Abstract
BACKGROUND: Early diagnosis benefits lung cancer patients with higher survival, but most patients are diagnosed after metastasis. Although cell-free DNA (cfDNA) analysis holds promise, its sensitivity for detecting early-stage lung cancer is unsatisfying. We leveraged cfDNA fragmentomics to develop a predictive model for invasive stage I lung adenocarcinoma (LUAD). METHODS: 292 stage I LUAD patients from three medical centers were included together with 230 healthy controls whose plasma cfDNA samples were profiled by whole-genome sequencing (WGS). Multiple cfDNA fragmentomic motif features and machine learning models were compared in the training cohort to select the best model. Model performance was assessed in the internal and external validation cohorts and an additional dataset. FINDINGS: A logistic regression model using the 6bp-breakpoint-motif feature was selected. It yielded 98·0% sensitivity and 94·7% specificity in the internal validation cohort [Area Under the Curve (AUC): 0·985], while 92·5% sensitivity and 90·0% specificity were achieved in the external validation cohort (AUC: 0·954). It is sensitive for early-stage (100% sensitivity for minimally invasive adenocarcinoma, MIA) and <1 cm (92·9%-97·7% sensitivity) tumors. The predictive power remained high when reducing sequencing depth to 0·5× (AUC: 0·977 and 0·931 for internal and external cohorts). INTERPRETATION: Here we have established a cfDNA breakpoint motif-based model for detecting early-stage LUAD, including MIA and very small-size tumors, shedding light on early cancer diagnosis in clinical practice. FUNDING: National Key R&D Program of China; National Natural Science Foundation of China; CAMS Initiative for Innovative Medicine; Special Research Fund for Central Universities, Peking Union Medical College; Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences; Beijing Hope Run Special Fund of Cancer Foundation of China.