A fibroblast-rich transcriptomic axis reflects stromal remodeling within the CMS4 phenotype in colorectal cancer

Bulk-tumor expression signatures can reflect cellular composition as well as tumor-cell programs. We evaluated a frozen, equal-weight 87-gene score in colorectal cancer using 11 Gene Expression Omnibus (GEO) cohorts, TCGA COAD/READ, and E-MTAB-12862. In GEO (1,783 tumors; 441 progression-related events), the hazard ratio per score standard deviation was 1.33 (95% confidence interval, 1.19-1.49). The estimate was similar in seven disease-free-survival cohorts. In 376 TCGA tumors, the hazard ratio was 1.25 (1.03-1.52) for progression-free interval. The association did not reproduce in E-MTAB-12862: among 948 stage I-III tumors with 436 recurrence-free-survival events, the hazard ratio was 1.05 (0.95-1.16). The frozen score correlated with CMS4, cancer-associated fibroblast, and stromal measures. The stromal-remodeling module produced a GEO estimate similar to the full score, whereas removing that module attenuated the association. Composition residualization did not detect additional prognostic information after accounting for the measured CMS4 and stromal components. Single-cell and spatial analyses localized the gene program to fibroblast-rich compartments. The 87-gene score mainly reflects a fibroblast-rich, matrix-remodeling component of CMS4 rather than an independent prognostic subtype. Its prognostic association varied across cohorts, and current evidence does not support clinical use.