Transcriptional states reveal drug vulnerabilities and predict EGFR inhibitor response in HPV-negative head and neck cancer
Background: Linking molecular classification to drug response
A study has linked transcriptional states in human papillomavirus (HPV)-negative head and neck squamous cell carcinoma (HNSCC) to gene dependencies and explored gene-expression-based prediction of erlotinib response. The researchers trained a model using cell lines and tested it in six patient-derived tissue samples. The study assessed drug sensitivity in ex vivo tissue; it did not establish clinical efficacy following treatment in patients.
Existing molecular subtypes of HNSCC describe biological differences among tumors but have had limited utility in guiding treatment selection. Epidermal growth factor receptor (EGFR) expression alone also does not adequately explain response to its inhibitors. This study examined whether broader gene expression programs could account for drug response.
Key findings: State-specific dependencies and a 13-gene prediction model
The researchers assembled 753 RNA sequencing (RNA-seq) samples from five cohorts and analyzed the 727 that passed quality criteria. The data also included normal tissue and HPV-positive tumors. They used this integrated dataset to distinguish major transcriptional states in HPV-negative tumors.
The main states were associated with proliferation, epithelial differentiation, epithelial–mesenchymal transition (EMT), and mitochondrial metabolism. The researchers calculated the similarity of pathway activity between tumors and cell lines, selected 18 representative HPV-negative cell lines, and linked these states to DepMap dependency data based on existing CRISPR gene-knockout experiments and drug-screening results. Gene dependency indicates how much removing a particular gene impairs cell survival or proliferation.
The proliferative state showed dependencies related to autophagy and mitosis, while the epithelial-differentiation state showed dependencies related to ERBB/PI3K signaling and adhesion. The mitochondrial-metabolism state showed dependencies involving oxidative phosphorylation and mitochondrial translation. The EMT state showed dependencies involving G2/M cell-cycle regulation, integrins, Notch, and other pathways. These findings provide a basis for exploring treatment candidates for each state. The effects of inhibiting these pathways in patients remain to be tested separately.
To predict response to EGFR inhibitors, the researchers used gene-expression and drug-response data from 26 cell lines. A support vector regression (SVR) model developed through feature selection and cross-validation was reported in the paper as a 13-gene signature. Independent validation used three-dimensional microtumors derived from six prospectively collected HPV-negative HNSCC tissue samples. After treatment with 2.5 μM erlotinib, tissue viability was assessed by luminescence over 3–7 days and normalized to a vehicle control.
The main text reports a Pearson correlation coefficient of R=0.93 between predicted and observed values and a mean absolute error (MAE) of 21.22. MAE indicates the magnitude of the viability-prediction error; a high correlation does not mean that predictions exactly match observations. The caption for Figure 6F gives different validation performance values, R=0.88 and MAE=18.44, leaving the figures to be clarified. In a comparison using EGFR alone, mRNA expression was only weakly correlated with erlotinib response (R=−0.14).
Implications and outlook: Hypotheses for treatment selection and validation needs
The study presents a framework that links tumor transcriptional states with functional dependencies and drug sensitivity. When the model was applied to the integrated tumor data, greater erlotinib sensitivity was predicted for epithelial-differentiation and cell-cycle-related states. However, the model was trained on 26 cell lines and independently validated in only six tissue samples, limiting conclusions about how well it generalizes.
Cell-line dependencies do not capture the full tumor microenvironment, and predictions of tissue viability must be distinguished from patients’ treatment responses or improved survival. Follow-up studies will need to reproduce performance in larger independent cohorts and assess its association with actual treatment outcomes. The study’s methods and results are available in the original paper.
Oncology & Systems Biology Core, Published May 2026. DOI: [Source Generated Data]Summary: Resolving the clinical translation gap where historical head and neck squamous cell carcinoma (HNSCC) transcriptomic subclassifications remained purely descriptive, this mechanistic platform converts multi-omic diversity into drug-actionable states. By synthesizing multi-cohort RNA-seq metadata from 727 independent tumors across five diverse datasets with genome-scale CRISPR dependency maps, the framework isolates four selective survival circuits characterizing HPV-negative HNSCC: a proliferative axis, an epithelial-differentiated state, an EMT-like matrix, and an oxidative metabolic baseline. To clinically exploit these targets, a machine learning pipeline was deployed to extract a 13-gene transcriptomic predictor of EGFR-inhibitor response. Validated prospectively in fresh patient-derived 3D microtumors, the signature identifies erlotinib susceptibility with a high fidelity coefficient (R = 0.93) by mapping the structural penetrance of epithelial differentiation, establishing a generalizable computational standard for programmatic biomarker optimization and precision stratification.
The study raises the possibility of assessing drug sensitivity that EGFR expression alone cannot readily explain by considering tumor transcriptional state as well. In follow-up research, RNA expression from surgical tissue could be used to predict response, with the predictions then compared with ex vivo drug testing of the same tissue to evaluate candidate biomarkers for clinical trials and criteria for stratifying patients. Gene-dependency data could also help prioritize treatment candidates to investigate in each transcriptional state. However, validation in only six tissue samples and the discrepancy between performance values reported within the paper mean the current results are insufficient to guide drug selection in clinical care or justify use as an approved companion diagnostic.