The epithelial-mesenchymal transition (EMT) is an important process during metastasis in various tumors, including colorectal cancer (CRC). Thus, the study of its characteristics and related genes is of great significance for CRC treatment. In this study, 26 EMT-related gene sets were used to score each sample from The Cancer Genome Atlas program (TCGA) colon adenocarcinoma (COAD) database. Based on the 26 EMT enrichment scores for each sample, we performed unsupervised cluster analysis and classified the TCGA-COAD samples into three EMT clusters. Then, weighted gene co-expression network analysis (WGCNA) was used to investigate the gene modules that were significantly associated with these three EMT clusters. Two gene modules that were strongly positively correlated with the EMT cluster 2 (worst prognosis) were subjected to Cox regression and least absolute shrinkage and selection operator (LASSO) regression analysis. Then, a prognosis-related risk model composed of three hub genes GPRC5B, LSAMP, and PDGFRA was established. The TCGA rectal adenocarcinoma (READ) dataset and a CRC dataset from the Gene Expression Omnibus (GEO) were used as the validation sets. A novel nomogram that incorporated the risk model and clinicopathological features was developed to predict the clinical outcomes of the COAD patients. The risk model served as an independent prognostic factor. It showed good predictive power for overall survival (OS), immunotherapy efficacy, and drug sensitivity in the COAD patients. Our study provides a comprehensive evaluation of the clinical relevance of this three-gene risk model for COAD patients and a deeper understanding of the role of EMT-related genes in COAD.

Background: Prostate cancer (PCa) is a leading cause of cancer-related mortality worldwide, highlighting the need for improved prognostic tools. The integration of artificial intelligence (AI) and machine learning (ML) with multi-omics data offers new opportunities for biomarker discovery and risk stratification. Methods: We integrated bulk transcriptomic data from GSE116918 (training, n = 248) and three cross-cohort consistency evaluation cohorts (TCGA-PRAD, GSE70769, GSE46602), focusing on 1087 epithelial–mesenchymal transition (EMT)-associated genes. Using consensus clustering, weighted gene co-expression network analysis (WGCNA), and 91 machine learning algorithm combinations (including Random Forest, Lasso, and CoxBoost), we constructed a prognostic signature. SHAP analysis was used for model interpretability. Single-cell RNA sequencing (scRNA-seq, GSE268307, 10,672 cells) and spatial transcriptomics (10× Genomics Visium FFPE) provided hypothesis-generating evidence; spatial analysis was based on one tissue section. Results: A three-gene signature (INHBA, FAP, ITGBL1) effectively stratified patients into high- and low-risk groups, with the high-risk group showing significantly worse metastasis-free survival (HR = 1.61, 95% CI: 1.39–1.87; 4-year AUC = 0.93 in the training cohort; external AUCs ranged from 0.62 to 0.77). CytoTRACE inferred high differentiation potential of COMP+ fibroblasts, and Monocle3 inferred a transcriptional transition from COMP+ toward NELL2+ fibroblasts. BayesPrism deconvolution suggested that high inferred COMP+ fibroblast abundance was associated with poor prognosis and advanced T stage. NicheNet analysis prioritized BMP7 as a key upstream ligand, with downstream targets enriched in TGF-β signaling and stem cell pluripotency pathways. Conclusions: This study presents a machine learning-based multi-omics framework for prostate cancer risk stratification. The three-gene signature provides a new exploratory prognostic model while inferring a COMP+ to NELL2+ transcriptional transition. These findings may inform future hypothesis-driven studies of treatment sensitivity, pending experimental validation, and demonstrate the value of AI-driven multi-omics integration for precision oncology.