An APOD-EFEMP1-COL5A2 fibroblast programme marks a matrix-remodelling, vascular and myeloid-associated niche linked to adverse gastric cancer outcomes

Gastric cancer stromal niches influence matrix architecture, vascular organisation and immune-cell localisation, but compact markers that connect fibroblast state, tissue context and clinical outcome remain incompletely defined. We integrated single-cell RNA sequencing, spatial transcriptomics and multicohort bulk transcriptomic data to evaluate an APOD-EFEMP1-COL5A2 fibroblast programme in gastric cancer. In a discovery atlas of 40 samples and 156,111 quality-controlled cells, the programme preferentially localised to a purified fibroblast/cancer-associated fibroblast compartment containing 2,415 high-confidence CAFs. Within this compartment, 616 programme-high, 1,177 programme-middle and 622 programme-low CAFs were identified, with programme-high cells represented in 34 samples. Paired pseudobulk analysis in 21 samples detected 261 significant genes, including 187 upregulated and 74 downregulated genes, and supported matrix, collagen, focal-adhesion and vascular-associated biology. Independent single-cell and spatial datasets provided supportive but sample- and resolution-bounded evidence, including GSE163558 CAF validation and ten GSE251950 tissue sections with 31,202 spots. An OS-only random-effects meta-analysis of TCGA-STAD and GSE15459 associated a higher programme score with worse overall survival (pooled HR 1.80, 95% CI 1.40–2.33, P = 6.24×10^-6, I2 = 0%), while GSE26253 separately supported an association with shorter recurrence-free survival (HR 1.59, 95% CI 1.18–2.14, P = 0.00249). Prognostic modelling showed modest but reproducible discrimination and incremental information beyond available clinical and generic CAF features. Mechanism analyses nominated candidate matrix-integrin-vascular and SP1-MIF/CD74-myeloid branches, while virtual perturbation prioritised COL5A2 and MIF/CD74-related concepts. The study therefore links a reproducible CAF-associated programme to tissue organisation, outcome association and testable mechanism hypotheses across complementary public datasets. These perturbation and intervention results are computational hypotheses and require experimental validation before mechanistic or therapeutic interpretation.