Bulk transcriptome data revealed upregulated expression of muscle contraction genes in LNM-positive samples

To identify genomic and transcriptomic determinants of LNM in patients with OSCC, we first analyzed whole-exome sequencing (WES) and RNA sequencing (RNA-seq) data from OSCC patients studied in The Cancer Genome Atlas (TCGA) project. For this study, we excluded human papilloma virus (HPV)-positive cases, as described in the Methods and Discussion. This resulted in the analysis of a total of 201 HPV-negative OSCC cases:115 patients with LNM (tumor-node-metastasis [TNM] stage, N1-N3) and 86 patients without LNM (TNM stage, N0; Figs 1A and S1A–S1B and S1 Table). The most commonly detected cancer-driver mutations were tumor protein p53 (TP53) mutations (74%; n = 148), followed by mutations in titin (TTN; 32%, n = 64), FAT atypical cadherin 1 (FAT1; 23%; n = 47), cyclin-dependent kinase inhibitor 2A (CDKN2A; 21%; n = 42), notch receptor 1 (NOTCH1; 17%; n = 34), phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha (PIK3CA; 17%; n = 34), mucin 16 (MUC16; 15%; n = 30), spectrin repeat-containing nuclear envelope protein 1 (SYNE1; 15%; n = 30), and caspase 8 (CASP8; 14%; n = 28; Fig 2A). When we analyzed the correlation between LNM status and the driver mutation profiles, none of the driver mutations showed a significant association with LNM status (Fig 2B and S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic overview of multimodal OSCC data integration.

(A) Schematic overview of this study, which integrated multimodal human oral squamous cell carcinoma (OSCC) data from bulk whole-exome sequencing (WES), RNA sequencing (RNA-seq), microarray, single-cell RNA sequencing (scRNA-seq), spatial RNA sequencing-(spRNA-seq), and immunohistochemical (IHC) analyses. Created in BioRender.

https://doi.org/10.1371/journal.pgen.1011791.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Bulk analysis of metastasis-related factors in OSCC.

(A) Oncoplot displays the 20 most frequently mutated genes in 201 patients with oral squamous cell carcinoma (OSCC) and is based on TCGA-head and neck squamous cell carcinoma (HNSCC) WES data. Rows denote genes arranged by mutation frequency. Gene names are on the left and mutation frequencies are on the right. Columns represent individual patients, and a color bar beneath the oncoplot provides patients’ clinical details, including the tumor-node-metastasis (TNM) stage (lymph node metastasis [LNM] status is indicated as an N stage, and tumor size is indicated as a T stage), primary tissue, and age categorization (under 60 years [y]). To the right, stacked bars show mutation tallies, colored by mutation type, for each gene. A key above these bars identifies the mutation types. Within the main plot, rectangles signify diverse somatic mutations. The bars at the top reflect the tumor mutation burden (TMB) of each patient, with the colors corresponding to the mutation types in the heatmap. (B) Forest plot depicts the associations between the 9 most frequently mutated genes (TP53, TTN, FAT1, CDKN2A, NOTCH1, MUC16, CASP8, SYNE1, and PIK3CA) and LNM status among 201 patients with OSCC. Squares represent odds ratios (ORs), and bars display 95% confidence intervals (CIs). ORs were resolved from a multiple logistic regression analysis. (C) Volcano plot shows 260 upregulated differentially expressed genes (DEGs) in 201 patients (LNM-positive or LNM-negative) with OSCC. The genes were identified using Tag Count Comparison (TCC) with iDEGES/edgeR normalization and the Fisher exact test (false discovery rate [FDR] < 0.01). LNM-positive DEGs are red, LNM-negative DEGs are blue, and the top DEGs are labeled. (D) Heatmap shows the top enriched pathways from the Metascape analysis of the DEGs associated with LNM status. Each row represents a pathway. The heatmap illustrates the top nonredundant enrichment pathways using a discrete color scale to indicate statistical significance. The pathway R-HSA-397014, the muscle contraction pathway, was the most enriched pathway in the LNM-positive patients. Abbreviations: CASP8, caspase 8; CDKN2A, cyclin-dependent kinase inhibitor 2A; FAT1, FAT atypical cadherin 1; FFPE, formalin-fixed, paraffin-embedded; ITF, invasive tumor front; MUC16, mucin 16; NOTCH1, notch receptor 1; OS, overall survival; PIK3CA, phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha; SYNE1, spectrin repeat-containing nuclear envelope protein 1; TCGA, The Cancer Genome Atlas TP53, tumor protein p53; TTN, titin.

https://doi.org/10.1371/journal.pgen.1011791.g002

We then analyzed RNA-seq data to identify the differentially expressed genes (DEGs) associated with LNM status and found 260 significantly upregulated DEGs (Fig 2C). Pathway analysis of these DEGs revealed that the genes involved in muscle contraction, including actin gamma 2 (ACTG2), ATPase sarcoplasmic/endoplasmic reticulum Ca2+ transporting 1 (ATP2A1), myosin heavy chain 6 (MYH6), myosin heavy chain 8 (MYH8), myosin heavy chain 11 (MYH11), myosin light chain 3 (MYL3), myosin light chain 4 (MYL4), and myosin light chain 7 (MYL7), were the top significant genes associated with LNM-positive status (Fig 2D). Since the enrichment of muscle contraction genes could have been driven by differences in the tumor purity of the biopsied samples, we used the Estimation of Stromal and Immune Cells in Malignant Tumor Tissues Using Expression Data (ESTIMATE) tool (The University of Texas MD Anderson Cancer Center) to deconvolute RNA-seq data and infer the tumor purity of the samples [15]. When comparing the primary tumor samples from patients with or without LNM, we found no significant differences in the abundance of tumor stromal cells (P = 0.95; 95% CI, −205.59-249.37) or immune cells (P = 0.87; 95% CI, −247.47-203.15; S1C Fig), suggesting that the observed DEGs in the LNM-positive samples were not caused by differences in sampling.

Single-cell transcriptome data revealed a high abundance of and enhanced intercellular communication between myCAFs in the LNM-positive OSCC microenvironment

To better understand the cellular origin of the observed DEGs in the bulk RNA-seq data, we then analyzed the publicly available single-cell RNA-seq (scRNA-seq) data of patients with OSCC (source: GSE103322 [7]). These data included a total of 5,884 cells from 17 patients with OSCC (12 LNM-positive patients and 5 LNM-negative patients; Figs 1A and S2A–S2D and S3 Table).

We first compared the DEGs of pseudo-bulk cells from the primary tumor site using the samples from the 17 patients with OSCC. Consistent with the bulk RNA-seq results from TCGA (Fig 2D), the DEGs highly expressed in the cells from the LNM samples were involved in pathways related to muscle contraction (S2E Fig). Next, we mapped the scRNA-seq data using the uniform manifold approximation and projection method and subsequently identified clusters of patients with OSCC. For each patient with LNM, the primary tumor and metastatic sites were clustered together, suggesting that the transcriptomic status of tumor cells between the primary tumor and metastatic sites was similar (Figs 3A–3C and S2F–S2H). We then examined the expression of the muscle contraction–related genes for each cell type, and the myCAFs showed the highest expression of muscle contraction genes (Fig 3D). Although we did not find a statistically significant difference in the relative fractions of each cell type between LNM-positive and LNM-negative samples, there was a numerically higher fraction of myCAFs in patients with LNM compared with those without (median fractions: 13% [interquartile range, 5%-24%] vs. 4% [interquartile range, 1%-16%], respectively; P = 0.15; Fig 3E). The lack of statistical significance was likely due to the small sample size. These results were further corroborated with the bulk TCGA RNA-seq data, in which computational deconvolution using CIBERSORTx revealed significant enrichment of myCAFs in patients with LNM progression (P = 0.049; Spearman correlation = 0.13; Figs 3F–3G and S2I). These results suggest that a higher abundance of myCAFs is associated with LNM in OSCC.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Enhanced intercellular interaction of myCAFs within the OSCC microenvironment in LNM.

(A-C) Uniform manifold approximation and projection (UMAP) visualizations of 5,884 cells from 17 oral squamous cell carcinoma (OSCC) samples analyzed using single-cell RNA sequencing. (A) UMAP visualization of primary (pri) tumor sites without lymph node metastasis (LNM) (blue), primary tumor sites with LNM (red), and metastatic (met) sites (green). (B) UMAP visualization of cells using sample identifier numbers. (C) UMAP visualization of cells grouped into 10 key cell types: B cells, dendritic cells (DCs), endothelial cells (ECs), macrophages (Macs), mast cells (Msts), myocytes (Myos), OSCC cells, T cells, myofibroblastic cancer-associated fibroblasts (myCAFs), and secretory/matrix-remodeling cancer-associated fibroblasts (sCAFs). See S2F–S2H Fig for further details. (D) Box plots illustrate the average expression of 211 muscle contraction–related genes across 8 distinct cell types, excluding those representing less than 1% of the total cell population. The genes were sourced from the REACTOME_MUSCLE_CONTRACTION database, v2023.1. The y-axis displays the mean expression levels of these genes. Black circles indicate individual cells. The center lines represent the medians, the box borders represent the interquartile ranges (IQRs), and the whiskers represent ± 1.5 × the interquartile ranges (IQRs). (E) Box plots show the relative fractions of OSCC cells, myCAFs, sCAFs, and T cells within samples, distinguished between cases with LNM (red) and without LNM (blue). Black circles represent individual samples. No statistical significance (P < 0.05) was observed in any comparison, as determined using a 2-sided Mann-Whitney U test. (F) Relative cellular fractions for 10 cell types in each sample from 201 patients with OSCC. Cellular fractions were estimated via a CIBERSORTx analysis. (G) Relative fractions of OSCC cells, myCAFs, sCAFs, and T cells within each sample were compared across different LNM stages. Significance (indicated with an asterisk) was determined using a Spearman correlation and was observed for the myCAFs (P = 0.049).

https://doi.org/10.1371/journal.pgen.1011791.g003

Next, to assess the functional activity of stromal cells—including myCAFs—in OSCCs, we used CellChat [16], which generates in silico inferences for cell-cell interactions using scRNA-seq data. Overall, the LNM-positive samples showed enhanced cell-cell interactions compared with the LNM-negative samples (number of inferred interactions: 4,902 [95% CI: 3,616-5,001] vs. 2,731 [95% CI: 2,004-3,529]). Moreover, the strength of the intercellular interactions was greater in the LNM-positive samples compared with the LNM-negative samples (interaction strength: 120.37 [95% CI: 96.63-146.83] vs. 88.07 [95% CI: 58.48-123.13]; Fig 4A). Among the cell types, the myCAFs exhibited the highest differential outgoing interaction when the LNM-positive and LNM-negative samples were compared (Figs 4B and S2J). Specifically, we observed that the myCAFs expressed higher levels of ECM-related genes, including those in the collagen gene family (collagen type 1 alpha 1 chain [COL1A1], collagen type 1 alpha 2 chain [COL1A2], collagen type IV alpha 1 chain [COL4A1], collagen type IV alpha 2 chain [COL4A2], collagen type VI alpha 1 chain [COL6A1], and collagen type VI alpha 2 chain [COL6A2]), FN1, and the laminin gene family [LAMA5, LAMB1, LAMB2], in LNM-positive samples (S2K Fig). We also found that OSCC cells upregulated ECM-associated receptors, specifically ITGAs, ITGBs, SDCs, and CD44 (Figs 4C and S2L). CD44 is a well-known marker for cancer stem cells [17] and has been associated with poor prognosis and treatment resistance in OSCC [18–20]. We found that the OSCC cells’ cancer stemness scores [21,22] were significantly higher in LNM-positive samples than in LNM-negative samples (P < 0.001 and 95% CI, 0.06-0.1 vs. P < 0.001 and 95% CI, 0.02-0.05, respectively; S2M–S2N Fig). Further analysis using CRISPRGeneEffect data from DepMap [23] found that ITGB1, ITGA3, and CD44 were among the most essential genes for the survival and growth of OSCC cells (S2O Fig). Taken together, these findings suggest that myCAFs were not only abundant in LNM-positive samples but also functionally activated and providing ECM-related signals to OSCC cells. The ECM-related signals may have subsequently activated stemness-related genes such as CD44 in OSCC cells.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Differential intercellular communication in OSCC by LNM status.

(A) Bar plots showing the total number (left panel) and overall strength (right panel) of inferred intercellular interactions in lymph node metastasis (LNM)-positive and LNM-negative OSCC samples. Data are presented with 95% bootstrap CIs derived from 10,000 resamples. (B) Scatter plot comparing interaction strengths for each cell type in LNM-positive vs. LNM-negative samples. Outgoing interactions are shown on the x-axis, and incoming interactions are shown on the y-axis. (C) Heatmaps comparing intercellular communications in LNM-positive vs. LNM-negative samples for each cell type. The bars at the top illustrate outgoing interactions, while the bars on the left display incoming interactions for each cell type. Red indicates LNM-positive interactions, blue indicates LNM-negative interactions, and the color intensity indicates the number and strength of the inferred interactions. The communications received by OSCC cells are depicted below the heatmaps.

https://doi.org/10.1371/journal.pgen.1011791.g004

Spatial transcriptomic analysis revealed colocalization of myCAFs and OSCC cells at the ITF

Next, to investigate how cellular interactions between OSCC cells and myCAFs are spatially organized, we performed a spatial transcriptomic analysis of 4 surgical specimens collected from 2 patients with OSCC treated at Hirosaki University Hospital. The clinical characteristics of these 2 patients (HUH001 and HUH002) are described in the S1 Methods and S4 Table. Patient HUH001 had LNM (stage T2N3bM0), whereas patient HUH002 did not (stage T2N0M0). Two primary tissue samples (HUH001-P1 and HUH001-P2) and 1 metastatic tissue sample (HUH001-met) from patient HUH001 and 1 primary tissue sample from patient HUH002 (HUH002-P) were analyzed using the Visium formalin-fixed, paraffin-embedded (FFPE) Spatial Gene Expression platform (Figs 5A,5B and S3A–S3G and S5 and S6 Tables).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Exploration of OSCC microenvironment heterogeneity with a spatial transcriptomic analysis.

(A) Schematic overview of the spatial transcriptomic analysis. Created in BioRender. (B) Illustration presenting the clinical appearance and various imaging modalities used for patients with oral squamous cell carcinoma (OSCC), including positron emission tomography/computed tomography (PET/CT), CT, and magnetic resonance imaging (MRI). Four FFPE tissue slides from 2 patients with OSCC, patients HUH001 and HUH002, were analyzed using the Visium formalin-fixed, paraffin-embedded (FFPE) Spatial Gene Expression platform. The 4 slides included 2 primary tissues (HUH001-P1 and HUH001-P2) and 1 metastatic tissue (HUH001-met) from patient HUH001, who had LNM. Additionally, 1 primary tissue (HUH002-P) from patient HUH002, who did not have LNM, was included. Specialist histopathological annotations identifying the invasive tumor fronts (ITFs) in HUH001-P1 and HUH001-P2 were provided. The adjacent sections from the same specimen were used for diagnostic hematoxylin and eosin/ immunohistochemical images and Visium images, respectively (See also the FFPE sample preparation and hematoxylin and eosin—stained images for spatial transcriptome analysis subsection in the Methods). Abbreviations: CA_WD_MD, cancer, well-differentiated/moderately differentiated; CA_MD_WD, cancer, moderately differentiated/well-differentiated; CA_Inv_Les, cancer, invasive lesions; CA_Fib_Lym, cancer, fibrosis, lymphocytes; CA_PD_Fib, cancer, poorly differentiated fibrosis; CA_PD_Mus, cancer, poorly differentiated muscle; CA_PD, cancer, poorly differentiated; Inf_Fib, inflammation, fibrosis; Inf_Mus, inflammation, muscle; Adi_Tis, adipose tissue; Bld, blood; Con_Tis_Ves, connective tissue, vessels; Epi, epithelium; Gran_Tis, granulation, tissue; LN_Cap, lymph node capillaries; Lym, lymphocytes; Mus, muscle; Neu, neurons; Neu_Tum, neuron tumor; Uncat, uncategorized; Ves, vessels. The depicted histopathological annotations are broadly categorized into shades of green, representing tumor regions; shades of yellow, indicating peritumor regions; and shades of blue, denoting nontumor regions. A 2-mm scale bar is included for reference. For additional information, please refer to S3 Fig.

https://doi.org/10.1371/journal.pgen.1011791.g005

To investigate the spatial localization of OSCC cells and myCAFs, we deconvoluted cells within each spot using Tangram [24]. This approach maps single cells to lower-resolution spatial transcriptomic profiles using marker genes identified via scRNA-seq data (S4A–S4E Fig). Qualitatively, in the LNM-positive samples (HUH001-P1 and HUH001-P2), the myCAFs spatially colocalized with the OSCC cells at the ITF, which is the pathologically defined edge of a malignant tumor where cancer cells are actively invading into the surrounding normal tissue (Figs 6A,S3A,S4A and S4B). Conversely, in the LNM-negative sample (HUH002-P), the myCAFs spatially segregated from the OSCC cells in the primary tumor site, and only OSCC cells were present within the ITF (Figs 6A and S4D). Quantitatively, the co-localization of the OSCC cells and myCAFs was significantly higher in the HUH001 (LNM-positive) specimens than in the HUH002 (LNM-negative) specimens (P = 0.003; odds ratio, 2.17; 95% CI, 1.40-3.38; Fig 6C–6D). Immunohistochemical (IHC) analysis of the corresponding FFPE samples also confirmed this spatial co-localization of myCAFs and OSCC cells (Fig 6B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Spatial mapping and colocalization analyses in primary tumor tissues with or without LNM, as well as IHC confirmation.

(A) Single-cell spatial localization in 2 lymph node metastasis (LNM)-positive samples (HUH001-P1 and HUH001-P2) and a LNM-negative sample (HUH002-P) were analyzed on the basis of the marker genes identified via spatial RNA sequencing. The Tangram method was used to map the predicted annotations of oral squamous cell carcinoma (OSCC) cells and myofibroblastic cancer-associated fibroblasts (myCAFs) onto Visium Spatial Gene Expression data, and the lower resolution of spatial transcriptomic profiles was addressed at the spot level. The legend shows the estimated cell numbers. See S4A–S4E Fig for more details. (B) The same formalin-fixed, paraffin-embedded (FFPE) samples used for the Visium Spatial Gene Expression analysis underwent immunohistochemical (IHC) staining with cytokeratin/keratin (CK) AE1/AE3 (a marker for OSCC cells), and α-smooth muscle actin (α-SMA, a marker for myCAFs). A 1-mm scale bar is provided for reference. (C, D) Log odds ratios display the spatial relationships among the 8 cell types predicted within the OSCC microenvironment. Spatial colocalization is represented in red, and mutual exclusivity is represented in blue. The odds ratio (OR) was determined using a 2 × 2 contingency table based on the colocalization counts of cell pairs. The Haldane-Anscombe correction was applied for zero counts. Initially, 10 cell types were identified within the OSCC microenvironment, but those cell types appearing in less than 4% of the total spots were excluded to ensure the robustness and reliability of the statistical analysis by mitigating the influence of sparse data. The level of statistical significance was determined by the false discovery rate (FDR). (C) Primary tumor tissues with LNM (HUH001-P1 and HUH001-P2). (D) Primary tumor tissue without LNM (HUH002-P).

https://doi.org/10.1371/journal.pgen.1011791.g006

To validate these findings in a larger cohort, we then conducted an IHC analysis of 90 additional samples of ITFs from 18 patients with OSCC (9 with LNM and 9 without LNM; Fig 7A). All selected patients had stage T2 primary tumors as determined by the tumor sizes and locations (S4 Table). Consistent with the findings from the spatial transcriptomic analysis, significantly higher absolute α-smooth muscle actin (α-SMA)–positive areas per square millimeter (myCAF densities) were detected in the ITFs of samples from patients with LNM compared with those of samples from patients without LNM (median: 19.14% [95% CI, 15.92%-21.99%] vs. 10.70% [95% CI, 9.06%-13.12%], respectively; P < 0.001; Fig 7B). Moreover, histopathological analyses revealed that the myCAF density did not significantly correlate with the pathological grade or the pattern of invasion (POI; Spearman correlation = −0.16, P = 0.14 and Spearman correlation = 0.06, P = 0.60, respectively; S4F Fig). Importantly, in the multivariable logistic regression adjusted for the pathological grade and the POI, myCAF density was identified as an independent predictor of LNM (odds ratio, 1.45; 95% CI, 1.21–1.72; P < 0.001; Fig 7C), underscoring the association between myCAF presence within the ITF and LNM risk.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. IHC analysis of the ITFs in 18 OSCC patients by LNM status.

(A) Representation of the IHC analysis conducted on 90 invasive tumor front (ITF) samples. These samples were obtained from 18 patients with OSCC (patients HUH003 to HUH020), with an equal division between patients with and without lymph node metastasis (LNM; see also the IHC analysis of 90 ITF samples from T2 OSCC patients subsection in the Methods and S4 Table). (B) Boxplots illustrate the proportion of IHC-positive regions per unit area across 90 ITF samples, distinguished between those from patients with LNM (represented in red) and those from patients without LNM (in blue). The metrics are the absolute OSCC density (CK AE1/AE3-positive area per square millimeter), the absolute myCAF density (α-SMA-positive area per square millimeter), and the relative myCAF proportion (α-SMA-positive area relative to the sum of the α-SMA- and CK AE1/AE3-positive areas). Each ITF sample is denoted by a black circle. The center line represents the median, the box borders represent the interquartile range (IQR), and the whiskers indicate ± 1.5 × IQR. Statistical significance was assessed using a 2-sided Mann-Whitney U test: *P < 0.05; ***P < 0.001. (C) Forest plots depict the associations between spatial metrics (OSCC density, myCAF density), pathological differentiation grades, and patterns of invasion (POI) and LNM in OSCC samples. Circles represent odds ratios (ORs), and horizontal bars display 95% CIs. ORs were determined by univariate and multivariate logistic regression analyses. Statistical significance is indicated by a 95% CI that does not cross 1.0.

https://doi.org/10.1371/journal.pgen.1011791.g007

Next, we examined the ligand-receptor pairs (myCAFs and OSCC cells) in the spatial transcriptome data using the scRNA-seq data described above. The LNM-positive samples (HUH001-P1 and HUH001-P2) showed enhanced spatial gene expression of inferred receptors such as ITGAs, ITGBs, CD44, and SDCs in OSCC cells within the ITF (S4G–S4I Fig). A further spatial interaction analysis confirmed that these LNM-positive samples exhibited increased spatially organized communication compared with the LNM-negative sample (HUH002-P); the identified signaling pathways involved the collagen gene family, CD44, and transforming growth factor β (TGFβ; S4J Fig).

To validate the proposed myCAF-OSCC cell interactions via the ECM-CD44 axis, we analyzed 12 additional, publicly available, spatial transcriptomic OSCC samples (source: GSE208253 [25]) and in vitro co-culture datasets (sources: GSE279481 [26] and GSE178153/GSE178154 [27]). The validation spatial cohort consistently showed a spatially organized ECM-CD44 axis. In the LNM-positive samples, this reflected active collagen signaling from myCAFs to OSCCs within the ITFs (S4K–S4N Fig). This spatial organization occurred alongside enhanced myCAF-OSCC colocalization (P = 0.01) and significantly higher spatial CD44 expression in LNM-positive samples compared with LNM-negative samples (P = 0.02; S4O–S4P Fig). Analyses of the GSE279481 co-culture dataset provided further support for these findings. A direct comparison of CD44 expression in OSCC cells cultured with or without myCAFs revealed a numerically higher median expression in OSCC cells co-cultured with myCAFs compared to OSCC cells cultured alone, although this difference was not statistically significant (median: 7.48 [interquartile range, 6.62-8.28] vs. 5.46 [interquartile range, 5.25-6.87], respectively; P = 0.40; S4Q Fig). However, hierarchical Bayesian modeling provided evidence supporting a global positive association between myCAF-derived ECM genes and CD44 expression in OSCC cells (posterior slope mean = 0.27; 95% highest density interval, −0.1-0.63; S4R Fig). Mechanistic modeling of this data further indicated significant indirect effects of myCAF-ECM signaling on CD44 expression in OSCC cells, mediated by ITGAs/ITGBs (indirect effect β = 0.29; 95% CI, 0.17-0.40) and SDCs (indirect effect β = 0.47; 95% CI, 0.34-0.57; S4S Fig). Finally, in an independent microarray co-culture dataset (GSE178153/GSE178154), patient-derived head and neck squamous cell carcinoma (HNSCC) cell lines from LNM-positive tumors showed significantly higher CD44 expression than those from LNM-negative tumors (P = 0.04; S4T Fig). These findings from multiple independent datasets confirmed the intercellular interaction data from the scRNA-seq analysis and provided evidence that these interactions are driven by the spatial colocalization of myCAFs and OSCC cells within the ITF.

Spatial transcriptomic analysis of matched primary and metastasis samples revealed a transcriptional program shared between primary and metastatic sites

By analyzing the spatial transcriptomic data of the matched primary tumor sites (HUH001-P1 and HUH001-P2) and metastatic site (HUH001-met) from the patient with LNM, we sought to identify the origins of the metastatic cells in the primary tumor site. We used single-cell Variational Inference (scVI; scvi-tools v0.14.5) [28], which is based on probabilistic models and deep neural networks, to integrate the spatial transcriptome data and perform unsupervised clustering. This resulted in 12 clusters that were subsequently restored to their spatial coordinates in the primary tumor (samples HUH001-P1 and HUH001-P2) and metastatic (HUH001-met) sites (Figs 8A, S5A and S5B). Additionally, copy number variations (CNVs) data [29] were used to confirm the origin of metastatic cells in the primary tumor sites at the genomic level (S5C Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Integrated spatial transcriptomic analysis of primary tumor and metastatic sites and positional reconstruction.

(A) Uniform manifold approximation and projection (UMAP) visualization shows integrated spatial transcriptome data from 3,637 spots from 2 primary tumor sites (HUH001-P1 and HUH001-P2) and a metastatic site (HUH001-met) from patient HUH001, who had lymph node metastasis (LNM). These samples come from the same patient, patient HUH001, and represent matched primary and metastatic sites. The visualization identifies unsupervised clusters, and 12 clusters are mapped back to their original spatial coordinates. In this visualization, each data point corresponds to a spatial spot. (B) Stacked bar graph shows the samples’ integrated distribution in spatial transcriptome clusters, with HUH001-P1 in blue, HUH001-P2 in orange, and HUH001-met in green. The color coding at the top of the bars displays the clusters in the primary tumor sites in blue (clusters 3 and 12) and the clusters in the metastatic primary tumor sites in red (clusters 1, 2, and 4-11). (C) Spatial visualization showcases the primary tumor lusters (clusters 3 and 12) in blue and the metastatic clusters (clusters 1, 2, 4-11) in red. These clusters are mapped onto the spatial coordinates of samples from HUH001-P1 and HUH001-P2. For information regarding pathologic annotation, see the note in Fig 5B. (D) Volcano plot shows 298 upregulated differentially expressed genes (DEGs) in the metastatic cluster, shown in red. The DEGs in the primary tumor cluster are shown in blue. These DEGs were identified using Scanpy with a 2-sided Mann-Whitney U test (false discovery rate [FDR] < 0.05 and log₂ fold change > 0.15). The top 20 DEGs are labeled. (E) Heatmap displays the most enriched pathways from the Metascape analysis of 298 upregulated DEGs in the metastatic cluster. M5930: Hallmark epithelial mesenchymal transition was the top enriched pathway. (F) Box plots illustrate the average expression of the top 9 DEGs (COL1A2, COL3A1, DCN, TPM2, IGKC, PCOLCE, PRRX1, ACTA2, and RNF213) in 8 distinct cell types in the metastatic cluster, as determined using single-cell RNA sequencing data with the exclusion of cell types representing less than 1% of the total cell population. The black circles indicate individual cells, the center lines represent the median, the box borders represent the interquartile ranges (IQRs), and the whiskers represent ± 1.5 × IQR. (G) Forest plot shows the association between the spatial coordinates of clusters and colocalization with oral squamous cell carcinoma (OSCC) and myofibroblastic cancer-associated fibroblasts (myCAFs) in primary tumor tissues with LNM (HUH001-P1 and HUH001-P2). Circles represent odds ratios (ORs), and bars display the 95% CI.

https://doi.org/10.1371/journal.pgen.1011791.g008

We identified 2 main clusters: the clusters predominant in the primary tumor sites (the primary tumor cluster, consisting of clusters 3 and 12) and the clusters shared between the primary tumor sites and the metastatic site (the metastatic cluster, consisting of clusters 1, 2, and 4–11; Figs 8B,8C, S5D and S5E). To explore tumor clusters associated with LNM, we conducted a pseudotime trajectory analysis using stLearn (S5F–S5G Fig) [30]. The metastatic clusters 1, 5, 7, and 9 exhibited shorter pseudotimes than did primary tumor cluster 12 and had elevated CD44 expression indicating enhanced stemness (all adjusted P < 0.001; S10 Table and S5H Fig). DEG analysis revealed unique molecular signatures for these spatial clusters. Cluster 1 showed enrichment of the myCAF‑like profile, including desmin (DES), troponin I1 (TNNI1), and collagen type XI alpha 1 chain (COL11A1). Cluster 5 was characterized by elevated serpin family B member 3 (SERPINB3), S100 calcium binding protein A9 (S100A9), and mesothelin (MSLN) expression, indicating highly invasive tumor cells. Cluster 7 demonstrated high expression of markers for class‑switched immunoglobulins, notably immunoglobulin heavy constant gamma 4 (IGHG4), immunoglobulin heavy constant alpha 1 (IGHA1), and immunoglobulin lambda variable 3–1 (IGLV3‑1). Cluster 9 was distinguished by markers typically associated with CXCL14-positive basal‑like epithelial cells, including C-X-C motif chemokine ligand 14 (CXCL14), serine peptidase inhibitor kazal type 6 (SPINK6), and keratin 1 (KRT1). Interestingly, cluster 1, which had the highest proportion of cells from primary tumor sites among the metastatic clusters, showed enrichment of myCAFs and activated COL11A1, suggesting molecular myCAF/OSCC crosstalk though the ECM-CD44 axis. These data depict a continuum of spatial states in the ITF and reflect the migratory, inflammatory, and stem-like characteristics potentially underpinning LNM in OSCC (refer to S11 Table for a full list of cluster characteristics).

To understand the gene-expression profiles unique to metastatic cells, we compared the gene-expression profiles in the 2 main clusters (the primary tumor cluster and the metastatic cluster) and identified 298 significantly upregulated DEGs, including members of the collagen gene family (COL1A2 and collagen type III alpha 1 chain [COL3A1]), decorin (DCN), tropomyosin 2 (TPM2), immunoglobulin kappa constant (IGKC), procollagen C-endopeptidase enhancer (PCOLCE), paired related homeobox 1 (PRRX1), actin alpha 2 (ACTA2), and ring finger protein 213 (RNF213; Figs 8D and S5I). The pathway analysis of these DEGs showed significant enrichment of the genes involved in hallmark epithelial-to-mesenchymal transition and hallmark myogenesis (Figs 8E and S5J). Consistent with this finding, these top DEGs showed the highest expression in myCAFs per scRNA-seq data (Fig 8F) and within ITFs of LNM-positive samples from the validation spatial cohort (P < 0.001; S5K Fig). Furthermore, we found a significant correlation between the colocalized regions with myCAFs and OSCC cells and those of the ITFs in the metastatic cluster (P < 0.001; odds ratio, 2.24; 95% CI, 1.78-2.81; Fig 8G).

Spatially resolved metastatic cluster genes predicted the clinical outcomes of patients with OSCC

We hypothesized that the 298 DEGs that were found in the spatial analysis and were unique to the metastatic cluster could function as predictive biomarkers for LNM progression and the overall prognosis of patients with OSCC. To test this hypothesis, we first developed a LASSO-penalized Cox proportional hazards model using clinical data and RNA-seq data from TCGA [31,32]. Using this model, we developed a gene signature–based risk score. This approach helped us to reduce our list of DEGs of interest to 23 core signature genes (Fig 9A and S12 and S13 Tables). This 23-gene signature stratified OS and disease-specific survival (DSS) well in the cohort from TCGA (P < 0.001, Fig 9B; P < 0.001, Figs 9C and S6A–S6C and S14 and S15 Tables). Additionally, higher scores showed a significant positive correlation with LNM progression in individual samples from 214 patients with OSCC (P = 0.03, 95% CI, 0.04-0.61, Fig 9D; P = 0.03, Spearman correlation = 0.24, Fig 9E). Furthermore, the scores were significantly higher in LNM-positive samples in the analysis of OSCC cells from 17 OSCC samples (P < 0.001, 95% CI, 0.76-1.45, Fig 9F) and in samples from the validation spatial cohort of 12 OSCC samples (P = 0.03, 95% CI, 0.06-1.35, Fig 9G).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Validation of spatially resolved 23-gene signature with external transcriptome data.

(A) Heatmap shows the Z-score expression of spatially resolved 23 core genes across 214 patients with oral squamous cell carcinoma (OSCC). Each row represents a gene, and each column represents a patient. The color bar above the heatmap shows clinical information, including lymph node metastasis (LNM) represented as an N stage, tumor size represented as a T stage, and risk scores from the 23-gene signature. On the left, another color bar shows whether each gene coefficient was unfavorable (red) or favorable (blue). (B, C) Kaplan-Meier curves show overall survival (OS) and disease-specific survival (DSS) based on the risk scores from the 23-gene signature and across different platforms. The high-score and low-score groups were determined on the basis of the median risk score. Statistical significance was assessed using log-rank tests. For additional information, please refer to S6A–S6C Fig. (B) OS for 214 patients from The Cancer Genome Atlas (TCGA). (C) DSS for 209 patients from TCGA. (D-G) Box plots depict risk scores from the 23-gene signature by LNM status. Black circles represent (D, E) 214 individual samples from patients with OSCC and (F) 5,467 OSCC cells across 17 OSCC single-cell RNA sequencing samples. (G) Samples from the validation spatial cohort of 12 OSCC samples. The center line represents the median, the box borders represent the interquartile ranges (IQRs), and the whiskers represent ± 1.5 × IQR. Statistical significance for the group differences was assessed using a 2-sided Mann-Whitney U test, and the strength of association was evaluated using the Spearman correlation: *P < 0.05; ***P < 0.001. (H-M) Kaplan-Meier curves show OS and DSS based on the risk scores from the 23-gene signature and across different platforms. The high-score and low-score groups were determined on the basis of the median risk score. Statistical significance was assessed using log-rank tests. For additional information, please refer to S6E–S6N Fig. (H) OS for 96 patients from GSE41613, (I) DSS for 76 patients from GSE41613, (J) OS for 50 patients from GSE42743, (K) DSS for 37 patients from GSE42743, (L) OS for 146 patients from GSE41613 and GSE42743 combined, and (M) DSS for 113 patients from GSE41613 and GSE42743 combined.

https://doi.org/10.1371/journal.pgen.1011791.g009

We then validated the prognostic significance of the 23-gene signature in 2 independent datasets (data sourced from the studies GSE41613 and GSE42743 [33]). In the GSE41613 data, the 23-gene signature score was significantly prognostic for both OS (P = 0.002; Fig 9H) and DSS (P < 0.001; Figs 9I and S6D–S6F and S15 Table), whereas for the GSE42743 data, the score showed a trend of predicting worse OS (P = 0.326; Fig 9J) and was prognostic for DSS (P = 0.021; Figs 9K and S6G–S6H and S15 Table). To address the limitations arising from facility differences and the constrained sample size, we merged the 2 microarray datasets. After harmonizing for batch effects using pyComBat [34,35] and standardizing the combined expression data with a Z-score, we consistently found significant associations between the 23-gene signature and both OS (P = 0.002; Fig 9L) and DSS (P < 0.001; Figs 9M and S6I–S6L and S15 Table). These results suggest that the 23-gene signature that was extracted from the metastatic cluster in the spatial transcriptomic analysis can be used clinically to predict LNM status and poor OS in patients with OSCC.

Samples, ethics, and consent for publication

This study was conducted in accordance with the principles of the Declaration of Helsinki. Clinical specimens from patients with untreated primary OSCC were used for the spatial gene expression analysis. Approval was obtained from the Human Research Ethics Committee of the Hirosaki University Graduate School of Medicine (approval # 2021-081, jRCT1020210053). Written informed consent was obtained from all participants, and consent included the use of all de-identified patient data for publication. Participants were not compensated. Four FFPE samples from 2 patients with OSCC (patients HUH001 and HUH002) were analyzed using Visium. The tissues from patient HUH001, who had LNM, consisted of 2 samples from primary sites (samples HUH001-P1 and HUH001-P2) and 1 sample from a metastatic site (HUH001-met). The tissue from patient HUH002, who did not have LNM, consisted of 1 sample from the primary site (HUH002-P). Case details are provided in the method details section. In the IHC analysis, we included 18 additional patients with OSCC and stage T2 primary tumors (patients HUH003 to HUH020), 9 of whom had LNM and 9 of whom did not (S4 Table). Permission to conduct the IHC analysis was granted by the Human Research Ethics Committee of the Hirosaki University Graduate School of Medicine (approval # 2022-025), which waived the need to obtain informed consent, given the noninterventional, retrospective design of the IHC analysis.

Bulk WES and RNA-seq data collection and processing

Data from 502 head and neck squamous cell carcinoma cases were obtained from TCGA’s database (TCGA-HNSC) [52]. We excluded 268 cases; in 242 cases, the primary tumor site was not the oral mucosa, and in 26 cases, the presence or absence of LNM was pathologically unclear (S1A Fig). We included 234 OSCC cases for which the primary tumor site was the oral mucosa and a pathological diagnosis of LNM or no LNM had been obtained. Open-access data for clinical characteristics, single-nucleotide variations, and transcriptome profiles were obtained using the gdc-client v1.6.0 from the Genomic Data Commons (GDC) Data Portal (RRID:SCR_014514). We reviewed 234 cases of oral cancer from TCGA (RRID:SCR_003193) according to the bulk OSCC data analysis inclusion criteria, and 201 HPV-negative patients were selected, which included LNM (n = 115, TNM stage: N1–N3) and non-LNM (n = 86, TNM stage: N0) cases (patient selection is described in the S1 Methods and S1A Fig).

RNA-seq differential expression analysis

After filtering out low-expression genes according to the standard TCC v1.26 (an acronym for Tag Count Comparison; RRID:SCR_001779) procedure [53], iDEGES/edgeR normalization was performed using the iteration option with the number of iterations in the normalization pipeline set to 3. DEGs were detected internally using the edgeR v3.28.1 exact test [54–56] with a cutoff false discovery rate [FDR] of less than 0.01. An enrichment analysis of the genes with variable expression was performed with Metascape (RRID:SCR_016620, https://metascape.org/) [57] using the default parameters (including an enriched term with at least 3 candidates, a P-value ≤ 0.01, and an enrichment factor ≥ 1.5) and the analysis species set to Homo sapiens. The Gene Ontology biological process, Kyoto Encyclopedia of Genes and Genomes, and Reactome options were selected.

RNA-seq deconvolution analysis

For each sample, calculations were performed using the ESTIMATE website (https://bioinformatics.mdanderson.org/public-software/estimate/) [15], which scores the purity of tumor tissue, presence of stromal cells, and level of immune cell infiltration based on RNA-seq data. Deconvolution of bulk RNA-seq was performed using CIBERSORTx (RRID:SCR_016955, https://cibersortx.stanford.edu/) [58,59]. The OSCC single-cell RNA-seq dataset (GSE103322) was obtained from the GEO database. A signature matrix was created using the OSCC single-cell RNA-seq dataset, and the parameters were set as replicates, 5; sampling, 0.5; fraction, 0.75; kappa, 999; q-value, 0.01; number of barcode genes, 300; and max, 500). Cell fractions were created by normalizing the OSCC RNA-seq raw data with transcripts per kilobase million, using the signature matrix and 1,000 permutations for statistical analysis. We used the relative cell fractions (the default parameter) for each cell type in the mixed tissues.

Single-cell transcriptome data processing

The OSCC scRNA-seq dataset (GSE103322) [7], which contains 5,884 cells from 17 of the 18 cases (no transcriptome data were obtained from 1 patient), was used (S2A Fig). Five patients had no LNM (TNM stage: N0) and 12 had LNM (TNM stage: N1-N3), of whom 5 had both primary and metastatic sites. We organized the sample IDs and cell-type annotations on the basis of the source data annotation [7] using Scanpy v1.7.1 (RRID:SCR_018139) [60] (details are provided in S1 Methods and S3 Table and S2F–S2H Fig).

Cellular scRNA-seq interaction network analysis

Using CellChat (RRID:SCR_021946) [16], which quantitatively measures networks using CellChatDB [61] data on ligand-receptor complexes, we estimated cell-cell interactions in the TMEs of OSCC samples from the metastatic (n = 12) and nonmetastatic (n = 5) groups. A signaling pathway refers to a set of ligand-receptor pairs belonging to a specific signaling pathway. Following standard CellChat procedures, we first calculated the probability of communication between each cell type using the computeCommunProb function’s default parameters to estimate the communication network. Subsequently, we set the filterCommunication function at min.cells = 10 (the default parameter) and filtered cell-cell communications for cell types with fewer than 10 cells. Finally, we used the computeCommunProb pathway function to estimate intercellular communication at the signaling pathway level.

Gene dependency analysis

For the gene dependency analysis, we used CRISPRGeneEffect data from the DepMap Public 23Q2 primary dataset, which includes dependencies for 17,931 genes across 1,864 cell lines (available at DepMap Portal, https://depmap.org/portal/download/all/). For this analysis, OSCC cells (n = 49) and other cell lines, excluding head and neck squamous cell carcinoma cells (n = 1,024), were considered. Adjusted effect sizes for gene dependency were determined using the cdsrmodels::lin_associations function. Methodological details can be found on the DepMap Community Forum (https://forum.depmap.org/). Statistical significance was assessed using a 2-sided Mann-Whitney U test, and the Benjamini-Hochberg method was applied for P-value correction.

FFPE sample preparation and hematoxylin and eosin—stained images for spatial transcriptome analysis

For each surgical specimen, samples were fixed in 10% buffered formalin for 12 h to 24 h. To accommodate the specific requirements of the Visium platform (6.5 mm × 6.5 mm or smaller), 2 adjacent FFPE tissue blocks were prepared from the same specimen: 1 designated for routine pathological diagnosis (including hematoxycin and eosin [H&E] and immunohistochemical staining) and 1 specifically dedicated for Visium spatial transcriptome analysis. The FFPE blocks were stored at 4 °C away from direct sunlight. We first evaluated the messenger RNA quality in the samples from the Visium-dedicated block using an RNeasy FFPE Kit (Qiagen, Hilden, Germany). A good predictor of success in establishing libraries using Visium FFPE is the percentage of 200-base RNA fragments (DV200) greater than 50%. All FFPE samples had high RNA quality with DV200s of greater than 80% (S5 Table). Next, a tissue adhesion test was performed on a test section from the Visium-dedicated block to ensure that the dissected tissue was affixed to the imaging slide. Then, Visium-dedicated FFPE blocks were immersed in ice water for 30 min for rehydration. The 5-μm-thick sections were placed on slides (Visium gene expression slides, 10x Genomics, Pleasanton, CA, USA), dried in a thermal cycler (Bio-Rad C-1000 Thermal Cycler, RRID:SCR_019688, Hercules, CA, USA) at 42 °C for 3 h, and placed in a desiccator to dry overnight. The dried slides were further heated in a thermal cycler at 60 °C for 2 h and deparaffinized using xylene (214736, Millipore Sigma, St. Louis, MO, USA) and ethanol (E7023, Millipore Sigma). Subsequently, the slides were washed with Milli-Q water and stained with hematoxylin (3 min; MSH16, Millipore Sigma), Dako Bluing Buffer (1 min; CS70230-2, Agilent Technologies, Santa Clara, CA, USA), and eosin (1 min; HT110116, Millipore Sigma). The slides were cleaned with Milli-Q water in between steps. Following the completion of staining and the addition of coverslips to the slides, images of the H&E–stained sections were acquired using the apochromatic option on an all-in-one BZ-X800 fluorescence microscope (Keyence, Osaka, Japan) at 10 × -20 × magnification (BZ-PA10-20). To minimize batch effects, sections were prepared on the same day by the same individual and using the same reagent lot.

Spatial RNA-seq using FFPE Visium

Quality checks of the FFPE Visium libraries were performed using the TapeStation High Sensitivity D5000 system (Agilent Technologies). Library concentrations were determined using a KAPA Library Quantification Kit (KAPA Biosystems, Wilmington, MA, USA). A pool of indexed libraries was sequenced on the NextSeq 550 platform using a High Output Kit v2.5 (Illumina, San Diego, CA, USA), which has a read length of 75 × 2 base pairs with paired-ends and covers at least 25,000 reads per spot, according to the manufacturer’s information booklet.

Spatial RNA-seq data processing and pathological annotation

Reads were demultiplexed and mapped to the reference genome using GRCh38 with Space Ranger software v1.3.0 (10x Genomics). Using the Space Ranger software for each sample, we aligned the H&E–stained images, mapped the entire spatial transcriptome of the FFPE samples, and obtained unsupervised clusters based on the spatial transcriptome (the default setting). Additionally, for all the samples, the whole transcriptomes per spot exceeded the expected metrics values for the sequencing, mapping, and spots, confirming that the quality was sufficient to proceed with the spatial transcriptome analysis (S6 Table). Count matrices and image data were loaded using the dedicated Spatial Transcriptomics Analysis Software Squidpy v1.1.2 [62] or stLearn v0.3.1. Three pathologists (T.Y., S.G., and H.K.) independently annotated spots using H&E–stained image data on a Loupe browser v6.00 (RRID:SCR_018555, 10x Genomics). Based on the pathological annotations on the H&E–stained images, we classified the spots into 3 major categories: tumor region, peritumor region, and nontumor region. Scanpy was used for quality control for each sample region. First, cells with low gene expression and genes expressed in only a few cells were excluded. We then used the p.filter_cells function to filter out cells with unique molecular identifier counts lower than 500 using the setting min_counts = 500 in the tumor region. For the peritumor and nontumor regions, the p.filter_cells function was set to min_counts = 100. Next, we set min_cells = 2 in the pp.filter_genes function to exclude rare genes detected in 2 or fewer cells. Then, using the p.normalize_total function, each cell was normalized to the total count of all genes, and we equalized the total count of each cell following normalization. Owing to their potential to strongly affect normalization, genes with total unique molecular identifier counts greater than max_fraction = 0.05 in at least 1 cell were considered highly expressed and were excluded from the normalization calculation (excluding_highly_expressed = true). We then used the pp.log1p function to convert the data matrix into natural logarithms.

Spatial ligand–receptor interaction analysis

Spatial interactions in the OSCC samples were estimated using Squidpy, which stores the distances between observations in a spatial graph and quantifies spatial interactions using a ligand-receptor interaction database [61,63] on the basis of spatial neighborhoods and gene expression levels. First, we performed spatial morphological expression normalization using stLearn v0.3.1 [30], which normalizes gene expression using spatial neighborhood information and morphological distance. Following standard procedures, image data were extracted using the pp.tiling function of stLearn for preprocessing and the cnn_base = exception in the pp.extract_featur function. Features were extracted from the image data using Xception [64], a convolutional neural network. A principal component analysis of the spatial gene expression data was performed using the setting svd_solver = arpack in the em.run_pca function of stLearn, with the logarithm-converted data normalized for quality control for each sample region in Scanpy v1.7.1 [60]. Next, the gr.ligrec function, implemented in Squidpy’s ligand-receptor analysis algorithm [61], was used with the setting interactions_params = {resources: CellPhoneDB}, and the ligand-receptor complex CellChatDB [61] was used to identify significant interactions. We set the number of test permutations to 1,000 (n_perms = 1,000) and the cutoff threshold for the percentage of cells needed per cluster to 0.01. We used the Benjamini–Hochberg method with α = 0.05 for the significance level of the FDR correction and the setting corr_method = fdr_bh. The ligand-receptor permutation test results were visualized using the pl.ligrec function (FDR < 0.001 and means_range > 0.95) in Squidpy.

Spatial RNA-seq mapping analysis

We investigated the spatial localization of cells in the ITFs at the single-cell level. We deconvoluted the cells within each spot according to the standard Tangram procedures for mapping single-cell-level annotations to spatial data [24,65]. We deconvoluted the cells within each spot. Using the OSCC scRNA-seq dataset (GSE103322), we detected signature markers in 10 principal cells using the tl.rank_genes_groups function with the Wilcoxon method in Scanpy v1.7.1 [60]. We identified a set of 338 genes common to both the single-cell and the spatial datasets and used the map_cells_to_space function in Tangram (mode = constrained, learning_rate = 0.1) and graphics processing units (GPUs) to map to the single-cell space.

IHC staining of the corresponding FFPE samples

IHC of the deparaffinized slides was performed using the standard avidin-oxidase complex method with an automated immunostainer (BenchMark XT; Ventana Medical Systems, Tucson, AZ, USA). In brief, deparaffinized slides were treated with Tris- EDTA (TE) buffer (pH 7.8) at 95 °C for 4 min. Additionally, slides were treated with 5% nonfat dry milk at 37 °C for 15 min to block endogenous peroxides and proteins. The slides were then incubated with primary antibodies for 60 min at room temperature (20 °C-25 °C). Antibodies against the following proteins were used in the indicated dilution ratios: cytokeratin/keratin (CK) AE1/AE3, an OSCC cell marker (Nichirei 412811; 1:1; clone names, AE1 and AE3) and α-SMA, a myCAF marker (Abcam Cat# M0851; RRID:AB_2313736; Dako; 1:100; clone name, 1A4). Images of CK AE1/AE3 and α-SMA were acquired using an all-in-one BZ-X800 fluorescence microscope (Keyence).

IHC analysis of 90 ITF samples from T2 OSCC patients

To measure the rates of CK AE1/AE3 positivity and α-SMA positivity in ITF samples, we performed an IHC analysis of 90 ITF samples from 18 patients with T2 primary tumors: 9 patients (45 samples) had LNM, and 9 patients (45 samples) did not (S4 Table). FFPE tissue sections representative of tumor cross-sections underwent standard immunostaining procedures. Images of the tumor sections (stained H&E, CK AE1/AE3, and α-SMA) were acquired using an all-in-one BZ-X800 fluorescence microscope (Keyence). We randomly selected 5 ITF areas (each 1.0 × 1.0 mm²) from a wide tumor area. The same regions were extracted from the specimens stained with α-SMA. We binarized the images using ImageJ software (RRID: SCR_003070, Java 1.6.0_24 [64 bit]; National Institutes of Health, Bethesda, MD, USA). The binarized images showed either positive or negative immunostaining. We calculated the percentage of CK AE1/AE3–positive cells (OSCC cells) and α-SMA–positive cells (myCAFs) per unit area (1.0 × 1.0 mm²).

Spatial transcriptome integration analysis

To integrate the spatial transcriptome data, we extracted the tumor regions from the primary tumor sites (samples HUH001-P1 and HUH001-P2) and the cervical LNM site (HUH001-met) from the samples from the patient with cervical lymph node metastasis (HUH001). Next, using scVI [28] (scvi-tools v0.14.5 [66]), we corrected the sample-specific batch effects on the basis of probabilistic models and deep neural networks. As the scvi-tools model requires raw counts of the spatial transcriptome data rather than normalized expression data, we again used Scanpy v1.7.1 [60] to extract data for the respective tumor regions from samples HUH001-P, HUH001-P2, and HUH001-met. We filtered the cellular and genetic data using the setting min_counts = 500 in the pp.filter_cells function to remove cells with unique molecular identifier counts lower than 500, and we used the setting min_cells = 3 in the pp.filter_genes function to remove genes detected in fewer than 3 cells. The data from samples HUH001-P, HUH001-P2, and HUH001-met were combined using the concatenate function in anndata v0.7.8. Using anndata, we identified 17,943 genes in the combined data, and we selected the top 4,000 highly expressed differential genes using the p.highly_variable_genes function (n_top_genes = 4,000; flavor = seurat_v3) in Scanpy. We used the data setup from the anndata function in scvi-tools to correct the batch effect for each sample (batch_key = sample). A model integrating spatial transcriptome data was constructed using the model.SCVI function in scvi-tools. The parameters of the model were as follows: number of nodes per hidden layer, 128 (default setting); dimensionality of the latent space, 30 (n_latent = 30); number of hidden layers used for encoders and decoders, 2 (n_layers = 2); dropout rate for the neural network, 0.2 (dropout_rate = 0.2); and negative binomial distribution (dispersion = gene, gene_likelihood = nb). The trained model can be found on figshare (https://doi.org/10.6084/m9.figshare.20279025.v1). With the model.get_latent_representation and model.get_normalized_expression functions of scvi-tools, the model output was returned to anndata, allowing interoperation with Scanpy. Using the standard procedure for clustering in Scanpy, the neighborhood graph was computed using data integrated with scvi-tools in the pp.neighbors function [67]. Unsupervised clustering was performed at a resolution of 0.5 in the tl.leiden function [68].The tl.umap function [69] was executed with default parameters to embed the neighborhood graph.

CNV estimation for spatial RNA-seq

The CNV was estimated using inferCNV v1.11.1 (RRID:SCR_021140, the inferCNV of the Trinity CTAT project; https://github.com/broadinstitute/inferCNV) [29]. We created gene and chromosome position files from the bed file probe set (GRCh38) used for mapping. Furthermore, we created sample annotation files from the tumor and nontumor regions on the basis of the pathological annotations. Additionally, inferCNV calculated the dynamic threshold for noise removal on the basis of the standard deviation of the average residual expression values. The cutoff options for inferCNV were set to 0.1, and genes with an average expression count of less than 0.1 were filtered out. The denoising filter options (default settings) were used to reduce noise (residual expression values of normal cells) while retaining tumor cell expression values.

LASSO-penalized Cox proportional hazards model

Following a previously described approach [70,71], we employed a LASSO-penalized Cox proportional hazards model for 220 patients with OSCC using the CoxPHFitter function (patient selection is described in the S1 Methods). Parameters for the lifelines v0.27.8 package [72] were set at l1_ratio = 1.0, penalizer = 0.02, and baseline_estimation_method = ‘breslow’(default). We first obtained a curated TCGA-HNSCC survival dataset [31] (n = 604), along with other phenotype data (n = 612) and HTSeq Counts data in a log₂(x + 1)-transformed format (n = 546), from the University of California, Santa Cruz’s Xena platform (RRID:SCR_018938, https://xenabrowser.net/datapages/) [32]. We selected the 220 OSCC patients from the GDC sample sheet of bulk RNA-seq analysis. Subsequently, we processed the HTSeq Counts data by first reverting the log₂(x + 1) transformation and then using transcripts-per-million normalization, re-applying the log₂(x + 1) transformation, and finally standardizing the Z-scores for each gene to adjust the data for downstream analyses. For the model training, we used survival outcomes—specifically, OS and its duration—and 298 upregulated DEGs identified in clusters of metastatic primary tumor sites from the spatial transcriptome analysis of matched LNM samples. Through 1,000 iterations of 10% leave-out cross-validations, we evaluated the gene frequencies in each iteration and included those genes chosen in over 90% of the iterations. We averaged the coefficients of these genes with an absolute value exceeding 10⁻² cross the 1,000 iterations. From the 298 DEGs, our model selected 23 genes. We resolved risk scores using their averaged coefficients as weighted expression levels.

Risk score validation

To assess the association between risk scores and prognosis, we conducted survival analyses for OS and DSS using the Kaplan-Meier method and the log-rank test. Initially, the risk score was evaluated in the training cohort from TCGA. Subsequently, to validate its external reliability, we examined the independent OSCC microarray datasets GSE41613 and GSE42743 using survival data from the GEO database. For these datasets, the Affymetrix human genome U133 Plus 2.0 Array probes were used, and 21 genes were detected out of 23. We excluded patients across all datasets on the basis of the following criteria from previous publications [73–75]: perioperative death (within 30 days after surgery), current treatment with neoadjuvant chemoradiotherapy, positive margins, and recurrent lesions. We identified 221 cases of oral cancer in TCGA (RRID:SCR_003193) using the GDC sample sheet of bulk RNA-seq analysis, and 220 patients were selected (patient selection is described in the S1 Methods and S6A Fig). We then preprocessed and removed missing values for survival time and events. We performed a logarithmic transformation and Z-score standardization of TCGA’s RNAseq data after transcripts-per-million normalization. Conversely, after we had verified the gene expression distribution, we standardized the publicly available, normalized microarray datasets using Z-scores, enabling cross-platform comparisons. Apart from this, we combined the 2 normalized microarray datasets after harmonizing them for batch effects using pycombat [34,35] and after standardizing them using Z-scores. We then calculated a risk score for each patient using a resolved linear equation and the median risk score as the cutoff of each cohort on the basis of the balanced data and robustness of risk score distribution as was done in previous studies [70,71]. Two distinct Kaplan-Meier survival curves were created. The first was based on the entire observation period. The second used bootstrapping to assess internal validity, representing the average survival curve with a 95% bootstrap CI. Specifically, survival functions were estimated for each of the 10,000 bootstrap samples. To evaluate the uncertainty of the original dataset, each bootstrap sample was divided into high-score and low-score groups based on the median risk score, followed by log-rank tests. We displayed P-values from these bootstrap samples as histograms and contrasted them with the P-values from the original dataset. As each bootstrap sample’s survival function might possess different time points, all time points were acquired, and survival functions were interpolated to align them to common time points. Due to the increased uncertainty regarding survival arising from our examination of the entire observation period, we limited the bootstrap average survival curve to 1 year after the last event to enhance the curve reliability and make the data easier to interpret.

Statistics and reproducibility

Statistical analyses were performed using R v3.6.2. No statistical method was used to predetermine the sample size for the bulk genome, transcriptome, single-cell transcriptome, and spatial transcriptome analyses, as these were exploratory studies in which the effect sizes of factors related to LNM were not predetermined. The sample size for the IHC analysis was estimated using G*Power v3.1.9.7 (RRID:SCR_013726) [76] (2-sided Mann-Whitney U test; significance level [α] = 0.05; target power [1 − β] = 0.8; allocation ratio N2/N1 = 1.0; effect size [d] = 0.8). Univariate logistic regression analyses were initially performed. Subsequently, multivariable logistic regression was used to determine the independent predictive value of myCAF density for LNM, after adjusting for pathological grade and the POI, in a clinical cohort (refer to S4 Table for patient characteristics and the specific cohort used for this analysis). A multiple logistic regression model was used to analyze the association between each gene mutation and LNM in 201 patients with OSCC. The number of explanatory variables in the model was set to 9 [77], representing the 9 most frequently mutated genes (TP53, TTN, FAT1, CDKN2A, NOTCH1, PIK3CA, MUC16, SYNE1 and CASP8). We performed a post-hoc analysis for the model with a sample size of 201 patients using G*Power (F test; effect size [f²] = 0.15; significance level [α] = 0.05; total sample size = 201; number of predictors = 9), and the power was 0.98. Continuous variables were evaluated for statistical significance using a 2-sided Mann-Whitney U test. For categorical variables, statistical significance was evaluated using a 2-sided Fisher exact test, and odds ratios were calculated to determine the strength and direction of the association. Furthermore, a Spearman correlation was calculated to assess the association between LNM stage progression and cell proportions. In the Spearman correlation analysis, given that the N3 category was comprised of only 2 patients, we added the 20 patients of N2c, a category signifying more severe, bilateral neck LNM, into the N3 category. This adjustment resulted in a total of 22 patients classified as N3. Two-sided P-values less than 0.05 were considered significant. The complete P-value data can be found in the publicly available code provided. We calculated 95% bootstrap confidence intervals using 10,000 bootstrap resamples and the 2.5th and 97.5th percentiles. Multiple testing corrections were performed using the Benjamini–Hochberg/FDR method.

Plots and graphs

We primarily used R v3.6.2 or Python v3.7.12 to generate Figs 1A and 5A were crafted with BioRender.com. The oncoplot (Fig 2A) was constructed using the maftools [78] in R. The volcano plots (Figs 2C and 8D) were constructed using the ggplot2 (RRID: SCR_014601) v3.3.5 in R. The forest plots presented in (e.g., Figs 2B and 8G) and (e.g., Fig 7C) were created using Forester v0.3.0 and forplo v0.2.0, respectively. The heat maps (Figs 2D and 8E) were developed using Metascape. Figs 3A–3C and 8A were created using Scanpy v1.8.2 [60] in Python, and the uniform manifold approximation and projection clustering was executed using the default parameters in leidenalg v0.8.8 [68,79]. Some box-and-whisker diagrams (Figs 3E, 3G, 9D and 9E) were created using ggplot2; the remainder (Figs 3D, 7B, 8F, 9F and 9G) and bar plot (Fig 4A) were created with Matplotlib v3.2.2 (RRID:SCR_008624) and Seaborn v0.11.2 in Python. CellChat v1.4.0 [16] was used to visualize the cell-cell interactions in Fig 4B and 4C. The spatial plots (Figs 5B, 6A, 8A and 8C) were formed using Scanpy [60] and Squidpy v1.1.2 [62] in Python. The pairwise and stacked bar graphs (Figs 3F, 6C, 6D and 8B) were created using ggplot2. The Kaplan-Meier curves (Fig 9B, 9C and 9H–9M) were drawn using KaplanMeierFitter and lifelines v0.27.8 [72], matplotlib, and seaborn. For additional data processing tasks, we used R packages (tidyverse v1.3.1, tidyr v1.1.3, tibble v3.1.4, dplyrr v1.0.7, purrr v0.3.4, and stringr v1.4.0) and Python modules (numpy v1.21.5, pandas v1.3.5, anndata v0.7.8, stLearn v0.3.1, and scvi-tools v0.14.5 [66]).