Background

Locally advanced oral cavity squamous cell carcinoma (OCSCC) ranks among the most prevalent malignancies affecting the head and neck region. Locally advanced OCSCC constitutes a diverse group including all stage III/IV oral cavity tumors without distant metastasis.1 The standard therapeutic approach involves surgery followed by adjuvant radiation or chemoradiation therapy,2 yet recurrence remains a formidable risk for a majority of patients.

In 2016, immune checkpoint blockades (ICBs)—nivolumab and pembrolizumab, which target programmed death 1 (PD-1), were approved by the US Food and Drug Administration for the treatment of patients with head and neck squamous cell carcinoma (HNSCC). However, the approval was restricted to patients with recurrent or metastatic diseases3 and durable clinical responses were attained in only 13–18% of these patients.4 5 Notably, the utilization of ICBs in the adjuvant setting remains limited.

Immune checkpoints are a key regulator of the immune response in the tumor immune microenvironment (TME). Components of the TME have been linked to the response to ICBs,6 as well as outcomes7 in patients with HNSCC. For instance, higher tumor-infiltrating CD8 counts have been associated with improved overall survival in patients with HNSCC.8 However, the heterogeneity within the broader category of head and neck cancers, including tumors from diverse anatomical locations such as the oral cavity, hypopharynx, larynx and oropharynx,9 has often been overlooked. Recent research has underscored the distinct immune landscape from different anatomical sites in the head and neck.10

Hence, an in-depth immunological characterization of OCSCC and its association with clinical outcomes is currently lacking. This study aims to bridge this gap by conducting a comprehensive examination of the tumor microenvironment in OCSCC, focusing on those at a locally advanced stage. The objective is to identify predictive biomarkers of patient outcomes, thereby facilitating the customization of therapeutic strategies for this patient cohort.

Materials and methodsPatient cohort and sample collection

At enrollment, clinical information was recorded. A comprehensive review of all available medical records and patient questionnaires was conducted for the cohort. The clinical database for this cohort was last updated in December 2015. The study analyzed 132 patients with OCSCC, who met the inclusion criteria and were treated between 2007 and 2010, referred to as the “VCCC cohort”. Tumor samples were collected during surgical resections and were preserved as formalin-fixed, paraffin-embedded (FFPE) sections for further analysis. A thorough pathological review was undertaken to confirm tumor presence. Samples were classified into early-stage (stage I/II, n=78) and locally advanced stage (stage III/IV without distant metastasis, n=45) based on the eighth edition of the American Joint Committee on Cancer tumor, node, metastases system.11

For validation purposes, this study leveraged publicly available data from The Cancer Genome Atlas (TCGA) HNSCCs cohort.12 From this cohort, tumors originating from the oral cavity and classified as stage III/IV without distant metastasis, were selected for analysis. This subset included 163 samples and was designated as the “TCGA – locally advanced OCSCC cohort” (online supplemental table 1). Access to the data was facilitated through the Genomic Data Commons data portal and included RNA sequencing counts, T-cell receptor (TCR), virus status, and mutations, and the analysis performed using the R package TCGAbiolinks.13 Data was downloaded in March 2020.

Single-cell RNA sequencing (scRNA-seq) data from 18 patients with OCSCC were downloaded in October 2023 from a published study14 via Gene Expression Omnibus (GEO), through accession number GSE103322. This cohort encompassed data not only for stromal cells but also cancer cells, presenting an opportunity to delve into the characteristics of cancer cells expressing varying levels of PD-L1 and their interaction with immune cells.

Tumor-infiltrating lymphocytes assessment

Four µm sections were cut and stained with H&E from each FFPE block in the study. The H&E slides were evaluated by a trained pathologist to define “tumor” and “stromal” regions. Tumor-infiltrating lymphocytes (TILs) were evaluated on H&E-stained slides, applying the criteria set by the International Immuno-Oncology Working Group.15 This standardized approach facilitates the consistent and reproducible assessment of TILs. The criteria encompass both tumorous TILs (tTILs), located within the confines of the tumor nests or cell islands, and stromal TILs (sTILs), situated in the stromal areas surrounding the tumor cells. A trained pathologist scored the percentages of tTILs and sTILs for each sample in the study (online supplemental table 2). The data obtained from the TILs assessment were then correlated with patient survival outcomes.

Multiplex immunohistochemistry

A total of 58 samples were selected for multiplex immunohistochemistry (mIHC) staining, comprising 22 early-stage and 36 locally advanced-stage samples (online supplemental table 3). All patients in the early-stage group underwent adjuvant therapy, aligning their treatment with that received by patients in the locally advanced-stage group. mIHC was performed on tumor sections using the OPAL POLARIS 7-color Automation IHC Kit (Akoya BioSciences, California, USA) as per manufacturer’s instructions, to identify CD3 (SP7, Thermo Fisher), CD11c (EP157, Dako), CD68 (514H12, Biocare Medical), PD-L1 (SP263, Roche Diagnostics), P40(BC28, Biocare Medical) and DAPI (Invitrogen) (online supplemental table 4). Automated staining was performed on the sections using Leica Biosystems BOND RX autostainer (Wetzlar, Germany) according to the manufacturer’s instructions. For detailed steps please refer to the method section in Jacquelot et al paper.16 All slides were subsequently imaged on a Vectra Polaris quantitative pathology imaging system (Akoya Biosciences). High-powered images at 20× were selected using Phenochart (Akoya Biosciences) and imported into inForm Software V.2.4 (Akoya Biosciences) for deconvolution. The deconvoluted files from inForm V.2.4 were exported into HALO (Indica Labs, New Mexico, USA) and merged to perform tissue and cell segmentation as well as cell phenotyping using the HighPlex V.3.0.3 module on HALO software. Cell segmentation was assessed based on all cells counter-stained with 4′,6-diamidino-2-phenylindole (DAPI). Manual annotations in HALO segmented the tissue into “tumor” and “stromal” regions guided by pathologist markups of H&E stained sections. Cell phenotyping on inForm was undertaken by selecting at least five representative cells per phenotype, and then performing reiterations until at least 20 representative cells per phenotype were selected. Cell phenotyping on HALO was performed by thresholding the individual markers in the panel for each slide. The density of cell subsets was subsequently calculated per mm2.

Differential gene expression, gene sets enrichment, immune cell deconvolution analyses of TCGA—locally advanced OCSCC cohort

Trimmed mean of M-values normalized counts from bulk RNA-seq data was processed using edgeR.17 FilterByExpr was used to filter out lowly expressed genes and normalized the gene expression to be counts per million (cpm) based on normalized library sizes. High versus low expression of PD-L1 (PD-L1hi vs PD-L1lo) were defined based on the top and bottom quartiles of messenger RNA (mRNA) expression, using cut-off cpm values of 21.365 and 4.5968, respectively. Differential gene expression analyses were performed between PD-L1hi and PD-L1lo samples using edgeR (V.3.40.1). Genes with a false discovery rate <0.05 were determined as differentially expressed. We quantified the pathway enrichment scores of hallmark gene sets collection18 for each sample using gsva function under the parameters of “Gaussain” kcdf and “gsva” method from GSVA R package (V.1.46.0).19 Spearman’s correlation was performed to assess the correlation between individual pathway enrichment scores and PD-L1 mRNA expression across all samples using the cor.test function from stats R package (V.4.2.1).20 The abundance of immune cell types in each sample was estimated using quanTIseq21 and CIBERSORT22 deconvolution algorithms implemented in immunedeconv R package.23

Single-cell RNA sequencing analysis

Published scRNA-seq data14 from the GEO database (Accession Number GSE103322), available on National Centre for Biotechnology Information (NCBI, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE103322) were used in this study. We excluded all single-cell samples from the lymph node and only used those from the primary tumor. A total of 18 samples were in the scRNA-seq cohort, with 15 samples having cancer cells and 3 samples (samples 8, 9 and 23) with no cancer cells evaluable. We used the TPM (transcript per million) normalized expression values and cell type annotation provided by the authors. Single-cell data was processed using Seurat (V.5.0.2).24–26 All cancer cells were subset from the 15 samples. ScaleData was used to normalize the gene expression across cells and determine the top 2,000 variable genes to perform principal component analysis (PCA). Uniform manifold approximation and projection (UMAP) reductions and cell neighbors were calculated using 25 dimensions from the PCA reduction. Clusters were detected with a resolution of 0.8. Five samples (samples 7, 10, 12, 13 and 24) were dropped for further analyses due to the limited number (cell number less than 10) of cancer cells. The remaining 10 samples were grouped based on their PD-L1 expression on cancer cells and tumor stages. To compare the gene expression profiles between PD-L1hi locally advanced tumors and PD-L1lo locally advanced tumors, FindMarkers was performed to explore the differentially expressed genes (DEGs) using the default parameters. Then gene set enrichment analysis (GSEA) was performed using the GSEA function with minGSSize of 15 and maxGSSize of 500 from clusterProfiler (V.4.10.1)27 28 on the hallmark gene set collection18 to find the significantly enriched pathways based on DEGs. To compare the gene expression profiles between PD-L1lo early-stage tumors and PD-L1lo locally advanced tumors, we performed the same analysis as above. T-cell populations annotation was performed using SingleR (V.2.4.1)29 based on the cell type reference of MonacoImmuneData30 downloaded from celldex (V.1.12.0).29 All non-T cell groups were grouped as “other cells”, and all minor T-cell subtypes with fewer than 10 cells were grouped as “other T cells”. All further analyses and plots were generated in R (V.4.3.3) using tidyverse,31 ggplot2 (V.3.4),32 ComplexHeatmap,33 scCustomize (V.2.1.2)34 and EnhancedVolcano (V.1.20.0).35

Statistical analysis

All statistical analyses were performed using the GraphPad Prism software (V.9.2.0, GraphPad, San Diego, California, USA) unless stated otherwise. Statistical analyses of quantifications were performed with the two-tailed Mann-Whitney U test between two groups as appropriate. For survival analyses, Kaplan-Meier plots were drawn, and statistical differences were evaluated using the log-rank Mantel-Cox test. For the correlation analyses, Pearson correlation was employed to assess relationships between cell densities and PD-L1 expression. Spearman’s correlation was performed to assess the correlations between individual hallmark gene sets and PD-L1 mRNA expression. A p≤0.05 was considered statistically significant.

ResultsElevated PD-L1 expression in locally advanced OCSCC indicates unfavorable outcomes after standard treatment

In light of the clinical relevance of TILs in breast cancer,15 we employed the same TILs evaluation method proposed by the International Immuno-Oncology Working Group36 to evaluate TILs in OCSCC. No significant difference in overall survival was observed when assessing patients across all clinical stages (figure 1A left panel). Surprisingly, when the outcomes of early-stage and locally advanced stages were compared, we observed divergent survival outcomes. TILs, more specifically sTILs, were associated with prolonged survival in the early-stage patients (p<0.05, figure 1A middle panel), but with poor outcomes in patients with locally advanced tumors (p<0.05, figure 1A right panel). We hypothesized that the presence of TILs in the tumor microenvironment may reflect distinct immune states that, in turn, bring about varying survival outcomes. Adjuvant radiation or chemoradiation is routinely employed following surgical resection of locally advanced OCSCC, and the improved patient survival noted in our study cohort (online supplemental figure S1A), aligned with previous studies.37 Notably, the prognostic significance of sTILs remained consistent irrespective of the type of adjuvant therapy the patients underwent (online supplemental figure S1A).

Figure 1

Elevated PD-L1 expression in locally advanced OCSCC indicates poor outcome. (A) Survival stratification based on stromal TILs (sTILs) assessment on H&E slides from patients with OCSCC with all stages (n=132), early-stage (n=78) or locally advanced stage (n=45). (B) Survival stratification based on CD3+, CD68+, CD11c+ or PD-L1+ cell densities (per mm2 in early-stage OCSCC (n=22). (C) Survival stratification based on CD3+, CD68+, CD11c+ or PD-L1+ cell densities (per mm2 in locally advanced OCSCC (n=36). (D) Differences of CD3+, CD68+, CD11c+ or PD-L1+ cell densities (per mm2 between patients with good or poor outcomes in locally advanced OCSCC (n=36). (E) Survival differences between PD-L1hi (top 25% n=41) and PD-L1lo (bottom 25% n=41) in the TCGA—locally advanced OCSCC cohort. (F) Differences in PD-L1 messenger RNA expression between patients with good or poor outcomes in the TCGA—locally advanced OCSCC cohort (n=163). Data are expressed as mean±SEM. OCSCC, oral cavity squamous cell carcinoma; PD-L1, programmed death ligand 1; TCGA, The Cancer Genome Atlas; TILs, tumor-infiltrating lymphocytes.

To further investigate the composition of immune cells in the tumor microenvironment, we performed mIHC to identify immune cells as CD3+ (T cells), CD68+ (macrophage), CD11c+ (dendritic cell), PD-L1+ and P40+ (cancer cells). All 36 tumor blocks of locally advanced OCSCC along with 22 early-stage tumors from the “VCCC cohort” were retrieved. All patients in the early-stage group underwent adjuvant therapy, aligning their treatment with that received by patients in the locally advanced-stage group. In the early-stage tumors, increased CD3+ cell density in the tumor was associated with significantly superior patient survival (p<0.05, figure 1B), while no significant differences in survival were observed using CD68+, CD11c+ or PD-L1+ cell densities individually (figure 1B). In contrast, CD3+, CD68+ or CD11c+ cell density did not stratify patients’ survival in the locally advanced OCSCC (figure 1C). Instead, patients with increased PD-L1+ cell infiltration had a poor survival outcome (p<0.05, figure 1C). When we further stratified the locally advanced OCSCC into those with good versus poor outcomes, significantly more PD-L1+ cells were found in the poor outcome group (p<0.05, figure 1D), while no differences were observed in CD3+, CD68+, and CD11c+ cells between the two groups (figure 1D).

To validate the clinical relevance of PD-L1 expression in locally advanced OCSCC, we identified 163 locally advanced OCSCC in the TCGA cohort. Consistent with our findings in mIHC, PD-L1hi patients experienced inferior outcomes compared with PD-L1lo patients (p<0.05, figure 1E). The PD-L1 (CD274) expression in the patients with poor outcomes was significantly higher than in the good outcome group (p<0.05, figure 1F), affirming the clinical relevance of PD-L1 as a potential prognostic marker.

PD-L1hi tumors display an active immune response phenotype within the tumor microenvironment

In light of the adverse prognostic value of PD-L1 in locally advanced OCSCC, we further characterized the tumor microenvironment in the “VCCC cohort” and the “TCGA-locally advanced OCSCC cohort”, respectively. In the VCCC cohort, we observed higher infiltration of CD3+ cells (p<0.001) and CD68+ cells (p<0.005), but not CD11c+ cells, in the PD-L1hi group compared with the PD-L1lo group (figure 2A). A positive and significant correlation was observed between CD3+ T cells and PD-L1+ cells, (r=0.49. p<0.05) as well as CD68+ macrophage cells and PD-L1+ cells (r=0.55, p<0.05) (online supplemental figure S1B). These positive correlations between PD-L1 expression and the density of T cells or macrophages in the tumor suggest the PD-L1 expression likely represents a result of dynamic immune cell interactions.

Figure 2

PD-L1hi tumors demonstrate an active immune response in the tumor microenvironment in locally advanced OCSCC. (A) Differences in CD3+, CD68+ and CD11c+ cells densities (per mm2 between PD-L1hi (n=18) and PD-L1lo (n=18) locally advanced OCSCC. (B) Differentially expressed genes between PD-L1hi (n=41) and PD-L1lo (n=41) samples from the TCGA—locally advanced OCSCC cohort. (C) Spearman’s correlation was performed to assess the correlations between individual hallmark gene sets and PD-L1 mRNA expression. Gene sets with rho values above 0.40 were illustrated. (D) Differences in immune cell composition using quanTIseq deconvolution analysis between PD-L1hi (n=41, left) and PD-L1lo (n=41, right) samples. (E) Differences of B cell, M1, M2, monocyte, neutrophil, NK cell, CD4+ non-regulatory cell, CD8+ T cell, Tregs, and myeloid dendritic cell populations between PD-L1hi (n=41) and PD-L1lo (n=41) samples using quanTIseq method. (F) Differences of total TCR, TCR-α, and TCR-β reads, Shannon’s entropy index, number of clones and neoantigens between PD-L1hi (n=41) and PD-L1lo (n=41) samples. Data are expressed as mean±SEM. DEGs, differentially expressed genes; FDR, false discovery rate; NK cell, natural killer cell; OCSCC, oral cavity squamous cell carcinoma; PD-L1, programmed death ligand 1; TCGA, The Cancer Genome Atlas; TCR, T-cell receptor; Tregs, T regulatory cells.

In the TCGA—locally advanced OCSCC cohort, we observed a list of 6,784 genes differentially expressed between the PD-L1hi and PD-L1lo groups (online supplemental table 5), with a majority of immune-related genes upregulated in the PD-L1hi tumors (figure 2B). Furthermore, a Spearman’s correlation was performed to assess the correlation between individual hallmark gene sets and PD-L1 mRNA expression. The gene sets related to the immune response were positively correlated with PD-L1 expression (figure 2C), including interferon-gamma response (rho=0.75), interferon alpha response (rho=0.70), IL-6 JAK STAT3 signaling (rho=0.67), inflammatory response (rho=0.65), allograft rejection (rho=0.63), complement (rho=0.62), and IL-2 STAT5 signaling (rho=0.55). Interestingly, gene sets including TNF-α signaling via NF-κB (rho=0.55), apoptosis (rho=0.54), KRAS signaling up (rho=0.46), apical surface (rho=0.45), PI3K AKT mTOR signaling (rho=0.42), and P53 pathways (rho=0.41), also had a weak positive correlation with PD-L1 expression. A heatmap of 50 hallmark gene sets in 163 samples from the TCGA—locally advanced OCSCC cohort based on PD-L1 expression is shown in online supplemental figure S2). To further characterize immune cell composition in the tumor microenvironment in the PD-L1hi and PD-L1lo tumors, we used the immune deconvolution method—quanTIseq21 (online supplemental table 6). This showed that PD-L1hi (figure 2D left) and PD-L1lo (figure 2D right) tumors have distinct immune cell compositions. More specifically, PD-L1hi tumors exhibited an active immune response phenotype with increased M1 macrophages (p<0.0001), CD8+ T cells (p<0.0001) and T regulatory cells (Tregs, p<0.0001), while PD-L1lo tumors showed a suppressive phenotype with more M2 macrophage (p<0.005), CD4+ non-regulatory T cells (p<0.05) and myeloid dendritic cells (p<0.05) (figure 2E). No significant differences in B cell, monocyte, neutrophil, or natural killer cell were observed between the two groups (figure 2E). Similar observations were made using the CIBERSORT method (online supplemental table 7), indicating increased percentages of M1 macrophages (p<0.0001) and CD8+ T cells (p<0.0001) in PD-L1hi tumors (online supplemental figure 3A). Further analysis of TCR profiles (online supplemental table 7) revealed an increase in the number of TCR-α (p<0.0001), TCR-β (p<0.0001), and total TCR (p<0.0001) reads in PD-L1hi tumors (figure 2F). Notably, not only was there an increase in the number of T-cell clones (p<0.0001), but also a greater diversity (p<0.0001) within the T-cell repertoire, as evidenced by Shannon entropy measurements, in PD-L1hi tumors (figure 2F). However, the enhanced complexity of the TCR repertoire in PD-L1hi tumors did not stem from an increase in virus pathogens (online supplemental table 8), including cytomegalovirus (CMV), Epstein-Barr virus (EBV), human herpesvirus (HHV), and human papillomavirus (HPV), or immunogenic mutations or neoantigens (online supplemental table 8 and figure 2F). These findings suggest that alternative mechanisms, other than viral pathogens or genetic mutations, may contribute to the significant increase and diversification of TCRs in PD-L1hi tumors. For instance, cytokines in the tumor can affect TCR diversity and number by altering T-cell survival and differentiation. Taken together, PD-L1hi tumors manifest an active immune response phenotype within the tumor microenvironment. The PD-L1 expression in these tumors likely reflects an ongoing antitumor immune response.

PD-L1 expression on cancer cells, rather than immune cells, is associated with adverse outcomes in the locally advanced OCSCC

The expression of PD-L1 on cancer cells and non-cancer cells has been shown to impact patient outcomes.38–40 mIHC provided the opportunity to identify which cell type expressed PD-L1 on the cell surface. Interestingly, we observed that PD-L1 on cancer cells (p<0.01), but not immune cells, was significantly elevated in the poor outcome group (figure 3A). In light of this, we next stratified the patients’ survival on the cell types expressing PD-L1. An increased number of PD-L1+ cancer cells (p<0.05) in the tumor was associated with poor outcomes, however, this was not the case for PD-L1+ CD3, CD68 or CD11c cells, (figure 3B). Examples of high-resolution images of mIHC for patients with good and poor outcomes are shown in figure 3C.

Figure 3

PD-L1 expression on the cancer cells, not immune cells, is associated with poor outcomes in the locally advanced OCSCC. (A) Differences in CD3+PD-L1+, CD68+PD-L1+, CD11c+PD-L1+ and P40+PD-L1+ cell densities (per mm2 between patients with good and poor outcomes (n=36). (B) Survival stratification based on CD3+PD-L1+, CD68+PD-L1+, CD11c+PD-L1+ and P40+PD-L1+ cell densities (per mm2 in locally advanced OCSCC (n=36). (C) Examples of multiplex staining for CD11c (magenta), CD68 (yellow), PD-L1 (green), CD3 (red) and P40 (white) in patients with good (n=2) and poor (n=2) outcomes. Scale bar, 50 µm. (D) Survival stratification based on stromal TILs (sTILs) and PD-L1 densities in locally advanced OCSCC tumors (n=36). Data are expressed as mean±SEM. OCSCC, oral cavity squamous cell carcinoma; PD-L1, programmed death ligand 1; TILs, tumor-infiltrating lymphocytes.

The PD-1/PD-L1 immune checkpoint is a key pathway in the regulation of antitumor immunity.41 This discovery played a pivotal role in the development of the revolutionary ICB.42 Recent studies revealed that cancer cell-intrinsic PD-L1 signaling operates independently of PD-1 engagement or interactions with immune cells.40 This insight elucidates why certain immune-cold tumors exhibit upregulated PD-L1 expression. Taking this into account, we used data on TILs to categorize tumors as either immune “hot” or “cold”. Through the amalgamation of sTILs assessment and PD-L1 expression, we stratified patients into PD-L1hi or PD-L1lo and “hot” or “cold” tumors. Intriguingly, PD-L1hi tumors were associated with inferior outcomes when compared with PD-L1lo tumors, irrespective of their “hot” or “cold” classification based on sTILs assessment (figure 3D). This finding indicates signaling pathways that downregulate PD-L1 expression on cancer cells may confer protective effects on patients undergoing adjuvant radiation or chemoradiation therapy.

Complex regulatory networks other than immune interactions are involved in PD-L1 expression on cancer cells in locally advanced tumors

Subsequently, our interest turned to understanding the features of cancer cells that are associated with PD-L1 expression. By using the scRNA-seq cohort from the study conducted by Puram et al,14 we successfully identified OCSCC samples exhibiting either PD-L1hi or PD-L1lo cancer cells. First, cancer cells from 15 samples were clustered based on Sample-IDs (figure 4A), and the expression of PD-L1 on the cancer cells was further determined (figure 4B and C). Notably, 10 of these samples—5, 6, 16, 17, 18, 20, 22, 25, 26, and 28—showed a significant presence of cancer cells (figure 4C). In contrast, samples 7, 10, 12, 13, and 24 had fewer cancer cells (n<10) and were therefore excluded from further analysis. Figure 4D illustrates the remaining 10 OCSCC tumors, detailing their tumor stages, PD-L1 expression levels on cancer cells, and cell compositions. Based on the clinical and pathological information illustrated in online supplemental table 2 from Puram et al,14 and considering that all samples included are without metastasis, sample 5 (T2N1), sample 16 (T2N0), and sample 22 (T1N0) were categorized as early-stage tumors. In contrast, the rest of the samples, such as sample 6 (T4aN2c), sample 17 (T4aN0), sample 18 (T3N1), sample 20 (T4aN2c), sample 25 (T3N1), sample 26 (T4aN2c), and sample 28 (T2N2c), were categorized as locally advanced tumors. Combing the tumor stages and PD-L1 expression levels on cancer cells, we then categorized the samples into three groups: samples 6 and 20 as PD-L1hi locally advanced tumors; samples 17, 18, 25, 26 and 28 as PD-L1lo locally advanced tumors; and samples 5, 16 and 22 as PD-L1lo early-stage tumors.

Figure 4

Complex regulatory networks other than immune interactions are involved in PD-L1 expression on cancer cells in locally advanced tumors. (A) Cancer cell clusters based on sample IDs from the single cell RNA sequencing cohort. (B) PD-L1 expression on individual cancer cells from different tumor samples. (C) Violin plot of PD-L1 expression on individual cancer cells from different tumor samples. (D) Single-cell compositions in the samples (n=10) with evaluable cancer cells (cell number >10). (E) Top 50 differentially expressed genes (DEGs) in cancer cells between samples 6 and 20 (PD-L1hi locally advanced tumors) and samples 17,18, 25, 26 and 28 (PD-L1lo locally advanced tumors). (F) Volcano plot of 6,475 variables from the comparison in panel E with immune-related genes labeled. (G) Enriched hallmark gene set analysis of DEGs from the comparison in panel E. NES, normalized enrichment score; PD-L1, programmed death ligand 1; UMAP, uniform manifold approximation and projection.

Interestingly, samples 6 and 20, despite lacking T cells (n=0), exhibited elevated PD-L1 expression on cancer cells. While this observation is based on a limited number of samples (n=2), it suggests that the PD-L1 expression on cancer cells in these two specimens may be influenced by factors other than the interaction with T cells in the tumor microenvironment. These two samples provided a unique opportunity to elucidate the genes or pathways influencing PD-L1 expression in cancer cells. We then conducted differential gene expression (figure 4E and F and online supplemental table 9) and pathways enrichment analyses (figure 4G and online supplemental table 10) by comparing cancer cells from samples 6 and 20 against samples 17, 18, 25, 26 and 28. Notably, in the PD-L1hi samples (6 and 20), genes from the MAGE family (MAGEA12, MAGEA2B, MAGEA2), the UDP glucuronosyltransferase family (UGT1A5, UGT1A10, UGT1A7), and other genes including CPXM1, DMRT2, SPP1, SCN9A, CARD11, DSCR8, DNAJB3, FGFBP2, CYP4F3, FREM2, and GSTM1 were upregulated. Conversely, genes upregulated in PD-L1lo samples varied, but genes such as HLA-DRB6, SULT1E1, HEPHL1, WISP2, VAV3, PDZK1IP1, and LOC645638 were consistently elevated in all five samples (figure 4E). Interestingly, most immune-related genes including genes associated with antigen presentation (HLA-DRB1, HLADRB6, HLADRA) and T-cell trafficking (CXCL9, CXCL10, CXCL11) were prevalent in the PD-L1lo samples, while genes other than CD274, such as LEF1, PTPRC, SELL, PDCD1LG2 were elevated in the PD-L1hi samples (figure 4F). This may explain the limited T-cell presence observed in samples 6 and 20 (figure 4D).

Further Gene Set Enrichment Analysis revealed most immune-related pathways, including complement, IL-6 JAK STAT2 signaling, allograft rejection, inflammatory response, interferon-gamma response, and interferon alpha response were enriched in the PD-L1lo samples, alongside pathways like ultraviolet (UV) response down, estrogen response, hypoxia, Kirsten rat sarcoma virus (KRAS) signaling up, transforming growth factor (TGF)-β signaling, apoptosis, coagulation, epithelial-mesenchymal transition (EMT) and tumor necrosis factor (TNF)-α signaling via nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) (figure 4G and online supplemental table 10). In contrast, the reactive oxygen species (ROS) pathway was the only pathway enriched in the PD-L1hi samples (figure 4G and online supplemental table 10). This emphasizes that immune interactions are not essential for PD-L1 upregulation in cancer cells.

PD-L1lo locally advanced tumors have a more complex pathway enrichment and diverse T-cell infiltrate compared with early-stage tumors

The scRNA-seq cohort offered a valuable opportunity to examine the characteristics of PD-L1lo cancer cells in both early-stage and locally advanced OCSCC tumors. We performed differential gene expression (figure 5A and B and online supplemental table 11) and pathway enrichment analyses (figure 5C and online supplemental table 12) by comparing cancer cells from samples 5, 16, and 22 to those from samples 17, 18, 25, 26, and 28. Interestingly, the number of DEGs between early and advanced stage tumors was significantly lower than those between PD-L1lo and PD-L1hi in locally advanced tumors, with 1,572 and 6,475 DEGs, respectively. Similar patterns were observed in immune-related genes, as depicted in figure 5B. These observations suggest PD-L1 expression might be a more influential factor in representing gene expression changes than the stage of the tumor itself. Further Gene Set Enrichment Analysis identified that inflammatory responses, KRAS signaling downregulation, hypoxia, and TNF-α signaling via NF-κB were predominantly enriched in the locally advanced tumors (samples 17, 18, 25, 26, and 28, shown in figure 5C and online supplemental table 12). These findings indicate these pathways may be involved in modulating the tumor microenvironment when tumor progresses.

Figure 5

More complex pathway enrichment and diverse T cells in PD-L1lo locally advanced tumors compared with early-stage ones. (A) Top 50 differentially expressed genes (DEGs) in cancer cells between samples 5, 16 and 22 (PD-L1 lo early-stage tumors) and samples 17,18, 25, 26 and 28 (PD-L1lo locally advanced tumors). (B) Volcano plot of 1,572 variables from the comparison in panel A with immune-related genes labeled. (C) Enriched hallmark gene set analysis of DEGs from the comparison in panel A. (D) Composition of cell phenotypes of T cells from samples 17, 18 and 25 (PD-L1lo locally advanced tumors) using the Monaco method. (E) Composition of cell phenotypes of T cells from samples 5 and 16 (PD-L1lo early-stage tumors) using the Monaco method. (F) Differences in T-cell composition between PD-L1lo locally advanced tumors (samples 17, 18 and 25) and PD-L1lo early-stage tumors (samples 5 and 16). NES, normalized enrichment score; PD-L1, programmed death ligand 1; Th cells: T helper cells; UMAP, uniform manifold approximation and projection; Vd2 gd T cells: V-delta 2 gamma delta T cell.

Interestingly, samples 6, 20, 26, 28 and 22 exhibited minimal T-cell presence (figure 4D), indicating a “cold” tumor phenotype, while samples 5, 16, 17, 18 and 25 displayed abundant T cells (figure 4D), reflecting a “hot” tumor phenotype. Additionally, all these five “hot” tumors demonstrated low PD-L1 expression on cancer cells (figure 4C). To further understand the composition of these T cells, we next annotated the T cells from PD-L1lo locally advanced tumors (samples 17, 18, and 25) and PD-L1lo early-stage tumors (samples 5 and 16) based on Monaco30 reference data sets (figure 5D and E). The T cells in the PD-L1lo locally advanced tumors were notably more diverse than those in the PD-L1lo early-stage tumors (figure 5F). Specifically, T cells from samples 17, 18, and 25 comprised 11 subsets (figure 5D) whereas only 7 subsets were identified from samples 5 and 16 (figure 5E). Both groups exhibited CD8+ T cells, including central memory and effector memory cells, and CD4+ T cells, including follicular helper cells, Tregs and Th17 cells, although with different percentages(figure 5F). Notably, in PD-L1lo locally advanced tumors, distinct subsets such as terminal effector CD8+ T cells, Th1 and Th2 cells CD4+ T cells, and Vδ−2 γδ T cells were present (figure 5F). The increased diversity of T cells in locally advanced tumors compared with early-stage ones indicates a more complex immune response as the tumor progresses.

Conclusion and discussion

Our study systematically evaluated the prognostic value of PD-L1 in patients with locally advanced OCSCC. The finding that high PD-L1 expression correlates with poorer outcomes after adjuvant therapy challenges existing assumptions and emphasizes the complex role of PD-L1 in cancer progression and treatment response. To understand potential mechanisms to explain this finding, we used multiple cohorts and advanced technologies in this study to unravel the intricate relationships between PD-L1 expression, immune responses, and molecular pathways within the tumor microenvironment.

Our investigations revealed that T cells are indicative of favorable outcomes in early-stage OCSCC, while their prognostic significance diminishes in locally advanced tumors. Conversely, high PD-L1 expression is a strong predictor of poor outcomes in these advanced cases. Both early-stage and locally advanced tumor cohorts assayed with mIHC staining in this study underwent adjuvant therapy, this suggests the influence of radiation on patients’ immune response varies between early-stage and locally advanced tumors. Additionally, scRNA-seq on cancer cells revealed that T-cell populations in early-stage tumors were less diverse than those in advanced tumors. Collectively, these findings led us to hypothesize that in early-stage OCSCC, adjuvant radiation might provoke a more robust immune response, resulting in more diverse immune cell populations and potentially prolonging patient survival. Supporting this, recent studies suggest radiation acts as an “accelerant”, similar to in situ vaccination, by killing tumor cells and eliciting a systemic immune response.43 The most representative example is the abscopal effect: radiation on one site may cause tumor regression at remote and distant non-irradiated sites.44

In contrast, in the locally advanced OCSCC tumors, our study revealed that PD-L1hi tumors exhibit an active immune response phenotype within the tumor microenvironment, as indicated by mIHC and TCGA data. This suggests the preservation of systemic immune response in PD-L1hi locally advanced tumors. However, the benefits from adjuvant