Integrating Plasma Cell-Free DNA Fragment End Motif and Size with Genomic Features Enables Lung Cancer Detection

Abstract

Early detection of lung cancer is important for improving patient survival rates. Liquid biopsy using whole-genome sequencing of cell-free DNA (cfDNA) offers a promising avenue for lung cancer screening, providing a potential alternative or complementary approach to current screening modalities. Here, we aimed to develop and validate an approach by integrating fragment and genomic features of cfDNA to enhance lung cancer detection accuracy across diverse populations. Deep learning-based classifiers were trained using comprehensive cfDNA fragmentomic features from participants in multi-institutional studies, including a Korean discovery dataset (218 patients with lung cancer and 2,559 controls), a Korean validation dataset (111 patients with lung cancer and 1,136 controls), and an independent Caucasian validation cohort (50 patients with lung cancer and 50 controls). In the discovery dataset, classifiers using fragment end motif by size, a feature that captures both fragment end motif and size profiles, outperformed standalone fragment end motif and fragment size classifiers, achieving an area under the curve (AUC) of 0.917. The ensemble classifier integrating fragment end motif by size and genomic coverage achieved an improved performance, with an AUC of 0.937. This performance extended to the Korean validation dataset and demonstrated ethnic generalizability in the Caucasian validation cohort. Overall, the development of a deep learning-based classifier integrating cfDNA fragmentomic and genomic features in this study highlights the potential for accurate lung cancer detection across diverse populations. Significance: Evaluating fragment-based features and genomic coverage in cell-free DNA offers an accurate lung cancer screening method, promising improvements in early cancer detection and addressing challenges associated with current screening methods.