Transcriptional states define dependencies and therapeutic vulnerabilities in head and neck cancer.

Molecular heterogeneity in head and neck squamous cell carcinoma (HNSCC) is well recognized, yet existing subtype frameworks remain largely descriptive and have not translated into therapeutic decision-making. Here, we establish a mechanistic platform that converts transcriptomic diversity into drug-actionable tumor states. Integrating multi-cohort RNA-seq from 727 tumors across five independent datasets, genome-scale CRISPR dependency maps, and pharmacologic screening, we define distinct tumor survival circuits across HPV-negative HNSCC and nominate subtype-matched therapeutic strategies. These circuits encompass a proliferative axis (MYC, MET/FAK, inflammatory and translational programs), an epithelial-differentiated/adhesion program, an EMT-like state with stromal activation, and mitochondrial/oxidative metabolic states, each mapping to selective liabilities (e.g., mitotic/autophagy control, ERBB/PI3K and cadherin signaling, OXPHOS/mitochondrial translation, and G2/M-integrin-Notch pathways, respectively). We then develop a transcriptomic predictor of EGFR-inhibitor response using machine learning and validate it in prospectively collected, fresh patient-derived 3D microtumors. The resulting 13-gene signature identifies erlotinib-responsive tumors (R = 0.93) and maps biologically to an epithelial-differentiated state, outperforming EGFR expression alone. Our study establishes a subtype-to-dependency-to-therapy framework, enabling precision stratification and providing a clinically feasible path for prospective biomarker deployment.