Objective: Gastric cancer (GC) exhibits profound heterogeneity, yet the contribution of lactate metabolism reprogramming to this diversity and its cellular basis remain incompletely understood. This study aimed to dissect GC heterogeneity through lactate metabolism-related genes (LMRGs), with a focus on the cellular origins and context-specific functions of key genes. Methods: We performed consensus clustering of TCGA GC samples (n = 375) using a curated set of 49 LMRGs. A multi-step screening strategy was employed to identify hub genes. Single-cell RNA-seq data were integrated to map the cellular sources of key genes. Subtype-specific analyses of mutation, expression, and prognosis were conducted. A prognostic model was constructed, but its cross-platform generalizability was critically evaluated to explore the functional heterogeneity of its constituent genes. Results: We identified two distinct GC subtypes: G1, a glycolytic and immunosuppressive subtype associated with poor prognosis, and G2, an immune-activated subtype with better prognosis. Crucially, single-cell analysis revealed that the hub gene HK2 is enriched in NK cells and pDCs, while FABP4 exhibits a dual cellular origin, being expressed in both proliferative CD8+ T cells and fibroblasts. This dual origin provides a mechanistic basis for the gene’s context-dependent behavior: while FABP4 appeared protective in overall models, it acted as a significant risk factor within the G2 subtype (HR = 1.71, p = 0.017) and in an external validation cohort (HR = 2.64). A derived prognostic model failed external cross-platform validation, a phenomenon driven by the reversal of risk effects for FABP4 and other genes across different populations. Conclusions: This study uncovers two distinct metabolic-immune subtypes of GC and demonstrates that the prognostic effect of FABP4 is not fixed but is highly dependent on its cellular source and the tumor microenvironmental context. Our findings generate a testable hypothesis regarding FABP4’s role in balancing anti-tumor immunity and stromal promotion. More broadly, the failure of our cross-platform model serves as a cautionary tale on the limitations of single-cohort-derived signatures and underscores the necessity of integrating single-cell resolution to unravel the biological complexity underlying prognostic biomarkers.
Xiaoxuan et al. (2026) studied this question.