An epithelial mesenchymal transition associated gene signature stratifies risk of ulcerative colitis associated colorectal cancer Authors: Nevin Belder Published: 2026 ABSTRACT BackgroundPatients with long-standing ulcerative colitis (UC) are at increased risk of developing colitis-associated colorectal cancer (UC-CRC). However, the molecular transition from chronic inflammation to neoplasia remains incompletely defined, and biomarkers for high-risk disease are still lacking.MethodsTranscriptomic datasets from UC and CRC cohorts were analyzed independently to identify reproducible transcriptional alterations. Differential expression and pathway enrichment analyses were followed by stepwise prioritization of candidate genes based on expression in UC-associated neoplasia, prognostic relevance in CRC, and classification performance. Receiver operating characteristic analysis and composite z-score modeling were used to evaluate signature performance, whereas gene co-expression analysis and single-sample gene set enrichment analysis were used to examine its biological context.ResultsEpithelial–mesenchymal transition (EMT) emerged as a consistently enriched pathway among upregulated genes in UC and CRC. Intersection of EMT-associated genes identified 24 shared candidates, of which 14 were significantly upregulated in UC-associated neoplasia compared with quiescent UC and showed consistent upregulation across independent active UC datasets. These genes showed strong discriminatory performance (AUC > 0.85), and Random Forest analysis further refined a seven-gene panel (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1). A composite z-score derived from this panel distinguished neoplastic from quiescent UC and showed a stepwise increase across disease states. Functional analyses linked the signature to EMT, hypoxia, angiogenesis, coagulation, and inflammatory signaling.ConclusionsThis study identified a seven-gene EMT-associated signature linked to UC-associated neoplastic transformation. The signature may support molecular risk stratification in UC and provides a framework for future validation in surveillance settings.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05330-4. Background Ulcerative colitis (UC) is a major subtype of inflammatory bowel disease (IBD), characterized by chronic, relapsing inflammation primarily affecting the colon and rectum [1]. Although its precise etiology remains incompletely understood, UC is widely recognized as a multifactorial disorder involving genetic susceptibility, immune dysregulation, microbial imbalance, epithelial barrier dysfunction, and environmental influences. Clinically, inflammation typically originates in the rectum and extends proximally in a continuous pattern, with bloody diarrhea as a hallmark symptom. Diagnosis is established through endoscopic assessment supported by histopathological evaluation [1–3]. Prolonged inflammatory exposure in UC promotes structural remodeling of the colon, characterized by the development of pseudopolyps, strictures, and progressive mucosal shortening. These cumulative alterations are thought to create a permissive environment for neoplastic transformation and are associated with a markedly increased risk of colorectal cancer (CRC) [4]. Colorectal cancer represents one of the most prevalent malignancies and a major global health burden, ranking third in incidence and second in cancer-related mortality worldwide [5]. Given the recent 30-fold surge in IBD incidence, the anticipated rise in UC-associated colorectal cancer (UC-CRC) has become a pressing oncological concern [6]. Ulcerative colitis–associated colorectal cancer is considered the most severe complication of inflammatory bowel disease, arising through a distinct inflammation–dysplasia–carcinoma sequence that serves as a compelling model for inflammation-driven tumorigenesis. In this setting, dysplastic lesions typically evolve from indefinite to low- and then high-grade dysplasia before progressing to invasive cancer, although direct transformation from low-grade dysplasia without a high-grade intermediate can also occur. Unlike sporadic adenomas, these lesions are frequently flat and poorly demarcated, posing significant challenges for early endoscopic detection [4, 7]. While UC-CRC accounts for only 1–2% of all colorectal cancer cases in the general population, it is responsible for approximately 10–15% of IBD-related mortality [8]. Despite improvements in surveillance and medical therapy, UC-CRCs remain clinically aggressive, with patients exhibiting poorer prognoses and distinct molecular profiles compared to those with sporadic CRC [4, 7]. Furthermore, the absence of reliable biomarkers that predict malignant transformation remains a significant clinical gap. Elucidating the molecular mechanisms that govern the transition from chronic inflammation in UC to colorectal malignancy is crucial for identifying robust diagnostic, prognostic, and therapeutic targets [7]. Changes in gene expression reflect dynamic shifts in cellular function and molecular activity. Advances in high-throughput transcriptomic profiling, including microarray and RNA sequencing technologies coupled with sophisticated bioinformatics approaches, have enabled the systematic identification of molecular markers and prognostic gene signatures associated with malignant transformation [9–13]. In this context, several studies have reported shared molecular features between UC and CRC. For instance, TIMP1, which is consistently overexpressed during the initiation and progression of ulcerative colitis–associated colorectal cancer, has been identified as a potential prognostic biomarker in colorectal cancer, underscoring its relevance in inflammation-driven tumorigenesis [14]. Guan et al. (2015) employed a multidimensional systems biology framework and their integrative systems biology analyses have further identified overlapping genes, transcription factors, and microRNAs that may bridge chronic inflammation and malignancy in UC [15]. More recently, integrative transcriptomic studies combining GEO and TCGA datasets have identified candidate drivers such as CCT2, reinforcing the notion that UC and CRC share convergent molecular programs [16]. Nevertheless, despite these advances, robust and clinically actionable biomarkers capable of stratifying UC patients by neoplastic risk remain lacking. To address this unmet clinical need, the transcriptomic landscape of the colitis-to-cancer continuum was systematically characterized through comparative analysis of multiple independent datasets. This analytical approach highlighted epithelial mesenchymal transition related signaling as a prominently enriched pathway associated with colitis-associated malignant transformation. Building upon this biological framework, a seven-gene signature (FAP, TIMP1, SPP1, INHBA, COL12A1, CALU, and CTHRC1) was identified and validated for its ability to distinguish neoplastic UC from quiescent disease. Functional enrichment analyses further demonstrated that high-risk samples showed enrichment of tumorigenic pathways, including hypoxia, TGF-β signaling, and inflammatory responses. Collectively, these findings highlight a molecular framework that may support risk stratification in UC and provide a basis for future studies aimed at improving early detection and surveillance strategies. Methods This study employed a systematic bioinformatics analysis of publicly available microarray datasets obtained from the Gene Expression Omnibus (GEO) database [17]. The analyses were designed to identify transcriptomic signatures associated with neoplastic risk in ulcerative colitis (UC). An overview of the analytical workflow is shown in Fig. 1.Fig. 1Study workflow. Publicly available transcriptomic datasets from ulcerative colitis (UC) and colorectal cancer (CRC) cohorts were analyzed independently to identify significantly upregulated genes in each disease. Pathway enrichment analysis performed on these gene sets identified epithelial–mesenchymal transition (EMT) as the most consistently enriched shared pathway. Intersection of EMT-associated upregulated genes across UC and CRC yielded a core set of 24 candidates. These genes were validated in independent cohorts and further evaluated in UC-associated neoplasia for model construction. Machine learning–based prioritization and ROC analysis defined a seven-gene panel, which was subsequently assessed using z-score modeling, co-expression analysis, and single-sample gene set enrichment analysis to characterize its functional and biological relevance Study workflow. Publicly available transcriptomic datasets from ulcerative colitis (UC) and colorectal cancer (CRC) cohorts were analyzed independently to identify significantly upregulated genes in each disease. Pathway enrichment analysis performed on these gene sets identified epithelial–mesenchymal transition (EMT) as the most consistently enriched shared pathway. Intersection of EMT-associated upregulated genes across UC and CRC yielded a core set of 24 candidates. These genes were validated in independent cohorts and further evaluated in UC-associated neoplasia for model construction. Machine learning–based prioritization and ROC analysis defined a seven-gene panel, which was subsequently assessed using z-score modeling, co-expression analysis, and single-sample gene set enrichment analysis to characterize its functional and biological relevance Data acquisition Publicly available gene expression datasets related to colorectal cancer (CRC) and ulcerative colitis (UC) were retrieved from the GEO database. Dataset selection was based on predefined criteria to ensure biological relevance and analytical consistency. Eligible datasets were required to include (i) clearly defined case–control or disease-state comparisons (e.g., UC vs control, CRC vs normal, or UC-associated neoplasia vs quiescent UC), (ii) Homo sapiens samples, (iii) adequate sample size, and (iv) availability of relevant clinical annotation. Datasets with unclear group definitions or limited sample size were excluded. Platform compatibility was also considered to minimize technical variability across analyses, particularly during differential expression analysis. The CRC datasets included: GSE110224 (GPL570), comprising 17 CRC samples and 17 matched normal tissues; GSE22598 (GPL570), containing 17 paired cancer and adjacent non-cancerous tissues; and GSE21815 (GPL6480), with 132 CRC samples and nine normal colon epithelial tissues. The UC datasets were: GSE83687 (GPL16791), including 32 UC samples and 60 controls; GSE92415 (GPL13158), with 162 UC samples and 21 controls; and GSE87473 (GPL13158), comprising 106 UC samples and 21 controls. A curated epithelial-mesenchymal transition (EMT) gene set consisting of 200 genes was obtained from the Hallmark EMT collection of the Molecular Signatures Database (MSigDB) [18, 19]. Data preprocessing and identification of differentially expressed genes Raw gene expression data were processed using the GEO2R web tool, which utilizes the limma (Linear Models for Microarray Data) package (v3.26.8) within the R/Bioconductor framework. Differentially expressed genes (DEGs) between disease groups (CRC or UC) and corresponding control samples were identified using thresholds of |log2 fold change|> 0.58 and an adjusted p-value < 0.05, with multiple testing correction performed using the Benjamini–Hochberg method. These criteria were selected to balance the detection of subtle but biologically relevant changes. DEGs were identified within each dataset separately, and no cross-platform merging of raw expression data was performed. Common signatures were subsequently derived by intersecting statistically significant gene sets across cohorts. This strategy was adopted to minimize potential batch effects and platform-specific biases. Overlapping DEGs, as well as commonly upregulated and downregulated genes within each disease, were identified using Venny 2.11. To further characterize the transcriptomic landscape, volcano plots were used to visualize the distribution of differential expression signals. Gene expression patterns across samples were assessed by unsupervised hierarchical clustering using average linkage and Euclidean distance, and visualized as heatmaps. These analyses and visualizations were performed in R using ggplot2 and pheatmap. Functional enrichment analysis and identification of epithelial–mesenchymal transition pathway for gene signature development Functional enrichment analysis was performed using Enrichr [20] with the Hallmark gene sets from the Molecular Signatures Database. Analyses were performed separately for the UC and CRC datasets using upregulated DEGs. Upregulated genes were prioritized to identify shared activated biological programs in both diseases, given their relevance to disease progression and potential therapeutic targeting. Significantly enriched pathways (adjusted p < 0.05) were identified for each disease, and pathways shared between UC and CRC were identified by Venn diagram analysis. Hallmark pathway enrichment was visualized as horizontal bar plots based on –log₁₀(adjusted p-values) in R. Enrichment analysis revealed the epithelial-mesenchymal transition (EMT) pathway as the most significantly enriched shared Hallmark pathway. Considering its consistent association with inflammation-driven remodeling and neoplastic progression, the EMT pathway was prioritized for downstream gene signature development. To refine the EMT-related gene set, genes that were both significantly upregulated and included in the Hallmark EMT pathway were identified in UC and CRC, and their intersection yielded 24 shared genes. Cross-validation of EMT-related gene candidates in independent datasets As the analysis focused on upregulated genes, only candidates showing consistent overexpression across disease contexts were retained for validation. To assess the robustness of these expression patterns, transcriptional profiles of 24 EMT-related genes were examined across four independent datasets, including two colorectal cancer tumor-normal cohorts (GSE5206 and GSE32323) and two ulcerative colitis versus control cohorts (GSE47908 and GSE59071). Group-wise comparisons were performed using Student’s t-test or the Mann–Whitney U test, according to distributional properties of the data. Results were visualized as boxplots generated in GraphPad Prism (v8.4.2). Protein–protein interaction (PPI) analysis was performed for the 24 EMT-associated genes using the STRING database (version 12.0) [21]. Analyses were restricted to Homo sapiens, and only interactions with a confidence score greater than 0.700 were included. Network connectivity was evaluated using the STRING PPI enrichment p-value, which tests whether the observed number of interactions exceeds that expected for a random protein set of comparable size. Gene selection and model construction Gene expression comparison between UC-neoplasia and quiescent UC To identify genes specifically associated with malignant transformation, the expression levels of the 24 EMT-related candidates were compared between UC-associated neoplasia and quiescent UC samples using the GSE37283 dataset. Comparisons were performed using Student’s t-test for approximately normally distributed data and the Mann–Whitney U test otherwise. The final visualization was generated using the ComplexHeatmap package in R. Data were row-scaled using Z-score transformation to highlight intra-gene expression dynamics across the disease continuum. Hierarchical clustering was applied to rows using Euclidean distance and the complete linkage method to identify functionally co-regulated gene modules. Genes with significantly higher expression in UC-associated neoplasia (p < 0.05) were retained for subsequent analyses, yielding a subset of 14 genes. Validation of EMT-related genes in active UC To further assess the disease relevance of the 14 candidate genes identified in the UC-neoplasia versus quiescent UC comparison, their expression profiles were examined in two independent datasets including control mucosa, inactive UC, and active UC samples (GSE179285 and GSE75214). This validation step was designed to determine whether these genes showed consistent overexpression during the transition from quiescent to active inflammatory states. Normalized expression values for each candidate gene were extracted from the respective cohorts. Group-based statistical comparisons were performed in R. Statistical significance was defined as p < 0.05. Prognostic relevance of EMT-related gene candidates To evaluate the association between gene expression and clinical outcome, univariate Cox proportional hazards regression analyses were performed for the 14 candidate genes using the GSE17536 colorectal cancer cohort. Genes with p-values < 0.05 in univariate analysis were retained for further evaluation. Multivariate Cox regression analyses were subsequently conducted to assess independence of associations, adjusting for available clinical covariates including age, sex, and tumor stage. The clinical impact of these genes was further characterized through Kaplan–Meier survival analysis using the survminer package in R (v4.2.0). Patients were dichotomized into high- and low-expression groups based on the median transcript level, and the log-rank test was used to assess differences in survival outcomes. Genes that showed significant survival separation (p < 0.05) across CRC datasets were considered to have additional clinical relevance. Furthermore, the association between these candidate genes and disease recurrence was examined by comparing their expression levels between recurrent and non-recurrent CRC patients. ROC analysis for UC neoplasia classification The ability of each gene identified through prognostic and recurrence analyses (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, CTHRC1, and COL4A1) to distinguish UC-associated neoplasia from quiescent UC was assessed using receiver operating characteristic (ROC) curve analysis. ROC analysis was performed in R using pROC, and the area under the curve (AUC) was calculated for each gene to quantify its individual discriminatory performance. All genes met the predefined performance criterion (AUC > 0.85) and were therefore retained for subsequent analyses. Random forest classification and final gene panel selection To refine the candidate set into a high-performance predictive panel, a Random Forest classification model was constructed using the randomForest package in R. The binary outcome variable was defined as "Neoplasia" versus "Quiescent", and gene-expression levels of the nine candidate genes were used as input features. The model was trained with 1000 trees (ntree = 1000) and default parameters. Feature importance was evaluated using the Mean Decrease in Gini index. Genes with a MeanDecreaseGini ≥ 0.5 were considered to contribute meaningfully to the model. Default parameters were retained given the exploratory nature of the analysis and the limited sample size, to avoid overfitting. Based on these criteria, COL4A1 was excluded from the final framework as it fell below the importance threshold. This process resulted in a final seven-gene panel comprising FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1, which was selected for downstream modeling and validation. Z-score modeling and validation in CRC datasets Z-score signature construction For the UC-associated neoplasia analysis and model construction, datasets enabling stepwise comparison across disease states, including normal mucosa, quiescent UC, and UC-associated neoplasia, were prioritized. GSE37283 was selected as the core dataset, as it enables comparison across disease states under conditions of minimal inflammatory activity. To evaluate the combined activity of the selected seven-gene panel (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1), a composite z-score–based signature was used on the GSE37283 dataset, which includes both quiescent UC and UC-associated neoplasia samples. For each sample, individual gene expression values were standardized via z-score transformation: z = (x − μ) / σ where x represents the expression level of a specific gene, while μ and σ denote the mean and standard deviation of that gene across the reference cohort (GSE37283). An integrated composite score (MeanZ) was subsequently derived for each sample by calculating the arithmetic mean of the standardized values across the seven genes. This composite score was used to stratify samples into risk categories and assess discriminative performance. Group comparisons of MeanZ between quiescent UC and neoplastic UC samples were performed using the Mann–Whitney U test. Results were visualized using box-and-jitter plots, with two-tailed p-values indicated. All analyses and visualizations were performed using R (v4.2.0). Furthermore, ROC curve analysis was performed in R to evaluate the signature's ability to distinguish UC-associated neoplasia from quiescent UC. The area under the curve was used to quantify the model’s classification accuracy and clinical utility. External validation in colorectal cancer cohorts To assess the generalizability of the seven-gene z-score signature trained on the GSE37283 dataset, the model was applied to three independent CRC datasets: GSE15781, GSE39582, and GSE37364. Expression values were standardized using gene-specific means and standard deviations derived from the training cohort, preserving model consistency without dataset-specific recalibration. A composite z-score (MeanZ) was subsequently calculated by averaging per-sample z-scores across all seven genes. Group-wise comparisons of MeanZ were performed using the Mann–Whitney U test (GSE15781 and GSE39582) or Kruskal–Wallis test followed by pairwise Wilcoxon tests (GSE37364). The discriminative performance of the signature was independently validated through ROC analysis and AUC computation. All statistical procedures and visualizations were executed in R. Gene co-expression analysis in UC and CRC cohorts To evaluate the coherence and potential modular behavior of the seven-gene signature, co-expression analysis was performed across inflammatory, neoplastic, and tumor contexts. This assessment aimed to delineate the internal transcriptional relationships among the signature components (FAP, TIMP1, SPP1, INHBA, COL12A1, CALU, and CTHRC1) during disease progression. Pairwise co-expression patterns were quantified using Pearson correlation coefficients. In the GSE37283 dataset, correlations were computed independently for quiescent UC and UC-associated neoplasia samples to capture transition-specific regulatory shifts. The analysis was further extended to CRC tumor samples in the GSE17536 and GSE39582 cohorts to validate the stability of the co-expression structure in advanced neoplastic contexts. Correlation matrices were constructed based on pairwise complete observations and visualized as circular correlograms using the ggcorrplot package in R. Single-sample gene set enrichment analysis (ssGSEA) To evaluate the functional relevance of the z-score–based risk model, single-sample gene set enrichment analysis (ssGSEA) was performed on GSE17536 (CRC) and GSE47908 (UC) using Hallmark gene sets from MSigDB. Samples were stratified into high- and low-risk groups based on the median z-score. Differentially enriched pathways were identified using the Wilcoxon rank-sum test (p < 0.05) and visualized as clustered heatmaps. Spearman correlations between MeanZ scores and ssGSEA scores were computed and visualized, with only statistically significant associations (p < 0.05) displayed. Analyses and visualizations were performed in R using GSVA, pheatmap, and corrplot. Statistical analysis All statistical analyses were performed using R (v4.2.0) and GraphPad Prism (v8.4.2). Group-wise comparisons were conducted using Student’s t-test or Mann–Whitney U test, as appropriate, based on assessment of data normality using the Shapiro–Wilk test. Multi-group comparisons employed Kruskal–Wallis and Wilcoxon rank-sum tests. ROC analyses were conducted with the pROC package. Gene–gene correlations used Pearson coefficients and were visualized with ggcorrplot. Statistical significance was set at p < 0.05. Results Comparative transcriptomic profiling of ulcerative colitis and colorectal cancer To delineate the transcriptomic landscape of ulcerative colitis (UC) and colorectal cancer (CRC), six microarray datasets were analyzed independently (UC: GSE83687, GSE92415, GSE87473; CRC: GSE110224, GSE22598, GSE21815). Using predefined filtering criteria (adjusted p < 0.05 and |log2FC|≥ 0.58), differentially expressed genes (DEGs) were identified across all cohorts. In the ulcerative colitis datasets, 6633 differentially expressed genes (DEGs) were identified in GSE83687 (3391 upregulated; 3242 downregulated), 3853 DEGs in GSE92415 (2224 upregulated; 1629 downregulated), and 3707 DEGs in GSE87473 (2141 upregulated; 1566 downregulated). For colorectal cancer, GSE110224 yielded 2216 DEGs (1171 upregulated; 1045 downregulated), GSE22598 yielded 5622 DEGs (3053 upregulated; 2569 downregulated), and GSE21815 yielded 7010 DEGs (3794 upregulated; 3216 downregulated) (Supp. Table 1). The distribution of DEGs was visualized using volcano plots (Supp. Figure 1 and 2A-C), where genes meeting the significance thresholds are shown in red (upregulated) and green (downregulated) for CRC and UC, respectively. Additionally, unsupervised hierarchical clustering of DEGs (Supp. Figure 1 and 2A-C) demonstrated clear separation between disease and control samples across both UC and CRC cohorts. To further characterize transcriptional alterations within each condition, the top ten upregulated and downregulated genes were identified for each dataset based on adjusted log₂ fold change. In CRC, genes such as MMP7, DPEP1, and CEMIP were recurrently upregulated, while CLCA4, ZG16, and CA1 were frequently downregulated (Supp. Table 2). In UC, MMP3, CHI3L1, and DUOX2 emerged among the top upregulated genes, whereas AQP8, CLDN8, and SLC51A were consistently downregulated (Supp. Table 3). Cross-disease comparison highlighted MMP3 and AQP8 as shared top deregulated genes, alongside epithelial barrier–related genes (CLDN8, CA1, GUCA2A) that were commonly downregulated, pointing to overlapping mechanisms in CRC and UC pathogenesis. Across datasets, 1144 DEGs were consistently shared among the CRC cohorts, while 1,708 DEGs were detected in all UC datasets (Supp. Figure 3A–B). Given their potential biological relevance, subsequent analyses focused on genes that were consistently upregulated across datasets within each disease. This approach identified 575 commonly upregulated genes in CRC (Fig. 2A) and 1,014 commonly upregulated genes in UC (Fig. 2C).Fig. 2Commonly upregulated genes and pathway enrichment in CRC and UC. A Venn diagram showing overlap of significantly upregulated genes across three colorectal cancer datasets. B Hallmark pathway enrichment analysis of the shared upregulated CRC genes (n = 575). C Venn diagram depicting commonly upregulated genes across three ulcerative colitis datasets. D Enriched pathways among the shared upregulated UC genes (n = 1014), highlighting epithelial–mesenchymal transition as a major shared process Commonly upregulated genes and pathway enrichment in CRC and UC. A Venn diagram showing overlap of significantly upregulated genes across three colorectal cancer datasets. B Hallmark pathway enrichment analysis of the shared upregulated CRC genes (n = 575). C Venn diagram depicting commonly upregulated genes across three ulcerative colitis datasets. D Enriched pathways among the shared upregulated UC genes (n = 1014), highlighting epithelial–mesenchymal transition as a major shared process Epithelial–mesenchymal transition is a commonly enriched pathway among upregulated genes in UC and CRC Significantly upregulated genes in CRC and UC were independently analyzed using Enrichr with the Hallmark gene sets (MSigDB), revealing disease-specific enriched biological processes. Enrichment analysis of upregulated genes in CRC revealed significant overrepresentation of proliferation- and cell cycle–associated pathways, including E2F targets, G2M checkpoint, and MYC targets. Additional enrichment was observed in mTORC1 signaling and epithelial–mesenchymal transition (EMT), suggesting activation of both proliferative and invasive transcriptional programs in colorectal cancer (Fig. 2B). In ulcerative colitis, epithelial–mesenchymal transition emerged as the most significantly enriched pathway among upregulated genes. Additionally, enrichment analysis highlighted strong activation of immune and inflammatory responses, including interferon alpha and gamma signaling, inflammatory response, and TNF-α signaling via NF-κB. Further enrichment was observed in IL6–JAK–STAT3 signaling, hypoxia, mTORC1 signaling, and allograft rejection, reflecting the coexistence of immune dysregulation, epithelial remodeling, and metabolic stress responses characteristic of UC (Fig. 2D). Comparative enrichment analysis identified 19 molecular pathways that were commonly activated across both diseases (Supp. Figure 4, Table 1). Notably, epithelial–mesenchymal transition, inflammatory response, TNF-α signaling via NF-κB, interferon alpha/gamma responses, IL-2/STAT5 signaling, hypoxia, and mTORC1 signaling were significantly enriched in upregulated gene sets from both diseases (Table 1). These overlapping signatures suggest a shared transcriptional landscape characterized by inflammation, immune dysregulation, epithelial remodeling, and stress signaling. Moreover, several genes, including INHBA, CXCL1, CXCL8, MMP1, MMP3, COL1A1 and S100A8 were commonly upregulated across datasets, supporting convergence of inflammatory and tissue-remodeling pathways in UC and CRC (Table 1).Table 1Common Molecular Pathways and Genes Implicated in Both Ulcerative Colitis and Colorectal CancerTermUC Adjusted P-valueCRC Adjusted P-valueGene SymbolsEpithelial Mesenchymal Transition2.29177E-575.94696E-12SPARC, CXCL8, SERPINE1, COL12A1, CXCL1, LAMC1, LOXL2, SPP1, PMEPA1, TIMP1, TGM2, CTHRC1, MMP1, MMP3, INHBA, COL1A1, COL1A2, FAP, COL5A1, COL4A1, COL5A2, CALU, TGFBI, CD44Interferon Gamma Response2.80207E-470.012460791IFITM3, IFITM2, RIPK2, STAT1, PSMB9, CXCL10, CXCL11, MTHFD2, NAMPT, EPSTI1, CFBInflammatory Response4.38209E-432.44057E-05IFITM1, CXCL8, TNFSF15, RIPK2, RGS16, SERPINE1, OSM, INHBA, CXCL10, CXCL11, GNA15, NAMPT, PDPN, TIMP1TNF-alpha Signaling via NF-kB5.7588E-380.000746215CEBPB, RIPK2, SERPINE1, CXCL1, INHBA, CXCL10, CXCL11, DRAM1, PLAU, NAMPT, PMEPA1, FJX1, PHLDA1, CD44Complement1.35077E-260.02592177MMP12, CEBPB, SERPINE1, GZMB, CXCL1, TIMP1, CFB, C2, PSMB9Allograft Rejection1.35077E-260.049496425RIPK2, STAT1, GZMB, TIMP1, INHBA, C2KRAS Signaling Up7.75195E-200.005161851RGS16, INHBA, CXCL10, TMEM158, PLAU, SPP1, CFBInterferon Alpha Response2.28708E-170.000375139IFITM3, CXCL10, IFITM1, CXCL11, IFITM2, RIPK2, EPSTI1, PSMB9Coagulation8.5704E-164.32792E-05SPARC, MMP7, MMP1, SERPINE1, CFI, MMP3, C2, DUSP14, KLF7, PLAU, TIMP1, CFBIL-2/STAT5 Signaling9.00698E-150.005161851IFITM3, RGS16, CXCL10, SPP1, PHLDA1, CD44, TGM2Angiogenesis1.94133E-110.001574012COL5A2, SPP1, TIMP1Apoptosis8.45082E-110.014830538IFITM3, GNA15, LEF1, PMAIP1, TIMP1, CD44, CDC25BUV Response Dn9.62168E-110.044262571COL1A1, COL1A2, SERPINE1, COL5A2, LAMC1Hypoxia2.12418E-080.049496425SLC6A6, KLF7, COL5A1, STC2, SERPINE1, TGFBI, TGM2mTORC1 Signaling2.12418E-085.94696E-12HSP90B1, NAMPT, RRM2, ELOVL5, SLC6A6, SLC7A5, MTHFD2, PSAT1, SCDEstrogen Response Late8.877E-074.83566E-07ELOVL5, TMPRSS3, SNX10, KLK10, SLC7A5, CD44Estrogen Response Early2.75189E-061.86256E-06ELOVL5, TMPRSS3, KLK10, SLC7A5, STC2, SLC39A6, PMAIP1, CD44, TGM2Androgen Response0.0001009830.016721127ELOVL5, SCD, PMEPA1Unfolded Protein Response0.000408454.17426E-06CEBPB, HSP90B1, SLC7A5, MTHFD2, PSAT1, STC2 Table 1Common Molecular Pathways and Genes Implicated in Both Ulcerative Colitis and Colorectal CancerTermUC Adjusted P-valueCRC Adjusted P-valueGene SymbolsEpithelial Mesenchymal Transition2.29177E-575.94696E-12SPARC, CXCL8, SERPINE1, COL12A1, CXCL1, LAMC1, LOXL2, SPP1, PMEPA1, TIMP1, TGM2, CTHRC1, MMP1, MMP3, INHBA, COL1A1, COL1A2, FAP, COL5A1, COL4A1, COL5A2, CALU, TGFBI, CD44Interferon Gamma Response2.80207E-470.012460791IFITM3, IFITM2, RIPK2, STAT1, PSMB9, CXCL10, CXCL11, MTHFD2, NAMPT, EPSTI1, CFBInflammatory Response4.38209E-432.44057E-05IFITM1, CXCL8, TNFSF15, RIPK2, RGS16, SERPINE1, OSM, INHBA, CXCL10, CXCL11, GNA15, NAMPT, PDPN, TIMP1TNF-alpha Signaling via NF-kB5.7588E-380.000746215CEBPB, RIPK2, SERPINE1, CXCL1, INHBA, CXCL10, CXCL11, DRAM1, PLAU, NAMPT, PMEPA1, FJX1, PHLDA1, CD44Complement1.35077E-260.02592177MMP12, CEBPB, SERPINE1, GZMB, CXCL1, TIMP1, CFB, C2, PSMB9Allograft Rejection1.35077E-260.049496425RIPK2, STAT1, GZMB, TIMP1, INHBA, C2KRAS Signaling Up7.75195E-200.005161851RGS16, INHBA, CXCL10, TMEM158, PLAU, SPP1, CFBInterferon Alpha Response2.28708E-170.000375139IFITM3, CXCL10, IFITM1, CXCL11, IFITM2, RIPK2, EPSTI1, PSMB9Coagulation8.5704E-164.32792E-05SPARC, MMP7, MMP1, SERPINE1, CFI, MMP3, C2, DUSP14, KLF7, PLAU, TIMP1, CFBIL-2/STAT5 Signaling9.00698E-150.005161851IFITM3, RGS16, CXCL10, SPP1, PHLDA1, CD44, TGM2Angiogenesis1.94133E-110.001574012COL5A2, SPP1, TIMP1Apoptosis8.45082E-110.014830538IFITM3, GNA15, LEF1, PMAIP1, TIMP1, CD44, CDC25BUV Response Dn9.62168E-110.044262571COL1A1, COL1A2, SERPINE1, COL5A2, LAMC1Hypoxia2.12418E-080.049496425SLC6A6, KLF7, COL5A1, STC2, SERPINE1, TGFBI, TGM2mTORC1 Signaling2.12418E-085.94696E-12HSP90B1, NAMPT, RRM2, ELOVL5, SLC6A6, SLC7A5, MTHFD2, PSAT1, SCDEstrogen Response Late8.877E-074.83566E-07ELOVL5, TMPRSS3, SNX10, KLK10, SLC7A5, CD44Estrogen Response Early2.75189E-061.86256E-06ELOVL5, TMPRSS3, KLK10, SLC7A5, STC2, SLC39A6, PMAIP1, CD44, TGM2Androgen Response0.0001009830.016721127ELOVL5, SCD, PMEPA1Unfolded Protein Response0.000408454.17426E-06CEBPB, HSP90B1, SLC7A5, MTHFD2, PSAT1, STC2 Common Molecular Pathways and Genes Implicated in Both Ulcerative Colitis and Colorectal Cancer Independent pathway analyses of upregulated genes in colorectal cancer and ulcerative colitis consistently identified epithelial–mesenchymal transition as one of the most significantly enriched and commonly shared biological processes (Table 1). Based on this convergence, EMT was selected for downstream analyses. To identify shared transcriptional features within the EMT program, EMT-related upregulated gene sets from both diseases were intersected. This analysis revealed 24 shared genes (Supp. Figure 5), representing a core EMT signature potentially involved in the pathogenesis of inflammation-associated colorectal cancer. Independent validation of EMT-associated candidates across CRC and UC cohorts To evaluate the clinical and biological robustness of the 24 identified EMT candidates, independent validation was performed across four additional datasets (CRC: GSE5206, GSE32323; UC: GSE47908, GSE59071). Consistent with the discovery analyses, most genes exhibited significant overexpression in CRC tumors relative to matched normal tissues, with LAMC1 as the only exception (p < 0.05) (Fig. 3A, B). Additionally, all 24 candidates were significantly elevated in UC compared with controls in both validation datasets (p < 0.05) (Fig. 3A, B). Taken together, these findings support a conserved EMT-associated transcriptional program that is upregulated in both ulcerative colitis and colorectal cancer and highlight a 24-gene signature shared across disease contexts.Fig. 3External validation and interaction analysis of the 24-gene EMT signature across UC and CRC cohorts. A Expression changes of the 24 EMT-related candidates in CRC (GSE5206, GSE32323; tumor vs. normal) and UC (GSE47908, GSE59071; UC vs. control) datasets. For each gene, log2 fold-change (log2FC) values and corresponding p-values are shown for each cohort. *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant. B Heatmap summarizing log2FC values for the same 24 genes across CRC and UC datasets. Color intensity reflects the direction and magnitude of log2FC. Asterisks indicate statistical significance as defined in panel A. Unsupervised hierarchical clustering is shown by the row dendrogram. C Protein–protein interaction (PPI) network of the 24 EMT-related genes. Nodes represent proteins encoded by the candidate genes, and edges indicate known or predicted interactions External validation and interaction analysis of the 24-gene EMT signature across UC and CRC cohorts. A Expression changes of the 24 EMT-related candidates in CRC (GSE5206, GSE32323; tumor vs. normal) and UC (GSE47908, GSE59071; UC vs. control) datasets. For each gene, log2 fold-change (log2FC) values and corresponding p-values are shown for each cohort. *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant. B Heatmap summarizing log2FC values for the same 24 genes across CRC and UC datasets. Color intensity reflects the direction and magnitude of log2FC. Asterisks indicate statistical significance as defined in panel A. Unsupervised hierarchical clustering is shown by the row dendrogram. C Protein–protein interaction (PPI) network of the 24 EMT-related genes. Nodes represent proteins encoded by the candidate genes, and edges indicate known or predicted interactions To determine whether the validated EMT-associated genes form an interconnected molecular network, protein–protein interaction analysis was subsequently performed. Integration of the 24 validated EMT-associated genes into a high-confidence protein–protein interaction network demonstrated a significantly higher number of interactions than expected for a random gene set of comparable size (51 observed edges vs. 2 expected; PPI enrichment p < 1.0 × 10⁻16). The network exhibited an average node degree of 4.25 and a local clustering coefficient of 0.552, indicating substantial connectivity among the included genes. MCL-based clustering analysis identified discrete network modules. A densely connected cluster comprising multiple collagen family members showed extensive interactions with matrix-associated genes, including MMP1, MMP3, and TIMP1. In addition, CD44 and SPARC exhibited high connectivity within the network (Fig. 3C). Identification of genes associated with UC progression To investigate whether EMT-related genes are associated with the progression from quiescent UC to UC-associated neoplasia, expression profiles of the 24 shared EMT genes were examined in the GSE37283 dataset, which included control, quiescent UC, and UC neoplasia samples. Fourteen genes (INHBA, CXCL1, FAP, TGFBI, COL4A1, COL12A1, MMP1, MMP3, SPP1, TIMP1, TGM2, CTHRC1, CALU, CD44) were found to be significantly upregulated in neoplastic UC compared to quiescent UC (p < 0.05) (Fig. 4). Among these, FAP, INHBA, CALU, SPP1, CTHRC1, TGFBI, COL12A1, and TIMP1 displayed a stepwise increase in expression across control, quiescent UC, and neoplastic UC samples.Fig. 4EMT-associated gene expression across disease stages in ulcerative colitis, including neoplastic transformation. Heatmap depicting expression patterns of EMT-related genes in the GSE37283 cohort, including control, quiescent UC, and UC-associated neoplasia samples. Columns represent individual samples grouped by clinical category, and rows correspond to genes. The color scale indicates relative expression levels within each gene (blue, lower; red, higher). The side annotation marks genes showing statistically significant differences between quiescent UC and UC-associated neoplasia (p < 0.05) EMT-associated gene expression across disease stages in ulcerative colitis, including neoplastic transformation. Heatmap depicting expression patterns of EMT-related genes in the GSE37283 cohort, including control, quiescent UC, and UC-associated neoplasia samples. Columns represent individual samples grouped by clinical category, and rows correspond to genes. The color scale indicates relative expression levels within each gene (blue, lower; red, higher). The side annotation marks genes showing statistically significant differences between quiescent UC and UC-associated neoplasia (p < 0.05) These findings refine a subset of EMT-associated genes linked to neoplastic UC. To further validate the disease relevance of the 14 EMT-related genes identified in UC-associated neoplasia, their expression was assessed in two independent datasets encompassing inactive and active UC samples (GSE179285 and GSE75214). Consistent with the earlier findings, all 14 genes exhibited significant and reproducible upregulation in active/inflamed UC compared with inactive/uninflamed UC and control colonic mucosa (Fig. 5A-B). Together, these analyses define a subset of EMT-associated genes that display consistent expression changes across UC disease states and neoplastic transformation.Fig. 5Validation of EMT-associated gene expression across ulcerative colitis disease states. A Normalized expression levels of the 14 EMT-associated genes across healthy control, inactive UC, and active UC samples in the GSE75214 dataset, and B Normalized expression levels of the same genes in the GSE179285 cohort, comparing healthy control, uninflamed UC, and inflamed UC samples. *p < 0.05, **p < 0.01, ***p < 0.001. The corresponding P-values are labeled accordingly Validation of EMT-associated gene expression across ulcerative colitis disease states. A Normalized expression levels of the 14 EMT-associated genes across healthy control, inactive UC, and active UC samples in the GSE75214 dataset, and B Normalized expression levels of the same genes in the GSE179285 cohort, comparing healthy control, uninflamed UC, and inflamed UC samples. *p < 0.05, **p < 0.01, ***p < 0.001. The corresponding P-values are labeled accordingly Refinement of UC progression–associated genes based on colorectal cancer prognosis The prognostic relevance of the 14 genes associated with UC progression was evaluated in a colorectal cancer cohort (GSE17536) using time-to-event and recurrence-based analyses. Univariate Cox proportional hazards analysis was first applied and identified seven genes (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1) whose elevated expression was associated with reduced overall survival in colorectal cancer (p < 0.05) (Fig. 6A). All seven genes exhibited hazard ratios greater than 1.0, consistent with an adverse prognostic association. Multivariate Cox regression was subsequently performed to assess whether these associations were independent of available clinical covariates. This analysis supported the prognostic association for the same subset of genes (p < 0.05) (Fig. 6B). In addition to Cox regression analyses, Kaplan–Meier survival analysis was used to visualize survival differences between expression-defined groups, revealing significant separation for the gene set identified by Cox regression, whereas SPP1 showed a consistent but non-significant trend (p = 0.065) (Fig. 6C). In parallel, recurrence-based analyses revealed higher expression levels of INHBA, COL4A1, COL12A1, TIMP1, CALU, CTHRC1, and FAP in recurrent compared with non-recurrent tumors (Fig. 6D).Fig. 6Prognostic relevance of 14 EMT-associated gene candidates in colorectal cancer. Forest plot of (A) univariate Cox proportional hazards regression analysis and (B) multivariate Cox regression analysis for the 14 EMT-associated genes in the GSE17536 colorectal cancer cohort, (C) Kaplan–Meier overall survival curves for each gene. Red curves indicate high gene expression, and black curves indicate low gene expression, defined by the median expression level. (D) Comparison of gene expression levels between non-recurrent (blue) and recurrent (red) colorectal cancer cases in the GSE17536 cohort. *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant. The corresponding p-values are labeled accordingly Prognostic relevance of 14 EMT-associated gene candidates in colorectal cancer. Forest plot of (A) univariate Cox proportional hazards regression analysis and (B) multivariate Cox regression analysis for the 14 EMT-associated genes in the GSE17536 colorectal cancer cohort, (C) Kaplan–Meier overall survival curves for each gene. Red curves indicate high gene expression, and black curves indicate low gene expression, defined by the median expression level. (D) Comparison of gene expression levels between non-recurrent (blue) and recurrent (red) colorectal cancer cases in the GSE17536 cohort. *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant. The corresponding p-values are labeled accordingly Integration of survival and recurrence analyses guided prioritization for downstream modeling. SPP1 was retained based on its consistent association with overall survival in univariate Cox regression, despite limited support from Kaplan–Meier and recurrence analyses. COL4A1 was included owing to its reproducible significance in multivariate Cox regression, survival stratification, and recurrence analyses, despite not reaching significance in the univariate model. Conversely, MMP3 was excluded, as its association was confined to recurrence status and lacked support from survival-based analyses. Collectively, these results refine the UC progression–associated gene set to eight candidates (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, CTHRC1, and COL4A1) that also convey prognostic information in colorectal cancer, supporting a molecular link between inflammatory disease progression and adverse oncologic outcome. A novel seven-gene EMT signature discriminates UC-associated neoplasia from quiescent disease The EMT-associated genes prioritized through prognostic and recurrence analyses were next evaluated for their ability to distinguish UC-associated neoplasia from quiescent UC. Receiver Operating Characteristic (ROC) curve analysis was performed, and the area under the curve (AUC) was calculated for each candidate gene in the GSE37283 cohort. Each of the eight genes showed strong individual discrimination, with AUC values ranging from 0.86 to 0.98 (p < 0.05), supporting effective separation of neoplastic and quiescent UC samples (Fig. 7A).Fig. 7Discrimination performance of EMT-associated genes for UC-associated neoplasia. A) Receiver operating characteristic (ROC) curves for individual EMT-associated genes distinguishing UC-associated neoplasia from quiescent UC in the GSE37283 cohort. Black curves represent gene-specific ROC performance, and the red dashed diagonal indicates the line of no discrimination. Area under the curve (AUC) values and corresponding p-values are shown within each panel. (B) Random Forest variable importance plot showing the contribution of each gene to classification performance, measured as Mean Decrease in Gini. Bars represent the relative importance of each gene, with higher values indicating greater contribution to node purity. (C) Summary table reporting AUC values, p-values from ROC analysis, and Mean Decrease in Gini scores for each gene Discrimination performance of EMT-associated genes for UC-associated neoplasia. A) Receiver operating characteristic (ROC) curves for individual EMT-associated genes distinguishing UC-associated neoplasia from quiescent UC in the GSE37283 cohort. Black curves represent gene-specific ROC performance, and the red dashed diagonal indicates the line of no discrimination. Area under the curve (AUC) values and corresponding p-values are shown within each panel. (B) Random Forest variable importance plot showing the contribution of each gene to classification performance, measured as Mean Decrease in Gini. Bars represent the relative importance of each gene, with higher values indicating greater contribution to node purity. (C) Summary table reporting AUC values, p-values from ROC analysis, and Mean Decrease in Gini scores for each gene To further assess their combined relevance and relative contribution, Random Forest classification was performed using the GSE37283 cohort. Variable importance was assessed based on Mean Decrease in Gini. Seven genes (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1) exhibited consistently higher importance scores, each exceeding the predefined threshold (MeanDecreaseGini ≥ 0.5) (Fig. 7B-C). In contrast, COL4A1 showed lower importance within the Random Forest model (Mean Decrease in Accuracy = 0.044; Mean Decrease in Gini = 0.394) (Fig. 7B-C). On this basis, COL4A1 was excluded, and the remaining seven genes were retained as the final feature set for downstream analyses. Z-score–based seven-gene signature across UC and CRC cohorts To quantify the collective transcriptional activity of the seven-gene panel, a composite z-score (MeanZ) was derived for each sample in the GSE37283 cohort. MeanZ values were significantly elevated in UC-associated neoplasia relative to quiescent UC (Mann–Whitney U test, p = 0.0059; Supp. Figure 6). Inclusion of healthy controls revealed a stepwise increase in the MeanZ signature across the continuum from normal mucosa to quiescent UC and, ultimately, UC-associated neoplasia (Kruskal–Wallis test, p = 0.00067; Fig. 8A). ROC analysis further supported the diagnostic precision of this signature, yielding an AUC of 0.95 in discriminating neoplastic transformation from quiescent disease (p = 0.00586; Fig. 8A).Fig. 8Z-score–based evaluation of the seven-gene signature across UC and CRC cohorts. (A) Distribution of the composite z-score across normal, quiescent ulcerative colitis, and UC-associated neoplasia samples in the GSE37283 cohort. Box plots display median and interquartile range. Receiver operating characteristic (ROC) analysis of the z-score was performed to assess discrimination between quiescent UC and UC-associated neoplasia; the area under the curve (AUC) and corresponding p-value are indicated. (B) Assessment of the seven-gene z-score signature in independent colorectal cancer cohorts (GSE15781, GSE39582, and GSE37364). Box plots show z-score distributions across normal and tumor samples (GSE15781 and GSE39582) and across disease stages (GSE37364). ROC curves show classification performance for the indicated comparisons, with AUC values reported. p < 0.05, *p < 0.01, **p < 0.001; ns, not significant Z-score–based evaluation of the seven-gene signature across UC and CRC cohorts. (A) Distribution of the composite z-score across normal, quiescent ulcerative colitis, and UC-associated neoplasia samples in the GSE37283 cohort. Box plots display median and interquartile range. Receiver operating characteristic (ROC) analysis of the z-score was performed to assess discrimination between quiescent UC and UC-associated neoplasia; the area under the curve (AUC) and corresponding p-value are indicated. (B) Assessment of the seven-gene z-score signature in independent colorectal cancer cohorts (GSE15781, GSE39582, and GSE37364). Box plots show z-score distributions across normal and tumor samples (GSE15781 and GSE39582) and across disease stages (GSE37364). ROC curves show classification performance for the indicated comparisons, with AUC values reported. p < 0.05, *p < 0.01, **p < 0.001; ns, not significant The performance of the seven-gene signature was subsequently evaluated across multiple independent colorectal cancer datasets. In the GSE15781 cohort, tumor tissues exhibited a marked increase in MeanZ values compared to normal controls (p = 5.7 × 10−8; AUC = 0.94; Fig. 8B), a trend consistently recapitulated in GSE39582 (p = 2.1 × 10−13; AUC = 0.99). Finally, the GSE37364 cohort was examined to assess the signature’s performance across the adenoma-carcinoma sequence. MeanZ values tracked closely with malignant progression, increasing significantly from normal mucosa to adenoma and invasive carcinoma (Kruskal–Wallis test, p < 2 × 10−16; Fig. 8B). Notably, ROC analyses demonstrated effective discrimination between normal and adenoma samples (AUC = 0.94; p = 2.6 × 10⁻12), and adenoma and colorectal carcinoma samples (AUC = 0.94; p = 2.44 × 10⁻1⁰) (Fig. 8B). Rewiring of the stromal microenvironment during ulcerative colitis–associated tumorigenesis To elucidate the functional coordination within the seven-gene EMT signature, pairwise correlation analyses were performed separately in quiescent ulcerative colitis, UC-associated neoplasia, and colorectal cancer tumor samples. In quiescent UC, the signature showed a highly coordinated pattern, with strong correlations between COL12A1–CTHRC1 (r = 0.98), COL12A1–CALU (r = 0.97), TIMP1–CTHRC1 (r = 0.97), and FAP–TIMP1 (r = 0.97) (Fig. 9A). These associations are consistent with a homeostatic program centered on matrix stabilization in the setting of chronic inflammation. In UC-associated neoplasia, this coordination was partially altered. Several gene pairs that were tightly correlated in quiescent UC showed reduced associations (r ≈ 0.43–0.68), particularly those involving COL12A1 and CALU. In contrast, new high-strength correlations emerged involving SPP1 and INHBA, including SPP1–CTHRC1 (r = 0.91) and TIMP1–SPP1 (r = 0.83) (Fig. 9A), suggesting reorganization rather than uniform loss of transcriptional coupling, with INHBA-centered interactions first emerging at this stage.Fig. 9Co-expression patterns of the seven-gene signature across inflammatory and neoplastic contexts. (A) Correlation plots of the seven-gene signature in the GSE37283 cohort, shown separately for quiescent ulcerative colitis (UC) and UC-associated neoplasia samples. Circle size and color intensity reflect the magnitude of Pearson correlation coefficients. (B) Correlation plots of the same seven genes in colorectal cancer tumor samples from two independent cohorts (GSE17536 and GSE39582), based on Pearson correlation coefficients. (C) Heatmap summarizing Pearson correlation coefficients between INHBA and selected stromal-associated genes across quiescent UC, UC-associated neoplasia, and CRC cohorts (GSE17536 and GSE39582). Color scale indicates the strength of correlation (Pearson r) Co-expression patterns of the seven-gene signature across inflammatory and neoplastic contexts. (A) Correlation plots of the seven-gene signature in the GSE37283 cohort, shown separately for quiescent ulcerative colitis (UC) and UC-associated neoplasia samples. Circle size and color intensity reflect the magnitude of Pearson correlation coefficients. (B) Correlation plots of the same seven genes in colorectal cancer tumor samples from two independent cohorts (GSE17536 and GSE39582), based on Pearson correlation coefficients. (C) Heatmap summarizing Pearson correlation coefficients between INHBA and selected stromal-associated genes across quiescent UC, UC-associated neoplasia, and CRC cohorts (GSE17536 and GSE39582). Color scale indicates the strength of correlation (Pearson r) In CRC cohorts (GSE17536 and GSE39582), the stromal transcriptional network showed a stable and highly coordinated pattern. Notably, the correlation between the TGF-β–associated mediator INHBA and CTHRC1 was strengthened (r = 0.89), while FAP maintained broad connectivity across the signature (r > 0.70) (Fig. 9B, C). Consistent elevation in CTHRC1–FAP (r = 0.84–0.92) and INHBA associations with CTHRC1 and COL12A1 (r = 0.84–0.89) (Fig. 9B, C) support the presence of a tumor-associated stromal transcriptional network. Collectively, these results demonstrate progressive rewiring of stromal gene coordination from quiescent ulcerative colitis to neoplasia and ultimately colorectal cancer. Shared functional pathways associated with the seven-gene risk signature in UC and CRC To characterize the functional context of the seven-gene z-score–based signature, ssGSEA was performed in CRC (GSE17536) and UC (GSE47908) cohorts using Hallmark gene sets. Samples were stratified into high- and low-risk groups based on the median MeanZ score. In the ulcerative colitis cohort, elevated MeanZ values were associated with higher ssGSEA scores for inflammatory pathways, including Inflammatory Response, TNF-α signaling via NF-κB, and IL6–JAK–STAT3 signaling (Fig. 10A). In parallel, ssGSEA scores for EMT, Angiogenesis, Coagulation, and Hypoxia were also higher in samples with increased MeanZ values, similar to the pattern observed in colorectal cancer. Peroxisome gene sets showed a consistent inverse association with MeanZ in UC, with additional negative associations observed for Oxidative Phosphorylation and Fatty Acid Metabolism (Fig. 10B). In the colorectal cancer cohort, unsupervised hierarchical clustering of ssGSEA scores showed that samples with higher MeanZ values were characterized by higher enrichment of EMT, Hypoxia, Angiogenesis, Coagulation, and TGF-β signaling pathways (Fig. 10C). Spearman correlation analysis further demonstrated positive associations between MeanZ and ssGSEA scores for EMT, Hypoxia, Angiogenesis, and Coagulation, whereas Peroxisome-related gene sets displayed an inverse relationship with MeanZ (Fig. 10D).Fig. 10Single-sample gene set enrichment analysis of the seven-gene z-score signature. A Heatmap of ssGSEA enrichment scores for Hallmark pathways in ulcerative colitis samples (GSE47908), stratified into high- and low-risk groups based on the median seven-gene z-score. Rows represent Hallmark gene sets and columns represent individual samples; color intensity reflects relative pathway enrichment. B Spearman correlation matrix between the composite z-score (MeanZ) and Hallmark ssGSEA scores in the UC cohort. Circle size and color indicate the strength and direction of correlation. C Heatmap of ssGSEA enrichment scores for Hallmark pathways in colorectal cancer samples (GSE17536), stratified by z-score–defined risk groups, displayed as in panel A. D Spearman correlation matrix between MeanZ and Hallmark ssGSEA scores in the CRC cohort, visualized as in panel B Single-sample gene set enrichment analysis of the seven-gene z-score signature. A Heatmap of ssGSEA enrichment scores for Hallmark pathways in ulcerative colitis samples (GSE47908), stratified into high- and low-risk groups based on the median seven-gene z-score. Rows represent Hallmark gene sets and columns represent individual samples; color intensity reflects relative pathway enrichment. B Spearman correlation matrix between the composite z-score (MeanZ) and Hallmark ssGSEA scores in the UC cohort. Circle size and color indicate the strength and direction of correlation. C Heatmap of ssGSEA enrichment scores for Hallmark pathways in colorectal cancer samples (GSE17536), stratified by z-score–defined risk groups, displayed as in panel A. D Spearman correlation matrix between MeanZ and Hallmark ssGSEA scores in the CRC cohort, visualized as in panel B Across both cohorts, the seven-gene z-score signature showed concordant associations with ssGSEA scores of EMT, Hypoxia, Angiogenesis, and Coagulation pathways, accompanied by increased inflammatory pathway activity, while displaying inverse associations with Peroxisome-related gene sets. Discussion Ulcerative colitis is a chronic, relapsing inflammatory bowel disease characterized by continuous mucosal inflammation that typically initiates in the rectum and extends proximally along the colon [22–24]. Long-term exposure to chronic inflammation serves as the primary risk factor for colorectal cancer development in patients with ulcerative colitis [4, 25–27]. Accordingly, colitis-associated colorectal cancer represents one of the most clinically consequential complications of UC and is estimated to account for approximately 15% of all-cause mortality among individuals with inflammatory bowel disease [28]. Recent studies further underscore this risk; a contemporary meta-analysis reported a pooled CRC incidence of 1.47 per 1,000 person-years and an overall prevalence of 1.54%, corresponding to a 2.48-fold increased malignancy risk relative to the general population [29]. Moreover, Ikebata et al. (2025) reported reduced 5-year recurrence-free survival for colorectal cancers arising within inflamed ulcerative colitis segments compared with non-inflamed regions (about 75% vs 88%), and demonstrated a graded association between endoscopic inflammatory activity at diagnosis and poorer cancer-specific survival [30]. In this context, identifying UC patients at risk of malignant progression remains a major clinical challenge, underscoring the need for objective biomarkers to refine colonoscopic surveillance and personalized management. Addressing this gap requires a deeper understanding of the molecular events that accompany the transition from chronic inflammation to neoplasia in the UC colon [6, 31]. Within this framework, transcriptome-wide profiling offers a systems-level view of disease progression by capturing coordinated changes across epithelial and microenvironmental compartments, thereby facilitating the identification of biologically relevant biomarkers and therapeutic targets [6, 32–34]. Accordingly, a stepwise transcriptomic framework was applied to delineate molecular programs shared between ulcerative colitis and colorectal cancer that may contribute to inflammation-associated tumorigenesis. Comparative enrichment analysis identified 19 molecular pathways commonly activated across both diseases, notably epithelial–mesenchymal transition, inflammatory response, TNF-α signaling via NF-κB, interferon alpha and gamma signaling, IL-2/STAT5 signaling, hypoxia, and mTORC1 signaling. These overlapping programs point to a transcriptional landscape shaped by inflammation, immune dysregulation, tissue remodeling, and stress-associated cellular responses. Among these pathways, epithelial–mesenchymal transition emerged as the most consistently enriched biological process. Consistent with this pattern, several genes involved in inflammatory signaling and extracellular matrix dynamics, including INHBA, CXCL1, CXCL8, MMP1, MMP3, COL1A1, and S100A8, were recurrently upregulated across both ulcerative colitis and colorectal cancer datasets. The progression of colitis-associated colorectal cancer from ulcerative colitis is a dynamic and heterogeneous process driven by persistent inflammatory injury and complex interactions between immune mediators and cytokine signaling pathways. This multifactorial sequence involves the transition from non-dysplastic mucosa to low- and high-grade dysplasia and ultimately invasive carcinoma, reflecting the cumulative effects of inflammation-associated molecular alterations rather than a linear genetic progression [22, 26, 35, 36]. In line with this model, although inflammatory signaling was present in both ulcerative colitis and colorectal cancer, epithelial-mesenchymal transition emerged as the most consistently enriched shared program, accompanied by hypoxia-related and mTORC1 signaling, highlighting stromal and extracellular matrix-associated transcriptional remodeling beyond inflammation alone. Recent multi-omics studies have advanced understanding of inflammatory bowel disease pathogenesis by highlighting the interplay between genetic susceptibility, immune dysregulation, and environmental cues [37]. Chronic inflammation is maintained by immune cell populations, including macrophages, neutrophils, and lymphocytes, through sustained release of cytokines and chemokines [38], and has been linked to abnormal T-cell activation, Th17 imbalance, and excessive inflammatory mediator production in ulcerative colitis [39–41]. Notably, mTOR signaling has been implicated in promoting Th17 differentiation via HIF-1α–dependent mechanisms, thereby sustaining inflammatory activity in ulcerative colitis [41]. Beyond immune dysregulation, hypoxia has emerged as an active regulator of inflammatory and tissue remodeling processes, with HIF- and NF-κB–dependent pathways shaping cellular behavior within the inflamed microenvironment [42]. Consistent with this framework, Yeh et al. showed that interferon-driven inflammatory signaling induces functional HIF-1α activation through PI3K/AKT/mTOR-dependent mechanisms, promoting epithelial–mesenchymal transition, invasiveness, and other tumorigenic properties in cancer cells [43]. These findings provide mechanistic evidence for a crosstalk between inflammatory and hypoxic microenvironmental programs, in which immune-mediated inflammation engages HIF-1α–dependent signaling through pathways that include PI3K/AKT/mTOR activation, thereby linking EMT to malignant progression. Consistent with these observations, inflammation, hypoxia, and tumor–stroma interactions have been identified as important drivers of the invasive cascade [44]. A recent study demonstrated that endoscopic disease activity in ulcerative colitis is associated with increased mRNA expression of inflammatory and tissue remodeling markers, including TGF-β, vimentin, FSP1, and α-SMA, not only in inflamed mucosa but also in adjacent endoscopically uninflamed regions. Notably, the authors proposed that mRNA-based profiling of inflammatory and epithelial–mesenchymal transition markers captures disease activity within the colonic mucosal microenvironment with greater sensitivity than histological assessment alone. These molecular activity patterns provide insight into the pathophysiology of ulcerative colitis and highlight the microenvironment as a tractable target for therapeutic modulation [38]. Collectively, these observations provide a biological context for the findings, in which chronic inflammation may establish a permissive microenvironment, while EMT, hypoxia, and mTOR-associated signaling programs mediate stromal remodeling and epithelial plasticity. Concordant with this model, EMT did not emerge as an isolated pathway but as the most consistently enriched shared program across ulcerative colitis and colorectal cancer cohorts, with associated genes linked to extracellular matrix organization and stromal reprogramming. Based on the consistent enrichment of epithelial–mesenchymal transition observed across ulcerative colitis and colorectal cancer DEG lists, EMT was selected for focused downstream analyses. Subsequent prioritization across neoplastic UC, active UC, and colorectal cancer cohorts refined the initial EMT-associated gene set to a seven-gene signature comprising FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1. This signature showed progressive elevation across the UC-neoplasia continuum and retained discriminatory performance across independent colorectal cancer cohorts, suggesting that EMT-associated stromal gene expression patterns captured by this stepwise analytical framework may be relevant across active inflammation and neoplastic transition. The EMT-associated signature identified in this study should be interpreted as an observational transcriptomic pattern associated with neoplastic transformation rather than a direct driver of disease progression. To determine whether the signature operates as a coherent program rather than seven independent markers, internal connectivity within the signature genes was evaluated across disease states. The co-expression analyses indicate that the seven-gene signature does not behave as a simple sum of independent transcripts but instead reorganizes as disease advances. In particular, increasing INHBA-centered connectivity across quiescent UC, neoplasia, and CRC was consistent with progressive reorganization of a stromal/ECM-related module along the UC-CRC sequence. Rather than reflecting inflammation alone, these genes may capture a fibroblast- and TGF-β–aligned program that becomes selectively reinforced in neoplastic tissue and more evident in colorectal cancer. Several members of the signature have been independently implicated in stromal activation and inflammation-driven tumorigenesis. A growing body of experimental and translational studies in diverse solid tumors including CRC [45–65] has independently linked members of this panel, including INHBA, FAP, COL12A1, TIMP1, CTHRC1, CALU, and SPP1, to fibroblast activation, extracellular matrix remodeling, and TGF-β–aligned stromal programming. Notably, emerging evidence supports a role for individual components of this panel within the inflammatory and stromal context of UC-associated carcinogenesis. In particular, SPP1 has been linked to macrophage-driven signaling and disease severity [66–69], while TIMP1 has been associated with extracellular matrix remodeling and immune infiltration [14, 23, 70]; however, evidence for other genes remains limited. This supports interpretation at the level of a coordinated stromal program rather than individual gene effects. Taken together, these independent studies appear to converge with the findings presented here, placing stromal activation, fibroblast-linked programs, and TGF-β/EMT signaling within the biology that accompanies the transition from chronic colitis to neoplasia and ultimately colorectal cancer. In parallel, an increasing number of studies support the idea that stromal activation contributes to disease processes in IBD and colorectal cancer. Honan et al. (2024) reported that circulating stromal-like cells are more abundant in patients with UC or Crohn’s disease than in healthy individuals, express podoplanin and other fibroblast-linked markers, and track with changes in both lymphoid and myeloid compartments, suggesting that stromal and immune responses evolve together during intestinal inflammation [71]. Supporting a possible role in neoplastic progression, Björner et al. (2026) showed that colitis-associated colorectal cancers with a high stromal component had substantially poorer survival than stroma-low tumors, and occurred more often in younger patients [72]. Zhou and colleagues (2025) extended these observations in experimental colitis, demonstrating that stromal subsets expand or contract depending on the inflammatory and therapeutic context, and influence epithelial repair and T-cell activation [73]. In addition and important point in this context, Zhou et al. (2023) further highlighted that colitis-associated colorectal cancers frequently map to the CMS4 transcriptional subtype, a mesenchymal program and characterized by epithelial–mesenchymal transition, TGF-β signaling, angiogenesis and marked stromal infiltration. CMS4 tumors are enriched for cancer-associated fibroblasts and have been associated with poorer clinical outcome and treatment resistance, suggesting that stromal remodeling could be a recurring feature of CAC biology[4]. Shah and Itzkowitz have also pointed out that colitis-associated cancers appear to show a relative absence of CMS2 tumors and a shift toward CMS4-linked biology, characterized by epithelial–mesenchymal transition pathways, prominent stromal infiltration and increased CD4⁺ T-cell and monocyte signals, suggesting that colitis-driven tumorigenesis may favour a more mesenchymal transcriptomic profile [74]. To examine whether samples with high composite scores showed convergence on shared functional pathways, ssGSEA was performed in the context of pathway-level associations observed across UC and CRC. The ssGSEA analyses indicate that the seven-gene signature does not operate in isolation but co-segregates with broader pathway programs shared between UC and CRC. Functionally, elevated signature activity was associated with enrichment of EMT, hypoxia, angiogenesis and coagulation pathways, accompanied by increased inflammatory pathway activity, while showing inverse associations with peroxisome-related gene sets. Epithelial–mesenchymal transition is a developmental program that can be aberrantly reactivated in cancer, conferring migratory, invasive, and stem-like properties while promoting resistance to apoptosis [44, 75]. Accumulating evidence indicates a reciprocal relationship between EMT and chronic inflammation, whereby inflammatory mediators such as cytokines, oxidative stress, and hypoxia induce EMT, while EMT-activated tumor cells amplify local inflammation through secretion of cytokines, chemokines, and matrix-remodeling enzymes. This bidirectional EMT–inflammation axis has been linked to early tumorigenesis, metastatic progression, and poor clinical outcome across multiple cancer types, including colorectal cancer [75]. In this context, recent single-cell studies have provided important insights into the cellular heterogeneity underlying inflammatory bowel disease. Kong et al. (2023) demonstrated that chronic inflammation in Crohn’s disease is accompanied by marked transcriptional reprogramming of epithelial and stromal cells, while immune alterations were largely compositional. Their single-cell atlas further identified disease-associated fibroblast activation programs, highlighting the contribution of stromal remodeling to intestinal pathology [76]. Zhou et al. identified SPP1⁺ macrophages and CHI3L1⁺ fibroblasts as key communication hubs in ulcerative colitis, linking stromal–immune interactions to disease severity and UC-associated colorectal cancer [68]. The findings of the present study extend these cell-resolved observations by demonstrating that inflammation-associated stromal and EMT-related programs converge at the bulk transcriptomic level into a coordinated gene network that persists across disease stages, from quiescent ulcerative colitis to neoplasia and colorectal cancer. Consequently, through systematic prioritization across multiple independent cohorts, a seven-gene stromal signature (FAP, TIMP1, COL12A1, CALU, INHBA, SPP1, and CTHRC1) was identified and validated, capturing this transition with high discriminatory accuracy. Previous studies have proposed several biomarkers [77–80] and gene expression signatures [81] for UC-associated colorectal cancer, many of which are based on inflammatory, immune, or epigenetic alterations and rely on either a limited number of genes or broader transcriptomic profiles. Early microarray studies identified gene signatures associated with neoplastic progression in UC [81], while methylation-based panels such as those evaluated in the ENDCaP-C study have shown utility in detecting dysplasia and neoplasia [82]. More recent transcriptomic analyses have also described gene sets linked to UC–CRC progression, frequently reflecting immune and inflammatory pathways [83]. However, these approaches have not yet been incorporated into routine clinical practice for risk stratification or surveillance and largely remain at the research stage, with current practice still relying primarily on endoscopic and histopathological assessment. Compared with these previous approaches, the signature identified in the present study captures a coordinated stromal and microenvironmental transcriptional program rather than isolated gene-level changes or predominantly epithelial or immune signals. To the best of available knowledge, the collective behavior of these seven genes has not been systematically evaluated. The analysis presented here highlights the value of evaluating these genes collectively, which may provide a reproducible expression pattern in the context of ulcerative colitis progression to colorectal cancer. In addition, CALU, CTHRC1, INHBA, and COL12A1 are reported here for the first time as genes specifically associated with UC-related colorectal cancer, highlighting a previously underappreciated stromal axis in inflammation-associated tumorigenesis. The consistent association of this signature with EMT, hypoxia, angiogenesis, coagulation, and TGF-β signaling across UC and CRC cohorts underscores its biological coherence and relevance to microenvironmental remodeling. The involvement of pathways related to epithelial–mesenchymal transition, extracellular matrix remodeling and hypoxia further suggests that stromal processes may be more prominent in this transition than traditionally assumed. Finally, as colitis-associated cancers have been reported in limited studies to align with a CMS4-like profile characterized by mesenchymal signaling and stromal expansion, the findings of the present study may fit within a broader picture suggesting that stromal reprogramming may be associated with the context in which UC-associated neoplasia develops. This possibility remains hypothesis-generating and will require prospective validation. From a translational perspective, the ability of this focused gene panel to distinguish UC-associated neoplasia from quiescent disease, and to track malignant progression across the adenoma–carcinoma sequence, suggests potential utility for risk stratification and surveillance in UC patients. These findings aim to support earlier detection and improved prevention strategies for UC-related colorectal cancer by providing a molecular readout of stromal reprogramming that precedes overt malignancy. This signature may serve as a molecular aid for risk stratification and could support current endoscopic surveillance by helping identify UC patients more likely to progress toward colorectal cancer. Although this study identifies an EMT-associated signature for stratifying the risk of UC-associated neoplasia, several limitations should be acknowledged. All analyses were conducted on retrospective, publicly available bulk transcriptomic datasets and therefore represent a computational discovery framework rather than a clinically validated predictive model. The proposed seven-gene signature should therefore be considered hypothesis-generating and requires validation in independent UC-associated neoplasia cohorts. Although each dataset was processed and analyzed independently within its respective platform, residual technical variability across platforms cannot be entirely excluded, and may influence relative expression estimates and cross-cohort comparisons of effect magnitude. In addition, the analysis relies on preprocessed GEO pipelines rather than a fully customized raw data workflow. The analysis was also restricted to upregulated genes and therefore does not represent the full spectrum of transcriptional alterations across the UC-CRC sequence. Another limitation is the small sample size of the dataset used to construct the core model. Although the discovery phase included multiple large-scale CRC and UC cohorts to enhance statistical power and generalizability, the dataset modeling the transition from quiescent UC to UC-associated neoplasia (GSE37283) is limited in size (n = 20; 11 neoplastic UC, 4 quiescent UC, and 5 controls). While this appears to be the only publicly available human dataset directly capturing the transition from normal mucosa to quiescent UC and UC-associated neoplasia, the limited number of quiescent and neoplastic UC cases constrains statistical power and may affect feature stability as well as the precision of discrimination metrics. Furthermore, the study lacks independent prospective or experimental validation, as well as multi-omic integration (e.g., proteomic, epigenetic, and single-cell transcriptomic data) and direct correlation with endoscopic and histopathological features, such as lesion morphology or dysplasia grade. Validation in clinically well-annotated cohorts spanning the UC–CRC sequence will therefore be essential to confirm coordinated upregulation of the seven-gene panel at both transcript and protein levels and to determine whether its discriminatory performance is reproducible beyond the discovery datasets. Finally, as the analyses are based on bulk transcriptomic data, the cellular origin of the signature cannot be directly resolved and should be further investigated using tissue-resolved approaches to confirm stromal localization within the mesenchymal and CMS4-like context identified here. Supplementary Information Additional file1 (PDF 539 KB) Additional file2 (PDF 245 KB) Additional file1 (PDF 539 KB) Additional file2 (PDF 245 KB)