Abstract A065: Fast and cost-effective prediction of treatment response in head and neck cancer by characterizing the tumor microenvironment from routine H&E slides

Abstract Background: Treatment response prediction in head and neck squamous cell carcinoma (HNSCC) remains a major clinical challenge due to spatial heterogeneity within the tumor microenvironment (TME). While spatial transcriptomics (ST) technology offers a powerful means to profile TME architecture, its clinical utility is constrained by high cost, tissue requirements, and limited scalability, restricting its application to large retrospective cohorts. To address these limitations, we developed an AI-framework to infer spatial gene expression profiles directly from routine H&E-stained histology slides to identify spatial biomarkers predictive of treatment response. Methods: We trained an AI model on an HNSCC ST cohort (21,466 spots, 10 slides from 10 patients) to predict spot-level spatial gene expression from H&E-stained images and validated its performance on two independent ST cohorts (11,485 spots, 8 slides from 8 patients), covering diverse anatomical sites and sample types. Results: First, we established that the trained model reliably predicts spatial expression for 2,772 genes (Pearson R > 0.4 predicted vs measured ST gene expression). Second, we employed the model to infer spatial gene expression from a large TCGA-HNSC cohort consisting of 445 HNSCC patients. Third, by spatial clustering of inferred ST expression, we identified 11 shared spatial clusters representing TME composition within each patient. Fourth, we utilized these spatially defined clusters to characterize the TME landscape of both adjuvant-treated (TCGA-HNSCC, n=226) and immunotherapy-treated patients (external cohort, n=94). Fifth, we quantified each spatial cluster's proportional representation per patient and trained two separate models using five-fold cross-validation to predict response for the adjuvant and immunotherapy-treated cohorts. The spatial cluster compositions are predictive of adjuvant therapy response, achieving an AUROC of 0.67±0.09. When combined with clinical variables (age and HPV status), performance improved further to an AUROC of 0.74±0.12. Notably, non-responders are predominantly enriched for a spatial cluster that exhibits a KRAS-driven, inflammatory-yet-immunosuppressed transcriptional signature associated with treatment resistance. Similarly, in the immunotherapy cohort, the model achieved an AUROC of 0.71±0.14. The spatial cluster having the strongest association with immunotherapy treatment response, characterized by an immunoreactive immune phenotype (T-cell-inflamed, IFN-γ-high tumor state) consistent with favorable immunotherapy outcomes. These results demonstrate that model-derived spatial features can reliably predict treatment response across diverse therapeutic modalities in HNSCC, with distinct spatial signatures distinguishing responders. Conclusions: We present the first AI-framework to derive spatial TME features from routine H&E slides that are predictive of treatment response in HNSCC. Our approach provides a scalable, clinically feasible framework for spatial biomarker discovery in precision immuno-oncology. Citation Format: Sumona Biswas, Sumeet Patiyal, Amos Stemmer, Saugato Rahman Dhruba, Eldad Shulman, Chi-Ping Day, Tien-Hua Chen, Muh-Hwa Yang, Eytan Ruppin. Fast and cost-effective prediction of treatment response in head and neck cancer by characterizing the tumor microenvironment from routine H&E slides [abstract]. In: Proceedings of the AACR Immuno-Oncology Conference (AACR IO): Discovery and Innovation in Cancer Immunology: Revolutionizing Treatment through Immunotherapy; 2026 Feb 18-21; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Immunol Res 2026;14(2 Suppl):Abstract nr A065.