Background Gastric cancer (GC) is a highly prevalent and deadly malignancy, necessitating timely diagnosis and intervention. However, current diagnoses predominantly hinge on gastroscopy, limited by invasiveness and low uptake rates. We aimed to develop diagnostic models for GC utilizing non-invasive plasma metabolic biomarkers.

Methods We conducted a two-phase study involving 647 participants, comprising 277 GC and 370 non-GC. Candidate differential metabolites were identified in the discovery and validation phases using ultra-performance liquid chromatography-mass spectrometry, and a diagnostic model was developed using machine-learning algorithms. Then, mendelian randomization (MR) analysis was used to examine the causal association between metabolic biomarkers and the risk of GC. These metabolic biomarkers were validated in the GC tissue by comparing them with tumor-adjacent non-malignant paired tissue.

Results Twenty-six replicated plasma metabolites were identified in the discovery and validation dataset. Six features were selected to construct a metabolic panel with excellent diagnostic performance (AUCs of 0.947–0.982 in the discovery dataset and 0.920–0.951 in the validation dataset). The sensitivity of the panel (0.900–0.940) significantly outperformed traditional clinical protein biomarkers (0.020–0.240). The panel also exhibited promise in early GC detection, with AUCs of 0.914–0.961 in the discovery dataset and 0.894–0.940 in the validation dataset. Among the identified metabolites, eight were traced differentially expressed in GC and paired adjacent tissues, and two (2-hydroxy-3-methylvalerate, isovalerylcarnitine(C5)) were causally linked with GC in MR analysis.

Conclusions This study identifies promising metabolic signatures for GC diagnosis and develops a reliable diagnostic model. The findings underscore the potential of metabolic analysis for accurate screening and early detection of GC.