Add info Plasma metabolomics offers significant potential for non-invasive biomarker discovery in gastric cancer (GC), yet conventional analytical workflows face challenges in absolute quantification and biological interpretability, hindering clinical translation. We present an innovative multi-phase hybrid framework integrating untargeted metabolomics with relative- and absolute-quantitative targeted metabolomics, coupled with a custom interpretability-driven algorithm for de novo biomarker identification. Metabolic profiling was performed using 1,706 plasma samples from multicenter cohorts. The relative-quantitation phase identified 84 key metabolites significantly enriched in caffeine metabolism and primary bile acid biosynthesis pathways. By applying the custom algorithm to absolute quantitation data, we established a 12-metabolite panel covering multiple functional metabolic modules. Machine learning-based diagnostic models using this signature achieved an area under the curve of 0.951 in validation cohort. Together, our study provides a robust and interpretable framework for translational metabolomics and establishes a GC detection biomarker panel, laying the foundation for future mechanistic research and clinical application.