Epigenetic and Transcriptional Programs Define Osteosarcoma Subtypes and Establish Targetable Vulnerabilities

Early osteoblast–derived and late- osteoblast–derived subtypes of osteosarcoma are characterized by distinct epigenetic transcriptional circuits and have unique vulnerabilities that suggest a new paradigm for treatment of this disease.

Abstract

Osteosarcoma is a genomically complex tumor characterized by widespread structural rearrangements. This complexity has limited the development of therapeutic strategies informed by molecular mechanisms of oncogenesis. We hypothesized that epigenetic mechanisms could drive distinct subtypes of osteosarcoma. Through analysis of chromatin accessibility, we identified an “early osteoblast–derived” cell state, characterized by upregulation of transcription factors associated with early bone development, and a “late osteoblast–derived” state, characterized by upregulation of genes involved in late bone development. We then defined core regulatory circuitries governing the underlying gene expression programs in these two cell states. Multiomic single-cell analysis indicates that these cell states coexist in a single tumor. Finally, using a panel of patient-derived xenograft models, we identified differential drug responses dependent on these cellular states. These findings create opportunities for developing new combination therapy strategies for osteosarcoma treatment and underscore the value of defining epigenetic subtypes in highly genomically complex cancers.

Significance:

This study identifies two distinct cellular states in osteosarcoma, driven by specific transcription factor circuitries linked to normal bone development. These epigenetically defined states demonstrate differential drug responses, are identifiable in patient samples, and are correlated with survival.

Introduction

Osteosarcoma is an aggressive cancer that primarily affects children, adolescents, and young adults. Treatment requires local control, followed by highly toxic chemotherapy regimens that have not changed in more than 40 years (1). Although this approach is sometimes effective for early-stage disease, osteosarcoma remains one of the pediatric cancers with the lowest survival rates, largely because of the frequent emergence of metastatic disease (2). The main challenges in osteosarcoma research, drug discovery, and clinical treatment stem from the fact that this disease lacks clear oncogenic driver genes; rather, it is characterized by high genomic complexity and significant intra- and intertumoral heterogeneity (3). Thus, although copy-number gains, deletions, and structural rearrangements in specific oncogenic drivers are common, recurrent point mutations in targetable oncogenes are few (4). Furthermore, intratumoral heterogeneity has impeded efforts to subclassify osteosarcoma based on underlying molecular features, optimize treatments based on key molecular characteristics, and predict therapeutic response. Although recent studies provided evidence that patients with osteosarcoma can be clustered based on transcriptional programs (5–7), the underlying molecular basis for this clustering remains unclear.

Epigenetic programs have been shown to play an important role in the initiation and progression of many cancers (8). DNA methylation and posttranslational modification of histones result in changes in chromatin structure and accessibility, thus regulating gene expression. Furthermore, the global analysis of chromatin accessibility has the power to identify previously unrecognized DNA regulatory elements, as well as the long-range genomic interactions involved in regulating specific cancer phenotypes. For example, a large-scale study used assay for transposase-accessible chromatin using sequencing (ATAC-seq; ref. 9) to map the chromatin accessibility landscape of 23 human cancers and demonstrated that this technique could reveal novel cancer subtypes (10). However, this analysis did not address the chromatin landscape of pediatric cancers and/or genomically complex tumors specifically. The value of chromatin analysis in pediatric sarcomas was recently underscored by work demonstrating the existence of core regulatory circuitries in rhabdomyosarcoma and the role of specific transcription factors (TF) in recruiting enhancers to activate cell type–specific transcription (11).

To understand the epigenetic landscape of osteosarcoma, we applied ATAC-seq to a combination of patient samples, patient-derived xenografts (PDX), and PDX-derived cell lines (PDX-CL). This identified two previously unknown osteosarcoma subtypes with distinct chromatin accessibility and enhancer regulation. By integrating ATAC-seq with histone 3 lysine27 acetylation (H3K27ac) chromatin binding analysis, we identified critical superenhancers (SE) for these subtypes and defined the core regulatory circuitries (CRC) that govern the underlying gene expression programs. We demonstrate that there is considerable plasticity in these cell states and that specific TFs regulate the transition from one state to another. Single-cell analysis strongly suggests that these two subtypes can coexist with the same tumor. Critically, we validate that the two epigenetically defined cell states have differential drug response to targeted therapies. We further demonstrate that these distinct epigenetic states are seen in clinical samples and that metastatic tumors from the same patient can exist in either of these two cell states, findings that are directly relevant for informing new strategies for combination therapy. Therefore, our study could pave the way toward subtype-specific therapeutic strategies for osteosarcoma.

Results

Epigenetic Landscape Identifies Two Distinct Cellular States in Osteosarcoma

A comprehensive epigenetic analysis of osteosarcoma models and patient samples was performed as outlined in Fig. 1A. First, we performed ATAC-seq on a panel of samples from patients with osteosarcoma (n = 8), PDX models (n = 17), and PDX-CLs (n = 11). This included samples derived from either metastatic or primary lesions. For comparison, we also evaluated publicly available osteosarcoma cell lines (n = 9; Supplementary Fig. S1A; Supplementary Table S1). Principal component analysis (PCA) of the top 10,000 open genomic regions with highest variance identified two clusters in the first principal component (Fig. 1B). K-means clustering and Uniform Manifold Approximation and Projection for Dimension (UMAP; ref. 12) embedding confirmed the existence of these two clusters (Supplementary Fig. S1B–S1D). To examine whether these clusters were related to trajectories of bone development, we incorporated chromatin accessibility data obtained from early human skeletal stem cells (hSSC; ref. 13) or mature osteoblasts (hFOB1.19; ref. 14). One of the osteosarcoma clusters shared similarities with chromatin accessibility observed in hSSCs, whereas the other was more similar to hFOBs (Fig. 1B).

Figure 1.

Epigenetic landscape identifies two distinct cell states in osteosarcoma. A, Schematic of overall approach. B, PCA of ATAC-seq data (n = 85) identifies two dominant subtypes (EOD and LOD). C, DARs between subtypes: 6,267 peaks in EOD and 6,790 peaks in LOD. D, Distribution of distance between peak–gene association among DARs. E, GO biological process enrichment of genes associated to DARs in EOD (5,643 genes) and LOD (5,717 genes). F, ATAC-seq profiles and gene expression from RNA-seq (n = 38) of EOD. Top, A distal regulatory region is accessible in EOD associated with CTNNB1 (β-catenin). Bottom, IBSP locus shows open chromatin at the promoter and intragenic regions in EOD. G, ATAC-seq profiles and gene expression from RNA-seq (n = 38) of LOD. Top, A distal regulatory region is accessible in LOD associated with integrin subunit α3 (ITGA3). Bottom, TM4SF1 locus shows open chromatin at the promoter and 3′ untranslated region in LOD. The central line within each box represents the median, whereas the top and bottom of the boxes correspond to the first and third quartiles, respectively. P values were derived from the Wilcoxon test. See also Supplementary Fig. S1. scATAC-seq, single-cell ATAC-seq; scRNA-seq, single-cell RNA-seq.

These results suggest that the ATAC-seq–identified clusters correspond to subtypes of osteosarcoma that share properties with either early or late bone development. We therefore labeled these osteosarcoma subtypes as “early osteoblast–derived” (EOD) or “late osteoblast–derived” (LOD). The EOD subtype included most of the MYC-amplified PDX-CLs, whereas the LOD subtype included most previously established long-term osteosarcoma cell lines. Furthermore, PDX-CLs in the EOD subtype showed a generally more aggressive behavior in mouse models than those in the LOD subtype (15). The proliferation rate of all the cell lines was similar, indicating that these subtypes are not driven by the cell cycle (Supplementary Table S1; Supplementary Fig. S1E). The samples analyzed included patient samples, PDX samples, and PDX-CLs derived from the same patient (trio samples), as well as PDX samples and PDX-CLs derived from the same patient (duo samples; Supplementary Table S1). These duos and trios tended to classify within the same subtype, although PDX-CLs in some cases diverged from the clustering of their parental PDX (Fig. 1B; Supplementary Table S1). As osteosarcoma is characterized by widespread structural rearrangements associated with large copy-number alterations (4), we evaluated whether the observed subtypes were driven by the underlying copy-number structure. The separation of the samples into two subtypes was maintained after normalization for copy number (Supplementary Fig. S1F), indicating that these subtypes are due to changes in chromatin accessibility and not simply due to the underlying copy number.

We used differential accessible region (DAR) analysis to identify biological processes enriched in EOD and LOD. There were 5,643 genomic regions that were significantly more accessible in EOD versus 5,717 more accessible in LOD (see “Methods”; Fig. 1C; Supplementary Table S2; ref. 16). Most DARs for both EOD and LOD were located in distal regions [±50–500 kb from the transcription start site (TSS); Fig. 1D]. The accessible regions in EOD were associated with genes related to mesenchymal cell proliferation, osteoblast differentiation, and Wnt signaling, among others (Fig. 1E; Supplementary Fig. S1G). For example, open chromatin in the promoter of CTNNB1 (β-catenin) correlated with higher expression of this gene in EOD (Fig. 1F). Wnt/β-catenin signaling affects the proliferation and differentiation of osteoblast progenitor cells (17). Similarly, IBSP, which is involved in bone formation (18, 19) and is upregulated in several cancers (20), was also highly expressed and associated with more accessible chromatin in EOD (Fig. 1F). By contrast, LOD-accessible regions were enriched for genes related to extracellular matrix (ECM) organization and cell migration (Fig. 1E; Supplementary Fig. S1G). For example, ITGA3, which encodes integrin subunit α3, a cell surface adhesion protein that interacts with ECM proteins (21), showed higher accessibility in distal regulatory elements and its promoter in LOD (Fig. 1G). Greater chromatin accessibility was also seen at the promoter of TM4SF1, a gene that encodes for a member of a family of integral membrane proteins previously implicated as drivers of tumor proliferation and cell migration (Fig. 1G; refs. 22, 23). The increased accessibility in both genes correlates with higher gene expression in LOD (Fig. 1G). Taken together, mapping the epigenetic landscape of osteosarcoma samples allowed us to define two osteosarcoma subtypes, one that shares features of skeletal stem cells and early bone development stage (i.e., EOD) and the other that exhibits similarity with mature osteoblasts (i.e., LOD). These two cell states display profound differences in their chromatin accessibility, suggesting that, at molecular level, they have clearly distinguishable molecular drivers.

Osteosarcoma Cell States Are Regulated by a Unique Set of TFs

To define the active TFs regulating gene expression in EOD versus LOD, we analyzed TF-binding motifs enriched in differentially accessible peaks (see “Methods”; refs. 24, 25). EOD-accessible regions were highly enriched for RUNX2, RUNX3, MEOX2, MEF2A/B/C, and other TFs known to be involved in bone stem cell differentiation (26, 27). On the other hand, LOD-accessible regions were enriched in AP-1 TF-binding sites, including FOSL1, FOSL2, and JUN, TFs which are known to play important roles in bone homeostasis (Fig. 2A; Supplementary Fig. S2A; Supplementary Table S3; refs. 26, 28, 29).

Figure 2.

A distinct set of TFs regulate the two epigenetic cell states. A, TF enrichment analysis of DARs (6,267 peaks in EOD and 6,790 peaks in LOD). The deviation score was generated by chromVAR. The x axis indicates the bias-corrected deviations in accessibility. For each motif this value represents how different the accessibility for peaks with that motif is from the expectation based on all cells being equal, corrected for biases. B, Pearson correlation matrix based on RNA-seq data of PDX-derived and established cell lines (n = 19) evaluating the expression of top 24 enriched and expressed TFs. C, Immunoblotting of differential protein expression using total lysates. D, Differential TF binding to chromatin. Western blots of the isolated cytosolic (Cyto), nuclear (Nuc), and chromatin (Chr) fractions. E, Immunoblot analysis of PDX panel of specific markers and TFs for EOD/LOD. F, Immunoblotting of osteogenic markers. G, Analysis of the TFs as subtype-specific dependencies utilizing the DepMap CRISPR dependency score (DepMap Public 25Q2; n = 17; ref. 97). The central line within each box represents the median, whereas the top and bottom of the boxes correspond to the first and third quartiles, respectively. P values were derived from a two-sided t test. See also Supplementary Fig. S2.

We then used gene expression to explore the existence of connected TF modules. The expression of RUNX2, MEF2C, DLX6, MEF2A, HOXA5, MEF2D, and VAX2 was upregulated and correlated in EOD, whereas FOSL1, JUN, JUNB, and FOSL2 expression was highly correlated in LOD (Fig. 2B). In agreement with these results, MEF2C, c-MYC, RUNX2, and DLX2 protein levels were also higher in EOD, whereas FOSL1, FOSL2, and JUN protein expression was higher in LOD (Fig. 2C; Supplementary Fig. S2B and S2C). This also corresponded to increased chromatin binding of these TFs (Fig. 2D; Supplementary Fig. S2D). Higher protein expression of SP7, SATB2, and MEF2C was also seen in EOD PDXs (Fig. 2E and F). Conversely, LOD PDXs had higher protein expression of FOSL1/2 and JUN. One PDX (OS526) showed enrichment of both EOD and LOD markers, suggesting that cells from both states can coexist in the same tumor or that some tumor cells have an intermediate phenotype (Fig. 2E). To evaluate whether these TFs could lead to subtype-specific vulnerabilities, we assessed whether there were dependencies in osteosarcoma cell lines previously analyzed in the pediatric DepMap resource (30). In the DepMap data, RUNX2 was an EOD dependency, whereas FOSL1 and JUN were LOD dependencies (Fig. 2G; Supplementary Fig. S2E). Together, this analysis nominates specific TFs that define the epigenetic landscape of osteosarcoma.

Epigenetic Subtypes Are Driven by Superenhancer Regulation

CRCs are networks of interdependent TFs that collectively determine and maintain the identity and function of specific cell states (31, 32). These CRCs play a critical role in driving and maintaining the oncogenic programs that define the malignant state. However, whether CRCs exist in osteosarcomas is unknown. CRCs can be identified by their association with SEs, genomic regions marked by long stretches of H3K27ac (33–36). To determine whether the TFs enriched in EOD and LOD osteosarcoma subtypes are indeed part of CRCs, we mapped H3K27ac to identify enhancer elements (see “Methods”; ref. 37). H3K27ac binding was profiled in the complete panel of PDX-CLs (Supplementary Table S1; n = 11). Using the top 10,000 acetylated regions with the highest variance, we observed a separation between EOD and LOD samples that was consistent with the ATAC-seq clustering described above (Supplementary Fig. S3A). SEs were defined using Rank Ordering of Super Enhancers (ROSE; refs. 36, 38, 39) to “stitch” adjacent H3K27ac peaks into potential SE regions. Separation into EOD and LOD was observed using the top 1,000 SEs per cell line (Fig. 3A; Supplementary Fig. S3B).

Figure 3.

Epigenetic subtypes are driven by acetylated regions. A, PCA of the top 1,000 stitched peaks by SE score per cell line (n = 11 PDX cell lines) with the highest variance (n = 28 H3K27ac CUT&RUN libraries plus their respective 28 IgG CUT&RUN libraries). PCA is color-coded by ATAC-seq subtype. PC, principal component. B, Differential analysis of the SEs per EC (n = 5,431 SEs). The top TFs potentially regulated by each SE are labeled. C, Left, The union of all SEs detected among osteosarcoma cell lines was expanded to a width of 500 kb and split into 50 bins, and the total number of H3K27ac reads minus IgG background reads among pooled samples was quantified per bin. SEs are ordered on the basis of their ability to distinguish EOD from LOD PDX-CLs. Right, column 1: log₂ fold enrichment of H3K27ac signal among SEs between EOD/LOD subtypes, with those significantly enriched in EOD or LOD samples color-coded by subtype. Column 2: the presence of EOD/LOD genes based on the enrichment of RNA transcripts within 500 kb of each SE is indicated by subtype color-coded lines. Genes are significantly more likely to be found within the range of SEs enriched in the same subtype (Fisher exact test P value < 2.2e−16). Column 3: the presence of EOD-/LOD-enriched TFs is annotated with subtype color-coded lines, respectively, if they are within 500 kb of each SE. TFs are significantly more likely to be found within the range of SEs enriched in the same subtype (Fisher exact test P value 7.6e−8). Select TFs are annotated on the right, and rounded rectangles around TF labels indicate that the TF-binding motif is enriched among DARs from the same subtype. D and E, Subtype-specific SEs. Top, Acetylated regions. Center, Chromatin accessibility. Bottom, Position of the SE in the genome showing its potential target. D, EOD-specific SE in the DLX5–DLX6 loci. E, LOD-specific SEs in the NFATC2 locus. See also Supplementary Fig. S3.

To identify genes associated with SEs for each subtype, SEs among all cell lines were combined into a unified set and assigned to the nearest gene within a 1-Mb window, provided that the gene promoter overlapped an H3K27ac peak (Supplementary Table S4). Differential analysis between SE regions identified 1,615 SEs with a stronger signal in EOD and 1,447 SEs with a stronger signal in LOD (Fig. 3B; FDR = 0.05). Cell lines of the same subtype had similar SE signatures, and genes with subtype-enriched RNA levels are significantly more likely to be found within 500 kb of SEs enriched in the same subtype (Fisher exact test P value < 2.2e−16; Fig. 3C). Enriched genes associated with SEs in EOD were associated with skeletal system development, skeletal system morphogenesis, and osteoblast differentiation, consistent with the suggestion that EOD is a less differentiated state. In contrast, enriched genes associated with SEs in LOD are mostly associated with cell migration, cell motility, and positive regulation of cell differentiation (Supplementary Fig. S3C). Furthermore, we assessed which TFs with subtype-specific expression were associated with SEs. MEF2C, RUNX2, and DLX5 (40) were associated with EOD SEs (Fig. 3C and D). Specific LOD SEs were associated with TFs with differential gene expression (DGE) between subtypes (e.g., NFATC2; Fig. 3E). Collectively, these results suggest that SEs are involved in regulating gene expression unique to specific osteosarcoma subtypes.

Osteogenic TFs Are Core Regulatory Circuitry Members of the EOD Subtype

To define the essential TF interactions that drive the osteosarcoma transcriptional program in each subtype, we constructed models of CRCs by searching for TF motifs within regions of accessible chromatin among SEs (Fig. 4A; see “Methods”); the presence of a significant TF motif within such regions is considered evidence of TF binding in that region. For a given TF, the number of other TFs which target its promoter and/or assigned SEs was tallied as its “in-degree,” a measure of its importance downstream of regulatory pathways. The genes which a given TF targets via its promoters and SEs are tallied as its “out-degree,” a measure of its importance upstream of regulatory pathways (32). TFs that are part of a CRC can be identified by the circular interdependencies between multiple TFs (cliques; ref. 33), and the centrality of a TF is indicated by high in- and out-degrees (Fig. 4B; Supplementary Fig. S4A; ref. 32). The CRC score of a given TF indicates the fraction of cliques containing that TF. Clustering samples based on the TF CRC scores again yielded clear EOD and LOD subtypes (Fig. 4C; Supplementary Fig. S4B). For example, ZNF148 both regulates and is regulated by TFs only in EOD cell lines, whereas NFATC2 both regulates and is regulated by TFs only in LOD cell lines (Fig. 4C; Supplementary Fig. S4B). We identified the TFs with the most differential CRC scores between subtypes, as well as TFs that were associated with SEs (Fig. 3C) but without a differential CRC score between subtypes, as these may be playing different roles in each cell state. Thus, potential CRCs per epigenetic subtype were defined by integrating TFs that were differentially associated between LOD and EOD states, as identified through both CRC inference and chromatin accessibility analyses linked to super- enhancers (Supplementary Fig. S4C).

Figure 4.

Distinct CRCs drive osteosarcoma subtypes. A, Schematic of enhancer-based CRC analysis. For each sample, the top 1,000 enhancers as determined by ROSE2 were extracted. For every TF associated with a top enhancer, in-degrees are assessed by motif analysis within ATAC-seq peaks overlapping the enhancer; out-degrees are assessed for each TF associated with a top enhancer by determining all other bound enhancers at TF gene loci. Node connections between TFs were used to discern autoregulatory cliques. B, Degree plots of four representative PDX-CLs (two EOD and two LOD). C, A heatmap of clique enrichment scores for the TFs associated with top enhancers per subtype. Clique enrichment scores are calculated by the percentage of cliques within each sample of which that TF is a constituent. TFs and samples are clustered by Euclidean distance to reveal the most consistent highly connected TFs. D, Evaluation of the expression of all TFs after the indicated knockdown (siSP7, siRUNX2, siSATB2, or siZNF148). Values >1 indicate negative regulation, whereas values <1 indicate positive regulation. E, qRT-PCR of OS742 (EOD) treated with siRNA to each member of the CRC—SATB2, SP7, RUNX2, and ZNF148—resulted in the decreased expression of all the CRC members. The line represents the mean with SD. F, Regulatory network validated by knocking down the TFs with transient siRNAs and evaluating the effect on the circuitry components for EOD. The TFs are connected if the knockdown has an effect of gene expression with FC <0.5 to indicate a positive regulation or FC >1.5 to indicate a negative regulation with respect to the nontargeting control siRNA transfection. Solid lines indicate negative regulation, whereas dotted lines (arrows) indicate positive regulation. The targets of the CRC for EOD are marked by different colors to clarify the connections. G, SATB2, SP7, RUNX2, and ZNF148 form a positive interconnected co-regulatory loop. H, Heatmap demonstrating co-occupancy of the four EOD-CRC TFs as well as chromatin accessibility and H3K27ac patterns based on CUT&RUN sequencing. The union of all TF peaks is flattened and resized to a 4-kb window about the region center (heatmap rows), and reads are counted in 80 bp bins and then converted to RPM. Windows are ordered by total RUNX2 RPM values, whereas scales are illustrated at top. Top 20,000 regions are shown. I, CUT&RUN tracks for the four TFs forming the EOD-CRC and the histone marks H3K27ac and H3K4me3 in the locus of RUNX2 (left) and SP7 (right). Each TF has a different scale based on its signal. J and K, TF enrichment analysis of DARs. The deviation score was generated by chromVAR. J, DARs between nontargeting control and after RUNX2 knockdown. K, DARs between nontargeting control and after SP7 knockdown. See also Supplementary Fig. S4. For J and K, the x axis indicates the bias-corrected deviations in accessibility. For each motif this value represents how different the accessibility for peaks with that motif is from the expectation based on all cells being equal, corrected for biases.

To interrogate the functional interactions within these CRCs, we knocked down the corresponding TFs using siRNA and evaluated the expression of other CRC components in two PDX-CLs (EOD: OS742 and LOD: OS526) by qRT-PCR (Supplementary Table S5). Transient siRNA-mediated knockdown led to decreased expression levels for all tested TFs (Supplementary Fig. S4D and S4E; see “Methods”; Supplementary Table S5). EOD TFs formed an intricate network in which four master regulators—SATB2, SP7, RUNX2, and ZNF148—exerted positive regulation on most of the EOD TFs and negative regulation on most of the LOD TFs (Fig. 4D and E). SATB2, SP7, and RUNX2 are known to regulate early stages of osteoblast differentiation and bone formation, in which their dysregulation may lead to osteosarcoma (41, 42). ZNF148 has been previously implicated in differentiation and tumorigenesis (43) but to our knowledge has not previously been associated with the other three. These four master regulators interact with each other, forming a feed-forward autoregulatory CRC for EOD (Fig. 4F and G). To further evaluate this CRC, we used CUT&RUN analysis to characterize the binding of each TF to chromatin. This demonstrated that the EOD CRC members bind together to regions of accessible chromatin marked by enhancers across the genome (Fig. 4H and I).

In contrast, the regulatory network of LOD did not show a straightforward interaction between core TFs by qRT-PCR. However, SP7, an EOD TF, is negatively regulated by most of the LOD TFs (Supplementary Table S5). As FOSL1 and JUN are dependencies for LOD (Fig. 2G), we evaluated the co-binding across the genome of the members of the AP1 complex and TEAD1 previously defined by ATAC-seq (Fig. 2C). We observed that these four TFs (FOSL1, FOSL2, JUN, and TEAD1) bind together to regions with accessible chromatin and marked by enhancers (Supplementary Fig. S4F and S4G), suggesting they act together to drive the LOD cell state. RUNX2, an EOD CRC member, does not show binding enrichment in the genomic regions regulated by the LOD TFs (Supplementary Fig. S4F).

To evaluate chromatin remodeling after disruption of the EOD CRC, we knocked down either RUNX2 or SP7 and performed ATAC-seq to identify the extent to which the chromatin landscape is dependent on these two TFs. Loss of either RUNX2 or SP7 led to an enrichment of LOD TF motifs in accessible chromatin (FOSL1/2, JUN, and FOS; Fig. 4J and K), suggesting that cells in the EOD state have significant plasticity and can transition to LOD if the EOD CRC is disrupted. Moreover, we observed that overexpression of the subtype-specific TFs RUNX2 (EOD) and FOSL1 (LOD) is sufficient to shift osteosarcoma cell states. The overexpression of RUNX2 induced an EOD-like program in LOD cells, measured by the decrease in FOSL1 protein abundance. Conversely, the overexpression of FOSL1 in EOD cells increased the levels of phosphorylated ERK (p-ERK), a marker of LOD (Supplementary Fig. S4H and S4I). Taken together, these data establish the existence of CRCs in osteosarcoma and provide evidence for epigenetic plasticity of osteosarcoma cells.

Subtype-specific Gene Signatures Are Detected in Patient Samples and Are Associated with Prognosis

We hypothesized that EOD and LOD may represent cell states that are identifiable in clinical samples using transcriptional signatures. An unsupervised PCA of RNA sequencing (RNA-seq) samples showed separation of the two subtypes, although this was less clear than the ATAC-seq–based clustering (Supplementary Fig. S5A and S5B). To define a comprehensive set of genes that define each subtype, we integrated bulk ATAC-seq and gene expression data (RNA-seq) of the complete panel of osteosarcoma cell lines (n = 19) using the approach outlined in Fig. 5A. For the RNA-seq analysis, each sample was assigned to one subtype (Supplementary Fig. S5B), and a supervised DGE analysis was performed [log fold change (|log FC|) > 1; 298 genes from EOD and 352 genes from LOD]. A gene signature of 237 genes was identified (126 EOD-enriched and 111 LOD-enriched genes) by overlapping the differentially expressed genes with the genes associated with accessible chromatin (Fig. 5A; Supplementary Table S6). RNA-seq data between primary patient samples (n = 38) exhibited separation into three clusters: 11 (29%) patient samples showed higher expression of EOD signature transcripts, 13 (34%) showed higher expression of the LOD signature, and 14 (37%) exhibited an intermediate pattern with the expression of genes from both subtypes (“blend” group; Fig. 5B; Supplementary Table S7). These results support the hypothesis that the EOD and LOD subtypes are present within primary human tumors, highlighting the potential clinical relevance of these epigenetic subtypes.

Figure 5.

Subtype-specific gene signature separates patient samples into three groups. A, Integration of the differential chromatin accessibility and DGE analysis to define the subtype-specific gene signatures. Gene signatures were defined using the complete panel of cell lines (n = 19). B, Heatmap of the hierarchical clustering (WarD) of gene expression data of patient samples using the 237 genes defined in the gene signature. First cluster of patients is EOD high, second cluster of patients is LOD high, and third cluster is blend. C and D, Kaplan–Meier plot of OS (n = 37) and event-free survival (EFS; n = 38) of the three groups classified by the hierarchical clustering shown in B: EOD, LOD, and blend. Log-rank and likelihood ratio test P values of the overall model are indicated on the plot. Dotted lines indicate the time point at which 50% of the individuals in a specific group have experienced the event of interest (death for OS and death or relapse for event-free survival). Below each plot is a risk table presenting the number and percentage of patients at risk for each group at various time points. The risk table shows the individual HR with P values and 95% CIs for each group comparison, using the LOD group as the reference. The covariate metastasis at diagnosis (metdxY) was included for OS. OS analysis is shown in C and event-free survival analysis is shown in D. E, Immunofluorescence of 15 patient samples contained in a TMA. SP7 was used as an EOD marker and c-JUN as an LOD marker. The positive cells were quantified as the % of cells that overlap with DAPI. The size of the circles indicates the double-positive cells. F, Representative images of the immunofluorescence experiment in E. See also Supplementary Fig. S5.

Survival analysis revealed significant differences among the three groups (likelihood ratio P = 0.002). Compared with LOD, both EOD- and blend-classified patients demonstrated worse overall survival (OS). EOD-classified patient samples showed an HR for OS of 2.15 [95% confidence interval (CI), 0.61–7.54; P = 0.23], whereas the blend-classified patients demonstrated a more pronounced risk with HR = 4.52 (95% CI, 1.20–17.06; P = 0.03; Fig. 5C). Event-free survival demonstrated a similar trend, with the LOD-classified patients showing significantly better event-free survival (likelihood ratio P = 0.025). Compared with the LOD group, both EOD and blend had worse event-free survival, with blend showing a more pronounced risk (HR = 6.68; 95% CI, 1.39–32.21; P = 0.02), whereas the EOD showed an elevated but nonsignificant HR of 2.72 (95% CI, 0.49–15.02; P = 0.25; Fig. 5D). These results suggest that tumors in an intermediate state between EOD and LOD could have more plasticity and possibly more drug tolerance.

These findings were validated with an independent cohort using the Therapeutically Applicable Research to Generate Effective Therapy (TARGET) osteosarcoma dataset. RNA-seq from patient samples from TARGET were stratified using the same gene signature (Fig. 5A) and exhibited similar gene stratification (EOD, LOD, and blend; Supplementary Fig. S5C and S5D). As with the first cohort, we stratified the sample clustering based on the expression pattern from the heatmap into three groups (LOD, EOD, and blend). Survival analysis again revealed significant differences among the three patients classified in each of these three groups (likelihood ratio P = 0.025). Both EOD and blend groups demonstrated worse OS, with the EOD group showing a more elevated HR of 5.48 (95% CI, 0.73–41.16; P = 0.1), whereas the blend group showed an HR of 1.52 (95% CI, 0.14–16.82; P = 0.73; Supplementary Fig. S5D).

We then examined whether different samples from the same patient exhibited the same or different subtype-specific gene signatures. We identified four patients with samples that display both signatures (patients OS457-A, OS661-B, OS52-C, and OS357-D), with the EOD signature found in the biopsy samples and the LOD signature occurring later in the resection or metastatic samples (Supplementary Fig. S5E; Supplementary Table S7). The patient samples with an enriched EOD signature show high expression of the EOD-CRC members (Supplementary Fig. S5E).

Furthermore, we evaluated the heterogeneity of 15 patient samples in a tumor microarray (TMA) using one marker per subtype (EOD = SP7 marker and LOD = JUN marker) by immunofluorescence. Fourteen of fifteen patients showed an enrichment of the subtype-specific marker, consistent with the subtype they were predicted to belong to (Fig. 5E and F; Supplementary Table S8). Most of the evaluated tumors (11 of 15) showed both cell states in the same sample, consistent with their blend classification in the previous heatmaps (Fig. 5B; Supplementary Fig. S5E). These findings suggest that patients with osteosarcoma may have tumors that exhibit characteristics of both subtypes, either within the same tumor or in different tumors, a finding with potentially profound implications for the development of new therapies for this disease.

Multiomic Analysis Suggests Intratumoral Coexistence of EOD and LOD Subtypes

The existence of tumors with a gene expression profile that is not clearly defined as EOD or LOD (Fig. 5B) suggests the presence of an intermediate phenotype, which could be due to the coexistence of cells from both subtypes in some tumors. To interrogate the heterogeneity of cell states and to analyze the relationship between open chromatin and gene expression at single-cell resolution, we performed simultaneous single-nucleus RNA-seq (snRNA-seq) and single-nucleus ATAC-seq (snATAC-seq) in five PDXs (four EOD-PDXs and one LOD; total cells = 47,928) and two PDX-CLs, OS833 (EOD) and OS384 (LOD; total cells = 19,592; see “Methods”; Fig. 6; Supplementary Figs. S6A–S6F and S7A–S7H; Supplementary Table S1). Using the gene expression signatures defined above (Fig. 5A), we identified 126 genes for EOD and 111 genes for LOD that met a defined expression cutoff and used these as subtype-specific signature gene sets. Individual cells were then classified as either EOD or LOD (see “Methods”; Fig. 6A and B; Supplementary Fig. S7A and S7B). Coexistence of both subtypes was observed in PDX526, PDX384, and PDX107, consistent with their positioning in bulk ATAC-seq PCA (Figs. 1B and 6C; Supplementary Fig. S6A) and immunoblot analysis (Fig. 2E). In contrast, PDX833 showed predominantly EOD cells (95.9%), and PDX774 contained 47.7% EOD cells with minimal LOD representation (Fig. 6C; Supplementary Fig. S6A). A significant fraction of cells did not have a strong signature for either EOD or LOD and were labeled as “undetermined” (Fig. 6B and C; Supplementary Fig. S7C; see “Methods”). In contrast to these results in PDXs, the PDX-CLs were more homogeneous with regard to cell state (Supplementary Fig. S7B and S7C). This suggests that cell lines tend to drift to either LOD or EOD, whereas primary tumors that have not been exposed to tissue culture conditions are more likely to contain both subtypes.

Figure 6.

Single-cell analysis indicates coexistence of LOD and EOD in a single tumor. ATAC-seq + RNA-seq. A and B, An integrated UMAP single-cell multiome (ATAC-seq + RNA-seq) from five osteosarcoma PDX (total = 47, 928 cells; EOD = 21,747 cells; and LOD = 14,548 cells). A, Cells are annotated with PDX IDs. B, Cells are annotated based on gene signatures for EOD/LOD (Fig. 5A; 237 genes). Gray cells represent cells that cannot be confidently classified as EOD or LOD. C, Stacked bar plots showing the percentage of cells enriched for EOD, LOD, or undetermined in each PDX. D, Latent time overlay on UMAP with density contours. Mean latent time values are labeled for EOD (0.454) and LOD (0.671) populations. The color scale represents velocity-inferred latent time. E, Density distributions of latent time by cell state. Dashed lines indicate mean values for each group. F and G, Transcriptional and chromatin dynamics along the osteosarcoma differentiation trajectory. Module scores for EOD (RUNX2, SP7, MEF2C, SATB2, and ZNF148) and LOD (FOSL1, FOSL2, JUN, ATF4, and NFATC2) gene programs plotted against velocity-derived latent time. Lines represent LOESS-smoothed trends with 95% CIs (shaded), whereas colors represent different cell states. F, Transcript module score plotted against latent time. G, Chromatin accessibility at TF-binding motifs across latent time. H, Scheme showing the association of peaks with genes and the identification of genes marked by DORC per epigenetic subtype (EOD vs. LOD). I, Scatter plot representing genes with the most correlated chromatin accessibility peaks and their expression levels. The y-axis indicates the total number of correlated cis-elements associated with each gene. The x-axis ranks the genes from those with the fewest to the most correlated peaks. Colored dots indicate genes with the most correlations (defined as the top 90^th percentile); DORCs that are associated with genes that are significantly different in EOD (top) and in LOD (bottom). Genes marked as “high” represent genes with the associated DORCs but are not assigned to one subtype. J and K, Peak–gene associations for two representative genes: (J) TRPS1 (EOD gene) and (K) COL6A3 (LOD gene). Arches at the bottom indicate any peaks that are significantly correlated with gene expression and are drawn connecting the predicted TSS of the gene. The color indicates the score of the peak–gene correlation. Relative gene expression per subtype is shown on right. See also Supplementary Fig. S6.

We next sought to infer the ability of cells to transition from LOD to EOD or from EOD to LOD. RNA velocity describes the rate of gene expression change for an individual gene based on the ratio of spliced and unspliced mRNA. It can therefore be used to estimate the trajectory of individual cells from one cell state to another (44, 45). We performed RNA velocity analysis to investigate the developmental relationship between EOD and LOD cells (Fig. 6D and E; see “Methods”). EOD cells were estimated to have an earlier developmental trajectory with a mean latent time of 0.454 compared with 0.671 for LOD cells (P < 0.001, Wilcoxon rank-sum test; Supplementary Fig. S6B–S6D). The undetermined population displayed the earliest mean latent time (0.382), suggesting that these cells occupy a less differentiated state. Velocity vector analysis suggested bidirectionality between EOD and LOD states (Supplementary Fig. S6C). Although some regions showed EOD→LOD transitions, the predominant observed pattern was LOD→EOD, indicating interconversion but with directional bias toward the EOD state. Velocity magnitude measurements showed that EOD cells (mean = 0.0148) exhibited higher transcriptional dynamics than LOD cells (mean = 0.0123). Undetermined cells displayed the highest velocity magnitude (0.0165), suggesting a more active transcriptional dynamic (Supplementary Fig. S6B–S6D).

To examine state-specific transcriptional programs along the latent trajectory, we calculated module scores using TFs playing a role in defining the cell states (Figs. 2–4). The EOD module consisted of RUNX2, SP7, MEF2C, SATB2, and ZNF148, and the LOD module included FOSL1, FOSL2, JUN, ATF4, and NFATC2. EOD cells maintained high early-module scores through latent time 0.75, followed by a sharp decline. This pattern suggests a progressive loss of early-module programming with increasing latent time, consistent with our model of state transitions. LOD cells showed low early-module expression throughout the trajectory while maintaining high late-module scores across all latent time points (Fig. 6F). Analysis of TF-binding motifs revealed distinct temporal dynamics between early- and late-module regulatory elements. Early-module motifs showed high initial accessibility in EOD cells (score ∼2.5) that decreased rapidly to baseline levels by latent time 0.5. LOD cells maintained relatively low accessibility at early-module motifs throughout the trajectory. For late-module motifs, EOD cells began with negative accessibility scores that increased between latent times 0 to 0.5, reaching peak values around 0.25 to 0.5 before stabilizing. LOD cells maintained elevated accessibility at late-module motifs across all time points, with the highest values at latent time 0.25 to 0.5. The undetermined population showed intermediate patterns for both motif sets (Fig. 6G).

Together, these multimodal analyses support a dynamic model of osteosarcoma cell state plasticity. Established EOD and LOD cells can maintain their transcriptional identities while undergoing chromatin and transcriptomic alteration, which may prime cells for state interconversion. The predominance of LOD→EOD velocity vectors, combined with progressive chromatin remodeling and altered expression levels, indicates that differentiated cells retain capacity for transitions toward earlier developmental states. This plasticity seems to be mediated in part through dynamic epigenetic mechanisms, providing a framework for understanding how osteosarcoma maintains cellular heterogeneity between LOD and EOD states and potentially evades therapeutic interventions. The undetermined cell population, characterized by the highest velocity magnitude, the earliest latent time, and intermediate expression profiles, may represent either uncommitted progenitors or cells actively undergoing state transitions. However, lineage-tracing studies would be required to establish their functional role in osteosarcoma development in later studies.

Next, we examined significant correlations between chromatin accessibility and gene expression (Fig. 6H). We identified 3,750 correlated peak-to-gene events, suggesting a robust interplay between chromatin accessibility and gene expression dynamics (Supplementary Table S9). Further analysis focused on identifying a subset of genes within regions with high correlations, termed domains of regulatory chromatin (DORC; ref. 46). DORCs are defined as genomic areas with higher concentrations of regulatory elements that are more likely to have a role in controlling gene expression. Genes marked by DORCs are likely to be involved in lineage determination (46). DORCs were defined as genes that exhibited correlations in the highest 90th percentile, i.e., genes in this region had at least nine significant peak-to-gene pairs. We identified 294 genes with membership within DORCs. Of these, 140 showed significantly higher chromatin accessibility and gene expression in the EOD cells versus LOD, whereas 117 genes were significantly higher in the LOD cells (Fig. 6I). EOD DORCs included GLI2 (Hedgehog signaling; ref. 47) and TRPS1 (skeletal development; Fig. 6J; ref. 48), genes involved in the regulation of osteoblast differentiation and endochondral ossification, as well as genes associated with histone methyltransferase activity signaling pathways (Supplementary Fig. S6E; Supplementary Table S9). Conversely, LOD DORCs included COL6A3 and COL6A2, genes associated with ECM organization and collagen organization (Fig. 6K; Supplementary Fig. S6F; Supplementary Table S9; ref. 49). We observed similar peak-to-gene pairs in the two PDX-CLs. Thirty-eight genes showed higher chromatin accessibility in EOD, whereas 274 genes were significantly higher in the LOD cells (Supplementary Fig. S7D–S7F). EOD DORCs included genes related to cell fate commitment and stem cell differentiation (Supplementary Fig. S7G; Supplementary Table S10), whereas LOD DORCs contained genes related to ECM and cell migration (Supplementary Fig. S7H; Supplementary Table S10). Taken together, the multiome data support the hypothesis of an epigenetically driven developmental progression within EOD and LOD subtypes.

Epigenetically Defined Cell States Have Differing Drug Responses

To test whether the epigenetically defined cell states described above affect drug response, we evaluated the results of drug testing using a library of 40 targeted therapies in seven PDX-CLs and two previously established cell lines (Supplementary Fig. S8A; Supplementary Table S11). This screen was done without the prior knowledge of the corresponding epigenetic subtype of these cell lines. To compare drug sensitivity between ATAC-seq–defined clusters while accounting for drug-specific variance patterns, we employed a linear mixed-effects model with random intercepts and slopes: IC₅₀ ∼ ClusterByATAC + (ClusterByATAC|Drugs). Drug-specific pairwise comparisons identified seven drugs meeting the significance threshold of raw P values < 0.05 and FDR-corrected q values < 0.2 (Supplementary Fig. S8B). Notably, a subset of cell lines responded to the MEK inhibitor (MEKi) trametinib, whereas other cell lines responded to an AURKB inhibitor (AURKBi, AZD1152; Fig. 7A–C). All trametinib responders were LODs, whereas most of the AZD1152 responders were EODs. As many EOD cell lines are c-MYC high, we noted that response to AURKBi is strongly associated with c-MYC levels (Fig. 7C and D). However, the expression of c-MYC was not altered after AZD1152 treatment (Supplementary Fig. S8C).

Figure 7.

EOD and LOD osteosarcomas have differential drug response. A, IC₅₀ curves showing the response to trametinib in vitro (n = 16). The cell lines are colored based on the subtype (EOD/LOD). Drug response was measured after 7 days of treatment. Cell lines are listed in the order they appear, highest to lowest, at the highest concentration. B, Differential drug response to trametinib (n = 15). Mean with SD with a Mann–Whitney test P value (**, 0.0022). C, IC₅₀ curves showing the response to AZD1152 in vitro (n = 16). The cell lines are colored based on the c-MYC protein levels (c-MYC high/c-MYC low). D, Relationship between the c-MYC protein level and the IC₅₀ of the response to AURKBi (AZD1152). E, Cell injection and treatment schedule of EOD/LOD tumors. F, EOD tumors respond to AURKBi (AZD1152) in vivo in a subcutaneous model. Eight vehicle tumors, nine AZD1152-treated tumors, and nine trametinib-treated tumors. Tumor volume (mm³) measured over time (left). Waterfall plot showing the decrease in tumor volume (center). Analysis of end point (right). Mean with SD with a Mann–Whitney test P value (***, 0.0001. G, LOD tumors respond to trametinib. Ten vehicle tumors, 12 AZD1152-treated tumors, and 12 trametinib-treated tumors. Tumor volume (mm³) measured over time (left). Waterfall plot showing the change in tumor volume (center). Analysis of end point (right). Mean with SD with Wilcoxon test P values = (**, 0.0025; *, 0.0184). H, Strategy to evaluate drug combination in the PDXs. I, PDX (PDX833) shows a better response to combination therapy targeting both epigenetic cell states (AZD1152 and trametinib). Eight vehicle tumors, nine AZD1152-treated tumors, eight trametinib-treated tumors, and eight tumors treated with AZD1152 and trametinib. Tumor volume (mm³) measured over time (left). Waterfall plot showing change in tumor volume (center). End point analysis comparing combination therapy with monotherapy (right). Mean with SD with Mann–Whitney test P values (***, 0.0007 & 0.0002; *, 0.0206). J, PDX (PDX774) shows a better response to combination therapy over time. Mean with SD with Mann–Whitney test P values (**, 0.0047). K, PDX (PDX526) also shows a better response to combination therapy over time. Mean with SD with Mann–Whitney test P values (**, 0.0043). L, PDX (PDX107) also shows a better response to combination therapy over time. Mean with SD with Mann–Whitney test P values (*, 0.0182). See also Supplementary Fig. S8.

The specific response of LOD cells to a MEKi is consistent with the observed upregulation of AP-1 TFs as these are regulated by the ERK/MAPK pathway (Supplementary Fig. S8D; refs. 50, 51). To explore the mechanism of response to MEK inhibition in LOD, we evaluated downstream targets of PI3K/AKT/mTOR and MEK/ERK pathways (29, 52) by immunoblotting. When the cells were treated with trametinib, p-ERK and FOSL1 decreased, whereas phospho-AKT increased (Supplementary Fig. S8E). These findings are consistent with the earlier observation that FOSL1 is an LOD dependency.

Differential drug responses were further validated in a subcutaneous model in vivo (Fig. 7E). EOD tumors responded to AURKBi (AZD1152) but showed little response to trametinib (Fig. 7F). As histone H3 phosphorylation is important for chromosome condensation and is controlled by AURKB (53, 54), we evaluated the protein expression of phospho-histone 3 (p-H3) after treatment by immunoblotting. p-H3 decreased after treatment with AZD1152 (Supplementary Fig. S8F). As p-ERK protein basal level is low in the vehicle in these subcutaneous tumors, we did not observe a decrease after treatment with trametinib. However, when the AZD1152 treatment was stopped and the tumors were regrowing, p-ERK protein expression was increased (Supplementary Fig. S8F). In contrast, LOD tumors responded to trametinib but not to AZD1152 (Fig. 7G). The level of p-ERK decreased after the trametinib treatment as expected (Supplementary Fig. S8G). A decrease in FOSL1 was not observed in trametinib-treated tumors (Supplementary Fig. S8G). These results indicate that cell states in osteosarcoma are correlated with distinct therapeutic responses.

Chromatin accessibility can change due to cell plasticity, suggesting that EOD and LOD are dynamic cell states and that transition between states can influence therapeutic resistance. Therefore, we evaluated whether targeting both cell states in the same tumor would improve therapy response (Fig. 7H). PDX833 is a xenograft classified as EOD by bulk ATAC-seq. As expected, the corresponding PDX-CL (OS833) responded better to AURKBi monotherapy than to trametinib monotherapy in vitro (Fig. 7A–C). However, in vivo combination therapy led to a much stronger response with tumor regression, which was not seen with either agent alone (Fig. 7I). In addition, we evaluated four more PDXs to validate the effect of the combination therapy in vivo. PDX774 (EOD) responded to AURKBi as predicted by the PDX-CL (Fig. 7J; Supplementary Fig. S8H). Importantly, the in vivo response to combination therapy was significantly stronger. The other PDX that was tested was PDX526, which shows higher heterogeneity with regard to TFs; EOD TFs (SATB2 and SP7) were highly enriched, as well as LOD TFs (FOSL1/2 and JUN; Fig. 2E). PDX526 also shows a stronger response to the combination than single-drug treatment (Fig. 7K; Supplementary Fig. S8I). The response to the combination was stronger in two more PDXs, including one LOD PDX (PDX107) and an EOD PDX (PDX566; Fig. 7L; Supplementary Fig. S8J–S8L). We observed preliminary evidence in which after the AZD1152 treatment (which targets EOD), there is an increase in the LOD markers, such as FOSL1 and p-ERK, suggesting that the cells have the capacity to go back and forth to the other cell state to survive (Supplementary Fig. S8M). This result highlights the heterogeneity in the xenografts and suggests that both epigenetic cell states coexist in the same tumor. Furthermore, it also suggests that combining LOD-specific and EOD-specific targeted therapies may improve efficacy and potentially prevent drug resistance.

Discussion

Mechanisms of cancer cell plasticity are well described for many carcinomas (55). However, for sarcomas, it is less clear what role cell plasticity plays in driving tumor progression and response to therapy. Osteosarcoma is widely known to be a disease characterized by a complex genome that is highly heterogeneous. We hypothesized that, despite this complexity at the level of DNA, chromatin accessibility might provide additional insights into developmental origins and mechanisms of tumor evolution and plasticity. By evaluating chromatin accessibility using ATAC-seq, we discovered two subtypes of osteosarcoma with distinct transcriptional drivers and cell signaling. Crucial to this discovery was the availability of a number of recently developed PDX and PDX-CL models.

We identified two cell states that exhibit clear similarities with either early or late stages of bone development (28). Given these similarities, we labeled these two cell states as EOD or LOD. The accessible chromatin of EOD tumors has a markedly higher enrichment for motifs corresponding to TFs involved in early stages of osteoblast development, including RUNX2, SP7, and DLX2/5/6 (56–59). SP7/osterix is a key TF of specifying the osteoblast cell fate and is recruited to osteoblast enhancers by DLX members (60). RUNX2 is essential for establishing open chromatin in osteoblasts (61). In addition, it promotes bone metastasis in several adult cancers, including prostate (62) and breast cancers (63).

In contrast, the accessible chromatin of LOD tumors was enriched for motifs corresponding to TFs belonging to the AP-1 family. The TFs with binding motifs enriched in the accessible chromatin of LOD, FOSL1/2, JUN, and ATF4 are involved in the late stages of development in the osteoblastic lineage (26, 29). Corresponding to these changes in motif enrichment between the two subtypes, we also found coexpression of TFs and corresponding elevation of protein levels. In addition, several of these TFs were subtype-specific vulnerabilities in DepMap. These results are somewhat surprising given the dramatic heterogeneity observed in whole-genome sequencing (WGS) and suggest that despite this heterogeneity, osteosarcomas (and perhaps by extension, other cancers with complex genomes) only have a defined number of ways to maintain the underlying genomic architecture required for oncogenesis and that these mechanisms are more readily identifiable by epigenetic analysis compared with direct assessment of DNA sequence.

Notably, the two epigenetically defined subtypes we initially discovered by ATAC-seq are also readily identifiable by analysis of H3K27ac binding to chromatin. More specifically, these two subtypes also demonstrate similar patterns of enrichment of H3K27ac at SEs. By overlapping the H3K27ac data and ATAC-seq data at these SEs, we were then able to more precisely define the CRCs driving EOD and LOD. Functional studies further supported the coregulated nature of these CRCs, especially for EOD. For EOD, our analysis strongly supports that SATB2, SP7, RUNX2, and ZNF148 are tightly and coordinately controlling a large part of the transcriptional program of this subtype. To our knowledge, a role for ZNF148 in osteosarcoma has not been previously described. Importantly, we also demonstrate that loss of RUNX2 alone is sufficient to drive the chromatin accessibility of osteosarcoma toward LOD. Whether this is a unique capability of RUNX2 and whether EOD tumors are uniquely dependent on RUNX2 will require further investigation. LOD seemed to have less tight regulation of the corresponding CRCs.

The identification of the two epigenetic subtypes described above relied mostly on the analysis of PDX and PDX-CL models. To begin to define whether these subtypes are present in primary tumors, we used RNA-seq analysis combined with ATAC-seq to define a list of genes mostly highly associated with either the LOD or EOD subtypes. We then used these gene signatures to evaluate primary tumor data for which clinical outcomes are available. The epigenetic-derived gene signatures stratified patients with osteosarcoma into three subtypes: EOD, LOD, and intermediate (i.e., “blend”) subtypes. This result first hinted at the possibility that there is significant heterogeneity within some tumors in regard to the presence of EOD and LOD cells. Unexpectedly, the gene expression signature analysis suggested a prognostic value to these signatures in human samples. Further work will need to be done to establish how this classification compares with others which have been recently described for osteosarcoma (7, 64–69). In addition, we also observed—in the limited number of cases for which more than one sample was available from a given patient—that some tumors were predominantly EOD, whereas others classified as LOD. This again supports the idea that these subtypes are pliable and that tumors can potentially transition from one to another, a process that would likely have important implications for cancer therapy.

The studies above all relied on bulk analysis of tumor samples. However, co-immunofluorescence staining (Fig. 5E and F) had already suggested that a single tumor contained cells with markers of both subtypes. To more directly assess whether EOD and LOD cells can be present within the same tumor, we used single-cell multiomic analysis to simultaneously assess chromatin accessibility and transcriptional profiles of single cells. Notably, this required significant technical refinement of existing workflows as osteosarcoma cells have delicate nuclei that are easily disrupted. Nevertheless, we successfully obtained high-quality data from five PDX models. As these were derived directly from patient tumors (“never touched plastic”), these are a more reliable reflection of the heterogeneity of a human tumor compared with the derived cell lines. Notably, of the five PDX models assessed, three had both LOD and EOD cells within them. The other two PDX were composed almost entirely of EOD cells. Importantly, RNA velocity analysis suggested that cells can interconvert from LOD to EOD and from EOD to LOD. Formally testing that possibility will require additional cell lineage-tracing approaches. This would be consistent with data seen in neuroblastoma in which overexpression of PRRX1 caused the transition from an adrenergic state to a mesenchymal state (70).

Another clinically relevant outcome of our work is the demonstration that different osteosarcoma subtypes respond to different targeted therapies in cell-based assays and in vivo. The EOD subtype responded to AURKBi, and LOD cells responded to MEKi. Intriguingly, both subtypes responded better to the drug combination, which suggests the presence of both cell types in this tumor or the ability of cells to transition from one to another as a therapy-escape mechanism. Although our data suggest that the observed transcriptional changes following treatment may reflect a cell state transition, we acknowledge alternative explanations. Specifically, these could include (i) clonal selection of a preexisting resistant subpopulation, (ii) a transient stress response induced by the inhibitors, or (iii) true plasticity leading to a stable reprogramming of cell identity. Further experiments will help to determine whether combination therapy targeting dependencies of both subtypes simultaneously may represent a valuable strategy for osteosarcoma treatment. Nevertheless, our work indicates that targeted therapies that use a combination approach to target subtypes of osteosarcoma may be needed.

Taken together, we demonstrate that knowledge of epigenetic states can be used to identify both prognostic expression signatures and specific druggable vulnerabilities in osteosarcoma. These studies also identified the underlying molecular circuitries that are required for these epigenetic states, suggesting additional strategies for the development of osteosarcoma-specific therapies. Beyond osteosarcoma, our work indicates that similar types of analyses, focused on epigenomic changes rather than genomic ones, may enable stratification and discovery of targetable vulnerabilities in other genetically complex cancers.

Methods

Cell Culture

All cell lines were routinely tested for Mycoplasma and short tandem repeat tested to confirm identity. Cells were cultured in standard DMEM (Gibco, #11965-092) supplemented with 10% bovine growth serum (HyClone, #SH30541.03) and 1% penicillin–streptomycin–glutamine (Gibco, #10378-016). PDX-CLs and established commercially available cell lines are shown in Supplementary Table S1.

Generation of PDXs and PDX-CLs

For PDX generation, 1-mm³ pieces of fresh or frozen-thawed patient tumor were dipped in Matrigel (Corning, #356237) and placed under the renal capsule of NOD/SCID gamma (NSG) mice for PDX formation. Abdominal palpation was used to detect tumor progression in mice for up to a year after implantation. Tumors that successfully engrafted were allowed to grow to 1 to 2 cm³ before the mouse was killed, and the tumor was removed. Tumors were digested with collagenase (Sigma-Aldrich, #C6885) and filtered using a 70-μm filter. Cells from single-cell suspension were implanted subcutaneously in the flank of NSG mice (5 × 10⁵ cells) in 30 mL of minimum essential medium alpha (Gibco, #12561-056) and 20 mL Matrigel for PDX passaging. PDX-CLs were generated as described previously (4).

Patient Samples

Written informed consent was obtained from each patient (or from a parent or legal guardian for those under 18 years of age) in accordance with recognized ethical guidelines (Belmont Report) and with institutional review board approval at each participating center. All analyzed samples were reviewed by a pathologist at the time of diagnosis and confirmed to be osteosarcoma.

ATAC-seq

ATAC-seq libraries were prepared as described previously with minor modifications (71). Briefly, 1 × 10⁵ cells were lysed on ice to prepare nuclei. Nuclei were incubated with the Tn5 transposase (Illumina, #15027865) at 37°C for 30 minutes in a thermomixer shaking at 1,000 rpm. PCR was performed to amplify the libraries for six to 11 cycles, and the libraries were purified with SPRI beads for size selection (Beckman Coulter, #B23318). The libraries were sequenced at Center for Advanced Technology facility at University of California San Francisco (UCSF) using the HiSeq 4000 instrument and 100-bp paired-end sequencing. When fresh/frozen PDXs or patient samples were used as input, the ATAC-seq libraries were prepared as described previously (dx.doi.org/10.17504/protocols.io.6t8herw). Libraries were quantified using the High Sensitivity D5000 ScreenTape system (Agilent, #5067-5592) using the High Sensitivity D5000 reagents (Agilent, #5067-5593) on the TapeStation 4200 (Agilent, RRID: SCR_018435).

Bulk RNA-seq

RNA was extracted using the RNeasy Mini Kit (Qiagen, #74106) and on-column DNA digestion with DNase I (Qiagen, #79256) according to the manufacturer’s protocol. RNA-seq libraries were made using the TruSeq Stranded mRNA Kit (Illumina, #RS-122-2101) in accordance with the manufacturer’s instructions. All manufacturer controls were used in preparation of the libraries. Libraries were quantified and quality-checked using the High Sensitivity D1000 ScreenTape (Agilent, #5067-5584) on the TapeStation 4200 (Agilent, RRID: SCR_018435), the High Sensitivity DNA (Agilent, #5067-4626) on the Bioanalyzer 2100 (Agilent, RRID: SCR_018043), or the High Sensitivity NGS (Agilent, #DNF-474-22) on the Fragment Analyzer. Sequencing was performed on an Illumina HiSeq System at the Stanford Functional Genomics (RRID: SCR_002050) and on an Illumina HiSeq or NovaSeq 6000 system at the Center for Advanced Technology at UCSF. Chemistry for sequencing includes 2 × 75 bp, 2 × 100 bp, and 2 × 150 bp for the respective samples.

Western Blotting

Protein extraction from cell lines was performed using RIPA buffer (Sigma-Aldrich, #R0278), protease inhibitor (Roche, #11697498001), phosSTOP (Roche, #4906845001), and phenylmethylsulfonyl fluoride (Sigma-Aldrich, #93482) on ice incubation. Xenograft tumor fragments were snap-frozen and kept at −80°C until use. Protein extraction was performed by grinding the tumor in a liquid nitrogen–cooled mini mortar and pestle set and lysing the tumor with the same buffer as previously described. Protein quantification was performed using the Pierce Bicinchoninic Acid Protein Assay Kit (Thermo Fisher Scientific, #23227) on 4% to 15% Criterion TGX gradient gels (Bio-Rad, #5671084), and proteins were then transferred to polyvinylidene difluoride membranes. Western blot membranes were incubated in 5% milk (MP Biomedicals, #902887) with TBS with Tween (Bioland Scientific, #TBST01-10) or 5% BSA (Sigma-Aldrich, #A4503) with TBST for primary and secondary antibodies. Clarity Western ECL Substrate (Bio-Rad, #170-5061) was used for chemiluminescence. Imaging was performed in the Odyssey M LI-COR system (LI-COR, RRID: SCR_025709).

Isolation of Nuclear, Cytosolic, and Chromatin Fractions

To isolate nuclear and cytoplasmic fractions, 1 × 10⁷ cells were harvested and washed twice with PBS (Gibco, #10010-023). The pellet was resuspended in 200 μL buffer A [10 mmol/L HEPES (Corning, #25-060-Cl) pH 7.9, 10 mmol/L KCl (Sigma-Aldrich, #P9541), 1.5 mmol/L MgCl₂ (BioWorld, #41320004), 0.34 mol/L sucrose (Fisher BioReagents, #BP220-1), 10% glycerol (Fisher Chemical, #G33-1), 1 mmol/L DTT (Sigma-Aldrich, #646563), and 0.1% Triton X-100 (Sigma-Aldrich, #T8787)] with protease inhibitor (Roche, #11697498001) and phosSTOP (Roche, #4906845001) and incubated for 5 minutes on ice. Centrifugation was performed at 13,000 × g for 5 minutes at 4°C. Supernatant (S1) was separated from the pellet (P1), and S1 was clarified by high-speed centrifugation at 20,000 × g for 5 minutes at 4°C to collect the second supernatant (S2), which contained the cytosolic fraction, and P2 was discarded. P1 was washed once with buffer A and twice with PBS before lysis for 30 minutes in 100 μL buffer B [3 mmol/L EDTA (Research Products International, #E58100), 0.2 mmol/L EGTA (Sigma-Aldrich, #324626), and 1 mmol/L DTT] with protease inhibitor and phosSTOP. P1 in buffer B was centrifuged at 1,700 × g for 5 minutes at 4°C to obtain the nuclear fraction from the resulting supernatant (S3; ref. 72). To isolate chromatin, P3 from the previous centrifugation was washed once with buffer B and resuspended in 150 μL of E3 buffer [500 mmol/L Tris-HCl (Teknova, #T1075), 500 mmol/L NaCl (Sigma-Aldrich, #S3014), and protease inhibitor; ref. 73]. Then, the samples were sonicated at 25% amplitude with 1 minute of effective time and centrifuged for 5 minutes at 20,000 × g to obtain the chromatin fraction.

Determination of IC₅₀In Vitro

Cells were seeded 24 hours prior to drug treatment at 2,500 cells/well in 96-well plates. Drugs were diluted in DMSO (Sigma-Aldrich, #D8418) at 10 µmol/L. Drugs were added at eight concentrations in one-third dilutions across the plate with a top dose of 10 μmol/L. Barasertib-HQPA (AZD2811 or AZD1152, #HY-10126) and trametinib (#HY-10999) were purchased from MedChemExpress. Cell viability was evaluated after 7 days using CellTiter-Glo 2.0 (Promega, #G9243). IC₅₀ values were calculated using Prism 10 software.

Osteosarcoma commercial cell lines hFOB 1.19 (CRL-11372), IMR-90 (CCL-186), 143B (CRL-8303), MG-63 (CRL-1427), SJSA-1 (CRL-2098), Saos2 (HTB-85), and U-2 OS (HTB-96) were purchased from ATCC. Drugs were purchased from Selleckchem or MedChemExpress. Cells were plated in 96G-well plates (Greiner Bio-One, 655098) and allowed to attach overnight. For the single-agent drug screen, drugs were added at 10 concentrations in one-third dilutions across the plate with a top dose of 10 μmol/L using a BioTek fluid handler (BioTek EL406). Cells were incubated with the drug for 7 days, and viability was assessed using CellTiter-Glo 2.0 (Promega, G9243) and measured on EnVision XCite (PerkinElmer) or BioTek (Synergy Neo2). Drug response curves and IC₅₀ determination were done using GraphPad Prism v9.5.1 (RRIS: SCR_002798).

Drug Sensitivity Analysis

We log-transformed IC₅₀ values and compared drug difference between EOD versus LOD cell types across 40 drugs. Initial two-way ANOVA (IC₅₀ ∼ drugs × ClusterByATAC) revealed significant heteroscedasticity (Bartlett test P < 0.001), so a mixed-modeling approach was utilized instead. We fitted a linear mixed-effects model with random intercepts and slopes: IC₅₀ ∼ ClusterByATAC + (ClusterByATAC|Drugs), allowing drug-specific variance in cluster effects. Drug-level pairwise t tests were performed between clusters, with P values adjusted using the Benjamini–Hochberg FDR method. Drugs were considered differentially sensitive if raw P values < 0.05 and FDR < 0.2 threshold were met.

In Vivo Drug Treatment

All animal studies were performed in accordance and with UCSF Institutional Animal Care and Use Committee (work approved under protocol AN197629). Animals were maintained in a barrier facility at UCSF. NSG mice were used for all experiments (NOD.Cg-Prkdc^scidIl2rg^wm1Wjl/SzJ, strain #005557, The Jackson Laboratory, RRID: IMSR_JAX:005557). PDX-CLs (1–2 × 10⁶ cells) were injected subcutaneously. Tumor volume was determined using the formula tumor volume (mm³) = (length × width²)/2. Tumor growth was measured using digital calipers; when the average tumor volume of the cohort reached 100 mm³, the mice were randomized into the different arms. Barasertib-HQPA (MedChemExpress, AZD2811 or AZD1152, #HY-10126) was dosed at 25 mg/kg 4 days per week and trametinib (MedChemExpress, #HY-10999) was dosed at 1 mg/kg 5 days per week. Tumor volume was measured twice per week.

IHC and Immunofluorescence

IHC was performed on formalin-fixed, paraffin-embedded PDX or patient TMA sections as previously described (15), with minor modifications. Tissue was blocked with 5% goat serum diluted in TBST (Bioland Scientific, TBST01-10). Sections were incubated overnight at 4°C in a humidified chamber with anti–c-Jun primary antibody (Cell Signaling Technology, #9165, RRID: AB_2130165) or SP7 (Abcam, ab209484, RRID: AB_2892207) diluted 1:100 in blocking solution. The following day, slides were washed and incubated with biotinylated goat anti-rabbit secondary antibody (Thermo Fisher Scientific, #A-11037, RRID: AB_2534095) diluted 1:1,000 for 30 minutes at room temperature. Slides were counterstained with Harris hematoxylin (Sigma-Aldrich, #HHS32), dehydrated, cleared in xylene, and mounted using Acrytol (Leica Biosystems, #3801720). Images were acquired using a Leica DMi8 Inverted Microscope (RRID: SCR_026672).

For immunofluorescence, tissue sections were deparaffinized and rehydrated by sequential immersion in xylene (Sigma-Aldrich, #247642) and decreasing concentrations of ethanol (Koptec, #V1101), followed by PBS (Gibco, #10010-023). Antigen retrieval was performed by boiling slides in sodium citrate buffer [sodium citrate salt (Fisher Scientific Education, #S25545) and 0.05% Tween-20 (Thermo Fisher Scientific, #BP337-500), pH adjusted with HCl (Sigma-Aldrich, #H1758)]. Sections were blocked in 5% donkey serum (diluted in TBST) for 1 hour at room temperature. Primary antibodies were diluted 1:100 in blocking buffer and applied overnight at 4°C in a humidified chamber: anti–c-Jun (rabbit mAb, Cell Signaling Technology, #9165, RRID: AB_2130165) and anti-SP7/osterix (mouse mAb, R&D Systems, #MAB7547, RRID: AB_3658831). Secondary antibodies [goat anti-rabbit Alexa Fluor 594 (Thermo Fisher Scientific, #A11037, RRID: AB_2534095) and goat anti-mouse Alexa Fluor 488 (Thermo Fisher Scientific, #A11029, RRID: AB_2534088)] were diluted 1:200 in blocking buffer and incubated for 1 hour at room temperature in the dark. Nuclei were stained with Hoechst 33342 (Thermo Fisher Scientific, #H3570) diluted 1:2,000 in PBS. Coverslips were mounted using ProLong Glass Antifade Mountant (Thermo Fisher Scientific, #P36980) and cured at room temperature for 24 hours before sealing. Imaging was performed using a fluorescence microscope. Fluorescence microscopy images were analyzed using Fiji. To assess co-expression, binary masks were generated for SP7 and c-Jun signals by thresholding individual channels. Each mask was then overlapped with the DAPI mask to identify SP7+ and c-Jun+ nuclei. The overlap of SP7 and c-Jun with DAPI was quantified to determine the proportion of double-positive cells relative to the total number of DAPI ⁺ cells.

TF Chromatin Binding

CUT&RUN was performed using the CUT&RUN Assay Kit (Cell Signaling Technology, #86652) according to the manufacturer’s protocol. Per antibody, 1.0 × 10⁵ cells were used. Briefly, cells were harvested, washed, and bound to activated Concanavalin A–coated magnetic beads (Cell Signaling Technology, #93569S) and permeabilized. The bead–cell complex was incubated overnight with either an acetyl-histone H3 (Lys27) antibody (Cell Signaling Technology, #8173, RRID: AB_10949503), mono-methyl-histone H3 (Lys4) antibody (Cell Signaling Technology, #5326, RRID: AB_10695148), tri-methyl-histone H3 (Lys4) antibody (Cell Signaling Technology, #9751, RRID: AB_2616028), or an IgG antibody control (Cell Signaling Technology, #66362, RRID: AB_2924329) at 4°C. Cells were then washed with digitonin buffer (Cell Signaling Technology, #16359S), resuspended in 50 μL of pAG/MNase (Cell Signaling Technology, #57813S), and incubated for 1 hour at 4°C. Spike-in DNA (5 ng, Cell Signaling Technology, #36598S) was added to each reaction. DNA from enriched chromatin samples was purified using the DNA Purification Buffers and Spin Columns kit (Cell Signaling Technology, #14209). DNA libraries were prepared using the NEBNext Ultra II DNA Library Prep Kit for Illumina (New England Biolabs, #E7645L) and NEBNext Multiplex Oligos for Illumina (New England Biolabs, #E7600S). Fragment size and quality of the libraries were evaluated by TapeStation (4200 system, Agilent, RRID: SCR_018435). Libraries were sequenced on a NextSeq 500 system (Illumina).

Knockdown Using siRNA

ON-TARGETplus Human siRNAs were ordered from Horizon Discovery Biosciences. Cells were transfected using RNAiMAX Transfection Reagent (Thermo Fisher Scientific, #13778150), targeting identified TF members of CRCs in a six-well format according to the manufacturer’s protocol. A nontargeted sequence was used as a control.

Overexpression of TFs in an Inducible System

pDONR plasmids were used to clone RUNX2 (clone ID: HsCD00829534) and FOSL1 (clone ID: HsCD00001664) open reading frames in the receptor plasmid pCW57.1 (RRID: Addgene_41393). Gateway LR Clonase II Enzyme mix (Invitrogen, 11791-020) was used to prepare the LR reaction according to the manufacturer’s protocol. LentiX HEK were transfected with the packaging plasmids, the target vectors (pCW57.1::RUNX2 or pCW57.1::FOSL1), and TransIT-293 transfection reagent (Mirus Bio 2705). Cells were transduced and selected with antibiotics. Cells were evaluated after transduction and were induced with doxycycline.

qRT-PCR

RNA was extracted after 48 hours of lipofection using the RNeasy Mini Kit (Qiagen, #74106) and on-column DNA digestion with DNase I (Qiagen, #79256) according to the manufacturer’s protocol. cDNA was synthesized from 1 μg RNA with the Maxima First Strand cDNA Synthesis Kit (Thermo Fisher Scientific, #K1642). Gene expression was evaluated by qPCR using PerfeCTa SYBR Green FastMix ROX (Roche, #95073-05 K), with the primers listed in Supplementary Table S5. The data were analyzed using the ΔΔCt method and plotted as a percentage of transcripts compared with the control group. All experiments were completed in triplicate.

Single-Cell Multiome

We used protocol CG000365 from 10x Genomics for the PDX-CLs. Cells were harvested using accutase (Innovative Cell Technologies Inc., #AT104-500). We optimized each PDX-CL by modifying the lysis buffer composition and the incubation times. The quality of nuclei was evaluated by microscopy using cell lines expressing GFP in the nucleus in the optimization process. A total of 500,000 cells were harvested when ∼70 to 80% confluent, followed by the nuclei isolation protocol. OS384 was incubated for 5 minutes in lysis buffer [(NP40)_f = 0.05%] and OS833 was incubated for 3 minutes in lysis buffer [(NP40)_f = 0.045%]. We used protocol CG000505 from 10x Genomics for nuclei isolation from tissue for the PDXs. We used the Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Reagent Bundle (#PN-1000285) as instructed by the manufacturer. Single-cell ATAC-seq was sequenced in NovaSeq X 10B 100 (51 × 12 × 24 × 51, Illumina, RRID: SCR_024569) and single-cell RNA-seq was sequenced in NovaSeq X 10B 100 (28 × 10 × 10 × 90).

Quantification and Statistical Analysis

ATAC-seq

Reads were trimmed using NGSUtils (RRID: SCR_001236; ref. 74). Trimmed reads were aligned to the GRCh38.p5 assembly of the human genome using Bowtie2 (RRID: SCR_016368; ref. 75) with the “–very sensitive” parameter. Duplicates and mitochondrial reads were removed using NGSUtils (74), and reads with high mapping quality (>q30) were retained using Samtools (76). Peak calling was performed with MACS2 (RRID: SCR_013291; ref. 77). All default options were used, with the following flags explicitly set: “-nomodel –keep-dup all.” Peaks overlapping blacklisted features as defined by the ENCODE project (RRID: SCR_006793; ref. 78) were discarded. Only peaks with FDR q ≤ 0.01 were retained. Differential analysis was performed using DESeq2 (RRID: SCR_015687; ref. 79). Limma “-removeBatchEffect” was used to correct for the source of origin when using patient samples, PDXs, and cell lines (80).

TF Analysis

We used the DAR between epigenetic subtypes (EOD/LOD) to identify the TFs. Enriched TFs per subtype were identified using chromVAR (RRID: SCR_026570; ref. 24). The differential TF footprints were identified by a de novo method HINT-ATAC, which employs hidden Markov model and corrects for strand-specific Tn5 cleavage bias (25).

Gene Ontology (ATAC-seq Data and SE Signature)

We assigned the DARs to genes using the basal plus extension method (5 kb upstream, 1 kb downstream, and 500 kb max extension; ref. 16). Enrichments for Gene Ontology (GO) biological process and GO molecular functions were calculated within GREAT (RRID: SCR_005807) using default settings with whole genome as background regions.

Subtype-Specific Gene Signatures

Differential chromatin accessibility analysis was integrated with DGE data (RNA-seq) using the cell line panel (Fig. 5A; n = 19): (i) Genes with DARs ((ATAC-seq genes): DARs were associated with genes using the GREAT algorithm (RRID: SCR_005807; ref. 16). The association from peaks to genes is a two-step process. First, every gene is assigned a regulatory domain. Then, each genomic region is associated with all genes whose regulatory domain it overlaps with (16). Shared genes between subtypes were removed. (ii) Genes with DGE (RNA-seq genes): Each sample was assigned to one subtype based on the cell state (Fig. 1B; Supplementary Fig. S5B), and DGE was performed. (iii) Gene signature: ATAC-seq genes were overlapped with RNA-seq genes to obtain the genes with differentially open chromatin and differentially expressed genes. This gene signature contains 237 genes (Supplementary Table S6).

RNA-seq Analysis

Raw FASTQ files were obtained directly from the sequencing facility, and the Trimmomatic tool (v0.36, RRID: SCR_011848; ref. 81) was used to trim adapters and low-quality regions. The processed reads were then aligned to the human reference genome (hg38) using the STAR aligner (v2.5.1b, RRID: SCR_004463; ref. 82). To quantify gene-level expression, we employed the “quantMode” feature of the STAR aligner with gene annotations from GENCODE (RRID: SCR_014966) p5. Quality control (QC) metrics, such as counts of uniquely aligned reads, ratios of unique-to-multiple alignments, and exon-to-intron ratios, were assessed using NGSUtilsJ (v0.3-2180ca6, RRID: SCR_001236; ref. 74). Further QC evaluations and subsequent analyses were conducted using R (v3.5.3, RRID: SCR_001905). Aligned reads were normalized by the trimmed mean of M values method from the EdgeR package (v3.24.3, RRID: SCR_012802; ref. 83), and counts were transformed to log₂ (CPM + 1), where CPM stands for counts per million. To detect statistically significant changes in gene expression, we used the voom function from the limma package to estimate the mean–variance relationship and provide precision weights. These were incorporated into the empirical Bayes linear modeling framework of limma (v3.38.3, RRID: SCR_010943; ref. 80) to compute statistical outputs, including P values, adjusted P values, and log FC.

Correlation Matrix of Gene Expression of TFs

Gene expression from RNA-seq was analyzed for the complete panel of cell lines (n = 19). Pearson correlation plot was generated using the R “cor()” function with default settings.

CUT&RUN Preprocessing and Normalization

CUT&RUN samples were processed using a custom pipeline scripted with CGPipe (RRID: SCR_027465). FASTQ files were paired and trimmed using Trim Galore (RRID: SCR_011847) to eliminate adapters and filter low-quality reads (quality scores <30) and then aligned to both Homo sapiens (hg38) and Saccharomyces (sac3) genomes (84) using Bowtie2 (RRID: SCR_016368; ref. 75), with settings to permit “dovetail” read alignments (–dovetail). Reads were determined to originate from human or yeast spike DNA using the NGSUtilsJ (RRID: SCR_001236) bam-best function (74), and the ratio of yeast-to-human reads was reported for use when comparing sample signals between cell lines. Yeast-aligned reads were removed (NGSUtilsJ; ref. 74), and the remaining reads were filtered to retain those which were (i) not PCR duplicates (Picard tools’ MarkDuplicates, RRID: SCR_006525), (ii) shorter than 120 bp (Samtools, RRID: SCR_002105; ref. 76), and (iii) did not align to ENCODE (RRID: SCR_006793) hg38 blacklisted regions (bedtools intersect, RRID: SCR_006646; refs. 85, 86). Peak calling was performed using Model-based Analysis of ChIP-Seq version 3 (MACS3; RRID: SCR_013291): BAM files for replicate epitope samples were merged within the MACS3 callpeak function, as were BAM files for IgG background samples, and a q value cutoff of 0.01 was used for narrowPeak calling (77).

To generate co-occupancy heatmaps in Fig. 4H and Supplementary Fig. S4F, the collapsed union of CRC TF peaks (RUNX2, SATB2, SP7, and ZNF148 for OS742; FOSL1, FOSL2, JUN, TEAD1, and RUNX2 for OS526) was generated and resized to 4-kb windows about each region center using GenomicRanges (1.54.1). These windows were then tiled into 50 80 bp bins [also using GenomicRanges (1.54.1)], and reads were counted in each bin using GenomicAlignments (1.38.2) and RSamtools (2.18.0) and then saved as a SummarizedExperiment (1.32.0) object. Read counts per bin were pooled for replicates and converted to reads per million (RPM) for each epitope pictured (reads/combined library size)*1E6, and windows were ordered by the sum of RPM values for the left-most TF (RUNX2 for OS742; FOSL1 for OS526). Heatmaps were generated using ggplot2 (3.5.2), and scales were cut off at 0.5 RPM.

SE Identification and Assignment

SEs were detected using the ROSE algorithm (RRID: SCR_017390; ref. 36) on H3K27ac CUT&RUN data. Sample replicates were pooled and normalized by library size, and narrowPeaks were detected using MACS3 (RRID: SCR_013291) as outlined above. These peaks were provided as the enhancer set to the ROSE algorithm (via argument “-i”), which was then run using default settings with specific arguments “-g HG38” (human genome hg38), “-s 12500” (stitching distance 12.5 kb), and “-t 10000” (no SE annotations in peaks within 10 kb of a TSS although SE scores are still calculated). BAM files for IgG controls were merged into a single file using “samtools merge,” and then provided to ROSE via the “-c” argument.

For differential enrichment analysis among SEs, all identified SEs among all samples were unified and merged into a single SE set, with overlaps expanded to cover the greatest width. Total read depth per SE was then calculated for each sample using the summarizeOverlaps function from the GenomicAlignments R package (RRID: SCR_024236; ref. 87). For each cell line, the average copy number of each peak region was calculated based on raw WGS read depth in 10 kb bins (bin WGS reads/average binned WGS reads genome wide = copy number), and peak copy numbers were calculated per cell line as the sum of the copy numbers of the bins they overlapped multiplied by the fraction of the SE’s area each bin represents. The contribution of copy number to CUT&RUN signal was approximated for every sample and peak combination using linear modeling [specifically the lm() function from the stats R package] and then subtracted. The resulting copy number–normalized counts were rounded up to integer values, and then DESeq2 (RRID: SCR_015687; ref. 79) was used to compare these values between EOD and LOD subtypes with consideration for library size and multiple testing. EOD- and LOD-enriched SEs were annotated as those with significantly higher copy number–normalized read depth in EOD and LOD samples, respectively.

CRCs were explored as suggested by Ott and colleagues (32). In each cell line, SE “elements” were defined as regions within the unified SE set which also correspond to an ATAC-seq peak in the same sample. For each cell line, sequences from these elements were extracted for motif analysis using the MEME Suite (RRID: SCR_001783; ref. 88); specifically, enrichment of motifs between EOD and LOD SE elements was calculated using the Analysis of Motif Enrichment function (Fisher exact test, default settings; ref. 89). Genomic locations of motifs within each element were determined using the Finding Instances of Motif Occurrences (FIMO) algorithm (default settings, including P value < 0.0001; ref. 90) to search for all motifs within the Homo sapiens Comprehensive Model Collection (version 11, RRID: SCR_005409; ref. 91). FIMO hits with scores less than the median (10) were discarded. Active TF promoters were next defined for each cell line based on the presence of at least one H3K27ac CUT&RUN peak and at least one ATAC-seq peak over that promoter (upstream 2 kb and downstream 0 kb), and FIMO was used to locate TF motifs within these active promoters as outlined above. Active TF promoters were next “assigned” SEs in each cell line by the following criteria, in order of precedence: (i) The SE overlaps that TF promoter, (ii) that SE is within 100 kb of that promoter, and (iii) the promoter is the closest active promoter to the SE within 1 MB. Note that in cases 2 and 3, multiple SEs can be assigned to the same gene promoter. For a given TF, its in-degree was next defined as the number of unique TF motifs (including its own) within its assigned SE(s), and its out-degree was defined as the number of unique TFs with SE(s) containing its motif. A network of TFs and their assorted in- and out-degree dependencies/targets was assembled into an adjacency matrix with TFs as nodes and dependencies as edges. CRCs were identified as cliques among these networks (sub-networks containing at least four nodes–i.e., TFs–that all target/are targeted by one another). Cliques were identified using the igraph R package (RRID: SCR_019225), and for each cell line, a clique enrichment score was determined for all TFs based on the fraction of all cliques containing that TF.

Survival Analysis

Survival analysis was conducted using patient stratification based on hierarchical clustering assignments or with a gene signature defined in Fig. 5A and B. The gene signature was derived as described above. Both OS (time to death) and progression-free survival were evaluated. For the survival analysis using patients assigned via hierarchical clustering, a Kaplan–Meier analysis and Cox proportional hazards model were performed using the survival package in R (RRID: SCR_026244). This analysis compared three patient groups: EOD, LOD, and blend. The significance of survival differences was assessed using the log-rank test for univariate analyses and the likelihood ratio test for multivariable Cox regression models. Additionally, for analyses concerning OS, metastasis at diagnosis was included as a covariate in the Cox regression model. Furthermore, a time-to-death analysis was conducted using an independent TARGET dataset to further validate our findings. The TARGET dataset consists of N = 72, whereas patients with osteosarcoma from our OS set consist of N = 38 patient samples. For the OS set, both OS (time to death) and progression-free survival were evaluated, with all events censored at 60 months. TARGET survival data were downloaded from the Xena UCSC website (RRID: SCR_018938) and re-censored to 60 months to match those of the OS set. All counts used in the analysis were in transcripts per kilobase million (TPM). TARGET counts were downloaded from the UCSC RNA-seq Treehouse Childhood Cancer Initiative Public Data repository (https://treehousegenomics.soe.ucsc.edu/public-data/), whereas TPM counts for the ASC set were generated with the UCSC Toil RNA-seq pipeline (https://github.com/BD2KGenomics/toil-rnaseq) to ensure harmonization of the two datasets.

Single-Cell Multiome

Raw sequencing data from the single-cell multiome experiment were processed using the CellRanger ARC pipeline (10x Genomics, version 2.0.0, RRID: SCR_023897) to generate gene expression and chromatin accessibility matrices. The CellRanger ARC count command was used with a pre-built human reference genome (10x Genomics, GRCh38-2020-A-2.0.0, RRID: SCR_023672). The resulting count matrices were then used for downstream analysis in R (4.3.2, RRID: SCR_001905). Both RNA-seq and ATAC-seq data were imported into R and processed using a combined Seurat (v5.0, RRID: SCR_016341; ref. 92) and Signac (1.11, RRID: SCR_021158; ref. 93) framework to jointly analyze both modalities. Signac was used to create additional assay slots such as genomic range with peak accessibility and DNA motifs. For each sample, poor-quality cells were filtered based on the number of genes (>200 or <4000), percentage of reads arising from the mitochondria (<30%), ATAC-seq fragments (>1,000 or <100,000), nucleosome_signal <2, and TSS.enrichment >1. Peaks for each sample were called with MACS2 (RRID: SCR_013291), and normalization was completed with “SCTransform” (RRID: SCR_022146), regressing for percent mitochondria.

After normalization, dimensionality reduction was performed using PCA (RRID: SCR_014676) on the SCT-transformed data, selecting the top 30 principal components. Subsequent steps included applying UMAP (RRID: SCR_018217; ref. 12) to the PCA-reduced data for visualization and identifying cell clusters using the “FindNeighbors” and “FindClusters” functions in Seurat, with a resolution of 1.5. The resulting clusters were visualized on the UMAP plot to delineate distinct cell populations. Batch correction across all samples was completed with Harmony (v1.1, RRID: SCR_022206; ref. 94).

Cell annotation was accomplished using the scGate R package (95), utilizing the rank-based method UCell to evaluate the strength of gene expression signatures in each cell. Specifically, EOD and LOD enrichment scores were calculated utilizing the full signatures based on the clusters defined in Fig. 5, EOD with 126 genes and LOD with 111 genes signatures, rather than individual marker genes. Each signature was evaluated independently and classified by applying optimized positivity thresholds tuned separately for EOD and LOD signatures. To address single-cell RNA-seq sparsity, we applied k-nearest neighbor smoothing with upward-only adjustment (smooth.up.only = TRUE), computing average UCell scores across each cell’s transcriptomic neighborhood. Cells with enrichment scores exceeding the positivity threshold for one signature were classified accordingly. For cells meeting threshold criteria for multiple signatures, the annotation with the highest UCell score was assigned. Cells lacking sufficient enrichment for either signature were designated undetermined.

Velocity Quantification

To maintain consistency, we exported cell identifiers, UMAP coordinates, cluster assignments, and sample IDs from the integrated, quality-controlled Seurat object as CSV files to serve as a reference for velocity analysis. Spliced and unspliced transcript counts were quantified from 10x Genomics CellRanger possorted BAM files and GTF annotations using velocyto v0.17.17 (96). This generated sample-specific loom files which were imported for RNA velocity analysis using scVelo v0.3.3 (44). Cell barcodes from velocyto output were reformatted by extracting the barcode sequence from the full identifier string, prepending the corresponding sample name (e.g., “PDX_OS107_”), and standardizing the “-1” suffix to ensure exact matching with the cell identifiers in the merged Seurat object. After barcode reformatting, individual samples were concatenated.

This then underwent standard scVelo preprocessing: genes detected in fewer than 20 cells were removed, counts were normalized per cell with enforcement across both spliced and unspliced layers, and the top 2,000 highly variable genes were selected using dispersion-based filtering. Data were log-transformed, and a k-nearest neighbor graph was computed (k = 30; n_pcs = 30). First- and second-order moments were calculated for velocity estimation.

RNA velocity and latent time were inferred using the dynamical model. Velocity vectors were projected onto the UMAP embedding space imported from Seurat, and cell-to-cell transition probabilities were calculated to construct the velocity graph. Velocity vectors and latent time values were imported back into R. Cell state transitions were visualized by overlaying velocity vectors onto UMAP coordinates, with arrows indicating predicted future states. For improving visualization only, velocity vector map projections were downsampled to 1,500. Latent time distributions were compared between cell states using Kruskal–Wallis tests, followed by pairwise Wilcoxon rank-sum tests with Bonferroni correction.

To assess state transition between cells, we calculated the Euclidean distance from the UMAP position of each cell to the centroid of the opposite cluster and then compared this to the projected distance after applying velocity vectors scaled by a factor of 10. Cells were classified as transitioning if their projected position moved closer to the opposite cluster centroid. Transition probabilities were calculated as the proportion of cells within each cluster exhibiting transition-directed velocities.

To characterize molecular changes for each state, gene module scores were constructed from core TFs of both EOD and LOD. EOD markers (RUNX2, SP7, MEF2C, SATB2, and ZNF148) and LOD markers (FOSL1, FOSL2, JUN, ATF4, and NFATC2) were scored using Seurat’s AddModuleScore function. Chromatin accessibility dynamics were assessed in two ways: (i) TF-binding motifs from JASPAR 2020 were scored using chromVAR, with mean z-scores calculated across motifs matching early and late marker genes and (ii) ATAC-seq peak accessibility was quantified at genomic loci within 100 kb of marker genes aggregated by summation. Relationships between latent time and molecular features were visualized using LOESS regression with 95% CIs, stratified by cell state.

To identify differentially accessible peaks, the “FindMarkers” function with the “logistic regression” was employed for statistical testing, and the total number of peaks per cell (nCount_peaks) was included as a latent variable to control for differences in sequencing depth. The “ClosestFeature” function was used to associate the genomic peaks with the nearest genes, and any peaks not within 10,000 base pairs of a gene were removed to ensure relevance to gene regulation. For DGE analysis, we also used the “FindMarkers” function but with the Wilcoxon rank-sum test (“wilcox”). The final statistical filter for both the differential accessibility and DGE tests was based on an adjusted P value threshold of 0.05. Peak correlation analysis was conducted to link chromatin accessibility peaks to gene expression profiles. This was performed using the “LinkPeaks” function from the Signac package (93). The analysis utilized expression data transformed via SCTransform, setting a threshold that required a peak to be present in a minimum of 30 cells.

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, E.A. Sweet-Cordero (alejandro.sweet-cordero@ucsf.edu).

Supplementary Material

Supplementary Table S1

Supplementary Table S1. Summary of the samples used in this study related to Figure 1.

Supplementary Table S2

Supplementary Table S2. Differential accessible regions (DAR) related to Figure 1B.

Supplementary Table S3

Supplementary Table S3. TFs per subtype related to Figure 2.

Supplementary Table S4

Supplementary Table S4. SEs per PDX-cell line related to Figure 3.

Supplementary Table S5

Supplementary Table S5. Primers and data from qRT-PCR used to generate networks related to Figure 4E-4F.

Supplementary Table S6

Supplementary Table S6. Gene signature related to Figure 5A.

Supplementary Table S7

Supplementary Table S7. Information about the patients related to Figure 5B and Supplementary Figure S5C.

Supplementary Table S8

Supplementary Table S8. IHC TMA quantification related to Figure 5E-5F.

Supplementary Table S9

Supplementary Table S9. Detailed peak and gene information from PDX- single cell multiome experiment related to Figure 6.

Supplementary Table S10

Supplementary Table S10. Detailed peak and gene information from CLs – single cell multiome experiment related to Supplementary Figure S7.

Supplementary Table S11

Supplementary Table S11. Drug screen tested in the PDX-cell lines panel related to Supplementary Figure S8A.

Supplementary Figure S1

Supplementary Figure S1 shows the characterization of the epigenetic subtypes, related to Figure 1.

Supplementary Figure S2

Supplementary Figure S2 shows the characterization of the enriched TF set per subtype, related to Figure 2.

Supplementary Figure S3

Supplementary Figure S3 shows the superenhancer signatures and pathway analysis, related to Figure 3.

Supplementary Figure S4

Supplementary Figure S4 shows the core regulatory circuits in OS PDX-derived cell lines, related to Figure 4.

Supplementary Figure S5

Supplementary Figure S5 shows the clustering and survival analysis including the TARGET osteosarcoma dataset, related to Figure 5.

Supplementary Figure S6

Supplementary Figure S6 shows the Multiomics for single nucleus ATAC-seq + RNA-seq from PDXs, related to Figure 6.

Supplementary Figure S7

Supplementary Figure S7 shows the Multiomics for single nucleus ATAC-seq + RNA-seq from cell lines.

Supplementary Figure S8

Supplementary Figure S8 show the analysis of drug response, related to Figure 7.

Data Availability

All sequencing data generated in this study have been deposited in the Gene Expression Omnibus under the SuperSeries GSE270506, including bulk ATAC-seq (GSE270502; Fig. 1), CUT&RUN profiling (GSE270503 and GSE309031; Figs. 2–4), multiome snRNA-seq/snATAC-seq from PDX-derived cell lines (GSE270504; Supplementary Fig. S7) and PDXs (GSE307240; Fig. 6), bulk RNA-seq (GSE270505), and ATAC-seq following TF perturbation (GSE308957; Fig. 4J–K).

Authors’ Disclosures

E. López-Fuentes reports grants from Make it Better Foundation, University of California Institute for Mexico and the United States, The Pablove Foundation, and Rally Foundation during the conduct of the study. E.A. Sweet-Cordero reports grants from NCI, St Baldrick’s Foundation, Alex’s Lemonade Stand Foundation, and Hyundai Hope on Wheels during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

E. López-Fuentes: Data curation, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. A.S. Clugston: Resources, data curation, formal analysis, visualization, methodology, writing–review and editing. A.G. Lee: Resources, data curation, formal analysis, funding acquisition, visualization, methodology, writing–review and editing. L.C. Sayles: Investigation, methodology, writing–review and editing. N. Sorensen: Methodology. M.V. Pons Ventura: Investigation, methodology. S.G. Leung: Investigation, methodology. T. Dinh: Investigation, methodology. M.R. Breese: Investigation, methodology. E.A. Sweet-Cordero: Conceptualization, resources, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing.