Multi-center investigation of an early detection model utilizing cfDNA fragmentomics for breast cancer screening.

Abstract

e15033 Background: Breast cancer is the most prevalent malignancy among women globally, and early detection significantly improve the prognosis of BC patients. Circulating free DNA (cfDNA) fragmentomics presents a promising, minimally invasive alternative through tumor-specific fragment analysis. In this study, we developed a deep learning model using cfDNA ultra-low pass whole-genome sequencing (LP-WGS) for early breast cancer screening. Additionally, we investigated its potential for molecular subtype classification and lymph node metastasis prediction, aiming to provide novel insights into personalized breast cancer diagnosis and treatment. Methods: This prospective, multicenter clinical study (NCT06016790) enrolled 664 breast cancer patients (79.5% early-stage, stages 0–IIa) and 289 benign controls across seven tertiary-grade A hospitals and one community hospital. Plasma cfDNA samples were collected following standardized protocols and analyzed using the TuFEst (Tumor Fraction Estimator), a deep neural network-based machine learning model designed for cfDNA LP-WGS. Model performance was evaluated through cross-validation using both internal and external validation cohorts. RNA-seq was performed on corresponding tumor bulk tissues (n = 90). Results: The TuFEst-based model demonstrated strong predictive performance for early breast cancer detection. In the training cohort (n = 412), 10-fold cross-validation yielded a sensitivity of 95.0% and specificity of 78.3% (AUC: 0.942). In the internal validation cohort (n = 224), sensitivity was 93.2%, specificity was 69.7% (AUC: 0.937), and in the external validation cohort (n = 156), sensitivity was 95.8%, specificity was 86.9% (AUC: 0.968). The model also showed robust ability to stratify molecular subtypes (training cohort: ER+PR+HER2- AUC: 0.928, HER2+ 0.921, TNBC 0.928; validation cohort: ER+PR+HER2- AUC: 0.899, HER2+ 0.925, TNBC 0.896) and identify patients with or without lymph node metastasis (training cohort AUC:0.9, internal validation cohort AUC:0.874, external validation cohort: 0.893). Additionally, this model successfully predicted the molecular subtype of metastatic sites in 80.9% (17/21) of patients with oligometastatic lesions. Transcriptome analysis revealed significant upregulation of immune signatures in tumors with higher cancer scores (based on TuFEst algorithm). Conclusions: This study represents the largest multicenter investigation of cfDNA fragmentomics in a Chinese population, predominantly consisting of early-stage patients. The TuFEst model shows promise as a tool for early detection, molecular subtyping and lymph node metastasis prediction in breast cancer, addressing key limitations of current screening practices. Future research will focus on large-scale validation to further assess its potential for population-level screening and treatment decision-making. Clinical trial information: NCT06016790 .