Evaluation of Raw Cell-Free DNA Sequences for Gastric Cancer Detection.

Abstract

Gastric cancer (GC) is a significant global health challenge, often diagnosed late, limiting treatment success. Early detection through biomarkers is critical for better outcomes. This study developed a noninvasive, cost-effective tool using circulating cell-free DNA (cfDNA) sequencing and advanced modeling for early GC detection. We performed low-coverage whole-genome sequencing on plasma from 404 patients with GC and 428 non-GC participants (healthy donors and patients with colorectal, esophageal, or liver cancer). Participants were divided into a training set (579) and a validation set (253). A deep learning model, GastricAI, was trained on raw cfDNA reads using Mamba, an advanced sequence model enhancing the structured state space model (S4) with adaptive parameters. Performance was evaluated in a second validation set (73 GC patients, 94 controls). In the first validation set, GastricAI achieved an area under receiver operating characteristic curve (AUROC) of 0.886 (95% CI, 0.842 to 0.930), accuracy of 0.846, sensitivity of 0.789, and specificity of 0.900. Across TNM stages, it recorded sensitivity/specificity of 0.920/0.738 (stage I), 0.846/0.900 (stage II), 0.811/0.900 (stage III), and 1.000/0.623 (stage IV). In the second validation set, it showed an AUROC of 0.850 (95% CI, 0.793 to 0.907), accuracy of 0.772, sensitivity of 0.973, and specificity of 0.617. GastricAI consistently detected intestinal, diffuse, and mixed GC subtypes per Lauren's classification and outperformed conventional biomarkers carbohydrate antigen 724 (<i>P</i> < .001) and carcinoembryonic antigen (<i>P</i> = .037). GastricAI provides a promising noninvasive method for GC detection via cfDNA sequencing, with high sensitivity and reasonable specificity. It holds potential for early screening to improve prognosis, although further prospective studies are needed to validate and refine its clinical utility across diverse populations.