Abstract
Background: Colorectal cancer (CRC) is molecularly heterogeneous across stages, yet most transcriptomic classifiers treat it as static and rarely test whether the importance of their genes changes with disease stage. Methods: A stage-stratified explainable machine-learning pipeline was applied to the Cancer Genome Atlas colon adenocarcinoma cohort (286 tumour, 41 normal). A ten-gene panel was derived by cross-validated stability selection; six classifiers were benchmarked under a leakage-free protocol; and exact SHapley Additive exPlanations (SHAP) were computed within each American Joint Committee on Cancer (AJCC) stage. Findings were tested in three independent cohorts (GSE50760, GSE39582, GSE17536) and corroborated by literature, GEPIA2, and four-platform pathway enrichment. Results: The ten-gene panel classified tumour versus normal tissue with internal AUC up to 1.000 and external AUC 0.969 (95% CI 0.896–1.000; area under the precision-recall curve [AUPRC] 0.980), robust to class imbalance. Stage-stratified SHAP showed a monotonic shift in feature importance across stages (cross-stage Spearman rho 0.66–0.93) that replicated at three of four stages in an independent cohort (rho up to 0.94); ESM1 rose to top-two late-stage importance in both cohorts, consistent with its role in CRC angiogenesis. The panel recovered known cancer biology, nominated novel candidates, and proved complementary to network-hub genes, and its enriched pathways matched four of five frequently altered Saudi CRC pathways. Survival discrimination was at chance (C-index 0.50–0.58), defining the panel as a classifier rather than a prognostic model. Conclusions: This is, to our knowledge, the first evidence in CRC that an explainable classifier’s feature-importance structure is itself stage-dependent and externally replicable. Coupling exact interpretability with reproducible, transparently scoped validation, the panel generalises as a classifier but not a prognostic model, and its convergence with Saudi CRC alterations defines conditions for prospective population-specific validation. Code and processed data are publicly available.