Novel early-detection model based on cfDNA methylation and fragmentation features for liver cancer.

4133 Background: In China, the 5-year survival rate of liver cancer patients is only 14%, far lower than the average of 43.7% for all cancer types. Early diagnosis and treatment are essential for survival. Traditional screening methods like AFP combined with abdominal ultrasound have low sensitivity. Recent studies suggest that blood cell-free DNA (cfDNA) characteristics could be a new screening approach for liver cancer. This study aims to compare methylation and fragmentation signals among liver cancer, hepatitis, cirrhosis patients, and healthy individuals, innovatively using these signals to construct an early-detection model which could improve patient prognosis. Methods: From July 2023 to November 2024, 315 blood samples were prospectively collected from five Chinese hospitals. The sample set included 105 liver cancer patients and 210 non-liver cancer controls (33 hepatitis, 30 cirrhosis, 147 healthy). This multi-center, multi-disease-controlled collection provides a robust data basis. Targeted enzymatic methyl sequencing detected over 600,000 methylation sites, enabling precise exploration of liver-cancer-related methylation. Beyond methylation, novel fragmentation features like break-point motifs, end motifs, arm-level count, fragment-size distribution and ratio were obtained. These, combined with methylation data, offer a multi-dimensional view for studying liver cancer pathogenesis and biomarkers. A gradient-boosted tree model, integrating 3840 methylation DMR features and fragmentomic model-predicted probabilities, was built. A nested cross-validation framework was used to optimize the model and ensure result accuracy. Results: The model achieved a high AUC of 0.97(95%CI:0.95-0.99) in liver cancer detection. At 96.2% specificity, the model had a 91.4% sensitivity for overall liver cancer detection, with 83.7% and 95.8% sensitivity for stage I and II respectively. Among 63 patients with hepatitis or cirrhosis, the model accurately predicted negative results in 88.9% of patients. Notably, for patients hard to identify by traditional tumor markers like AFP and DCP, the model showed high detection rates. When AFP < 400 ng/ml, the detection rate was 88.9%, and with concurrent DCP < 40 ng/ml, it reached 87.0%. When AFP < 20 ng/ml, the detection rate was 89.5%, and with DCP < 40 ng/ml simultaneously, the detection rate was 83.3%. Conclusions: This study established an early-detection model for liver cancer by leveraging cfDNA methylation and fragmentation signals. The model demonstrated remarkable performance, particularly in detecting liver cancer patients who are difficult to identify through conventional methods. It blazes a new trail for the early detection of liver cancer and could significantly enhance patient prognosis.