Integrated Multiomic Profiling Enhances Risk Stratification and Prognostication in Canine Osteosarcoma

Abstract

Purpose:

Osteosarcoma is a heterogeneous and aggressive primary bone malignancy that affects both canines and humans. Standardized treatment regimens prescribed to both species do not address the complexity of the disease and thus have resulted in stagnant patient outcomes for more than 30 years.

Experimental Design:

In this study, we present the first multiomic dataset created from a large outcome-linked biobank of canine osteosarcoma treatment-naïve primary tumors, utilizing a computational framework designed to interrogate each dataset individually and to compare and integrate findings.

Results:

This exploratory work suggests that the presence of MYC amplification is a poor prognostic indicator in canines and highlights alterations in DNA damage repair, metabolism, and cell-cycle genes that are shared with humans. Furthermore, we show relationships between the local tumor immune microenvironment, TP53 mutations, MYC and PTEN status, and global gene methylation patterns.

Conclusions:

This work highlights the complexity of the disease and provides new insight into the utility of prognostic biomarkers and potential druggable targets for future study.

Introduction

Osteosarcoma, the most prevalent primary bone malignancy affecting both humans and pet dogs, presents significant clinical challenges, with limited progress in outcomes despite extensive research efforts over the past three decades. Osteosarcoma is characterized by its complex and heterogeneous nature that lacks clear genomic drivers and exhibits substantial intra- and interpatient variability in molecular and cellular features. This heterogeneity is reflected in diverse histologic patterns, variable immune infiltration, tumor microenvironment (TME) composition, and chaotic chromosomal events. In humans, traditional tumor-derived biomarkers, such as the percentage of tumor necrosis after neoadjuvant chemotherapy, genomic alterations such as MYC amplification (AMP; ref. 1), circulating tumor DNA levels (2), and patient-specific factors (i.e., tumor location and size), often fall short in accurately predicting the risk of disease progression in patients with localized disease following frontline therapy (3). The clinical and molecular similarities between human and canine osteosarcoma make the canine patient a valuable resource for investigating the human disease, particularly given that the incidence of osteosarcoma is 20 to 50 times higher in dogs. Both species share clinical, histologic, genomic, transcriptomic, and tumor microenvironmental features, positioning canine osteosarcoma as an important model in comparative oncology (4). The canine genome contains ~19,000 genes that are orthologous to human genes, and dogs exhibit more than 650 Mb of ancestral DNA sequence that is shared with humans but not with rodents, further strengthening the dog as a translational model species (5, 6). This comparative approach enables the discovery and testing of predictive biomarkers and novel therapeutic strategies in pet dogs with spontaneous osteosarcoma, which can then be translated into human clinical research and practice.

In this study, we leverage multiomic computational techniques with genomic, epigenomic, transcriptomic, and associated clinical data from canine patients enrolled in NCI’s Comparative Oncology Trials Consortium (COTC) osteosarcoma clinical trials (7). Our analysis of canine osteosarcoma samples reveals genomic variation in cell cycle and DNA damage repair pathway genes, somatic missense mutations in critical oncogenes, and recurrent AMP and deletion (DEL) of specific segments on canine chromosomes 13 and 26. Additionally, our study highlights features of canine osteosarcoma that are shared with human osteosarcoma, including (i) a significant correlation between the presence of TP53 mutations and an immune-depleted TME (8); (ii) upregulation of oxidative phosphorylation, glycolysis, and G₂–M checkpoint pathways in outcome-linked MYC-amplified samples (1, 9–15); and (iii) associations among global hypomethylation, MYC AMP, PTEN DEL, and poor patient outcomes (16, 17). These exploratory findings, which must be validated in future prospective studies, open avenues for investigating genomic variation involving these genes in human osteosarcoma to assess their prognostic value for patients while also exploring new clinical biomarkers and potential therapeutic targets.

Materials and Methods

Clinical data collection

Demographic and clinical outcome data, including information on canine patient characteristics, tumor location, and serum alkaline phosphatase (ALP) levels, were collected from patient cohorts enrolled in the NCI’s COTC-021/022 (7), 026 (18), and 030 (19) osteosarcoma clinical trials. The study participants were privately owned pet dogs of diverse ages, breeds, and sexes, recruited from participating COTC veterinary academic institutions. Each institution adhered to and maintained its own Institutional Animal Care and Use protocol, which ensured informed written owner consent, standardized trial procedures, and harmonized clinical monitoring for all canine patients. All tumor samples were obtained prior to treatment and evaluated by board-certified anatomic veterinary pathologists affiliated with the COTC institutions to confirm the diagnosis of osteosarcoma.

Biological samples

The discovery cohort consisted of tumor and matched normal tissue samples from pet dogs enrolled in the COTC-021/022 clinical trial, which evaluated the clinical effect of adjuvant sirolimus therapy added to standard-of-care (SOC) therapy (n = 55 tumor/normal paired samples). The samples were acquired at the time of SOC amputation surgery. The tissue was placed in RNAlater and stored at −80°C until processing.

The expansion cohort consisted of tumor samples only and was acquired from pet dogs enrolled in the COTC-021/022 (7), 026 (18), and 030 (19) clinical trials (n = 209 tumor samples). The tissue was harvested at the time of SOC amputation surgery, placed in RNAlater, and stored at −80°C until processing (Supplementary Fig. S1).

Nucleic acid isolation and quality assurance/quality control

As previously described (20, 21), DNA and RNA were isolated from all tumor and normal tissues preserved in RNAlater using the QIAGEN AllPrep DNA/RNA Mini Kit (cat. #80204). A Qubit 2.0 Fluorometer with the Qubit dsDNA BR assay (Thermo Fisher Scientific) and the TapeStation Genomic DNA assay (Agilent Technologies) were used to assess the DNA quantity and quality. Samples with a 260:280 ratio >1.8 and with >200 ng of total DNA were utilized for DNA sequencing. Furthermore, a NanoDrop 8000 (Thermo Fisher Scientific) and an Agilent 4200 TapeStation with RNA ScreenTape (cat. #5067–5576) and RNA ScreenTape sample buffer (cat. #5067–5577) were used to assess the total RNA quality and quantity. Samples with RNA integrity number (RIN) >8 and a total RNA quantity >100 ng were used for bulk RNA sequencing (RNA-seq) analysis.

Whole-genome sequencing

Normal tissue, tumor, and cell line DNA samples that passed Quality Assurance/Quality Control (QA/QC; 260:280 ratio >1.8; >200 ng of total DNA) were converted into libraries using TruSeq Nano DNA library prep, pooled, and sequenced on NovaSeq 6000 S4 (RRID: SCR_016387) flow cells in paired-end sequencing mode. The samples were mapped, and variants were called using DRAGEN. The percentage of total mapping against the reference canine genome (canFam4, GSD_1.0) was about 99%, and uniquely mapped reads were above 70%. Library complexity (i.e., the percentage of unique reads) was determined by measuring the percentage of unique fragments in the mapped reads using the MarkDuplicates utility. Coverage statistics were also measured using DRAGEN. The mapped sequencing depth coverage (after alignment and marking duplicates) was between 22x and 134x (discovery cohort) and 1x and 6x (expansion cohort).

Copy-number variation

To determine DNA copy-number levels, we used the GATK copy-number variation (CNV) caller for the discovery cohort (version 4.6.0.0; RRID: SCR_001876; ref. 22). For the expansion cohort, which consisted of tumor-only samples, the sequencing depth was insufficient to analyze with the GATK CNV caller; thus, CNVkit was used (version 0.9.11; RRID: SCR_021917; ref. 23). We called AMP and DEL using a threshold of the log₂ ratio ≥0.2 and ≤−0.2 (24). The detailed parameters and steps used are shown in Supplementary Fig. S2. The osteosarcoma marker genes (n = 83) were collected from the My Cancer Genome repository and grouped based on biological pathways (25). Significant changes in marker genes across the tumor samples were qualitatively represented using ComplexHeatmap (version 2.22.0; RRID: SCR_017270; refs. 26, 27). The GATK segmented copy-number profile was used as an input for Genetic Identification of Significant Targets in Cancer (GISTIC 2.0, version 2.0.23, RRID: SCR_000151; ref. 28) to identify significantly amplified or deleted regions and obtain gene-level estimates of copy number in the discovery cohort (n = 55 tumor/normal paired samples). GISTIC was run with a 0.99 confidence level, and the threshold for copy-number AMPs and DELs was 0.3. Aberrant regions with FDR q values ≤0.1 were considered significant.

Single-nucleotide and structural variation

Somatic single-nucleotide variants (SNV) in paired tumor/normal samples were identified by Mutect2 (version 4.6.0; RRID: SCR_026692; ref. 29), Seurat (version 2.6; ref. 30), and Strelka2 (version 2.9.10; ref. 31); variants called by two or more callers were considered for final analysis. Variants were annotated using SnpEff (version 5.2f; RRID: SCR_005191; ref. 32). Additionally, germline mutations were identified by GATK Haplotype Caller (22). Common single-nucleotide polymorphisms (SNP) and insertions/deletions (InDel) from 1,929 samples of canFam4 were collected from iDog 2.0 (33). Variant effect was annotated using variant effect predictor version 109 (RRID: SCR_007931; ref. 34). Structural variants (SV) were called from paired whole-genome sequencing (WGS) data by Delly (version 0.7; https://github.com/dellytools/delly; RRID: SCR_004603), and the somatic regions that passed QC with MAPQ score ≥40, paired end ≥10, and split read ≥10 (35) were collected. SVs identified by Delly are classified as translocations (BND), inversions (INV), DEL, and duplications (DUP). The detailed parameters and steps used are shown in Supplementary Fig. S2.

Catalogue of Somatic Mutations in Cancer signature identification

In mutational signature analysis, gene mutations in each canine sample (n = 55 paired; discovery cohort) were classified into 96 trinucleotide patterns according to the neighboring nucleotide context of mutations, using MutationalPatterns (36) in the R package. The resulting mutational pattern matrix (96 trinucleotide patterns by 55 paired samples) was then broken down using non-negative matrix factorization (NMF). The best number of components for this factorization was found by maximizing the logarithmic likelihood.

Five signatures were determined in the discovery cohort based on the cophenetic metric and compared with the mutational signatures in Catalogue of Somatic Mutations in Cancer (COSMIC; version 3.4; RRID: SCR_002260; ref. 37) via cosine similarity (>0.85). The distribution of single-base substitutions (SBS) in each sample was shown using ComplexHeatmap (version 2.22.0; RRID: SCR_017270; refs. 26, 27). The Pearson correlation was calculated to assess the relationship between the age of dogs and the relative contribution value of each COSMIC signature using R. The relative contribution of COSMIC signatures between MYC-amplified and MYC-unamplified stratified samples was compared using the Wilcoxon rank test and visualized in bar plots using the ggplot2 package (version 3.5.1; RRID: SCR_014601).

Tumor purity and ploidy estimation

The estimated tumor content from a representative hematoxylin and eosin (H&E)–stained section for all COTC tumors varied from 20% to >90%. However, there was no consistent physical relationship between the tumor samples placed in formalin for paraffin embedding and histologic review and those samples placed in RNAlater for sequencing. Thus, paired tumor and normal WGS data (n = 55; discovery cohort) were used to estimate the tumor purity and ploidy using the ichorCNA tool (version 0.3.2; RRID: SCR_024768; ref. 38). HMMcopy (version 1.50.0; ref. 39) was used to generate read depth counts in 1-Mb intervals across the genome. Reference files for Guanine-Cytosine (GC) and mappability correction were generated using the hmmcopy_utils package (https://github.com/shahcompbio/hmmcopy_utils; RRID: SCR_026464). Putative centromere locations were assigned as regions with 80% or more repetitive sequence, based on a genome scan for four centromeric repeats using a 5-kb window size (canFam4; https://github.com/Chao912/Mischka/blob/master/GSD1.0_CanFam4.centromere.bed). To enhance accuracy and minimize noise, the analysis incorporated a panel of normals from the discovery cohort samples. The detailed parameters and steps used are shown in Supplementary Fig. S2.

Tumor mutational burden analysis

Tumor mutational burden (TMB) was quantified as the frequency of somatic base substitutions and small InDels within the callable coding sequence, expressed per megabase, and has emerged as a significant prognostic and predictive biomarker for assessing response to immunotherapy, particularly with checkpoint inhibitors (40). To calculate the TMB, the total number of mutations counted was divided by the size of the Coding DNA Sequence (CDS) region (~36 Mb). In this study, variants observed across multiple samples were included, whereas less impactful variants, specifically those classified under the “Modifier” category, were explicitly excluded from the analysis.

Whole-genome bisulfite sequencing

Primary osteosarcoma (n = 57) tumor tissue selected from both the discovery and expansion cohorts and normal bone samples from ribs and vertebrae (n = 4) from NCI-COTC-021/022 samples were sequenced on eight NovaSeq 6000 S4 (RRID: SCR_016387) runs using the EZ DNA Methylation-Gold Kit/Accel NGS Methyl-Seq DNA library preparation and paired-end sequencing. The samples were mapped using DRAGEN. The percentage of total mapping against the reference canine genome (canFam4, GSD_1.0) is approximately 95%, and uniquely mapped reads are above 60%. Library complexity (i.e., the percentage of nonduplicate reads) was determined by measuring the percentage of unique fragments in the mapped reads using the MarkDuplicates utility. The percentage of duplicated reads is between 8% and 34%. Coverage statistics were also measured using DRAGEN. The mapped sequencing depth coverage (after alignment and marking duplicates) was between 15x and 118x. Methylation calling was completed using DRAGEN. The mapping efficiency (the number of sequences with a unique best alignment) against the reference genome is 83% to 91%. There are 42.39% to 83.64% of cytosines methylated in the CpG context, 0.36% to 1.91% of cytosines methylated in the CHG context, and about 0.35% to 2.27% of cytosines methylated in the CHH context. The low levels of non-CpG methylation (CHG and CHH) were used as a proxy to estimate bisulfite conversion efficiency, which was approximately 98% to 99.6%.

Unsupervised clustering and differential methylation pattern

For clustering analysis, CpG context files underwent preprocessing using methylKit (version 1.34.0; ref. 41) in R version 4.5.0 (RRID: SCR_005177). Methylation calls were filtered based on a minimum coverage threshold of 10 reads per CpG position. To minimize PCR bias, CpG sites exceeding the 99.9th percentile of coverage were excluded. Subsequently, the methylation percentage was computed for each sample (n = 61). To discern inherent groupings, the prepared data were subjected to various clustering methodologies. First, principal component analysis was performed on the methylation matrix, followed by k-means clustering to partition CpG sites into k distinct clusters based on their features. Second, consensus clustering was performed using the ConsensusClusterPlus (42) package (RRID: SCR_016954) to validate the stability of the clusters. Lastly, we used the Monte Carlo reference-based consensus clustering (M3C) algorithm (43). M3C uses a multicore-enabled Monte Carlo simulation to generate null distributions along the range of K, which are used to select its value. M3C uses the relative cluster stability index and P values to decide on the value of K and reject the null hypothesis, K = 1. Differential DNA methylation was calculated by comparing the proportion of methylated CpGs in tumor samples (n = 57) relative to a control (n = 4) using the calculateDiffMeth function of methylKit. Bases or regions with more than a 25% methylation difference between tumor and control sample groups and a q value of at least 0.01 were selected.

Bulk mRNA-seq

As described previously (21), RNA samples from 190 primary osteosarcoma tumors from the NCI-COTC-021/022 clinical trial were pooled and sequenced on NovaSeq Xp_S1 (RRID: SCR_024569) using Illumina TruSeq stranded mRNA library prep and paired-end sequencing. The samples have 17 to 115 million pass filter reads, with more than 89% of bases above the quality score of Q30. Reads of the samples were trimmed for adapters and low-quality bases using Cutadapt before alignment with the reference canine genome (canFam4, GSD_1.0) and the annotated transcripts using STAR (RRID: SCR_004463). The mapping statistics are calculated using Picard software (RRID: SCR_006525). Library complexity is measured in terms of unique fragments identified after UMI-based deduplication using Picard’s MarkDuplicates utility. In addition, the gene expression quantification analysis was performed for all samples using STAR/RSEM tools.

Gene set enrichment and cell deconvolution

Normalized raw count data and the MYC-amplified and MYC-unamplified grouping information for n = 35 tumors were provided as input to the edgeR (version 4.6.1; ref. 44) software in R (RRID: SCR_012802). The edgeR was then used to rank all genes based on their estimated log fold change (FC) in expression. To visualize the differential expression results, including the estimated log₂ FC in the MYC-amplified and MYC-unamplified samples and the corresponding P values, the volcano plots were generated using the ggplot2 package (version 3.5.2; RRID: SCR_014601). Gene set enrichment analysis (GSEA) was performed using fgsea (version 1.34.0) in R, with P value estimation based on an adaptive multilevel split Monte Carlo scheme. The top 10 upregulated and downregulated pathways are listed by their normalized enrichment score (NES) and ordered by P value.

To estimate the abundance of different immune cell types between the MYC-amplified and MYC-unamplified samples, MCPcounter (version 1.2.0; ref. 45) was used. The list of marker genes used is shown in Supplementary Table S1. The resulting MCP-counter scores were scaled, and the “Wilcoxon rank test” was adopted for significant differences, which were then visualized in bar plots using the ggplot2 package (version 3.5.1; RRID: SCR_014601).

TME subtype and survival analysis

Our laboratory previously developed a machine learning model to classify the TME based on its inferred composition using mRNA-seq profiles of canine tumors from the NCI-COTC-021/022 clinical trial (21). In this prior report, bulk transcriptomic data from 245 pet dogs with treatment-naïve appendicular osteosarcoma were analyzed using deconvolution to characterize the TME components. We incorporated a portion of this data in the current study to explore relationships between transcriptomically defined subtypes and other genomic features. Based on the cell type composition of bulk tumor samples, TME subtypes of canine osteosarcoma can be classified as immune-enriched (IE), immune-enriched dense extracellular matrix-like (IE-ECM), and immune desert (ID). The fraction of samples grouped in each TME subtype is shown in a heatmap generated using the ggplot2 package (version 3.5.1; RRID: SCR_014601).

To perform Kaplan–Meier and Cox proportional hazards survival analysis of the NCI-COTC clinical trial datasets and patient subgroups defined by genomic alterations (see Supplementary Fig. S1 for details on which trial patients contributed to each dataset), we utilized the survival (version 3.8.3) and survminer (version 0.5.0) packages implemented in R.

We estimated hazard ratios (HR) and 95% confidence intervals (CI) between patient subgroups using Cox proportional hazards models adjusted for ALP status, age, sex, weight, treatment type (SOC with or without adjuvant sirolimus), peripheral blood monocyte count at diagnosis (below or equal to/above 400 cells/μL; ref. 46), and TME subtype. Both survfit (for Kaplan–Meier) and coxph (for Cox regression) utilize the na.action argument, which defaults to na.omit. This means any row with a missing value (NA) in the survival time, event status, or any covariate included in the formula is automatically dropped from the analysis.

Post hoc power for the Cox regression analysis was calculated using the powerSurvEpi package (version 0.1.5) in R, based on the observed HR for the patient subgroups, the total number of events in each group, and a significance level of 0.05.

Results

Clinical and molecular features of DOG2 samples

Using the tumor samples described above and associated clinical metadata, datasets were collected and integrated to facilitate a multiomic study of pet dogs with appendicular osteosarcoma. All tumor samples are treatment-naïve and from the primary appendicular site (Fig. 1A) and were collected as part of COTC canine clinical trials (Fig. 1B; refs. 7, 18, 19). Broadly, these datasets were generated from a total of 301 primary osteosarcoma samples; the numbers of samples with overlapping multiomic data are depicted in Fig. 1C and Supplementary Fig. S1. The median age of dogs at osteosarcoma diagnosis was 8 years (range 0.9–15.6 years). Of these, 127 (42.1%) were female, and 174 (57.8%) were male, and 118 females and 164 males were spayed or neutered, respectively. The most common breeds included mixed breeds (54/301), labrador retriever (44/301), golden retriever (21/301), and greyhound (19/301). Additional demographic and clinical outcome data, including information on canine patient characteristics, tumor size/location, peripheral blood monocyte count at diagnosis, and serum ALP levels of each sample, are shown in Supplementary Table S2. The baseline monocyte count was only available from 39/55 dogs assessed in the Cox proportional hazards analysis.

Figure 1.

Data summary and WGS analysis pipeline. A, Pet dogs spontaneously develop appendicular osteosarcoma, depicted here in the distal femur. B, Sample preprocessing and multiomics data collection from the NCI-COTC-021/022 clinical trial, which forms the basis for the NCI’s DOG2 project. C, Sample cohort with sequencing platform, with the corresponding UpsetR plot showing the number of dogs with multiple sequencing information. D, Flowchart demonstrating the sequential use of tools in the evaluation of paired normal tissue and tumor OS WGS samples. The algorithm with version information is shown underneath the type of variants analyzed. DOG2, Decoding the Osteosarcoma Genome of the Dog; WGBS, whole-genome bisulfate sequencing. [A, Created in BioRender. Garg, A. (2026) https://BioRender.com/lp561qd.]

Genomic features of canine osteosarcoma mirror those of human patients

The analysis of WGS data from the discovery cohort (n = 55) revealed alterations in many genes known or implied to play a role in osteosarcoma (Fig. 1D; Supplementary Table S3). Genes within oncogenic pathways involving TP53, RAS, PI3K, MYC, and cell-cycle control were commonly affected (Fig. 2; Supplementary Fig. S3; refs. 10, 35). The assessment of somatic copy-number variants (using log₂ FC cutoff ≤−0.2 and ≥0.2) revealed copy-number gains in AKT1 (31/55 tumors, 56.36%), MYC (31/55 tumors, 56.36%), ARHGAP39 (29/55 tumors, 52.73%), RECQL4 (29/55 tumors, 52.73%), KIT (23/55 tumors, 41.82%), PDGFRA (23/55 tumors, 41.82%), CCND3 (21/55 tumors, 38.18%), and AURKA (16/55, 29.09%; Fig. 2). As also noted in prior studies (10, 35), copy-number losses of genes were identified in our cohort, namely DLG2 (36/55, 65.45%), CDKN2A (33/55, 60%), methylthioadenosine phosphorylase (MTAP; 31/55, 56.36%), CDKN2B (30/55, 54.55%), PIK3CG (26/55, 47.27%), RB1 (24/55, 43.64%), RAD50 (22/55, 40%), BARD1 (18/55, 32.73%), PTEN (21/55, 38.18%), TP53 (17/55, 30.90%), SETD2 (15/55, 27.27%), PIK3R1 (11/55, 20%), and RAD51D (11/55, 20%; Fig. 2). GISTIC 2.0 analysis of the GATK profiles confirmed significant AMP or DEL of several key genes in the discovery cohort (q < 0.1). Recurring AMPs include AKT1, MYC, CCND3, and AURKA. Recurring DELs include CDKN2B, MTAP, DLG2, BARD1, and PIK3R1 (Supplementary Table S4). The overall mutational burden was a mean of 0.71 per Mb, which is comparable to prior literature in dogs and humans, further highlighting similarities between the mutational landscape of human and canine osteosarcoma (10, 35, 47, 48). The depth of sequencing and lack of matched normal samples precluded performing this analysis in the expansion cohort.

Figure 2.

Genomic landscape of canine osteosarcoma. Genetic profiling of discovery cohort patients (n = 55). Dog identification numbers are listed along the x-axis, with selected genes organized by pathway listed along the left y-axis and the frequency of abnormality denoted along the right y-axis. Each column represents an individual dog, and each row represents genes organized based on their biological pathways. Significant copy-number gain (log₂ FC > 0.2; red) and loss (log₂ FC < −0.2; blue) are shown for each gene. (Top) Stacked bar plots show the frequency of somatic SNVs and SVs across each sample. (Bottom) Associated clinical metadata for each sample are presented, including serum ALP level (normal = 0, elevated above the laboratory reference interval = 1), age, tumor purity, ploidy, sex, TME subtype assignment, tumor location, and DFI.

The missense point mutations and chromosomal BNDs were the most frequent somatic mutation types in primary osteosarcoma tumors within the discovery cohort (46/55, 83.64%; Figs. 2 and 3A). A median of 48 different complex chromosomal BNDs (range: 0–207), 27.6 DELs (range: 0–119), 16.8 DUPs (range: 0–84), and 33.7 INVs (range: 0–171) were identified in this sample set (Fig. 3B; Supplementary Table S5). Genes affected by SVs across the discovery cohort WGS samples (RRID: SCR_016637) are described in Supplementary Table S5. Eight tumors (out of nine samples with purity <0.1) did not demonstrate somatic variations, which may be attributed to the low tumor purity of the analyzed sample (Fig. 2; Supplementary Table S3). Each gene with copy-number variation (CNV), SNV, and SV in each discovery cohort WGS sample is described in Supplementary Table S3 (please refer to Sheet 2).

Figure 3.

Mutational profile of canine primary osteosarcoma. A and B, Bar plots summarize the total number of somatic SNVs and SVs in each discovery cohort sample. Most of the samples exhibit missense (83.6%) and synonymous (85.5%) somatic SNVs, with BND being the most common SV in 76.3% of canine patients. C, Schematic structure of TP53 and its different domains, highlighting that mutations frequently occur within the DNA-binding domain. Cys143fs is the most frequent somatic mutation, present in 20 osteosarcoma samples, whereas the codon shown in red was found in one normal sample. * and fs represent stop gained and frameshift mutations in the codon, respectively. The Kaplan–Meier survival plots show overall survival (D) and DFI (E) of TP53-mut (n = 30) and -WT (n = 25) samples. The shaded area around each curve represents the 95% CI, and tick marks indicate censored cases. F, Heatmap depicting the fraction of TP53-stratified patients (n = 34) grouped into TME subtypes (IE, IE-ECM, and ID). G, Distribution of SBS in each primary osteosarcoma sample is shown, with the x-axis representing the sample identification number with age information and the y-axis showing specific COSMIC signatures. Relative contribution values were scaled from −2 to 2. H, Correlation between the COSMIC-1 signature and the age of the dog.

We further analyzed the germline mutations of the discovery cohort tumor/normal paired samples (Supplementary Table S6). Germline variants were identified in the AURKB, CHEK2, DLG2, FANCD2, FANCE, KRAS, RICTOR, TP53, and TSC1 genes but were not associated with any single type of dog breed. We also found that several high-impact splice and frameshift variants were recurrent, including variants in AURKB, KRAS, RICTOR, and DLG2, with allele frequencies of up to ~33% (Supplementary Table S6).

TP53 also demonstrated both somatic SNV and structural variations. Further in-depth analysis suggests that one of the normal samples showed a germline mutation in the TP53 DNA-binding domain (at position 141, frameshift mutation), and 20/55 osteosarcoma tumor samples show the same somatic SNV frameshift mutation at position 143 within the DNA-binding domain (Fig. 3C). The list of all the variations we found that involved TP53 is shown in Supplementary Table S7. To test the hypothesis that TP53 mutations may have an effect on the outcome, we next stratified the cohort based on the presence [TP53-mutant (TP53-mut), 30/55] or absence [TP53-wild type (TP53-WT), 25/55] of “high-impact” TP53 mutations in tumor tissue (Supplementary Table S3; ref. 32). No significant difference was observed between groups for either disease-free interval (DFI; HR, 1.09; 95% CI, 0.62–1.94; log-rank P = 0.76, post hoc power = 0.05) or overall survival (HR, 1.14; 95% CI, 0.61–2.1; log-rank P = 0.68, post hoc power = 0.06; Fig. 3D and E). We then utilized a multivariate Cox proportional hazards regression model to assess the predictive value of the TP53-stratified samples in the context of other clinically measured variables. We found that the primary TME subtype had a significant effect on overall survival but not DFI after adjusting for other clinical factors (Supplementary Fig. S4).

Using bulk mRNA-seq–based TME subtyping described previously (21), we compared TME subtype with TP53 mutation status and found that 13 of 19 (68.4%) TP53-mut samples had an ID TME designation (Fig. 3F). TP53-WT primary osteosarcoma samples were a mixture of ID [7 of 15 (46.6%)] and IE-ECM subtypes [6 of 15 (40%); Fig. 3F].

To further address the potential driving mutational processes of osteosarcoma, we applied NMF and identified five predominant signatures, which were similar to COSMIC 1, 8, 9, 40, and 17b signatures (Fig. 3G). COSMIC-1 is related to G:T mismatches in double-stranded DNA, which result from spontaneous deamination of 5-methylcytosine and thus C → T transitions and is known as the “aging signature” (49). In our dataset, this signature shows a significant correlation between the contribution of COSMIC-1 and the age of the dog (Spearman correlation: r = 0.42, P = 0.001; Fig. 3H). COSMIC-8 (unknown etiology) and COSMIC-17b (associated in some human cases with fluorouracil chemotherapy and reactive oxygen species damage) have been reported in human and canine osteosarcoma (8, 35, 47, 50, 51). COSMIC-9 (unknown etiology) has been reported previously in canine but not human patients (48). We also assessed COSMIC-40, an additional mutational signature of unknown etiology associated with aging in some human cancers as well as canine osteosarcoma (50, 52).

Further analysis of somatic copy-number variants in the expansion cohort (n = 209) also revealed copy-number gains in AKT1 (64/209, 32.62%), MYC (119/209, 56.94%), PDGFRA (103/209, 49.28%), KIT (102/209, 48.80%), ARHGAP39 (99/209, 47.37%), RECQL4 (99/209, 47.37%), and CCND3 (92/209, 44.02%) and losses of similar genes as noted for the discovery cohort, namely PTEN (151/209, 72.24%), DLG2 (131/209, 62.68%), PIK3CG (120/209, 57.41%), SETD2 (100/209, 47.84%), CDKN2B (99/209, 47.39%), MTAP (96/209, 45.93%), RB1 (95/209, 45.45%), RAD50 (76/209, 36.36%), RAD51D (73/209, 34.92%), and PIK3R1 (59/209, 28.22%; Supplementary Fig. S5; Supplementary Table S8).

Methylation patterns of primary tumors and clinical associations

A total of 57 primary osteosarcoma samples from COTC-021/022 and four normal bone samples were used in the genome-wide methylation analysis. The workflow used to identify the differentially methylated regions between these groups of samples is shown in Fig. 4A. Unsupervised consensus clustering revealed two patient subgroups (Fig. 4B) with strikingly different focal methylation patterns, in which cluster 1 (n = 27) was largely hypermethylated relative to cluster 2 (n = 34; Wilcoxon rank test, P < 2.22e–16, Fig. 4C). Based on similar work in humans, we hypothesized that cluster assignment based on global methylation patterns may carry prognostic value in dogs (53). The overall survival between each cluster was analyzed (Fig. 4D and E). Patients in cluster 2 demonstrated shorter DFI (HR, 2.17; 95% CI, 1.17–4.02; log-rank P = 0.01; post hoc power = 0.67, Fig. 4D) and poor survival (HR, 2.42; 95% CI, 1.25–4.70; log-rank P = 0.007; post hoc power = 0.72, Fig. 4E), which overlaps with that reported in human osteosarcoma (10). We also used a multivariate Cox proportional hazards regression model to assess the predictive value of the methylation-stratified samples in the context of other clinically measured variables. In this analysis, we found that a peripheral blood monocyte count >400/μL had a significant effect on DFI and overall survival after adjusting for other clinical factors (Supplementary Fig. S7). Our analysis also found an association between global hypomethylation (cluster 2) and the ID TME subtype (Fig. 4F; refer to Methods for detailed TME subtype designation). Gene-specific methylation status is shown in Supplementary Table S9.

Figure 4.

Methylation patterns in primary osteosarcoma and associations with clinical outcomes and other genomic features. A, Computational workflow demonstrating the processing of control and primary osteosarcoma WGBS samples, including unsupervised clustering and differential methylation region identification. B, Three independent clustering algorithms: consensus matrix (refer to Supplementary Fig. S6 to see multiple clustering runs), K-means, and Monte Carlo simulation reveal two stable clusters. C, A significant difference (P < 2.22e–16) was seen in the methylation levels between the two unsupervised clusters, with cluster 1 (n = 23 primary and 4 control) exhibiting significant hypermethylation compared with cluster 2 (n = 34). D, Overall survival and (E) DFI curves of unsupervised clusters indicate that cluster 2 exhibits a poor outcome compared with cluster 1 (P = 0.01 and P = 0.007, respectively). The shaded area around each curve represents the 95% CI, and tick marks indicate censored cases. F, The heatmap shows the fraction of samples from each cluster grouped into TME subtype; 27 of 32 primary osteosarcoma samples in cluster 1 are classified as immune ECM, and 15 of 18 samples in cluster 2 are classified as ID. TME information is unavailable for 7 WGBS primary osteosarcoma samples. DM, differential methylation; RCSI, relative cluster stability index.

MYC AMP association with poor survival

The copy-number segmentation of the discovery cohort demonstrates that 31 of 55 (56.36%) tumors demonstrate MYC AMP and 24 of 55 (43.64%) are without MYC AMP (Fig. 5A). Based on recent data (54) indicating MYC AMP as a negative indicator of outcome in human osteosarcoma, we hypothesized that MYC AMP may also be associated with poor outcomes in dogs. Interestingly, MYC-amplified patients show shorter DFI (HR, 1.68; 95% CI, 0.94–3.0; log-rank P = 0.078; post hoc power = 0.4, Fig. 5B) and poorer survival (HR, 2.01; 95% CI, 1.07–3.8; log-rank P = 0.028; post hoc power = 0.54, Fig. 5C) and are enriched for the ID TME subtype (Fig. 5D). We additionally used a multivariate Cox proportional hazards regression model to assess the predictive value of the MYC-stratified samples in the context of other clinically measured variables. We found that the primary TME subtype has a significant effect on overall survival but not on DFI, even after adjusting for other clinical factors (Supplementary Fig. S8). We next investigated the epigenomic and transcriptomic changes within these MYC-stratified samples to determine whether other molecular features were associated with these patient subgroups. Fifteen primary osteosarcoma samples with MYC AMP had paired whole-genome bisulfate sequencing (WGBS) information. Further integrative analysis showed that MYC-amplified tumors exhibited hypomethylation compared with MYC-unamplified ones (Fig. 5E). Furthermore, transcriptomic analysis of MYC-amplified samples showed significant upregulation and downregulation of the following genes: CDK1 (log FC = 0.84, P = 0.000089), CHEK1 (log FC = 0.54, P = 0.02), glyceraldehyde-3-phosphate dehydrogenase (GAPDH; log FC = 0.53 and P = 0.02), triose phosphate isomerase (TPI) 1 (log FC = 0.34 and P = 0.03), ACKR1 (log FC = −5.089, P = 2.49E-07), CLEC3A (log FC = −4.49, P = 5.14E-06), and GZMK (log FC = −4.1983411, P = 0.00010402; Fig. 5F; Supplementary Table S10). Transcriptomic analysis comparing MYC-amplified and MYC-unamplified tumors revealed upregulation of mTOR, MYC targets, and oxidative phosphorylation pathways in MYC-amplified osteosarcoma, whereas inflammatory responses, IFNα/γ signaling, and KRAS signaling pathways were downregulated (Fig. 5G; Supplementary Table S11). Based on these findings, we next hypothesized that tumor microenvironmental components may differ between MYC-amplified and MYC-unamplified tumors. Cell deconvolution analysis using bulk mRNA-seq data suggests that there is downregulation of B-cell lineage genes (P = 0.026) and monocytic lineage genes (P = 0.022) in MYC-amplified samples (Fig. 5H; Supplementary Table S12). Additional observations, including the comparative analysis of SBS signatures, suggest that COSMIC-1 (P = 0.043), COSMIC-8 (P = 0.0013), COSMIC-9 (P = 0.00035), and COSMIC-40 (P = 0.00043) were significantly enriched in MYC-amplified cohorts (Supplementary Fig. S9). Given the high proportion of canine osteosarcoma tumors with an immune-depleted TME, we evaluated the CNV landscape for geographically associated gene losses or gains that could contribute to this phenotype. One example that we evaluated is MTAP, a gene that is frequently codeleted with CDKN2A/B and is associated with an ID phenotype in many human cancers, including osteosarcoma (55, 56). Consistent with these observations, we found frequent MTAP loss in canine osteosarcoma (discovery cohort: 56.36%; expansion cohort: 45.93%), which was codeleted with CDKN2A/B on CFA11 in 29 of 31 tumors and was enriched in the ID TME subtype (15/31, 48.39%, Supplementary Fig. S10).

Figure 5.

MYC AMP associated with poor survival in canine primary osteosarcoma. A, MYC copy number–based sample stratification. Based on copy-number gain (shown in Fig. 2), 31 and 24 dogs were grouped into MYC-amplified and MYC-unamplified cohorts, respectively. The numbers of WGS overlapping samples with WGBS and bulk mRNA-seq are shown in colored boxes (designed with BioRender). B and C, Kaplan–Meier curves show a nonsignificant (P = 0.78) and significant (P = 0.028) difference between DFI and overall survival of dogs with MYC-amplified (n = 31) and MYC-unamplified (n = 24) tumors, respectively. The shaded area around each curve represents the 95% CI, and tick marks indicate censored cases. D, Within the discovery cohort, 15 of 21 MYC-amplified samples are grouped under the ID TME subtype, whereas 7 of 13 MYC-unamplified samples are grouped as IE-ECM. E, Box plots represent differences in the methylation levels of the MYC-amplified (n = 15) and MYC-unamplified (n = 14) samples. MYC-amplified samples exhibited significant hypomethylation compared with the MYC-unamplified (P = 0.0068). F, The volcano plot shows the differential gene expression in MYC-amplified (n = 21) and MYC-unamplified (n = 14) primary osteosarcoma samples. A horizontal line is drawn at −log₁₀(10–6) and vertical lines at −0.5 and 0.5. G, GSEA between MYC-stratified samples. H, Estimation of the abundance of infiltrating immune cells across MYC-amplified and MYC-unamplified tumor samples using cell deconvolution based on the associated mRNA-seq data. I and J, The Kaplan–Meier curve shows a significant difference between DFI (P = 0.01) and overall survival (P = 0.0039) of dogs with PTEN (n = 21) and without PTEN (n = 34) loss, respectively. The shaded area around each curve represents the 95% CI, and tick marks indicate censored cases. K, Within the discovery cohort, 14 of 15 samples with PTEN loss are grouped under the ID TME subtype, whereas 11 of 19 samples without PTEN loss are grouped as IE-ECM. L, GSEA between PTEN-stratified samples. TME information is unavailable for 21 discovery cohort samples. Datasets used in B; C, D, I, J, and K; E; and F, G, H, and L are derived from WGS, WGBS, and bulk mRNA-seq platforms, respectively. [A, Created in BioRender. Garg, A. (2026) https://BioRender.com/ry92k1e.]

We were also interested in exploring the effect of PTEN alterations in our dataset, using the multiomic nature of the data to orthogonally compare gene dose and expression of this important tumor-suppressor gene. We observed that the PTEN gene segment on CFA26 was lost in 38.18% (21/55) and 72.25% (151/209) of the discovery and expansion cohorts, respectively. Notably, canine patients with PTEN loss exhibited significantly low PTEN expression (log FC = −1.95, P = 6.89e–09, Supplementary Table S13), shorter DFI (HR, 2.15; 95% CI, 1.18–3.93; log-rank P = 0.001; post hoc power = 0.78, Fig. 5I), and poorer survival (HR, 2.57; 95% CI, 1.33–4.99; P = 0.0039; post hoc power = 0.89, Fig. 5J), along with enrichment for the ID TME subtype (14/21 discovery cohort, 66.67%, Fig. 5K). Using a multivariate Cox proportional hazards regression model to assess the predictive value of the PTEN-stratified samples in the context of other clinically measured variables, we found that ALP and sex had a significant effect on overall survival even after adjusting for other clinical factors (Supplementary Fig. S11). PTEN loss–stratified samples also showed upregulation of G2M checkpoint pathways (P = 3.328840e–14, NES = 2.36, Fig. 5l; Supplementary Table S14).

Discussion

Using multiomic computational approaches, our study integrates extensive clinical, genomic, epigenomic, and transcriptomic data to characterize the molecular landscape of canine osteosarcoma (7). Our comprehensive analysis confirms that similarities exist between canine and human osteosarcoma genomes, nominates outcome-linked candidate genomic alterations in osteosarcoma-bearing pet dogs, and presents a focused examination of TP53, MYC, and PTEN and their relationships with tumor microenvironmental components in canine osteosarcoma. Furthermore, our analysis of gene dosage, expression levels, and chromosomal alterations offers valuable insights into the genetic and molecular mechanisms underlying osteosarcoma while facilitating the exploration of deployable clinical biomarkers. It is important to point out that these exploratory analyses and the resulting data were evaluated retrospectively and require prospective investigation and validation as definitive biomarkers for clinical outcomes, as well as generalizability to patients receiving therapies other than those described here.

In Figure 2, we present an overview of the hallmark genomic changes in treatment-naïve primary canine osteosarcoma, with associated clinical metadata, derived from a canine clinical trial. It is noteworthy that a subset of samples (sample identification numbers: 1407, 2303, 1402, 308, and 335) did not demonstrate noticeable CNVs, SVs, and SNVs, which is possibly attributable to low tumor purity (<0.1), thereby highlighting the rigorous criteria applied in this work to discover a high-confidence mutational landscape. Future studies such as this should attempt to match samples selected for H&E review and sequencing as closely as possible.

The Cox proportional hazard analyses were completed to incorporate major clinical prognostic factors for canine osteosarcoma (ALP status, age, sex, and weight) in addition to other molecular and clinical features that have been proposed by our group and others (monocyte count at diagnosis and TME subtype). The impact of treatment was also included (COTC-021/022 trial; SOC vs. SOC with sirolimus). Four different analyses were conducted based on the discovery cohort (n = 55 dogs from the COTC-021/022 trial) and thus represent stratification of a small number of patients based on a selected group of molecular determinants (MYC AMP, PTEN loss, TP53-mut vs. TP53-WT, and methylation cluster assignment). In TP53-stratified samples, primary TME subtype showed a significant association with overall survival after adjustment for other clinical factors. In methylation-stratified samples, a baseline monocyte count >400/μL was significantly associated with both DFI and overall survival. In PTEN-stratified samples, ALP status and sex were significantly associated with overall survival after adjustment for other clinical factors. An observation consistent with prior data from our group was the decreased risk of poor survival with the IE-ECM TME subtype in dogs when stratified by MYC status. These analyses provide interesting observations about the increased risk of a poor outcome in specific patient subsets but require additional validation to determine causal links, not just statistical associations, in a larger set of cases.

The major findings within this WGS dataset indicate several similarities shown to occur in human osteosarcoma and are largely consistent with prior studies in canine osteosarcoma. Beginning with TMB (mean 0.71 per Mb), as well as CNVs, AMPs of the MYC, PDGFRA, and KIT loci on CFA13 were the most prevalent AMPs identified in our dataset, aligning with previous reports (10, 35). The AMP of MYC was initially documented in osteosarcoma nearly 30 years ago via Southern blotting (57) and has since been implicated in various solid tumors, such as neuroblastoma (58) and medulloblastoma (59). MYC overexpression was a defining feature of the most aggressive medulloblastoma subtype (60, 61) and has been recently shown to be a promising prognostic biomarker in osteosarcoma (9). In human osteosarcoma, MYC oncogene alterations have been linked to disease biology through effects on the MAPK pathway (62). Other notable DELs affecting critical genes, including CDKN2A/B, DLG2, CCNE1, BRCA2, RB1, PIK3CG, TP53, and PTEN, were identified in our cohort and likely contribute to disease progression (10, 35).

Our cohort also exhibited alterations in genes critical for DNA damage repair pathways, such as BRCA1/2, the RAD51 complex, PALB2, ATR, CHEK1, CDK12, NBN, RAD50, SLX4, BARD1, FANC family genes, BRIP1, RECQL4, and MCPH1. Although the prevalence and functional consequences of these genetic changes in canine osteosarcoma need further investigation, their occurrence mirrors trends observed in human osteosarcoma, indicating potentially shared therapeutic targets (63). Indeed, current clinical investigations in human osteosarcoma are exploring DNA damage response–targeted therapies, such as the combination of olaparib as a PARP inhibitor and ceralasertib as an ATR/CHK1 inhibitor, particularly in tumors with genomic instability (64). Furthermore, we also observed frequent alterations in several cell cycle–related genes, including CDKN2A (DEL, 60.0%), CDKN2B (DEL, 54.6%), CCNE1 (AMP, 47.3%), CCND3 (AMP, 38.2%), CCND1 (DEL, 36.4%), CDK4 (AMP, 21.82%), and CDK6 (AMP, 20%). These recurrent alterations highlight the central role of cell-cycle dysregulation in this disease and suggest potential targets for combination therapy approaches, such as CDK inhibitors (targeting CDK4/6 and CDK2) to block cell-cycle progression, various DNA-damaging chemotherapeutic drugs (such as doxorubicin and cisplatin) that interfere with DNA replication in rapidly dividing cells, and novel targeted agents that disrupt specific cell-cycle pathways, such as CCNE1 (65, 66).

It is worth noting that as an example, a previous clinical investigation in a limited group of extensively pretreated human osteosarcoma patients (N = 23) did not result in any significant responses to the CDK4/6 inhibitor palbociclib given as a single agent, even with the presence of cell-cycle gene alterations (67). This outcome emphasizes the difficulty in targeting isolated gene mutations in a disease like osteosarcoma, characterized by widespread genomic heterogeneity, while also highlighting the opportunity to leverage canine datasets that comprise linked genomic alterations and clinical outcomes to determine what combinations of alterations are responsible for sensitivity and/or response to specific therapies.

We examined specific mutations and associations with other features of the canine tumor cohort for which other data types were available (Fig. 3). In our study, somatic missense mutations were most frequently identified in MCPH1 (63.64%), TP53 (58.18%), and CDKN2B (58.18%). The prevalence of TP53 mutations ranges from 47% to 90% in humans (68) and 24% to 47% in canine osteosarcoma cases (69). Mutations in TP53 can result in reduced or aberrant TP53 activity, including loss of tumor-suppressive functions (8, 70, 71), and are linked to poor prognosis in human osteosarcoma patients (72). Previous work has demonstrated the importance of robust immune infiltration in improved outcomes for canine and human osteosarcoma patients (20, 21). Furthermore, TP53 DEL or mutation impairs T-cell recruitment and activity, contributing to immune evasion by cancer cells (71, 73). In line with prior findings, our study also observed a significant correlation between TP53 mutation and the ID TME subtype, which is the most common TME designation in canine osteosarcoma (8). Prior work from our group has demonstrated the independent effect of TME subtype on outcomes, which carries implications for the use of immunotherapy in both canine and human patients. Finally, the presence of TP53 mutations has been suggested to work synergistically with RB1 alterations (74). Although RB1 somatic mutations are somewhat common in human osteosarcoma (30%–75%), RB1 copy-number loss seems to be more common in canine osteosarcoma (29%; refs. 35, 52). In our cohort, we identified RB1 copy-number losses in 43.64% of the canine osteosarcoma samples and 16.36% of samples with DELs in both RB1 and TP53 genes. We have observed that canine osteosarcoma less commonly carries direct RB1 mutations, suggesting that RB pathway disruption is occurring through alternative mechanisms such as CDKN2A DEL, TP53 mutation, or PTEN loss, all of which were common findings in our dataset. These differences may reflect species-specific evolutionary routes to achieve the same biological outcome.

Aligning with previous observations (35, 48, 52, 75), our study identifies significant alterations in canine osteosarcoma that mirror those in human osteosarcoma, including DLG2 and DMD DELs and SETD2 BNDs. These shared oncogenic gene variations highlight the value of osteosarcoma-bearing pet dogs as translational models for advancing therapeutic development for the benefit of both canine and human patients. Further assessment of somatic mutation patterns can offer valuable insights into the causes of cancer. For instance, Tseitline and colleagues (76) used somatic mutation patterns as functional biomarkers, developing a machine learning classifier to predict ERCC2 deficiency based on genome-wide mutation distributions. This approach suggests that pattern-based diagnostics could be used in the future to develop biomarkers for osteosarcoma. Although SNV and SV analysis was not performed for the expansion cohort, the main findings from the discovery cohort remain robust. Future studies will expand mutation profiling to additional samples, with efforts to ensure coverage across all analytical platforms for every sample.

The evolutionary history of a tumor and the factors that drive its development and progression can be identified by COSMIC signatures (37). Our study revealed that the most frequent mutation signature was COSMIC-1, a pattern commonly found in many human cancers and associated with the individual’s age (77, 78). Patient age was also significantly correlated with this signature in our cohort (Fig. 3H). The predominance of COSMIC-1 here contrasts with reports from pediatric and adolescent human osteosarcoma, in which the most prominent mutational signatures are more reflective of SV and DNA repair deficiency (79). This may reflect species differences in disease natural history, in which osteosarcoma arises in both species at a comparable absolute age, but in relatively older dogs, which potentially allows greater accumulation of age-associated mutations, a decrease in immune function (80, 81), and potential differences in cancer-initiating events. Further comparative analyses are needed, ideally across a wider age range of dogs with osteosarcoma, but this raises intriguing questions about the natural history and biology of osteosarcoma across species. Other signatures (i.e., COSMIC-8, COSMIC-9, COSMIC-17b, and COSMIC-40) identified in our data showed similarities to those previously reported in human osteosarcoma (82), indicating shared characteristics between the mutational signatures across both species.

The methylation analysis presented in Fig. 4 not only identifies potentially prognostically relevant epigenetic subgroups in canine osteosarcoma but also reveals key similarities to human cancers, highlighting shared mechanisms of disease progression and immune evasion. In both species, global hypomethylation is associated with more aggressive tumors, immune exclusion, and poor clinical outcomes (16, 17). Similarly, global hypermethylated subtypes tend to correlate with better prognosis (83) and more favorable TME characteristics (84). Recently, the Assay for Transposase-Accessible Chromatin using sequencing profiling identified early osteoblast-derived and late osteoblast-derived osteosarcoma subtypes, reflecting genes associated with early and late stages of osteoblast differentiation, which coexist within individual tumors and exhibit differential responses to therapy (80). These findings provide insight into our differentially methylated patient populations and may elucidate the underlying, potentially druggable programs that define these clusters of patients (85).

The loss of PTEN, a critical tumor suppressor, is frequently observed through mutation or loss in various human cancers. In human osteosarcoma, homozygous copy-number loss of PTEN has not been identified; instead, increased methylation of the PTEN promoter exhibits an association with elevated cell-cycle gene transcripts (86). Another study also showed that osteosarcoma samples often display absent or low PTEN expression, typically indicating an advanced tumor stage and a poorer prognosis (87). This provides further evidence, as has been suggested in prior reports, that PTEN-deleted osteosarcoma in dogs and PTEN-intact but silenced osteosarcoma in humans might represent convergent evolutionary mechanisms that result in clinical convergence as evidenced by the strong histologic and clinical similarities in osteosarcoma between species. Furthermore, as indicated in prior work (51) about the comparative features of PTEN between canine and human osteosarcoma, the hypermethylation of the PTEN promoter in humans versus the loss of the canine chromosomal 26 segment that houses PTEN in dogs highlights an important therapeutic implication: that humans would potentially benefit from demethylating therapy that could restore PTEN function, whereas dogs would not.

This study allowed us to use multiomic data integration to explore other relationships between gene dose and expression. For example, we found significant hypomethylation of CHEK1 (P = 0.0000024) in primary osteosarcoma (Table S8), with copy-number loss in 27% of samples. As expected, this hypomethylation was accompanied by increased mRNA expression (log FC = 0.54, P = 0.02) and G₂–M checkpoint pathway enrichment in our MYC-amplified cohort, which is discussed further below. However, other hypomethylated genes, such as PIK3CG (CNV loss, 47% of samples), DMD, RET, and FANCF, demonstrated low expression, suggesting an alternative regulatory mechanism. Further investigation of specific methylation patterns within the context of the multiome may provide additional insights into the mechanisms that facilitate osteosarcoma progression. Furthermore, the tumor suppressor gene DLG2 shows copy-number loss in the discovery cohort (36/55, 65.45%) as well as in the expansion cohort (131/209, 62.68%), along with SV (DEL, 18/55, 32.72%) and decreased mRNA expression (log FC = −0.639, P = 0.05).

In this study, we also present a focused examination of MYC (Fig. 5), a proto-oncogene that encodes a transcription factor that influences the expression of >15% of the human genome (88). Growing evidence indicates that MYC plays an essential role in the regulation of metabolic reprogramming in cancer cells, facilitating the rapid production of energy substrates and building blocks necessary for sustained, uncontrolled proliferation (12–15). MYC also regulates aerobic glycolysis, indirectly controlling key genes like GAPDH and TPI (12). MYC has been shown to enhance mitochondrial oxidative phosphorylation and the generation of reactive oxygen species, processes vital for sustaining cancer stem cells in breast cancer (89). Our transcriptomic analysis of canine primary osteosarcoma tumor samples supported these observations but in the context of osteosarcoma, demonstrating significant upregulation of oxidative phosphorylation (NES = 1.89 and P = 0.00000062) and glycolysis pathways (NES = 1.38 and P = 0.006), as well as increased expression of GAPDH (log FC = 0.53 and P = 0.02) and TPI1 (log FC = 0.34 and P = 0.03). Furthermore, MYC has been reported to regulate the G₂–M transition by activating key cell-cycle proteins, including cyclin B1 and CDK1 (90). In our study, MYC-amplified samples demonstrated upregulation of CDK1 (log FC = 0.84 and P = 0.000089) and G₂–M checkpoint pathway genes (NES = 2.82 and P = 3.52e-23). Suppression of immune response pathways and reduced enrichment for B and monocyte immune lineage cells was observed in MYC-amplified samples. These findings align with existing research suggesting that an immunosuppressive TME is driven by MYC, which may contribute to immune evasion and progression (91–94). Another potentially contributing observation is that the gene-dense region commonly lost on canine chromosome 11 in our dataset, which includes CDKN2A/B and MTAP, may contribute to an immunosuppressive program in canine osteosarcoma. The loss of MTAP results in altered tumor metabolism through the accumulation of methylthioadenosine, which collectively affects the TME by altering cell signaling, epigenetic processes, and chemokine/cytokine production (95). This observation may open lines of investigation into targeting the immunometabolism of osteosarcoma, including synthetic lethal therapeutic strategies that target glutamine, methionine, and purine metabolism (55, 95). Indeed, prior work from our group (96) suggests that pharmacologic targeting of glutaminase-1 via CB-839 affected metastatic progression in mouse models of osteosarcoma metastasis, but further functional validation in immunocompetent models is needed.

Similar to that described in pediatric, adolescent, and young adult osteosarcoma patient cohorts (1, 9–11), our analysis identified a significant association between MYC AMP and unfavorable outcomes in canine osteosarcoma patients, positioning it as a potential biomarker for risk stratification in clinical trials and, upon validation, in clinical practice. Furthermore, patients with MYC AMP may benefit from potential therapeutic strategies that involve indirect inhibition through CDK7 inhibitors and BET bromodomain inhibitors (1, 9, 97, 98), alongside translation inhibitors and metabolic interventions targeting MYC-driven states (99). Second, inhibitors against cell-cycle regulators such as CCNE1, CDK2, and CDK4/6 could be effective when combined with DDR inhibitors and/or chemotherapy or chemoradiation. Lastly, therapeutics that could improve antitumor immunity, such as STING agonist therapy and/or metabolic reprogramming of the TME, could potentially restore immune surveillance and enhance the immunotherapy response (100, 101).

In conclusion, we present a comprehensive genomic landscape of osteosarcoma, identifying distinct subtypes of canine osteosarcoma based on an integrated analysis of genomic, transcriptomic, and epigenomic data. The exploratory analyses presented here further highlight similarities between human and canine osteosarcoma and provide interesting hypothesis-generating comparisons worthy of future study in prospective canine clinical trials. These subtypes are characterized by consistent patterns of oncogenic alterations, immune microenvironment composition, and genomic instability. This work offers an introductory framework for future investigations into mechanisms, therapeutic development, and patient stratification, benefiting both veterinary and comparative oncology research relevant to human osteosarcoma. Although further functional validation and clinical correlation are necessary and generalization of the findings beyond dogs receiving SOC with or without adjuvant sirolimus may not be possible, this study provides a robust comparative platform for osteosarcoma research and can inform the design of stratified clinical trials in canine patients with translational value for humans.

Supplementary Material

Supplementary data for this article are available at Clinical Cancer Research Online (http://clincancerres.aacrjournals.org/).

Translational Relevance.

Multiomic profiling of canine osteosarcoma provides new avenues for the creation of a molecular framework for the disease and further highlights the value of osteosarcoma-bearing pet dogs as translational patient models for humans. Future studies should focus on the validation of candidate biomarkers for risk stratification and the clinical assessment of novel therapeutics that could benefit both canine and human patients.

Data Availability

The data generated in this study are available within the article and its supplementary data files, as well as public repositories. WGS data: discovery cohort (PRJNA1374203) and expansion cohort (PRJNA1367604); WGBS data: GSE311005; and expression profile data: GSE238110. All datasets have been submitted to SRA and GEO and will be publicly available upon publication. Code used in this study is available at https://github.com/Anjaligarg006/multiomics_OS.git.