Genetic Screen in a Pre-Clinical Model of Sarcoma Development Defines Drivers and Therapeutic Vulnerabilities

Abstract

Purpose:

High-grade complex karyotype sarcomas are a heterogeneous group of tumors with a uniformly poor prognosis. Within complex karyotype sarcomas, there are innumerable genetic changes but identifying those that are clinically relevant has been challenging.

Experimental Design:

To address this, we utilized a pooled genetic screening approach, informed by TCGA data, to identify key drivers and modifiers of sarcoma development that were validated in vivo.

Results:

YAP1 and wildtype KRAS were validated as drivers and transformed human mesenchymal stem cells into two distinct sarcoma subtypes, undifferentiated pleomorphic sarcoma (UPS) and myxofibrosarcoma (MFS), respectively. A subset of tumors driven by CDK4 and PIK3CA reflected leiomyosarcoma (LMS) and osteosarcoma (OS) demonstrating the plasticity of this approach and the potential to investigate sarcoma subtype heterogeneity. All generated tumors histologically reflected human sarcomas and had increased aneuploidy as compared to simple karyotype sarcomas. Comparing differential gene expression of TCGA samples to model data identified increased oxidative phosphorylation signaling in YAP1 tumors. Treatment of a panel of soft tissue sarcomas with a combination of YAP1 and oxidative phosphorylation inhibitors led to significantly decreased viability.

Conclusions:

Transcriptional co-analysis of TCGA patient samples to YAP1 and KRAS model tumors support that these sarcoma subtypes lie along a spectrum of disease and adds guidance for further transcriptome-based refinement of sarcoma subtyping. This approach can be used to begin to understand pathways and mechanisms driving human sarcoma development, the relationship between sarcoma subtypes and to identify and validate new therapeutic vulnerabilities for this aggressive and heterogeneous disease.

Statement of Translational Relevance

High-grade complex karyotype sarcomas are a heterogeneous group of tumors with poor outcomes. This is due to the development of metastatic disease, indicating a high rate of primary treatment failure. In sarcomas, there are few mutations and targeted therapy options are limited. To identify and test new primary therapies, we have developed models to study sarcoma that allow for investigations of genetic drivers and critical underlying pathways. Using this approach and comparing these models to the human disease, we have identified a subset of tumors that are sensitive to treatment with a combination of YAP and oxidative phosphorylation inhibition. These studies create the foundation for a shift in sarcoma treatment from a histology based to a genetic based approach.

Introduction

High-grade complex karyotype sarcomas are a heterogeneous group of tumors with a 65% survival rate at 5 years despite aggressive multimodality treatment with surgery, radiation, and chemotherapy. The heterogeneity of the disease is evidenced by the more than 70 sarcoma subtypes with varying histology, genetics, and patient demographics leading to challenges in treatment and clinical trial design ¹.

High-grade sarcomas can be genetically subcategorized into two distinct groups. First, those that have a simple karyotype where the biology is driven by a known translocation ². Second, complex karyotype sarcomas that display an aneuploidy phenotype with numerous chromosomal gains and losses ³. In adults, 85% of sarcomas are complex karyotype. Contrary to translocation driven sarcomas that have a clearly defined oncogenic driver, pathways and factors leading to the development of complex karyotype sarcomas are poorly understood. Further contributing to the complexity of sarcoma management is the rarity of sarcomas as they cumulatively constitute less than 1% of all adult tumors. This leads to limited biological material for study and small sample sizes for clinical trials, resulting in a generalization of treatments for a heterogeneous set of diseases ⁴.

To gain insight into the biology of complex karyotype soft tissue sarcomas, The Cancer Genome Atlas (TCGA) genetically characterized the most common sarcomas in adults. These tumors were found to have a low mutation burden and a high rate of aneuploidy ⁵. The most common mutations and chromosomal alterations were loss of known tumor suppressors RB1 or TP53 which have established roles in sarcoma biology ^3,6,7. Mesenchymal stem cells (MSCs) give rise to the connective tissues of the body and are the presumed cell of origin for bone and soft tissue sarcomas ⁸. Prior models of human sarcoma using MSCs as the cell of origin have relied primarily on viral proteins such as SV40 or mutated RAS to drive transformation, and these have resulted in the formation of a single undifferentiated sarcoma subtype ^9–13. More recently, human MSCs have been transformed to osteosarcoma using RB1 silencing and c-Myc overexpression or by the addition of c-JUN to E6/E7 immortalized MSCs ^14,15. While these models have provided valuable insights, the ability to make comparisons among human sarcoma subtypes within a common model system has not been possible.

Here, we describe the development and characterization of a new pre-clinical model of human sarcoma. This system transforms mesenchymal stem cells into high-grade sarcomas by introducing genetic alterations observed across human sarcomas. Histologically, this results in the formation of four distinct high grade sarcoma subtypes, namely, undifferentiated pleomorphic sarcoma (UPS), myxofibrosarcoma (MFS), osteosarcoma (OS), and leiomyosarcoma (LMS). Transcriptional analysis demonstrated that UPS and MFS reflect the human disease and comparisons of the model to human tumors identified clinically relevant subpopulations and new therapeutic opportunities providing a platform to study the heterogeneity of this aggressive disease.

Materials and Methods

Cloning of CRISPR/Cas9 Constructs

Guide RNA sequences targeting RB1 (GCTCTGGGTCCTCCTCAGG), TP53 (CCGGTTCATGCCGCCCATGC), or control sequences (GTAATCCTAGCACTTTTAGG, GTTCCGCGTTACATAACTTA) were individually cloned into the LRG vector as previously described ¹⁶. Sequences were previously used in a genome wide screen by Wang et al ¹⁷. To combine these guides into one vector, the Gibson cloning method was used. The U6 promoter, guide sequence, and scaffold were amplified with custom primers (IDT) that contained a tag overlapping the LRG sequence flanking the EcoRI site. LRG vectors containing the control sequence or RB1 targeting sequence were cut with EcoRI and then ligated with the amplified sequence using the NEB HiFi Assembly kit (Cat#E5520).

Lentivirus Production

To generate lentivirus, 293T cells (ATCC CRL-3216, RRID:CVCL_0063) were transfected with the 3^rd generation lentivirus packaging plasmids pMDL, pREV, and pVSVg ¹⁸ and a lentivirus construct using calcium chloride transfection. Media containing virus was collected 48 hours after transfection, filtered through a 0.22μm filter and concentrated by centrifugation using Amicon 100kDa centrifugal filter units (Millipore Cat # UFC9100). Virus was titered using the Lenti-X p24 Rapid Titer Kit (Takara Cat #632200).

Cell Lines

To generate the ASC52telo (ATCC SCRC-4000, RRID: CVCL_U602) cell lines with RB1 and TP53 gene knockout or controls, ASC52 cells were cultured in MSC basal media (Lonza Cat #PT-3001). ASC52 was validated using STR profiling. Lentivirus was generated from pCW-Cas9 (Addgene plasmid #50661, RRID:Addgene_50661), construct with doxycycline inducible Cas9, as described above and this was added to cells using polybrene and spinfection ¹⁷. Cells were selected in puromycin and then single cells were seeded into 96 well plates. The clones were expanded and tested by western blot for Cas9 in the presence or absence of doxycycline. The clone with the tightest regulation of Cas9 was used for cell line generation. Lentivirus constructs containing CRISPR guides were transduced into cells using spinfection and polybrene. Cells were expanded and then sorted for green fluorescent protein (GFP) positivity. Individual GFP positive cells were sorted into 96 well plates to create single cell clones. Clones were expanded and tested by western blot to determine which had loss of RB1 and/or TP53. The DNA level mutation leading to the knockout was determined by PCR of the region containing the CRISPR target sequence and subsequent TOPO cloning and sanger sequencing (Invitrogen K2800J10). For drug studies, GCT (ATCC TIB-223 RRID: CVCL_1229), SKLMS (ATCC HTB-88 RRID: CVCL_0628), and SW872 (ATCC HTB-92 RRID: CVCL_1730) were purchased. NIH UPS1 (Lot 12654-015-R-J1-PDC) and NIH UPS2 (Lot 317291-083-R-J1-PDC) are patient derived tumor cultures provided by the NCI Patient Derived Models Repository (PDMR). Sheffield Undifferentiated Pleomorphic Sarcoma Cell Line 1 and 2 (SH1 and SH2) were obtained from the University of Sheffield and previously characterized and validated ¹⁹. To create YAP knockout cell lines from GCT, UPS2, and SW872, Cas9 expressing lines, the same procedure was used with guides targeting YAP1. All lines were cultured in provider recommended medias and tested regularly for mycoplasma contamination (last in June 2024). All studies were performed within seven passages of thawing.

Flow Cytometry

Human mesenchymal stem cells (ATCC PCS-500-012) or ASC52telo cells were suspended in 5%FBS/PBS solution and the following antibodies were added: APC-CD14 (BD 561708, RRID:AB_398596), APC-CD19 (BD 561742, RRID:AB_398597), PE-CD29 (Biolegend 303003, RRID:AB_314319), APC-CD34 (BD 560940, RRID:AB_398614), PE-CD44 (BD 555479, RRID:AB_395871), APC-CD45 (BD 560973, RRID:AB_398600), PE-CY7-CD73 (BD 561258, RRID:AB_10643764), PERCP-CY5.5-CD90 (BD 561557, RRID:AB_10712762), PE-CD166 (Biolegend 343903, RRID:AB_2289303). Percentage positive was determined as compared to an unstained control using FloJo analysis of FACS plots.

In Vitro Differentiation

To differentiate mesenchymal stem cells in vitro, the Lonza protocol was used. Briefly, adipogenic, osteogenic, and chondrogenic medias were prepared by addition of supplements (Lonza cat #PT-3002, PT-3003, PT-3004). ASC52telo cells were plated at 1x10⁴ per cm² for adipogenic differentiation and 1.5x10³ per cm² for osteogenic differentiation. Cell numbers used were less than recommended by the manufacturer because these cells lack contact inhibition. For chondrogenic differentiation, cells were resuspended at a concentration of 1.6x10⁷ and 5ul was used per pellet. Differentiation medias were added and changed every three days for 14-21 days. Cells were fixed in 10% PBS buffered formalin then stained with oil red o, alizarian red, and alcian blue as previously described for visualization ²⁰.

Western Blot

Cells were grown in adherent conditions and either left untreated or treated with 20uM doxorubicin (Fisher AAJ64000MA) for 24 hours. Cells were collected and lysed in 8M urea buffer containing protease inhibitors (Prometheus 18-427). Protein was quantified using a BCA assay (Pierce Rapid Gold BCA Protein Assay PIA53226). Blots were probed using antibodies against P53 1C12 (Cell Signaling 2524, RRID:AB2714036) or RB1 4H1 (Cell Signaling 9309, RRID:AB_823629) followed by Beta Actin BA3R (Invitrogen MA515739, RRID:AB_10979409). Bands were visualized using chemiluminescence (Supersignal West Pico PLUS PI34577). For additional studies, YAP D8H1X (Cell signaling 14074, RRID_AB2650491) and KRAS EPR23474-76 (Abcam 275876) antibodies were utilized.

Lentivirus Library Production

LentiORF clones were purchased from Sigma containing genes of interest in the TRC3 backbone (Millipore Sigma). All constructs are V5 tagged except for KRAS. Each construct was verified by western blot using a V5 D3H8Q (Cell signaling 13202, RRID:AB_2687461) antibody. Plasmids were maxiprepped, quantified, then pooled together for lentivirus production. Lentivirus was produced using the 3^rd generation system as described above. In brief, 293T cells were transfected, media containing virus was collected after 48 hours and filtered. The virus was concentrated by centrifugation and titered as described aboe.

Primary Screen

All in vivo work was ethically performed and approved by the UC Davis Institutional Animal Care and Use Committee with oversight from the veterinarian staff. Cells lines with RB1 and TP53 targeted or a control line were transduced with the lentiviral library at MOI of 20 using spinfection and polybrene. This MOI was used because of the difficulty in transducing primary cells, at this MOI, approximately 60-70% of cells are transduced. Cells were expanded in culture for 3-5 days and then trypsinized, counted, and suspended in 60ul of a 1:1 mix of HBSS and Matrigel (Corning 354234). These cells were injected subcutaneously into 6-8 week old immuocompromised Non-SCID Gamma (NSG) mice (Jackson Labs Cat#005557). Tumors were monitored by palpation and once tumors reached 1cm, frozen and formalin fixed samples were collected.

Secondary Screen

To test the ability of individual oncogenes to drive tumor transformation, individual TRC3 constructs (Sigma) were transduced into cells using spinfection. These were grown in culture and expanded then trypsizined, counted, and suspended in a 1:1 mix of HBSS and Matrigel and injected subcutaneously into NSG mice.

Barcode Sequencing

For analysis of barcodes in tumor samples, DNA was extracted from paraffin embedded samples. Tumors were sectioned and stained by H&E to ensure the presence of viable tissue. The subsequent section was removed from the slide and DNA was extracted (Qiagen, Cat#56404). Barcodes were amplified using flanking primers (CTTGAAAGTATTTCGATTTCTTGGC and TCCAGAGGTTGATTGTCGAC) and KAPA HiFi HotStart ReadyMix (Cat#K2602) to form a 104bp product. Adapters for next generation sequencing were ligated to the PCR products using the KAPA Hyper Plus Kit and KAPA dual indexed adapter kit (Cat#K8727 and Cat#KK8514). Libraries were evaluated using a bioanalyzer and samples were submitted for next generation sequencing to UC Berkeley QB3 genomics and sequenced using Illumina technology. Barcodes were extracted from sequenced library fastQs by correcting for read direction using grep to sort, then taking the first 24bp of each read, and then determining the top 100 unique most frequent barcodes in each library using uniq in unix. Sequences were checked for matches within 1 levenshtein distance by building a dictionary of possible sequences in R software and then searching for matches to known transgene barcodes. The total of each transgene barcode detected per library was tabulated from this data and then visualized using the pheatmap (V1.0.12; RRID:SCR_016418) library in R.

Histology

Paraffin embedding, sectioning, and H&E staining was performed by the UC Davis Genomic Pathology Lab. SATB2, myogenin, and alpha SMA staining was performed by the UCLA TPCL facility. Vimentin staining was done according to the following protocol. Immunohistochemistry (IHC) staining was performed on formalin-fixed paraffin embedded (FFPE) tissue sections. Sections were de-paraffinized in Xylene, then re-hydrated. Antigen Unmasking Solution (Vector Laboratories, Cat# H-3300), citrate-based pH 6.0, was used for antigen retrieval. Endogenous peroxidase activity was blocked by incubating sections with BLOXALL (Vector, Cat# SP-6000) for 10 minutes. The M.O.M Mouse IgG Blocking Reagent from the M.O.M ImmPRESS Polymer Kit (Vector, Cat# MP-2400) was used to block endogenous staining. The Vimentin primary antibody (Agilent, Cat#M072529-2) was diluted 1:75 in 2.5% horse serum. Secondary antibody staining was done with the M.O.M ImmPress Horse Anti-Mouse IgG for 10 minutes. The ImmPACT DAB EqV (Vector, Cat# SK-4103) reagent was used for visualization. Sections were counterstained with hematoxylin and mounted with VectaMount AQ Mounting Medium (Vector, CAT# H-5501). All slides were interpreted by Dr. Morgan Darrow, a board-certified pathologist with expertise in sarcoma.

RNA Sequencing

Total RNA was isolated from human sarcoma models and controls using the RNeasy Mini Kit (Qiagen 74101). Library preparation was performed by the Functional Genomics Laboratory (FGL), a QB3-Berkeley Core Research Facility at UC Berkeley. Total RNA samples were evaluated by Bioanalyzer assay (Agilent) for quality. Only high-quality RNA samples (RIN > 8) were used. At the FGL, Oligo (dT)₂₅ magnetic beads (Thermofisher) were used to enrich mRNA. The treated RNAs were rechecked on the Bioanalyzer for their integrity. The library preparation for sequencing was done on Biomek FX (Beckman) with the KAPA hyper prep kit for RNA (now Roche). Truncated universal stub adapters were used for ligation, and indexed primers were used during PCR amplification to complete the adapters and enrich the libraries for adapter-ligated fragments. Samples were checked for quality on an AATI (now Agilent) Fragment Analyzer. Samples were then transferred to the Vincent J. Coates Genomics Sequencing Laboratory (GSL), where quantitative PCR was used to calculate sequence-able molarity with the Kapa Biosystems Illumina Quant qPCR Kits. Libraries were pooled evenly by molarity and sequenced on an Illumina NovaSeq6000 150PE S4. Raw sequencing data were converted into fastq format, sample-specific files using the Illumina bcl2fastq2 software (RRID:SCR_015058).

Gene Expression PCA and PLSR Analyses

RNA-seq fastq files were processed through the TOIL pipeline (V3.12.0; RRID:SCR_024391). TOIL processed and batch effect normalized gene expression data of TCGA patient tumors were acquired from the XenaBrowser (RRID:SCR_018938). For visualization, RSEM expected count data was upper quartile normalized and log2 transformed. Unsupervised principal component analysis (PCA) and supervised Partial Least Squares Regression (PLSR) were performed on protein-coding genes. PCA was performed centered and unscaled using the prcomp function in R (V4.3.3; RRID:SCR_001905). PLSR was performed unscaled using the pls function in R. Projections onto PCA/PLSR frameworks were done by multiplication of the projected sample expression profiles by the rotation matrix.

To systematically select a range of primary cancer types from the TCGA to be included in the pancancer gene expression comparisons, a PCA of only protein coding genes was performed with all 33 TCGA cancer types. (disease = # of samples, acute myeloid leukemia = 173, adrenocortical cancer = 77, bladder urothelial carcinoma = 426, brain lower grade glioma = 523, breast invasive carcinoma = 1212, cervical & endocervical cancer = 308, cholangiocarcinoma = 45, colon adenocarcinoma = 330, diffuse large B-cell lymphoma = 47, esophageal carcinoma = 195, glioblastoma multiforme = 171, head & neck squamous cell carcinoma = 564, kidney chromophobe = 91, kidney clear cell carcinoma = 603, kidney papillary cell carcinoma = 321, liver hepatocellular carcinoma = 421, lung adenocarcinoma = 574, lung squamous cell carcinoma = 548, mesothelioma = 87, ovarian serous cystadenocarcinoma = 427, pancreatic adenocarcinoma = 183, pheochromocytoma & paraganglioma = 185, prostate adenocarcinoma = 546, rectum adenocarcinoma = 103, sarcoma = 205, skin cutaneous melanoma = 470, stomach adenocarcinoma = 449, testicular germ cell tumor = 154, thymoma = 121, thyroid carcinoma = 571, uterine carcinosarcoma = 57, uterine corpus endometrioid carcinoma = 204, uveal melanoma = 79). Sarcoma samples were filtered to only include samples that were confirmed by central pathology (DDLPS = 49, MFS = 17, MPNST = 5, SS = 10, STLMS = 53, ULMS = 27, UPS = 43). A Euclidean distance was then calculated between the TCGA sarcoma cluster and every other cancer type to measure overall transcriptome likeness. The distance was calculated to the 11^th component, accounting for >1/2 of the variance (51.11%). The distance contributed from each component was multiplied by that component’s percent variance explained to weigh components with a higher proportion of variance more significantly. Cancer types were then ranked closest (skin cutaneous melanoma) to farthest (brain lower grade glioma) and every fifth was selected. To calculate the distance from the tumor models cluster to the subset of TCGA clusters, the same protocol was performed to the 8^th component (50.78% variance). For the PCA containing only the model data with sarcoma and skin cutaneous melanoma clusters, the distance was calculated to the 15^th component (50.72%).

Differential Expression Analyses

Differential expression analysis was performed on raw RSEM expected count data of protein-coding genes using DESeq2 (V1.42.1; RRID:SCR_015687). DESeq2 was used to compare MFS (n = 17) versus UPS (n = 44) amongst patients, and then YAP1 (n = 3) versus KRAS (n = 3) within the human transformation model. Ggplot2’s stat_density_2d (V3.5.1; RRID:SCR_014601) was used to highlight regions of high and low overlap in the rank-rank plot. Rank rank hypergeometric overlap analysis was performed using the R package RRHO (V3.19; RRID:SCR_014024). The R package lattice (V0.22-6; RRID:SCR_015662) was used to plot the hypergeometric p-values.

Gene Set Enrichment Analysis (GSEA) and GSEA-Squared

Gene set enrichment analysis (GSEA) was performed using the fgsea (V1.28.0; RRID:SCR_020938) package with minSize = 3, maxSize = 50000, eps = 1e-20. Gene ontology (GO), canonical pathways (CP), and hallmark (H) gene sets from MSgidDB were included. Genes were ranked by the signed and log2 transformed p-value calculated by DESeq2. Only pathways with a p-adj <0.05 were individually plotted. To investigate trends in related enriched or de-enriched gene sets, GSEA-Squared analysis was performed. After identifying broad pathways of interest from top Normalized Enrichment Score (NES)-ranked GSEA results (e.g., Oxidative Phosphorylation, DNA Damage), related key terms of interest were queried. All gene sets were ranked by NES and marked for if they contained a key term in the pathway name. To assess the distribution of a category of terms, KS tests were performed using ks.test.2 (RRID:SCR_025636). The keywords used are listed (Supplemental Table 1A). Names of the pathways individually plotted were adjusted to save space using specific formatting (Supplemental Table 1B).

Whole Exome Sequencing

gDNA was isolated from human sarcoma models and controls using the QIAamp DNA Mini Kit (Qiagen). Library preparation was performed by the Functional Genomics Laboratory (FGL), a QB3-Berkeley Core Research Facility at UC Berkeley. The DNA quality was checked on Fragment analyzer (Agilent) and fragmented to the desired length using Bioruptor pico (Diagenode). The library preparation for sequencing was done on Biomek FX (Beckman) with the KAPA hyper prep kit for DNA (now Roche). Truncated universal stub adapters were used for ligation, and indexed primers were used during PCR amplification to complete the adapters and enrich the libraries for adapter-ligated fragments. Samples were checked for quality on an AATI (now Agilent) Fragment Analyzer. Samples were then transferred to the Vincent J. Coates Genomics Sequencing Laboratory (GSL), where quantitative PCR was used to calculate sequence-able molarity with the Kapa Biosystems Illumina Quant qPCR Kits. Libraries were pooled evenly by molarity and sequenced on an Illumina NovaSeq6000 150PE S4. Raw sequencing data were converted into fastq format, sample-specific files using the Illumina bcl2fastq2 software (RRID:SCR_015058).

Copy Number Variation Analyses

Copy number variation (CNV) analysis of the human models was performed using CNVKit (V0.9.9; RRID:SCR_021917). CNVKit was run using an in-house pooled normal reference from normal samples sequenced under the same conditions at the Technology Center for Genomics & Bioinformatics (TCGB) at UCLA. Copy number segmentation data of TCGA samples were downloaded from XenaBrowser (RRID:SCR_018938). (disease = # of samples, acute myeloid leukemia = 197, adrenocortical cancer = 90, bladder urothelial carcinoma = 411, brain lower grade glioma = 527, breast invasive carcinoma = 1099, cervical & endocervical cancer = 298, cholangiocarcinoma = 36, colon adenocarcinoma = 455, diffuse large B-cell lymphoma = 48, esophageal carcinoma = 185, glioblastoma multiforme = 592, head & neck squamous cell carcinoma = 524, kidney chromophobe = 66, kidney clear cell carcinoma = 529, kidney papillary cell carcinoma = 289, liver hepatocellular carcinoma = 372, lung adenocarcinoma = 518, lung squamous cell carcinoma = 503, mesothelioma = 87, ovarian serous cystadenocarcinoma = 602, pancreatic adenocarcinoma = 185, pheochromocytoma & paraganglioma = 180, prostate adenocarcinoma = 493, rectum adenocarcinoma = 166, sarcoma = 206, skin cutaneous melanoma = 472, stomach adenocarcinoma = 477, testicular germ cell tumor = 156, thymoma = 123, thyroid carcinoma = 512, uterine carcinosarcoma = 56, uterine corpus endometrioid carcinoma = 541, uveal melanoma = 80). Sarcoma samples were filtered to only include samples that were confirmed by central pathology (DDLPS = 50, MFS = 17, MPNST = 5, SS = 10, STLMS = 53, ULMS = 27, UPS = 43) ⁵.

Integrated CNA score ²¹ for each sample was defined as:

$${\mathrm{\Sigma }}_{\mathit{segments}}\left|\mathit{segment\; end}-\mathit{segment\; start}\right|\ast \left|\mathit{segment\; mean}\right|$$

The copy number of a gene was calculated as follows with an assumed purity of 0.95 and ploidy of 2:

$$\mathit{AbsCN}=\left[\left(2^{\mathit{log}2\left(\mathit{copy\; ratio}\right)}\ast \left(2\ast \left(1-\mathit{purity}\right)+\mathit{ploidy}\ast \mathit{purity}\right)-2\ast \left(1-\mathit{purity}\right)\right)\right](\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\mathit{purity}$}\right.)$$

Viability Analysis

For drug treatments, cells were plated into 96 cell plates on day 1 and treated with inhibitors or DMSO on day 2. The following day, relative proliferation was analyzed using Cell Titer Glo (Promega). Briefly, cells were lysed in the plate using the provided reagent and light was measured using a plate reader. Data was normalized to the DMSO control and plotted as a percentage of this value. IM156 (HY-136093A), Verteporfin (HY-B0146), and IAG933 (HY-158384) were purchased from Med Chem Express.

Data Availability Statement

All data are available in the main text or the supplementary materials. RNA sequencing data is publicly available at the Gene Expression Omnibus (GEO) (RRID:SCR_005012) repository accession GSE228213. Whole exome sequencing data is available on the Dataview server under bioproject (RRID:SCR_004801) number PRJNA948835. Lentiviral plasmids containing guides targeting RB1 and/or TP53 are available at Addgene (accession numbers 225873, 225874, 225875, 225876). Raw data can be obtained by contacting the corresponding author.

Results

Genetic Screen Results in the Development of High-Grade Sarcomas

To develop an in vivo model of sarcoma development with the goal of creating different subtypes for studies of sarcoma heterogeneity, we began with human MSCs and aimed to introduce genetic changes that are commonly observed across high grade sarcomas (Figure 1A). While mouse MSCs can be readily transformed to sarcomas, human MSCs have been more challenging ²². A putative mechanism for this variation between species is due to differences in telomere maintenance ^11–13. In TCGA data, human telomerase (hTERT) was amplified in a subset of human sarcoma ⁵. Therefore, as a basis for these studies, we utilized ASC52telo, a MSC line that has been immortalized by hTERT ²³. Despite immortalization, these cells have identical cell surface marker expression to non-immortalized control mesenchymal stem cells (Supplemental Figure 1A). These cells can undergo in vitro differentiation into adipocytes, osteocytes, and chondrocytes. Adipogenic differentiation is limited compared to non-immortalized MSCs due to impaired contact inhibition and chondrogenic pellets were larger and more diffuse due to an increased proliferative rate (Supplemental Figure 1B).

Figure 1. Human Mesenchymal Stem Cells can be Transformed to High-Grade Sarcomas.

(A) Schematic diagram of forward transformation model of sarcoma development. (B) Western blot of generated cell lines were evaluated for RB1 and P53 expression. Given the low basal levels of P53, cells were treated with doxorubicin to induce expression. Beta actin is shown as a loading control. Experiment was performed in duplicate (C) H&E staining demonstrates high power views of sarcomas formed by addition of the lentiviral vector library to each of the genetic backgrounds shown. Scale bar represents 50um. (D) Distinct areas of necrosis were seen as evidenced by an immune infiltrate, scale bar represents 50um. N=8 per genotype.

RB1 and TP53 are commonly inactivated or mutated in human soft tissue sarcomas ³. While mutations in these genes are uncommon, they can be lost at the chromosomal level or amplifications of negative regulators can lead to functional suppression⁵. Therefore, we targeted RB1 and TP53 in ASC52telo using a CRISPR-Cas9 system ²⁴. A lentivirus containing doxycycline-inducible Cas9 was transduced into the ASC52telo cell line ¹⁷. An inducible system was used to mitigate the potential in vitro and in vivo toxicity of high levels of Cas9 ²⁵. Guides targeting either RB1, TP53, both genes, or a control sequence were inserted into a lentiviral vector containing GFP and cells were sorted and expanded (Supplemental Figure 1C) ¹⁶. Gene knockout or knockdown was verified at the DNA and protein levels (Figure 1B and Supplemental Figure 1D). This resulted in four cell lines; wildtype (RB1+/+TP53+/+), loss of RB1 alone (RB1−/−TP53+/+), loss of TP53 alone (RB1+/+TP53+/−) and loss of both genes (RB1−/−TP53+/−). Interestingly, although numerous clones screened showed heterozygous loss of TP53, we did not observe homozygous allele loss. Despite this, the P53 expression is low in response to stress, as evidenced by response to doxorubicin treatment (Figure 1B).

ASC52telo cells with loss of RB1, TP53, or both genes failed to yield tumors within 90 days of subcutaneous injection into mice, suggesting this is not sufficient to drive robust tumor formation (0/12 tumors formed). We hypothesized that additional genetic alterations were needed for transformation. To aid in the assessment of multiple candidates, we generated a lentiviral library that would allow for a genetic screen approach. All genes were wildtype given the low mutation rate observed in human samples. Genes upregulated or amplified in the TCGA soft tissue sarcoma dataset such as JUN, YAP1, and CDK4 were included ⁵. In addition, genes implicated in the development of sarcomas such as the Wnt antagonist DKK1 that has been shown to transform human MSCs expressing SV40 to sarcoma were included ¹⁰. Factors regulating the stem cell population were also utilized along with genes involved in the growth of sarcomas such as FOXM1 ²⁶. Barcoded lentiviral constructs containing twenty-seven genes were used in the screen (Supplemental Figure 1E).

In murine syngeneic models, loss of p53 is sufficient for tumor formation but when using human MSCs to induce transformation to high-grade sarcoma, the requirements for RB1 and TP53 are not known ^27,28. Therefore, all four lines, RB1+/+TP53+/+, RB1−/−TP53+/+, RB1+/+TP53+/− and RB1−/−TP53+/− cells were transduced with the lentiviral gene library and implanted subcutaneously into immunocompromised non-SCID gamma (NSG) mice. After 8 weeks, 7/16 (44%) of RB1−/−TP53+/+, 12/16 (75%) of RB1+/+TP53+/− and 14/16 (88%) of RB1−/−TP53+/− cells transduced with the library formed tumors. Control cells (RB1+/+TP53+/+) did not yield tumors, indicating that the lentiviral library alone is not capable of transforming MSCs to sarcomas. Tumors from these different genetic backgrounds were indistinguishable and demonstrated highly pleomorphic cells with numerous mitotic figures and areas of necrosis consistent with high grade sarcoma (Figure 1C–D). Interestingly, loss of heterozygosity at the TP53 locus was not observed. Taking this data together, both targeting of a tumor suppressor and activation of an oncogene are needed for transformation to high-grade sarcoma. Based on the tumor penetrance data, the contribution of RB1 to the system may be relatively small but, RB1 loss has a clear role in sarcoma pathophysiology and it is possible that specific oncogenes may require RB1 inactivation for transformation. Therefore, for subsequent studies, RB1−/−TP53+/− cells were used to create the most robust and permissive environment for sarcoma development.

Tumors Histologically and Phenotypically Mimic the Human Disease

In adults, undifferentiated pleomorphic sarcoma is the most frequently diagnosed high-grade complex karyotype sarcoma. This tumor type is characterized as having pleomorphic cells and a lack of immunohistochemical staining that would indicate differentiation towards a specific lineage such as bone, muscle, or fat ^1,29. Further histologic examination of tumors formed from the transduction of the lentiviral library to RB1−/−TP53+/− cells showed primarily areas of high-grade undifferentiated pleomorphic sarcoma although, infrequent areas contained two distinct histologies (Figure 2A). The most abundant histology was consistent with undifferentiated pleomorphic sarcoma (UPS) while the second area contained features of aggressive sarcoma and a lack of differentiation but also abundant myxoid accumulation, consistent with high-grade myxofibrosarcoma (MFS) (Figure 2B–C). Immunohistochemistry of these tumors showed strong positivity for vimentin with alpha smooth muscle actin marking endothelial cells and infrequent tumor cells. Myogenin was consistently negative indicating the tumors did not demonstrate evidence of muscle differentiation consistent with rhabdomyosarcoma (Figure 2D). Taken together, these tumors represent undifferentiated pleomorphic sarcoma histology intermixed with myxofibrosarcoma.

Figure 2. Primary Screen Identifies YAP1, KRAS, and DDIT3 as Potential Drivers of Sarcoma Development.

(A) Low power view of H&E staining of tumors formed that demonstrated two distinct histologies. Scale bar represents 300um. High power view of the two distinct histologies labeled “B” and “C” demonstrate (B) Undifferentiated pleomorphic sarcoma and (C) Myxofibrosarcoma. Scale bar represents 50um. Primary screen was performed in triplicate with 4-8 implants per experiment (D) Immunohistochemistry of tumors formed show strong positivity for vimentin, primarily endothelial cell staining and infrequent tumor cells showing positivity for smooth muscle actin, and negative staining for myogenin. Left side shows low power views with scale bars representing 300um and right side shows high power views with 50um scale bars. N=6, two tumors from each experiment were stained for the panel of IHC markers. (E) Representative lung sections show the presence of metastatic sarcoma cells with low (top) and high (bottom) power views and scale bars representing 300um and 50um respectively. N=12 animals total in duplicate experiments, of those analyzed, 5 demonstrated lung metastasis (42%) (F) Heat map representing barcode sequencing counts from the primary screen. See also supplemental figure 2 for percentage of each barcode per sample and calculated p-values.

To determine if these tumors phenotypically mimic the human disease, cells were injected into the gastrocnemius muscle of immunocompromised mice. In humans, the majority of sarcomas are found in the extremities and metastasize almost exclusively to the lungs ³⁰. Consistent with this, implantation of RB1−/−TP53+/− cells with the lentiviral library into the gastrocnemius gave rise to lung metastases in 42% (5/12) of animals without lesions in the lymph nodes, liver, or spleen (Figure 2E). Using this defined and highly efficient system, mesenchymal stem cells can be transformed into high grade sarcomas that histologically and functionally reflect the human disease.

Primary Screen Identifies YAP1, KRAS, and DDIT3 as Potential Drivers of Sarcoma Development

To identify genes in the lentiviral library that drive high-grade sarcoma formation, the introduced genes enriched in the transformation process were identified by sequencing of their associated unique vector barcodes. Of note, all lentiviral library constructs were expressed although at varying reference levels prior to implantation and there was a minimal change in the relative levels over five days of cell culture, indicating significant in vitro selection had not occurred (Supplemental Figure 2A). Barcode sequencing of the tumor outgrowths from three independent experiments were analyzed and three genes, YAP1, DDIT3, and KRAS were enriched for and expressed at high levels across multiple tumor samples (Figure 2F, Supplemental Figure 2B). While KRAS and DDIT3 were strongly represented in the starting lentiviral pool, YAP1 was not and showed significant enrichment in 13/15 tumors during the process of transformation (Supplemental Figure 2C).

YAP1 and KRAS Drive the Formation of Histologically Distinct Sarcoma Subtypes

YAP1 or Yes associated protein is a key downstream mediator of Hippo signaling and has been shown to be overexpressed or amplified in sarcomas as well as other cancer types ^26,31,32. DNA damage inducible transcript 3 (DDIT3) is a member of the C/EBP family of transcription factors and regulates adipogenesis. The FUS-DDIT3 translocation leads to the formation of myxoid/round cell liposarcoma and amplification of DDIT3 is associated with a myxoid liposarcoma like histology ³³. Rare mutations in oncogenic HRAS and KRAS have been noted in genomic analysis of sarcomas and implicated in transformation, but a role for wildtype KRAS has not been established ^11,34. These genes have not been experimentally demonstrated to drive human sarcoma development.

Next, we determined if either YAP1, KRAS, or DDIT3 alone were capable of transforming RB1−/−TP53+/− cells into sarcomas. RB1−/−TP53+/− cells were transduced with a negative control, blue fluorescent protein (BFP), each lentivirus alone (YAP1, KRAS, and DDIT3) or the combination as a positive control. These were injected into equal numbers of male and female 6–8-week-old immunocompromised mice. As expected, the combination of all three genes was sufficient to drive transformation. Interestingly, YAP1 and KRAS alone led to tumor formation within 8-16 weeks while addition of DDIT3 did not after 26 weeks. Tumors can be propagated in vivo with second passage cells giving rise to tumors within 2 weeks with identical histology (Supplemental Figure 3A). Of note, the BFP negative control led to 12.5% implants forming tumors after 26 weeks that histologically represented UPS or osteosarcoma indicating that spontaneous transformation can occur at a long latency in RB1−/−TP53+/− cells.

UPS and MFS exist on a disease spectrum ⁵. YAP1 tumors were histologically consistent with undifferentiated pleomorphic sarcoma as evidenced by vimentin positivity and a lack of expression of markers of differentiated cell types (Figure 3A). KRAS tumors had the distinct histological appearance of myxofibrosarcoma (Figure 3B). The latency of tumor development in the secondary screen was longer than that observed for the primary screen indicating that cooperation amongst genes can likely accelerate sarcoma development. In support of this, while DDIT3 alone did not give rise to tumors within 26 weeks, the addition of DDIT3 to KRAS tumors led to a more aggressive phenotype with increased pleomorphism and shorter latency (Supplemental Figure 3B) indicating that DDIT3 is likely a modifier or enhancer of sarcoma development.

Figure 3. Drivers of Histologically Distinct Sarcoma Subtypes.

(A) Histology from YAP1 driven tumors are shown with lower power on the left column, scale bars 300um and high power on the right, scale bars 50um. Tumors histologically represent undifferentiated pleomorphic sarcoma with strong staining for vimentin and negative for SMA and myogenin. (B) KRAS driven tumors are shown in this panel with histology consistent with myxofibrosarcoma. Low power images are shown on the left with high power on the right. Representative western blots of tumors are shown underneath. Secondary screen was performed in duplicate with n=8 for the first experiment and n=4 for the replicate. (C) Secondary screen summary table showing the gene that was added to RB1−/−P53+/− cells, the amount of time until the tumor reached 1cm (latency) and the number of tumors that formed from the injections (efficacy) as well as the histology of the outgrowths. P-values representing the relative enrichment is shown in the table, for KRAS, YAP1, and DDIT3. Latency is shown as the mean +/− the standard deviation (D) H&E and immunohistochemistry of PI3KCA driven tumors that show osteosarcoma confirmed with SATB2 expression. Low power is shown on the left and higher power on the right. (E) H&E and immunohistochemistry of CDK4 driven tumors showing leiomyosarcoma with positive alpha smooth muscle actin staining. For all low power images, scale bars represent 300um and higher power shows scale bars representing 50um.

Identification of Additional Drivers and Modifiers of High-Grade Sarcoma Formation

We next considered that additional factors in the lentiviral library could lead to transformation, but this was masked by a proliferative advantage conferred by KRAS, YAP1, or DDIT3. Therefore, the primary screen was repeated removing KRAS, YAP1, and DDIT3 from the lentiviral library. Histologically, the resulting tumors were most consistent with undifferentiated pleomorphic sarcoma (Supplemental Figure 3C). Barcode sequencing of these tumors identified several potential drivers. Genes whose barcodes comprised >10% of represented barcodes in one tumor or >5% in two outgrowths were tested individually in a secondary screen (Supplemental Figure 3D).

BCL2 and SOX2 transduced individually gave rise to undifferentiated pleomorphic sarcoma albeit with a significantly prolonged latency and lower penetrance as compared to YAP1 and KRAS. Given that the latency and penetrance are similar to the spontaneous transformation rate seen with the BFP control, we labeled these as “potential drivers” (Figure 3C). CCND1 gave rise to one tumor after 264 days and zero tumors in a second experiment after 272 days indicating either a low penetrance or more likely, that the single tumor was a spontaneous transformant and that CCND1, like DDIT3, is a modifier of sarcoma development. JUN robustly drove the formation of high-grade sarcoma with an average latency of 71 days and 100% penetrance. Amplification or overexpression of JUN has been shown to inhibit key drivers of adipocytic differentiation ³⁵. Of note, these tumors at <1cm in size were primarily necrotic. Examined sections of viable cells had uniform V5 expression but a highly heterogenous appearance with areas histologically demonstrating undifferentiated pleomorphic sarcoma as well as pleomorphic leiomyosarcoma as evidenced by positive SMA staining and a more spindled morphology (Supplemental Figure 3E).

Rarely, mutations of PIK3CA have been noted in soft tissue sarcomas and these are associated with poor survival ³⁶. Addition of PI3KCA to RB1−/−TP53+/− cells led to formation of tumors although with a lower efficacy and longer latency as compared to cells expressing YAP1 or KRAS (Figure 3C). The histology of 50% of these tumors was consistent with osteosarcoma, confirmed by strong staining for SATB2 (Figure 3D), while 50% represented UPS.

Alterations in cyclin dependent kinase 4 (CDK4) have been previously noted in large sarcoma sequencing datasets ³⁷. In TCGA data, CDK4 was shown to be amplified in up to 18.5% of samples ⁵. These tumors had a significantly longer latency and a decreased penetrance. Of those, 60% of tumors displayed strong staining of SMA and had a more spindled appearance consistent with pleomorphic leiomyosarcoma while the remaining tumors were histologically UPS (Figure 3E). While the histology of these tumors is not uniform, it does demonstrate the potential of this model and that additional sarcoma subtypes can be generated using this approach. The latency of PIK3CA and CDK4 tumors are similar to BFP but, the penetrance of the tumors are above 50% which is significantly higher than BFP indicating these are drivers of sarcoma development albeit weaker than YAP1 and KRAS.

Tumor Models Demonstrate Chromosomal Changes and YAP1 Amplification

We then sought to determine if the tumors formed genetically reflect the human disease and contain the pathognomonic complex karyotype or aneuploidy phenotype ¹. To investigate this, YAP1, KRAS, CDK4, and PIK3CA driven tumors were evaluated by whole exome sequencing with ASC52telo and RB1−/−TP53+/− pre-transformation parent cells as controls. JUN was not included given the histologies were mixed within individual tumors and most tumors were necrotic. Copy number variation (CNV) analysis revealed ASC52telo cells harbor a notable amount of aneuploidy with CNV regions ranging in size from single genes to megabase pairs (Figure 4A). CNV changes are not uncommon in in vitro maintained stem cell cultures, but generally tend to be less than that seen in aneuploid tumors or models³⁸. In addition to the CNVs present in ASC52telo cells, which are highly conserved in the tumors, there were tumor model-specific CNVs. For example, YAP1 driven tumors displayed single-copy number loss of portions of chr 4, 5, 19, and 21 (Supplemental Figure 4A). The KRAS driven model displayed single-copy number loss of chr 5 and a portion of chr 19 (Supplemental Figure 4B). The PIK3CA driven model displayed single-copy number loss of chr 15 and various smaller deletions (Supplemental Figure 4C). Lastly, the CDK4 driven model displayed a two-copy number gain of a portion of chr1 where parental RB1+/+TP53+/+ cells displayed a single-copy number gain, indicating that regions of the genome which already contain CNVs in the RB1+/+TP53+/+ cells remain dynamic (Supplemental Figure 4D).

Figure 4. Tumor Models Demonstrate an Aneuploidy Phenotype and YAP1 Amplification.

(A) Heat map showing chromosomal gains (red) and losses (blue) in tumor models and cell lines. Three samples are shown from each tumor type and chromosomes are represented on the X axis. (B) Quantification of aneuploidy by iCNA scores. Thyroid carcinoma and kidney chromophobe are shown on the axis to demonstrate the extremes of the TCGA data. (C) Graph demonstrating chromosome 11q amplification in CDK4 and PIK3CA driven tumors and exogenous amplification of YAP1 in the YAP1 driven tumors. (D) Table summarizing the mRNA expression levels of YAP1 across the models and the relative copy number.

To quantify the degree of aneuploidy, an integrated copy number alteration (iCNA) score was calculated ²¹. To serve as a relative scale, iCNA scores were also calculated for all 33 TCGA primary tumors (Supplemental Figure 5A). The TCGA data were then ranked by median iCNA score, with the highest and lowest ranking disease, along with sarcoma, being plotted alongside the models and controls (Figure 4B). The tumor model iCNA scores fall within the range of human sarcomas. When plotted by subtype, the TCGA sarcoma data does not show any significant trends, other than synovial sarcoma, a simple karyotype sarcoma, demonstrating lower scores than controls, tumor models, and complex karyotype sarcomas (Supplemental Figure 5B). From these data, we can conclude that there are chromosomal level changes that evolve during the process of transformation. Yet, given the high degree of chromosomal changes in the starting material and the extent of aneuploidy noted in the human samples, the contribution of these changes to oncogenic fitness is unclear.

YAP1 has been shown to have copy number gains across a subset of high grade sarcomas ⁵. In this pre-clinical model, amplification of chromosome 11q in a 1.5 Mbp region which contains the YAP1 oncogene was observed (Figure 4C). This amplification was noted across CDK4 and PIK3CA driven tumors with approximately 40-50 copies of the gene present (Figure 4D). Notably, KRAS driven tumors did not show amplification or overexpression of YAP1 as compared to cell lines indicating that the development of these tumors is mechanistically distinct (Figure 4C). Models with endogenous (CDK4 and PIK3CA driven) and exogenous (YAP1 driven) amplification of YAP1 show similar YAP1 gene expression values, while the controls and KRAS driven samples are much lower (Figure 4D). It is possible that the YAP1 amplification occurred de novo during transformation or was present as a rare clone in the pre-transformation cells. Nevertheless, this selection for YAP1 signaling indicates it is a key tumor fitness-conferring pathway in sarcoma development and growth.

Transcriptome of YAP1 and KRAS Driven Tumors Overlaps Human Undifferentiated Pleomorphic Sarcoma and Myxofibrosarcoma

To investigate the transcriptome of the tumors formed in our models and make comparisons to human sarcoma samples, RNA sequencing was performed. In an unsupervised, pan-cancer approach, principal component analysis (PCA) was performed on gene expression data of all 33 TCGA primary diseases (Supplemental Figure 6A). To unbiasedly select a subset of TCGA cancers for comparison, Euclidean distances in the PC space were calculated to rank each cancer type by overall expression-based closeness or likeness to sarcoma (Supplemental Table 1C). To achieve a range of diseases, every fifth disease plus sarcoma was selected for further analyses. We performed PCA on the TCGA subset and projected the tumors and controls onto the PCA-defined gene expression space (Figure 5A). We observed that both the model tumors and pre-transformed control cells have RNA profiles that are close to sarcoma, indicating the transcriptome of the model reflects that of patient tumors. The TCGA sarcoma data does not separate when plotted by subtype (Supplemental Figure 6B). In addition to sarcoma, our system also overlapped with skin cutaneous melanoma which had ranked number one in distance to sarcoma in the pan-cancer analysis (Supplemental Table 1C). Euclidean distance calculations within this space further suggested that cutaneous melanoma and sarcoma were the most similar to our tumor models (Supplemental Table 1D). To investigate if the model system preferentially aligned with one of the two cancer types, the model tumors were projected onto a PCA space defined only by skin cutaneous melanoma and sarcoma (Supplemental Figure 6C). Our tumor models projected closer to sarcoma, suggesting their transcriptomes most closely reflect tumors of sarcoma patients (Supplemental Figure 6C–D and Supplemental Table 1E).

Figure 5. YAP1 and KRAS Driven Tumors Transcriptionally Reflect Human Undifferentiated Pleomorphic Sarcoma and Myxofibrosarcoma.

(A) PCA plot of human cancers from TCGA data with projection of the tumor models and cell lines onto the space showing overlap with human sarcomas. (B) PCA plot of tumor models and cell lines with projection of TCGA samples into the space demonstrating that the models cluster with tumors and are distinct from the cell lines. (C) PLSR plot of UPS and MFS transcriptome with KRAS and YAP1 models projected. (D) Co-rank plot showing overlapping signal in KRAS-YAP1 tumor model differentially expressed (DE) gene signature and TCGA MFS-UPS DE gene signature with Pearson correlation coefficient. (E) Co-rank plot showing overlapping signal in KRAS-YAP1 tumor model DE gene signature and TCGA MFS-UPS PCA signature with Pearson correlation coefficient. (F) GSEA comparing YAP1 tumors, human UPS, and PCA loadings (+) to KRAS tumors, human MFS and PCA loadings (−). (G) Distribution of oxidative phosphorylation and DNA damage related pathways in gene sets ranked by normalized enrichment score (NES). All listed categories are nominally significant (p<0.001) by Kolmogorov-Smirnov (KS) test.

As there was a significant overlap between the models and controls in the pan-cancer projection, we investigated the degree to which the model tumors were transcriptionally unique and reflected a disease state. PCA performed on the models and pre-transformed control cells revealed a significant difference between the two groups (Figure 5B). Projection of the TCGA tumor subset onto this space resulted in a tight cluster on the tumor model side. This result supports that the separation between tumors and controls along PC1 is due to the process of transformation and that these models transcriptionally reflect a disease state as compared to controls. It was also observed that while YAP1 and KRAS driven tumors co-occupy a space, CDK4 and PIK3CA driven tumors are transcriptionally distinct.

Further analyses were performed to characterize the CDK4 and PIK3CA driven models. At this time, we are unable to conclude that these models preferentially reflect leiomyosarcoma and osteosarcoma patient tumors respectively, but simply, most closely resemble sarcoma as a whole when compared to other cancer types (Figure 5A–B, Supplemental Figure 6C–D, Supplemental Table 1C–E). While these tumors were histologically distinct, it is possible that the areas sequenced may contain a mixed histology or these drivers histologically recreate the human disease while the transcriptome is not fully concordant. Gene expression signatures can guide future genetic and other adjustments to the model to improve recapitulation of core human tumor features.

Tumors driven by YAP1 and KRAS consistently demonstrate undifferentiated pleomorphic sarcoma (UPS) and myxofibrosarcoma (MFS). Morphologically, patient tumors from these subtypes are often difficult to distinguish from one another with both tumor types staining negative for markers of differentiation by immunohistochemistry. Previous efforts to molecularly differentiate UPS and MFS have been unsuccessful, suggesting these two phenotypes occupy a disease spectrum rather than two separate entities ⁵. To investigate if this spectrum is reflected in our models, MFS and UPS patient data were projected onto a partial least squared regression (PLSR) plot of the YAP1 and KRAS driven models. Qualitatively, the transcriptomes of the two models appear to sit at either endpoint of a UPS/MFS disease spectrum (Figure 5C). In support of the model’s capability to define a translationally relevant spectrum, patient samples exhibited similar distributions in both the model-defined PLSR space and in an unsupervised patient-defined PCA space (Supplemental Figure 7A–B). We additionally observed a subset of UPS samples that do not overlap with the MFS samples, termed “UPS-high,” in both the model and patient-defined spaces (Supplemental Figure 7A). Of note, the UPS-high samples projected closely to our YAP1 model, further supporting these models as end points of the disease spectrum.

Differential gene expression analysis was performed to further investigate the overlap between models and patient samples (Supplemental table 1F–G). A co-rank plot showing differential gene expression between UPS and MFS patients and between the YAP1 and KRAS driven models identified a statically significant overlap in the patient UPS-MFS and model YAP1-KRAS based signatures. Genes highly expressed in the YAP1-driven model were enriched for genes highly expressed in UPS patient tumors, and a corresponding enrichment was observed for the KRAS- driven model and MFS patient tumors (Figure 5D). Signature overlap analysis of the co-rank plot using rank-rank hypergeometric overlap (RRHO) yielded a statistically significant signal, further supporting that the YAP1 and KRAS driven models occupy endpoints of the UPS-MFS disease spectrum (Supplemental Figure 7C) ³⁹.

Further motivated by the challenge of distinguishing MFS and UPS samples histologically, we investigated if gene loadings from the patient-defined PCA space could provide an additional, unbiased and transcriptome-based signature to stratify samples (Supplemental Table 1H). A co-rank plot between patient-defined PC1 gene loadings and the model YAP1 versus KRAS-based signature resulted in an even higher degree of statistical overlap (Figure 5E). RRHO analysis yielded a strong signal, approximately three times the significance of the histology-based UPS versus MFS signature (Supplemental Figure 7D). Data from both the unbiased and model-based approach supports a spectrum-based approach to classifying tumors in the UPS and MFS subtypes. YAP1 and KRAS model tumors are clearly distinct from the starting untransformed cells and represent a transformed state with complex karyotypes and transcriptomes that recapitulate key aspects of the human disease state of these two distinct sarcoma subtypes.

Comparison of Model Transcriptome to TCGA Identified New Therapeutic Vulnerabilities

Gene set enrichment analysis (GSEA) revealed the YAP1 driven model and UPS patient tumors both showed significant enrichment of oxidative phosphorylation related pathways while the KRAS driven model and MFS patient tumors showed enrichment for DNA damage response related pathways (Figure 5F–G, Supplemental Figure 8A, Supplemental Table 1I–J). Similar enrichment was also observed in the PCA loadings signature (Figure 5F–G, Supplemental Figure 8B, Supplemental Table 1K). Further examination of the genes contributing to the oxidative phosphorylation pathway enrichment showed upregulation of multiple members of the NDUF (NADPH:Ubiquinone Oxioreductase) family which make up complex I of the electron transport chain. To determine if targeting of oxidative phosphorylation or YAP1 was clinically relevant, we treated a panel of high-grade soft tissue sarcomas with the Hippo signaling inhibitor Verteporfin or an oxidative phosphorylation inhibitor, IM156. Treatment of cell lines with either drug led to decreased viability (Figure 6A and Supplemental Figure 9A and 9B). YAP expression levels did not correlate to IC50 values, high YAP expression did not predict response (Figure 6B). We then asked if inhibition of both YAP1 and oxidative phosphorylation would have a synergistic effect in sarcomas. The drug concentrations tested were based on the average IC50 value across the cell lines. As seen in Figure 6C, treatment of cell lines with the combination of IM156 and verteporfin led to an additive effect in four out of five lines. Verteporfin has been shown to have off target effects outside of YAP1 inhibition⁴⁰. To address this, we have utilized a second YAP1 inhibitor, IAG933. Interestingly, this inhibitor did not have an impact on cell viability when used as a single agent. Yet, concomitant treatment of cells with both IAG933 and IM156 led to a significant decrease in cell viability in two verteporfin sensitive lines, GCT and UPS1, when compared to IM156 alone (Figure 6D). In addition, GCT cells with YAP1 KO by CRISPR demonstrated increased sensitivity to IM156 as compared to controls (Figure 6E). This was also noted in UPS2 KO cells but not SW872 (Supplemental Figure 9C). Interestingly, YAP1 could not be knocked out in UPS1 cells, this led to widespread apoptosis indicating YAP1 is essential for this cell line. This data indicates that targeting of Hippo signaling and oxidative phosphorylation inhibits the growth of a subset of sarcomas and should be further investigated.

Figure 6. Treatment of Sarcoma Cell Lines with Hippo, Oxidative Phosphorylation or Dual Inhibition.

(A) Top panel shows GCT (UPS cell line) proliferation with the addition of increasing concentrations of either the YAP1 inhibitor verteporfin or the complex I inhibitor IM156. Bottom panel provides relative IC50 values for each of the tested cell lines (also see supplemental figure 6). (B) Western blot of cell lines demonstrating YAP1 expression levels. (C) Treatment of sarcoma cell line panel with 30uM of IM156, 20uM of verteporfin, or the combination. *p<0.05 as compared to the DMSO only control. **p<0.05 as compared to IM156 and verteporfin alone. N=3 (D) GCT and UPS1 cells treated with YAP/TEAD inhibitor IAG933 and IM156 or the combination. Verteporfin treatment is shown as a comparator group. *p<0.05. N=3 (E) Treatment of control or YAP1 knockout GCT cell lines with IM156. *p<0.05. UPS1 KO lines could not be generated as cells did not tolerate YAP1 loss. N=4.

Discussion

High-grade sarcomas are a heterogeneous group of tumors that likely represent a spectrum of disease states rather than distinct entities. We have developed a system that allows for rigorous investigations into this complex set of malignancies and provides the opportunity to identify clinically relevant pathways and factors.

Modeling Human Sarcoma Development, Evolution, and Metastasis

Mesenchymal stem cells (MSCs), the presumed cell of origin for sarcomas have presented significant challenges when used as starting material for studies of human sarcomagenesis. Human MSCs have been historically difficult to transform compared to mouse MSCs presumably due to telomere maintenance and senescence ²². In addition, there is variability across human MSC populations reflecting the heterogeneity of the human population. To overcome these challenges, we utilized ASC52telo, an immortalized mesenchymal stem cell line as the starting material for transformation ²³. While there are inherent limitations to the use of immortalized cell lines, this allowed for reliable and reproducible results across a series of experiments. Interestingly, in our genomic analysis it was observed that this cell line has baseline aneuploidy which has been previously observed in human MSCs in culture ³⁸. This raises the possibility that these cells are shifted closer to a neoplastic fate, due to specific acquired changes or a general tolerance of genomic instability. Nevertheless, these cells are in a pre-transformed state as defined by the inability to grow as a tumor in vivo. In the transcriptome analysis it was observed that these cells are distinct from the tumors that form and as a result, provide a starting material with proper characteristics to study the critical steps of transformation.

From available literature, it is known that the process of sarcoma formation requires multiple hits ²². The MSC-based forward genetics transformation model recreates this aspect of sarcoma biology. Both the loss of a transcription factor and the addition of an oncogenic gene were required for transformation. This provides the opportunity to study the cellular evolution of sarcoma development. In this study, twenty-seven factors were added but this can be scaled to include large classes of genes, specific pathways, or the whole genome. Despite the heterogeneity of primary sarcoma tumors, they uniformly metastasize to the lung which is the primary cause of mortality and this phenotype was observed in the model ³⁰. We can use this system to address key clinical questions regarding the affinity of sarcoma to the lung, screen for drivers and regulators of metastasis and identify and test therapeutic targets.

The model system has limitations as well. The use of an immortalized cell line is not completely reflective of the human disease and limits the ability to expand this approach to other cell types such as mesenchymal progenitor cells that have been implicated in sarcoma development ^41,42. RB1−/−TP53+/− cells were noted to spontaneously transform in vivo at long latencies, complicating the evaluation of less robust drivers of sarcoma development that require a long latency to drive tumor formation. Within the model, we also noted that PIK3CA and CDK4 drive tumors that histologically represent sarcoma subtypes but, transcriptional analysis was not aligned with the human subtypes. Further studies are needed to understand and refine the development of osteosarcoma and leiomyosarcoma in this system. While the model has translational relevance, the use of human cells in an immunocompromised mouse limits the pre-clinical potential because immune based therapies cannot be tested. Despite these limitations, the ability to form multiple subtypes from the same starting material is a critical advance in sarcoma modeling and provides the opportunity to better understand the relationships between these subtypes.

YAP1 and KRAS are Drivers of Sarcoma Development

YAP1 and KRAS robustly transform MSCs into two common high-grade sarcomas, UPS and MFS. Both have been implicated in sarcoma growth and progression ^6,43. In a well characterized genetically engineered mouse model, expression of mutant kras G12D, in conjunction with p53 loss results in the development of high-grade sarcomas ⁶. YAP1 has been shown to interact with the oncogenic transcription factor FOXM1 to drive sarcoma proliferation ²⁶. In our model, tumors are either YAP1 or KRAS dominant but, further work to evaluate the connection and interdependence is needed. The relationship between YAP1 and KRAS has been extensively evaluated in pancreatic cancer. In this setting, mutant KRAS is dependent on YAP1 for tumor development and resistance to KRAS inhibitors is mediated through YAP1. This indicates that these pathways may have redundant effects. Extrapolating from other tumor types, YAP1 is known to promote stem cell properties by inhibiting differentiation and driving proliferation which may result in the undifferentiated pleomorphic sarcoma subtype. Interestingly, KRAS has been shown to be mutated or aberrant in a genetic characterization of human myxofibrosarcoma ⁴⁴. While KRAS has also been shown to regulate proliferation and promote a stem like state, activation of MAPK signaling is noted in myxoid liposarcomas and myxofibrosarcoma⁴⁵. This raises the possibility that downstream MAPK signaling from KRAS may promote myxoid development and deposition and further delineation of this pathway may lead to myxofibrosarcoma specific therapeutic options.

YAP1 Amplification in Sarcoma Formation and Growth

Aneuploidy is the hallmark of non-translocation driven sarcomas such as UPS and MFS. The recapitulation of aneuploidy and copy number variations in general, and the involvement of specific patient sarcoma-associated amplification events (YAP1), are two key features of this model. In cancer in general, aneuploidy has been implicated in key cellular processes such as tumorigenesis and metastasis⁴⁶. In sarcoma, the impact of aneuploidy has not been extensively studied but will provide insight into important clinical questions surrounding drug resistance and metastasis. The evaluation of aneuploidy in this model is complicated by the ASC52telo parental MSC line demonstrating a high degree of chromosomal changes compared to simple karyotype sarcomas. Further investigation into the evolution of aneuploidy in this model are needed. Interestingly, it was observed that chromosome 11q which contains YAP1 was amplified in tumors driven by CDK4 and PIK3CA and this amplification was not detected in the cell lines used as starting material. These results support a tumor fitness benefit of YAP1 amplification, with the amplification either occurring de novo or being selected for from a rare clone during the transformation process or during tumor growth. Based on the TCGA patient sarcoma data, there are a subset of human tumors that have amplified YAP1, and likely this is a key pathway driving the pathophysiology of a subset of high-grade sarcomas. YAP1 is a transcription factor in the Hippo signaling pathway that has been shown to regulate mesenchymal stem cell and sarcoma proliferation, motility, and differentiation ⁴³. Further investigation into the role of YAP1 in driving sarcoma development and transformation may provide new mechanistic insights.

Heterogeneity of Human Sarcomas is Reflected in the Model

The heterogeneity of sarcomas is multilayered. In addition to histology and patient related heterogeneity, intratumoral and intrasubtype heterogeneity also exist. This has been observed clinically with patients within one subtype such as undifferentiated pleomorphic sarcoma (UPS) having vastly different responses to the same therapies. It has been observed that sarcomas can contain more than one histology indicating they have undergone heterologous differentiation into another subtype. For example, a tumor that is primarily UPS but contains areas that are generating bone and appear histologically to be osteosarcoma ⁴⁷. This is not the case in the majority of patient tumors, but these cases are representative of the fact that sarcoma subtypes are likely a spectrum of disease with considerable plasticity⁵. In this model system, we see parallels to the heterogeneity that is seen across patients. While we observed UPS and MFS consistently in the YAP1 and KRAS driven tumors, in those driven by CDK4, PI3K, and JUN, multiple or mixed histologies were seen. This is either driven by the heterogeneity of the starting material or to the evolution of the sarcomas over time which is being addressed in ongoing studies. In this study we have also observed that multiple genes can give rise to the same histology, UPS which likely represents an additional disease spectrum. Understanding the heterogeneity of those tumors and if this will provide further insight and overlap onto the human UPS disease spectrum warrants further investigation.

Translational Potential of the Developed Model System

Within complex karyotype sarcomas, there are numerous genomic and transcriptomic changes and it is unclear which of these are driving the pathophysiology. The comparison of the tumors generated in the model to their patient counterparts allowed us to identify key pathways for sarcoma growth. We identified YAP1 and oxidative phosphorylation as relevant clinical targets in a subset of high-grade sarcomas. Currently, YAP/TAZ and TEAD inhibitors are being tested in clinical trials and have shown promising early data in mesothelioma and tumors driven by Hippo pathway dysregulation⁴⁸. Single agent targeted therapies often result in resistance, particularly in metastatic disease which is highly heterogeneous, making combination treatment a favorable approach pending tolerability of multiple agents. IM156, the oxidative phosphorylation inhibitor has been tested in a phase I trial and led to stable disease in 7 out of 22 patients with limited side effects⁴⁹. Based on the data from our model and the panel of high-grade sarcomas tested, these therapies could be effective for a subset of sarcoma patients warranting further investigation and consideration for early phase trials in localized tumors or as an adjuvant approach in high-risk patients. We are working to develop a method of predicting tumors that would respond to this treatment approach.

Through transcriptomic analysis comparing model tumors to patient tumors supported that YAP1- and KRAS-driven tumor lie near the endpoints of an unbiasedly defined spectrum of patient UPS and MFS tumors. The sarcoma community acknowledges that within each there are yet unidentified subsets or continuous spectrums that could direct treatment decisions. It is feasible that the identified YAP1-like UPS-high subset may be highly sensitive to both inhibition of Hippo signaling and oxidative phosphorylation. Using this model, we can continue to expand our analysis and identify and define additional subgroups within high-grade sarcomas that may have unique vulnerabilities. While further perspectives and comparisons will need to be investigated, both unbiased and model-based transcriptome analyses stand to improve the current practice of classifying tumors in the undifferentiated MFS and UPS subtypes, and support a spectrum-based rather than binary-based approach.

Here, we have generated four histologic sarcoma subtypes from a single cell of origin, undifferentiated pleomorphic sarcoma, myxofibrosarcoma, leiomyosarcoma, and osteosarcoma indicating that these subtypes have common underlying biology and may exist together on a disease spectrum. This model allows for additional insights into sarcoma biology and more broadly, can serve as a system for those that are tumor agnostic to understand critical aspects of cancer biology. The opportunities for further investigation and studies of high-grade complex karyotype sarcoma biology are vast and this preclinical model will serve as the basis to understand and target pathways and mechanisms driving the pathophysiology of this devastating disease.