Genome-wide cell-free DNA fragmentomes from blood and saliva enable early detection of head and neck cancers

INTRODUCTION: Over a billion people globally are impacted by risk factors for head and neck squamous cell carcinoma (HNSCC) - primarily smoking and smokeless tobacco, alcohol and human papillomavirus (HPV) - and may benefit from screening programs if they were available (five-year survival more than doubles for early-stage compared to metastatic disease across tumor sites but currently only 30% of cancers are diagnosed in early stages). We previously developed non-invasive blood tests for multiple cancer types using cell-free DNA (cfDNA) fragmentation profiles (DELFI) and repeat landscapes (ARTEMIS). These tests use cost-effective low-coverage whole-genome sequencing (WGS) to analyze millions of cfDNA fragments whose length, composition, and genome-wide distribution reflect cancer-related genetic and epigenetic changes. Here, we expand these analyses to another non-invasively sampled biofluid, saliva, and demonstrate their use for HNSCC detection. METHODS: We established a prospective single-center collection of blood and saliva samples from individuals with and without HNSCC (n=329). These include 68 healthy individuals without high-risk exposures, 62 healthy individuals who use chewing tobacco, 17 individuals with high-risk leukoplakia, and 182 individuals with cancer, including 39, 37, 27, and 79 with stages I-IV of HNSCC respectively. We extracted cfDNA from blood (all, n=329) and saliva (subset, n=96), performed low-coverage WGS, and split individuals into Discovery (n=244) and Validation (n=85) cohorts. We evaluated genome-wide fragmentation profiles (DELFI) and repeat landscapes (ARTEMIS) in blood and saliva. We cross-validated three ARTEMIS-DELFI classifiers using blood, saliva, or features from both biofluids when available in the Discovery Cohort, and evaluated the locked models in the held-out Validation Cohort. RESULTS: Genome-wide fragmentomes were concordant between blood and saliva samples from the same patients (median correlation coefficient r=0.76, IQR 0.64 - 0.87), and reflected cancer-related alterations including chromosomal changes such as 8q gain and 3p and 9p loss. In the Discovery Cohort, individuals with HNSCC were detected with high performance (AUC=0.87, 95% CI=0.83-0.91 and AUC=0.85, 95% CI=0.76-0.94 for plasma and saliva, respectively), and a combined model using features from both biofluids achieved an AUC of 0.89 (95% CI 0.82-0.97). In the Validation Cohort, at locked thresholds corresponding to 90% specificity, 65%, 71%, and 71% of cancers were detected by the blood, saliva and combined models, respectively. Across cohorts, sensitivity for Stage I tumors was >= 50% for all models. CONCLUSIONS: We demonstrated that cfDNA fragmentation profiles and repeat landscapes can be obtained from both blood and saliva. These approaches may enable accessible, high-performance screening for HNSCC in global high-risk populations.