A highly resolved integrated single-cell atlas of HPV-negative head and neck cancer

Abstract

Background

Head and Neck Squamous Cell Carcinomas (HNSCC) are the seventh most prevalent form of cancer and are associated with human papilloma virus infection (HPV-positive) or with tobacco and alcohol use (HPV-negative). HPV-negative HNSCCs have a high recurrence rate, and individual patients’ responses to treatment vary greatly due to the high level of cellular heterogeneity of the tumor and its microenvironment.

Methods

Here, we describe a HPV-negative HNSCC single cell atlas, which we created by integrating six publicly available datasets encompassing over 230,000 cells across 54 patients. We classify cell types, subpopulations, and their expression programs in the immune, mesenchymal, endothelial and epithelial compartments. We interrogate the relationship between cell types through hierarchical clustering, cell-cell communication analysis and correlating populations changing together across patients.

Results

We resolve the myeloid and fibroblast compartments, revealing an IL1B+ myeloid population previously unexplored in HNSCC and clarifying two immune cancer associated fibroblast populations that are frequently conflated, identify sex-associated changes in cell type proportions, and a unique interaction between CXCL8-positive fibroblasts and vascular endothelial cells.

Conclusions

We utilize the atlas to contextualize the relationships between existing signatures and cell populations, harmonize nomenclature across studies, and show the power of this large-scale resource to robustly identify associations between transcriptional signatures and clinical phenotypes that would not be possible to discover using fewer patients. Beyond our findings, the atlas serves as a public resource for the high-resolution characterization of tumor heterogeneity of HPV-negative HNSCC.

Kroehling et al. build an integrated single-cell RNA-sequencing atlas of HPV-negative head and neck squamous cell carcinoma comprising over 230,000 cells from 54 patients. It revels distinct immune, stromal, and tumor populations, harmonizes annotations and links transcriptional profiles to clinical features.

Plain Language Summary

HPV-negative head and neck squamous cell carcinomas (HNSCC) are a type of cancer that develops in and around the mouth and is often linked to smoking and drinking alcohol. They have a high recurrence rate and respond differently to treatment due to the high level of variation between patients in both their tumor cells and surrounding cells. In this study, we combined cell-level data from many people with HPV-negative HNSCC to create a large, unified dataset of the cell types found in these tumors. Using computational analyses, we identified and characterized the different cell types and their features. We then related those cellular features to how the tumors behaved in different patients. These results advance our understanding of how particular combinations of tumor and surrounding cells can influence cancer growth and treatment response. This knowledge could help guide more precise, personalised approaches to treating HNSCC.

Introduction

Head and Neck Squamous Cell Carcinomas (HNSCCs) arise from the mucosal epithelium lining of the oral cavity, larynx, and pharynx^1,2. The two major subtypes are HPV-positive and HPV-negative HNSCC, the former more frequently occurring in the pharynx and associated with human papilloma virus infection, especially the HPV-16 strain^3,4, and the latter more frequently occurring in the oral cavity and larynx and associated with carcinogens in alcohol, tobacco and betel quid, among others.

Amongst all cancers, HNSCCs have the seventh highest incident rate, accounting for almost 7% of cancers worldwide, with 75% being HPV-negative^3,5,6. According to the GLOBOCAN database, cancers of the lip and oral cavity are the leading cause of cancer death for men in India⁶. HPV-negative cancers tend to have higher mutational burdens than HPV-positive cancers and lower levels of immune infiltration in the tumor microenvironment (TME). Together, these factors result in clinicopathological differences and a significant difference in 5-year survival rates between the two subtypes, with a 75-80% rate for HPV-positive HNSCCs and a 55% rate for HPV-negative HNSCCs^4,7–9. Consequently, the two subtypes have now been separated in the American Joint Committee of Cancer (AJCC) staging system and are commonly considered distinct diseases^7,10.

Current treatments for the disease include radiation, chemotherapy, combination therapy, and immune checkpoint inhibitors (ICIs). However, response to treatment as well as mode of effective treatment varies greatly across individuals due to the high degree of heterogeneity both in the genetic profiles and plasticity of the cancer cells, as well as in the cell type composition of the TME, even within HPV-negative HNSCCs^11,12. HNSCC patients have a poor overall response to ICI therapy, with less than 20% presenting a durable response^11,13.

Significant progress has been made in characterizing the HNSCC TME. HPV-negative tumor epithelial cells are commonly characterized by tumor suppressor inactivating mutations in genes TP53, CDKN2a, and PTEN, and amplifications in CCND1, which together enable the progression of the disease^4,14. Previously identified transcriptional signatures in tumor cells include response to hypoxia, stress, epithelial-to-mesenchymal transition (EMT), and partial EMT (pEMT), a molecular phenotype in which tumor cells become more plastic to acquire a mesenchymal-like state and enhanced migratory capabilities¹⁵. Additionally, immune and stromal cells present in the TME play a major role in affecting tumor growth through TME interactions. For example, cancer-associated fibroblasts (CAFs) have been shown to affect tumor cell migration by remodeling the extracellular matrix¹⁵. Natural killer (NK) and CD8 + T cells have been found to take on effector phenotypes and kill tumor cells, whereas myeloid-derived suppressor cells (MDSCs) and Tregs secrete cytokines that suppress the immune system⁴. Further delineation of these complex dynamics in the non-HPV TME is critical for discovering new potential avenues of therapeutic intervention that may drive heterogeneous patient responses to ICI.

Single-cell RNA-sequencing (scRNA-seq) has emerged as the technology of choice to profile the individual transcriptomes of tumor cells and the surrounding TME. scRNA-seq enables the characterization of tumor heterogeneity within and across patients through the identification of cell types – here defined as categories of cells that perform a specific function (e.g., T cell, Fibroblast), and specific cell states – here defined as transient molecular changes in gene expression in response to endogenous and exogenous cues that multiple cell types can share (e.g., dysfunction, proliferation)^15–18. While individual single-cell HNSCC datasets have produced invaluable findings, given the relatively small sample size of single studies, and the relatively small number of publicly available single-cell HNSCC datasets, we hypothesized that their integration and combined analysis would facilitate contextualization of findings from individual studies, and provide increased statistical power to identify novel associations between molecular and clinical phenotypes¹⁹. Other HNSCC atlases have been previously developed, however the focus has been on a single cell type^20,21, contrasting the HPV-positive and HPV-negative subtypes¹⁶, or on the stepwise progression of the disease^22,23, leaving a need for the comprehensive characterization of the HPV-negative subtype.

To this end, we have integrated six publicly available datasets (Puram et al.¹⁵, Cillo et al.²⁴, Kurten et al.²⁵, Peng et al.²⁶, Choi et al.²³, Quah et al.²⁷) for the high-resolution characterization of the cells in the TME of HPV-negative HNSCC. Creation of the atlas allowed us to perform four types of analyses. Contextualization, in which we taxonomically organize different cell types and identify how closely related previously identified populations are to each other, and identify cell populations whose proportional changes are associated across patients. Harmonization, in which previously described populations with the same characteristics but differently labeled in the individual studies are assigned a consistent nomenclature. Identification and characterization of subpopulations and interactions, in which we leverage the increased number of cell profiles in the atlas to identify and annotate cell types and predict interactions. Association with clinical phenotypes, in which the increased sample size of the atlas increases the statistical power to detect significant associations between cell type compositional changes or cell state signatures, and clinical phenotypes such as stage and sex. In summary, we leverage published data to create and annotate an HPV-negative HNSCC transcriptomic atlas comprised of over 50 patients and 230,000 cells that may serve as a public resource. The atlas integrates six existing datasets to produce a comprehensive landscape of HNSCC tumors and their microenvironment, supporting novel findings, providing a reference for generating and contextualizing new hypotheses, and connecting molecular phenotypes to clinical end points.

Methods

Data acquisition

Six datasets were integrated in this study, totaling 54 patients and 232,015 cells. For five datasets, the processed matrix files containing raw counts and gene and barcode tsv files were downloaded from GEO: Peng²⁶ (GSE172577), Choi²³ (GSE181919), Cillo²⁴ (GSE139324), Quah²⁷ (GSE225331), and Kurten²⁵ (GSE164690). Due to availability, processed normalized data were downloaded for the last dataset, Puram¹⁵ (GSE103322). HPV-positive samples were removed to retain only HPV-negative samples for use in the atlas. The studies were approved by the following Institutional Review Boards. Puram: Patients at the Massachusetts Eye and Ear Infirmary (MEEI) were consented preoperatively to take part in the study following Institutional Review Board approval (Protocol #11-024H). Quah: The study of patient tumor samples was approved by SingHealth Centralized Institutional Review Board (CIRB: 2014/2093, 2018/2512 and 2016/2757) with each patient’s written consent. Choi: All the experiments with patient samples were performed under the approval of the Ajou University Institutional Review Board, using the approved protocol AJIRB-BMR-SMP-18-150 with the written informed consent of all patients. Cillo: All patients provided informed written consent, and the study was approved by the Institutional Review Board (University of Pittsburgh Cancer Institute, Protocol 99-069). Kurten: For all human patient samples, informed written consent was obtained prior to donation. The University of Pittsburgh Cancer Institute Review Board (Protocol 99-069) approved the study. Peng: The study was approved by the Institutional Review Board of Sun Yat-Sen Memorial Hospital, China (No. SYSEC-KY-KS-2021-040). Informed written consent was obtained from each participant.

Dataset pre-processing

Before integration, each dataset was pre-processed individually using Seurat v4²⁸ (RRID:SCR_007322) to eliminate low-quality cells and annotate the cells by cell type. If not pre-filtered in the originating study, cells with fewer than a minimum number of genes detected per cell (nFeatures ≤ 200) or more than a maximum percent of reads coming from mitochondrial genes ( ≥ 20%) were filtered out, as summarized in Table 1. These cutoffs were largely consistent with those chosen by the individual studies’ authors. We retained all cells for Puram, Quah, and Choi, which were already prefiltered.

Table 1.

Processing Steps for Pre-Integration

Dataset	Puram	Kurten	Cillo	Peng	Quah	Choi
Individual Dataset Pre-processing
Filtering	Published data prefiltered	Filtered	Filtered	Filtered	Published data prefiltered	Published data prefiltered
nGenes/Cell	2000, or average housekeeping expression level <2.5	200	200	200	200	200
Percent mito	NA	20%	20%	20%	20%	10%

To annotate cells, each dataset was normalized by library size and log transformed using Seurat’s NormalizeData function, except for Puram, which was already normalized. Data was clustered using the Louvain algorithm with 2000 variable features and 30 PCs at a resolution of 0.3. Cells were annotated at a single-cell level using SingleR v2.6.0²⁹ (RRID:SCR_023120) with both the Blueprint^30–32 (RRID:SCR_003844) and Human Primary Cell Atlas (HPCAD)^32,33 databases, accessed through the celldex v1.13.3 package³². Each cluster was assigned a broad cell label (Epithelial, Endothelial, Fibroblast, T/NK, Immune, B Cell, or Mast), based on the consensus majority label amongst Blueprint and HPCAD, and when available, the paper-provided annotations.

Integration

Integration was performed via Seurat’s reciprocal principal component analysis (rPCA) method³⁴. For each pre-processed dataset, SCTransform was used to normalize the raw counts using method = “glmGamPoi”, except for Puram, in which the already normalized counts were used. SelectIntegrationFeatures was used to identify 3000 integration features (genes highly variable amongst all datasets). The FindIntegrationAnchors function was used to identify 20 anchors in 30 dimensions. The integration was then performed using IntegrateData, producing a single Seurat object containing two slots; an integrated slot which contains integrated-corrected counts for the 3000 integration features, used only for clustering, and an RNA slot containing the raw counts for all genes. The raw data in the RNA slot was normalized by library size and log transformed using NormalizeData, as well as saved to a new SCT slot in which it was normalized using SCTransform, with vars.to.regress = ‘dataset’ (again, the pre-normalized Puram data was copied into the SCT slot and not renormalized). After integration, dimensionality reduction and clustering were performed on the integrated slot using 30 PCA dimensions, using the Louvain algorithm, at a resolution of 0.4. scDblFinder v1.21.2 (RRID:SCR_022700) was used to identify doublets using the random based method and dbr.ds = 1. No cluster contained more than 20% doublets.

Signature scoring

Transcriptional signatures were collected from multiple sources (Table 2 Supplementary Data 2). Select genesets from the KEGG³⁵ (RRID:SCR_012773), Hallmarks³⁶, and Reactome³⁷ (RRID:SCR_003485) compendia were also used. Signatures of HNSCC tumor core versus leading edge were compiled from Arora et al.³⁸.

Table 2.

Signatures used in atlas

Publication	Derivation	Signature Type	nSignatures
Szabo et al.	Pan-cancer	T cell	7
Chu et al.	Pan-cancer	T cell	9
Luen et al.	Pan-cancer	T cell	4
Rose et al.	Human blood	Memory T cell	2
Mulder et al.	Cross-tissue	Myeloid	7
Coulton et al.	Pan-cancer	Myeloid	1
Luoma et al.	HNSCC	Myeloid	1
Wei et al.	Pan-cancer	Myeloid	2
Coulton et al.	Pan-cancer	Myeloid	1
Cheng et al.	Pan-cancer	Myeloid	15
Ma et al.	Pan-cancer	Fibroblast	1
Cho et al.	Pan-cancer	Fibroblast	11
Cords et al. Review	Pan-cancer	Fibroblast	5
Cords et al.	BrCa	Fibroblast	2
Puram et al.	HPV + , HPV- HNSCCs	Epithelial states,TME,Cell Types	9,43,22
Arora et al.	Spatial HNSCC	Epithelial	3
Kegg		General
Hallmarks		Tumor
Reactome		General

For all gene expression analyses, the normalized data slot of the RNA assay was used. In each compartment, cells were scored for a given signature using Seurat’s AddModuleScore, whereby the average expression level of the genes in the signature is computed at the single cell level, subtracted by the aggregated expression of randomly selected control feature³⁹. Thus, taking as input the gene-by-cell normalized expression matrix and a list of signatures, signature scoring outputs a signature-by-cell matrix S, with entry s_ij representing the signature score for a signature i in cell j.

Analysis of variance (ANOVA) was performed on the signature scores across T cell clusters to test for significant differences between clusters. Post-hoc pairwise comparisons were conducted using Tukey’s Honest Significant Difference (TukeyHSD) test using the stats v4.4.0 (RRID:SCR_025678) package to identify differences between specific clusters.

Creation of effector cell score and correlation with stage

All T cells were scored using the dysfunctional and cytotoxic signatures from Luen et al.,⁴⁰ using the same signature scoring procedure described above. The dysfunctional state is characterized both by increased expression of immune inhibitory receptors such as TIGIT and LAG3 and loss of cytotoxic genes such as IFNG and IL2⁴¹. It’s been shown that T cells can express genes that are characteristic of both the cytotoxic and dysfunctional states simultaneously, suggesting a continuous transition between these states as opposed to a binary switch. Because of this, we created an additional score to infer the dominant state of the T cells that we termed ‘effector capacity score’ by subtracting the dysfunctional score from the cytotoxic score for each cell, so that the higher the score, the more purely cytotoxic the cell is assumed to be. CD8+T cells were selected by taking clusters 0, 1, 8, 15, and 23 from the immune compartment clustered at resolution 0.8. The relationship between the effector capacity score and stage was determined using a mixed effect model using lme4⁴² (RRID:SCR_015654), with dataset and patient as nested random effects across the CD8+T cells.

$$EffectorCapacityScore~Stage+\left(1\mid dataset\right)+\left(1\mid dataset:patient\right)$$

The association of effector capacity score with stage was also performed separately within each dataset using the same model, but with just patient as a random effect.

$$EffectorCapacityScore~Stage+\left(1\mid patient\right)$$

Identification of marker genes

Identification of marker genes per cluster was performed using Seurat’s FindAllMarkers with the parameters test.use = “MAST” (RRID:SCR_016340), and latent.vars = “dataset” to control for dataset as a fixed effect⁴³.

Gene set enrichment analyses

Gene set enrichment analyses were performed by hypergeometric testing with the KEGG³⁵ (RRID:SCR_012773), Hallmarks³⁶, Wikipathway⁴⁴ (RRID:SCR_002134), and Biocarta⁴⁵ (RRID:SCR_006917) compendia using hypeR v2.4.0⁴⁶, with all genes in the atlas used as the background.

Taxonomic organization of cell type clusters

K2Taxonomer v1.0.7⁴⁷ was used to elucidate the relationship between different clusters and define transcriptional programs enriched at data-driven cluster subgroups. K2Taxonomer was applied separately to the immune and the epithelial compartments, using the scaled, integrated slot of the dataset. The preprocessing step was performed using featMetric = “F”, nBoots = 400, and clustFunc=cKmeansWrapperSubsample. The K2 dendrogram was produced using the main K2tax function with default settings. The normalized RNA slot data was used to perform downstream differential expression and gene set enrichment analyses, with the dataset source included as a covariate.

Correlation of cell type proportions across patients

The correlations of cell type proportions were calculated on the immune (after removal of PBL cells) and non-immune compartments separately to identify clusters manifesting coordinated proportional changes across patients. Per patient cluster proportions were estimated as the number of cells in a cluster divided by the total number of cells in the given compartment. Spearman correlation was calculated among all clusters across all patients. FDR corrections were applied to the lower-triangle of the correlation matrix. Correlations with a FDR < 0.05 and absolute value of the coefficient greater than 0.3 were considered significant.

Inference of copy number variations

InferCNV v1.20.0⁴⁸ (RRID:SCR_021140) was used to detect copy number variations (CNVs) in the epithelial cells, and was run on each dataset separately. Epithelial cells were considered cancer cells and used as ‘observations’. A set of 1000 cells randomly sampled from T cells, B cells, and vascular endothelial cells were used as reference cells. As a negative control, a subset of 100 of the 1000 reference cells was removed and added to the observation set of epithelial cells in which CNVs should not be found. The gene position file hg38_gencode_v27.txt was used for gene locations. Raw counts were used as input, and inferCNV was run using cutoff = 0.1, cluster_by_groups = TRUE, denoise = TRUE and HMM = FALSE.

The output of InferCNV is a gene-by-cell matrix containing CNV scores centered at one, where a value less than one corresponds to a loss in copy number and a number above one corresponds to a copy number gain. A malignancy score for each cell was calculated by subtracting one from the copy number score to center at zero, taking the absolute value of the sum of scores per cell, and dividing by the number of genes.

$$MalignancyScore=\frac{\sum ∣CNVscore-1∣}{nGenes}$$

To identify if epithelial and salivary cells were cancer cells or not, we calculated the top 10% of the malignancy scores of the normal control cells for each dataset. All epithelial and salivary cells that lie within or above this top 10% were defined as cancer cells, and all those below were defined as normal cells.

Inference of cell potency

CytoTrace2 v1.0.0⁴⁹ was run separately on each dataset, and on the immune and non-immune compartments separately. The counts slot of the RNA assay was used as input. The final potency annotation was added to the atlas for visualization across datasets.

Detection of cell type proportion changes across sex

Sccomp v1.99.18⁵⁰ was used to detect changes in cell type proportions between males and females. Sccomp was applied to the immune and non-immune compartments of the atlas separately to avoid artefacts arising from CD45+ and CD45- cell being sorted and sequenced separately which was the case in many of the individual datasets. Sccomp was first applied to each compartment using all datasets in the atlas. Composition and variability models were both included, where dataset was included as a random effect in both cases, and with bimodal_mean_variability_association = TRUE. Sccomp was then applied to the immune and non-immune compartments on each dataset separately using the same parameters, but without including dataset as a random effect.

Cell cell communication analyses

CellChat v2.1.2⁵¹ (RRID:SCR_021946) was used to perform cell cell communication analyses on the RNA slot between clusters as specified in Supplementary Fig 7a, b. The communication probability pathway was computed using the “truncatedMean” method with a cutoff of 0.15, requiring a minimum of 10 cells. A second CellChat analysis focused on signaling between the specific myeloid clusters was performed which used the clustering specified in Supplementary Fig 7a, b, but in which the macrophage and mac/mono/neutrophil clusters were resolved at a higher resolution as detailed as the six clusters using the “trimean” method. Lastly, a CellChat analysis focused on signaling occurring between the fibroblast populations was performed using the clustering shown in Supplementary Fig. 7a, b, but in which clusters Fib1, Fib2, and Fib3 were resolved at the higher resolution of 0.7, again using the “trimean” method.

Statistics and reproducibility

95% confidence intervals were used for all linear regression models. FDR-corrected hurdle p-values were used for all differential expression testing through MAST (RRID:SCR_016340).

Analysis of variance (ANOVA) was performed on the signature scores across T cell clusters to test for significant differences between clusters. Post-hoc pairwise comparisons were conducted using Tukey’s Honest Significant Difference (TukeyHSD) test to identify differences between specific clusters.

GLMs were used to model the effector score with respect to stage. FDR-corrected p-values < 0.05 were used to identify significant associations.

For cell type proportion correlation across patients, spearman correlations were used. P-values from the AS 89 algorithm were FDR-corrected, and significant correlations were defined as those with FDR < 0.05 and the absolute value of the coefficient > 0.3.

Cell-cell communication analyses used empirical p-values in dotplots showing specific ligand-receptor interactions between cell types. P-values are calculated through the CellChat package as the fraction of (cell type label) permuted interaction scores that are as large or larger than the original score.

FDR q-values are used for all differential cell type abundance analyses calculated with the sccomp package.

Results

Generation of an integrated atlas

Six publicly available datasets of human HPV-negative HNSCC tumors were integrated, encompassing 54 patients and a total of 232,015 cells (Table 3). Aggregated patient metadata is provided in Supplementary Data 1. All patients were treatment-naïve, and cells were derived from primary tumors and, in some cases, from local lymph node metastases. The integrated UMAP demonstrated both effective dataset mixing and preservation of cell-type classification: cells assigned to a given cell type prior to integration clustered together after integration (Fig. 1a, b). Clusters comprised cells across sexes, anatomical sites, ages, disease stages, and datasets (Supplementary Fig 1a, b). Despite known biological differences between anatomical sites⁵², cells from the larynx and pharynx were well-mixed with those from the oral cavity, which contributed the majority of cells (Supplementary Fig 1a). Canonical marker genes for broad cell types showed specificity in their expression patterns (Fig. 1c). The datasets varied substantially in the number of cells recovered from each patient, reflecting both biological and technical factors. For example, the Puram dataset, generated with Smart-seq2, yielded the fewest cells while the other datasets employed 10X Genomic platforms. The distribution of recovered cell types also differed between patients, likely due to tumor microenvironment heterogeneity as well as dataset-specific sampling strategies (e.g., Cillo and Kurten also profiled peripheral blood lymphocytes). (Fig. 1d and Supplementary Fig 1b).

Table 3.

Summary of datasets used in atlas

Dataset	GEO	# Patients HPV(-)	# Cells	Sorting	Age Range	Sex (#F/#M)	LNM (primary/LNM)	T Stage (1/2/3/4)
Puram	GSE103322	10	4199		41–88	4/6	5/5	1/3/2/4
Kurten	GSE164690	12	80523		35-80–66	5/7	12/0	1/2/7/2
Cillo	GSE139324	6	20692	CD45+ only	43–66	1/5	6/0	1/1/1/3
Peng	GSE172577	6	57984		32–72	2/4	6/0	0/2/2/2
Quah	GSE225331	7	53459		18–70	2/5	0/7	0/0/4/3
Choi	GSE181919	13	15158		53–77	4/9	12/1	2/8/0/3
Total		54	232015		18–88	18/36	49/13	5/15/16/18

Fig. 1. Creation of integrated HPV-negative scRNAseq atlas reveals inter- and intra-tumor heterogeneity.

a Merged but nonintegrated UMAP (left) and rPCA integrated UMAP (right) colored by dataset. b Integrated UMAP colored by cell type. c DotPlot showing scaled average expression of canonical marker genes by cell type. d Bar plots showing distribution of cell types per patient (bottom), and total number of cells per patient (top).

Separation of immune and non-immune compartments

Because some datasets were sorted prior to integration (e.g., Cillo et al. contained only CD45+ cells), we divided the atlas into the immune and non-immune compartments for downstream analyses. To this end, we first performed global clustering on the entire integrated dataset at a resolution of 0.4 and identified major cell populations within. The non-immune compartment was defined as the union of cells classified as adipocytes, endothelial cells, epithelial cells, fibroblasts, and muscle cells. The immune compartment was defined as the union of cells classified as T cells, NK cells, B cells, Myeloid (labeled broadly as “immune”), and Mast cells. The resulting immune and non-immune subsets were reclustered at resolutions of 0.8 and 0.2, respectively (Supplementary Fig 2).

Immune compartment

Within the immune compartment we identified 30 clusters which we classified into five broad cell types: myeloid, T/NK cells, B cells, Plasma cells, and Mast cells (Fig. 2a, b). This classification was based on a consensus annotation that integrated singleR classification with canonical marker gene expression. Expression of CD14 and CD16 (FCGR3A) separated monocytes from macrophages within the myeloid group, while CD4, CD8A, and CD160 expression clearly separated CD4 + T cells, CD8 + T cells and NK cells within the lymphoid cell cluster (Fig. 2b). We further identified plasma cells (IGHG1-positive), B cells (CD20/MS4A1-positive), and Mast cells (TSAB1-positive). Enrichment analysis of marker genes and signatures further resolved clusters corresponding to cDCs (clusters 14, 25, 26), pDCs, (19), macrophages (5, 7), and monocytes (2, 17, 24) (Supplementary Fig 3a, b).

Fig. 2. Characterization of cell types in the immune microenvironment.

a Immune compartment UMAP clustered at resolution 0.8, across 53 patients. b Identification of broad cell types by expression of marker genes. FCGR3A and CD14 expression on myeloid cells (left), CD160, CD8A and CD4 expression on T and NK cells (center), and IGHG1, TPSAB1 and MS4A1 expression on B cells, plasma cells, and mast cells (right). Cells highlighted are those expressing more than the average gene expression across the immune compartment for the given gene. c UMAP of immune compartment with cluster names based on signature enrichment. d UMAP of immune compartment colored by CytoTRACE2 potency classifications. e K2 Taxonomer dendrogram of immune compartment clusters at resolution 0.8. f Heatmap of correlation of immune cell type proportions across patients. Spearman correlation was calculated on 16 sample-level immune subcluster frequencies. Positive co-occurrence patterns are shown in yellow to red colors and negative cooccurrence patterns are shown in blue colors, with color intensity proportional to the spearman correlation coefficient. Asterisks indicate statistical significance based on FDR-adjusted two-sided p values P < 0.05 from the AS 89 algorithm, and |R | ≥ 0.3. n Patients used is 53.

Annotation of the T cell compartment and association of cellular programs with clinical phenotypes

To resolve cell subtypes higher resolution within the broad T/NK cell group, we scored T cells using published signatures (Supplementary Data 2) and evaluated their average expression at the cluster level, as described in Methods (Supplementary Fig 3c). This analysis identified clusters corresponding to naïve T cells, T follicular helper cells (Tfh), regulatory T cells (Tregs), and central memory T cells (Tcm) within the CD4+ compartment, and cytotoxic, central memory, and dysfunctional cells within the CD8+ compartment, alongside populations of NK and proliferating T cells (Fig. 2a–c and Supplementary Fig 3d).

Naïve T cells were represented by clusters 4, 9, 10, 20 and 21; Tregs by cluster 6; and Tfh cells by cluster 12, each expressing CD4 (Supplementary Fig 3c). Across published studies, the naïve population has been referred to by various terms – such as naïve (Choi²³), CD4+ (Kurten²⁵), naïve-like (Quah²⁷), conventional T-helper (Puram¹⁵, Cillo²⁴), and TCF7+ cells (Peng²⁶) – despite expressing a consistent set of genes (IL7R, LTB, CCR7, SELL, TCF7).

Natural Killer (NK) and CD8 + T cells fall on a trajectory from cytotoxicity to dysfunction

Next, we focused on CD8+ and NK cells. Clusters 13 and 16 showed significantly higher expression of the NK cell signature compared to all other clusters (Supplementary Fig 3c and Supplementary Data 3). Clusters 0, 1, 8, and 23 were identified as CD8 + T cell clusters. Each cell was further annotated by dysfunctional and cytotoxic scores as previously described⁵³. CD8 + T cells mediate cytotoxic killing of tumor cells but can transition to a dysfunctional state following prolonged stimulation by tumor antigens. This state is characterized by reduced cytokine production and effector killer function, alongside upregulation of immune checkpoint markers such as LAG3, TIGIT, CTLA4, and PD-L1⁵⁴. To model this transition, we calculated an overall effector capacity score by subtracting the dysfunctional score from the cytotoxic score, with higher values indicating greater cytotoxicity and reduced dysfunction. NK cells (clusters 13, 16) exhibited the highest cytotoxic score, consistent with their known cytotoxic role, followed by CD8 + T cell clusters 0, 8, and 23. Notably, CD8 + T cell cluster 0 displayed the highest dysfunctional score, reflecting a multi-state phenotype characterized by expression of both cytotoxicity and dysfunction programs.

The balance of T cell cytotoxicity and dysfunction is associated with tumor stage

We investigated the relationship between the T cell effector capacity score (cytotoxic minus dysfunctional) and tumor stage. After isolating the CD8 + T cells, we applied a linear mixed effects model to regress the effector capacity score on stage. We found a significant positive association with stage T1 (t = 0.318, FDR = 0.018), suggesting a cytotoxic T cell response at the earliest disease stage. (Supplementary Fig 3e, Supplementary Data 4). When performing the same analysis within the individual datasets, a significant association with stage T1 is not achieved. Within the Peng dataset there was an almost significant (t = 0.326, FDR = 0.06) association between the effector capacity score and stage T2, which was not observed in other datasets. This highlights the utility of the large sample size in the atlas to either boost or temper confidence in identifying associations between cellular programs and clinical phenotypes.

Contextualization of immune cell populations

We next sought to elucidate the relationship between cell-type annotated clusters within the immune compartment (Fig. 2c).

CytoTRACE2 is an interpretable AI method for predicting cellular differentiation potential from scRNA-seq data originally developed in a developmental biology context. We applied CytoTRACE2 to the immune compartment and found that most immune cells were in a fully differentiated state (Fig. 2d). In contrast, naïve T cells and proliferating T cells were classified as unipotent or oligopotent, reflecting their ability to differentiate into various T cell subtypes, such as effector and memory cells, upon antigen stimulation⁴⁹. Interestingly, some Treg cells were also annotated as unipotent. This finding is consistent with previous studies showing that Tregs exhibit high plasticity and may switch between states in response to different cues^55,56.

Application of K2Taxonomer (K2T) to the immune cell compartment revealed the taxonomic relationships between the clusters therein⁴⁷ (Fig. 2e). The top-level partition broadly separated myeloid and lymphoid subsets, with the exception of mast cells, which clustered with lymphoid cells, likely due to their lack of expression of some myeloid markers, such as LYZ and LST1. The four monocyte clusters (2, 7, 17, 24, and 29) formed a hierarchical group. Cluster 5 (macrophages) was most closely related to cDC clusters 14, 22, and 26, due to elevated expression of complement genes shared by macrophages and cDCs relative to other immune cells (Supplementary Fig 4a,b). Within the lymphoid lineage, T cells and NK cells grouped hierarchically and further segregated by cell states: naïve (4, 9, 10, 20, 21), cytotoxic (8, 13, 16, 23), and dysfunctional (0).

Concurrently, we evaluated the correlation of immune cell type proportions across patients to identify populations that co-vary, which may suggest cellular interactions, although it does not provide direct evidence. (Fig. 2f, Supplementary Data 5) Naïve T cells were highly correlated with cytotoxic T/NK (R = 0.67, FDR < 0.05) and CD14+ monocytes (R = 0.58, FDR < 0.05), and both populations have previously been shown to be associated with better prognosis^57,58. Tregs, Tfh, and dysfunctional T cells were highly correlated with each other (R ≥ 0.6 for each pairwise correlation, FDR < 0.05) and have been linked to worse survival outcomes^58–60. They were also positively correlated with CD8 + T cells (R > 0.4 for each pairwise correlation, FDR < 0.05). Naïve T cells showed a positive, albeit marginally nonsignificant, correlated with B cells, (R = 0.34, FDR = 0.063), consistent with findings from Peng et al., where TCF7+ naïve T cells localized in and around tertiary lymphoid structures (TLSs) composed of CD20 + B cells²⁶.

Macrophage annotation resolves subtypes

Recent studies have revealed new polarization states of macrophages beyond the M1/M2 axis^61–64. Leveraging the large sample size of our atlas, we first recapitulated these states and then further characterized macrophages at high resolution. Clusters 5 and 7 were identified as putative macrophages, cluster 5 by its expression of C1QB and macrophage gene signatures, and cluster 7 by proximity to cluster 5 in the UMAP and expression of both macrophage and monocyte signatures (Supplementary Fig 4b). These macrophage populations were reclustered at resolution 0.3, resulting in six subpopulations (Fig. 3a).

Fig. 3. Identification of myeloid subclusters and signaling pathways.

a UMAP of immune compartment clusters 5 and 7 (from Fig. 2a) reclustered at resolution 0.3. Six clusters were identified. N patients = 49. b UMAPs of marker gene expression for the six clusters, CXCL9, SPP1, CXCL10, IL1β, HCAR3, and S100A8. c Heatmap of top 10 marker genes per cluster. Expression is scaled row-wise. d KEGG signature enrichment for macrophage subclusters based on signature genes using hypergeometric test with FDR-corrected significance values. e Cell-cell communication analysis by CellChat showing pooled signaling probability across pathways. Signals shown are sent from the CXCL9, SPP1, and IL1β, clusters (outgoing) to the other clusters (incoming). Bar plots on the top and right of the heatmap show total signaling per cluster.

Interrogation of the top genes enriched in each cluster revealed two clusters marked by high expression of canonical macrophage markers C1QB, CD4, and CD163 (Fig. 3b, c and Supplementary Data 6). One cluster was defined by CXCL9 expression and enriched for antigen presentation-related genes (Fig. 3b–d). The other cluster expressed high levels of SPP1 and metalloproteinases (MMPs), but showed a relative paucity of KEGG or Hallmark pathway enrichments (Fig. 3c, d). Notably, these macrophage clusters were distinguished more by CXCL9 and SPP1 than by canonical M1/M2 markers (Supplementary Fig 4c), supporting an alternative macrophage categorization by polarization states in HNSCC recently described by Bill et al.⁶¹.

The remaining four clusters lacked expression of canonical macrophage markers C1QB, CD4, APOE and CD163 (Supplementary Fig 4a). Among these, one cluster was CXCL10-positive, also overexpressing interferon induced proteins with tetratricopeptide repeats (IFIT) genes and CCL2, and was enriched for interferon alpha (IFN-α) and gamma (IFN-γ) response pathways (Fig. 3c, d). This cluster is similar to a population described by Zhang et al. in inflammatory conditions, expressing genes such as IDO1 and GBP1, and stimulated by IFN-γ and TNF-α⁶⁵.

Another distinct cluster exhibited features consisted with myeloid-derived suppressor cells (MDSCs)⁶⁶, characterized by high expression of inflammatory cytokines including IL1B, but lacking macrophage markers SPP1 and CXCL9. This population expressed inflammatory genes such as the aforementioned IL1B, known to promote drug resistance through induction of ICAM1 on tumor cells⁶⁷, as well as IL6 and IL10, cytokines implicated in Treg-mediated immunosuppression⁶⁸. It was also enriched for the cytokine-cytokine receptor interaction KEGG pathway (Fig. 3d). This cluster shares partial transcriptional similarity with previously described HNSCC tumor-associated macrophages (TAMs) classified as M2-like macrophages⁶⁹. However, it lacks expression of several canonical M2 markers such as a CD163 and complement genes (Supplementary Fig. 4a)⁷⁰.

To contextualize the IL1B+ population within the broader myeloid landscape, we compared it with signatures from pan-cancer macrophage datasets (Supplementary Data 2). This cluster showed transcriptional similarity to VCAN- and EREG-expressing TAMs associated with angiogenesis identified in a pan-cancer macrophage atlas (AngioMac)⁶³, a pan-tissue IL1B-expressing monocyte population⁷¹, a pan-cancer CXCL3 + TAM cluster⁶⁴, and a HNSCC CXCL8 + TAM cluster⁷² (Supplemental Fig. 5d). While nomenclature referencing this cluster varies, sometimes referring to these cells as macrophages or monocytes, this population emergers consistently across datasets. While studies indicate comparable levels of macrophages (identified using CD68 as a marker) between HPV-positive and -negative patients⁷³, elevated levels correlate with worse prognosis, suggesting this population may play important roles in both contexts.

Given the IL1B+ cluster’s enrichment for cytokine and chemokine signaling pathways, we examined signaling pathway activity unique to this cluster relative to other myeloid subtypes (Fig. 3d, e). Thrombospondin (THBS) signaling was specific to the IL1B cluster, and tumor necrosis factor (TNF) and IL−1β signaling pathways were elevated compared to the SPP1+ and CXCL9+ clusters (Fig. 3e). Although IL-1β and TNF-α secretion by SPP1-positive macrophages has been linked to HNSCC progression⁷⁴, our atlas highlights the IL1B + /SPP1–population as a distinct subpopulation associated with unique signaling pathways.

Finally, we identified a cluster of Neutrophils characterized by low expression of HLA-DRA, HLA-DRB1, and CD74 expression, and positive expression of IFITM2, S100A8, PROK2, HCAR3, CMTM2, and CSF3R, among others (Fig. 3c)^75,76. Gene set enrichment analysis confirmed enrichment for cytokine-cytokine receptor interactions in this cluster (Fig. 1d), and cell-type classification using HPCAD and Blueprint databases corroborated these findings (Supplementary Fig 4d, Supplementary Data 7). The myeloid subcluster annotations are shown in the context of the larger immune compartment in Supplementary Fig 4e, and cluster-specific gene signatures are provided in Supplementary Data 6.

Non-immune compartment

Clustering analysis of the non-immune compartment based on Seurat at resolution 0.2 identified 16 clusters. These included distinct clusters of vascular and lymphatic endothelial cells, three fibroblast clusters, seven tumor epithelial cell clusters, and four salivary cell clusters, each defined by canonical marker genes and cell type signatures (Fig. 4a, b).

Fig. 4. Non-Immune compartment characterization and contextualization.

a UMAP clustered at resolution 0.2 for a total of 16 clusters across 47 patients. b DotPlot showing z-scaled average expression of marker genes (black labels) and z-scaled average signatures scores of cell type signatures (green labels) across all non-immune cells. c Same as in (b), but across the seven epithelial clusters. d K2Taxonomer organization of clusters. e Heatmap of spearman correlations of cell type proportions across patients. Asterisks indicate statistical significance based on FDR-adjusted two-sided p values P < 0.05 from the AS 89 algorithm and |R | > 0.3. N Patients = 47. f Heatmap of total number of interactions being sent from and received by each cell type from the cell-cell communication analysis by CellChat. g Heatmap showing signaling pathways (rows) being sent from the seven epithelial clusters to themselves and the rest of the clusters in the TME. The color intensity represents the pooled signaling strength per pathway across. The bar plots represent total signaling per cell type (top), and per pathway (right). All pathways with cumulative signaling probabilities above 0.01 are shown. h Same as in (g), but looking at incoming signals being received by the epithelial clusters.

Epithelial compartment characterization and tumor cell identification

We used InferCNV to detect copy number variations (CNVs) in epithelial and salivary cells, as described in Methods. An aggregated malignancy score was computed per cell, defined as the total CNV burden in a cell regardless of direction (amplification or loss), with higher scores indicating more alterations. Within each dataset, epithelial and salivary cells exceeding the 90th quantile of malignancy scores derived from “normal” cells were designated cancer cells; others were classified as normal (Supplementary Fig. 5a, b, and Supplementary Data 8). Malignancy scores varied significantly by cluster, with patient-specific clusters Epi2 and Epi3 displaying the highest and lowest scores, respectively, within the epithelial compartment (Supplementary Fig 5c).

Normal epithelial cells segregate on a differentiation continuum

Annotation of the epithelial compartment revealed a gradient of differentiation signatures, with cluster Epi1 representing the least differentiated state and increasing differentiation across Epi6, Epi5, and Epi4 (Fig. 4c). This trajectory reflects the maturation of non-tumor epithelial cells from basal cells to squamous cells, the latter being the tumor’s cells of origin. Epi6 expressed high levels of KRT5 and KRT14, markers of basal cells⁷⁷, indicating its position at the early stage of the differentiation trajectory. Epi5 showed high expression of suprabasal markers KRT13, KRT4, and SPRR3⁷⁸. Epi1 expressed high levels of CD44, a marker of cancer stem cells (CSCs)⁷⁹, and KRT18, associated with progenitor and malignant epithelial cells⁸⁰. The pEMT program was also enriched in Epi1, further suggesting its malignant and plastic phenotype.

To further characterize differentiation status, we applied CytoTRACE2 to the non-immune compartment, which predicts cellular developmental potential on a continuous scale from 0 (fully differentiated) to 1 (totipotent)⁴⁹. While originally developed in a developmental biology context, this tool can be applied to cancer datasets with the ‘potency’ score quantifying a tumor cell’s degree of plasticity, or a normal cell’s differentiation state. Most the tumor cells were predicted to be oligopotent (plastic capability), whereas other tumor microenvironment cell types such as endothelial cells were largely fully differentiated (Supplementary Fig 5d). Within epithelial clusters, Epi1 exhibited the highest CytoTRACE2 potency scores, followed by Epi6, whereas Epi4 and Epi5 had lower scores. This suggests an inverse correlation between potency and non-tumor epithelial cell differentiation, and positive correlation with tumor cell plasticity and pEMT. Interestingly, some cancer-associated fibroblasts (CAFs) were also predicted to be oligopotent (defined as a lineage-restricted potential to differentiate into 2–3 downstream cell types), which may reflect their capacity for reprogramming⁸¹.

Some epithelial clusters are patient-specific

Hierarchical organization of the non-immune compartment using K2T first partitioned cells primarily into stromal and epithelial groups, with epithelial cluster pairs Epi4 and Epi5 grouped together, consistent with their differentiation state relative to the tumor cell of origin (Fig. 4d). Interestingly, Epi7 grouped within the stromal partition. Examination of patient distribution across epithelial clusters revealed that clusters Epi2, Epi3, and Epi5 were largely patient-specific (Supplementary Fig 5e). These clusters showed unique pathway enrichments, such as the pentose phosphate pathway and xenobiotic drug metabolism in Epi2, and translation in Epi3, with Epi7 consistent with a cycling cell population (Fig. 4c). Notably, Epi2 uniquely expressed KRT19, a gene shown to reprogram CSCs in breast cancer toward a more drug-sensitive state⁸².

More plastic cells are more likely to exist on the tumor edge and interact with the TME

Correlation analysis of cell type proportions across patients showed that cluster Epi1 correlated more strongly with the stromal clusters than with other epithelial clusters, although this did not reach significance (Fig. 4e). To test the hypothesis that Epi1 cells exhibit increased interactions with the stromal cells and distinct spatially localization, we leveraged an external spatial transcriptomics dataset, Arora et al.³⁸, which defined tumor leading-edge and core signatures in HNSCC. Projection these signatures onto our epithelial clusters showed that Epi1 was enriched for the leading-edge signature, whereas Epi4, and to a lesser extent Epi5, were enriched for the tumor core signature (Supplementary Fig 5f). These findings align with prior reports locating pEMT-expressing tumor cells at the invasive front, facilitating interaction with CAFs¹⁵. Additionally, Epi1 strongly expressed TGFBI, a secreted ECM protein linked to aggressive cancers^83,84.

Detailed characterization of epithelial cells

Reclustering epithelial cells at higher resolution (0.4) identified 18 distinct clusters (Supplementary Fig 5g). Marker genes were shared across multiple clusters suggesting that the initial lower resolution clustering was sufficient to capture the most salient cell differences (e.g., ACKR1C3 in 0 and 4) (Supplementary Fig 5h). Gene set enrichment analysis using the Hallmarks compendium and signatures from Puram et al.¹⁶ revealed pathways co-enriched in multiple epithelial cell clusters (Supplementary Fig 5i). These included cholesterol homeostasis, known to be important for proper keratinization, and epithelial cell differentiation, both enriched in clusters 7, 8, 9 and 13⁸⁵. Clusters 3, 11, and 17 were enriched for translation, Notch signaling, apoptosis and the p53 pathway, suggesting Notch signaling is driving changes in these clusters, as notch signaling has been shown to interact with the p53 pathway to inhibit apoptosis and drive aberrant translation^86,87. Clusters 10, 12, 15, and 16 showed enrichment for hypoxia, TGF-β and IL6 signaling, consistent with hypoxia-induced TGF-β and IL6 signaling in cancer^88,89. Oxidative stress and the NRF2 pathway, which are induced by cigarette smoke⁹⁰, were enriched in clusters 0, 4, and 10, implicating these clusters as ones which may be more prevalent in HPV-negative patients.

Epithelial cell signaling

To examine the global patterns of interactions among clusters, we performed cell-cell communication analysis and looked at the total number of interactions between clusters in a pairwise manner (Fig. 4f). Epi1, Epi2, Fib1, Fib3 and vascular endothelial clusters exhibited the highest number of estimated outgoing signals, whereas epithelial cells generally received the most incoming signals. We next characterized signaling pathways exchanged between epithelial clusters and the TME (Fig. 4g, h). Epi1 and Epi2 were unique in sending extracellular matrix (ECM) and tissue physiology signals, including elevated laminin, fibronectin (FN1, primarily from Epi1), collagen (mainly from Epi1), and thrombospondin (THBS, exclusive to Epi1), compared to the other Epithelial clusters (Fig. 4g). Incoming signals to epithelial clusters featured collagen and laminin pathways predominantly originating from the fibroblasts, with increased strength to Epi1 and Epi2 (Fig. 4h). Unique incoming signals to Epi1 and Epi2 included angiopoietin (Angpt) signaling from endothelial cells, and visfatin. Notably, only Epi1 received TGF-β signals, consistent with the EMT pathway enrichment in Epi1 subclusters (1, 5, 15, 14) (Supplementary Fig. 5i), and aligning with TFGβ’s known role in EMT induction and poor prognosis in HNSCC^84,91.

Mesenchymal cell characterization and nomenclature harmonization

Next, we examined mesenchymal cells, a heterogeneous population that includes fibroblasts, pericytes and cancer associated fibroblasts (CAFs), which play critical roles in tumor progression^15,92–94.

Clustering of the non-immune compartment at a resolution of 0.7 identified seven mesenchymal populations (Fig. 5a). Expression of canonical CAF markers (FAP, PDPN, CTGF) was observed primarily in clusters 0, 15, 16, suggesting these are CAFs as opposed to resting fibroblasts, while myofibroblasts markers (ACTA2 and MYL9) marked clusters 10 and 14. (Fig. 5b)¹⁵. We then assessed expression of CAF signatures gathered from pan-cancer and HNSCC-specific studies (Table 4 and Supplementary Data 2) to classify subpopulations (Fig. 5c and Supplementary Data 9). Clusters 0 and 15 expressed classic fibroblast¹⁶ and matrix-associated CAF (mCAF) signatures. Clusters 10 and 14 expressed myofibroblast and vessel-associated CAF (vCAF) signatures, with cluster 14 uniquely expressing vCAF markers BCAM and MYH11 (Fig. 5c, Supplementary Fig 6a). Cluster 16 expressed the inflammatory CAF (iCAF) signature. Notably, while cluster 0 expressed chemokines such as CXCL1, CXCL8, and IL11, cluster 16 uniquely expressed canonical iCAF genes CXCL12 and CXCL14, suggesting that the iCAF signature^95,96 may encompass at least two distinct subtypes in HNSCC. Some genes such as CCL2, CXCL2 and IL6 were shared between both clusters (Supplementary Fig. 6a). Interestingly, cluster 0 also expressed the metabolic CAF (meCAF) signature (Fig. 5c), first identified in pancreatic ductal adenocarcinoma⁹⁵. Although antigen-presenting CAFs (apCAFs) are common in several cancer types^93,97, there is a lack of a strong localized expression of the corresponding signature in the atlas (Supplementary Fig. 6a). Cluster 22 likely represents myocyte muscle cells due to their expression of myosin (MYLPF) (Fig. 5d). Cluster 32 expressed both fibroblast markers (COL1A2, PDGFRB) and vascular endothelial marker PECAM1, suggesting this may be a cluster of doublets.

Fig. 5. Fibroblast classification and signaling.

a UMAP clustered at resolution 0.7 for a total of 34 clusters. n Patients = 47. b Violin plots showing normalized expressed values for FAP, PDPN, CTGF, ACTA2, MYLK, and MYL9 across the fibroblast clusters. n Patients with cells in c0:n = 45; c15:n = 36; c16:n = 40; c10:n = 38; c22:n = 34; c14:n = 38; c32:n = 20. n cells per fibroblast cluster: c0:n = 8514; c15:n = 2588; c16:n = 2575; c10:n = 3110; c22:n = 1621; c14:n = 2596; c32:n = 119. c Boxplots showing the average signature scores per fibroblast cluster. Boxes represent median ± interquartile range; whiskers represent 1.5× interquartile range. d Heatmap of top five DE genes per fibroblast cluster. Values shown are z-scaled normalized expression values. e Cell-cell communication analysis by CellChat showing pathways involved in sending signals from fibroblast clusters 0, 15, and 16 (outgoing) to the vascular and lymphatic endothelial cells (incoming). Color intensity represents the sum of the signaling strength of all interactions in each pathway. Bar plots represent total signaling strength per cluster (top), and per pathway (right). f Specific ligand and receptor pairs involved in CXCL pathway interactions occurring between fibroblast clusters 0, 15, and 16 and the vascular endothelial cells. P-values shown represent empirical p values. g Expression of CXCL8 and ACKR1 in non-immune compartment.

Table 4.

Fibroblast nomenclature harmonization

Cluster	0	15	16	14	22	10	32
Dataset
Atlas	Imm-CAF (CXCL8)	ECM-CAF	Imm-CAF (CXCL12)	vCAF	Myocytes	Myofibroblast-ECM	Fib-Endo Doublets
Puram	CAF1	CAF1	CAF2	Myofibroblasts	NA	CAF1	Myofibroblasts
Choi	cf4	cf4	cf2,cf4	cf3	Myocytes	cf3	cf3,cf4
Kurten	CAF	CAF	Elastic	Pericytes	Pericytes	Pericytes	Pericytes
Quah	CAF	CAF	CAF	Myofibroblast	Myofibroblast	Myofibroblast	CAF,Myofibroblast
Peng	No Fib analysis
Cillo	No CD45- cells

To contextualize atlas fibroblast clusters with CAF populations identified in the originating studies, we scored atlas clusters using the published CAF gene signatures (Supplementary Fig 6b). Cluster 15 was consistently classified as CAFs by most studies^15,23,25,27. while clusters 10, 14, 22 and 32 were classified as myofibroblasts, and were termed either myofibroblasts or pericytes in most studies^15,23,25,27 Cluster 0 and 16, the two immune-related fibroblast populations were variably termed ‘elastic’ or ‘resting’ fibroblasts^15,25. Recent studies identified CXCL8+ inflammatory CAFs¹⁶ and IL11 + CAFs⁹⁸ in HNSCC, prompting our harmonization of nomenclature across studies to define equivalent populations in our atlas (Supplementary Fig 6b, Table 2). Cluster-specific gene signatures were defined using a log2 fold change cutoff > 1 and FDR < 0.05 (Supplementary Data 10).

Cell-cell communication

Due to the high expression of multiple chemokines in CAF clusters 0, 15, and 16, and the observed high correlation between fibroblast cluster 1 and the vascular endothelial cell cluster across patients (R = 0.5, FDR = 0.008) (Fig. 5e), we examined fibroblast-endothelial cell signaling. Fibroblasts sent collagen signals preferentially to lymphatic, but not vascular, endothelial cells (Fig. 5e). Cluster 0 showed increased signaling strength to vascular endothelial cells via the CXCL pathway via enhanced CXCL1-ACKR1 and CXCL8-ACKR1 interactions (Fig. 5f). Interestingly, CXCR1 and CXCR2, the canonical receptors for CXCL1 and CXCL8, were not expressed in any atlas cells, whereas ACKR1, an alternative receptor for these ligands⁹⁹, was exclusively expressed in vascular endothelial cells (Fig. 5g). Similarly, cluster 16 showed increased signaling strength to the vascular endothelial cells through CXCL2-ACKR1 interactions (Fig. 5f).

Changes in cell type proportions with sex

The large sample size of the atlas enabled us to evaluate potential associations between TME heterogeneity and sex, recognizing that the extent and composition of immune cell infiltration in the TME can play an important role in tumor aggressiveness. Differential cell type composition analyses were performed within the immune and non-immune compartments, separately (Supplementary Data 11)⁵⁰. For simplicity, related clusters were merged: naïve, Treg, CD4+ Tcm and Tfh clusters were combined into a single CD4 + T cell cluster; the cytotoxic and dysfunctional CD8 + T cells were merged into one CD8 + T cell cluster; and the four salivary clusters were consolidated (Supplementary Fig 7a, b). Several significant sex-associated changes were observed within the immune compartment. Male patients exhibited increased proportions of macrophages, and CD8+ and proliferating T cells, whereas female patients had higher proportions of Plasma cells, Monocytes and NK cells (Fig. 6a). A pan-cancer study of changes in cell type proportions using deconvolution in TCGA datasets similarly reported increased CD8 + T cells in male patients, alongside multiple immune populations enriched in female patients¹⁰⁰. In the non-immune compartment, the proportions of Epi1 and Epi4 were greater in male patients (Fig. 6b). Equivalent analyses performed within the individual studies largely failed to identify significant sex-associated changes (Supplementary Fig 7c,d), underscoring the advantage of the integrated atlas in detecting such associations thanks to the increased statistical power afforded by a large number of patients.

Fig. 6. Sex-specific cell type proportion changes.

a Cell type proportion changes identified by sccomp within the immune compartment. Significance shown is based on FDR-corrected p-values. Female n = 18. Male n = 35. Error bars represent 95% credible intervals. b Cell type proportion changes identified by sccomp within the non-immune compartment. Significance shown is based on FDR-corrected n-values. Female n = 16. Male n = 31. Error bars represent 95% credible intervals.

Discussion

In this study, we integrated six publicly available HNSCC scRNA-seq datasets to establish a comprehensive HPV-negative HNSCC atlas. This resource is thoroughly annotated with patient metadata, cell types, and their corresponding genes signatures. The large cohort size attained enabled statistically powered evaluation of associations between gene expression, cell type composition, and molecular and clinical phenotypes.

Recent work uncovered the SPP1/CXCL9 polarization axis that characterizes distinct macrophage populations⁶¹. However, other myeloid populations remain less well defined in HNSCC. Our atlas recapitulates SPP1+ and CXCL9+ macrophage subsets and further identifies an IL1B+ population variably labeled in prior studies as CXCL3 TAMs, monocytes, CXCL8 TAMs, or AngioMacs, highlighting the ongoing challenge of inconsistent nomenclature across studies. We find that this IL1B+ cluster secretes immunosuppressive cytokines (e.g., IL10) and interacts with other TME cells via IL-1β and TNF signaling pathways, both implicated in HNSCC progression⁷¹, as well as through unique pathways such as thrombospondin signaling. The role of the thrombospondin pathway in HNSCC is complex; it has been reported both to promote invasion¹⁰¹ and inhibit angiogenesis¹⁰², leaving its net effect on tumor progression unclear. However, CD47, a major thrombospondin receptor, is a promising therapeutic target in HNSCC, with anti-CD47 treatment shown to delay tumor growth and stimulate effector T cells¹⁰³. Prior studies have linked macrophage populations to poor prognosis, specifically SPP1+ macrophages in HNSCC⁶¹ and M2 macrophages across cancers, which have also been effectively targeted by therapies⁷⁰. Furthermore, recent evidence showed macrophage-secreted IL-1β to promote docetaxel resistance⁶⁷. Given that these macrophage clusters are major sources of IL1B and immunosuppressive cytokines, their inhibition may improve both drug efficacy and immune response, although further validation is needed to demonstrate their precise role in the HNSCC TME.

The landscape of CAFs has also been increasingly refined, both in pan-cancer contexts and specifically in HNSCC^97,104, and we were able to extend this characterization further within the atlas. We distinguish two separate iCAF populations – CXCL8+ and CXCL12 + iCAFs—that are frequently conflated in the literature^97,105,106. We also identify vCAFs, which have been described in pan-cancer studies but are frequently termed myofibroblasts in HNSCC. Jenkins et. al. recently reported enrichment of both myCAFs and iCAFs (the iCAFs align with the CXCL8 + iCAFs) in HPV-negative HNSCC, while a population termed FRC-like CAFs (which align with the CXCL12 + iCAFs) was more abundant in HPV-positive tumors⁹⁸. Interestingly, their spatial transcriptomics data revealed CXCL12 + iCAFs correlate with CD4 + T cells and B cells and are associated with better survival, whereas CXCL8 + iCAFS correlate with IL1B+ inflammatory monocytes but show no survival association.

We identify a specific interaction between CXCL8 + CAFs and ACKR1 on vascular endothelial cells that, based on current literature, may be associated with worse outcome through promoting angiogenesis, although experimental validation is needed. The CXCL8+ population was previously characterized by Quah et al,²⁷ who demonstrated its induction by galectin-7 (LGALS7) expression in malignant epithelial clusters. A recent spatial transcriptomics study of HNSCC also showed co-expression of CXCL1, CXCL8 and ACKR1³⁸, corroborating our finding that CXCL8 + CAFs and vascular endothelial cell abundances are correlated across patients. Prior studies also showed that CXCL8 and CXCL1 promote the angiogenic potential of endothelial cells and affect their proliferation and survival in-vitro¹⁰⁷, and a similar myCAF population expressing CXCL8 and IL24 was recently linked to metastasis and angiogenesis in esophageal squamous cell carcinoma¹⁰⁸. Both CXCL1 and CXCL8 have been associated with poor patient survival in HNSCC¹⁰⁹. Collectively, these findings suggest that the CXCL8 + iCAF population may contribute to worse outcomes for HNSCC patients by stimulating angiogenesis and could represent an effective target for anti-angiogenic therapies. However, studies in other contexts have reported that ACKR1 may function as a decoy receptor limiting angiogenesis¹¹⁰. Thus, experimental validation is essential to ascertain ACKR1’s role in HNSCC.

A central challenge in HNSCC management is overcoming resistance to treatment. It has been suggested that many cancers, including HNSCCs, maintain a population of CSCs or tumor cells with stem-like properties capable of self-renewal and therapy resistance^{109,111–113}. Our integrated analysis of epithelial and tumor cells across datasets identified shared transcriptional programs across patients, exemplified by the Epi1 cluster, which exhibits stem-like properties, pEMT and EMT signatures, and preferential localization to the tumor edge where it interacts with stromal populations, recapitulating key findings from Puram et al.¹⁵ Cell-cell communication analyses revealed extensive bidirectional signaling involving extracellular matrix components (e.g, laminin, collagen) which influences tumor stiffness, cell adhesion and migration^114,115, with increased signaling directed to Epi1 compared to other epithelial clusters. Uniquely, Epi1 receives TGF-β signals, consistent with its EMT program enrichment. Since TGF-β has been implicated in treatment resistance¹¹⁶, co-tageting TGF-β alongside ECM-related pathways may enhance therapeutic response while suppressing tumor progression. As additional drug resistance signatures emerge¹¹⁷, they can be mapped onto the atlas to identify specific cell types and transcriptional programs underpinning resistance.

Sex-associated differences in cellular composition across cancers have been noted but remain understudied in HNSCC. Leveraging the atlas, we observed that male patients harbor higher proportions of PDL1-expressing CD8 + T cells, which correlates with improved response to immune checkpoint inhibitors (ICIs)¹¹⁸, suggesting potential sex-based differences in ICI efficacy. Within the non-immune compartment, the Epi1 cluster, enriched for both the pEMT signature and proliferating tumor cells, was more prevalent in male patients, consistent with a study demonstrating increased carcinogen-induced proliferation in male mice, leading to more aggressive and metastatic cSCC¹¹⁹.

Limitations of this study include variability in cell types collected and sequenced per patient, as well as heterogeneity in tissue anatomical origin. Although the current cohort enables robust analyses, fully disentangling technical from biological variation will require continued expansion of the atlas with new datasets. Moreover, integration updates require manual reprocessing. Scalable integration strategies such as non-negative matrix factorization¹²⁰ or transfer learning approaches such as scArches¹²¹ may streamline future expansions. Beyond transcriptional profiling, future efforts may aim to integrate spatial transcriptomics and multi-modal datasets capturing paired transcriptomic and epigenetic information to deepen understanding of HNSCC heterogeneity.

Author contributions

S.M and L.K. designed the study. L.K. performed all analyses and wrote the manuscript. A.C. and E.R. provided computational guidance and feedback. A.S, M.K., and X.V. provided biological guidance. S.M. oversaw the project, provided feedback, and guided the research. The manuscript is edited by all authors.

Peer review

Peer review information

Communications Medicine thanks Vanessa Porter, Noah Sasa and the other, anonymous, reviewers for their contribution to the peer review of this work.

Data availability

The following publicly available datasets (raw count.mtx, gene and cell barcode files) were accessed via GEO through the accession numbers: Peng²⁶ (GSE172577), Choi²³ (GSE181919), Cillo²⁴ (GSE139324), Quah²⁷ (GSE225331), and Kurten²⁵ (GSE164690). Due to availability, processed normalized data were downloaded for the last dataset, Puram¹⁵ (GSE103322). The atlas has been deposited to cellxgene for online visualization by the research community through the cellxgene repository, an interactive data explorer, at the following link. The full atlas is also available as a Seurat.rds object for download on zenodo [10.5281/zenodo.17880437]¹²² using the following link. Source data is provided with this publication on Zenodo. Source data for Fig. 1a–d, Fig. 2a–f, Fig. 3a–e, Fig. 4a–h, Fig. 5a–e, g–f, Fig. 6a, b, Supplementary Figs. 1a, b, Supplementary Figs. 3b,c, Supplementary Figs. 4a–f, Supplementary Figs. 5a–f, Supplementary Fig. 6a, Supplementary Figs. 7c, d, are in SourceData.xslx. Source data for Supplementary Fig. 5h is in SourceData Fig5h.csv. Source data for Supplementary Fig. 3e is in SuppData4_Teffstage_data_SFig3e.xlsx.

Code availability

All code is available on github and zenodo.¹²³

Competing interests

The authors declare no competing interests.

Consent to publish

Not applicable. All datasets included are already published.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01401-3.