Abstract B051: Elastic net discovery of DNA methylation biomarkers for non-invasive diagnosis and recurrence detection in head and neck cancer

Abstract Background: Head and neck squamous cell carcinoma (HNSCC) is diagnosed in nearly 60,000 Americans each year. Despite therapeutic advances in recent decades, survival rates have remained largely unchanged. A major contributing factor to poor outcomes is the lack of a non-invasive screening test for HNSCC, as well as the high recurrence rates within two years of curative therapy. In this study, we aimed to leverage elastic net penalized regression to identify and validate DNA methylation biomarkers for HNSCC using non-invasive oral rinse samples. Methods: Oral rinse methylation data from the Collaborative Study of Head and Neck Diseases (CoHANDS), a large population-based case-control study of HNSCC in the greater Boston-area, was used for biomarker discovery. Data was split 3:1 into training and test datasets. Elastic net hyperparameter optimization with five-fold cross validation and model fiting was conducted on the training data set. Performance of the resulting model was evaluated in the test data. Data splitting and model fitting was repeated over 1,000 iterations. Performance metrics and the features selected by the model were collected for each iteration. Two approaches were employed: (1) a locus-by-locus analysis and (2) an analysis based on the average methylation values across CpG-dense regions identified via Hidden Markov modeling. Biomarker panels for each approach were selected to optimize predicted performance and parsimony. The identified biomarker panels were externally validated for discriminatory performance using tissue biopsy methylation data from The Cancer Genome Atlas (TCGA). Additionally, the biomarker panels were evaluated for their ability to predict recurrence in post-treatment oral rinse samples collected at the University of Cincinnati, which were interogated using the Illumina HumanMethylation EPIC BeadChip. Results: To optimize accuracy while maintaining model simplicity, biomarker panels derived from an α = 0.7 within the elastic net model were selected for further analysis. The individual CpG panel consisted of 14 CpG loci, while the CpG-dense region panel contained 10 CpG-dense regions. Validation using HNSCC methylation data from TCGA demonstrated high discriminatory performance, with AUCs of 0.97 (95% CI: 0.92–1.00) for the CpG-dense region panel and 0.99 (95% CI: 0.97–1.00) for the individual CpG panel. Prospective prediction of recurrence was more modest, with AUCs ranging from 0.83 (95% CI: 0.64 – 1) for the CpG-dense region panel to 0.76 (95% CI: 0.40 – 1) for the individual CpG panel when restricted to patients with a recurrence < 6 months prior to the last provided sample. Conclusions: The biomarker panels demonstrated excellent performance in distinguishing tumor tissue from paired normal tissue using TCGA data, with promise in predicting recurrence in post-treatment oral rinse samples. These DNA methylation biomarker panels may provide a novel screening method for HNSCC diagnosis or early identification of recurrence. Citation Format: Alexander C. Sprague, Damaris Kuhnell, Wei-Wen Hsu, Trisha Wise-Draper, Scott Langevin. Elastic net discovery of DNA methylation biomarkers for non-invasive diagnosis and recurrence detection in head and neck cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B051.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.