Spatial Transcriptomic Landscape of Canine Oral Squamous Cell Carcinoma

ABSTRACT

Canine oral squamous cell carcinoma (COSCC) is the second most common oral tumor in dogs and the most relevant for comparative human trials as a spontaneous large animal model of disease. Historical genomic work has focused primarily on bulk sequencing. The present study describes the complete transcriptomic landscape of COSCC with spatial distinction between the surface tumor, deep invasive tumor, peritumoral dysplastic epithelium, and tumor microenvironment compared to matched normal oral samples. Each region demonstrated distinct molecular signatures. Genes related to epithelial growth factor (EGFR) and epithelial‐mesenchymal transformation (EMT) were upregulated in both peritumoral dysplasia and surface cancer. Additionally, the KRAS gene set, KRT17, and SSP1 were enriched in cancer. We identified five genes that represent dysplastic lesion with high potential for malignant transformation (FZD4, GAS1, HACD2, NOG, and SLC39A6). Also, three genes, SFRP4, FZD1, and IL34 represented a specific signature of the invasive portion of the COSCC that should be explored for prognostic value as a biomarker of malignancy. Lastly, we verified the immunomodulatory tumor microenvironment detecting an increase in macrophages and an abundance of IL‐10 secretion. The other predominant leukocytes were T‐cells, with CD4+ T‐cells being the most prevalent. CD4+ T cells expressed transcripts for both stimulatory (Inducible T‐cell Co‐Stimulator (ICOS) and inhibitory molecules (CTLA4). The observed high CTLA4 suggests that this inhibitory signal may be preventing a robust antitumor immune response. Taken together, this study identified multiple targets to be explored for biomarkers of malignancy, prediction of tumor behavior, and potential targets for development of novel therapies.

Abbreviations

COSCC

canine oral squamous cell carcinoma

DEGs

differentially expressed genes

EGFR

epithelial growth factor receptor

EMT

epithelial‐mesenchymal transformation

GSEA

gene set enrichment analysis

HNSCC

head and neck squamous cell carcinoma

LMM

linear mixed model

ROI

regions of interest

TAM

tumor associated macrophage

TME

tumor microenvironment

1. Introduction

Canine oral squamous cell carcinoma (COSCC) is the second most common oral cancer in dogs [1]. It presents in a biphasic disease model, with peaks in juvenile and geriatric populations [2, 3, 4]. Surgery is the standard of care, and if performed in early‐stage disease, has the potential to result in long‐term remission [2, 3]. However, in more advanced cases, surgery is associated with the potential for high functional and cosmetic morbidity [5] and, despite aggressive surgical dose, may still result in incomplete surgical margins and local recurrence [6]. Further, there is a subset of lesions with complete resection (clean margins) that will recur, likely due to molecular changes in the peri‐tumoral tissues that are not detected with traditional pathology techniques (e.g., field cancerization or pre‐malignant high‐grade dysplasia) [7, 8]. An improved understanding of molecular changes that are occurring in the tumor and peri‐tumoral dysplastic epithelium, surrounding tumor microenvironment (TME), including immune cells, mesenchymal stroma, and signaling molecules, relative to matched normal tissue could allow tailored interventions for COSCC and identification of early biomarkers of pre‐malignant disease. This may ultimately help to improve treatment paradigms and inform novel treatment strategies.

Historical molecular work in COSCC has shown that the primary upregulated genes are associated with cell cycle alterations, protein kinase activity, RAS signaling, TGF‐β signaling, and epithelial‐mesenchymal transformation (EMT), similar to their human counterparts [9, 10, 11]. However, use of bulk molecular techniques in historical work reflects a mixture of all cells present; thus, these techniques cannot accurately differentiate epithelial cancer cells from the remaining non‐neoplastic cells. A recent study utilized laser capture microdissection to ensure that only neoplastic epithelial cells were captured for transcriptomic profiling and compared them to matched normal counterparts, effectively circumventing the limitations of bulk sequencing [12]. However, questions remain about tumor heterogeneity (i.e., surface tumor vs. deep invasive front) and whether different tumor areas possess distinct molecular signatures or contributions by the TME. Further, the immune environment of COSCC has also been recently described using immunohistochemistry and NanoString nCounter gene expression analysis [13], yet these findings were not assessed in the context of the remainder of the tumor and peritumoral regions. Thus, critical gaps remain in our understanding of the molecular and regional landscape of COSCC.

This study aimed to add to the existing body of literature by utilizing spatial transcriptomics to define high‐resolution molecular signatures of different spatial phenotypes of COSCC, including the surface of the tumor, its deep invasive front, the TME, and peritumoral epithelial dysplasia compared to patient‐matched normal tissue.

2. Materials and Methods

2.1. Clinical Samples

This study evaluated nine formalin‐fixed paraffin‐embedded (FFPE) samples from dogs with naturally occurring COSCC that were surgically excised. All animal procedures were conducted upon institutional protocol review and approval (UC‐Davis IACUC # 23110). Before pathologic sectioning the soft tissue was removed from the underlying bone, ensuring tissues remained in formalin for only 24–48 h to preserve high quality RNA at the time of paraffin embedding and allowing the entire soft tissue specimen to be serially sectioned every 5 mm. Each serial section was routinely processed, stained with hematoxylin and eosin (H&E), and reviewed by a board‐certified pathologist (B.G.M. and N.V.). This approach allowed us to select slide sections that featured the surface of the tumor, its deep invasive front, the TME, peritumoral dysplasia, and normal tissue on a single block. Patient characteristics are shown in Table 1.

Table 1.

Patient and tumor information for formalin‐fixed paraffin‐embedded samples utilized for digital spatial transcriptomic profiling sequencing.

	Dog signalment	Variant	Tumor stage	Tumor location	Locoregional metastasis, staging performed	Distant metastasis, staging performed	All phenotypes sequenced
1	6 year MC Siberian Husky	Conventional SCC	T1	Rostral mandible	No, histology SLN per ICTL	No, thoracic CT	No, peri‐tumoral dysplasia not sequenced
2	8 year FS Labrador Retriever Mix	Conventional SCC	T2	Rostral mandible	No, histology SLN per ICTL	No, thoracic CT	Yes
3	2 year MC Australian Shepherd	Papillary SCC transformed from viral papilloma	T3	Caudal mandible	Unknown, owner declined LN screening due to low likelihood of spread with papillary variant	No, thoracic CT	No, no normal (all papilloma) and no deep tumor
4	3 year MC Golden Retriever	Papillary SCC	T1	Rostral maxilla	No, cytology bilateral MLN	No, thoracic CT	Normal only
5	7 year MC Labrador Retriever	Conventional SCC	Scar Revision, T1 before excisional biopsy at VMTH, no sutures	Caudal mandible	No, histology of SLN on ICTL	No, thoracic CT	No, no surface of deep tumor present for sequencing
6	10 year FS Chinese Crested	Conventional SCC	T2	Caudal mandible	No, histology of SLN on ICTL	No, thoracic CT	Yes
7	13 year FS Maltese	Conventional SCC	T3	Rostral‐caudal mandible	No, histology from CLND	No, thoracic CT	Yes
8	12 year MC Boxer Mix	Conventional SCC	T3	Rostral mandible	No, histology from CLND	No, thoracic CT	Yes
9	10 year MC Pit Bull	Conventional SCC	T2	Rostral maxilla	Unknown, owner declined screening	Unknown, owner declined screening	No, surface tumor was cut off on nanostring slide preparation

Abbreviations: CLND, complete lymph node dissection; CT, computed tomography; FS, female spayed; ICTL, indirect computed tomography lymphangiography; MC, male castrated; SLN, sentinel lymph node; SCC, squamous cell carcinoma; T1, < 2 cm; T2, 2–4 cm; T3, > 4 cm.

2.2. Spatial Transcriptomics

Spatial transcriptomic analysis of COSCC and surrounding nonmalignant tissues was performed on the NanoString GeoMx Digital Spatial Profiler platform [14] utilizing the GeoMx Canine Cancer Atlas assay to assess the expression of 1962 genes and 48 control targets. H&E‐ stained slides were reviewed by a board‐certified pathologist (NVA, BM) to choose the most appropriate section per patient capturing all relevant tissue phenotypes. Unstained tissue Section (5 µm) were then cut from the corresponding FFPE block and mounted on Superfrost slides (Thermo Fisher, Waltham MA). GeoMx DSP analysis was performed by the Genomics Shared Resource (Department of Pathology and Laboratory Medicine, University of California, Davis School of Medicine) and according to NanoString's standard procedures for NGS readout. Briefly, semi‐automated slide preparation was conducted on a Leica BOND RX stainer for deparaffinization, epitope retrieval, and proteinase K treatment to expose RNA targets. The GeoMx Canine Cancer Atlas probe set (UV‐cleavable DNA oligonucleotide barcode‐coupled RNA target probes) was hybridized in situ to the tissue sections by overnight incubation at 37°C. Subsequently, tissue sections were stained with fluor‐conjugated morphology marker antibodies specific for pan‐cytokeratin (PanCK; Alexa 532, Cy3/568 nm channel), CD45 (Alexa 594, Texas Red/615 nm channel), CD3 (Alexa 647, Cy5/666 nm channel), and the SYTO‐13 nucleic acid stain (FITC/525 nm channel, for instrument focus). Slides were scanned at high magnification (20×) using the following exposure parameters: 100 ms for SYTO‐13 and 300 ms for PanCK, CD45, and CD3. Morphology markers facilitated the selection of regions of interest (ROIs) containing approximately 300 cells, as defined by the nuclear stain. Four ROIs were selected for each phenotype (5 phenotypes: surface normal (normal surface epithelium), surface tumor (epithelial surface of the tumor), peri‐tumoral dysplasia (surface dysplastic epithelium next to surface tumor), deep invasive tumor (deep invasive tumor epithelium), and TME (peritumoral mesenchymal stromal tissue infiltrated with CD45+ cells), leading to a total of 20 ROIs per patient (Figure 1). In cases where a certain phenotype was not visualized well due to sample characteristics (i.e., biological) or tissue mounting (i.e., technical), then this phenotype was excluded from analysis. In total, 141 non‐segmented ROIs were analyzed. ROI aspirates were collected following UV photocleavage and ROI‐specific, indexed libraries were prepared from the DSP DNA oligonucleotides using high‐fidelity PCR with GeoMx Seq Code Illumina‐compatible primers. Validated libraries (i.e., having expected amplicon size of 162 bp) were sequenced on an Illumina NovaSeq X Plus Sequencing System using paired‐end sequencing with index reads for the UDIs (2 × 27 bp paired‐end, i7/i5 reads = 8 bp) and with a target sequencing depth factor of 100 clusters/mm².

Figure 1.

ROI selection utilizing visual markers. Representative immunofluorescently‐labeled tissue section encompassing the normal oral epithelium, surface tumor, invasive neoplastic epithelium, and TME (peritumoral mesenchymal stroma). For each tissue phenotype, four regions of interest (ROI) were specified (including on average 300 cells) followed by the gene expressing analyses in these particular ROIs. Blue: nuclear stain (syto 13), green: panCK, yellow: CD45.

2.3. Data Analysis

Raw FASTQ sequence reads were processed through the GeoMx NGS Pipeline Software for conversion to digital counts (DCC files) and subsequent downstream analyses with the GeoMx DSP Data Analysis Suite. Summary statistics through the NGS pipeline included mean raw (6,794,863.81), trimmed (6,762,693.48), stitched (6,704,821.26), aligned (6,475,159.06), and de‐duplicated (100,119.94) reads per ROI. Quality control (QC) parameters were assessed to determine segments/ROIs qualified for analysis and to flag those that did not meet these criteria, including a raw read threshold of < 1000 reads, < 80% read alignment, < 50% sequencing saturation, negative probe count geomean < 10, and a No Template Control (NTC) count > 1000. ROIs were also flagged if the DSP parameters for nuclei count and surface area were less than 50 and 1600 mm², respectively. Biological probe QC was conducted to identify and exclude outlier probes, specifically low outlier probes that had average counts across all ROIs of ≤ 10% of the counts for all probes to that gene and/or global outliers that failed the Grubbs outlier test in ≥ 20 ROIs. Additionally, the limit of quantitation (LOQ) per ROI was set to be calculated at a confidence threshold of 2 standard deviations above the geomean of the negative probes. The data set was then filtered to remove ROIs and genes/targets with abnormally low signal, specifically to retain ROIs with ≥ 10% targets above the expression LOQ threshold and targets with ≥ 5% ROIs that are above the expression LOQ threshold. The filtered data set was normalized using third quartile (Q3) normalization applied to all targets and utilized for all downstream analyses and visualizations.

2.4. Gene Expression Analysis

Differential gene expression between different tumor regions (e.g., surface tumor, deep invasive tumor, TME), peri‐tumoral dysplasia, and normal surface epithelium was tested using a linear mixed‐effect model (LMM) with a random slope and random intercept (patient ID, scan/slide number), followed by Benjamini‐Hochberg correction for multiple comparisons. Differentially expressed genes (DEGs) and associated statistics (log2 fold changes, p values) were visualized with volcano plots (LabeledVolcanoPlot DSP R script). Hierarchical clustering and heatmap visualization of statistically significant DEGs (p < 0.05 or BH‐adjusted FDR p < 0.05) were performed using the pheatmap (Pretty Heatmaps, version 1.0.12) R package. Intersection analysis of DEG sets resulting from LMM analyses (i.e., surface cancer vs. normal epithelium, surface cancer vs. dysplasia, dysplasia vs. normal epithelium) was conducted with the VennDetail (version 1.22.0) R package to identify and visualize the shared and unique subsets and associated expression details. For mechanistic insight, functional pathway analysis with Gene Set Enrichment Analysis (GSEA) [15, 16] of ranked lists of DEGs (e.g., Gene Symbols with p values < 0.05) was performed using the GSEAPreranked tool with default parameters: weighted enrichment statistic, 1000 permutations, and the mean division (meandiv) metric for calculation of the normalized enrichment score (NES). Gene sets hosted at the Molecular Signatures Database (MSigDB) used in these analyses included Hallmark (H) and C6 oncogenic signatures. To visualize concordance and/or divergence in pathway enrichment patterns, enrichment results from selected analyses were compared and visualized with scatterplots (ggplot2, version 3.5.1) of the normalized enrichment scores enrichment scores for the gene sets in each analysis.

Immune cell abundance and relative immune cell proportion were estimated using spatial deconvolution for areas of normal tissue and areas of TME. The ROIs chosen to represent the TME were those determined by visual inspection as having negative staining for panCK and positive staining for CD45 cell infiltrate in the surrounding tumor stroma (Figure 1). Tumor immune cell deconvolution analysis of TME‐derived ROI expression data was performed with the SpatialDecon R script (NanoString) using the SafeTME cell profile matrix consisting of the expression profiles for 18 cell types determined from cell‐sorted and single‐cell RNA‐sequencing analyses [17]. The outputs of the SpatialDecon algorithm include estimated abundance and proportions for each cell type within each ROI. Immune cell proportions are calculated as an immune subset abundance score divided by the total immune cell abundance, where endothelial cells and fibroblasts are not considered immune cells. Principle component analysis was conducted based on immune cell proportions and PC1 loadings were used to determine the most variable immune cell subsets.

Paired comparisons between normal, deep, and TME phenotypes were limited to patients with three ROIs per phenotype (n = 6 patients). Differences in select gene expression or immune cell proportions were assessed with paired two‐tailed t‐test. When multiple comparisons were made, p‐values were corrected using the Bonferroni test. Correlations between gene expression and immune cell proportion were evaluated using Pearson's correlation coefficient. Statistical significance was defined as p < 0.05 for all tests. These statistical analyses were performed using Python 3.11.7 and R version 4.4.0.

3. Results

3.1. Differential Gene Expression Analysis (Log Fold Change > 2) Between Tumor Regions (Surface Cancer, Peri‐Tumoral Dysplasia, Normal Epithelium, and Deep Invasive Cancer)

The overall goal of this study was to define the transcriptomic signatures and associated functional implications of pathologically discrete regions of COSCC tumors using spatial transcriptomic analysis. The patient cohort consisted of nine dogs diagnosed with COSCC (Table 1). NanoString GeoMx digital spatial profiling was performed to measure the expression of immuno‐oncology‐associated genes selected from discrete regions including the surface normal, peri‐tumoral dysplasia, surface tumor, deep invasive tumor, and the TME (Figure 1).

The initial step in this analysis was to determine the gene expression alterations occurring in surface cancer compared to that of normal surface tissue. Differential expression analysis (linear mixed model, LMM) identified 144 differentially expressed genes (DEG) (LMM, p < 0.05), with 42 being upregulated and 102 downregulated in surface cancer, respectively. Subsequent hierarchical clustering of the DEGs and heatmap visualization demonstrated a distinct separation of the ROIs derived from the two regions into two major clusters. While the expression patterns differed significantly between the two groups, they did exhibit intra‐regional variability (Figure 2A). KRT17 (log2FC 4.64), SSP1 (log2FC3.86), and CXCL10 (log2FC3.80) had the highest magnitude of upregulation, while GDA (log2FC −5.05), PLA2G4E (log2FC‐3.80) and KLF4 (log2FC‐3.42) were the most highly downregulated genes. Following p‐value adjustment (FDR), 8 genes remained significant (p < 0.05), with GBP5 (log2FC 2.23), DSE (log2FC 1.80), and CASP4 (log2FC1.08) having the highest fold change (Table S1).

Figure 2.

Differentially expressed genes between surface cancer and normal tissue. Principal component analysis and heat maps showing ROI relationships/variations and clustering patterns of differentially expressed genes between surface cancer versus normal (A), peri‐tumoral dysplasia versus normal (B), and surface cancer versus dysplasia (C).

The next analyses were intended to define the transcriptomic profile of peritumoral dysplasia and determine the DEG signatures that distinguish dysplasia from surface normal and from surface cancer. Seventy‐one DEGs (p < 0.05) were identified when comparing peritumoral dysplasia to surface normal. Of these, 49 and 22 were up‐ and downregulated in peritumoral dysplasia, respectively. In both the results of principal component analysis (PCA) of normalized counts and hierarchical clustering of the DEGs, it can be noted that there is substantial overlap between gene expression between ROIs from the two tissue two histologic states, which is particularly apparent in the PCA plot with most peritumoral dysplasia ROIs blending with those comprising the normal tissue cluster. Similarly, hierarchical clustering did not exhibit distinct separation of the tissue types with only certain genes primarily driving differences in expression and in a heterogeneous manner (Figure 2B). MMP7 (log2FC 2.47), COL1A2 (log2FC 2.14), and KRT14 (log2FC 1.90) had the highest upregulated fold change. Following FDR adjustment, no genes retained significance.

When comparing surface cancer to surface peri‐tumoral dysplasia, 127 DEGs (p < 0.05) were identified, and of these, 26 were upregulated, and 101 were downregulated in surface cancer (Figure 2C). KRT17 (log2FC 4.07), SSP1 (log2FC 3.61), and AREG (log2FC 2.51) had the highest upregulated fold change. Following FDR adjustment, four genes retained significance. PYCR1 (log2FC 1.38) was upregulated, while CD59 (log2FC −1.26), RIN 1 (log2FC −1.14), and FZD4 (log2FC −1.91) were downregulated. Clustering of ROIs was more defined between surface cancer and dysplasia (e.g., separation into two branches of the dendrogram) in contrast to the results obtained for dysplasia and normal. Yet, substantial overlap is still displayed, and a separation is not as distinct as what is observed between surface cancer and normal tissue (Figures 2 and 3). The relatedness of gene expression alterations occurring in surface cancer and dysplastic tissue relative to that of normal were examined using intersection analysis of the DEGs identified above and depicted with a Venn diagram (Figure 3D). Of note, the five DEGs (log fold change > 2, p < 0.05) that were shared between “surface cancer versus peri‐tumoral dysplasia” and “peri‐tumoral dysplasia versus normal” that may represent dysplastic lesions at high risk of malignant change included FZD4, GAS1, HACD2, NOG, and SLC39A6.

Figure 3.

Differentially expressed genes that distinguish surface cancer, dysplasia, and normal tissue. Volcano plots depict the upregulated (red), downregulated (blue), and nonsignificant (grey) differentially expressed genes (LMM, p < 0.05) from group comparisons between surface cancer versus normal (A), peri‐tumoral dysplasia versus normal (B), and surface cancer vs. dysplasia (C). The log2(fold change) and −log10(p value) are indicated on the x‐ and y‐axis, respectively. The 30 most significant DEGs are labeled. Intersection analysis was conducted and the number of differentially expressed genes that overlap between surface cancer, dysplasia, and normal phenotypes is depicted in the Venn diagram (D).

It was also critical to examine the molecular features of the deep invasive tumor regions relative to that of surface cancer. This analysis identified 77 DEGs between the deep invasive front and surface cancer (p < 0.05; 61 upregulated, 16 downregulated). Although DEGs were detected, there was substantial overlap between the phenotypes, as shown by the PCA plot and hierarchical clustering results. (Figure 4) DCN (log2FC 1.88), TNC (log2FC 1.71), and SFRP4 (log2FC 1.56) had the highest upregulated fold change. Following FDR adjustment, 3 genes retained significance: SFRP4 (log2FC 1.56), FZD1 (log2FC 0.98), and IL34 (log2FC 0.92).

Figure 4.

Relatedness and differentially expressed genes between deep invasive and surface cancer. PCA projection, volcano plot, and clustering heatmap showing the differentially expressed genes between deep invasive and surface cancer.

3.2. Gene Set Enrichment Analysis Between Regions (Surface Cancer, Peri‐Tumoral Dysplasia, Normal, and Deep Invasive Cancer)

Gene Set Enrichment Analysis (GSEA) was performed to gain insight into the functional implications of the transcriptional alterations that were identified by comparative analyses of the different tumor regions and normal tissue. Notable enriched gene sets in surface cancer compared to matched surface normal were those associated with upregulated KRAS signaling, EMT, MEK upregulation, EGFR upregulation, MTOR upregulation, angiogenesis, and interferon (IFN) response. Significant enriched pathways in surface dysplasia were similar to surface cancer and included upregulation of MEK, EGFR, KRAS, IFN response, and EMT. Enriched pathways that were found in both surface cancer and dysplasia, but more substantially upregulated in surface cancer, included MEK, EGFR upregulation, IFN response, EMT, and angiogenesis. Notable pathway gene sets that were significantly enriched in surface cancer, but not seen in dysplasia included the hallmark KRAS gene set, and MTOR upregulation (Table 2, Figure 5). No notable clinical pathways separated deep from surface cancer, rather the same notable pathways were present, but just further upregulated within the deeper cancer subset (Table 2).

Table 2.

Gene set enrichment analysis results documenting clinically impactful upregulated pathways between different regions.

	Upregulated KRAS	EMT	Upregulated MEK	Upregulated EGFR	Upregulated MTOR	Angiogenesis	IFN response
Surface cancer vs. normal	Hallmark NES: 1.66 p = 0.02 FDR q = 0.1 C6 KRASS 600 NES: 2.49 p = 0 FDR q = 0.015	NES: 2.59 p = 0 FDR q = 2.38 × 10^−4	NES: 2.45 p = 0 FDR q = 0.018)	NES = 1.65 p = 0.06 FDR q = 0.25	NES = 1.92 p = 0.03 FDR q = 0.20	NES: 2.2 p = 0.01 FDR q = 0.012	IFN‐ alpha NES: 3.64 p = 0 FDR q = 0 IFN‐gamma NES: 3.22, p = 0, FDR q = 0
Dysplasia vs. normal	C6 KRAS 600 NES: 2.49 p = 0 FDR q = 0.015	NES: 2.41 p = 0 FDR q = 0.001	NES: 2.71 p = 0 FDR q = 0	NES: 2.1 p = 0.001 FDR q = 0.035	ns	ns	IFN‐ alpha NES: 3.27 p = 0 FDR q = 0; IFN‐gamma NES: 3.73 p = 0 FDR q = 0
Surface cancer vs. dysplasia	Hallmark NES: 1.8 p = 0.022 FDR q = 0.072 C6 KRAS 600 NES: 2.06 p = 0 FDR q = 0.08	NES: 2.39 p = 0 FDR q = 0.006	NES: 1.61 p = 0.07 FDR q = 0.18	ns	NES: 2.4 p = 0.002 FDR q = 0.02	NES: 2.32 p = 0.002 FDR q = 0.009	IFN‐ alpha NES 2.94 p = 0 FDR q = 0 IFN‐gamma NES: 2.08 p = 0 FDR q = 0.02
Deep vs. surface cancer	Hallmark NES: 2.0 p = 0.022 FDR q = 0.009; C6 KRAS 600 NES: 2.06, p = 0, FDR q = 0.08	NES: 2.70 p = 0.0 FDR q = 0.0	ns	ns	ns	NES: 1.32 p = 0.15 FDR q = 0.27	ns

Abbreviations: FDR, p‐value adjustment; NES, normal enrichment score; ns, not significant.

Figure 5.

Functional enrichment analyses. Differential expression analyses were performed for the indicate comparisons as described above and in Materials and Methods. GSEA was performed using the MSigDB Hallmark and C6 oncogenic gene signature collections. Scatterplots depict normalized enrichment scores (NES) from GSEA conducted with MSigDB Hallmark and C6 oncogenic signature gene sets across the indicated comparisons of (1) dysplasia compared to surface normal and (2) surface cancer compared to surface normal (upper panels) and (3) dysplasia compared to surface normal and (4) surface cancer compared to surface dysplasia (lower panels). Gene sets exhibiting statistical differential enrichment or depletion (NES FDR q‐value < 0.05) for the comparisons on the x‐axis are shown in red.

3.3. Tumor Microenvironment

COSCC TME was specifically characterized by selecting corresponding ROIs from each specimen based on pathological and morphological criteria. Since the TME possesses marked cellular heterogeneity, the diverse cellular components were evaluated by performing mixed cell deconvolution analysis of individual ROI data and assessing key immune/inflammatory‐associated genes. Cell deconvolution demonstrated that macrophages, fibroblasts, and endothelial cells predominated (Figure 6A). A high level of variability was seen among different patient tumors, and occasionally between ROIs from the same case, regarding total abundance of infiltrating leukocytes (e.g., macrophages, CD4 and CD8 memory T cells, memory B cells), with certain tumors being much more inflammatory than others (Figure 6A). Yet distinct patterns can be seen among all tumors. When the stromal cells are removed from deconvolution analysis, the primary immune cells that differentiate TME from deep tumor and normal tissue are macrophages (higher in TME, most negative contribution to PC1) and neutrophils (higher in normal, most positive contribution to PC1) (Figure 6B–D). When evaluating macrophage‐specific cytokines, it was found that IL‐10 secretion was significantly higher in TME compared to normal (fold change 2.24, p < 0.0001), and there was a significant correlation between macrophage abundance and IL‐10 secretion on a patient level (Pearsons correlation 0.46, p < 0.001). Conversely, IL‐6 secretion (fold change 4.37, p = 0.10) and PDL1 (fold change 1.5, p = 0.062) were not significantly different between TME and normal on a patient level.

Figure 6.

Immune cell composition of COSCC tumor microenvironment. (A) Mixed cell deconvolution analysis was conducted on the gene expression data for ROIs selected from the TME and surface normal tissue. Results are shown as stacked bar charts of proportion of fitted stromal and immune cells in each phenotype (tumor stroma; normal). (B) Principal component analysis of the scaled immune scores per regions of the tumor and normal tissue. (C and D) The abundance of macrophages (C) and neutrophils (D) in each patient in the tumor microenvironment compared to the deep invasive tumor and normal tissue, where each point represents the average value of three ROIs per patient. * p < 0.05; ** p < 0.001; ns, nonsignificant.

The other prominent infiltrating leukocytes were T cells (especially CD4+ T cells), which predominated over B cells (Figure 7A). CD4+ T‐cell scores were positively correlated with B cell scores (Pearson coefficient 0.52, p < 0.001). While CD4+ T‐cell scores were not significantly higher in the TME compared to normal phenotypes, there was a significant increase in ICOS gene expression (Figure 7A,B). This finding is strengthened by the correlation between CD4 memory score and ICOS expression (Pearson coefficient 0.38, p = 0.00023). Genetic markers of T‐cell exhaustion and inhibitory markers (CTLA4, PDCD1LG2) were also higher in the TME compared to normal (Figure 7C,D), with CD4+ T‐cells being positively correlated with CTLA4 expression (Pearson coefficient: 0.39, p = 0.00021) suggesting that CTLA‐4 may be inhibiting the antitumor immune response (Figure S1). The ratio of CD4:CD8 positive T cell scores were not substantially different between the TME and normal population.

Figure 7.

Immune cell genes and relationship with T cells and Macrophages in COSCC tumor microenvironment, deep tumor, and normal tissue. (A) CD4+ memory T cell frequency, (B) ICOS gene expression, and (C) CTLA4 gene expression in deep invasive tumor, normal tissue, and the TME, where each point represents the average value of three ROIs per patient. (D) Relative counts of clinically impactful immune genes in the TME compared normal epithelium, again averaging across the ROIs per patient. * p < 0.05; ** p < 0.001; ns, nonsignificant. (E and F) Macrophage score correlation with epithelial‐mesenchymal transformation (EMT), showing macrophages are negatively correlated with epithelial marker E‐cadherin (CHD‐1 expression) and positively correlated with the mesenchymal marker vimentin (VIM expression). * p < 0.05; ** p < 0.001; *** p < 0.0001; ns, nonsignificant.

Interestingly, the immune cell composition, particularly the macrophage score was associated with evidence of epithelial‐mesenchymal transformation. Specifically, the macrophage score was positively correlated with vimentin and negatively correlated with E‐Cadherin, suggesting that having a higher population of macrophages may be associated with more aggressive biological behavior and invasion (Figure 7E,F). There was an insufficient number of dogs with metastasis or disease progression to compare the predominant cells in the TME with patient outcomes.

4. Discussion

This is the first study to describe the spatial transcriptomic landscape of COSCC. This study reveals distinct molecular signatures, or lack there‐of, of distinct tumor and peri‐tumor phenotypes. Our findings directly support the recent work describing the transcriptomic signatures of COSCC and its similarities to human HSNCC [12], with additional molecular data spatially refined to the surface and deep invasive cancer, peri‐tumoral dysplasia and the tumor microenvironment.

We confirmed on a functional level (i.e., based on GSEA) that EMT is an essential process of COSCC, being enriched in deep tumor, surface tumor, and peri‐tumoral dysplasia (Figure 5, Table 2). Features of early EMT with distinct molecular signatures may represent dysplastic lesions with a high propensity for malignant change. The importance of EMT in COSCC was supported at a gene expression level in that the “classic” EMT transcription factors ZEB2, SNAI2, and TWIST 2 [18] were all increased in cancer (deep and surface) compared to normal, and absent in the dysplasia phenotype. Yet, the differential expression in these genes was not significant (p > 0.05). The lack of significance of the “classic” transcription factors may be nonrelevant as overall gene signatures of EMT were significantly upregulated. Specifically, the genes from the molecular signature that were primarily driving EMT in COSCC were SPP1, TNFSRF11B, FN1, and COL1A.

Conversely, EMT in peri‐tumoral dysplasia was primarily driven by expression of COL1A1, COL1A2, and COL3A31. It has been suggested in human HNSCC that not all genes in the hallmark gene set carry the same importance nor prognostic information, and a curated list for risk stratification may be more impactful [19]. In dogs, SSP1, specifically, may represent one of the most important genes for COSCC. SSP1 has been historically implicated as having a critical role in EMT in human cancers through the upregulation of TWIST, maintenance of cancer cell stemness, and modification of the surrounding microenvironment [20]. SSP1 has also been suggested to potentially differentiate COSCC from less aggressive canine tumors, including ameloblastoma [11]. SSP1 may play a critical role in differentiating high‐grade peritumoral dysplasia with early features of EMT that should be closely monitored for malignant change, from actual neoplastic cells. Another distinguishing feature between the peritumoral dysplasia and COSCC is the differential expression of cytokeratins. KRT17 was heavily upregulated in COSCC, mirroring historic canine and human genomic analysis [12, 21]. It has been suggested that upregulated KRT 14 and 17 are needed to sustain a stem cell like proliferative epidermal cell phenotype. Accordingly, expression of KRT17 was the primary gene differentiating surface cancer from dysplasia.

Other key pathways upregulated in COSCC included EGFR and KRAS (Figure 5, Table 2). It has been well documented that EGFR overexpression is ubiquitous in human HNSCC, although treatments focused on EGFR inhibition result in only moderate improvements in outcome [22, 23, 24]. In addition to utilizing small molecule therapeutics to target EGFR, a more promising capitalization of this feature is to bind upregulated EGFR with theragnostic agents to improve surgical outcomes for HNSCC [25, 26, 27, 28]. Similar options for targeting EGFR for optical imaging agents may be available in canine patients, given the similar upregulation of EGFR. Interestingly, in humans, the upregulation is usually confined to EGFR, with PI3KA occasionally being the downstream pathway affected. Downstream RAS pathway upregulation is rare [29]. This may represent a species difference or be an underrecognized impactful pathway, as The Cancer Genome Atlas (TCGA) is not focused at a spatial transcriptional level. Further, EGFR, but not the hallmark KRAS GSEA gene sets, were upregulated in dysplasia, suggesting this mechanistic change occurs late in the course of disease. The hallmark KRAS gene set, in combination with gene upregulation of SSP1 in the KRT17 (proliferative keratinocyte compartment), may help to distinguish high‐grade epithelial dysplasia from early COSCC. Additional genes that may help identify high‐grade dysplasia that is likely to shift to neoplasia include FZD4, GAS1, HACD2, NOG, and SLC39A6 (Figure 3D) based on these being the shared DEGs between dysplasia vs. normal and dysplasia vs. cancer. These genes should be further investigated as a potential biomarker of dysplastic tissues with potential to become COSCC, and a more extensive study size is required to validate a promising gene marker/set.

Interestingly, there was very little difference between the surface cancer and deep invasive front on a molecular level. This is promising from a biomarker and therapeutic perspective, as it does not elucidate different pockets of resistance within the tumor. However, despite the high resolution of the GeoMx DSP digital spatial transcriptomic platform, it is still not at the resolution of single‐cell sequencing; thus, it may underestimate the heterogeneity of the phenotypes.

Our findings on the predominant infiltrating leukocytes in the TME are similar to recent work describing the immune landscape of COSCC [13]. We confirmed that macrophages and T cells represented the highest proportion of tumor‐infiltrating leukocytes, although macrophages were more prominent in our population. (Figure 6) We also confirmed variability among different tumors, with some tumors being more inflamed (higher abundance of inflammatory immune cells) than others. Yet, on a matched patient level (both normal and TME in the same patient), distinct trends were identified regarding the macrophage, neutrophil, and CD4+ T cell population regardless of total immune cell abundance, suggesting a similar immune response despite patient variation.

Our findings suggest that within the TME, CD4+ T cells express both costimulatory (ICOS) and inhibitory (CTLA4) molecules similar to what is observed in the work by Boss et al., and in human HNSCC [30]. The strongest correlation between immune subsets existed between CD4+ T cells and B cells, suggesting a coordinated adaptive immune response. CD4+ T cells negatively correlated with neutrophils, suggesting a suppressive immune response. Further supporting this, TAMs appear to contribute to the immunosuppressive environment directly. The strong correlation between macrophage presence and IL‐10 transcript levels suggests that macrophages mediate immune suppression through cytokine production. Human HNSCC immune infiltrates are comprised primarily of M2 macrophages, and the level of M2 macrophages in the TME is associated with the presence of metastasis and patient survival [31, 32, 33]. Within our population, we found a correlation between the macrophage score and the presence of EMT, suggesting that TAMs are associated with more aggressive biological behavior. We were unable to compare the TME cell population and patient features (e.g., presence of lymphatic or pulmonary metastasis, short survival time) as metastasis were absent in this patient cohort. In addition, the majority of patients (8/9) included in the study are still alive, with only one dying from a non‐cancer‐related event; the role of TAMs in immunosuppression and patient outcomes should be further explored in COSCC.

Limitations of this study include a small sample size with heterogenous cancer and patient features as well as the capacity (1962 immuno‐oncology‐associated genes) of the Canine Cancer Atlas. A larger sample size focused on a distinct subset of COSCC, such as T1 conventional SCC only, may reveal further insight into the molecular landscape. Yet, although the sample size is small, this study reveals distinct molecular signatures (or lack there‐of) of tumor and peri‐tumor phenotypes and strengthens the current body of literature for COSCC.

5. Conclusions

In COSCC, peritumoral epithelial dysplasia shares molecular changes similar to surface cancer, specifically the presence of EMT and EGFR expression. Molecular signatures that help to differentiate dysplasia from cancer include the hallmark KRAS gene set, KRT17, and SSP1. Five genes that may identify dysplastic lesions that are highly likely to transform were identified by comparing shared DEGs between the phenotypes, including FZD4, GAS1, HACD2, NOG, and SLC39A6. Use of these genes as biomarkers for early cancer should be explored. TAMs and CTLA4 deficiency may contribute to immunosuppression in COSCC and represent prognostic indicators of more aggressive behavior and potential immunotherapeutic targets.

Author Contributions

Stephanie Goldschmidt: conceptualization, data acquisition, data analysis, primary manuscript draft, manuscript editing, funding, supervision. Clifford G. Tepper: data analysis, manuscript draft of bioinformatics methods, manuscript editing. Jack Goon: data analysis, manuscript draft of immune subset methods, manuscript editing. Maria Soltero‐Rivera: data analysis, manuscript editing. Robert Rebhun: data analysis, manuscript editing. Andrew C. Birkeland: conceptualization, manuscript editing, funding, supervision. Xiao‐Jing Wang: conceptualization, manuscript editing, funding, supervision. Ryan R. Davis: technical support, data analysis. Stephenie Y. Liu: technical support, data analysis. Iris Rivas: data acquisition, technical support. Brian Murphy: data acquisition, data analysis manuscript editing. Natalia Vapniarsky: data acquisition, data analysis, manuscript editing.

Supporting information

Supplemental Figure 1.

supplemental table 1.

Data Availability Statement

The data that support the findings of this study are openly available in Gene Expression Omnibus at https://www.ncbi.nlm.nih.gov/geo/, reference number GSE292226.