Cell lines are one of few available models used to interrogate EwS biology given the rarity of the disease and the lack of appropriately faithful animal models³⁵. One drawback of cell line use is potential loss of intra-tumoral heterogeneity compared to the original tumor due to serial passaging. We reasoned that profiling an array of EwS cell lines would capture the diversity of gene regulatory networks that may underlie the clinical and transcriptional heterogeneity of this disease. Because of the well-described direct influence of the EwS fusion on chromatin remodeling ³⁴, we utilized an assay capable of pairing single-cell RNA transcript counts with a chromatin assay for transposase-accessible chromatin (ATAC) (herein referred to as ‘multiomic’ sequencing) on 9 EwS and 5 non-EwS cell lines (Fig.1A, Fig.S1A). We recovered 30,105 profiles after filtering for poor quality cells (Fig.1A, Fig.S1A, Fig.S1B). Multiomic single-cell profiles self-assembled after dimensionality reduction according to cell line identity (Fig.1A).

First, to identify the extent of transcriptional heterogeneity of validated direct EWS::FLI1 target genes¹⁹ (EwS TGs) in individual cells, we calculated expression of EwS TGs at the single- cell level, subtracted by the aggregated expression of control gene sets³⁶. As expected, RNA expression of these EwS TGs was higher in EwS compared to non-EwS cell lines (Fig. S1C). As these published EwS TGs were experimentally derived from the A673 cell line¹⁹, it was unsurprising that A673 exhibited the highest average expression of these genes. To interrogate EwS TGs using chromatin-accessibility data, we generated gene activity scores³⁷ which model accessibility of regulatory elements in the vicinity of a given gene. These scores confirmed that chromatin accessibility of EwS TGs was higher in EwS cell lines compared to non-EwS cells (Fig. S1D). To compare EWS::FLI1 motif enrichment across cells, we calculated single-cell ChromVAR deviation scores³⁷ of GGAA-microsatellites (four or more consecutive GGAA repeats; abbrev. ‘GGAA- µSat’) in accessible regions. As anticipated, this analysis revealed low enrichment in non-EwS cell lines. Unexpectedly, stark differences in EWS::FLI1 motif enrichment were observed across EwS cell lines (Fig. 1B). In some cell lines accessibility of GGAA- µSats was similar across single cells (e.g. TC71, RDES) while others showed a broader range of accessibility (e.g. SKNMC, TC32, PDX305). CHLA10 cells showed a clear bimodal distribution of EWS::FLI1 motif accessibility (Fig. 1B). This hints that within these cell lines, heterogeneity exists among individual cells with respect to accessibility of GGAA- µSats. Next, we collated a list of publicly available EWS::FLI1 fusion binding sites³⁸ with or with a GGAA µSat present. We then assessed chromatin accessibility at these published EWS::FLI1 binding sites across all cell lines and discovered that accessibility at these loci varied in a cell-line-specific manner (Fig. 1C & Fig. S1D).

Given the heterogeneity in the magnitude of EWS::FLI1 motif enrichment across EwS cell lines, we hypothesized that EwS cell lines differ in the proportion of cells expressing high levels of the EWS::FLI1 fusion transcript. It has been reported that lower levels of EWS::FLI1 gene expression and activity induce mesenchymal phenotypes in EwS cell lines³⁹. To address this, we designed an approach to enrich EWS::FLI1 fusion transcript from the RNA components of our multiomic capture using biotinylated hybridization probes that tiled the EWSR1 and FLI1 gene loci. We then performed long-read sequencing on the product and counted fusion reads using novel fusion detection software designed specifically for long-reads (see Methods for details). As expected, EWS::FLI1 fusion transcripts were detected in EWS cells (Fig.1D). Scant non-ES cells showed low levels of the fusion, likely from PCR artifacts of low-complexity products. Overall, however, we observed that EwS cell lines exhibit a high degree of variability in detectible EWS::FLI1 fusion transcript (Fig. 1D & Fig. S1E), possibly reflecting a difference in expression levels among individual cells. Generally, the proportion of cells with detectible transcript correlated with values obtained using single-cell qPCR to assess fusion expression¹⁵. These findings are generally congruent with motif enrichment in that high fusion-expressing cell lines (e.g. A4573, SKNMC, and RDES) exhibit high EWS::FLI1 motif enrichment while cell lines bearing lower levels of fusion transcript and FLI1 product (e.g. PDX305, TC32, CHLA9) exhibit lower EWS::FLI1 chromatin signature (Fig.1B & Fig. 1D-E). By contrast, other cell lines displayed inverse trends of fusion transcript and product to chromatin enrichment (e.g. A673 and TC71).

To profile the global chromatin and transcriptional heterogeneity using an unbiased approach, we calculated top differentially expressed (DE) genes and differentially accessible (DA) regions across EwS and non EwS cells. As with the EwS TG-focused analysis, this approach showed that top DE/DA features readily distinguished EwS cells from non-EwS cells (Fig. 1F-G). We observed self-aggregation of EwS cells into distinct clusters. Specifically, in terms of top DA regions, CHLA9, CHLA10, TC32, and PDX305 cells were more similar to each other than other EwS cell lines (Fig. 1G). Altogether, these data show that EwS cell lines exhibit heterogeneous gene-regulatory profiles that segregate into groups that display distinct trends in accessibility, potentially reflecting unique underlying gene regulatory networks that are active in discrete cell contexts.

Next, to evaluate how cis-regulatory gene networks differ across EwS cells, we correlated transcript expression to regions of accessibility using the joint measurements of transcript count and accessibility in each cell (so-called “peak-to-gene linkages”)³⁷. Hierarchical clustering of these cis-regulatory networks with a k parameter equal to 3 clearly separated CHLA9, TC32, and PDX305 into one distinct network module and SKNMC, A4573, RDES, and TC71 into another (Fig. 2A), recapitulating what we observed using top DE/DA features (Fig. 1F-G). We selected k = 3 to extend the two-cluster structure observed in Fig. 1F–G, reasoning that A673 represented a clear outlier whose distinct regulatory program would be obscured at lower k. Cell subsets of CHLA10 separated into either cluster 2 or 3 (Fig. 2A) but more cells had cumulative expression of module 2 genes (Fig. 2B). We then assigned the 9 cell lines into three groups based on their dominant module signature (Fig. 2A-B), resulting in assignment of A673 cells as group 1, CHLA9, CHLA10, TC32, and PDX305 cells as group 2, and SKNMC, A4573, TC71, and RDES as group 3 (Fig. 2A). Enrichment analysis using Gene Ontology and Hallmark terms confirmed that the genes associated with these modules have functional associations with distinct cell behaviors (Fig. 2C). While there was some degree of overlap, module 2 showed a relative enrichment in GO terms related to mesenchymal identity, and modules 1 and 3 had enriched neuronal identity markers. Genes involved in oxidative phosphorylation and targets of mTOR were uniquely enriched in module 1, possibly due to the presence of a BRAF^V2600E mutation in A673 cells given the reported effects of this mutation on the mTOR pathway^40–42.

Next, we investigated if gene modules might harbor distinct TF binding sites. We performed differential motif enrichment using HOMER⁴³ and found a strikingly well-delineated, and module-specific enrichment of specific TF families across the cis-regulatory networks (Fig. 2D). The AP-1 family of motifs (FOS, FOSL2, JUN, AP1), which has been suggested to form ternary complexes with EWS::FLI1⁴⁴, enriched in modules 1 and 2, but not in module 3 (Fig. 2D). Zinc Finger (ZF) TFs (KLF4, KLF5, MAZ, SP1) were most enriched in modules 1 and 3 with limited evidence for activity in regulation of module 2 (Fig. 2D). Module 2 peaks uniquely exhibited motif enrichment for developmental regulators from the Homeobox family of TFs (HOXD13, HOXA9) and downstream transcriptional regulators of the Hippo and Wnt signal transduction pathways (TEAD1, TEAD4). To corroborate these findings, we performed an analysis which correlates, at the single-cell level, TF binding motifs in accessible regions with gene expression (Fig. 2E). This showed that TEAD1, HOXD13, and AP-1 TFs were highly enriched in group 2 cells (Fig. 2E-F). Group 3 lacked motif enrichment for AP-1 TFs but showed robust enrichment in ALX1/4, RUNX3, CEBPB and EN2 TFs (Fig. 2E-F). We assessed protein levels of selected TF regulators by Western blots to evaluate their heterogeneity across the cell lines. The protein abundance of these TFs exhibited substantial variability among the cells tested (Fig. 2G). Together, these multimodal analyses of single cells uncovered at least two dominant gene regulatory circuits. One of these (module 2) displays mesenchymal/EMT-like functionality that is likely to be mediated by high activity of TEAD, AP-1, and HOXD13 TFs. The other (module 3) has enriched neural program functionality most likely mediated by ZF TFs and neural and neural crest development TFs such as RUNX3 and HAND2.

We next sought to address whether the identified gene regulatory modules are directly or indirectly modulated by EWS::FLI1. We first measured EWS::FLI1, EWS::ERG (the second most common fusion in EwS tumors), and wild-type FLI1 motif enrichment and found all to be enriched in all three modules (Fig. 3A) suggesting that EWS::FLI1 is likely to regulate expression of linked genes across all cells irrespective of the dominant module. However, by annotating these sites using published EWS::FLI1 binding sites (39) and segregating into those with or without a GGAA microsatellite (GGAA- µSat), we observed enrichment of GGAA- µSats only in module 3 peaks (Fig. 3B). Thus, this analysis suggests that EWS::FLI1 binding is likely to directly influence gene regulatory circuits and its impact on transcriptional state depends on GGAA patterning, as has previously been suggested⁷.

To begin to address the impact of EWS::FLI1 activity on cis-regulatory modules, we sourced data from published lists of EWS::FLI1 “activated” and “repressed” genes and performed gene set enrichment³³ (Fig. 3C). Overall, EWS::FLI activated genes were enriched in all modules, whereas repressed genes were enriched only in module 2 (Fig. 3C). When combined with the lack of GGAA-uSat enrichment in module 2 (Fig. 3B), these observations are consistent with previous reports that repressed loci usually do not have long stretches of GGAA repeats^7,45. Altogether, these data reveal distinct and polarized fusion-mediated regulatory events across EwS cell lines that differentially depend on GGAA repeats, suggesting that the diversity of fusion function is dependent on the motif landscape at fusion binding sites.

To directly address how different gene modules are affected by EWS::FLI1, we performed single-cell multiomic sequencing on selected cell lines from each EwS subgroup after shRNA- mediated knockdown of the fusion (Fig. 3D). Following quality control (Fig. S2A), we recovered 11,614 cells expressing negative control shRNA (A673 shNS, CHLA10 shNS, PDX305 shNS, TC71 shNS) and 12,466 cells expressing shRNA specific to the FLI1 transcript (A673 shFLI1, CHLA10 shFLI1, PDX305 shFLI1, TC71 shFLI1). All shFLI1 cells showed decreased expression of EwS TGs compared to shNS cells (Fig. 3E), providing evidence that the knockdown effectively induced loss of EWS::FLI1 activity. EWS::FLI1 knockdown had specific, predictable, and reproducible effects on all cell lines. Importantly the same stratification of shNS cell lines into groups with specific module enrichment was observed, indicating that lentiviral transduction and selection did not alter dominant gene programs (Fig. 3F-H, Fig. S2B-D). In addition, each module was affected by EWS::FLI1 knockdown in a manner consistent with predictions based on module gene sets. Specifically, EWS::FLI1 knockdown resulted in decreased expression of module 1 and 3 genes (Fig. 3F,H, Fig. S2B,D) and increased expression of module 2 genes (Fig. 3G, Fig. S2C). These data confirm that EWS::FLI1 activates cis-regulatory networks found in modules 1 and 3 but represses module 2. The results also support our initial hypothesis that EWS::FLI1-regulated gene modules differ among commonly used EwS cell lines. They further show that specific and reproducible patterns of transcriptomic and phenotypic signatures dominate each cell line and these predictably respond to fusion knockdown.

Most EwS cell lines in this study are dominated by a single gene-regulatory program (modules 1, 2, or 3). A notable exception is CHLA10, which contains mutually exclusive sub-programs defined by differences in both EWS::FLI1 motif activity and module utilization (Fig. 1B, Fig. 2A). This distinction is particularly striking because CHLA10 and CHLA9 originated from tumors from the same patient, with CHLA10 being derived post-therapy and following disease relapse. To further investigate this intratumoral heterogeneity, we performed unsupervised Louvain clustering⁴⁶ on EwS cell lines (Fig. S3A) and computed the Pearson correlation coefficient of the most highly variable regions of accessibility. As expected, CHLA9, TC32, and PDX305 cells grouped together based on similarity (Fig. S3B). In contrast, CHLA10 segregated into two clusters (C22 & C23) that each showed greater transcriptional similarity to other cell lines than to each another (Fig 4A, Fig. S3B), highlighting the extent of its internal heterogeneity. Unsurprisingly, of all cell lines, CHLA10 C23 cells exhibited the greatest correlation to profiles from CHLA9.

Cell-cycle effects contributed to but could not fully explain CHLA10 heterogeneity (Fig. S3C-D). Instead, plotting peak accessibility and transcript abundance of each module revealed that C22 cells enriched for module 3 while C23 cells relied on module 2 (Fig. 4B). Motif enrichment of differentially accessible peaks further underscored the nature of the split: RUNX motifs were enriched in C22 cells whereas AP-1, TEAD, and HOX TF motifs predominated in C23 cells (Fig. 4C). Gene-set enrichment mirrored these findings, C22 expressed neural programs whereas C23 cells were enriched for EMT, migration, and TGF-β signaling gene sets (Fig. 4D-E). Pathway- activity inference analysis between cell clusters⁴⁷ further implicated TGF-β as a dominant upstream driver in C23 (Fig. 4F).

Collectively, the isogenic model CHLA9/CHLA10 encapsulates the spectrum of transcriptional states seen across the other EwS cell lines and pinpoints TGF-β signaling as key mediator of the module 2 transcriptional state.

Prior reports have shown that TGF-β promotes pro-angiogenic extracellular-matrix (ECM) gene expression in EwS⁴⁸. To interrogate the impact of TGF-β signaling on the newly defined transcriptional states in EwS, we performed multiomic profiling on A673 (module 1-dominant), CHLA10 (module 2 dominant), and TC71 (module 3 dominant) after treatment with recombinant TGF-β1 for 24 hours and compared them to vehicle-treated controls. After quality control filtering 19,611 cells were retained for analysis (Fig. S4A). Dimensionality reduction revealed substantial shifts in both gene expression and chromatin accessibility in A673 and CHLA10 following TGF- β treatment, whereas TC71 profiles remained relatively stable (Fig. 5A-C). In A673 and CHLA10, TGF-β treatment resulted in a clear polarization of cells towards module 2 usage (Fig. 5A-C), while expression patterns of modules 1 and 3 remained consistent with prior observations (Fig. S4B). The muted response of the TC71 cell line (module-3 dominant) to the influence of TGF-β (Fig. 5A-C, Fig. S4B) suggests that pre-existing transcriptional states may condition the sensitivity to TGF-β signaling.

We next examined TF expression and motif accessibility in clusters exhibiting the most pronounced or lowest enrichment of TGF-β-treated cells (indicated by arrows in Fig. 5A-F). Correlation analyses revealed a reciprocal enrichment pattern of many key TFs from Figure 2, where correlation of FOSL2 (but not RUNX3) gene expression and accessibility was generally highest in TGF-β responsive clusters. Western blot validation confirmed TGF-β-induced FOSL2 expression specifically in CHLA10 but not in the relatively insensitive TC71 cells, despite observed SMAD2 phosphorylation (Fig. 5G).

Our recent whole human transcriptome digital spatial profiling (DSP) study of CHLA10 tumor xenografts (Nanostring GeoMx) showed that transcriptionally distinct tumor cell subpopulations are spatially distributed²¹. We therefore interrogated these data to determine if the identified gene regulatory modules in CHLA10 cells were randomly or spatially distributed across tumors in vivo. We first selected genes from modules 1 & 3 that were significantly upregulated (average log₂ fold change > 0.25) by EWS::FLI1 as well as module 2 genes that were significantly repressed by the fusion (average log₂ fold change < -0.25) and applied these fusion-regulated gene sets to the DSP data. Intriguingly, regions of interest (ROIs) that were captured from tumor borders (ROIs 2, 6, 8 & 10) were enriched with tumor cells that displayed heightened module-2 expression, whereas densely packed tumor cores (ROIs 1, 3, 4, 5, 7, 11) favored modules 1 and 3 (Fig. 5H, Fig. S4C). Module 3 was also highly favored in subcutaneous compared to tail vein-derived disseminated tumors (Fig. 5H). Cells captured specifically from the border of a metastatic retroperitoneal tumor (ROIs 9 & 10) expressed the highest levels of FOSL2, TGF-β-response genes, and ECM-organization programs (Fig. 5I).

In summary, these findings highlight TGF-β-driven activation of module 2 gene expression in EwS cells via AP-1 transcription factors, particularly FOSL2. Furthermore, spatial analysis suggests module 2 programs are enriched at invasive tumor regions, underscoring the role of TGF- β signaling in mediating EwS cell heterogeneity and tumor progression dynamics.

We next sought to determine whether the gene-regulatory networks defined in our cell line panel are similarly deployed across the landscape of early passage patient-derived-xenograft (PDX) and primary patient biopsies. To determine how the three regulatory gene programs manifest in these tumors, we used the fusion-regulated gene sets as defined above. In a series of publicly available PDX tumor profiles¹⁹, module 1 was prominent, while modules 2 and 3 (Fig. 6A) showed more modest and heterogeneous enrichment. On the other hand, in six primary EwS tumor samples⁴⁹ (Fig. 6B, Fig. S4D), profiled by single-nucleus RNA-seq, every case contained discrete subpopulations that preferentially expressed either module 2 or module 3. In contrast to the PDX data, module 1-positive cells were comparatively rare in primary tumors.

We next asked whether these single-cell patterns could be detected in bulk transcriptome data that were generated from three independent patient datasets^9,50,51. To achieve this we employed a deconvolution method⁵² that identifies the overlap of signals from bulk measurements with a single-cell transcriptomic reference and captures any proportion of the bulk signal not represented by single-cell data in an ‘intercept’ term (schematized in Fig. 6C). We first validated the method on bulk RNA sequencing data taken from many of the same EwS cell lines we profiled in our experiments and found a high level of concordance with little signal captured in the intercept term (Fig. S4E). We then tested the transcriptomic overlap of EwS subgroups in 85 non-clinically annotated patient tumors that were profiled by bulk RNA sequencing in two independent cohorts (PRJNA253759, PRJNA261990). Group 2 signals were evident in nearly all tumors and group 3 cell signals were also present in more than half of tumors, confirming heterogeneity of gene network usage and cell state in primary tumors (Fig. 6D). Tumors with abundant group 2 signal also displayed significantly higher expression of TGF-β response genes and FOSL2 expression (Fig. 6E), linking our in-vitro observation that TGF-β drives module 2 activation to patient tumors in the clinical setting.

Finally, we examined microarray data from a series of 46 clinically annotated primary tumors (GSE63157) that were uniformly treated and for whom outcome data were available⁵¹. First, we observed significant correlation between expression of FOSL2, TGF-β response and module 2 expression (Fig. 6F). Furthermore, high expression of fusion-repressed module 2, and particularly of its key effector FOSL2, identified patients with markedly worse event-free survival, whereas modules 1 and 3 had little prognostic value (Fig. 6G-H, Fig. S4F). Module 2, linked to TGF-β responsiveness and FOSL2 activity, emerges as the predominant—and clinically adverse— program in primary EwS, underscoring its potential as both a biomarker of aggressive disease and a therapeutic target.

Our study examined male and female animals, and similar findings are reported for both sexes.

The EwS cell lines A673, SKNMC, CHLA10, CHLA9, TC32, A4573, RDES, and TC71 were obtained from ATCC and COG (https://www.childrensoncologygroup.org/) cell lines repositories. The osteosarcoma cell line U2OS and rhabdomyosarcoma cell line RD were also obtained from ATCC. 5064L and 240L MSCs were obtained from Dr. Darwin Prockop. H7 MSCs were generously gifted from Dr. Alejandro Sweet-Cordero. PDX305 was generated from a recurrent metastatic tumor, as previously described²¹. A673, TC32, A4573, TC71, and SKNMC cell lines were maintained in RPMI 1640 media (Gibco) supplemented with 10% FBS (Atlas Biologicals) and 2 mmol/L- glutamine (Life Technologies). RDES was maintained in RPMI 1640 media supplemented with 15% FBS. Isogenic cell lines, CHLA10 and CHLA9, were maintained in IMDM media (Fisher) supplemented with 20% FBS, 2 mmol/L-glutamine, and 1X Insulin-Transferrin-Selenium (Gibco). PDX305 was maintained in RPMI 1640 media supplemented with 10% FBS, 2 mmol/L-glutamine, and 1% Antimycotic-Antibiotic (Gibco). U2OS was maintained in McCoy’s 5A media supplemented with 10% FBS and 2mmol/L-glutamine and RD was maintained in DMEM with 10% FBS. All hMSC lines were maintained in MEM-alpha media supplemented with 20% FBS, 2 mmol/L-glutamine, and 1% Antimycotic-Antibiotic. Cells were cultured at 37°C with 5% CO2. Cells were all confirmed to be mycoplasma free and identities subject to STR-confirmation every 6 months.

pLKO.1 shNS (SHC002) and shFLI1 (TRCN0000005322) were used for FLI1 knockdown multiome experiments (Sigma). Knockdown plasmids were co-transfected with pCD/NL-BH*DDD (addgene, #17531) and pMD2.G (addgene, #12259) plasmids into 293FT packaging cells using PEI. 24 hours post-transfection, sodium butyrate was added to plates for 6 hours. Viral supernatant was collected 48 hours after transfection, filtered through 40 µM filters, and added the cells for 24-hours. Ewing sarcoma cells were selected by supplementing the media with 1.5 µg/mL puromycin (Gibco) for 48 hours before collection at 96 hours.

TC71, A673, and CHLA10 cells were plated at 500,000 viable cells per 10cm plate. 48 hours later recombinant human TGFβ1 (R&D 7754-BH-005, final dose 10 ng/mL in sterile 4 mM HCl and 0.1% BSA) or vehicle control was added. 24 hours later cells were trypsinized and >1×10⁶ viable cells were subjected to single cell library preparation and sequencing (details below).

Cells were washed with ice cold PBS, scraped and centrifuged, then lysed with RIPA buffer (Fisher Scientific) supplemented with protease and phosphatase inhibitors (Sigma) and sonicated (ten 30 second cycles). For TGF-β treated samples, recombinant TGFβ1 (10 ng/mL) or vehicle control (4 mM HCl, 0.1% BSA) were included in the media for either 30 minutes or 24hrs before lysis. Western blots were performed using the Bio-Rad Mini-PROTEAN Tetra System using 40 ug of lysate and 4-15% Mini-PROTEAN TGX polyacrylamide gels. Following transfer, nitrocellulose membranes were blocked in Odyssey Blocking Buffer (LI-COR) for 1 hour. Membranes were washed and incubated rotating overnight at 4°C with primary antibodies (FOSL2 CST #19967, HAND2 R&D AF3876, RUNX3 CST #13089, TEAD1 CST #12292, pSMAD2 S465/467 CST #3108, SMAD2 CST #3103, GAPDH Invitrogen AM4300, Vinculin CST #13901). Membranes were then washed four times in TBST for 5 minutes each and incubated with secondary antibodies (LI-COR IRDye 700CW or 800CW; 1:10,000) for 1 hour. After 4 additional 5 minute TBST washes membranes were scanned on a LI-COR Odyssey scanner.

Q3 normalized gene counts from GeoMx Digital Spatial Profiling (NanoString) of EwS cell line xenograft tumors formed by CHLA10 injection were obtained from previous work²¹. This processed data was used to query fusion regulated module gene sets with respect to tumor microenvironment and injection site. Gene sets values for each ROI were calculated by converting the DSP Q3 counts matrix into a Seurat object and using the AddModuleScore function.

The 78 direct EwS TGs were sourced from a previously published dataset¹⁹. Bulk RNAseq patient tumor samples were sourced from dbGaP under accession numbers PRJNA253759 and PRJNA261990. FASTQ files were aligned to reference genome (GRCh38) and using the piperna (https://github.com/furlan-lab/piperna) workflow which uses STAR⁷⁰ aligner and generates a count matrix using the Bioconductor summarize overlap function (https://www.bioconductor.org) with its default settings. Microarray data was accessed from GSE63157⁵¹. RNA probes were annotated using pd.huex.1.0.st.v2, and normalized using oligo.

Patient tumor biopsies were sourced and sequenced at St. Jude’s Children’s Hospital⁴⁹. 10X reads were demultiplexed and processed using the cellranger pipeline. Ambient RNA was filtered out using cellbender. Cells with poor RNA measurements were filtered using the following thresholds: percent mitochondrial reads < 15% and 3 < log₁₀(RNA counts) < 5.25. Dimensionality reduction and clustering was performed using standard workflow. Surrounding tissues cells were identified by using FindAllMarkers on Seurat clusters. Tumor enrichment was validated by choosing cells that express CD99, NR0B1, and EwS TGs. Identified tumor cells underwent normalization and dimensionality reduction with the same workflow. Average gene set scores were calculated using AddModuleScore().

Bulk RNAseq patient data were deconvoluted using cellSignalAnalysis (https://github.com/constantAmateur/cellSignalAnalysis). EwS subgroups 1-3 were used as transcriptomic references.

All statistical analyses were conducted using R or Python, with specific packages indicated in relevant sections of the Methods. Statistical significance was assessed using two-sided tests unless otherwise specified. For comparisons of gene set scores, motif deviations, or expression levels, Wilcoxon rank-sum or Student’s t-tests were used as appropriate based on data distribution and variance. Correction for multiple hypothesis testing was performed using the Benjamini-Hochberg method where applicable. Gene set enrichment analyses were conducted using the clusterProfiler package with MSigDB Hallmark and Gene Ontology annotations. For scRNA-seq and multiome analyses, differential expression and peak calling were performed using Seurat, ArchR, or Signac, and marker features were identified with adjusted p-values < 0.05 considered significant.

All work was approved by the Fred Hutchinson Cancer Center Institutional Review Board - Tumor Heterogeneity and Metastasis in Ewing Sarcoma - FHIRB0010627