Genome-wide variation in cell-free DNA end-motif entropy predicts immunotherapy response in head and neck cancer
Ravi Bandaru, Hailu Fu, Haizi Zheng, Jocelyn Liang, Li Wang, Shuchi Gulati, Benjamin H. Hinrichs, Mingxiang Teng, Bin Zhang, Masha Kocherginsky, De-Chen Lin, David A. Hildeman, Francis P. Worden, Matthew Old, Neal E. Dunlap, John M. Kaczmar, Maura L. Gillison, Dalia El-Gamal, Trisha Wise Draper, Yaping Liu
Ravi Bandaru, Hailu Fu, Haizi Zheng, Jocelyn Liang, Li Wang, Shuchi Gulati, Benjamin H. Hinrichs, Mingxiang Teng, Bin Zhang, Masha Kocherginsky, De-Chen Lin, David A. Hildeman, Francis P. Worden, Matthew Old, Neal E. Dunlap, John M. Kaczmar, Maura L. Gillison, Dalia El-Gamal, Trisha Wise Draper, Yaping Liu
View: Text | PDF
Clinical Research and Public Health Genetics Oncology

Genome-wide variation in cell-free DNA end-motif entropy predicts immunotherapy response in head and neck cancer

Abstract

BACKGROUND Minimally invasive biomarkers predicting the immunotherapy response in head and neck squamous cell carcinoma (HNSCC) remain an unmet clinical need.METHODS In a prospective, multi-institutional phase II trial, we performed whole-genome sequencing of 185 longitudinal plasma cell-free DNA (cfDNA) samples from 68 patients with locally advanced, surgically resectable HNSCC who received neoadjuvant and adjuvant pembrolizumab. We developed the regional motif diversity score (rMDS), a fragmentomic metric that quantifies the entropy of cfDNA 5′-end motifs across genomic regions.RESULTS Unsupervised analysis showed that rMDS robustly distinguished responders from nonresponders, outperforming established fragmentomic metrics and copy number alterations while remaining independent of technical confounders. Longitudinal rMDS changes localized to regions enriched for immune-, lectin-, and keratinization-related genes — hallmarks of squamous cell carcinoma — reflecting tumor–peripheral immunity interplay during treatment. The most dynamic regions clustered at telomere-proximal loci, suggesting a link between telomere biology and cfDNA fragmentation. An rMDS-based machine learning classifier achieved AUC 0.89–0.99 across validation settings, with the highest accuracy after treatment, outperforming PD-L1 expression and tumor fraction in matched samples. Predicted responders showed improved disease-free survival (log-rank P = 0.035; HR 2.67, 95% CI 1.03–6.92).CONCLUSION rMDS represents a biologically meaningful and clinically actionable biomarker for the immunotherapy response in HNSCC, and merits integration into future risk assessment frameworks.TRIAL REGISTRATION ClinicalTrials.gov NCT02641093.FUNDING National Human Genome Research Institute (NHGRI), NIH grant R56HG012360; startup funds from Cincinnati Children’s Hospital Medical Center, Northwestern University, and Robert H. Lurie Comprehensive Cancer Center; Science Olympiad Alumni Research Grant, Science Olympiad USA Foundation; Merck Sharp & Dohme Corp.

Authors

Ravi Bandaru, Hailu Fu, Haizi Zheng, Jocelyn Liang, Li Wang, Shuchi Gulati, Benjamin H. Hinrichs, Mingxiang Teng, Bin Zhang, Masha Kocherginsky, De-Chen Lin, David A. Hildeman, Francis P. Worden, Matthew Old, Neal E. Dunlap, John M. Kaczmar, Maura L. Gillison, Dalia El-Gamal, Trisha Wise Draper, Yaping Liu

×

Figure 1

Experimental design and analytical workflow for the HNSCC cohort.

Options: View larger image (or click on image) Download as PowerPoint
Experimental design and analytical workflow for the HNSCC cohort. Patien...
Patients with head and HNSCC underwent surgical tumor resection and pembrolizumab immunotherapy, followed by adjuvant radiation therapy (60–66 Gy over 6 weeks) and cisplatin (40 mg/m2) for high-risk pathological features. A total of 185 blood samples were collected at defined intervals throughout the course of treatment and follow-up from 68 patients across 6 institutions (UC, U-M, OSU, MUSC, MDACC, and UofL). Plasma-derived circulating cfDNA was isolated from blood samples collected at 3 different time points (screen, day 0, and adjuvant week 1) and subjected to WGS. Fragmentation features were extracted using FinaleToolkit, including MDS, single genome-wide summary score), rMDS, end-motif frequencies, DELFI, genome-wide coverage, and fragment length distributions. CV, cross-validation.