AI‐Driven Multimodal Transcriptomic Integration Identifies an Immune Microenvironment–Related Prognostic Signature and Exploratory Immune Checkpoint–Associated Subgroups in Colon Adenocarcinoma

Colon adenocarcinoma (COAD) exhibits substantial molecular and immune heterogeneity, which can limit prognostic stratification based on clinicopathological variables alone. The primary objective of this study was to develop an immune microenvironment–related overall‐survival (OS) signature for COAD, with secondary exploratory analyses aimed at characterizing its immune‐cell context using bulk and single‐cell transcriptomic data. RNA‐sequencing and clinical data from The Cancer Genome Atlas (TCGA) were integrated with ESTIMATE immune scores. Differentially expressed genes were subjected to weighted gene co‐expression network analysis (WGCNA), and least absolute shrinkage and selection operator (LASSO) regression was used to construct a 25‐gene prognostic signature. Among 430 patients with usable OS information, 215 were assigned to each risk group according to the median risk score; the high‐risk group had significantly poorer OS ( p < 0.001). Time‐dependent receiver operating characteristic analysis yielded areas under the curve of 0.7363, 0.7946, and 0.8498 at 1, 3, and 5 years, respectively. In multivariable Cox analysis, age (HR = 1.038, 95% CI 1.018–1.059; p < 0.001) and the risk score (HR = 8.737, 95% CI 5.087–15.006; p < 0.001) remained independently associated with OS. CIBERSORT, ssGSEA, and xCell demonstrated subset‐specific differences in immune‐cell estimates rather than a uniform increase in immune infiltration. Eight candidate hub genes (LZTS3, TNNT1, VAMP1, HDC, ARHGAP4, PLAC1, HOXC11, and ITLN1) were further examined, and single‐cell RNA sequencing was used to localize their expression across annotated cell populations. ARHGAP4 and ITLN1 were selected for an exploratory two‐gene consensus clustering analysis because they showed opposite prognostic directions in the LASSO model and immune‐associated cell‐level expression patterns. The resulting clusters differed in immune checkpoint expression; however, because no immunotherapy‐treated cohort or response labels were analyzed, these clusters should be regarded as hypothesis‐generating immune phenotypes rather than evidence of treatment sensitivity. Overall, this study provides an internally evaluated, interpretable multimodal framework for COAD prognostic stratification while defining the additional validation required before clinical translation.