Single-cell RNA profiling identifies diverse cellular responses to EWSR1/FLI1 downregulation in Ewing sarcoma cells

Abstract

Background

The EWSR1/FLI1 oncogene is the most common rearrangement leading to cell transformation in Ewing sarcoma (ES). Previous studies have indicated that expression at the cellular level is heterogeneous, and that levels of expression may oscillate, conferring different cellular characteristics. In ES the role of EWSR1/FLI1 in regulating subpopulation dynamics is currently unknown.

Methods

We used siRNA to transiently suppress EWSR1/FLI1 expression and followed population dynamics using both single cell expression profiling, CyTOF and functional assays to define characteristics of exponentially growing ES cells and of ES cells in which EWSR1/FLI1 had been downregulated. Novel transcriptional states with distinct features were assigned using random forest feature selection in combination with machine learning. Cells isolated from ES xenografts in immune-deficient mice were interrogated to determine whether characteristics of specific subpopulations of cells in vitro could be identified. Stem-like characteristics were assessed by primary and secondary spheroid formation in vitro, and invasion/motility was determined for each identified subpopulation. Autophagy was determined by expression profiling, cell sorting and immunohistochemical staining.

Results

We defined a workflow to study EWSR1/FLI1 driven transcriptional states and phenotypes. We tracked EWSR1/FLI1 dependent proliferative activity over time to discover sources of intra-tumoral diversity. Single cell RNA profiling was used to compare expression profiles in exponentially growing populations (si-Control) or in two dormant populations (D1, D2) in which EWSR1/FLI1 had been suppressed. Three distinct transcriptional states were uncovered contributing to ES intra-heterogeneity. Our predictive model identified ~1% cells in a dormant-like state and ~2–4% cells with stem-like and neural stem-like features in an exponentially proliferating ES cell line and in ES xenografts. Following EWSR1/FLI1 oncogene knockdown, cells re-entering the proliferative cycle exhibited greater stem-like properties, whereas for those cells remaining quiescent, FAM134B-dependent dormancy may provide a survival mechanism.

Conclusions

We show that time-dependent changes induced by suppression of oncogenic EWSR1/FLI1 expression induces dormancy, with different subpopulation dynamics. Cells re-entering the proliferative cycle show enhanced stem-like characteristics, whereas those remaining dormant for prolonged periods appear to survive through autophagy. Cells with these characteristics identified in exponentially growing cell populations and in tumor xenografts may confer drug resistance and could potentially contribute to metastasis.

1. Introduction

Genomic characterization of cancers has led to the realization that many malignancies once considered homogeneous entities can be sub-divided based upon mutations or translocations, or through expression profiles. Examples are breast cancer, based upon hormone receptor expression or HER2 [1], and studies conducted through The Cancer Genome Atlas (TCGA) network that have defined classic, pro-neural, neural and mesenchymal subtypes in glioblastoma [2]. Similarly, medulloblastoma, a brain tumor predominantly diagnosed in children, is now defined as encompassing at least four distinct entities based on genomic changes and expression profiles [3, 4]. While childhood cancers such as rhabdomyosarcoma have been recognized as having subtypes, embryonal and alveolar, based upon presentation, histology and the presence of either t(2;13) or t(1;13) chromosomal translocations, there has been no sub-typing for Ewing sarcoma, a tumor of bone or soft tissue most frequently diagnosed in adolescents. Ewing sarcoma (ES) is predominantly characterized by a chromosomal translocation, t(11;22)(q24;q12), that fuses the DNA binding domain of the ETS transcription factor FLI1 to the N-terminal domain of the RNA binding protein EWSR1. In most cases exons 1–7 of EWSR1 are fused to exons 6–9 of FLI1 (type 1 fusion), whereas type 2 fusions result from EWSR1 exons 1–7 joining to exons 5–9 of FLI1. The presence of type 1 tumor transcripts is associated with an improved outcome in patients with localized disease [5–7], although not in advanced or metastatic patients. Ewing sarcoma is the second most frequent bone cancer in children and young adults, with a survival rate following relapse below 20% [8–11]. Despite significant improvements in various therapeutic regimens that have been made, relapse remains a critical challenge [12–19]. Ewing sarcoma is notable for a significant rate of relapse later than 5-years after diagnosis. For individuals who survived their disease at least 5 years from the time of diagnosis, the cumulative mortality rate was 25% at 25 years following entry into the Childhood Cancer Survivor Study [20]. Late relapses in ES patients may indicate the presence of dormant cells that evade cytotoxic therapy.

Stem cell models have allowed an accurate description of the ways in which complex hierarchical systems and organs develop, as a result of variation in their basic elements, as in the case of stem cell cycling illustrated by the capacity for self-renewal [21], where hematopoietic systems first allowed the identification of cells with long-term repopulating activity [21] [22, 23]. Similarly, experiments on spleen formation revealed heterogeneity in the cellular composition of hematopoietic self-renewal and differentiation [21, 22, 24, 25]. In the cancer stem cell model, it is not clear how many sources contribute to the heterogeneity that arises as a consequence of gene expression changes in response to environmental factors [26–31]. Thus, these findings underscore the importance of incorporating concepts of regulatory mechanisms that lead to cell plasticity and intra-tumoral heterogeneity, into approaches to stem cell modeling. The advent of single-cell RNA-sequencing (RNA-seq) technologies has greatly facilitated lineage tracing and sophisticated studies of stem cell heterogeneity when applied to T-cells and stem cells, as well as breast cancer cells [32–36]and more recently ES cells [37, 38]. Single cell RNA-seq profiling can reveal gene expression variation between individual cells that play important roles in subpopulation dynamics [39–46], for example, in the late relapses of ES patients [47]. Many studies have revealed heterogeneity within cell populations [48–52], and single-cell RNA-seq provides a promising and unbiased approach to unravel key molecular programs leading to cancer heterogeneity at single cell resolution [53–58].

The molecular mechanisms driving normal or cancerous cells into dormant-like or G0 states remain largely undefined. Hematopoietic cells in the bone marrow microenvironment may enter a quiescent state or a proliferative state depending on the specific environmental niche [23]. The discovery of microenvironmental effects on the fate of stem cells [59–69] gave rise to the concept of “quiescent”- now widely considered to be “dormant-like” cells in the bone marrow [53, 54]. Studies on the interplay between the microenvironment and cancer cell proliferation have suggested additional prognostic factors such as cell cycle and metabolism-related chemo-resistant heterogeneity in cancers [55–57]. In T-cell acute lymphoblastic leukemia (T-ALL), for example, it has been found that various bone marrow niches differentially protect T-ALL cells from chemotherapy by inducing quiescence [57].

Here, we have applied single cell RNA-seq combined with CyTOF to interrogate heterogeneity in ES cell populations, with a focus on defining characteristics of cells that may represent either a stem-like population or remain in a dormant state following suppression of EWSR1/FLI1 expression. Previous studies have shown that the EWSR1/FLI1 oncogene can transcriptionally regulate many genes involved in cell cycle regulation, and those having an effect on survival and differentiation [70–72]. Single-cell RNA-seq has revealed substantial cell-to-cell heterogeneity in EWSR1/FLI1 expression and in the expression of motility gene markers [37]. As yet, however, it remains largely unknown whether EWSR1/FLI1 expression variation can regulate subpopulation dynamics. Thus, it is important to understand how variation occurs among seemingly identical cells resulting in diversified population-level responses, ultimately impacting therapy at the single-cell level.

2. Material and methods

2.1. Cell culture

Human Ewing sarcoma cells (EW-8, ES1, ES4, ES6 and ES8) were established in this laboratory. TC71 and A673 cells were obtained from the ATCC. Non-ES cell lines (Hela, U2OS and MCF-7) were a gift from Manjeet Rao (GCCRI). DBTRG-05MG glioblastoma cells were a gift from T. Nicolaides (UCSF). All cells were grown in antibiotic-free RPMI-1640 medium supplemented with 20% FBS (Sigma) and maintained at 37 °C with 5% CO₂ and 90% humidity. For single-cell RNA-seq, 5×10⁵ cells were seeded in 100-mm dishes and cultured overnight. After 24 h, cells were treated with 50 nM human On-TargetPlus si-EWSR1 (Cat # J-005119–11-0050; Dharmacon, General Electric) or 50 nM si-Control (Accel Control Pool, Non-Targeting, Cat # D-0019; Dharmacon, General Electric) in Dharmafect1 (Cat # T-2001–03; Dharmacon). The target sequence for EWSR1 was “GAGUAGCUAUGGUCAACAA” and for FLI1 was “CUUUGGAGCCGCAUCACAA[DT]”. Single cell RNA-seq data were generated with siRNA targeting EWSR1 in the EW-8 cell line, and downregulation of EWSR1/FLI1 expression was confirmed with siRNA targeting FLI1 in control (non-Ewing sarcoma) cell lines and in the EW-8 cell line.

2.2. C1 microfluidic-based single-cell RNA-seq and data analysis

Single EW-8 cells were isolated using a C1 microfluidic system. Three 96 well chips were used for each condition (Control, Dormant 1, Dormant 2) and a C1 SMART-Sq V4 protocol was used for the reaction mixes and for thermo-cycling. The Fluidigm modification of the Nextera XT kit was used to prepare libraries for 864 cells (Fluidigm 100–7168). Ethidium homodimer-1 (in LIVE/DEAD Viability/Cytotoxicity Kit, for mammalian cells, Life Technologies, PN L-3224) was used to identify dead cells before library preparation. ERCC RNA Spike-In Mix 1 (Cat#4456740; Life Technologies) was used to normalize raw sequencing data and reads aligned to UCSC hg19 were converted to transcripts per million (TPM) to quantitate gene expression levels. High quality cells were selected based on a threshold for the number of expressed genes detected in each individual cell (a minimum of 2000 genes per cell). Data were log-transformed [log (TPM+1)] for data analysis and visualization. Live single-cell libraries were sequenced using a HighSEQ 3000 apparatus (Illumina) and three technically matched culture replicates were prepared for all three-time points representing exponentially growing cells (si-Control) and two dormant populations D1 and D2 (D1 was derived 48 h after replating, and D2 at 120 h after replating). Each library was sequenced to an average depth of 2×10⁶ read pairs. Expression of a range of 7000–9000 genes was detected in each cell.

2.3. Single-cell profiling and cellular identity discovery

The assignment of novel cellular identities with discriminative markers was performed on transcriptionally perturbed cells with si-EWSR1/FLI1, and a random forest feature selection method [73] was applied to the data generated from two time points (D1 and D2). KEGG pathway enrichment analysis was performed using (http://www.genome.jp/kegg/pathway.html) and the top biological processes in the gene sets based on the KEGG database were reported. The pathways with p-values < 0.001 and fold change > 2 were used to determine significant cellular processes enriched in each sub-cluster.

2.4. Gene expression variation model parameter estimation

We assumed that gene expression follows a Gaussian mixture model with up to 2 modes: one from cells in which the gene is not expressed (log2 (TPM+1) ~ 0) and the other from cells in which the gene is activated (log2 (TPM+1) > 1). We then fitted the model by maximum likelihood, using the Expectation-Maximization (EM) algorithm, implemented by the fitgmdist function of MATLAB to estimate 3 nominal parameters, μ, σ² (for mean, variance) and α for each gene in each condition (si-Control, Dormant 1 and Dormant 2). Log2 (TPM + 1) was used for expression levels. We defined α value as the proportion of cells in which a transcript is expressed (or detected at level log2 (TPM+1) > 1).

2.5. Gene activation and induction scores

We used logistic regression to establish a relationship between the detection rate and expression level. Specifically, for each cell, we defined the following function [60]:

$$F_{c,g}=\frac{1}{1+e^{-(b_{0,c}+b_{1,c}{\mu }_g)}}$$

where $b_{0,c}$ and $b_{1,c}$ are the fitted parameters for the logistic regression for cell $c$, and ${\mu }_g$ is the expression level for gene $g$ in cell $c$.

Induction Score:

We then calculated average expression levels for all genes in the control population as ${\mu }_{control,g}$. For a given cell, we transferred expression levels into “induction value” through dividing by the average expression of the gene in the control population. To reduce technical variability, we weighted each gene by the activation function obtained from logistic regression reported before [42], or the final induction score $I(c,g)=w\left(c,g\right)·\frac{x_{c,g}}{{\mu }_{control,g}}$ for a given cell $c$ and gene $g$, where

$$w(c,g)=\left\{\begin{array}{cc}1,& if{x}_{c,g}>1\\ \left(1-F_{c,g}\right),& if{x}_{c,g}\le 1\end{array}\right.$$

and $x_{c,g}$ is the gene expression level in $\mathrm{l}\mathrm{o}\mathrm{g}2(\mathrm{T}\mathrm{P}\mathrm{M}+1)$.

Functional module induction score:

A functional module induction score $M(c,S)$ of a cell was then calculated as the weighted mean of “induction values” over a set of gene $S$ (functional gene set) within the cell, or

$$M(c,S)=\sum _{g\in S} I(c,g)/\sum _{g\in S} w(c,g)$$

In this study, 3 functional gene sets were considered, cell-cycle, stemness and dormancy (Supplemental Table S2).

2.6. Differential Induction score, heat map and functional enrichment

Student t tests were performed to identify dormant populations (si-Control, Dormant 1 and Dormant 2), or 4 class-labeled clusters (C0, C1, C2 and C3). Significant genes with differential induction scores were selected based on: (1) Benjamini-Hochberg multiple-test adjusted p-value < 0.05, (2) Induction score > 0.2 and (3), at least 2-fold-change in induction score. Genes with differential induction scores were collected to generate a heat map (highest induction score in red color, and low induction score in white). Genes and cells were sorted based on their average induction score within each group and across cells, respectively. In addition, genes were grouped to each group’s class-label if their mean induction scores were higher in the group than in other groups. Functional enrichment was performed using enrichr (https://maayanlab.cloud/Enrichr/). 3D PCA plots were produced using Mayavi 3D visualization library in Python with the PCA function. PCA1, PCA2 and PCA3 were used with all genes for class labels C0, C1, C2 and C3. To determine cell types for newly profiled EW-8 single cells, we used a program developed by Lin et al. [74]; the link to their retrieval method is http://sb.cs.cmu.edu/scnn/. Then we used K-Nearest Neighbor (KNN) to identify the closest cell type to each profiled single EW-8 cell.

2.7. Western blotting

Cells were lysed in RIPA buffer and run in 4–12% Mini-NuPage Bis-Tris precast protein gels (Thermo Fisher). The primary antibodies used were rabbit anti-EWSR1 (no.11910S; Cell Signaling), rabbit anti-FLI1 7.3 (SC-53826, LOT#0213; Santa Cruz), rabbit-anti-FLI1 (ab15289; Abcam) and rabbit anti-GAPDH (D16H11; Cell Signaling). The secondary antibody used was an anti-rabbit IgG HRP-linked antibody (7074S; Cell Signaling). Cell treatment, sample collection and Western blotting were repeated at least three times, and representative blots are shown in the figures.

2.8. Immunofluorescent imaging

Exponentially growing EW-8 cells were seeded and grown in individual chamber slides (MatTek; Thermo Fisher). The following day (day1) EWSR1/FLI1 siRNA transfections were initiated. Cells were incubated in transfection media to simulate D1 or D2 populations. Washed cells were fixed in 100% methanol and blocked with 5% goat serum. Primary antibodies used were Alexa Fluor 647 anti-human CD63 (Cat # 561983), LC3A/B (D3U4C) rabbit mAb Alexa-555 (131 Cell Signaling), anti-FAM134B (HPA012077; ATLAS) and anti-CELSR2-N-terminal (ab189045). As secondary antibody Alexa Fluor 488 (ab150077; Abcam) against FAM134B, SEC61B (sc-393633; Santa Cruz) was used. Cells were stained and mounted with ProLong Gold Antifade mounting media with DAPI (P36931; Thermo Fisher Scientific). Images were taken using a Flouview FV3000 confocal laser-scanning microscope (Olympus). Signal analysis was performed using FIJI software (GPL v2; ImageJ).

2.9. Preparation of single cell suspensions from xenografts

Fresh CDX tumor tissues were cut to 1~2mm pieces and then digested with collagenase (1 mg/ml) in DMEM for 2 h at 37°C. Digested tissues were pipetted repeatedly, and then filtered by 70~100 μm Cell Strainer to remove tissue debris. The cells in the pass-through fractions were collected in a 50 ml conical tube and washed with 20 ml HBS for 5 times and finally resuspended in 15~30 ml stem cell culture medium (DMEM/F12; Gibco;Life Technologies) to obtain single cell suspensions.

2.10. CyTOF and immunostaining

In vitro assessments of protein expression variations in response to EWSR1/FLI1 down-regulation were studied using CyTOF (Helios 2; Fluidigm Inc.) and the acquired data were analyzed using FlowJo software v10.2 (Tree Star Inc.) and Cytobank (Cytobank Inc.). EW-8 cells treated with si-EWSR1 were processed at D1 and D2 time points. Two vials of each line were obtained for the no staining control and functional control (si-Control). All cells were treated with 5 μM Cell ID Cisplatin (Fluidigm; Cat # 201064) in polypropylene round-bottom tubes, 5 ml capacity (Becton Dickinson 352235) for 5 min. The staining protocol for antigens located in the nucleus and cell surface markers was applied as per manufacturer’s protocols (Fluidigm) to 3×10 ⁶ cells in 100 μl staining volume. The primary antibodies used were human Cyclin B1–153Eu (Fluidigm; 3153009A), Cyclin D-141Pr (Cell Signaling; 2978BF), CD271–149Sm (Fluidigm; 3149017B), CD40–142Nd (Fluidigm; 3142010B), CD79a-158Gd (Cell Signaling; 13333BF), CD49b-161Dy (Fluidigm; 3161012B), CD49e-160Gd (Fluidigm; 3160015B), CD63–150Nd (Fluidigm; 3150021B), CDKN2A-155Gd (Abcam; Ab186932), CELSR2–165Ho (Abcam; Ab189045), FGFR1–159Tb (Cell Signaling; 9740BF), MKI67–147Sm (Cell Signaling; 9449BF), MYC-176Yb (Fluidigm; 3176012B), TGFBR2–170Er (Abcam; Ab 78419),TGFB1–163Dy (Fluidigm; 3163010B), SOX2–169Tm (Cell Signaling; 3579BF). A pilot study was designed to optimize the staining protocol using a panel of 15+ markers. Antibodies were titrated using EW-8 cells and controls were included when EWSR1/FLI1 was knocked down at the two dormant conditions. Cells were treated with Fc protein to block non-specific binding of Fc receptor expressing cells with antibody (Cat # 564220; BD Pharmingen) for 30 min before staining with 1.5 μl of each antibody over night at 4 °C. Cells were labeled with 125 nM Intercalation solution (Fluidigm; Cat # 201192B) for 1 h at the room temperature. Cells were washed after each step for 5 min and centrifuged at 350× g before cell fixation for 5 min and centrifuged at 1000× g after cell fixation. Pelleted cells were re-suspended in Maxpar water (Fluidigm; Cat # 201069) for data acquisition with mass cytometer Helios2 in the Flow Cytometry and Cellular Imaging Core Facility at the University of Texas MD Anderson Cancer Center. A list of antibodies and staining conditions is included in the supplemental section (Supplemental Table S1).

2.11. Cell proliferation and viability assays

Cells seeded at identical densities (5000 cells/well) were incubated in 96-well plates (Cat # 3596; Corning). Cell proliferation was assessed by measuring confluence (%) over 96 h using the IncuCyte system (Essen Bioscience Inc.) Images were taken every 4 h from 4 different areas of each well. To distinguish live from dead cells Cytox green reagent was added to each well at the end point to yield a final concentration of 250 nM. Images were taken from 3 technical replicate and assays were repeated at least 3 times.

2.12. Cell motility and chemotaxis assays

To study motility, cells were isolated from cell line-derived ES xenografts (CDX). Isolated cells were FACs sorted for CD63-High (Q4) or CD63-Low (Q1). Cells (2500/well) were seeded into the top layer wells of a ClearView Plate (Essen Bioscience Inc./IncuCyte) with serum depleted media for cell migration studies. In the reservoir 20% fetal Bovine Serum (FBS) was used as chemoattractant to study migration over 72 h for cells derived from CDX-ES1, CDX-ES4 and CDX-EW-8. Images were taken from cells on the top layer and bottom layer of the ClearView membrane (ESSEN Bioscience Inc./IncuCyte). The number of cells that migrated and adhered to the bottom-side of the membrane was quantified after applying the standard phase metric defined by Essen as “Phase Object Count (1/well). The number of objects per well normalized to (divided by) the initial object count over time is reported from each graph.

2.13. Functional stem-like studies, flow cytometry and cell sorting

Self-renewal studies were conducted on FACs sorted cells based on the CD40 signal. Cells isolated based on their CD40 expression were seeded at identical densities (50 cells/well) and incubated in 96-well plates (Ultra Low Cluster round bottom, ultra-low attachment polystyrene, Cat # 7007, Corning). Cells were grown in KnockOut Dulbecco’s Modified Eagle Medium/Nutrient Mixture F-12 or DMEM/F-12 (Gibco;Life Technologies; media without HEPES buffer or L-glutamine optimized for the growth of induced pluripotent stem cells and human embryonic cells) supplemented with 1% penicillin-streptomycin (Thermo Fisher Scientific) and with EGF, FGF growth factors and StemPro Neural supplement (Cat # A1050801;Thermo Fisher Scientific) for 7 days. At day 7, images were taken using the Celigo S imager (Nexcelom Bioscience) to assess the ability of each population to form spheroids with the Celigo sphere formation tool. For secondary spheroid formation, primary spheroids were disaggregated and replated and analyzed as for the primary spheroid formation.

2.14. Endoplasmic reticulum (ER) area measurements

ER area measurements were conducted using ImageJ as follows. Background thresholds were manually defined and set for all images. Borders of each cell were drawn, and the ER area was then calculated and presented as a fraction of the total cell area.

3. Results

3.1. Fluidic-based single Ewing sarcoma-cell transcriptomics profiles

To experimentally test whether variation in the expression of EWSR1/FLI1 by individual cells plays a role in subpopulation dynamics, we transcriptionally tracked EW-8 cells at single-cell resolution using the Fluidigm C1 single-cell system. The hypothesis was that high-level expression of EWSR1/FLI1 will be associated with proliferating cells, as it activates expression of E2F [69–72, 75], while low-level expression of EWSR1/FLI1 will be associated with cellular dormancy. However, it is not clear whether cells respond homogenously to dynamic changes in EWSR1/FLI1 oncogene activity overtime.

As doxycycline-regulated shRNA systems can be ‘leaky’, we used siRNA to down regulate EWSR1/FLI1 in several ES cell lines (Fig. 1A and Suppl. Fig. S1). In EW-8 cells, compared to the control siRNA, the siRNA directed at EWSR1 caused downregulation of both the endogenous EWSR1 protein (to 35% of control), and to a greater extent to the fusion protein (to < 10% of control) (Fig. 1A). EWSR1 was suppressed between 60–89% in non-ES cells and 33–78% in three additional ES cell lines (Suppl Fig. 1A, B). Endogenous FLI1 was not detected in any of the cell lines examined. Knockdown of the EWSR1/FLI1 oncogene varied between 54–87% in these ES cell lines (Suppl Fig. 1B). To examine the effect of EWSR1 knockdown on proliferation, cells were replated 48 h after siRNA transfection (time is denoted 0 h; Fig. 1B), and proliferation determined over a further 90–120 h. We found that the proliferation of a non-ES cell line, Rh30 (an alveolar rhabdomyosarcoma cell line expressing the Pax3-Foxo1 fusion gene) was not significantly suppressed by EWSR1 knockdown, whereas the proliferation of EW-8 cells was essentially abrogated for > 96 h. Knockdown of EWSR1 in additional non-ES cells had a small effect on proliferation in U2OS cells, but had essentially no effect on proliferation in HeLa or DTBRG (glioblastoma) cells. Of note, following knockdown, the greatest decrease in EWSR1 levels was found in DTBRG, but EWSR1 knockdown in this line showed the least effect on proliferation. In contrast, proliferation was essentially abrogated in A673 and ES-8 ES cells (Supplemental. Figs 1C, D). Similar results were obtained using GFP-labeled EW-8 cells, where si-EWSR1 inhibited EW-8 cell proliferation by > 85% (Suppl. Fig. 1E). As the effect of EWSR1 knockdown in 3 of 4 non-ES cell lines had negligible effects on proliferation, and in all ES cell lines essentially suppressed further proliferation completely, these data support that the predominant effect of the knockdown was due to suppression of the fusion oncogene, and not the partial suppression of endogenous EWSR1.

Figure 1. Transient EWSR1/FLI1 downregulation directs ES cells into a dormant-like state.

A, Western blot analysis showing that siEWSR1/FLI1 causes less downregulation of endogenous EWSR1 protein compared to the EWSR1/FLI1 fusion oncoprotein at D1; B, Quantitative curves of the relative confluence of Rh30 and EW-8 cells transfected with si-Control or si-EWSR1. After 48 h cells were replated, after which proliferation (confluence) was determined over 100 h. C, Distribution plots of genes, EWSR1 and FLI1 revealing that the majority of the cells exhibit a lower EWSR1 and FLI1 expression at D1 following si-EWSR1/FLI1 transfection, whereas the number of cells with higher EWSR1 and FLI1 transcripts increase in the D2 population; D, Cell cycle distribution following 48 h si-EWSR1/FLI1 or si-Control knockdown; E, (Violin PLOTS) Single-cell expression dynamics of known EWSR1/FLI1 targets, (NR0B1, CAV1, VIM, PRKCB, LIPI) and FLI1. Expression of the genes is suppressed in D1, whereas in D2 there is a population of cells that recover expression. F, Scheme describing three functionally distinct populations for single-cell RNA-Seq profiling from time point 0 to 110 h; the siControl population at 48 h was collected to study exponentially dividing cells (black line); the si-EWSR1/FLI1 populations were profiled during two time points, Dormant 1 and Dormant 2 (blue line).

To examine the heterogeneity in expression of EWSR1 and FLI1 at single cell resolution, EW-8 cells were transfected with si-Control or si-EWSR1, and then re-plated after 48 h. Cells were harvested after an additional 48 h or 120 h to obtain the Dormant1 (D1) and Dormant 2 (D2) populations, respectively. Our experimental model revealed heterogeneity in the expression of EWSR1 and FLI1 in EW-8 cells (si-Control population), and that both were suppressed at D1, following siEWSR1 transfection, with some recovery at D2 (Fig.1 C). To test the effect of knockdown on cell cycle progression, ES1, ES2 and EW-8 cells were transfected with control or EWSR1 siRNAs. After 48 h cell cycle analysis revealed that the predominant effect of EWSR1/FLI1 knockdown was an increase in the fraction of cells in the G1 and a decrease in the S phase, whereas a significant sub-G1 fraction was not observed (Fig. 1D).

Repeated experiments with EW-8 cells showed that after knocking down EWSR1/FLI1 there was a slight increase in cell confluence after a further 120 h incubation after re-plating the cells (designated D2), suggesting that a fraction of the cells were re-entering the proliferative cycle. This notion is supported by an increased expression of EWSR1/FLI1 and five EWSR1/FLI1-dependent genes, (NR0B1, CAV1, VIM, PRKCB, LIPI, FLI1) that were maximally suppressed at D1, but of which the expression had increased in the D2 population (Fig. 1E). This result suggests that EWSR1/FLI1 plays a direct role in preventing the ‘dormant-like’ state. Based on these observations, we defined a model in which there are three functionally distinct cell populations, exponentially growing, Dormant 1 (D1) and Dormant 2 (D2) cells (Fig. 1F).

3.2. EWSR1/FLI1 expression variation and subpopulation dynamics

To identify dependent and independent transcriptional programs that lead to different cell fates of EW-8 cells in response to heterogeneous EWSR1/FLI1 activity, distinctive proliferating conditions (si-Control, Dormant 1 and Dormant 2) were used to profile cells at single-cell resolution. We hypothesized that in the first dormant population (D1) cells become transiently proliferatively quiescent and in the second dormant population (D2) a fraction of the cells recover from the dormant state. In this analysis the expression of on average of 8970 genes was detected in each of 685 viable single cells. Reproducibility was observed between biological replicates (Biological: aggregate R = 0.97416667 ± 0.01) (Fig. S2A–D), with correlations plateauing once ~30 cells had been sampled. In the Dormant 2 (D2; 120 h sample), 100 sampled cells reached a similar level of correlation, suggesting greater heterogeneity in this population (Fig. S2E), supporting the suggestion of distinct populations with differing characteristics at this time point. The high correlation between EWSR1 or FLI1 with proliferative markers (PCNA, MCM2, KI67) and expression of NGFR (CD271) suggests that EWSR1/FLI1 expression drives ES cell proliferation (Fig. S4F).

To select functional features for each proliferative state (Control, D1, D2) an activation score was computed for three functional gene panels: cell-cycle regulatory genes, dormancy associated genes and stem cell associated gene features (Fig. 2A). The expression of cell cycle genes was most suppressed in D1, whereas the expression of dormancy associated genes and genes associated with ‘stemness’ was elevated in D2 relative to control EW-8 cells (Table S2). This suggests that at D2 there are two distinct populations of cells, one with stem-like properties and one with dormant properties. To identify distinguishing transcriptional programs regulated by EWSR1/FLI1 in EW-8 cells a three-step computational pipeline was developed (Fig. S3A, B). In doing so, four novel transcriptional states (Clusters C0-C3) were identified (Fig. 2B, Fig. S3C). A pathway analysis that distinguishes each population and subpopulation is presented in Fig. S4A, B. In unperturbed cells (si-Control) and in D1, two predominant populations were identified: Cluster C1 (“Cell-cycle committed cluster”) with high expression of proliferative markers and low expression of stem-like and survival genes, and Cluster C3 (“Transitional Dormant state”) with undetected expression of proliferative genes (Table 1). In D2, the predominant populations were C0 cells (“Stem-Like Cluster “) that had distinct patterns of expression for genes associated with stemness, and a higher expression of EWSR1/FLI1. This suggests that re-entry into the proliferative state was associated with a different expression profile being more ‘stem-like’, compared to proliferating cells in the control culture. The other major population was Cluster C2 (“Dormant-like Cluster”) with high expression of dormancy associated genes. Distinguishing gene features separating C2 from C3 are presented in Supplemental Table S3.

Figure 2. Single cell profiling in combination with machine learning reveas a role of EWSR1/FLI1 in subpopulation dynamics.

A, (Violin Plots) Functional analysis of gene pathways associated with cell cycle, dormancy and stemness (signatures were obtained from GO) interrogated in control (siControl) or Dormant 1 and Dormant 2 populations. Genes associated with cell cycle were significantly decreased in Dormant 1 cells, whereas genes associated with ‘stemness’ were increased in Dormant 2 cells. The same gene sets were interrogated in newly discovered subclusters; Cluster C3 shows reduced cell cycle-associated gene expression, whereas cluster C2 exhibits an enhanced gene expression profile for dormancy, and cluster C0 exhibits elevated levels of ‘stemness’ gene expression at all three time points. Statistical significance is shown with p values for functional induction between populations and subpopulations (see the material methods section for more details). B, (PCA PLOTS) PC1, PC2 and PC3 distinguish subpopulations regulated by EWSR1/FLI1 expression. In the unperturbed population (si-Control, n = 231) 4 cells in C0, 106 cells in C1, 1 cell in C2 and 120 cells in C3; In the knockdown population (D1, n = 196) 68 cells in C1, 2 cells in C2 and 126 cells in C3; In the recovery population (D2, n = 257) 134 cells in C0, 52 cells in C1, 62 cells in C2 and 9 cells in C3. C, (PCA PLOTS) PC1, PC2 and PC3 distinguish subpopulations with low EWSR1/FLI1 expression from subpopulations with high EWSR1/FLI1 expression from all EW-8 conditions and their phenotypes. High activity of Senescence, Autophagy and Cell cycle genes is shown in dark purple and low activity in light grey.

Table1.

Top distinguishing gene features for newly discovered cellular states in EWS.

Subpopulation ID	Active gene name
C0	PCP4,PCSK2,GDF6,DAB1,ETV1,CDH23,IRF1,GCH1,DCC,CXCR4, GFRA2,KLF2,S100B,KCNA2,CXCR4,FZD5,TOX3,NR0B1,CD86,SOAT2,AHR,SYN1,ARC,PKP2,FGFR4,HSD11B1,ALK,SLC29A4,CXCL10,HSPB8,NPY,PNOC,COL9A3,SFRP2,CORT,LAMB3,ALDH4A1, MBD2,SLC1A3,RAX,NTN4,BCL11B,IGF1,EGR2,SLC18A3,LOXL2, IRF1,SCARB1,ELK1,ALDH4A1,PAX7,SNCA,NEUROD6,ENG,NEURL,TOX2,KCNAB3,NXPH4,GABRA2,CELSR2,PYCARD,CD40,SOX2,MYC,CCND1,ALDH2A,KIT,KAT,FGFR1,HOMER2,ELK1,SNCA,RPGR,SAG,CDH23,PCP4,EPHB2.PCSK2,PDE1B,KCNA2,KCND1,ARC,PNOC,DTNA,SLC1A3,CORO1A,BCL11B,GNA12,FBF1,CASK,S100P,ITPKA,DCC
C1	E2F7,OIP5,MCM2,MCM5,RCC1,HMMR,MCM10,POLA2,TRIP13,TTK,SYCE2,CDC45,FANCA,FANCD2,TYMS,HAUS8,ATRIP,RECQL4,MKI67,ASPM,TACC3,KLF11,MYBBP1A,ATF5,POLE2,DTL,DCLRE1B,STIL,XRCC2,SPAG5,CENPI,FBXO43,PLD6,CLSPN,MAP2K6,CDCA5,PLK4,MYBL2,INCENP,ERCC6L,DLGAP5,NUF2,NCAPG2,GINS1,BLM,CDCA2,NEK2,NCAPG,KIF15,BUB1B,KIF14,FAM83D,CDT1,WDR62,KIF11,KIFC1,KIF2C,RAD51,TCF19,CAMK2B,E2F1,TEX15,NCAPH,CDCA8,CIT,E2F8,RFWD3,CCNA2,CEP55,KIF20A,PKMYT1,CCNF,ORC1,LIG1,SASS6,SKA3,CDC6,CDC25A,CDC25C,SKA1,AURKB,NGFR,E2F1,CDC25C,EME1,KIF20A,HIST1H2BH,CPNE7,FAM111B,E2F8,GADD45G,IL32,NRG1,SNORD44,SNORA68,CCK,NOTCH1
C2	VASN,TGFBR1,TGFBR2,TGFR3,LTBP1,PDGFA,COL1A1,COL3A1,COL4A1,COL5A1,IGF2R,IGFBP5,IGFBP7,FLT1,SRPX2,IL11RA,A2M,CYP1B1,CD63,SEC61,RTN4,GABARAP,GABARAPL2,BCAS3,AZGP1,ACP5,EMILIN2,GBP2,GBP3,ITGA4,JAM2,NEDD9,PTN,IL17RD,NEO1,SRPX2,DST,NEGR1,PCDH11X,PLCB1,PLAU,SRPX,GPNMB,CXCL14,IFI27,MERTK,PCDH18,VCAM1,AMIGO2,COL1A1,COL3A1,AMIGO,MEIS3,CXXC5,CNTN4,COL5A1,FNDC3B,NID2,COL6A3,NRCAM,PCDH7,APOD,PELI1,S100A1,BCL2L11,PCDHGA10,TNC,NT5E,CAMK4,L1CAM,PCDHB14,FNDC3B,NID,PI15,PCDHB12,PCDHB11,PELI1,CCDC80,APOD,PCDH20,TGFB1I1,TGFB2,TGFBI,CAMK4,RND3,SNED1,COL18A1,CD1D,CDH24,CSF1,IGFBP7,LPP,SERPINE1,SNAI2,CD24,MACF1,PCDH7,PHLDB2,CDK6,MKX,COL6A3,JAM2,NID2,PLAU,GPNMB,MERTK,ACP5,NRCAM,
C3	DFNA5,SNX10,CXXC4,DKFZP586I1420,ANKRD30A,DDAH1,DCX,FBP2,SNORA12,LOC440905,ANK3,KLRC2,RHBDL2,CAPN6,PRL,RASSF9

Novel candidate markers from each subpopulation were extracted to uncover driving features leading to heterogeneity in EW-8 cells from all single cells (si-Control, D1 and D2) as depicted in Fig. S4B. Single cell expression dynamics over time of 6 genes known to be direct targets of EWSR1/FLI1 (CAV1, LIP, NR0B1, PRKCB, VIM, FLI1) showed that EWSR1/FLI1 expression is induced in D2 after knockdown (Fig. 1E). Principal component analysis (PCA) was performed on the expression values of all detected genes from all cells at each of the three timepoints. We found that PCA of gene expression profiles from the two time points following EWSR1/FLI1 downregulation distinguish low from high EWSR1/FLI1 expressing cells (Fig. 2C, S3C).

3.3. Identification of ‘stem-like’ cells in exponentially proliferating ES cells

Discriminative gene features that define clusters C0-C3 were identified. Stem-like characteristics (high expression of SOX2, CD40, CELSR2), or characteristics suggesting dormancy (low expression of SOX2, ALDH1, ALDH2, NEUROD6, CELSR2) are shown in Fig. 3A. The frequency of detection of each cluster (C0-C3) over time after EWSR1/FLI1 knockdown is shown in Fig. 3B. Based on genes that identify ‘stem-like’ properties (CELSR2, SOX2 and CD40), stem-like cells were found to be increased from 1% in an unperturbed EW-8 population (si-Control) to 58% in the D2 population. Expression of EWSR1 and FLI1 was highest in C0 (stem-like) and C1 (cell cycle committed) cells, (Fig. 3C). This suggests that EWSR1/FLI1 fluctuation led to two distinctive transcriptional states (C0 and C2) after stress induced via EWSR1/FLI1 suppression and proliferative arrest in EW-8 cells. Of note, expression of CD40 correlated with a higher expression of stem-like gene features (Fig.3D; p < 0.0001).

Figure 3. Characterization and validation of distinct subpopulations.

A, Discriminative gene-features that define EW-8 subclusters for stem-like state (C0, Blue; high expression of SOX2, CD40, CELSR2); Cells committed to cell cycle (C1, Red; Low expression of CD40 and CELSR2); Dormant-like (C2, Green; low expression of SOX2, ALDH1, ALDH2, NEUROD6, CELSR2); Transitional dormant state (C3, Orange; low expression CELSR2, NEUROD6); B, Clusters C0 (52%) and C2 (32%) become predominant in D2, but are also detected with a very low frequency (C0~2%, C2~0.5%) in the control population (si-Control,); C, (VIOLIN PLOTS) Sub-population level study of EWSR1 and FLI1 expression; D, (VIOLIN PLOTS) Sub-population analysis based on CD40 expression shows induced stem-cell gene features in CD40-positive cells. (**** p < 0.0001); E, (VISNE PLOTS) Visualization plots of exponentially proliferating EW-8 cells highlight co-expression patterns of stem-like cell surface markers (CELSR2, CD40) with FGFR1, SOX2, CCND1 and MYC. Cells considered to have high expression are boxed; F, Single-cell analysis of Ewing sarcoma xenografts. Cells were isolated from two Ewing sarcoma xenografts (ES-1, EW-8), and stained for CD40, CELSR2, FGFR1, SOX2, CCND1, MYC and analyzed by CyTOF. VISNE plots show a low percentage of isolated cells that were considered to express all three proteins at high levels (boxed) in ES xenograft tissues; G, IHC staining for CELSR2 in tissue sections of ES xenografts. FFPE sections from ES4 and EW-8 xenografts were stained for CELSR2 expression. The percentage of cells scoring positive was deduced from scoring 3-high powered fields (40x).

Among the most highly associated markers, CD40 and CELSR2 appeared to show the most attractive features for a stem-like condition. To identify stem-like cells in exponentially proliferating cultures of si-Control EW-8 cells, the single cell mRNA data were searched for cells with the same profile as described for the stem-like Cluster C0. We classified 232 cells using new class-derived labels. The Neural Network (NN) that obtained 95% predictive precision and 96% success in recall, was used to predict the transcriptional states in si-Control EW-8 cell cultures (Area Under Curve (AUC): 0.998, Classification Accuracy (CA):0.949, F1 Score: 0.954). Using the expression signature identifying stem-like cell surface markers (CELSR2, CD40) with SOX2, MYC, FGFR1, and CCND1 (Fig.3A), four cells from 232 cells (1.73%) with a high expression in the siControl population demonstrated the stem-like expression signature (C0) (see Fig. 2B).

We next sought to determine the frequency of cells co-expressing CD40 and CELSR2 protein in two cell line-derived ES xenograft (CDX) models in scid mice. Using CyTOF, a fraction of cells (~4–5%) in CDX-ES1 and CDX-EW-8 xenografts with positive expression for CD40 and CELSR2 was identified, and these cells also showed higher levels for SOX2 (Fig. 3F). Cellular heterogeneity for CELSR2 in ES xenograft tumor sections was also examined by IHC detecting CELSR2 only in a small fraction of cells (Fig. 3G). Temporal changes in the expression of a fraction of cells identified by high stemness markers (CD40, CELSR2 and SOX2) at a single cell level with CyTOF revealed that CD40 and SOX2 downregulation in the Dormant 1 population was consistent with the transcription profiles and protein expression distribution showing heterogeneity in the expression of all three markers over time using μ (mean expression) and α (Fraction of cells positively expressing the marker) (Fig. S5).

3.4. EW-8 cells with a high CD40 expression exhibit increased self-renewal properties

We aimed to assess the stem-like capabilities of cells with a high expression of CD40. From our initial FACS analysis, we estimated that ~5% of EW-8 cells were CD40 positive, whereas > 90% were CD40 negative. To assess self-renewal capacity, cells were FACS sorted for high or low expression of CD40 and plated to form spheroids in neurosphere medium. Formation of spheroids was assessed after 7–10 days. To determine ‘self-renewal’ capacity, a stem-like characteristic, spheroids were disaggregated, after which the cells re-seeded under the same conditions to assess secondary spheroid formation (Fig. 4A). We found that CD40-positive cells formed spheroids more readily than CD40-negative cells, and further maintained the ability to generate secondary spheroids, indicating self-renewal capabilities (Fig. 4B). Interestingly, the association between CD40 expression with survival using Kaplan-Meier survival analysis on two independent cohorts clinically annotated as ES tumors [51] showed that a poor outcome appears to relate to the expression of CD40, but not SOX2 (Fig. 4C).

Figure 4. Characterization and validation of distinct subpopulations with stem cell properties.

A, 7-day self-renewal assay; EW-8 cells were sorted based on their CD40 expression profile after which 50 cells were plated in each of 96 wells in neurosphere medium without serum. Tumor-sphere formation was quantified and measured from a total of 2400 cells expressing high or low CD40 (total = 4800 cells); Upper panels: primary sphere formation at 7 days from CD40-negative subpopulation (left) and CD40-positive subpopulation (right). Bottom panels: secondary sphere formation following disaggregation of primary spheres (left) CD-40-negative, (right) CD40-positive cells (p = 0.001, unpaired t test); B, Quantification of primary and secondary spheroid colonies formed by CD40-positive and CD40-negative EW-8 cells. (mean ± SD); C, Kaplan Meier curves showing the relationship between CD40 or SOX2 expression in biopsies of Ewing sarcoma and patient outcome; cohort data: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE171205 D, In D2, after EWSR1-FLI1 knockdown cells in cluster C2 are characterized by increased expression of autophagic genes, TGFβ receptor and cyclin dependent kinase inhibitory genes; E, The relationship between CD63 and dormancy associated gene markers (TGFBR2 and TGFBR3), proliferation markers (MKI67, CDKN2A) and EWSR1 and FLI1 in each subpopulation (Pearson correlation). (C0: Stem-like (blue), C1: Committed to the cell-cycle (Red), C2; Dormant-like (Green), C3; Transitional dormant state (Orange).

3.5. Dormant-like state C2 shows characteristics of autophagy

Single cell expression profiling of D2 cells indicated two major populations: C0 having stem-like properties, and C2 having dormant-like characteristics (Fig. 2A). Single-cell gene expression data showed a distinct expression pattern for marker CD63 in cluster C2 (μ_C0 = 9.66 ± 0.5; μ_C1 = 9.18 ± 0.61; μ_C2 = 10.88 ± 0.68; μ_C3 = 9.73 ± 0.82, p = 0.0001, ANOVA) (Fig. 4D). Using Spearman correlation, the relationship between cell surface marker CD63 expression, a member of the tetraspanin family of membrane proteins, and dormancy associated gene features (TGFBR2, TGFBR3, CDKN2A) and EWSR1 and KI67 expression were explored (Fig. 4E). CD63 was negatively correlated with EWSR1 and KI67 expression (marking proliferation), but positively associated with TGFBR2, TGFBR3 and CDKN2A expression.

A population-level protein study of EW-8 cells revealed bimodal expression values for dormancy associated phenotypes (CDKN2A, TGFBR2) (Fig. 5A) following EWSR1/FLI1 downregulation. The variation in the expression levels of dormancy associated gene features in the dormant states (D1 and D2) was validated using CyTOF with previously described parameters. This single-cell protein study confirmed that cells in the proliferative quiescent populations (D1, D2) show induced levels of dormancy associated gene features, TGFBR2 (α = 2.7%_siControl; α = 34.91%_D1, α = 12.92%_D2) and CDKN2A (α = 2.07_siControl; α = 37.35% _D1, α = 11% _D2) (Fig. 5B). T-SNE visualization shows an increase in the fraction of cells with dormant features (D1, ~36%), and recovery from a dormant state (D2, ~12%) when cells are grouped based on their co-expression patterns for both dormancy-associated markers in each population (Fig. 5C). Using CyTOF, we next explored co-expression of TGFBR2 and CDKN2A with CD63 in cells isolated from two ES-cell line derived xenografts (CDX-ES1, CDX-EW-8). Approximately 1% of the cells isolated by FACs were detected with positive expression for all three features (Fig. 5D). Previously, cells with a low EWSR1/FLI1 expression have been associated with an increased motility [37]. As CD63 was negatively correlated with EWSR1/FLI1 expression, we explored migration of ES cells using a chemotaxis assay. Cells were isolated from ES xenografts (ES4, ES1, EW-8) and sorted by FACS based on the number of CD63 surface molecules per cell. We found that migration was higher in cells sorted for high-level expression of CD63 (t-test, p < 0.0001, n = 3) (Fig. 5E). Consistent with a previous report [37], we observed a higher mean level of motility genes (FN1, SERPINE2, TIMP1) and epithelial to mesenchymal regulatory genes (ITGA4, CALD1, COL3A1, COL5A2, GSN, MYL6, MYL12A, SPARC, TWIST1, VCL) in cluster C2 (t-test, p < 0.001–0.000001) (Fig. S6), suggesting that C2 cells transitioned to a more mesenchymal and motile state following EWSR1/FLI1 downregulation.

Figure 5. Identification and characterization of cells with dormancy characteristics in cell cultures and xenografts.

A, Distribution of cells with dormant features (TGFBR2⁺, CDKN2A⁺) in si-Control, D1 and D2 populations. B, Determination of the fraction of cells (α) with detectable protein expression values for dormancy gene markers (TGFR2⁺, CDKN2A⁺ and KI67⁻) by CyTOF showing a shift in the frequency of cells positive for dormancy markers in D1 and D2 populations. C, Enrichment of a fraction of EW-8 cells with positive dormancy gene features (CDKN2A, TGFBR2) determined by CyTOF in the D1 and D2 populations. Approximately 3×10⁶ EW-8 cells were stained for dormancy markers and 3×10⁵ cells were analyzed by CyTOF; D, Left, Illustration of the gating strategy for detecting cells with high levels of the membrane marker CD63 that are also expressing dormant features (high TGFRB2 and high CDKN2A). CD63-positive, TGFRB2^High, CDKN2A^High represents < 1% of the population of exponentially growing EW-8 cells. Center, and right panels: visualization of the rare fraction of cells isolated from dissociated CDX-ES1 and CDX-EW-8 xenografts with dormant characteristics with ViSNE plots. E, Cells dissociated from ES4, ES1 and EW-8 xenografts were FACS sorted for high or low CD63 expression, after which their ability to migrate was determined using a chemotaxis assay. Quantitative counts of the ratio (Bottom/Top) of cells migrated from the top side of the ClearView membrane insert to the bottom side over 72 h (p-values are from unpaired t test, two-tailed).

Under the dormant condition D2 induced by EWSR1/FLI1 downregulation, a network of genes regulating selective autophagy (FAM134B, MAPL3CB, GABARAP, GABARAPL2,) and ER response to stress (SEC61B, RTN4, CLIMP-63, CD63) was identified as encompassing genes differentiating the C2 cluster (Fig. 6A). Variation in the component genes regulating autophagy and ER protein expression reflects a discriminative feature for dormant-like C2. This cluster is defined by a failure to re-enter the cell cycle at D2, but with a common expression activity for survival genes (S100BP, S100PP, S100A10, S100A11, S100A13, S100A6, COX6A1, COX6C, COX7A2, COX8A) with cluster C0, the stem-like sub-population (Fig. S7A). Consistent with the dormant model induced via EWSR1/FLI1 downregulation, single cells sorted based on their relative transcript number for CD63 revealed statistically significant differentially expressed genes involved in cell cycle, senescence and drug resistance pathways (Fig. S7B).

Figure 6. Downregulation of EWSR1/FLI1 leads to increased autophagy and expanded ER in EW-8 cells.

A, Subpopulation studies reveal down-regulation of FAM134B and up-regulation of LC-3-like modifiers (MAPLC3B, GABARAPL2) and ER stress response proteins in cluster C2 (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001, one-way ANOVA, error bars indicate S.D.); B, Fluorescence images of EW-8 cells transfected with si-Control or si-EWSR1/FLI1. EW-8 cells were fixed in D1, immunostained and subsequently imaged via fluorescent microscopy to study the induction of the autophagy marker MAPLC3B and the lysosomal marker CD63; C, Quantification of FAM134B co-localization with LCB3, presented as Pearson’s correlation coefficient (r); **** p < 0.0001, t-test, n = 15 fields (~150 cells) in the dormant cells shown in A using the colocalization tool in imageJ; D, EW-8 cells were treated with si-EWSR1/FLI1 or si-Control 48 h before being stained for the ER resident protein SEC61B. The expansion of ER is shown following EWSR1/FLI1 downregulation (boxed areas). E, Quantification of cells with expanded ER after masking nuclei. **** p < 0.0001, t-test, error bars indicate S.D., n = 150 cells.

Co-localization of FAM134B with MAPL3CB was strongly increased in the induced cluster C2 at the D1 timepoint following EWSR1/FLI1 downregulation (Fig. 6B, C). Furthermore, dormant cells also exhibited increased expression of CD63, which is also a lysosome marker, compared to si-Control. These observations support the idea that cells with lower EWSR1/FLI1 levels which remained in a dormant state for an extended period of time showed higher levels of gene features involved in autophagy, as CD63 has been shown to associate with FAM134B and MAP1L3CB in autophagic cells [76]. The role of EWSR1/FLI1 in ER activity was investigated by monitoring the levels of the ER-resident protein SEC61B. Increased expression of SEC61B was observed in dormant cells (C2) with low levels of EWSR1/FLI1. The presence of an expanded ER, especially in the cell periphery, was further confirmed in dormant cells with lower FAM134B levels determined by IHC (Fig. 6D, E). These data show the link between a dormant model induced via EWSR1/FLI1 suppression and autophagy. Therefore, it appears that ER-autophagy is a cellular characteristic contributing to cell survival in a prolonged dormant-like state (C2) in EW-8 cells.

3.6. Identification of EWSR1/FLI1 heterogeneity through cell type assignment

Ranked distinguishing gene features can help to identify transcriptional programs which translate to different cell functions, as indicated above. We assessed the expression of top cancer gene drivers across the 4 novel clusters. Cluster C0 (“Stem-Like Cluster”) was compared to cluster C2 (“Dormant-like Cluster”) in D2 cells. We found that APC, DAXX, TP53, PIK3CA and FBXW7 were upregulated 1–2-fold (p = 0.0066, 0.0003, 0.02, 0.009 and 0.004, respectively), and there was a 4-fold increase in the expression of TCF7L2, ZFP36L2 (p < 000001, p < 0.000001) and SOX2 (p < 000001) in cluster C0 compared to C2 (Fig. 7A).

Figure 7. Analysis of EWSR1/FLI1 heterogeneity through the expression of known cancer gene drivers and cell type assignment.

A, Fold change difference in common cancer gene drivers between clusters (C0-C3). Each violin plot represents median and density distribution at single-cell level expression of top known cancer gene drivers, the color represents newly discovered cell cluster ID (Blue: C0, Red:C1, Green:C2, Orange:C3); B, Relationship between EWSR1 and FLI1 expression and 17 common cancer gene drivers with Pearson correlation (r); C, Overview of the frequency of each cell type in 4 discovered clusters C0-C3. Individual cluster estimate of the percentage of each cell type across all single EW-8 cells revealed gene expression signatures similar to Embryonic stem cells (ESC), Progenitors, and Cerebral Cortex cell types.

Comparison of the “Stem-Like Cluster” (C0) with Cluster C1 (“Cell-cycle committed Cluster”) showed that PTEN was downregulated (4-fold; p = 0.0001) and that TP53 and DAXX were downregulated at least 2-fold (p = 0.002, p < 0.000001). Comparing Cluster C2 (“Dormant-like Cluster”) to Cluster C1 (“Cell-cycle committed Cluster”) showed that FBXW7, B2M, TCF7L2 and ZFP36L2 were upregulated at least 2-fold (p = 0.001, p < 0.000001) in C2 cells. Expression of TP53, SOX2, DAXX and EWSR1 were highly positively correlated with expression of FLI1, whereas a negative correlation was found for APC, B2M and TCF7L2 (Fig. 7B).

To test whether there might be common regulatory networks between known cell type specific states and novel transcriptional states (C0-C3), we used published single-cell RNA-seq data (33 data sets) and compared the similarities of EW-8 single cell expression profiles with various previously profiled cell types (Table S4). We analyzed more than 17,000 single cells across 30 different cell types using K-Nearest Neighbor (KNN) to assign new cell IDs corresponding to known cell types. Progenitors and embryonic stem cells (ESC) were considered as less differentiated cell types compared to epithelial and cerebral cortex cells as the most differentiated cell types. If the proposed model is correct, we should identify a more diversified cell population dynamic in the recovery time and a predictive regulatory model defining cell fate in response to EWSR1/FLI1 expression. Of note, all profiled EW-8 cells regardless of their state (i.e., control, D1, D2) were divided into three major cell types (embryonic stem cells (ESC), progenitor cells and differentiated cells) (Fig. 7C, Fig. S8, Table 2). This approach showed that more differentiated cells were enriched in C3 (Transitional Dormant state) cells that exhibit lower levels of EWSR1/FLI1 expression. Consistent with the expression profile of EW-8 cells in D2, the fractions of progenitor cells and ESC cells were highly enriched in the stem-like C0 cluster

Table 2.

Cell Type Assignments Summary

Cell Type	Total %	% Control	% D1	%D2
Brain (Neuron, CNS, Cerebral Cortex	11.97	4.5	3.2	4.23
Epithelial	0.3	0	0.14	0.14
Embryonic Stem Cell (ESC)	72	24.4	21.78	26.16
Hematopoietic Stem Cells
1(T1-LT)	0.43	0.43	0.14	0
ZYGOT	0.73	0.14	0.14	0.43
Spleen and Lymph Nodes	4.67	1.60	1.46	1.60
Endothelial	0.3	0	0.14	0.14
Thymus	0.58	0.14	0	0.43
2HSPC	1.6	0.29	0.58	0.73
Progenitor	4	1.16	0.43	2.48
Cancer Lymphocytic	1.30	0.4	0.14	0.73

¹T1-HSC: CD31⁺VE-cadherin⁺CD45⁻CD41^lowc-Kit⁺ in E11 aorta–gonad– mesonephros Hematopoietic Stem Cells; LT-HSCs: Lin^–, c-Kit⁺, Sca1^+, CD34^–, Flk2^– Hematopoietic Stem Cells; T1-HSC: CD31⁺VE-cadherin⁺CD45⁻CD41^lowc- Kit⁺ in E11 aorta–gonad–mesonephros Hematopoietic Stem Cells

²HSPC: Hematopoietic Stem Cells and Progenitor Cells

4. Discussion

Genomic analyses of tumor types considered as homogeneous have been critical in identifying subtypes characterized by specific DNA alterations or expression profiles. For example, medulloblastoma, once considered a ‘homogeneous’ disease, is now considered to entail at least four molecular entities, and glioblastomas can be classified into three subgroups based on expression profiles. Single cell profiling offers the next level of resolution, identifying characteristics that have functional consequences. Single cell profiling has identified OLIG2 expressing glial progenitors as transit-amplifying at medulloblastoma initiation, but as quiescent cells later in tumor progression, yet enriched in recurrent tumors following treatment [77]. The value of single cell profiling in other malignancies and normal tissues has amply been shown, i.e., single cell profiling has revealed diversity among stem cell populations in hepatocellular carcinoma that have significantly different self-renewal abilities [78], and a role for KRAS in exit of cells from an epithelial state to a mesenchymal phenotype when exposed to TGFβ [79]. In contrast to brain and other tumors, genomic studies have suggested Ewing sarcoma to be homogeneous. However, recent studies using single cell expression profiling have demonstrated that intracellular levels of EWSR1/FLI1 are heterogeneous, and that functional phenotypes can be driven by the different levels of EWSR1/FLI1 expression [36, 37].

Our interest was to examine cellular fates when EWSR1/FLI1 was suppressed, but under conditions where the oncogene was re-expressed over a period of time. We found that the levels of EWSR1 were similar in both ES cells and non-ES cells. The siRNA used in this study targeted EWSR1, and reduced EWSR1 protein by 40–90% in both ES and non-ES cells. However, the effect of knockdown had modest or no effect on proliferation in non-ES cell lines. In contrast, si-EWSR1 significantly reduced the EWSR1/FLI1 protein level in all ES cell lines tested, with the greatest effect in EW-8 cells. EW-8 was used as a model cell line to assess the effects of oncogene modulation. Although si-EWSR1/FLI1 knockdown had a relatively modest effect on the proliferation in non-ES cells, proliferation was essentially completely suppressed in all ES cells tested. Thus, the effect of knockdown on cell proliferation appears to be predominantly a consequence of EWSR1/FLI1 suppression and not of suppression of endogenous EWSR1.

Following EWSR1/FLI1 knockdown, proliferation of EW-8 cells was arrested over a period of 4–5 days, with subsequent re-expression of EWSR1/FLI1 and that of oncogene-regulated genes in a subpopulation of cells. These studies revealed expression changes associated with specific subgroups of cells with different functional characteristics, at different time points. Downregulation of EWSR1/FLI1 led to proliferative quiescence, designated Dormant 1 [D1] and Dormant 2 [D2], respectively. Individual cells from three functionally distinct groups (Control, D1, D2) revealed heterogeneity in cellular responses to EWSR1/FLI1 downregulation and time. The computational pipeline developed revealed that changes in the average expression of genes over time can dissect discriminative behaviors contributing to diverse responses. Four distinct groups were identified based on expression profiles, and the proportions of each subpopulation changed with time. In the D2 population, cells with C0 profiles (stem-like cluster; 53%) or C2 profiles (dormant-like cluster; 24%) became predominant. Cluster C0, was characterized by upregulation of stem cell genes (SOX2, MYC, CCND1, ALDH2), whereas these genes were suppressed in the C2 subpopulation, compared to cells from control populations. Survival genes and autophagy regulatory genes were progressively induced from clusters C3 to C2, possibly indicating an important regulatory role in maintaining this dormant state. In addition to a higher expression of stem-like genes, the C0 subpopulation was characterized by expression of inflammatory genes and also showed induction of neural stem cell features defined by the expression of cadherin, EGF, LAG, seven-pass G-type receptor 2 (CELSR2) and neuron projection genes. Using expression profiles that discriminated C0-C3, we next interrogated exponentially proliferating EW-8 cell populations and ES xenograft tissues to determine whether cells with similar gene profiles could be identified. Compared to control, exponentially proliferating EW-8 cell populations harbored ~2% of cells with the C0 expression profile.

We coupled single cell expression profiling with determination of single cell protein content by CyTOF and functional assays. Rare cells positive for CELSR2 were detected by CyTOF in control EW-8 cell populations, and 4–5% of the cells disaggregated from ES xenografts were classified as having high levels of CD40, CELSR2 and SOX2 expression. Functionally, cells sorted for high expression of CD40 had a greater capacity to form primary and secondary spheroids in culture, characteristics of ‘stemlike’ cells, suggesting that cells in D2 with stem-like features (C0) play a role in the recovery of proliferation following EWSR1/FLI1 knockdown. Of interest, poor outcome of patients with Ewing sarcoma was associated with high expression of CD40, but not SOX2. In mice, CELSR2 expression is only required for the development of the forebrain at embryonic day 13.5 (E13.5) to E14.5 [80]. Additional studies on mammalian cells, report a role of CELSR2 with PRICKEL and Frizzled (FZD) proteins in the regulation of planer cell polarity (PCP) in epithelial sheets. Moreover, the C0 state showed a significantly higher level of PRICKEL3 compared to the other 3 identified cellular states (C1, C2, C3).

The second major cluster at D2 was C2, cells that showed a ‘Dormant-like’ profile. C2 cells displayed expression patterns for genes regulating autophagy and ER, characterized by GBA, CD63, GABARAPL2, SEC61B, RTN4 and CLIMP-63 expression. Consistent with a recent single cell analysis [37], C2 and C3 cells showed lower EWSR1/FLI1 levels and increased levels of actin-binding genes, cytoskeleton assembly genes and actin-based motility genes (MYL6, MYL12a, ACTN4, CFL1, GSN, VCL, PKP1, ITGA1, ITGA4). C2 (dormant-like cluster) cells exhibited an increased expression of genes associated with EMT and had a greater migration capability. Of interest, cells with the C2 profile were extremely rare in the control EW-8 cell population, as only one cell with this expression profile was identified in a heterogeneous (si-Control) population (~0.5%). One model, consistent with the changes in expression is that EWSR1/FLI1 fluctuation induces a stress response, and that a subpopulation of cells can transition to a C0 (stem-like) state, whereas C2 cells without the ability to suppress autophagy remain in a proliferatively quiescent state. These cells would be anticipated to be resistant to proliferation-dependent killing by cytotoxic drugs.

Although cells with C0 or C2 profiles predominate in the D2 population, the exact cues that determine cell fate remain to be identified. The ability of C0 cells to influence other cells via paracrine signaling through secretion of VEGFA, GDF6, BMPER, TNF, IGF1, CXCR4 or CXCL10, or to be modulated through receiving signals via receptors (FGFR4, TNFSF13B) could be an efficient survival strategy. This idea is consistent with observations made on the relationship between cell proliferation and cell density. It indicates variability in the secretion and sensing of signals by single cells and how they are tightly regulated by inflammatory cytokines. Thus, individual cells probably display a variable control over the expression of cytokines to determine cell fate under stress induced following EWSR1/FLI1 downregulation, although the experimental conditions used here do not address this. Consistent with a recent study [81], we further highlighted heterogeneity in clinically important genes with metastatic potential like YAP1/TAZ (Supplemental Fig. S9).

There are several areas in which our single-cell RNA-seq profiling and biological design may be improved. Advances in single cell sequencing technology now allow far greater numbers of cells to be profiled, although at lower resolution [82], that would allow a more accurate definition of rare populations following oncogene fluctuation. While our study used model systems (cell lines and xenografts), profiling of primary patient samples will be essential to confirm our results. Ewing sarcoma is a rare tumor representing ~2% of childhood cancers, hence patient-derived xenografts were used as a first approximation. Although downregulation of EWSR1 in non-ES cells had little effect on proliferation, we cannot exclude the possibility that downregulation of endogenous EWSR1 influenced cell fate outcome.

4. Conclusions

The important findings are that in the D2 phase, two major populations were identified, cells re-entering proliferation (C0), and a C2 population that remained quiescent. Of note, cells with expression profiles for C0 and C2 were rare in exponentially growing EW-8 cell populations. Specifically, the C0 population re-entering the proliferative cycle had greater stem-like properties than proliferating cells in control populations. In contrast the quiescent C2 population showed characteristics of autophagy, but increased motility. Since seemingly identical EW-8 cells responded differently to EWSR1/FLI1 fluctuation, it suggests that suppression of EWSR1/FLI1 expression confers new characteristics to ES cells that may increase tumorigenicity and confer both drug resistance or metastatic properties.

List of abbreviations

CA

Classification Accuracy

EM

Expectation-Maximization

EMT

Epithelial to Mesenchymal Transition

ER

Endoplasmic reticulum

ES

Ewing sarcoma

ESC

Embryonic Stem Call

ETS

E26 transformation-specific

F1

Feature 1

FACS

Flow Activated Cell Sort

FBS

Fetal Bovine Serum

FZD

Frizzled

GCCRI

Greehey Children’s Cancer Research Institute

KNN

K-Nearest Neighbor

PCA

Principle Component Analysis

PCP

Planer Cell Polarity

T-ALL

T-cell acute lymphoblastic leukemia

TPM

Transcript Per Million

UCSF

University of California San Francisco

Availability of data and material

The single-cell RNA-Seq datasets supporting the conclusions of this article are available in the [NCBI GEO] repository. https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE171205

Token: gdqtquomthubnkf

The images presenting single cells before mRNA extraction and CyTOF raw data from the current study are available from the author on reasonable request.

The cohort reference supporting the conclusions of this article are available in the the NCBI repository https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE63156