Abstract 6611: Genome-wide cfDNA fragmentomics combined with machine learning enhance non-invasive prostate cancer detection

Abstract The liquid biopsy approach focusing on DNA molecules offers a promising non-invasive method for prostate cancer diagnosis. Next-generation sequencing (NGS) of targeted gene panels is widely used for detecting somatic mutations but is limited by confounding clonal hematopoietic variants and narrow genome coverage. We propose a solution that encompasses genome-wide DNA fragmentation features for tumor fragment evaluation. This study collected plasma samples from 209 patients with prostate cancer and 49 healthy individuals as controls. Cell-free DNA (cfDNA) was extracted and tested using PredicineSCORE™, a low-coverage whole-genome sequencing assay designed to identify tumor-specific features. Tumor content assessed by mutation-based detection assays served as the reference indicator. Fragmentomics characteristics—including regional coverage, fragment length, and end-motif features—were profiled as tumor markers. Copy number burden and tumor fragment content were refined using the expectation-maximization algorithm. An ensemble machine learning model was then developed to estimate tumor fractions. We found strong correlations between estimated tumor fraction and fragmentomics features. The absolute correlation coefficients for tumor fraction with the relative depth at LYL1 binding sites, the insert size ratio of short reads, and the CCCA end motif were higher than 0.9, 0.75, and 0.7, respectively. To validate these findings, a total of 500 simulated samples with tumor content ranging from 0.1% to 50% were generated by mimicking tumor and normal reads. The ensemble model outperformed single-method approaches, with the XGBoost algorithm achieving high predictive performance (Pearson correlation r > 0.9, p < 0.01) for predicting tumor fraction in cfDNA. The model demonstrated a limit of detection (LoD) at 2%, surpassing previously proposed models in sensitivity. This study highlights the potential of plasma cfDNA fragmentomics profiling as a highly precise and non-invasive method for quantifying and detecting prostate cancer. The findings support its utility in advancing liquid biopsy applications with improved sensitivity and accuracy. Citation Format: Shuang Gan, Haoran Tang, Feng Xie, Yue Zhang, Hang Dong. Genome-wide cfDNA fragmentomics combined with machine learning enhance non-invasive prostate cancer detection [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6611.