Spatial biomarker-driven deep learning model via digital pathology predicts response to PI3K inhibitor buparlisib in head and neck squamous cell carcinoma
Antoine Desîlets, Minh Tri Le, Justin Lucas, Orit Matcovitch-Natan, Amit Bart, Avi Laniado, Meir Azulay, Ettai Markovits, Jennifer Kaplan Kerner, Amit Gutwillig, Hadar Yehezkeli, Lisa Francesca Licitra, Sunny Y. Lu, Kevin Dreyer, Ying Pan, Nanhai He, Archie N. Tse, Sandrine J. Faivre, Denis Soulières · medRxiv · 2025
ABSTRACT Purpose Buparlisib, a pan-class I PI3K inhibitor, combined with paclitaxel, demonstrated improved survival in the BERIL-1 trial for patients with recurrent/metastatic (R/M) head and neck squamous cell carcinoma (HNSCC). However, predictive biomarkers of benefit remain undefined. We evaluated spatial biomarkers derived from hematoxylin and eosin (H&E) images using artificial intelligence (AI)–based digital pathology. Materials and Methods Whole-slide H&E images (n=144) from BERIL-1 were analyzed using a deep learning model trained to segment tissue compartments and classify individual cell phenotypes. Three prospectively defined spatial features were evaluated: (1) tumor-infiltrating lymphocyte (TIL) density in the tumor area; (2) tumor microenvironment (TME) heterogeneity; and (3) granulocyte fraction in the tumor invasive margin (TIM). Cox proportional hazards model was used to evaluate biomarker–treatment interactions, with patients stratified by biomarker status. Results High TIL density (>10%) defined by deep learning–derived analysis of H&E was associated with a significantly improved overall survival with buparlisib versus placebo (HR□=□0.25; 95% CI, 0.01–0.64; p = 0.002), as were high TME heterogeneity (HR□=□0.47; 95% CI, 0.27–0.80; p = 0.005) and granulocyte enrichment in the TIM (HR = 0.51; p = 0.014); in a within-arm proximity analysis, higher granulocyte–tumor cell proximity correlated with improved OS on buparlisib (HR = 0.32; p < 0.001). AI-derived spatial metrics outperformed CD3 immunohistochemistry staining in stratifying survival outcomes. In patients with oropharyngeal tumors, human papillomavirus-positive cases were more frequent among those with high TILs. Conclusions Spatial features extracted from standard H&E slides using AI-driven digital pathology can predict OS benefit from buparlisib in R/M HNSCC. These cost-effective and scalable biomarkers support image-based patient selection strategies and are being prospectively evaluated in the ongoing BURAN phase 3 trial.