Gene Expression Signature for Distant Metastasis in Loco-regionally Advanced Head and Neck Squamous Cell Carcinoma

Abstract

Nearly two-thirds of head and neck squamous cell cancer (HNSCC) present with loco-regionally advanced disease. Despite improved local tumor control from advances in surgery, chemotherapy, and radiotherapy, the rate of distant failure has remained relatively unchanged. Distant metastasis is known to be driven by a series of biological processes that occur in a subset of malignant tumors. We hypothesized that a gene expression signature measured from primary HNSCC tumors could identify patients at highest risk of developing distant metastasis (DM). Patients with Stage III or IV HNSCC in The Cancer Genome Atlas (TCGA) HNSCC dataset with new tumor events of distant metastasis on follow-up were matched with non-metastatic controls in a 1:2 ratio. Differential gene expression, pathway analysis, and bulk tissue cell-type deconvolution were applied to characterize alterations in regulatory programs within the tumor microenvironment. Adaptive LASSO-penalized regression, a form of supervised machine-learning, was used to develop a transcriptional risk score (RS) for distant metastasis. Area under the ROC curve (AUC) assessed performance of the model with ten-fold internal cross-validation. The RS was applied to a separate validation cohort from our institution, with distant metastasis free survival (DMFS) and overall survival as primary endpoints. A subgroup analysis was performed based on HPV status. Twenty-four HNSCC cases with distant metastases and forty-eight matched non-metastatic controls were identified in TCGA dataset. 1100 genes were differentially expressed, composed largely of genes involved in epithelial-stromal remodeling, and cytokine-immune interactions, annotated based on pathway analysis. Single-cell deconvolution revealed a significant increase in fibroblasts and T-cell dysregulation within the tumor microenvironment among DM cases. A 28-gene risk score was developed with an AUC of 0.938. In the validation cohort of 122 patients, high-RS patients had significantly shorter DMFS (hazard-ratio [HR] 9.0; 95% CI, 2.6-31.7, P<0.0001). Overall survival was inversely associated with metastatic cases; however, this did not achieve significance (HR 1.5; 95% CI, 0.9-2.3, P = 0.10). RS was superior to T- and N- stage for DM prognostication, and remained statistically significant after multivariable adjustment. RS had improved predictive value for DMFS in the HPV-negative subgroup (HR 20.8; 95% CI, 2.6-166, P = 0.004). We have identified and validated a group of genes that predict distant metastasis among loco-regionally advanced HNSCC tumors. This score offers a prognostic tool for detecting a group of HNSCC patients at high risk for DM who may benefit from intensified monitoring and/or treatment.