A non-invasive cfDNA fragmentomics assay for differentiating malignant and benign thyroid nodules.

6110 Background: Thyroid cancer (TC) is the most common endocrine malignancy in the world. Although it has a high survival rate, distinguishing between the common benign thyroid nodules and malignant ones remains a clinical challenge. We introduce a non-invasive assay based on cfDNA fragmentomics, aimed at accurately identifying TCs, thus reducing the discomfort and risks linked to surgical biopsies. Methods: We conducted a study involving 322 participants, comprising 161 early-stage thyroid cancer patients (154 stage I and 7 stage II) and 161 patients with benign thyroid nodules, of whom 68 were high-risk (Ti-Rads: 4/5) and 93 were low-risk (Ti-Rads = 2/3). The training cohort contained 94 TC and 98 BN patients, while the independent test cohort included 67 TC and 63 BN patients. Malignant and high-risk benign conditions were confirmed pathologically from surgically removed biopsy samples. Pre-operative plasma samples were collected from all patients and used for cell-free fragmentomics feature generation through low-pass WGS. A stacked machine learning model was developed using three fragmentomics features and three machine learning algorithms. Results: The machine learning model excelled in differentiating thyroid cancer from benign nodules, achieving area under the curve (AUC) scores of 0.973 in the training cohort and 0.965 in the independent test cohort. The model demonstrated high sensitivity (94.7% in the training cohort and 95.5% in the test cohort) and specificity (89.8% in the training cohort and 77.8% in the test cohort) after applying an optimized cutoff determined by cross-validation of the training cohort. Notably, the model was particularly effective in detecting malignant nodules smaller than 1cm, a task challenging for conventional methods, with sensitivities of 92.4% in the training cohort and 91.6% in the test cohort. For nodules larger than 1cm, the model achieved a perfect 100% sensitivity in both cohorts. Our predictive model showed universal excellent performance for accurately detecting TC patients with different histology, sex, and age in both the training cohort and the independent test cohort. Finally, the high-risk BN group showed higher risk scores compared to the low-risk BN group as expected, further validating the robustness of our model. Conclusions: Our study presents a non-invasive liquid biopsy assay, which utilizes fragmentomics profiling derived from low-pass WGS for differentiating between malignant and benign thyroid nodules, showcasing remarkable sensitivity and specificity. Our assay was able to accurately detect thyroid malignancies, especially in nodules smaller than 1cm, illustrating great clinical potential and therefore offers a promising alternative to surgical biopsies, potentially reducing patient discomfort and associated risks.