Spatial analysis of head and neck cancer reveals two ecosystems with distinct modes of epithelial-to-mesenchymal transition

Background: Single-cell RNA sequencing has revealed diverse malignant programs in head and neck squamous cell carcinoma (HNSCC), including partial epithelial-to-mesenchymal transition (p-EMT), a state linked to invasion and poor outcome. However, how p-EMT is spatially organized within tumors and shaped by the tumor microenvironment (TME) remains unclear. Here, we used spatial transcriptomics to define tumor-wide spatial architectures of p-EMT in HPV-positive and HPV-negative HNSCC and to identify their microenvironmental drivers. Methods: We profiled 26 primary HNSCC tumors (18 HPV-negative, 8 HPV-positive) using 10x Visium spatial transcriptomics, with paired single-cell RNA-seq for 11 tumors. Cell states were defined using consensus expression meta-programs, and malignant regions were identified via inferred copy number aberrations. Spatial interactions were quantified using distance-based neighborhood analyses, and malignant regions were stratified into tumor edge and core compartments. Key findings were validated using high-resolution Visium HD on paired sections, 25-plex CODEX spatial proteomics in an independent cohort of 40 laryngeal squamous cell carcinomas, and cytokine perturbation assays in vitro . Results: HPV-positive tumors exhibited increased cellular density, an immune-enriched microenvironment, and malignant programs consistent with hypoxia, alongside relative depletion of p-EMT. In contrast, HPV-negative tumors showed prominent p-EMT that segregated into two distinct spatial architectures. The canonical p-EMT edge pattern localized to the invasive front, where p-EMT malignant cells spatially coupled to fibroblasts through TGFbeta-associated signaling. Notably, we identified an alternative p-EMT core architecture characterized by diffuse p-EMT throughout tumor cores and co-localization with immunosuppressive macrophages and neutrophils. This spatial niche-related transcriptional program was associated with worse overall survival in large external cohorts. Ligand-receptor analyses and in vitro validations implicated oncostatin M (OSM) as a key upstream driver of p-EMT in tumor cores. Both Visium HD and CODEX independently confirmed the spatial organization of p-EMT states and the corresponding TME interactions. Notably, both p-EMT architectures were consistent across multiple samples from the same tumor, indicating tumor-wide organizational states rather than region-specific variability. Conclusions: We identify two tumor-wide p-EMT architectures in HNSCC driven by distinct microenvironmental programs: fibroblast-derived TGFbeta at tumor edges and myeloid-derived OSM in tumor cores. These findings establish OSM as a novel regulator of p-EMT and suggest spatially organized EMT programs as clinically relevant determinants of tumor aggressiveness, with potential implications across cancer types.