Genomically Complex Human Angiosarcoma and Canine Hemangiosarcoma Establish Convergent Angiogenic Transcriptional Programs Driven by Novel Gene Fusions

This article is featured in Highlights of This Issue, p. 741

Sarcomas are diverse tumors that arise from cells of mesenchymal origin in soft tissues such as blood and lymphatic vessels, fat, bone, cartilage, muscle, and connective tissues. The heterogeneity of sarcomas has provided an impetus for developing molecular approaches to classify these tumors (1, 2), leading to their categorization into genomically simple and genomically complex sarcomas (1, 3). Angiosarcomas are rare, highly aggressive, genomically complex sarcomas of blood vessel–forming cells (3, 4). The five-year survival rate of angiosarcoma is approximately 40% (5–7), but half of patients have metastatic or unresectable disease with a median overall survival of less than 6 months (8). The events that drive progression are incompletely understood; previous studies have identified recurrent mutations of RAS, PTPRB, PLCG1, KDR (kinase insert domain receptor, also known as VEGFR2), TP53, PIK3CA, and FLT4 (VEGFR3) in human angiosarcomas (9–12). MYC gene amplification and alterations in the TP53, CDKN2, NF-κB/IL-6, and PIK3CA/AKT/mTOR pathways have also been reported (13); however, these studies represent a small case series, precluding definitive conclusions regarding pathogenic mechanisms that contribute to the genetic cause and to the progression of the disease.

Hemangiosarcoma is a malignant vascular tumor that is common in dogs with an estimated tens of thousands of cases diagnosed each year (14–16). Canine hemangiosarcoma shares clinical and morphologic features with human angiosarcoma, as well as aspects of its mutational landscape (17–20). We previously documented three molecular subtypes of hemangiosarcoma, characterized by angiogenic, inflammatory, and adipogenic transcriptomic signatures (21). These gene-expression signatures are conserved in hemangiosarcoma progenitor cells that show multipotency and self-renewal (21). Nevertheless, the transcriptional state of these hemangiosarcoma progenitor cells seems to be somewhat malleable, regulated by immune and metabolic reprogramming (22). Mutations in genes that regulate genomic integrity, such as TP53, can alter the intrinsic transcriptional program of tumor cells; however, genomic instability in the tumor can create even more dramatic changes by modulating transcriptional programs of heterotypic stromal cells in the tumor tissue, as well as in the composition of the niche (23, 24).

Chromosome translocations and the resulting fusion genes are important contributors to the pathogenesis of cancer, particularly in sarcomas and hematopoietic malignancies (25). However, the nature and frequency of these events in canine hemangiosarcoma and human angiosarcoma remains unclear. Here, we used next generation RNA sequencing (RNA-seq) data to identify fusion genes in 13 human angiosarcomas and 76 visceral hemangiosarcomas originating from 74 dogs, and we investigated the relationship of these fusions to the mutational landscape of the tumor. We identified ten novel protein-coding fusion genes including TEX2-PECAM1 and ATP8A2-FLT1 in 7 of 13 human angiosarcomas, and two of the fusion-detected tumors showed mutations of TP53 (R248Q and P250L). In canine hemangiosarcomas, we found novel protein-coding fusion genes in a subset of the tumors of the angiogenic subtype. These fusion genes cooccurred with TP53 mutations and were associated with gene enrichment for activated angiogenic pathways in the tumors. Our data suggest that genomic instability induced by mutations of TP53 creates a permissive environment for fusion genes, with selection for angiogenic molecular programs in malignant vasoformative tumors. Our data also demonstrate that human angiosarcoma and canine hemangiosarcoma maintain molecular programs that activate convergent signaling pathways to establish angiogenic phenotypes despite their genomic complexity.

Snap frozen and formalin-fixed paraffin-embedded (FFPE) tissues for human biospecimens were obtained from the University of Minnesota Biological Materials Procurement Network (UMN BioNet) and from the Cooperative Human Tissue Network (CHTN) under their standardized patient consent protocols. The demographic characteristics of human patients from whom we obtained angiosarcomas (n = 13) and normal tissue samples (n = 6) are summarized in Supplementary Table S1.

Seventy-six snap frozen and FFPE tissue samples were obtained from 74 dogs with hemangiosarcomas. Frozen and FFPE tissues samples from 10 dogs with splenic hematomas, which are benign lesions with enlarged vascular spaces lined by endothelial cells, were used as controls. Samples were obtained as part of medically necessary diagnostic procedures and were used for research with owner consent. The origin of these samples was reported previously (14, 21, 26–28), or they were collected from dogs with hemangiosarcoma or with splenic hematomas at the Veterinary Medical Center, University of Minnesota (St. Paul, MN). Procedures involving animal use were done with approval and under the supervision of the University of Minnesota Animal Care and Use Committee (protocols 1110A06186, 1507–32804A, 0802A27363, 1101A94713, 1312–31131A, and 1702–34548A). The demographic characteristics of dogs (n = 74) from whom we acquired hemangiosarcoma and nonmalignant splenic hematomas (n = 10) are summarized in Supplementary Table S2.

FFPE sections (4 μm) were stained with hematoxylin and eosin (H&E) and examined by veterinary pathologists to assign a histologic diagnosis of canine hemangiosarcoma. Solid, capillary, cavernous, or mixed histologic subtypes were assigned using accepted criteria (29); mitotic index (MI) was calculated per 1,000 cells in 5 to 10 random fields under 400× magnification (30). H&E slides were further reviewed for tumor content by two board-certified medical pathologists (M.A. Linden and P. Murugan; ref. 31), with the percent of sample containing viable nucleated cells corresponding to tumor recorded in a range of 0 to >90% based on the planar surface of the sections. Diagnostic and histopathology reports of human tissues were provided by the specimen providers, the UMN BioNet and the CHTN.

Total RNA was isolated from tissue samples using the TriPure Isolation Reagent (Roche Applied Science). The RNeasy Mini Kit (Qiagen) was used for clean-up according to the manufacturer's instructions. RNA-seq from 74 canine hemangiosarcoma tissues is published (17, 21, 32, 33) and an additional dataset was generated from two canine hemangiosarcoma tissues and from 10 nonmalignant splenic hematoma tissues. Total RNA was also extracted from 13 human angiosarcoma tissues and from 6 normal tissues. Two micrograms of total RNA from each sample were quantified and assessed for quality; RNA-seq libraries were generated as described in ref. 21 using the TruSeq RNA sample preparation kit (Illumina). Sequencing was performed using HiSeq 2000 or 2500 systems (Illumina). Each sample was sequenced to a targeted depth of 20–80 million paired-end reads with mate-pair distance of 50 bp. Primary analysis and demultiplexing were performed using CASAVA software version 1.8.2 (Illumina) to verify the quality of the sequence data. The end result of the CASAVA workflow was demultiplexed into FASTQ files for analysis. Bioanalyzer quality control and RNA-seq were performed at the University of Minnesota Genomics Center (UMGC) or at the Broad Institute.

The original FASTQ files prepared from thirteen human angiosarcomas and six nonmalignant tissues were mapped to the human reference genome (GRCh38). The FASTQ files generated from 76 canine hemangiosarcomas and 10 nonmalignant splenic hematomas were mapped to the dog reference genome (Canfam3.1). Sequencing quality was assessed by FastQC. The deFuse algorithm (34) was used to identify putative fusion events. To discriminate true fusion candidates from artifacts, we included fusion events with exon boundaries in both fusion partners and excluded events created from adjacent genes that showed breakpoint homology (>1). We also filtered highly recurrent fusion events that were found at implausible frequencies across tumor and nonmalignant tissue samples (35) and transcription-induced chimeras. The split sequences of the fusion genes were validated by de novo assembly using Trinity (36). TranscriptsToOrfs and deFuse-Trinity tools verified the deFuse fusion predictions with Trinity-assembled transcripts and open reading frames. TopHat2 was used to generate BAM files, and the Integrative Genomics Viewer (IGV 2.3; Broad Institute) was used to visualize the mate pair sequences of fusion genes. A protein translation tool in Expert Protein Analysis System (ExPASy; SIB Swiss Institute of Bioinformatics, Lausanne, Switzerland) was used to determine in-frame fusion proteins. Tumor purity and microenvironment scores were assessed using the bioinformatics tools ESTIMATE (37) and xCell (38).

Reverse transcription (RT-PCR) was performed to validate fusion transcripts identified by deFuse (39). Briefly, cDNA was synthesized using SuperScript VILO cDNA Synthesis Kit and Master Mix (Invitrogen). PCR amplification was performed using a conventional thermocycler with HotStarTaq DNA polymerase (Qiagen) or using a LightCycler 96 (Roche Applied Science) with FastStart SYBR Green Master Mix (Roche Applied Science) for quantitative real-time RT-PCR (40). PCR primer pairs used for fusion gene amplification are presented in the Results section. GAPDH was used as a control for RNA integrity and for the RT-PCR reactions. The forward and reverse primer sequences for GAPDH were 5′-GGA GTC CAC TGG CGT CTT CAC-3′ and 5′-GAG GCA TTG CTG ATG ATC TTG AGG-3′, respectively. Relative mRNA values were expressed as delta-C_t values normalized to GAPDH. Sanger sequencing was performed at the UMGC.

Fluorescence in situ hybridization (FISH) was performed to detect MYO16-PTK2 and GABRA3-FLT1 fusion genes by designing FISH probes derived from the genome-anchored canine CHORI-82 bacterial artificial chromosome (BAC) library (41). Single locus probes were used for proximal MYO16 at dog chromosome (CFA) 22:57,565,917–57,750,789 (clone 183H20), distal MYO16 at CFA 22:57,750,801–57,967,880 (clone 385H13), and PTK2 at CFA 13:35,302,679–35,483,060 (clone 451H13) with distinct fluorescent tags. For GABRA3-FLT1 fusion, break-apart FISH probes were used for proximal FLT1 at CFA 25:11,057,892–11,263,935 (clone 363B20) and distal FLT1 at CFA 25:11,274,078–11,471,538 (clone 235H9). The PureLink HiPure Plasmid Maxiprep Kit (Invitrogen) was used for BAC DNA extraction. For preparation and hybridization of FISH probes, BAC DNA probes were labeled by Nick Translation Kit (Abbott Molecular) using Green-500 dUTP, Orange-552 dUTP and Aqua-431 dUTP (Enzo Life Science). Labeled DNA was precipitated in COT-1 DNA, salmon sperm DNA, sodium acetate and 95% ethanol, then dried and resuspended in 50% formamide hybridization buffer. The Red-proximal MYO16, Green-distal MYO16 and Aqua-PTK2 probes were combined into one 3-color FISH probe for MYO16-PTK2 fusion. The Red-proximal FLT1 and Green-distal FLT1 break-apart probes were applied for the split FLT1 gene.

FFPE sections (4 μm) were processed according to the Dako IQFISH protocol; probes were applied to the slide and hybridized for 24 hours at 37°C in a humidified chamber. After hybridization, slides were washed and counterstained with DAPI. Fluorescent signals were visualized on an Olympus BX61 microscope workstation (Applied Spectral Imaging) with DAPI, FITC, Texas Red and Aqua filter sets. FISH images were captured using an interferometer-based CCD cooled camera (ASI) and FISHView ASI software. A total of 200 interphase cells were examined for each sample. Nonmalignant canine spleen tissues were used as controls for the FISH experiment.

A pipeline was developed to identify the bases present at locations defined as somatic mutations in the Tumor-Normal Exome calls (17). Briefly, RNA-seq data were mapped using the STAR-Mapper (42) with STAR-FUSION mapping settings (43) to the Canfam3.1 or GRCh38 genome. BAM files generated by STAR were sorted and indexed using Samtools (44). Starting from a file containing somatic mutation locations and a file containing a list of BAM file locations, the pipeline uses Samtools (44) functions to identify the bases present at each location at each file and then reports whether a variant is found at that location. For inclusion within the variant file, at least one sample must have at least three reads that support the variant and that represent greater than 10% of all reads at that location.

CLC Bio Genomics Workbench 10 (CLC Bio) and DESeq2 R packages were used to quantify gene expression and differential gene expression analysis as described previously (45). Briefly, paired-end RNA-seq data with mate-pair distance of 50 bp in FASTQ format were imported, and sequencing quality was determined. Transcriptomics analysis was then performed to generate the expression level of each gene presented as total reads by mapping the sequencing reads to Canfam3.1 or GRCh38. Heatmaps and hierarchical clustering based on average linkage were visualized using Cluster 3.0, Morpheus (https://clue.io/morpheus), or R packages. GO Enrichment Analysis (46–48) or Ingenuity Pathway Analysis software version 8.6 (Qiagen) were used to define biological functions, canonical pathways, and upstream regulators associated with differently expressed genes (DEG) between groups using Benjamini–Hochberg multiple testing corrections to evaluate significance. For gene-expression profiling, unsupervised principal component analysis (PCA) and hierarchical clustering were performed to define subtypes of canine hemangiosarcomas as described previously (21). Gene-expression data of human sarcomas in The Cancer Genome Atlas (TCGA) database were also compared with our data sets.

Canine tissue microarray (TMA) blocks were generated from 45 hemangiosarcoma tissues, including 32 tumors used for RNA-seq and eight nontumor tissues (six splenic hematomas and two nonmalignant liver samples). Tissue cores of 1-mm diameter in quadruplicate from each sample were assembled in random order in four TMA blocks. One TMA block with mouse tissues was generated for staining controls. Immunostaining with CD31, Vimentin and pan-Cytokeratin antibodies was evaluated to support tumor content estimates. A human TMA block was generated from ten angiosarcoma tissues and six nonmalignant tissues (submandibular gland, skin, breast adipose tissue, thigh skeletal muscle, spleen, and lung).

Unstained TMA sections (4 μm) were deparaffinized and rehydrated using standard methods for IHC. All of the IHC assays, including validation for antibodies, were performed and optimized at the UMN BioNet Histology Laboratory or the Veterinary Diagnostic Laboratory at the University of Minnesota. Antibodies used for IHC are summarized in Supplementary Table S3. The immunostaining score assigned to each case was a semiquantitative assessment derived from the product of two integers, ranging from 0 to 3 and from 1 to 3, that, respectively, reflect the percentage of positive cells in a sample and the intensity of staining at high-power magnification (400×) as described previously with some modifications (49). The percentage of positive cells was scored from 0 to 3+, where 0 reflected specific staining in <1% of the cells, 1+ reflected specific staining in >1% and <25% of the cells, 2+ reflected specific staining in 25%–75% of the cells, and 3+ reflected specific staining in >75% of the cells. The intensity was assessed as weak (intensity score 1), moderate (intensity score 2), or strong (intensity score 3). Immunostaining results were scored (ranging from 0 to 9) by multiplying the percentage of positive cells (score 0–3) by the intensity (score 1–3).

χ² or Fisher exact test was performed for contingency tables analysis. Continuous values were analyzed by Welch (Heteroscedastic) t test or Mann–Whitney U test. The statistical tests were two-tailed. Statistical analysis was performed using GraphPad Prism 6 (GraphPad Software, Inc.). P values are reported without inference of significance, consistent with the American Statistical Association's Statement on statistical significance and P values (50).

RNA-seq gene expression data generated from human sarcomas are available from the TCGA Research Network (https://www.cancer.gov/tcga). Exome-sequencing data from human angiosarcomas are available from The Angiosarcoma Project (https://ascproject.org), a project of Count Me In (https://joincountmein.org/). RNA-seq data from human angiosarcoma tissues are available through the Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/; accession number GSE163359). RNA-Seq data from canine hemangiosarcoma tissues are published (17, 21, 32, 33) and available through the GEO (accession number GSE95183) and the NCBI Sequence Read Archive (accession number PRJNA562916). All other data generated from this study are available upon request to the corresponding author.

Putative fusion gene events were identified from RNA-seq data as paired-end sequence reads that mapped connecting two distant genes (Supplementary Fig. S1). We identified novel in-frame protein-coding fusion transcripts for ten fusion events in 7 of 13 (53.8%) human angiosarcomas (Fig. 1A; Table 1). Two of the fusions were inter-chromosomal events and eight were intrachromosomal events. The fusions included TEX2-PECAM1, which contained the gene that encodes CD31, and ATP8A2-FLT1, a kinase fusion gene that encodes the VEGF receptor 1 (VEGFR1). None of the ten fusion events was seen in more than one tumor, and three of the seven fusion-positive tumors contained two distinct fusion events each.

In canine hemangiosarcoma, we found fifteen novel protein-coding fusion genes in eleven of 76 tumors (14.5%; Fig. 1B; Table 2). Ten of the fusions were interchromosomal events and five were intrachromosomal events. None of the fifteen fusion events was seen in more than one tumor, and four of the 11 fusion-positive tumors involved two distinct fusion genes each. One fusion partner in four of the translocations encoded either a protein kinase or a protein phosphatase associated with angiogenic signaling (MYO16-PTK2, GABRA3-FLT1, AKT3-XPNPEP1, and PTPRB-NOL10). Eight of the fusion genes were associated with kinase signaling or kinase binding activity, such as PI3K signaling, a MAPK, a receptor tyrosine kinase, a protein serine/threonine kinase, and an NAD+ kinase. Gene ontology annotations of the fusion partners for every translocation are described in Supplementary Table S4 for human angiosarcoma and Supplementary Table S5 for canine hemangiosarcoma. Fusion genes were not present in any of the human nonmalignant tissues (n = 6) or canine hematomas (n = 10) examined. Conserved driver translocations such as BCR-ABL and MYC-IGH that are present in both human and canine chronic myelogenous leukemia and Burkitt lymphoma, respectively (51), were not identified in human angiosarcoma and canine hemangiosarcoma. All the fusion events identified in the seven human angiosarcomas and eleven canine hemangiosarcomas involved different gene pairs, with the exception of the FLT1 gene, which created fusions with a different partner gene in one case from each species.

To determine whether the non-tumor components in the tumor tissue affected the detection of fusion genes, we quantified tumor content histologically and bioinformatically in canine hemangiosarcomas (Supplementary Fig. S2A–S2E). Seventy of the 76 hemangiosarcoma samples were histologically evaluated, and the tumor content was not different between hemangiosarcoma samples with fusion events (n = 11) and those without fusion events (n = 59). We used two independent bioinformatic tools, xCell and ESTIMATE, to predict stromal and immune cell components, and these algorithms generated consistent output scores (Pearson R = 0.84; R² = 0.71; P < 0.00001) that showed the presence of fusion genes was not associated with tumor purity. The detection of fusion genes was also independent of sequencing depth (Supplementary Fig. S2F). To rule out artifacts from the computational process, we validated the presence of the interchromosomal fusion gene, SCLT1-NIPBL, in the original human angiosarcoma sample where it was identified, in two additional samples where it was undetectable based on sequencing data, and in a nonmalignant tissue sample. The fusion transcript was detectable by quantitative real time RT-PCR amplification; PCR primer pairs were designed to amplify putative split sequences (up to 200 base pairs) involving the breakpoints identified by deFuse (Supplementary Fig. S3A). We confirmed that the junction sequences between the two genes producing the new fusion event were amplified by PCR (Supplementary Fig. S3B). Four representative fusion transcripts found in canine hemangiosarcomas (MYO16-PTK2; AKT3-XPNPEP1; AP4E1-BAIAP2; NOL10-PTPRB) were also detected, but only in the respective cases where they were identified in the sequencing data (Supplementary Fig. S3C–S3E). Each PCR amplification product was verified by Sanger sequencing. We then used RT-PCR to evaluate RNA-seq data from 63 canine tissue samples (53 hemangiosarcomas and 10 hematomas) for the presence of these four fusion transcripts. The results were consistent between RNA-seq and PCR, as we found neither false-positive nor false-negative events in the samples tested (Supplementary Table S6).

We then determined whether any of the fusion partner genes identified in our analysis were associated with DNA copy number alterations. Publicly available whole Exome-sequencing data generated from an independent data set of 36 human patients with angiosarcomas was used (12). Copy-number variations were found in twelve of the twenty (60%) fusion partner genes: nine genes were amplified, and five genes were deleted (Fig. 1C; Supplementary Fig. S4). TEX2 (39%), STEAP1B (25%) and PECAM1 (25%) were the top three genes where copy number gains occurred most frequently. For canine hemangiosarcoma, we used oligonucleotide array comparative genomic hybridization (oaCGH) in a larger hemangiosarcoma dataset (n = 123; ref. 19). Copy number gains were observed in 29 of the 30 (96.7%) fusion partners, and copy number losses were observed in 27 of the 30 (90.0%) fusion partners. Protein kinase–encoding genes, PTK2, FLT1, and AKT3 revealed a higher frequency of copy-number gain: 15.5% gain versus 0.8% loss for PTK2; 4.9% gain versus 0.8% loss for FLT1; and 4.1% gain versus 0.8% loss for AKT3, suggesting that copy-number alterations leading to dysregulation of downstream kinase signaling contribute at least partly to the angiogenic program in a subset of canine hemangiosarcomas (Fig. 1D).

We performed FISH to confirm that fusion genes were generated by chromosome translocations. We chose representative interchromosomal fusions, MYO16-PTK2 and GABRA3-FLT1 for cytogenetic validation because PTK2 and VEGFR1 are key molecules that regulate pathogenic signaling in vascular cancers, including canine hemangiosarcoma (52). Figure 2A illustrates the predicted structure of MYO16-PTK2 interchromosomal fusion between CFA 22 and CFA 13, based on deFuse and Sanger sequencing data (Fig. 2B). The predicted fusion gene comprises exons 1–32 of MYO16 (CFA 22) and exons 12–31 of PTK2 (CFA 13), with the putative junction joining MYO16 exon 32 and PTK2 exon 12. Breakage occurs between exons 32 and 33 of MYO16 at CFA 22:57,750,807 bp, and between exons 11 and 12 of PTK2 at CFA 13:35,397,284 bp. Independent FISH probes identifying the association between proximal and distal MYO16 and the breakpoint of PTK2 (Fig. 2C) confirmed the presence of the MYO16-PTK2 fusion gene between CFA 22 and 13 in archival FFPE samples from the same dog tumor. The MYO16-PTK2 fusion was identified by deFuse and RT-PCR (Fig. 2D). The t(CFA 13;CFA 22) translocation was present in interphase nuclei of 16.8% of the tumor cells, with a smaller subpopulation showing amplification of the fusion (Fig. 2E).

We used break-apart FISH to validate the presence of the GABRA3-FLT1 fusion gene (Fig. 2F). Split FLT1 probes were found in 36.7% of tumor cells in archival FFPE samples from the dog tumor in which the GABRA3-FLT1 fusion was identified by deFuse and RT-PCR. Interestingly, in this tumor the intact FLT1 gene showed consistent amplification (up to four copies), suggesting Flt-1 (also known as VEGFR1) activation in this tumor might have occurred through multiple mechanisms (Fig. 2G). We next used FISH analysis to assess recurrence of the MYO16-PTK2 fusion in a tissue microarray TMA comprised of 45 visceral hemangiosarcomas and eight nonmalignant tissues (six spleens; two livers). The MYO16-PTK2 fusion was once again present in the sample from the canine tumor in which it was discovered, but it was not seen in any other sample on the TMA. We also performed FISH to detect the ATP8A2-FLT1 fusion in human angiosarcoma using a break-apart FLT1 probe, but the fusion was undetectable in our FFPE sample. This might have been due to the small number of tumor cells that were likely to contain the fusion event in a heterogeneous clonal population. Since none of the fusion transcripts identified in our cohorts of human and canine tumors were recurrent, we sought to determine whether the fusion events were associated with other genetic and molecular programs.

To examine genomic aberrations associated with the fusion genes, we determined somatic variations and gene-expression profiles using RNA-seq data. In human angiosarcomas, TP53 mutations (R248Q and P250L) were observed in two of thirteen human angiosarcomas, which also had fusion genes (SMURF1-TMEM139 and AGO2-TRAPPC9 in one tumor, IRF9-THTPA in the other tumor). NRAS (Q61L; n = 1) or HRAS (Q61L; n = 1) mutations were also detected, and both were present in tumors that did not have fusion genes or TP53 mutations (Fig. 3A; Table 1). The RNA-seq data did not provide evidence of mutations in PIK3CA, PTEN, or KRAS in this group of thirteen angiosarcomas.

We established transcriptomic profiles of human angiosarcomas to identify molecular traits that regulate global gene expression. We identified 1,237 differentially expressed genes between angiosarcomas (n = 13) and nonmalignant controls (N = 6; FDR P < 0.05): 490 genes were upregulated and 747 genes were downregulated in angiosarcomas. Biological functions and pathway analysis revealed that upregulated genes in angiosarcomas were associated with cancer, angiogenesis, vasculogenesis, and development of vasculature (P < 0.00001; Supplementary Table S7). In addition, we performed cell type enrichment analysis using the xCell tool (38) to predict relative populations of cellular components that comprised the angiosarcoma tissues (Supplementary Fig. S5). We compared the cell type signature of angiosarcoma with that of sarcomas (n = 263) from the TCGA database, which did not include angiosarcoma. The results showed that gene signatures associated with endothelial cells and activated dendritic cells were highly enriched in angiosarcomas, while other sarcomas in the TCGA revealed gene enrichment of smooth muscle cells. Gene-expression profiles of nonmalignant samples indicated distinct tissue-specific patterns of submandibular gland, skin, breast adipose tissue, thigh skeletal muscle, spleen, and lung. The angiosarcomas showed upregulation of key angiogenic genes such as PECAM1 (CD31), FLT1 (VEGFR1), KDR (VEGFR2), and FLT4 (VEGFR3) compared with nonmalignant tissues (Fig. 3B). In addition, 6 of 20 (30.0%) fusion partners (including PECAM1) showed a higher level of expression in the tumors compared with nonmalignant tissues (P < 0.05; fold change in a range from 1.9–35.0; Fig. 3C). Collectively, our data showed enriched angiogenic molecular programs were present in human angiosarcomas. Furthermore, both tumors with TP53 mutations also harbored fusion genes.

In canine hemangiosarcomas, TP53 (n = 24/74; 32.4%), PIK3CA (n = 16/74; 21.6%), PIK3R1 (n = 5/74; 6.8%), and NRAS (n = 4/74; 5.4%) transcripts showed recurrent mutations that were consistent with those identified using tumor:normal Exome-sequencing (17). We found associations between mutations of TP53 (TP53^mt) and PIK3CA (PIK3CA^mt) and fusion genes (Fusion⁺) (Supplementary Table S8). Specifically, TP53^mt commonly cooccurred with PIK3CA^mt (P = 0.004) and with fusion genes (P = 0.004); and Fusion⁺ tumors were seen in the tumors with TP53^mt or PIK3CA^mt (P = 0.005) more frequently than would be expected by random chance. When fusion genes cooccurred with PIK3CA^mt, they invariably cooccurred with mutations of TP53 (P = 0.037), and they were not associated with PIK3CA^mt alone (P = 0.324).

Next, we sought to determine whether fusion genes were preferentially associated with specific molecular subtypes of canine hemangiosarcoma. We previously defined distinct angiogenic, inflammatory, and adipogenic molecular subtypes of canine hemangiosarcoma (21). To further validate this classification, we applied unsupervised PCA and hierarchical clustering to identify distinct groups in the sample cohort from this study (Supplementary Fig. S6). Our results show that the three molecular subtypes (58 angiogenic, 14 inflammatory, and 4 adipogenic) were reproducibly identified in the current dataset, as illustrated in the heatmap of 1,477 DEGs (FDR P < 0.001; fold change > |3|) shown in Fig. 3D. Interestingly, fusion genes were present only in tumors of the angiogenic hemangiosarcoma subtype (P = 0.046). Likewise, TP53 mutations were identified in 23 of 56 (41.1%) angiogenic hemangiosarcomas, and in one of 18 tumors from the two other molecular subtypes (P = 0.008). The somatic variants of TP53, PIK3CA, PIK3R1, and NRAS were found in 33 of 56 (58.9%) angiogenic hemangiosarcomas, and in 3 of 18 (16.7%) tumors from the two other subtypes (P = 0.002). The angiogenic subtype of hemangiosarcoma also showed upregulation of PECAM1 (CD31), FLT1 (VEGFR1), KDR (VEGFR2), and FLT4 (VEGFR3) compared with the other two hemangiosarcoma subtypes and to nonmalignant hematomas (Fig. 3E). We found that five of 32 dogs (16%) with angiogenic hemangiosarcoma lived longer than 5 months, while four of 10 dogs (40%) with inflammatory hemangiosarcoma survived longer than that (Supplementary Fig. S7). Eighteen of 30 (60.0%) fusion partners, including protein kinase–encoding genes such as FLT1, PTK2, and AKT3, showed higher levels of expression in hemangiosarcomas compared with nonmalignant controls (P < 0.05; fold change in a range from 1.3 to 6.0; Fig. 3F). Neither breed, sex, neuter status, age, nor affected organs were associated with the presence of fusion genes (Supplementary Fig. S8). There was also no association between the fusion events and histologic subtype or mitotic index (Supplementary Table S9).

Next, we analyzed DEGs (FDR P < 0.05; fold change > |2|) and gene pathways to examine gene signatures enriched in hemangiosarcomas that had both fusion genes and TP53 mutations. We classified tumors according to their mutations as summarized in Supplementary Table S10. Cooccurrence of fusion genes with TP53 mutations (i.e., TP53^mt/Fusion⁺/PIK3CA^wt or PF tumors) was associated with angiogenic and vascular signaling with enrichment of genes in pathways such as PI3K, VEGF, and PDGF (Supplementary Tables S11–S13). Thirteen genes that were commonly enriched in PF tumors were associated with activation of WNT3A as an upstream regulator (Supplementary Fig. S9). Figure 4 illustrates a model integrating the data from these findings to highlight potential pathogenetic contributions of fusion genes and recurrent mutations in canine hemangiosarcoma.

To determine whether the genetic and molecular features of human angiosarcoma and canine hemangiosarcomas contributed to the activation of functional pathways, we performed IHC in eleven human angiosarcoma tissues and in 44 canine hemangiosarcomas (Fig. 5; Supplementary Tables S14 and S15). First, we evaluated the effects of TP53 mutation on the presence and location of p53, phospho-p53 (Ser15), and phospho-p53 (Ser20) (Fig. 5A and B). In human angiosarcomas, nuclear expression of p53 was found in all of eleven (100%) tumors, showing various levels of expression. Nuclear expression of phospho-p53 (Ser15) was seen in seven of eleven (63.6%) tumors, showing low expression in five of seven (71%) tumors (IHC score ≤ 3). Nuclear and cytoplasmic expression of phospho-p53 (Ser20) protein was detected in all eleven (100%) tumors, and seven of those showed high expression (IHC score ≥ 7). In canine hemangiosarcomas, p53 protein was localized to the nucleus in 34 of 44 (77%) tumors with various levels of expression. Immunoreactivity of phospho-p53 (Ser15) was observed in the nuclei of tumor cells in 38 of 40 (95%) cases, with 34 (90%) showing low or medium expression (IHC score ≤ 6). Nuclear and cytoplasmic expression of phospho-p53 (Ser20) was seen in all of 40 hemangiosarcomas (100%), with 32 tumors (80%) showing high levels of expression. These data revealed that patterns of p53 and activated p53 were comparable in human angiosarcomas and canine hemangiosarcomas; especially, p53 was strongly phosphorylated at residue Ser20 in tumors from both species. We found no association between phosphorylated p53 and TP53 mutations or fusion genes (Supplementary Fig. S10), suggesting that DNA damage and cellular stress are widespread among these tumors, and they are likely to activate p53-mediated repair mechanisms independent of these genetic alterations.

The PI3K/AKT/mTOR signaling pathway is important for regulation of angiogenic, vascular, and energetic functions. To assess whether PI3K mutations resulted in higher levels of downstream pathway activation, we evaluated expression of AKT and phospho-AKT proteins in human angiosarcomas and canine hemangiosarcomas (Fig. 5C and D). Expression of nuclear and cytoplasmic AKT protein was observed in all eleven human angiosarcomas (100%) with 8 of 11 (73%) tumors showing medium or high expression. Expression of phospho-AKT (Thr308) was evaluated in 8 tumors; all of them (100%) showed weak or medium levels of expression. Similarly, AKT protein was detectable in the nucleus and cytoplasm of all forty (100%) canine hemangiosarcomas with 37 (93%) showing medium or high expression. Phospho-AKT (Thr308) was detected in the nuclei and cytoplasm of all evaluable 39 (100%) canine hemangiosarcomas, with 34 (87%) expressing medium or high level of the protein. Strong AKT immunoreactivity was also seen in scattered stromal cells in both human hemangiosarcoma and canine hemangiosarcoma. Neither mutations of TP53, PIK3CA, PIK3R1 nor the presence of fusion genes were associated with expression of AKT and phospho-AKT in human or canine tumors (Supplementary Fig. S11). Because PIK3CA and PIK3R1 mutations were undetectable in this set of 11 human angiosarcomas, we examined expression of mTOR and phospho-mTOR (Ser2448) proteins as surrogates to confirm activation of their downstream pathways in these tumors. mTOR protein was observed in the nuclei and cytoplasm of all eleven tumors (IHC score ≥ 6), and it was not associated with the presence of TP53 mutations or fusions (Supplementary Fig. S12A and S12B). However, nuclear and cytoplasmic expression of phospho-mTOR was higher in angiosarcomas that had TP53 mutations or that had fusion genes than it was in tumors without one of these genetic changes (Supplementary Fig. S12C and S12D). In summary, the immunostaining data suggest that human hemangiosarcoma and canine hemangiosarcoma have comparable activation of the p53 and PI3K/AKT/mTOR pathways, and these events are largely independent of their mutational states. Our results further suggest that these vasoformative tumors from both species activate convergent signaling pathways that contribute to their final architecture and organization with predictable enrichment of angiogenic gene signatures.

For this study, our objective was to identify novel fusion genes in human angiosarcomas and spontaneous canine hemangiosarcomas. We showed that novel protein-coding fusion genes were identified in approximately 50% of human angiosarcomas of which two had TP53 mutations. In canine hemangiosarcomas, protein-coding fusion genes were detectable in approximately 15% of tumors, and those were associated with p53 deficiency and enrichment of angiogenic gene signatures. Our data suggest that convergent molecular mechanisms associated with p53 inactivation and enhanced PI3K/AKT/mTOR signaling pathways are operational in genomically complex human angiosarcomas and canine hemangiosarcomas.

In the past decade, advances in next-generation sequencing and bioinformatics have enabled genome-wide identification of unbiased cancer-associated fusion transcripts in a variety of tumor types. Previous studies have reported 7,887 fusion transcripts identified across thirteen tumor types in TCGA datasets and 9,928 fusion genes with a 3% recurrence rate in the Mitelman Database of Chromosome Aberrations and Gene Fusions in Cancer (53–55). These findings illustrate the complexity of the cancer-associated fusion gene landscape, showing a relatively high rate of fusions with low recurrence, possibly arising from catastrophic chromosome rearrangements by chromothripsis (56) and chromoplexy (57). Despite this relatively high frequency of fusion genes, a solution to define their pathogenic significance remains elusive. One key finding from this work was that protein-coding fusion genes cooccurred with mutations of TP53 in the angiogenic molecular subtype of canine hemangiosarcoma, suggesting that genomic instability might create a predisposition for translocations and the resultant fusion genes and that, in turn, these fusion genes create unique transcriptional programs that promote angiogenic phenotypes in these p53-deficient backgrounds.

Specifically, kinase fusion genes involving FLT1, PTK2, and AKT3 can activate key convergent gene pathways associated with blood vessel formation and remodeling (58), and they represent potential therapeutic targets for kinase inhibitors (35, 59). When we consider that sarcomas have the highest frequency of kinase fusions in TCGA datasets (35), but that they also have extremely low recurrence, a more rational approach might be to develop agents that target these convergent angiogenic pathways instead of the products from the individual fusion genes themselves (60).

The two fusion genes that we confirmed by genomic structural evaluation in canine hemangiosarcomas were present in approximately 20% to 40% of cells in the tumor, both genes showing chaotic amplification. Several explanations could account for these observations. One is that histology and bioinformatics assays overestimated tumor content and tumor purity. Another is that fusions are epiphenomena arising from chaotic genomes with no influence on selection. A third, which we believe is most likely, is that translocation and the resulting fusion events occur stochastically in genomically unstable cells late in the course of tumor evolution. However, the enrichment of fusion genes and angiogenic transcriptional programs suggests that these traits endow tumor cells with selective growth and/or survival advantages that contribute to tumor progression by promoting proangiogenic environments. It is worth noting that the selective pressures in the angiosarcoma milieu favor not only fusion-positive clones, but also fusion-negative clones, as the establishment of a proangiogenic environment could improve survival of all the subpopulations within the tumor. Indeed, a similar mechanism might be operative in alveolar rhabdomyosarcomas, where PAX3-FOXO1A fusions are necessary for tumor initiation but have no effect on tumor recurrence (61, 62). Further work will be necessary to distinguish which among these nonmutually exclusive possibilities are operative, and to better understand the role of fusion genes in tumor evolution of human angiosarcoma and canine hemangiosarcoma and, potentially, in promoting clonal heterogeneity through the creation of a permissive niche.

Fusion genes have been reported in human angiosarcomas (10, 63–65). For instance, one study found a CIC-LEUTX fusion in one of 120 (0.8%) FFPE angiosarcomas examined (10); another found a CEP85L-ROS1 fusion in one of 34 (3.0%) angiosarcomas examined (63); and a third found an EWSR1-ATF1 fusion in one case of angiosarcoma (64). A NUP160-SLC43A3 fusion has also been reported in the ISO-HAS angiosarcoma cell line (65). However, none of these fusion genes has been identified recurrently in subsequent studies of angiosarcoma samples. While these observations are consistent with our stochastic hypothesis, we cannot completely exclude the possibility that fusion genes in human angiosarcomas, or for that matter in canine hemangiosarcomas, are nonpathogenic passenger aberrations.

A larger case series will be required to define the fusion gene landscape in human angiosarcoma, but canine hemangiosarcoma provides potential insights for what might be expected. Mutations of TP53 are largely mutually exclusive of mutations in KDR, PIK3CA, and RAS gene family in human angiosarcoma, and the mutational patterns seem to be associated with the location of the primary tumor (10, 12, 17, 66, 67). We see a similar pattern emerge in a subset of canine hemangiosarcoma, and from our data, we propose a model that can be used as a foundation to test mechanistic links between the mutational and transcriptional landscapes in malignant vascular tumors (Fig. 4) and determine their roles in tumor progression. In this model, inflammatory hemangiosarcomas harbor no mutations of PIK3CA, and only rarely of TP53, maintaining sufficient genomic stability that disfavors formation of fusion genes. Furthermore, the transcriptional programs in these inflammatory hemangiosarcomas are weakly angiogenic, and their permissive inflammatory environments restrain growth and metastasis. Conversely, angiogenic hemangiosarcomas harbor frequent mutations of PIK3CA and TP53, and fusion events. Mutations of PIK3CA in p53-proficient backgrounds promote proangiogenic environments, while in p53-deficient backgrounds, these mutations promote altered chromatin regulation and immunomodulatory transcriptional programs. Finally, mutations of TP53 enable genomic instability with formation of fusion genes. These events are stochastic, but fusion genes that promote proangiogenic transcriptional programs can enhance or even supplant the effects of PIK3CA mutations and create environments that accelerate tumor growth and metastatic propensity.

Molecular distinctions among human angiosarcomas could be driven by their clinical phenotype and potential therapeutic responses (9, 68, 69). For instance, a subset of angiosarcomas harbor gene amplifications of MYC and FLT4 which frequently cooccur in tumors associated with ultraviolet (UV) irradiation- or therapeutic radiation. Mutational signatures associated with UV exposure and high mutational burden might predict more favorable immunotherapeutic responses in patients with angiosarcoma, as they do in patients diagnosed with malignant melanoma; however, supportive clinical trials to test this premise are limited, and the use of immune checkpoint inhibitors in human patients with angiosarcoma thus far has yielded mixed results (70, 71). The mutational signatures in canine hemangiosarcomas are largely confined to the “aging” (cellular replication) signature (17), and total mutational burden is relatively low (17, 72), so this condition is unlikely to provide a model to address the utility of immunotherapy in this context. However, canine hemangiosarcoma could provide a suitable model to address other treatments, whether pharmacologic or immunologic, directed at the molecular programs that drive progression and maintenance of the tumors in both species (33). Such validation studies could alter the paradigms for diagnosis and treatment of human angiosarcoma and canine hemangiosarcoma, as well as of other aggressive, genomically complex sarcomas that affect humans and dogs alike.

J.H. Kim reports grants from NIH/NCI, AKC Canine Health Foundation, National Canine Cancer Foundation, Morris Animal Foundation, Swedish Cancerfonden, Swedish Research Council, and Oscar J. Fletcher Distinguished Professorship during the conduct of the study. K. Megquier reports grants from NCI during the conduct of the study. R. Thomas reports grants and personal fees from Canine Health Foundation during the conduct of the study. E.K. Karlsson reports grants from NIH during the conduct of the study. No disclosures were reported by the other authors.

J.H. Kim: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. K. Megquier: Data curation, formal analysis, validation, methodology, writing–review and editing. R. Thomas: Data curation, formal analysis, validation, investigation, methodology, writing–review and editing. A.L. Sarver: Data curation, formal analysis, validation, methodology, writing–review and editing. J.M. Song: Data curation, investigation, methodology, writing–review and editing. Y.T. Kim: Data curation, formal analysis, methodology, writing–review and editing. N. Cheng: Data curation, formal analysis, visualization, writing–review and editing. A.J. Schulte: Project administration, writing–review and editing. M.A. Linden: Data curation, formal analysis, writing–review and editing. P. Murugan: Data curation, formal analysis, writing–review and editing. L. Oseth: Resources, data curation, formal analysis, methodology, writing–review and editing. C.L. Forster: Resources, data curation, formal analysis, writing–review and editing. I. Elvers: Resources, data curation, validation, methodology, writing–review and editing. R. Swofford: Resources, data curation, methodology, writing–review and editing. J. Turner-Maier: Resources, data curation, methodology, writing–review and editing. E.K. Karlsson: Resources, methodology, writing–review and editing. M. Breen: Resources, funding acquisition, methodology, writing–review and editing. K. Lindblad-Toh: Resources, funding acquisition, methodology, writing–review and editing. J.F. Modiano: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, writing–review and editing.

The authors would like to acknowledge Dr. Corrie Painter for reviewing the manuscript and providing feedback. The authors acknowledge Mitzi Lewellen for assistance with inventory, database management, and editorial assistance. The authors would also like to thank Lauren Mills for processing of the next generation sequencing data and Dr. Douglas Yee, Director of Masonic Cancer Center, for assisting with the collection of human tissues. Human biospecimens were obtained from the UMN BioNet and from the CHTN. Tissue samples were provided by the CHTN which is funded by the NCI. Other investigators may have received specimens form the same subjects. This work was partially supported by grants 1R03CA191713-01 (to J.F. Modiano, A.L. Sarver, and J.H. Kim) and R37CA218570 (to E.K. Karlsson) from the NCI of the NIH, grants #422 (to J.F. Modiano) and 1889-G (to J.F. Modiano, M. Breen, and K. Lindblad-Toh) from the AKC Canine Health Foundation, grant JHK15MN-004 (to J.H. Kim) from the National Canine Cancer Foundation, grant D10-501 (to J.F. Modiano, M. Breen, and K. Lindblad-Toh) from Morris Animal Foundation, and a grant from Swedish Cancerfonden (to K. Lindblad-Toh). This work was also supported by an NIH NCI R50 grant, CA211249 (to A.L. Sarver). The NIH Comprehensive Cancer Center Support Grant to the Masonic Cancer Center, University of Minnesota (P30 CA077598) provided support for the cytogenetic analyses performed in the Cytogenomics Shared Resource. K. Megquier is supported by the NCI of the NIH under Award Number F32CA247088. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. M. Breen is supported in part by the Oscar J. Fletcher Distinguished Professorship in Comparative Oncology Genetics at North Carolina State University. K. Lindblad-Toh is supported by a Distinguished Professor award from the Swedish Research Council. J.F. Modiano is supported by the Alvin and June Perlman Chair in Animal Oncology. The UMGC (http://genomics.umn.edu) supported for generation of genomic sequencing data libraries, and the Minnesota Supercomputing Institute (MSI) at the University of Minnesota (http://www.msi.umn.edu) provided computational resources that contributed to the results in this study. The authors gratefully acknowledge donations to the Animal Cancer Care and Research Program of the University of Minnesota that helped support this project.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.