/0.40dry objective.The light was routed to the detectors with 560 nm, 495 nm, and 640 nm long-pass dichroic mirrors.The SHG signal was recorded with a 425/26 nm bandpass lter, the auto uorescence was recorded with a 650/60 nm bandpass lter.The pixel size was set to 0.746 µm, and 16x line averaging was used to improve the signal-to-noise ratio.Data were digitized in an 8-bit mode.The sample navigator software module was used to create autofocus support points and individual elds of view were tiled and stitched.

RNAseq
Tongue epithelia were isolated and RNA was prepared as described above.RNA samples passing purity, concentration, and integrity quality metrics by NanoDrop and TapeStation were submitted to Novogene for oligo-dT-based mRNA selection, cDNA library preparation, and sequencing on Illumina NovaSeq6000.
siRNA transfection in human cell lines All human cells were transfected at 60% con uency using Lipofectamine RNAiMAX reagent according to the manufacturer's instructions, using 20nM of each siRNA.Culture media was refreshed at 24 hours after transfection.Cells were placed under serum free conditions at 48 hours, and collected for experimentation at 72 hours post-transfection.

Immunoblot assay
Cells rinsed with ice cold PBS and lysates were harvested in RIPA buffer (50 mM Tris-HCl, 150 mM NaCl, 1 mM EDTA, 1% NP-40) supplemented with Halt TM Protease and Phosphatase Inhibitor Cocktail (#78440, ThermoFisher Scienti c) and cleared by centrifugation for 15 minutes.The concentration of supernatants was measured using Bradford colorimetric assay.Equal amounts of protein were loaded onto 10% polyacrylamide gels, subjected to electrophoresis in Tris/Glycine/SDS buffer, and transferred to PVDF membranes.The membranes were blocked with 5% milk in TBS with 0.1% Tween-20 (TBS-T) buffer for 1 hour, incubated with primary antibodies diluted in 5% BSA overnight at 4°C.After washing 3 times with TBS-T, the membranes were incubated with HRP-conjugated secondary antibodies diluted in 5% milk in TBS-T for 1 h at room temperature.Immobilon Western Chemiluminescent HRP substrate (Millipore, MA) was used for detection.

Cytokine array
were homogenized in RIPA lysis buffer with protease and phosphatase inhibitors, snap frozen, and sent to Eve Technologies for Mouse Cytokine/Chemokine 44-Plex and Mouse MMP 5-plex Discovery Assay® Arrays.

Generation of epithelial cell suspensions
Isolated epithelia were minced in 0.25% trypsin-EDTA (Thermo) and subjected to mechanical dissociation in the gentleMACS dissociator C tubes (Miltenyi #130-095-937) for 12 minutes at 37°C, followed by inactivation of trypsin and ltration.

Primary epithelial cell culture
Mouse tongue epithelial cells were isolated from mice following transgene induction as described above.Cells were grown on collagen coated plates in complete DermaCult keratinocyte basal expansion medium (STEMCELL Technologies).Medium contained the manufacturer's provided supplements, plus 5 options; normalization.method= "LogNormalize" and scale.factor= 10000.The normalized transgene expression arrays were merged with the Seurat object of the epithelial cell cluster by cell barcodes for downstream analysis.
Weighted Gene Co-expression Network Analysis (WGCNA) of scRNAseq data.
R package hdWGCNA version 0.2.03 (https://smorabit.github.io/hdWGCNA/) was used for WGCNA analysis in the scRNAseq dataset.Normalization of the integrated Seurat object containing cell-gene expression arrays of EY-genotype epithelial cells was performed using NormalizeMetacells using parameters k=10, max_shared=10, min_cells=20.A soft thresholding power was determined as 8 using the function TestSoftPowers and applied for estimation of coexpressing network in the EY-genotype scRNAseq dataset.Signi cantly co-expressed module genes and highly connected genes within each module (hub genes) were identi ed by computing eigengene-based connectivity (kME).The heatmap representing topology overlap matrix (TOM) of module genes was generated using R package ComplexHeatmap version 2.14.0.Genes with signed module eigengene-based connectivity measure (kME) greater 0.3 were considered as moderate to high con dence module genes.Modules were assigned functional annotations based on enrichment of member genes for biological processes using Enrichr 68,69 and MetaScape. 70quence: HPV16 E6-E7 Atgcaccaaaagagaactgcaatgtttcaggacccacaggagcgacccagaaagttaccacagttatgcacagagctgcaaacaactatacatgatataatattagaatgtgtgtactgcaagcaacagttactgcgacg Sequence: YAP1 S127A (S127A codon underlined and in bold)

Figures
Figure 1 YAP and E6-E7 activation is su cient to induce rapid tumor initiation in OEPCs Invasive carcinoma is preceded by the expansion of a stem-like cell population (a-d) Lineage tracing by uorescent microscopy using the H2B-GFP reporter transgene to track and quantify GFP + nuclei.(b) GSEA for gene sets associated with TEAD-and YAP-target gene expression, and mTOR activation.For gene set details, please see Supplementary Table 1.
(k) MMP gene expression across immune cell types.
(l) Experimental approach for depletion of LY6G + G-MDSCs in transgene induced EY mice.
(m) Number of in ltrative carcinoma lesions per examined tongue upon treatment in control mice and mice treated with anti-LY6G depleting antibody.
Genetically-de ned oncogenic and tumor suppressive pathway alteration in normal oral epithelial progenitor cells de nes tumor initiating events.

Supplementary Files
This is a list of supplementary les with this preprint.Click to download.

Figure 3 YAP
Figure 3 YAP promotes mTOR signaling (a) YAP target gene expression in transgenic epithelia transcriptomes.

Figure 7 CSC
Figure 7 CSC programs are enriched in HNSC and associated with poor prognosis (a) Left: Weighted gene co-expression network analysis (WGCNA) heatmap displaying topological overlap matrix (TOM) dissimilarity indices among genes in TP cluster cells.Right: Table of WGCNA modules and selected genes identi ed by Metascape and Enrichr.See also Supplementary Table 6.(b) Module enrichment in malignant tumors (T) compared to matched adjacent normal solid tissues (N) by single sample GSEA among TCGA-HNSC subjects (n=43 subjects with matched T and N samples).Two-tailed paired T-test: **p<0.01,****p<0.0001.(c-e) Kaplan-Meier plots demonstrating worse disease-free survival (n=393) among TCGA-HNSC subjects with greater median enrichment for the (c) G1/S, (d) Cell Motility & Migration, and (e) Focal Adhesion EY-modules.(f-h) Kaplan-Meier plots demonstrating worse overall survival (n=520) among TCGA-HNSC subjects with greater median enrichment for the (f) G1/S, (g) Cell Motility & Migration, and (h) Focal Adhesion modules.