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

Problem Statement

Gastric cancer prognosis is heavily influenced by the tumor microenvironment (TME), particularly the organization of the extracellular matrix (ECM), vascular architecture, and immune cell localization by cancer-associated fibroblasts (CAFs). However, current clinical understanding is limited because bulk stromal abundance measurements collapse biologically distinct fibroblast states into a single signal. There is a lack of compact, reproducible markers that connect specific fibroblast transcriptional states to tissue context and clinical outcomes without overfitting to specific datasets or technical artifacts. The study aims to define a specific, compact fibroblast programme that can be traced across single-cell, spatial, and bulk transcriptomic platforms to identify a niche linked to adverse outcomes.

Methodology

The study employs a multi-layered, cross-platform evidence hierarchy to evaluate an APOD-EFEMP1-COL5A2 three-gene programme. The analysis strictly distinguishes between biological units (cells, spots, samples, cohorts) to avoid pseudo-replication.

1. Discovery Single-Cell Atlas (GSE183904):

   • Integrated 40 gastric cancer samples (156,111 quality-controlled cells).
   • Identified eight major compartments. Fibroblasts were purified using UCell scoring for CAF identity genes (e.g., DCN, LUM, FAP) and exclusion of contamination signatures (epithelial, plasma, immune, endothelial).
   • Defined 2,415 high-confidence CAFs from 36 samples.
   • Programme Definition: Within this high-confidence CAF set, cells were stratified into "Programme-High," "Middle," and "Low" states based on sample-wise quartiles of the mean z-score of APOD, EFEMP1, and COL5A2.

2. Transcriptional Characterization:

   • Paired Pseudobulk Analysis: Aggregated raw counts from 21 samples containing paired high and low CAF states. Used edgeR with a paired design to identify differentially expressed genes (DEGs) and perform functional enrichment (GO, KEGG, Hallmark).
   • Independent Single-Cell Validation (GSE163558): Recalculated the programme score within a separate CAF object (255 cells, 6 samples) to assess detectability and directionality, noting sample concentration limitations.

3. Spatial Transcriptomics (GSE251950 & GSE246011):

   • Analyzed 10 tissue sections (31,202 spots) to test spatial colocalization.
   • Calculated a three-gene spatial expression index (mean of z-transformed genes) and correlated it with CAF-core, endothelial, angiogenesis, and myeloid feature scores within each section.
   • Performed neighbourhood enrichment analysis using permutation tests to assess local association with specific TME features.

4. Bulk Transcriptomic & Clinical Analysis (10 Cohorts):

   • Scored 10 independent bulk cohorts (including TCGA-STAD, GSE15459, GSE26253) using the programme formula.
   • Survival Analysis: Stratified patients by median programme score. Conducted Cox proportional-hazards regression for Overall Survival (OS) and Recurrence-Free Survival (RFS). Performed a random-effects meta-analysis for OS.
   • TME Correlation: Correlated programme scores with stromal, immune-exclusion, and vascular features within each cohort before summarizing.

5. Mechanism & Perturbation Inference:

   • Regulatory Network: Ranked transcription factors (TFs) based on integrated scores linking TF expression to programme activity.
   • Ligand-Receptor Analysis: Used an expression-product score (sender ligand × receiver receptor) to prioritize communication branches, specifically focusing on matrix-integrin and MIF/CD74 pathways.
   • Virtual Perturbation: Simulated knockdowns of programme genes (e.g., COL5A2) and communication nodes (e.g., MIF/CD74) to identify computational vulnerabilities. Drug mapping was limited to knowledge-based concepts due to lack of rankable IC50 data.

Key Results

1. Identification and Localization of the Programme

• The APOD-EFEMP1-COL5A2 programme preferentially localizes to a purified, high-confidence CAF compartment.
• In the discovery atlas, 616 Programme-High CAFs were identified across 34 samples. These cells were not confined to a single cluster but showed preferential enrichment in specific clusters (e.g., Cluster 0, OR 3.51).
• Programme-High CAFs exhibited a coherent transcriptional state characterized by upregulation of ECM, collagen, focal adhesion, and vascular-associated genes.

2. Transcriptional Identity

• Paired pseudobulk analysis (21 samples) identified 261 significant genes (187 up, 74 down).
• Key upregulated pathways included Collagen fibril organization (NES 2.44) and Blood vessel morphogenesis.
• The programme-High state is strongly associated with matrix remodeling, collagen/integrin signaling, and vascular features.

3. Independent Validation

• Single-Cell: The programme was detectable in the independent GSE163558 dataset, though the high-state cells were highly concentrated in a few samples (one sample contributed 75% of high-state cells). Directional consistency was observed for defining genes and ECM features.
• Spatial: In GSE251950, the three-gene spatial index showed positive correlations with CAF-core, endothelial, angiogenesis, and myeloid features across all 10 sections. Neighbourhood analysis confirmed enrichment of these features around high-index spots.

4. Clinical Association

• Survival: Higher programme scores were significantly associated with worse outcomes.
  • TCGA-STAD & GSE15459 (OS): Meta-analysis yielded a pooled Hazard Ratio (HR) of 1.80 (95% CI 1.40–2.33, P = 6.24×10⁻⁶).
  • GSE26253 (RFS): Programme-high tumors had shorter recurrence-free survival (HR 1.59, 95% CI 1.18–2.14, P = 0.0025).
• TME Phenotype: The programme correlated strongly with CAF-core, ECM/matrix remodeling, and stromal immune-exclusion signatures across all 10 bulk cohorts.

5. Prognostic Modeling

• The programme provided modest but reproducible prognostic information.
• Adding the programme score to clinical variables (e.g., age) improved the C-index by 0.060 in TCGA-STAD and 0.103 in GSE15459.
• The programme added incremental value beyond generic CAF features (e.g., CAF-core scores), with positive delta C-index values across cohorts.

6. Mechanistic Hypotheses

• Matrix-Integrin-Vascular Branch: Programme-High CAFs are computationally linked to matrix/collagen ligands (e.g., COL1A1, COL5A2) and integrin receptors (e.g., ITGB1) on endothelial and pericyte/smooth-muscle cells.
• SP1-MIF/CD74-Myeloid Branch: SP1 was identified as a top candidate upstream regulator. The programme is linked to a MIF-CD74 communication axis, with high expression-product scores for myeloid/macrophage receivers.
• Virtual Perturbation: COL5A2 knockdown showed the strongest predicted reversal of ECM remodeling and collagen-integrin signaling. MIF/CD74 blockade simulation prioritized myeloid/macrophage receivers.

Significance and Claims

The paper claims to establish a reproducible, biologically coherent stromal programme that links CAF matrix biology, vascular-associated tissue organization, and myeloid context to adverse gastric cancer outcomes.

• Evidence Boundaries: The study explicitly states it does not establish a universal biomarker, a ready clinical assay, or a causal mechanism. The findings are described as "computational hypotheses" and "supportive evidence" rather than definitive proof.
• Value Proposition: The primary contribution is a compact, three-gene axis that serves as a stable handle for a broader, complex stromal phenotype. It successfully bridges single-cell localisation, spatial context, and bulk clinical outcomes without conflating distinct biological units.
• Translational Potential: The study prioritizes specific testable hypotheses for experimental follow-up:
  1. Validating the matrix-integrin-vascular communication branch.
  2. Testing the SP1-MIF/CD74-myeloid regulatory and communication axis.
  3. Investigating COL5A2 and MIF/CD74 as potential intervention nodes.
• Limitations: The authors acknowledge limitations including the retrospective nature of public data, sample concentration in validation sets, spot-level resolution limits in spatial data, and the computational nature of perturbation and drug-mapping results. No formal drug efficacy or treatment-response predictions are claimed.

In summary, the paper provides a rigorous, multi-layered computational framework that defines an adverse CAF-associated niche in gastric cancer, offering a focused set of hypotheses for future experimental validation regarding matrix remodeling, vascular organization, and myeloid interactions.