Plasma metabolite biomarker identification study for the early detection of gastric cancer

Background: Gastric cancer (GC) is the fifth most prevalent and the fifth deadliest cancer worldwide, and timely diagnosis of GC contributes to an increased survival rate. However, current detection methods for GC mainly rely on gastroscopy examination, limited by relatively low compliance. We attempted to identify plasma metabolite biomarkers and develop a diagnostic model for GC. Methods: A total of 597 subjects, including healthy controls and GC patients were recruited from multiple centers in China. Ultra-performance liquid chromatography–mass spectrometry–based metabolomics methods were used to characterize the subjects’ plasma metabolic profiles and to screen and validate the GC biomarkers. Five machine learning algorithms (neural network, support vector machine, ridge regression, lasso regression and Naïve Bayes) were used to build a diagnostic model. We compared the performance of the metabolic panel with risk factors and clinical protein biomarkers (CA724, CA199, CA242, CA125, CEA and AFP), involving sensitivity, specificity, accuracy, AUC and clinical net benefit. Findings: A plasma metabolite biomarker panel consisting of 6 metabolites was constructed and identified for GC diagnosis. Among the five machine learning algorithms, the neural network algorithm demonstrated the best diagnostic performance, achieving AUC of 0.982 (95% CI: 0.965–0.998) and 0.951 (95% CI: 0.931–0.970) in the discovery and validation dataset, respectively. The panel's sensitivity, specificity, and accuracy (95% CI) were 0.940 (0.825–0.984), 0.936 (0.861–0.974), and 0.938 (0.881–0.969) in the discovery set, and 0.925 (0.881–0.954), 0.867 (0.814–0.907), and 0.896 (0.864–0.922) in the validation set, respectively. The panel also exhibited superior diagnostic performance in detecting early-stage GC, with the ridge regression algorithm achieving the best performance (AUC: 0.982, 95% CI: 0.965–0.998 and 0.951, 0.931–0.970 in the discovery and validation dataset). This panel significantly outperforms clinical protein biomarkers in sensitivity. For instance, CA724, the most sensitive clinical biomarker for GC, showed sensitivities of only 0.240 (95% CI: 0.131–0.382) in the discovery dataset and 0.148 (95% CI: 0.103–0.203) in the validation dataset. Interpretation: The discovered and validated serum metabolite biomarker panel exhibits good diagnostic performance for the early detection of GC, highlighting the potential in clinical practice for GC diagnosis and offering insights into the metabolic characterization of diseases including but not limited to GC. Funding: This study was supported by grants from the Medical and Health Research Project of Zhejiang Province (2024KY050), Zhejiang Cancer Hospital's National Natural Science Foundation Cultivation Fund (PY2022042) and Zhejiang Cancer Hospital Youth Research Fund (QN202201).