Abstract 3998: AI-driven spatial analysis of 1100 head and neck squamous cell carcinoma patients H&E slides uncovers three sub-types with markedly distinct tumor microenvironments and survival patterns

Abstract Background: The spatial organization of the tumor microenvironment (TME) is a critical determinant of tumor progression and patient outcomes in head and neck squamous cell carcinoma (HNSC). However, spatial transcriptomics (ST) technology remains limited by high costs and tissue requirements, which restrict its application to large clinical cohorts and hinder biomarker translation. Methods: We trained a deep learning model on a HNSC ST cohort comprising 21,466 spots from 10 slides to predict spot-level spatial gene expression directly from H 0.4 for predicted vs. actual ST gene expression) across all three ST cohorts. Second, we applied it to infer the spatial gene expression for 1,145 patients from two independent datasets (TCGA-HNSC, n = 445; HANCOCK, n = 700). Third, spatial clustering of inferred TCGA-HNSC profiles identified 11 shared spatial clusters representing TME composition within each patient. Fourth, hierarchical clustering of these TME compositions revealed three distinct patient subgroups, termed SpatioStates, each associated with unique biological features and survival patterns. The Proliferative SpatioState, dominated by hyperproliferative-hypoxic programs, represents an aggressive tumor niche associated with poor overall survival. In contrast, the Immune-Activated SpatioState exhibits strong inflammatory signaling and robust immune-cell engagement, reflecting active antitumor immunity and showing significantly improved prognosis compared with the Proliferative SpatioState (HR = 0.66, p = 0.023) after adjustment for age and stage. Meanwhile, the Metabolic SpatioState is characterized by high oxidative phosphorylation and stress-response pathways corresponding to metabolically adaptive tumor regions with intermediate survival outcomes. Remarkably, these SpatioStates are consistently repeated in the independent HANCOCK cohort, maintaining significant prognostic discrimination between Proliferative and Immune-Activated SpatioStates (HR = 0.46, p = 0.005), underscoring the robustness and generalizability of our model. Conclusions: This study establishes the first scalable computational framework for inferring prognostic spatial TME architecture from routine H Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3998.