Transcriptional profiling of canine osteosarcoma identifies prognostic gene expression signatures with translational value for humans

Abstract

Canine osteosarcoma is increasingly recognized as an informative model for human osteosarcoma. Here we show in one of the largest clinically annotated canine osteosarcoma transcriptional datasets that two previously reported, as well as de novo gene signatures devised through single sample Gene Set Enrichment Analysis (ssGSEA), have prognostic utility in both human and canine patients. Shared molecular pathway alterations are seen in immune cell signaling and activation including TH1 and TH2 signaling, interferon signaling, and inflammatory responses. Virtual cell sorting to estimate immune cell populations within canine and human tumors showed similar trends, predominantly for macrophages and CD8+ T cells. Immunohistochemical staining verified the increased presence of immune cells in tumors exhibiting immune gene enrichment. Collectively these findings further validate naturally occurring osteosarcoma of the pet dog as a translationally relevant patient model for humans and improve our understanding of the immunologic and genomic landscape of the disease in both species.

Fig. 1. Computational approach to mRNAseq datasets.

a General workflow for application of GS-1, GS-2 gene signatures, and downstream analyses. mRNA is filtered down to a gene expression signature followed by K-Means clustering independent of any associated outcome data. Clusters are subjected to several modes of secondary analysis to discern differences in clinical outcome, differentially expressed genes and pathways, and deconvolution of immune cell types with immunohistochemical validation in tissue. b General workflow for Iterative Search Algorithm (ISA) signatures devised from canine DOG² cohort mRNAseq data. K-means clustering is performed on ssGSEA results that are used as input to a secondary DEG analysis. ISA bi-clustering is then performed on the DEGs, creating new gene signatures. These new ISA signatures are then applied to the human TARGET dataset to determine if the signatures define clusters with meaningful differences in survival.

Fig. 2. Canine osteosarcoma gene signatures are prognostic within the DOG² dataset.

a Expression profile across GS-1 clusters on DOG² data. The two distinct clusters identified by GS-1 have significantly different Kaplan–Meier curves for disease free interval (DFI, b) and overall survival (OSv, c), both given in days from diagnosis. Expression profiles across GS-2 clusters on DOG² data (d). The clusters identified by GS-2 demonstrate significantly different DFI (e) and OSv (f).

Fig. 3. Canine osteosarcoma-derived gene signatures are prognostic in human osteosarcoma.

Expression profile of the canine-derived gene signature GS-1 applied to clusters on TARGET data (human osteosarcomas) (a). The groups identified by GS-1 have significantly different Kaplan–Meier curves for progression free survival (PFS, b) and overall survival (OSv, c). d The expression profile of the canine-derived gene signature GS-2. In contrast to GS-1, GS-2 clusters do not have a distinct difference in PFS (e) but do demonstrate significant differences in OSv (f). Median PFS and OSv is given in days from diagnosis.

Fig. 4. Transcriptionally-defined clusters are enriched for specific cellular processes.

Normalized enrichment scores for gene set enrichment analysis (GSEA) of the top 20 pathways over-represented in the favorable prognosis group when compared to the poor prognosis group when clustered by GS-1 in (a) DOG², (b) TARGET, (c) TARGET Non-metastatic patients, and (d) TARGET metastatic patients. Dot size is representative of the number of genes in the pathway and color is indicative of FDR-q value calculated by software, red line is indicative of significance cutoff. Similar plots for GS-2 and pathways over- represented in the poor prognosis group can be found in Supplemental Figs. 3–5.

Fig. 5. Osteosarcoma sub-populations are defined by distinct immune cell populations.

M0 Macrophage CIBERSORTx scores are significantly higher in the poor prognosis (PP) group compared to the favorable prognosis (FP) group in (a) DOG² and (b) TARGET. This is the opposite of what is observed for M2 macrophages where higher CIBERSORTx scores are consistently seen in the FP group in (c) DOG² and (d) TARGET. Spearman correlations between CIBERSORTx scores in the FP group for DOG² (e) suggest an adaptive immune response characterized by highly coordinated populations of T-cells. Likewise, the same can be seen for cytotoxic lymphocyte response (activated NK and dendritic cells, T-helper cells, and CD8 T cells) in the FP TARGET group (f).

Fig. 6. Immunohistochemical staining supports use of gene signatures as surrogate for immune infiltrates in canine osteosarcoma.

For immunohistochemical (IHC) analysis, a subset of cases from the GS-1 Poor Prognosis (Green) and Favorable Prognosis (Orange) clusters were selected. a Based on IHC labeling of antibodies listed along the x-axis, cases were categorized using a 3 point scale as immune high (3+), immune intermediate (2+), or immune low (1+). Iba1 was not scored in one sample (identified in gray) due to poor tissue sectioning. All other tissues were included as indicated. The adjacent heatmap shows gene expression for the IHC cases based on GS-1 hierarchical clustering of all samples and illustrates the relationship between the GS-1 clusters and IHC category. Asterisks indicate significant differences between clusters in IHC quantification with *p < 0.05 and **p < 0.005. b Example images of the IHC labeling from the Poor Prognosis (Green) and Favorable Prognosis clusters (Orange). c Example images of IHC labeling for CD3, CD20, and CD204 in human OS samples. Scale bar = 50 µm.*IHC labels for Figs. B and C represent IHC chromogen (either red or brown) used to label positive cells.

Fig. 7. Bi-clustering defines additional gene signatures and sub-populations in canine osteosarcoma.

DOG² gene expression profiles for (a) ISA signature 1, (b) ISA signature 2, (c) ISA signature 3, and (d) ISA signature 4. Corresponding Kaplan–Meier curves of ISA sample bi-cluster versus all other samples for ISA signature 1 (e), ISA sample bi-cluster 2 (f), ISA signature 3 (g), ISA signature 4 (h). Y-axis is Overall Survival, with medians reported in days from diagnosis.

Fig. 8. Novel canine ISA-derived gene signatures are prognostic in human osteosarcoma.

TARGET gene expression profiles for (a) ISA gene signature 1, (b) ISA gene Signature 2, and (c) ISA signature 3. Corresponding Kaplan–Curves for K-Means clusters formed from (d). ISA gene signature 1, (e) ISA gene signature 2, and (f) ISA gene signature 3. *ISA gene signature 4 was omitted as it did not form consistent clusters. Y axis is Progression Free Survival. Medians are reported in days from diagnosis.

Fig. 9. Novel canine ISA-derived gene signature are prognostic for progression free survival in human osteosarcoma in patients without metastatic disease at the time of diagnosis.

TARGET (non-metastatic patients) gene expression profiles for (a) ISA gene signature 1, (b) ISA gene Signature 2, and (c) ISA signature 3. Corresponding Kaplan–Curves for K-Means clusters formed from (d) ISA gene signature 1, (e) ISA gene signature 2, and  (f) ISA gene signature 3. *ISA gene signature 4 was omitted as it did not form consistent clusters. Y axis is Progression Free Survival. Medians are reported in days from diagnosis.