Stem-cell states converge in multi-stage cutaneous squamous cell carcinoma development

Abstract

Stem cells play a critical role in cancer development by contributing to cell heterogeneity, lineage plasticity, and drug resistance. We created gene expression networks from hundreds of mouse normal and tumor samples, and integrated these with lineage tracing and single-cell RNA-seq, to identify convergence of cell states in premalignant tumor cells expressing markers of lineage plasticity and drug resistance. Two of these cell states representing multilineage plasticity or proliferation were inversely correlated, suggesting a mutually exclusive relationship. Treatment of carcinomas in vivo with chemotherapy repressed the proliferative state and activated multilineage plasticity, while inhibition of differentiation repressed plasticity and potentiated responses to cell-cycle inhibitors. Manipulation of this cell state transition point may provide a source of potential combinatorial targets for cancer therapy.

Summary:

INTRODUCTION:

Human tumors arise as a consequence of exposure to environmental agents, including mutagens and tumor-promoting chemicals, but cancer incidence is also heavily influenced by complex genetic and lifestyle factors. We replicated this complex etiology in mice by generating a genetically heterogeneous population from crosses between two diverse strains, and exposing them to initiators and promoters of skin carcinogenesis. This controlled genetic complexity enabled us to create gene expression networks (metagenes) from hundreds of normal skin and tumor samples, enabling an analysis of how gene networks, rather than single genes, evolve during multistage carcinogenesis. We applied this approach to analysis of stem cells, which play an important role in normal tissue homeostasis but are also known to be highly expressed in cancer cells, where they contribute to the cancer hallmark lineage plasticity, drug resistance, and poor treatment outcomes.

RATIONALE:

Tumor cell plasticity has been attributed to the existence of cancer stem cells (CSCs), but there is presently no consensus as to how CSCs relate to their normal tissue counterparts, to the cells of origin of tumors, or even whether they exist. Stem cells have been extensively studied in both bulk-tissue samples and at the single-cell level, but hierarchical relationships between these cell populations in cancers have been difficult to identify. We reasoned that analysis of stem-cell networks rather than of single genes may reveal such relationships. We generated gene networks for a wide range of stem-cell markers from hundreds of normal skin, benign and malignant tumor samples, and visualized expression of these network metagenes in single cells derived from multiple stages of carcinogenesis. By combining these modalities, we leveraged bulk-tissue sampling breadth and single-cell resolution to track multi-sample stem-cell states as they evolved during the transition from normal skin to malignant carcinomas.

RESULTS:

Individually, stem-cell genes were sporadically expressed in single cells without clear patterns, but analysis of their corresponding network metagenes revealed a convergence of stem-cell states in early-stage tumors that was not seen in matched normal skins. Lgr6+ stem cells from tumors generated progeny that expressed lineage plasticity markers Sox2, Pitx1, Foxa1, and Cd44, as well as genes associated with quiescence and tumor suppression. This cell state was strongly anti-correlated with an alternative state expressing markers of cell proliferation and DNA damage responses. Transitions between these states could be induced by treating carcinomas with cisplatin, which reduced proliferation and activated the cell plasticity state, or by inhibition of the master regulator of differentiation Pp2a, leading to suppression of the plasticity state and sensitization to drugs that kill dividing cells.

CONCLUSION:

By tracing the expression of stem-cell networks at the single-cell level during multistage carcinogenesis, we identified cell populations expressing two alternate cell states with opposite phenotypes: one expressed a rapid cycling cancer hallmark and the other expressed the slow-cycling lineage plasticity cancer hallmark. We propose a simplified model of carcinogenesis that identifies convergent stem-cell states at different stages of tumor progression. Manipulation of the gene networks underlying these cell states influenced the transition between them, and may provide a rich source of potential targets for combinatorial cancer therapy.

Introduction

The development of phenotypic plasticity is a common, and likely universal, feature of cancer development recently recognized as an emerging cancer hallmark (1). The mechanistic basis for this plasticity, defined as the redeployment of lineage-specific gene expression programs along alternative cell fate trajectories, is presently unclear (2). Elucidation of this process has major practical implications for understanding and treatment of cancer, since cell plasticity has been associated with invasion, metastasis, and resistance to chemotherapy or targeted drugs (3–5). Tumor cell plasticity has been attributed to the existence of cancer stem cells (CSCs), but there is presently no consensus as to how CSCs relate to their normal tissue counterparts, to the cells of origin of tumors, or even whether they exist (6, 7). Different models suggest that CSCs may lie within a hierarchy of differentiation within tumors (8, 9) or they may show complete plasticity, being essentially interconvertible (10).

Here, we address these questions using mouse skin, which has over several decades been the most widely studied solid tissue for analysis of stem-cell function in normal homeostasis and during oncogenic transformation (11–13). In the mouse, a number of distinct stem-cell markers have been identified which contribute to homeostasis by repopulating specific compartments within the epithelium (14, 15). Certain healthy-skin stem-cell markers, including Lgr5 (16), Lgr6 (17), Twist1 (18), Sox2 (19), and Pitx1 (20), mark cancer stem cells (CSCs) in several tumor types (9, 21) To identify the cell plasticity states that arise during transitions from normal tissue through pre-neoplasia to malignant carcinomas, we developed a model system that encompasses both germline genetic and somatic genetic diversity to construct gene expression networks in each stage, and visualized these networks in single cells representing the continuum of steps in carcinogenesis. Our data identify connections between different stem-cell populations in the skin, leading to a more unified model of cell state transitions during tumorigenesis.

Results

Rewiring of stem-cell gene expression networks in tumors

Although progress has been made in unravelling gene expression networks in single cells (22), understanding global dysregulation of these networks in cancer continues to rely upon bulk-tissue data (23). Bulk-tissue data are useful when inferring gene networks since they encompass hundreds of independent samples and thus are likely general features of a system, occurring repeatedly across independent instantiations of carcinogenesis. On the other hand, features in single cells are usually assayed in a small number of samples because of the throughput trade-off between the number of samples and the number of cells that single-cell ‘-omics’ can measure. Thus, single-cell datasets often represent specific contingent outcomes of a particular sample rather than general system features (24). Together, the sampling breadth of bulk-tissue data and high resolution of single-cell data are complementary. We leveraged this complementarity by generating gene networks from hundreds of bulk samples and resolving their expression in single cells. Combined with stem cell biology, we are thus able to elucidate how these conserved features operate in single cells to drive initiated cells from normal homeostasis to cancer.

We previously carried out genetic and gene expression analysis of multistage chemically-induced carcinogenesis in genetically heterogeneous mice (25–27). These chemical carcinogenesis models capture a realistic view of cancer development in human populations since 1) tumors are induced by environmental insult, rather than genetic modification, and carry thousands of somatic mutations including many cancer driver mutations that are also seen in human tumors (26, 28, 29); 2) they are autochthonous to the animals in which they are generated, rather than transplanted into immunodeficient host animals; and 3) the treated mice are from a genetically heterogeneous mouse population that mimics human germline diversity. From this, we generated a transcriptomic database from several hundred samples that span progression from normal skin to benign pre-neoplasia (papillomas in the skin) to malignant tumors (squamous cell carcinomas) and finally metastasis (25, 26). A similar database has been used to infer functions for specific genes based on their network structures in bulk-tissue samples (25, 27, 30, 31). Here, we profiled gene expression of 106 normal skin and 157 carcinoma samples from interspecific Mus spretus x FVB/N backcross mice, generated as previously described (32) (Fig. 1A). These networks are based on gene-gene correlations across samples, and thus capture direct and indirect interactions among genes. For the present analysis, we considered indirect interactions to be the biologically relevant effects of a gene’s function, so we did not seek to limit network inference to specific physical interactions such as transcription factors and their targets (33) or among their protein-protein interactions (34). Furthermore, since bulk samples contain mixtures of different cell types, these networks capture the extended function of network seed genes expressed across different cell types (fig. S1).

Fig. 1. Bulk-tissue and single-cell expression data.

(A) Experimental pipeline showing creation of an interspecific backcross mapping population segregating genetic variation in resistance to inflammation, infection, and cancer. Mice were randomly selected for subsequent matched normal skin and carcinoma expression analysis. Tumor initiation occurred by DMBA application to dorsal back skin and promotion by twice weekly TPA treatment for 20 weeks. Matched normal skin and tumors were co-sampled for bulk gene expression assays by microarray. Co-expression gene networks for genes of interest were inferred from bulk expression data. Several normal skin, papilloma, and carcinoma samples from inbred FVB mice were profiled using scRNAseq. Network expression inferred from bulk samples was overlaid onto the single-cell data as tissue-specific metagenes: a normal-skin metagene generated from normal skin bulk-tissue samples and a carcinoma metagene generated from carcinoma bulk-tissue samples. Spearman rank correlations of selected stem cell-related genes in (B) 106 normal skin and (C) 157 carcinoma samples. (D) Single-cell data from normal skin, papilloma, and carcinoma from FVB mice displayed on a UMAP with cells colored according to cell type; OuB is for outer bulge; Isth isthmus; UnF undifferentiated follicular; InB inner bulge; SG sebaceous gland; UHF upper hair follicle; Bas basal cells; tK terminal keratinocytes; IfE interfollicular epidermal cells; Inf infundibulum; mEp mitotic epithelial cells; Er erythrocytes; Mel melanocytes; T T-cells; nEp neoplastic epithelia of the papilloma; LS lower spike cells (defined hereafter); US upper spike cells (defined hereafter); Sp spindle cells of the carcinoma; Sq squamous cells of the carcinoma; Neut neutrophils; LC Langerhans cells; Mac macrophages; End endothelial cells; Myo myofibroblasts.

To gain an unbiased view of the global gene expression network, we implemented the unsupervised inference algorithm WGCNA, which clusters genes sharing similar expression from bulk-tissue samples into gene modules (35) (Data S1). We then tested these modules for functional enrichment using biological process gene ontologies (GO) (Data S2). One large module in carcinomas, Module 3 (N_gene=1720), was highly enriched for genes associated with wound healing. This wound healing module included several known stem-cell marker genes that have been implicated in carcinogenesis, including Sox9, Psca, Pitx1, Krt15, Krt19, and Lgr6. Lgr6 is of particular interest in this system since Lgr6+ cells are highly clonogenic in squamous carcinomas (17), play an important role in wound healing (36), and can drive extreme lineage plasticity by repopulating all skin cell types (37).

We next examined the correlation structure of these well-characterized stem-cell marker genes in normal tissues and tumors. This revealed three clusters of genes that were correlated with each other in tumors but not in matched normal tissues (Fig. 1B,C). We refer to these groups of correlated stem-cell genes that are co-expressed at the bulk-tissue level as ‘Spearman groups.’ These Spearman groups recapitulated the gene modules identified by WGCNA (Data S3). Importantly, Spearman group 1, consisting of Bmi1, Sox4, and Lrig1, was strongly anti-correlated to Spearman group 3, the latter containing multiple known adult stem-cell markers.

Each stem-cell marker was then used as a “seed gene” to generate normal skin- and carcinoma-specific gene correlation networks (called “metagenes,” Supplementary Text, Data S4), which can capture the downstream function of specific genes in bulk-tissue samples (25, 27, 30, 31). The Lgr6 metagene showed major network rewiring between normal skin and carcinoma. In normal skin, the Lgr6 network included of several epidermal lineage markers including Krt15 (ρ=0.78) and Klf5 (ρ=0.63), as well as Znrf3 (ρ=0.51) and Rnf43 (ρ=0.55), the two E3-ligases that act as R-spondin coreceptors to mediate Wnt signaling (38). However, during carcinogenic rewiring, these correlations were lost, replaced by Sox9 (ρ=0.67), as well as other important drivers of cancer phenotypes including Tgfb1 (ρ=0.60) (table S1). Tgfb1 is a growth factor known to play multiple roles in tumor development, acting as a growth inhibitor at early stages and switching roles to become an inducer of the epithelial-mesenchymal transition (EMT) during tumor progression (39). It is also a major mediator of immunosuppression that, when inhibited with blocking antibodies, improves immunotherapy outcomes in mouse models (40, 41). Thus, changes in the correlation network architecture of the Lgr6 metagene during tumor development revealed biologically functional rewiring associated with cancer progression and immune escape.

Single-cell transcriptomes in multi-stage carcinogenesis

Since our goal was to understand how gene networks inferred from hundreds of bulk samples can be visualized in single cells, we then obtained single-cell RNA sequencing (scRNAseq) for 57,807 cells, of which 33,234 (57.5%) cells were normal skin from 4 mice, 15,280 (26.4%) were papilloma from 1 mouse, and 9,293 (16.1%) were carcinoma from 1 mouse (Fig. 1C), and plotted metagene expression within the resulting UMAP. Additionally, we included in this analysis single epithelial cells from normal skin, benign papillomas, and carcinomas that were selected by lineage tracing in the Lgr6-eGFP-tdTomato mouse strain (17), in order to mark and separate Lgr6+ stem cells (eGFP+) from their direct progeny (tdTomato). Later, we replicated these data with an additional 7,883 cells from 4 carcinomas and 7 papillomas (see below). Critically for our first analysis, to make their transitional relationships analytically tractable, the proportion of Lgr6+ and progeny cells were increased relative to unsorted cells by loading equal numbers of Lgr6+, progeny, and unsorted cells for scRNAseq. We identified 21 low-resolution single cell clusters by unsupervised clustering, of which cell clusters 1–4 dominated the assembly, predominantly populated by normal skin (clusters 1,2), papilloma (cluster 3), and carcinoma (cluster 4), which we consider to be tissue-specific parenchymal cells (fig. S2A,B). We classified normal skin cell types with canonical markers and re-analyzed the carcinoma parenchyma alone, which showed two distinct keratinocyte subpopulations corresponding to squamous (20.8% of carcinoma parenchyma) and spindle phases (79.2%) that had undergone an epithelial-mesenchymal transition (fig. S2C–E,S3) (42, 43).

Scaling gene networks from bulk tissue to single cells: metagenes quantify extended gene function

To visualize how metagenes are expressed at the single-cell level, we first examined Lgr6 and its closely related family member Lgr5. Both genes are known stem-cell markers in the skin, but Lgr6 is clonogenic in squamous cell carcinomas whereas Lgr5 is not (17). Lgr6 as an individual gene was expressed in the isthmus region of the upper hair follicle (arrow, Fig. 2A), as well as in the interfollicular epithelium (IfE). Prior lineage tracing studies have shown that Lgr6+ cells repopulate these regions during normal homeostasis and wound healing (37, 44, 45). These cells then differentiate into the overlying specialized epidermal strata. The Lgr6 normal-skin metagene (the aggregate expression of the 100 genes most highly rank-correlated with Lgr6 across multiple bulk-tissue samples of normal skin) reflects this developmental pattern, showing increasing expression across interfollicular keratinocytes that peaks in terminally keratinized cells from the granular layer of the epidermis (arrow, Fig. 2B). Since this metagene consists of a gene network derived from independent bulk-skin samples, which contain mixtures of interfollicular and follicular cells, we tested whether this metagene was expressed in hair follicle cells as well. Indeed, it was robustly expressed in normal skin (NSk) hair follicle cells (fig. S4A,B), although its expression was highest in terminal keratinocytes (fig. S4C). Furthermore, the individual genes that constitute the Lgr6 normal-skin metagene are enriched for keratinocyte proliferation and differentiation (Data S5), implying that this network captures Lgr6’s function to establish epidermal cell layer patterning more strongly than its complex functions in other skin compartments such as the sebaceous gland or hair follicle (46). The Lgr6 carcinoma metagene, on the other hand, shows high expression in the carcinoma parenchyma (arrow, Fig. 2C), reflecting the functional switch during tumor progression from normal homeostasis to malignancy (see also table S1).

Fig. 2. Single-cell expression of stem-cell seed genes and metagenes.

(A) Expression of Lgr6 alone in the single-cell UMAP. (B) Expression of the Lgr6 normal skin metagene defined as the top 100 genes correlated with Lgr6 in normal skin (NSk) bulk-tissue samples. (C) Expression of the Lgr6 carcinoma metagene defined as the top 100 genes correlated with Lgr6 in carcinoma (Car) bulk-tissue samples. (D) Expression of Lgr5 alone in the single-cell manifold. (E) Expression of the Lgr5 carcinoma metagene defined as the top 100 genes correlated with Lgr5 in Car bulk-tissue samples. (F) A comparison of metagene expression during carcinogenesis for the Lgr5 and Lgr6 carcinoma metagenes. Y-axis is log metagene expression for each stage-specific parenchyma cell; box-and-whisker plots are median+IQR with outliers (Q1–1.5×IQR or Q3+1.5×IQR) not shown. All within-plot pairwise comparisons are significant at FDR<0.05. (G) Exemplar carcinoma metagenes that localize to the upper spike (US, upper panels) and lower spike (LS, lower panels) of the papilloma. The location of the spikes is colored red in the panels next to the metagenes and corresponds to unsupervised cell clusters 19 and 33. Purple is low expression and yellow is high expression.

Lgr5 is highly expressed in the lower bulge region of the hair follicle, reflecting its known follicular stem-cell function (arrow, Fig. 2D). Unlike Lgr6, the Lgr5 carcinoma metagene shows low expression across all stages of carcinogenesis (Fig. 2E, fig. S5). Thus, metagene expression of the Lgr6 carcinoma network increases, while that of the Lgr5 carcinoma network decreases, in parenchymal cells during tumor progression (Fig. 2F). This demonstrates that metagene expression in single cells traces the functional role of these stem-cell markers in normal tissue and reflects functional rewiring that occurs during malignant transformation.

Convergent expression of multiple stem-cell metagenes in the same cell populations.

Since Spearman groups 1 and 3 of stem-cell genes were strongly anticorrelated in tumor samples (Fig. 1C), we reasoned that they may represent two mutually exclusive stem-cell states expressed in distinct single-cell populations. To test this hypothesis, we examined whether stem-cell markers and their metagenes were expressed in different cell populations in the single-cell data. The expression of individual stem genes was notably variable across single cells, with sparse and divergent expression in disparate cell populations (seed gene panels in fig. S6). However, metagene expression dramatically altered this landscape. Spearman group 3 carcinoma metagenes co-localized to a specific “spike” cell population corresponding to cell cluster 33 in the UMAP of the papilloma (lower spike, Fig. 2G). “Spike” refers to the appearance of this cell population in the UMAP. For example, carcinoma metagenes for Pitx1, Krt15, and Psca were highly and specifically expressed in this lower spike population.

To investigate possible overlapping functions of stem networks in the lower spike, we tested for gene ontology (GO) enrichment in genes of the ten stem-cell Spearman group 3 metagenes that co-localized to the lower spike. GO analysis showed that this gene set was highly enriched in functions related to oxidative stress (Duoxa1/2), skin barrier formation (Sprr1a, Sprr3), wound healing (Wnt4, Klf5), cell migration (Cd44, Ceacam1), apoptosis (Anxa1), and immune responses (Arg1) (table S2). Notably, normal wound healing is associated with oxidative stress and lineage infidelity, the latter indicative of stem cell plasticity and characterized by de novo co-expression of genes representing different hair follicle and epidermal lineages (21). Several genes associated with stem-cell plasticity were prominent in this set of lower spike metagenes (including Sox15, Foxa1, Klf5, Cd44, Wnt4), consistent with the development of lineage infidelity in the lower spike region. Despite the individual seed genes being expressed in disparate cell populations, these stem-cell markers converged at the metagene level with high expression in the same single-cell population, and this convergence cohered around lineage plasticity and wound healing.

In contrast, carcinoma metagenes for Bmi1 (Spearman group 1) and Lgr6 (Spearman group 2) were almost absent from the lower-spike cell population (fig. S6). The upper spike cells showed higher expression for Spearman group 1 metagenes, with distinct enrichment in expression of metagenes corresponding to markers of cell cycle progression, including E2f1 and Foxm1, known as a master regulator of proliferation and malignant progression (Fig. 2G) (47). The upper spike was composed of a much larger proportion of cells in the G2M phase than the lower spike or other neoplastic cells: 20.6% versus 0.52% and 13.29%, respectively (table S3, figs. S7,S8). Furthermore, mitotic metagenes such as E2f1, Foxm1, and Mki67 were significantly (p<0.01) more highly expressed in G2M cells than in G1- or S-phase cells (fig. S7E–G). Similar metagene patterns were seen for multiple markers of DNA damage/genomic instability, for example those corresponding to Atm and Atr (fig. S9). Together, we take this to indicate that the upper spike cells highly express gene networks related to DNA replication, mitosis, and cell division.

We then hypothesized that if the stem-cell Spearman groups 1 and 3, which were negatively correlated at the bulk-tissue level, were truly mutually exclusive gene programs at the single-cell level, then their anticorrelated genes would be expressed primarily in mutually exclusive cell populations. We refer to the set of a seed gene’s most highly anticorrelated genes as the ‘negative metagene.’ Strikingly, plotting Spearman group 1’s negative metagenes showed high expression in the lower spike region (fig. S10). This points to the mutually exclusive relationship between Spearman groups 1 and 3 carcinoma metagenes in these distinct single-cell populations.

Alternative stem cell populations arise from Lgr6-positive papilloma cells.

Having identified two distinct single-cell populations with high expression of alternate stem programs (the upper and lower spikes), we then asked how they relate to Lgr6, which is a major driver of clonogenicity in this system. We tested this directly through lineage tracing and immunofluorescent analysis of normal skin, papillomas, and carcinomas (Fig. 3). This showed that in normal skin Lgr6+ cells repopulated the upper hair follicle and interfollicular epidermis (Fig. 3A), but during neoplastic progression, expression was more widespread and disorganized as Lgr6 marked clone-initiating cancer stem cells (17). In papillomas, rare Lgr6-GFP+ cells were primarily located at the basement membrane, and their tdTomato+ progeny formed streaks that extended into the upper differentiated cell layers (Fig. 3B) (48). Conversely, carcinomas showed highly disorganized tissue-level architecture (Fig. 3C). In order to understand how the Lgr6 metagene differed in papillomas and carcinomas more precisely, we used a previously published bulk expression dataset of 68 papillomas and 60 carcinomas to generate papilloma- and carcinoma-specific Lgr6 metagenes (30). This showed Lgr6 network rewiring consistent with the tissue tracing: Lgr6 was strongly correlated in papillomas with genes representative of the papilloma lower spike including Foxa1 and Krt19, but these correlations were weaker in the carcinomas, where Lgr6 is more linked to self-renewal (Data S6). Furthermore, the Lgr6 papilloma metagene shows stronger expression in the lower spike, whereas the Lgr6 carcinoma metagene generated from this separate cohort of independent carcinomas shows the same expression pattern as the main cohort of carcinomas presented in this study (fig. S11).

Fig. 3. Immunofluorescent single-cell lineage tracing of Lgr6+ cells and their progeny.

Lgr6 lineage tracing in normal dorsal mouse skin, benign papilloma, and carcinoma tissue. Representative immunofluorescence images of Lgr6-driven lineage tracing for: (A) normal dorsal mouse skin; (B) benign papilloma tissue; (C) carcinoma, 10 days after topical 4-OH-tamoxifen treatment in vivo. Lgr6+ stem cells (green) are localized within (a) the hair follicle epithelium and epidermis of the normal dorsal skin; (b) predominantly the basal epithelium of the papilloma; (c) scattered throughout the carcinoma epithelium. These give rise to tdTomato+ (red) progeny within those tissues. Yellow boxes indicate the magnified region, dotted lines represent the epidermal-dermal border, and nuclei were counterstained with DAPI (blue), scale bar = 50 μm.

We then physically separated the Lgr6:GFP+ stem cells in normal skin, papillomas, and carcinomas from their respective tdTomato+ progeny cells by flow cytometry, and examined the distribution of these cells within the UMAP (Fig. 4, figs. S11 and S12). Firstly, Lgr6:GFP+ cells were considerably enriched in the upper spike (18.6% of cells) compared to the lower spike (3.4%) or to the remainder of the papilloma parenchyma (4.1%, Fig. 4A). Since these tumors initiate from Lgr6+ cells, this points to the upper spike as earlier in developmental time than other papilloma cells. We next performed a stage-specific differentially expressed gene (DEG) analysis comparing Lgr6+ and progeny cells (Data S7). At each stage (NSk, Pap, and Car), the DEGs that increased most in the progeny cells relative to Lgr6+ cells were most highly expressed in the lower spike (Fig. 4B). The overall intensity of expression, however, decreased from normal skin to papilloma, and then to carcinoma, in line with the known propensity for loss of differentiation capacity during malignant progression. We then tested how expression of specific stem-cell metagenes changed when Lgr6+ cells gave rise to their progeny in benign or malignant tumors. Specifically, Spearman group 3 metagenes increased in expression across the Lgr6→progeny trajectory (Fig. 4C). Conversely, Spearman group 1 metagenes (Bmi1, Sox4, Lrig1) showed decreased expression across the Lgr6→progeny trajectory. Taken together, these differential shifts in stem network expression across this empirically traced Lgr6→progeny lineage are strongly suggestive of a hierarchical relationship between stem-cell genes in these populations. In conjunction with the metagene patterns, these lineage-tracing results point to two highly divergent cellular fates: 1) the upper spike maintains Lgr6-associated self-renewal and proliferation, and 2) the lower spike shows increased lineage infidelity, plasticity, and commitment to different cell fates.

Fig. 4. Lgr6-progeny cells form the papilloma lower spike and express high levels of a Spearman group 3 program.

(A) Proportion of FACS-sorted Lgr6, progeny, and unsorted cells in each cell population. (B) Aggregated expression of top 100 DEGs by fold change from stem cell to progeny cells (tdTomato+/ Lgr6:GFP+) discovered for DEGs within normal skin parenchyma; papilloma parenchyma; and finally carcinoma parenchyma. (C) Carcinoma metagene expression means across individual cells of the Lgr6:GFP+ (L) and tdTomato+ Progeny (P) lineages taken from neoplastic parenchyma cells (pooled papilloma and carcinoma parenchyma); red and blue lines are significant differences at Bonferroni-adjusted p<0.01 by t-tests, and gray are non-significant; error bars are standard errors.

Evaluating the lower and upper spike paradigm in independent tumors

In order to ensure that the single-cell metagene patterns are not a specific and contingent outcome of the particular single-cell samples gathered in the original data, we performed additional experiments to generate new, independent tumor samples. Specifically, we analyzed four additional carcinomas from 4 mice and seven additional papillomas from 5 mice by scRNAseq. All additional samples contained lower and upper spike cells (fig. S12 table S4). Fig. S12B shows expression of a Pitx1 carcinoma metagene whose peak expression colocalizes with consensus lower spike cells (fig. S12C). Furthermore, when integrated with the original data, these additional samples maintain the upper spike as an early transitional population between normal skin and neoplastic cells (fig. S12). This position was robust to the removal of a mitotic expression signature, with the upper spike being intermediate between normal skin and the papilloma even after mitotic signal was regressed out (fig. S13). Together, the hundreds of independent bulk samples that generate metagenes combined with the recapitulation of metagene expression patterns in these additional independent tumors point to the external validity and generalizability of this paradigm.

The lower-spike cell state is implicated in drug resistance and conserved in mouse and human tumors

Since stem-cell plasticity in tumors has been commonly associated with resistance to chemotherapy or targeted drugs (49, 50), we sought to determine whether lower-spike cells expressed empirically derived markers of drug resistance identified from drug screens. We examined markers of drug resistance in several human tumor types including basal cell carcinoma (BCC), prostate adenocarcinoma (PAD), and lung adenocarcinoma (LUAD). BCCs from patients treated with vismodegib (51) showed elevated expression of resistance genes Tacstd2, Ly6d, and Lypd3. We generated carcinoma metagenes for these seed genes using our mouse bulk expression data and visualized their expression in our mouse single-cell data. These BCC drug-resistance metagenes were most highly expressed in the lower spike cells in papilloma (fig. S14). A high-plasticity cell state that contributed to cell heterogeneity, drug resistance, and poor patient prognosis was also identified in human and mouse lung cancers (52). The most specific markers for this cell state were Slc4a11 and Tigit, the latter being a marker of immune responses and a cancer drug target (53). We again generated carcinoma metagenes for Slc4a11 and Tigit and found that they too were highly expressed in the lower spike (fig. S14). Finally, Foxa1 has been identified as a marker of drug resistance that can be mutated in human prostate cancers and contributes to lineage plasticity during prostate adenocarcinoma development (54, 55). Although sporadically expressed as a single gene, its carcinoma metagene also specifically localizes to the lower spike (fig. S14). Together, these metagenes show that treatment-responsive genes discovered from a variety of human cancers were most highly expressed in the same lower spike population.

To test human orthology of the spike cell states directly, we examined scRNAseq from human prostate adenocarcinoma (PAD, Fig. 5). These data consisted of 13 prostate tumors comprising 36,424 cells and 11 control samples comprising 46,117 cells of healthy prostate (56). We first resolved human orthologs of our 100-gene lower-spike signature in PAD single cells, which localized to a small population in the PAD, which we hypothesized to be the human equivalent of the mouse lower spike (orange arrows Fig. 5B). To test this, we then reversed this process by examining a gene set that an independent study found to mark a “regenerative” drug-resistant stem-cell population in PAD (57). In the human PAD data, this gene set is most highly expressed in the same cell population as the LS signature (Fig. 5D). More conclusively, when we resolved these human genes in our mouse data, their joint expression was highest in the lower spike (Fig. 5C). Thus, the lower-spike cell state recapitulated in both directions along this mouse-human orthology, independent of whether the orthology initiated from the human or the mouse: the lower-spike cell state discovered in mice localized to the human regenerative PAD population, and the regenerative PAD cells discovered in humans localized to the mouse lower spike.

Fig. 5. Recapitulation of lower spike signature genes in orthologous human prostate adenocarcinoma and vice versa.

The top panels show expression of lower spike (LS) markers in A) mouse cells and B) human prostate adenocarcinoma (PAD). Bottom panels show expression of human prostate regenerative signature in C) mouse cells, and D) PAD cells. Orange arrows point to cells with peak expression. Black arrows show direction of orthologous transformation origin: LS signature was discovered in mouse SCC and was preserved in human PAD; the PAD regenerative signature was preserved in mouse SCC.

To functionally test whether genes that mark the lower spike are indeed upregulated during chemotherapy treatment, animals bearing primary carcinomas were treated in vivo with the commonly used chemotherapeutic agent cisplatin and profiled with scRNAseq. We tested cisplatin-vs-control carcinomas for DEGs, finding 2461 significantly upregulated and 1298 significantly downregulated genes (Fig. 6A, Data S8, q<0.05). Upregulated genes were significantly enriched for cellular responses to stress involved in drug resistance such as reduction-oxidation, mitotic exit, and anti-apoptotic genes (Data S9, p<0.05). We also interrogated these cell populations for potential molecular mediators of drug resistance by testing for significantly co-upregulated ligand-target partners, and found a number of immune-epithelial cross-talk pathways activated between the lower spike cells and components of the immune system (Data S10, p<0.03). Plotting the expression of the upregulated genes from these independent tumors in the original data showed that they were most highly expressed in the lower spike (Fig. 6B,C). We also tested whether the overlap between genes upregulated by cisplatin treatment and lower-spike marker genes was greater than expected by chance. This analysis (Fig. 6D) shows that upregulated cisplatin DEGs were the same as lower-spike markers more often than expected by chance. Thus, both cisplatin DEG expression and overlap with lower-spike markers orthogonally validate our interpretation of the lower spike as a chemotherapy-responsive cell population.

Fig. 6. Cisplatin activates genes that localize to the lower spike, whereas Pp2a inhibition elevates expression of upper spike genes.

(A) Differentially expressed genes in squamous cell carcinoma parenchyma cells treated with cisplatin in vivo. Data derived from scRNAseq of 2 control and 2 treated tumors. Positive values represent genes that are upregulated by cisplatin, and negative downregulated. Labelled genes are those mentioned in the main text. (B) Integrated expression of the top 100 genes significantly upregulated by cisplatin ordered by fold-change (p<0.05). (C) The intersection of 1) genes that define a cell cluster ordered from high to low marker score and 2) genes that are upregulated by cisplatin challenge ordered from high to low fold change. The x-axis refers to the number of genes concurrently tested for intersection size in each list. For example, 100 refers to the top 100 marker genes in a cluster and the top 100 positively upregulated genes, which are then tested for the number of overlaps (intersection size), which is displayed on the y-axis. The red line shows the results for lower-spike markers, the blue for upper-spike markers, and gray lines show 1000 randomly chosen sets of n = 100 genes to establish a null distribution of intersection curves between cisplatin DEGs and potential marker genes in the scRNAseq data. (D) Box and violin plots of integrated expression of the top 100 positively upregulated DEGs shown in panel B. nEp refers to all papilloma cells that are not in the spikes; US is for upper spike (cluster 19); and LS is for lower spike (cluster 33). All pairwise comparisons are significant at p<0.01. (E) Integrated expression of the top 100 genes significantly upregulated by Pp2a inhibition with LB100 ordered by fold-change (p<0.05). (F) Box and violin plots of integrated expression of the top 100 positively upregulated DEGs shown in panel E. nEp refers to all papilloma cells that are not in the spikes; US is for upper spike (cluster 19); and LS is for lower spike (cluster 33). LS cells are significantly different at FDR<0.01.

These data demonstrate that targeting mitotic cells with cisplatin reduced the population of upper spike cells and increased the expression of lower spike markers. We sought to determine whether the reverse is also true, by testing whether inhibition of lower spike differentiating cells could in turn activate the upper spike-cell state. We accomplished this by inhibiting Pp2a, a tumor suppressor gene and inducer of differentiation, whose loss is a key step in malignant transformation to anchorage-independent mitosis in a number of cancers (58). We treated 2 mouse SCC cell lines in vitro with LB100, a potent Pp2a inhibitor (59), quantified expression differences with RNAseq, and plotted the top positive DEGs within our original papilloma cells (Fig. 6E, Data S11). Indeed, this showed high expression in the upper spike with significantly lower expression in the lower spike (Fig. 6F,p<0.01). Together with cisplatin’s induction of a lower-spike expression signature, we interpret these two spike cell states as alternative fates that can be toggled by selective treatment pressures.

To further explore the therapeutic implications of this transition point, we treated SCC cells in vitro with the Pp2a inhibitor together with the mitotic toxin paclitaxel. We found that SCC colony growth was not significantly different from control when treated with LB100 or paclitaxel alone but was significantly reduced when sequentially treated with LB100 followed by paclitaxel (fig. S15, p<0.05). Indeed, this chemotherapy sensitization strategy with Pp2a inhibition has been found to be effective in a large number of human cancers (60). Our data elucidate the cell-state mechanism that underlies this apparently paradoxical phenomenon of treating cancer by suppressing a tumor suppressor: chemotherapy sensitization by Pp2a inhibition is caused by an induction of the upper-spike cell state, marked by increased mitosis, which is then vulnerable to chemotherapeutic toxicity.

Finally, in order to integrate these observations into a coherent evolutionary trajectory of single-cell-resolution carcinogenesis, we explored the single-cell data with two additional frameworks. First, we used pseudotime inference with inferred trajectories rooted in the normal skin (Fig. 7A). Globally, the pseudotime trajectory was accurate since the carcinoma was correctly identified as occurring later in developmental time than the papilloma, although in the UMAP the carcinoma appears more proximate to the normal skin. At a more regional level within the papilloma, the inferred trajectories with pseudotime again showed the upper spike to be the earliest neoplastic population. To ensure that the upper spike’s position was not simply a random and arbitrary outcome of UMAP data representation (61), we tested whether the upper spike maintained this transitional position by analyzing these data using force-directed layouts. These similarly showed that the upper spike’s intermediate position between normal skin and neoplasia was robust to these orthogonal frameworks (fig. S16).

Fig. 7. Whole-chromosome aneuploidy and pseudotime imply the upper spike is transitional between normal skin and neoplasia.

A) Inferred trajectories and pseudotime of parenchyma cells and the papillomas. B) UMAP showing cells with chromosome complements greater than regular diploid of any chromosome in parenchyma cells. C) Gains (red) of chromosome 7 in the papilloma parenchyma; D) chromosome 6; E) chromosome 10; F) chromosome 1.

Finally, to obtain further genomic support of the pseudotime trajectories, we inferred whole-chromosome duplications or aneuploid gains within our single cells (Fig. 7B–F). Bulk-tissue papillomas and carcinomas contain cells with chromosome 7 and 6 duplications, and carcinomas show complex additional patterns including gains of chromosomes 1 and 10 (26, 62, 63). In our single-cell data, the upper spike had very few aneuploid cells, suggesting that these are early pre-malignant cells, in agreement with the pseudotime trajectory. There was a clear increase in the density of aneuploidy in chromosomes 7, 6, 10, and 1 in the papilloma moving from right to left across UMAP1 (Fig. 7C–F). All of these chromosomes duplicated uniformly in the carcinoma, showing that the final malignant stage does result in co-occurring gains of all 4 chromosomes.

Together this leads to the following model of carcinoma development: initiated cells adopt a rapid cycling phenotype; these cells then enter a critical transition point in which they can 1) adopt a high-plasticity, quiescent stem state encapsulated by the lower spike or 2) maintain rapid cycling and acquire increased aneuploidy moving towards malignancy.

DISCUSSION

Identifying and quantifying cell plasticity is an emerging challenge in cancer biology (1, 2). Numerous studies have identified a wide array of specific genes that contribute to plasticity signatures, but a common theme has emerged: at the level of individual cells, there appears to be a relatively small number of developmental paths towards malignancy (52, 64, 65). Here, we directly explored plasticity states in vivo by identifying expression networks for known stem-cell genes and tracing their expression across a continuum of stages of carcinogenesis at the single-cell level. The mouse model we used encompassed germline genetic heterogeneity, somatic genomic events, and gene expression across multiple samples of normal, premalignant, and malignant tissues, to identify the steps necessary for rewiring of stem-cell networks at each stage. Tumors in this model were induced by sequential exposure to both mutagens and tumor promoters that elicit chronic inflammatory responses, factors increasingly seen to be critically important in human cancer etiology (66–68). We used hundreds of independent samples from this model to generate metagenes, thereby identifying general features of gene expression network rewiring in this system. We then analyzed single-cell expression patterns of these metagenes, revealing high expression in two distinct populations of pre-malignant cells associated either with quiescence and markers of stress-induced lineage plasticity (lower spike), or with proliferation and self-renewal (upper spike).

The biological relevance of these populations is that they represent two alternative cell states with opposite phenotypes, both of which must be considered if cancer is to be cured: the upper spike expresses the uninhibited cycling hallmark of florid cancer growth whereas the lower spike expresses the slow-cycling plasticity hallmark of treatment resistance. More importantly, these cell states appear to be mutually exclusive as expected for alternative cell fate decisions, shown by the strong negative correlations between markers of the upper and lower spikes (Fig. 1C, fig. S10). These cell populations occurred in pre-malignant tumors, but whether they progress or are culled at this stage may depend upon further genetic changes, such as loss of function of tumor suppressor genes, which occurs frequently during benign-malignant transformation (42, 69). For example, we observed the loss of Lgr6’s correlation with tumor-suppressing WNT degradation genes in the carcinoma network (table S1), and loss of function in these genes is known to unleash the WNT-β-catenin pathway to sustain uncontrolled cell growth in both mouse and human SCCs (70). Further, large-scale genetic changes quantified as copy number variants oriented evolutionary trajectories in the single-cell data and clarified the relationship between cell states seen in papillomas (Fig. 7). Together with expression data, they pointed to a critical transition point that emerged between upper spike cells that could maintain rapid cycling to move onto a more progressed state and lower-spike cell that could enter a high-stemness cell state. Together, this spike paradigm is reminiscent of bifurcations between proliferative and quiescent stem-like states that have been mechanistically traced in vitro to control of the cell cycle restriction point by balanced activities of Cdk2 and the tumor suppressor gene p21/Cdkn1a (71). Together, this two-spike paradigm provides a mechanistic basis for the observations that developmental paths to malignancy occur through a relatively small number of routes.

Although plasticity in single cancer cells can be driven by stem-gene expression, there remains considerable controversy surrounding the relationships between stem-cell populations in normal tissues, and those that are found in tumors (72). One view is that a tumor stem-cell hierarchy exists (73), while others suggest that the chaotic, stressful environment associated with tumor growth induces extreme plasticity and heterogeneity, with no clear evidence of a hierarchical relationship between stem cells and their progenitors (74). Since decades of empiric effort have conclusively demonstrated that the stem-cell genes we have analyzed constitute bona fide stem-cell markers in skin, we reasoned that we could test how their relationships changed both at the bulk level through expression reorganization, and at the single-cell level by measuring their expression in an empiric lineage-tracing system. At the bulk level, we uncovered at least three distinct tumor stem-cell co-expression groups exemplified by Spearman groups 1) Bmi1, Sox4, and Lrig1, 2) Lgr6 and Sox9; and 3) Sox2, Pitx1, Klf5, Psca, Cd44, Prom1, and other well-known stem-cell markers. Spearman groups 1 and 3 appear to represent alternative states, as their expression patterns are strongly anti-correlated, and as such may represent a major transition point in cells that undergo malignant conversion. In support of this possibility, at the single-cell level, markers of Spearman group 3 are specifically and highly expressed in a subpopulation of pre-malignant cells characterized by lineage infidelity, oxidative stress, and wound healing. However, these cells also express high levels of counterbalancing gene programs such as protective immunogenicity, apoptosis, senescence, and tumor suppressor activity. The alternative cell state linked to Spearman group 1 (Bmi1, Sox4, Lrig1) lacks this protective antagonistic pleiotropy and is marked by increased expression of cell cycle and DNA damage response genes. Both cell states may arise from Lgr6+ stem cells through alternative self-renewal and lineage commitment cell fate decisions (Fig. 8).

Fig. 8. Multistage chemical carcinogenesis model of initiation and promotion in single cells reveals simplified stem-cell paths to malignancy.

The model for carcinoma development in single cells based on co-expression networks from bulk-tissue samples. DMBA initiates carcinogenesis by inducing oncogenic Hras mutations in cells, which then lie dormant until promoted by the inflammatory promoter TPA. A single initiated Lgr6+ papilloma cell or its early progenitor clonally expands and begins to divide rapidly, developing towards the upper-spike cell state characterized by cell cycle activation and genomic instability. This results in specific aneuploid gains and subsequent movement towards a critical transition point in the papilloma. At this transition point several cell fates become available to the pre-malignant cells: 1) a lower-spike cell state characterized by oxidative stress, reduced cell growth linked to elevated Cdkn2a/b expression, increased lineage plasticity, and immune activation; 2) a high-Lgr6+ state that shows similar metagene expression patterns to a carcinoma; 3) a maintained upper spike cell state with increasingly heavy aneuploid burden. Chemotherapy treatment of neoplasia increases stress, causing a reversion towards the lower-spike high plasticity cell state and away from the upper-spike cell state; Pp2a inhibition causes the opposite patterns with movement away from the lower-spike cell state and towards the upper spike. This latter treatment renders upper-spike-like cells vulnerable to chemotherapeutic toxicity.

This does not imply that cells cannot travel in the opposite direction across this single Lgr6→progeny step in the developmental hierarchy during carcinogenesis. In fact, it is increasingly clear that cancer stemness is not bound by the same rigid unidirectional hierarchy that exists in most normal tissue since cancer stemness responds to complex microenvironmental induction and repression (75). Indeed, we demonstrate that a range of human tumor types, as well as malignant mouse skin primary tumors, respond to chemotherapy in vivo by returning to this conserved high-plasticity lower-spike state that is a major feature of untreated premalignant papillomas. Conversely, we induced the alternative upper spike cell fate by treatment with a Pp2a inhibitor. This induction of the upper-spike cell state then potentiated the effects of the anti-mitotic chemotherapeutic paclitaxel. Thus, these cell fates, which emerge by analysing metagenes from hundreds of independent tumors, emerge sporadically as general features of cancer development in this system, and provide a framework by which alternative cell fates can be manipulated to improve responses to treatment (2). Together, these constitute unexpectedly complementary and orthogonal findings that integrate disparate observations in the literature regarding cancer stem cells and drug resistance and provide a method that can be used to interrogate expression of dysregulated cancer gene networks, rather than just single genes, at the single-cell level.

Materials and methods summary

Mouse husbandry

Animals were housed under standard conditions, fed ad libitum, and treated in accordance with the rules and protocols stipulated by the UCSF Laboratory Animal Resource Center. All mouse experiments were pre-approved by the UCSF Institutional Animal Care and Use Committee (IACUC).

Chemical carcinogenesis experiments

For chemical carcinogenesis experiments, 8- to 12-week-old male and female mice were randomly assigned to control or treatment groups. Initiation was carried out using a single dose of the mutagen dimethylbenzanthracene (DMBA, Sigma, D3254, 25 μg per mouse in acetone) applied topically to shaved back skin. After 7 days, the mice were administered biweekly topical treatments of 12-O-tetra-decanoylphorbol-13-acetate (TPA, Sigma, P8139, 200 μl of a 10⁻⁴ M solution in acetone) for 20 weeks. The animals were then monitored for papilloma number and progression to carcinoma. Skin tumors were surgically resected in accordance with the IACUC protocol, and mice were humanely euthanized if they reached a health endpoint stipulated by IACUC (tumor size greater than 2 cm) or at the termination of the experiment, whichever was soonest. Dorsal skin, visible papillomas, and carcinomas were collected for analysis immediately post-mortem.

Bulk RNA expression analysis of genetically heterogeneous interspecific backcross mice

Male SPRET/Ei and female FVB/N mice were obtained from the Jackson Laboratory, and female F1 hybrids were crossed with male FVB/N mice to produce a 720-member heterogenous F1 backcross cohort for chemical carcinogenesis experiments. For bulk expression analysis of carcinomas, the entire lesion was excised, surrounding normal skin was trimmed away, and the sample was snap frozen and stored at −80°C. Normal skin from independent mice under the same experimental conditions was also sampled. To extract mRNA, frozen tissue was ground by chilled mortar and pestle, and suspended in TRIzol. TRIzol RNA extraction was then performed, followed by purification by Qiagen kit. mRNA was assayed on Affymetrix M430 2.0 chips according to manufacturer protocols. BisqueRNA v1.0.5 was used to deconvolute cell type in the bulk samples based on scRNAseq samples (described below). Our scRNAseq with cell type metadata was converted to a single-cell expression dataset, and was then used as the reference for R/BisqueRNA/ReferenceBasedDecomposition with no overlap.

Bulk RNA-sequencing of Pp2a inhibitor-treated carcinoma cell line

Two squamous cell carcinoma cell lines (CCK168 and 85) were treated with PP2A inhibitor LB100 at a concentration of 5 μM (or water as a control) for 24 h, and after that the cells were trypsinized, pelleted, and frozen. The RNA was extracted using ISOLATE II RNA mini kit (Bioline) according to the manufacturer’s recommended protocol. Libraries for RNA-Seq were prepared with KAPA Stranded mRNA-Seq Kit (Cat.KK8420), KAPA UDI Primers mix, and KAPA Universal UMI Adapters. The workflow consisted of mRNA enrichment and fragmentation, first strand cDNA synthesis using random priming followed by second strand synthesis converting cDNA:RNA hybrid to double-stranded cDNA (dscDNA), and incorporated dUTP into the second cDNA strand. cDNA generation was followed by end repair to generate blunt ends, A-tailing, adaptor ligation, and PCR amplification. Different adaptors were used for multiplexing samples in one lane. Sequencing was performed on Illumina NovaSeq 6000 for PE 2×50 run. Data quality check was done on Illumina SAV. Demultiplexing was performed with Illumina Bcl2fastq v2.19.1.403 software. Library preparation and sequencing were done at UCLA Technology Center for Genomics & Bioinformatics Core facilities. STAR aligner v2.7.11 was used to map reads to the mouse reference genome (mm39) with default settings. Sequencing QC reports were generated with QualiMap v2.3 with both bampqc and ranseq qc. Quantification of mapped reads was performed with Salmon v1.10.2 with gcBias and seqBias enabled. Read annotations was done with GenCode’s M31 basic annotation. We then used DESeq2 to perform differential expression analysis between Pp2a-inhibitor treated samples and vehicle-treated samples.

Tamoxifen-induced in vivo fluorescent Lgr6 stem cell-derived lineage tracing

To permanently fluorescently label Lgr6+ skin stem cells and all daughter progeny for lineage tracing experiments in skin and tumors, two doses of 4-OH tamoxifen (Sigma, T5648, 25 mg/ml in 100% ethanol; 5 mg per dose) were administered to shaved mouse back skin of Lgr6-EGFP-CreERT2 mice carrying the Rosa26-LSL-Tomato reporter (Lgr6eGFPtdTomato) (17), once in the morning and once at the end of the day. When skin tumors were present, the same tamoxifen dosage was topically applied directly over the lesions and onto the surrounding tumor-adjacent skin. Lineage-traced skin and tumor samples were collected 10 days after the initial tamoxifen treatment.

To preserve Lgr6GFP and tdTomato fluorescence within lineage-traced tissue, prior to embedding, a small piece of the excised skin / tumor used for analysis was fixed in 4% paraformaldehyde (Electron Microscopy Sciences, 15710) for 2 hours at 4°C and then placed in 30% sucrose (EMD Millipore, 573113, w/v in PBS) overnight at 4°C before freezing in OCT (Tissue-Tek, 4583) and storage at −80°C. Analysis of skin Lgr6eGFP+ stem cells (green fluorescence) and their tdTomato+ (red fluorescence) progeny was verified using cryosection analysis. For cryosectioning, fixed OCT embedded samples were sectioned at 10 μm thickness onto SuperFrost slides (Fisher, 12-550-15) prior to anti-eGFP and anti-tdTomato antibody staining and imaging with a Zeiss Axiovision microscope.

Immunofluorescent staining

Slides with 10 μm cryosections were brought to room temperature and washed in 1x PBS (Invitrogen, AM9625) containing 0.1% Triton X-100 (v/v in 1x PBS) for 30 min. Sections were then blocked using 10% normal goat serum (Jackson ImmunoResearch, 005000121) in 0.1% triton X-100 / PBS for 30 min at room temperature and then incubated with a primary anti-GFP antibody (Abcam, ab13970) for 3 hours at room temperature. Slides were then washed 3x in PBS containing 0.1% triton X-100, and then incubated with a fluorescent secondary antibody (Invitrogen, A11039) for 60 min at room temperature in the dark. Sections were then washed 3x in PBS containing 0.1% Triton X-100, counterstained with DAPI solution (Invitrogen, D3571) and mounted using Fluoromount mounting medium (Sigma, F4680).

Preparation of single cell suspensions for flow cytometry and scRNAseq

Murine dorsal skin tissue (with subcutaneous fat removed) and tumors from Lgr6eGFPtdTomato mice were processed immediately following procurement (fig. S17). Samples were processed from 8-week-old untreated dorsal back skin (3 pooled female back skins); 8-week-old, tamoxifen-treated dorsal back skin (with 10-day lineage tracing); DMBA/TPA treated back skin adjacent to skin tumors; benign papillomas (4 pooled tumors) and carcinoma tissue (from one single carcinoma, trimmed of visible normal skin) prior to flow cytometry. The adjacent skin, benign papillomas, and carcinoma tissue were collected from a single male Lgr6eGFPtdTomato mouse (following 10 days of lineage tracing).

Single-cell suspensions of dorsal back skin were prepared by carefully removing subcutaneous fat and floating the intact tissue dermal side down in 0.25% Trypsin/EDTA (Invitrogen, 25200–056) solution with gentle rocking for 1 h at 37°C. Trypsin activity was neutralized with 10% chelexed FBS in Ca²⁺ free PBS (Hyclone, SH30028.02). The epithelial skin layer was then scraped into suspension and agitated to dissociate cells. To obtain single-cell suspensions of skin tumors, lesions were finely chopped into very small pieces with a sterile scalpel and then enzymatically digested in 0.25% Trypsin-EDTA with gentle agitation for 1 hour at 37°C. Individually, each disaggregated cell suspension sample was then passed through a sterile 40 μm filter, centrifuged at 300 g for 5 min to pellet cells, and resuspended in 10% chelexed FBS in Ca2+ free PBS prior to flow cytometry. Cell samples were collected and analyzed using a FACSAria III, and appropriate gating was used to exclude doublets. DAPI staining was used to assess cell viability, and only viable, DAPI-negative cells were selected for subsequent analysis. Within each sample type, three populations of viable cells were evaluated: unsorted global cell populations, Lgr6eGFP+ stem cells (gated on positive eGFP expression), and tdTomato+ Lgr6+ stem cell-derived progeny cells (gated on positive tdTomato expression). FACS sorted cells were collected in Ca2+ free, PBS-containing 1% chelexed Ca2+ free FBS, centrifuged at 300 g to pellet the cells before final resuspension in 35 μl 1% chelexed Ca2+ free PBS, and stored on ice prior to scRNA-sequencing library preparation which occurred within 2–3 h of sample collection.

Cisplatin chemotherapy experiment

For the single-cell analysis of carcinomas treated by cisplatin, we generated four carcinomas as described above (“Chemical carcinogenesis experiments”) in 3 independent FVB female mice. Chemotherapy treatment began when visible carcinomas reached 1 cm in diameter and consisted of intraperitoneal injections. Two mice bearing primary carcinomas were treated with 100 μL of 0.8 mg/mL cis-diammineplatinum(II) dichloride (D3371, TGI Chemicals), and one control mouse with 2 primary carcinomas was treated with 100 μL of PBS. Treatments were carried out twice, separated by 4 days, for both control and cisplatin-treated mice. Treated carcinomas were harvested 3 days after the second treatment, and scRNAseq libraries were prepared and analyzed as other scRNAseq data (detailed below, Data S12). This entire protocol was performed twice to produce 8 carcinomas in total (4 cis-platin treated carcinomas and 4 control carcinomas).

Additional independent papilloma and carcinoma scRNAseq data

We gathered 7 additional papillomas using our chemical carcinogenesis protocol as described above (“Chemical carcinogenesis experiments”), and sequenced unsorted cells from them as below (Data S12). We used the control carcinomas from the chemotherapy experiment as additional independent carcinomas.

Single-cell RNA sequencing (scRNAseq) library preparation, alignment, assignment, and quality control

scRNAseq libraries from samples were prepared using the 10x Genomics Chromium extraction and preparation protocol according to the manufacturer’s instructions and sequenced on a Novaseq S4 sequencer (Illumina). Reads generated by 10x Genomics sequencing were demultiplexed using default settings by Cell Ranger’s mkfastq function (version 2.0, 10x Genomics). Demultiplexed reads were aligned and assigned to the mm10 (GRCm38) mouse reference provided by 10x Genomics and converted to sample-specific unique molecular identifier (UMI) count matrices using the count function with default settings. Count data for each sample were independently quality-controlled. Minimally, all cells were required to express at least 200 genes, with at least 20% of reads aligning, and retained genes were required to be detected in at least 3 cells. We filtered stressed or dying cells with more than 10–17% of UMIs generated from mitochondrial genes (see Data S13). We then analyzed sequencing curves of cells ordinated by UMI detection and gene number in order to detect lower elbows indicating low-complexity libraries (empty droplets) and upper elbows indicating an artifactually rapid increase in well-wise UMI detection (multiplets). We filtered cells falling below and above these thresholds, respectively. The lower threshold ranged between 250–750 and the upper between 2500–6000 (Data S13).

Normalization, batch correction, clustering

We integrated and normalized count data using R/Seurat v.2.3. To normalize, we used the NormalizeData function to scale UMIs by dividing each cell by the total number of counts per cell, multiplying by a scaling factor of 10000, and converting to a log scale. We then passed these data to the R software package Monocle3. We then included all genes in dimensionality reduction with principal components, limiting the reduced hyperspace to 100 dimensions. Together, the first 3 PCs explained 34% of the data, so the data were subsequently projected onto a 3-D manifold using the uMAP method (76). However, upon inspection of sample arrangement in this space, we observed that FACS was strongly confounded by batch. In particular, unsorted cells were colinear with the first batch, and sorted cells with the second batch. To reduce this technical variability while preserving biological variability, we implemented a mutual nearest neighbor method to identify hyperdimensional batch-specific planes and resolve these planes into a single mutually uniform plane (77). We then used these batch-corrected data for all downstream analyses (fig. S18). To detect unsupervised clusters of cells on different scales of transcriptomic similarity, we implemented the Leiden community detection algorithm within Monocle3 to call low-resolution clusters (referred to as “partitions” by the monocle software suite) and high-resolution clusters (referred to as “clusters”) (78)

Cell type annotation

First, we implemented an unsupervised, bottom-up approach to identify genes that are highly and specifically expressed by high-resolution cell clusters. To do this, we used R/monocle3/top_markers to implement logistic regression wherein each high-resolution cluster was a predictor, and examined the top 50 most specific genes for each cluster, manually reviewing these sets for cell type. Next, we compiled a custom database of dermal, epidermal, stromal, and immune marker genes based on a literature review of publicly available mouse marker sets (Data S14). With this database, we trained a hierarchical classifier based on cross-validating 100-n random partitions of the entire data, and then applied the trained classifier to the entire dataset simultaneously (79). Second, we used this database to construct cell-type gene signatures and examined the relative expression of these signatures in the manifold. Together, these methods consensus-classified 57571 (99.591%) cells. We carried out high-resolution cell type deconvolution for 1) the carcinoma parenchyma identifying its squamous and spindle phases separated in its own manifold; 2) macrophage-Langerhans cells to test for the presence of M1-M2 polarization and transition among these cell types (fig. S2D); and among T cells to test for Cd4/8 differentiation (fig. S2E).

WGCNA gene clusters, Spearman gene clusters, metagenes

We implemented WGCNA on our normal skin and carcinoma samples separately. For each, we tested soft-thresholding powers between 1 and 20, choosing 6 for both. Minimum module size was 30 and max was 4000, with default clade reassignment and cut height parameters. For our 16 seed genes of interest, we calculated the Spearman rank correlations between these genes and all others present in the transcriptome. Since the correlations were significantly stronger in Car than NSk ($\stackrel{-}{\rho }=0.21$ and 0.12, respectively [p=0.016, 1-sided t-test]), we searched for clustering patterns among seed genes in the Car correlation matrix using R/hclust with the complete method. This resulted in seed clusters 1–3, with the exception of Sox2 which was originally grouped in cluster 2 but which we considered to be in cluster 3 due to its similar single-cell metagene pattern, as explained in the text. Next, we inferred rank correlation co-expression networks from bulk data, terming their aggregate behavior as metagenes in the single-cell data. To do this, a seed gene was chosen and Spearman correlation coefficients to all genes in the microarray were calculated. The top 100 correlated genes were taken as the positive network, and the top 100 anticorrelated genes as the negative network. Next, we identified those genes expressed in the single-cell data, log-normalized their expression values with a pseudocount of 1, and scaled and summed them. This produced a single integrated value representing the aggregate expression of the 100n network, and we considered this to be the metagene expression value.

Seed and metagene visualization and comparison

We overlaid log₂ transformed individual gene and metagene expression on individual cells on dimensions 1 and 2 of the 3D manifold. In order to simultaneously compare expression of metagenes among specific cell groups and among metagenes, we first divided the parenchyma into tissue-specific cell groups shown in Fig. 2G; next we aggregated gene expression for all genes within those cell groups using R/monocle3/aggregate_gene_expression; we then summed group-aggregated expression values across all genes within a metagene, dividing this value by the number of constituent genes; finally, we centered and scaled each row (metagene) independently. To compare seed and metagene expression across the Lgr6→progeny hierarchy, we isolated stage-specific parenchyma cells; calculated metagene expression within each cell; tested for differences in the distribution of metagene and seed gene expression values with t-tests, adjusting significance for multiple testing with Bonferroni corrections.

Gene ontology analysis

We implemented R/topGO v2.44.0 (80) to test for functional enrichment in gene sets using Fisher exact tests, correcting for multiple testing with Benjamini-Hochberg, and comparing to genome-wide annotations provided by R/org.Mm.eg.db v3.8.2 (81). We also tested for enrichment of gene sets using multiple gene annotation compendia compiled by the STRING database (82), including KEGG, Reactome, UniProt, Pfam, SMART, and InterPro, and tested for enrichment using the R package STRINGdb (83). Since resulting gene annotations were largely similar, we report FDR significance in enriched terms taken from the STRING GO categories.

Chemotherapy scRNAseq analysis

We identified differentially expressed genes (DEGs) between control and cisplatin-treated carcinomas by first isolating carcinoma parenchyma as above. We then narrowed candidate genes to those expressed in at least 10% of parenchyma cells, and among those, to the top 5000 most variable genes using Seurat’s FindVariableFeatures with the vst selection method. Next, we used non-parametric Wilcox rank sum tests to compare expression between the control and cisplatin-treated carcinomas. To test whether DEGs from the cisplatin experiment were over-represented in marker genes for the spikes, we first used Monocle3’s top_markers function to identify markers for the lower spike (unsupervised high-resolution cell cluster 33) and the upper spike (unsupervised high-resolution cell cluster 19), comparing these against all other cells present in the experiment and testing one thousand genes per group (Data S15). We then ordered spike-specific marker genes from high to lower marker scores and ordered significant DEGs from high to low fold change (p<0.05). We then scanned each list for intersection at simultaneous rank, for example asking how many overlaps exist at the top 10 for each, then top 11, etc. To establish an empiric null distribution of overlaps that would be expected by chance, we repeated this procedure 1000 times with 1000 dummy marker sets consisting of randomly chosen ‘marker’ genes.

Differential gene expression in stage-specific parenchyma across empiric Lgr6+→progeny

Parenchyma cells from all stages shown in Fig. 1D were chosen. Parenchyma cells were those defined as expressing epithelial markers constituting the bulk of unsupervised clusters (fig. S2B). For normal skin, parenchyma cell types were interfollicular epithelium (IfE), infundibulum (Inf), outer bulge (OuB), upper hair follicle (UHF), sebaceous gland (SG), basal cells (Bas), isthmus (Isth), undifferentiated follicular (unF), proliferating epithelium (PrE), and inner bulge (InB). For the papilloma and carcinoma samples, these were neoplastic epithelium (nEp), upper spike (US), and lower spike (LS). The Lgr6+ vs progeny DEG analysis was done in the same manner as the chemotherapy DEG above. Namely, we narrowed candidate genes to those expressed in at least 10% of parenchyma cells, and among those, to the top 5000 most variable genes using Seurat’s FindVariableFeatures with the vst selection method. Next, we used non-parametric Wilcox rank sum tests to compare expression between the Lgr6 and progeny cells separately in normal skin, papilloma, and carcinoma parenchyma.

Orthologous human prostate adenocarcinoma (PAD) data

We obtained human prostate adenocarcinoma data from GSE141445, which had been already log-normalized. We then performed PCA reduction on these data to 100 dimensions and subsequently decomposed into a UMAP using R/monocle3/reduce_dimension with default parameters. We then resolved the L2 luminal regenerative signature consisting of the genes Sca1, Ly6a, Tacstd2, Psca, Krt4, and Cldn10 within both this dataset and our original mouse data. We also resolved our 100-gene lower spike marker set within these data (top 100 genes in Data S15 by marker score). We plotted the integrative expression of both of these gene sets as metagenes as described above.

Effects of Pp2a inhibitor (LB100) and paclitaxel treatment on colony formation in vitro

Mouse squamous cell carcinoma CCK168 cells were grown in DMEM supplemented with 10% FBS. One day after seeding, LB100 was added at 10 μM for 24 h, then it was washed off and paclitaxel was added at a concentration of 25 nM for 72 h. For the sequential combination, we selected the drug concentrations producing lowest effect on cell growth when used alone. After the drug treatment the cells were washed, trypsinized, counted, and replated at 400 cells per well in 6-well plates in triplicates. The cells were allowed to grow for 10 days until distinct colonies formed, then they were fixed in 70% ethanol and stained with Crystal Violet 0.05% (diluted in water from 0.4% in ethanol stock). The plates were dried and scanned with Keyance VHX-7000N microscope (Keyance Corp., USA) at 4x magnification using brightfield imaging. The images were stitched to produce the composite image of each well. The total colony area for each well was quantified using ImageJ software (NIH) and plotted relative to control group treatment. The results were reproduced in at least three independent experiments.

Supplementary Material

Supplementary Text

Figs. S1 to S22

Tables S1 to S5

Data S1 to S15

References (85–99)

Data and materials availability:

All code and processed data to reproduce analyses and figures in this study are available in Zenodo(84) and raw data are available in the Gene Expression Omnibus (GSE261743, GSE261766).