Skip to main content
NCBI home page
Journal of Inflammation Research logoLink to Journal of Inflammation Research
. 2026 Sep 3;19:633770. doi: 10.2147/JIR.S633770

Integrative Machine Learning and Single-Cell RNA Sequencing Reveals Timp1+ Stromal Cells as Key Drivers in Ulcerative Colitis

Xiaosong Xu 1, Lipeng Du 1, Ze Li 2, Xuemei Jia 2, Yao Du 2,3, Qian Yang 2,✉, Yanru Cai 2,3,✉
PMCID: PMC13549397  PMID: 42707758

Abstract

Background

Ulcerative colitis (UC) is a chronic relapsing inflammatory bowel disease (IBD). The role of stromal cells in UC pathogenesis is increasingly being recognized, yet their functional subsets and underlying molecular mechanisms remain poorly defined.

Objective

To integrate multiple transcriptomic datasets, combined weighted gene co-expression network analysis (WGCNA) with machine learning algorithms for the identification of robust core biomarkers of UC and to resolve the cell-subset-specific expression and pathogenic mechanisms of candidate genes using single-cell RNA sequencing (scRNA-seq).

Methods

Three GEO datasets (GSE38713, GSE3365, and GSE24287) were merged after batch correction. WGCNA, together with five machine learning algorithms, LASSO, random forest, Boruta, gradient boosting machine, and XGBoost, were applied to screen for core genes. External validation was performed using two independent datasets, GSE92415 and GSE179285, and a DSS- of DSS-induced chronic colitis. The cell- subset-specific expression and mechanisms of candidate genes were dissected by scRNA- sequencing.

Results

Three core genes, TIMP1, S100A12, and ALPL, were identified. External validation showed that TIMP1 and S100A12 were consistently upregulated in both independent datasets, whereas ALPL was significantly upregulated in GSE92415, but not in GSE179285. TIMP1 was selected for further investigation based on its differential significance. In the DSS-induced chronic colitis mouse model, the area of Timp1-positive colon tissue increased by 2.3-fold compared to in the controls. scRNA-seq revealed that Timp1 was specifically enriched in a stromal cell subset of the mouse colon. Pseudotime trajectory analysis indicated that Timp1⁺ stromal cells predominantly localized to late differentiation stages, with their expression dynamics undergoing a sequential transition from matrix remodeling and angiogenesis to immune regulation. Cell-cell communication analysis further identified the enrichment of the Ppia-Bsg signaling axis between Timp1⁺ stromal cells and epithelial cells.

Conclusion

Timp1 was specifically enriched in a colonic stromal cell subset in UC mice, and this subset exhibited a shifted differentiation trajectory and enhanced communication with epithelial cells. These findings provide candidate targets for the molecular diagnosis of UC and highlight the translational potential of targeting Timp1⁺ stromal cells or their downstream signaling pathways for stromal-directed therapeutic intervention.

Keywords: ulcerative colitis, stromal cells, tissue inhibitor of metalloproteases, TIMP1, machine learning, single-cell RNA sequencing, WGCNA analysis

Graphical Abstract

A flowchart of data integration, machine learning and validation processes in a research study. The flowchart illustrates a research study process divided into three main sections. The first section, ′Data integration & Batch correction,′ involves datasets GSE38713, GSE3365 and GSE24287, leading to an integrated dataset after batch correction. This section connects to ′WGCNA,′ showing gene dendrogram and module-trait relationships. The second section, ′Machine Learning,′ includes LASSO analysis, Random Forest, Boruta, GBM and XGBoost, identifying hub genes TIMP1, S100A12 and ALPL. The third section, ′External Validation,′ uses datasets GSE92415 and GSE179285, focusing on TIMP1. ′Experimental Validation′ involves various analyses like tSEN, Psudome and CellChat and a pathway showing immune cells, Timp1+ stromal, epithelial cells and UC. The flowchart connects these processes, highlighting the study′s comprehensive approach to data analysis and validation.

Open in a new tab

Introduction

Ulcerative colitis (UC) is a chronic, relapsing inflammatory bowel disease (IBD) characterized by persistent colonic mucosal injury, aberrant immune responses, and disruption of the epithelial barrier.1,2 In recent years, the global incidence of UC has continued to rise, particularly in newly industrialized countries.3 Although biologic agents and small-molecule drugs have substantially improved outcomes in a subset of patients, a considerable proportion still fail to respond adequately or develop resistance. Moreover, reliable molecular markers for precise disease stratification and personalized treatment remain lacking.4,5 Consequently, a systematic understanding of UC pathogenesis through the identification of robust core driver genes and exploration of their functions in specific intestinal cell subsets is of fundamental value and translational promise.

High-throughput transcriptomic technologies are powerful tools for investigating the mechanisms of UC. Numerous microarray and single-cell RNA sequencing (scRNA-seq) studies have identified differentially expressed genes (DEGs) implicated in inflammation, epithelial repair, stromal remodeling, and immune responses.6,7 However, owing to limited sample sizes, platform heterogeneity, biological variability, and batch effects, DEGs reported across studies often show poor reproducibility, and few markers have demonstrated cross-dataset robustness.8,9 In recent years, the combination of weighted gene co-expression network analysis (WGCNA) and multiple machine learning algorithms has emerged as a promising strategy to extract stable and reproducible core genes from high-dimensional omics data. WGCNA systematically captures co-expression modules, whereas algorithms such as LASSO, random forest, Boruta, gradient boosting machine, and XGBoost offer complementary advantages in controlling overfitting, handling high-dimensional features, and identifying critical variables.10–12 Nevertheless, most existing bioinformatics studies remain confined to gene lists and lack mechanistic validation that links candidate molecules to functional phenotypes in specific intestinal cell subsets. In particular, analyses at single-cell resolution that are capable of dissecting cellular heterogeneity and intercellular communication networks are conspicuously absent.13,14

Beyond epithelial barrier dysfunction and immune cell infiltration, the role of stromal cells in tissue remodeling, inflammation persistence, and repair processes in UC has gained increasing attention, recent studies have identified diverse stromal subpopulations in the inflamed intestine, including inflammatory fibroblasts, myofibroblasts, and pericytes, each exhibiting distinct transcriptional profiles and functional roles.15,16 However, the stromal subpopulations that are functionally relevant, and how they interact with epithelial or immune cells, remain largely unknown. Notably, although upregulation of stromal-related protease inhibitors has been reported in UC,17 their spatiotemporal expression dynamics and functional significance within specific stromal subsets have yet to be elucidated.

In this study, we integrated multiple independent UC transcriptomic datasets, rigorously corrected for batch effects, and combined WGCNA with five machine-learning algorithms to identify core genes that are robust across different cohorts. After external validation using independent datasets and a DSS-induced chronic colitis mouse model, we resolved the cell- subset-specific expression of these genes in the colon using scRNA- sequencing. In brief, this work identified key molecular markers of UC at the multi-dataset level and, at single-cell resolution, revealed the potential involvement of a specific stromal cell subset in UC pathogenesis, thereby providing a new foundation for molecular diagnosis and stromal-targeted therapeutic strategies.

Materials and Methods

Data Preprocessing

The microarray datasets used in this study were obtained from the Gene Expression Omnibus (GEO) database (version 5.12; https://www.ncbi.nlm.nih.gov/geo/), a publicly accessible repository of microarray and gene expression profiling data. Three major transcriptomic datasets were included in this analysis: GSE38713, GSE3365, and GSE24287. GSE38713 was generated on the GPL570 platform and contained samples from 30 patients with ulcerative colitis (UC) and 13 healthy controls.6 GSE3365, based on the GPL96 platform, comprises 26 patients with UC and 42 healthy controls.7 GSE24287, based on the GPL6480 platform, includes 27 patients with UC and 25 healthy controls.18 After data normalization to ensure data integrity and batch consistency, two additional datasets, GSE92415 and GSE179285, were selected as independent validation cohorts.19,20 Detailed information for all the datasets is summarized in Table 1.

Table 1.

Summary of All Datasets Used in This Study

Dataset Platform Title
GSE38713 GPL570 [HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array
GSE3365 GPL96 [HG-U133A] Affymetrix Human Genome U133A Array
GSE24287 GPL6480 Agilent-014850 Whole Human Genome Microarray 4x44K G4112F (Probe Name version)
GSE92415 GPL13158 [HT_HG-U133_Plus_PM] Affymetrix HT HG-U133+ PM Array Plate
GSE179285 GPL6480 Agilent-014850 Whole Human Genome Microarray 4x44K G4112F (Probe Name version)
Open in a new tab

Differential Gene Expression Analysis

DEGs between patients with UC and healthy controls were identified using the GSE38713, GSE3365, and GSE24287 datasets. Principal component analysis (PCA) was performed using the ggplot2 package in R (version 4.5.0) to assess global expression differences among the samples. DEG analysis was conducted using the limma package, with significance thresholds set at P < 0.05, and |log2 fold change (FC)| > 1. The ComBat function in the sva package was applied to correct batch effects across the three datasets, effectively reducing potential confounding factors.21 Normalized data are presented as box plots (Supplementary Figure 1) and PCA cluster plots before and after batch correction are shown in Supplementary Figure 2. The filtered DEGs were visualized as volcano plots and heatmaps using ggplot2 and pheatmap packages.

Construction of a Co-Expression Network and Identification of Hub Modules by WGCNA

WGCNA is an advanced systems biology method used to reveal the correlation patterns among genes across microarray samples. It facilitates the identification of gene sets exhibiting strong covariation and can uncover potential candidate biomarker genes or therapeutic targets by exploring the internal relationships within gene sets and their correlation with phenotypes.22 In this study, the WGCNA co-expression system was established using the WGCNA package.23 First, genes with a variance exceeding 25% across samples in the meta-dataset were selected as inputs for WGCNA. To enhance the robustness, outlier samples were removed via hierarchical clustering using the goodSamplesGenes function. A scale-free network was then employed, and the pickSoftThreshold function was used to estimate the appropriate soft-thresholding power. The correlation matrix was converted into an adjacency matrix, which was further transformed into a topological overlap matrix (TOM). Dynamic tree cutting was subsequently applied to identify the modules. Finally, Pearson’s correlation analysis was performed to assess the relationships between modules and clinical phenotypes, and the module with the highest correlation coefficient was selected for further investigation.

Machine Learning Methods and Hub Gene Screening

Five machine-learning methods were applied to the identified core genes: gradient boosting machine (GBM), Boruta algorithm, extreme gradient boosting (XGBoost), random forest (RF), and least absolute shrinkage and selection operator (LASSO). RF, GBM, and XGBoost are decision-tree-based ensemble learning models that effectively capture nonlinear relationships and high-order interactions among features.22,24,25 Boruta, a random-forest-based feature selection algorithm, directly identifies features that are truly correlated with the outcome variable.26 LASSO uses L1 regularization to achieve sparse variable selection, making it particularly suitable for high-dimensional omics data.27 These five algorithms were deliberately chosen to represent diverse methodological paradigms: LASSO provides sparse linear selection via L1 regularization; Random Forest and Boruta evaluate feature importance through tree-based and wrapper strategies; and GBM with XGBoost apply gradient boosting to iteratively refine feature weights. This multi-paradigm ensemble minimizes the bias inherent to any single approach and ensures that the identified genes are consistently prioritized across fundamentally different analytical perspectives.

Experimental Animals

All the animal experiments were performed in accordance with the principles of laboratory animal care. Male C57BL/6J mice (7–9 weeks old) were purchased from Beijing HFK BIOSCIENCE Co. Ltd. (Beijing, China). All animals were housed under standard pathogen-free conditions at a temperature of 21–24°C and a 12 h light/dark cycle. A mouse model of ulcerative colitis was established as previously described.28 Briefly, C57BL/6J mice were randomly divided into the control and model groups (n = 6 per group). Sample size was determined based on power analysis using data from our preliminary experiments (effect size for colon length reduction = 1.2, α = 0.05, power = 0.80). Mice in the model group received 1.5% dextran sulfate sodium (DSS) solution for 5 days followed by distilled water for 5 days per cycle, which was repeated for a total of five cycles to induce chronic colitis. This study followed the ARRIVE guidelines 2.0.29 The animal experimental protocol was approved by the Laboratory Animal Welfare and Ethics Committee of the Hebei Provincial Hospital of Traditional Chinese Medicine (IACUC-HPHCM-2026014).

Miniature Endoscopic Imaging

Isoflurane inhalation anesthesia was used for all colonoscopic examinations, delivered via standard small animal vaporizer. Mouse respiratory rate was continuously monitored throughout the anesthetic procedure to adjust anesthetic concentration and avoid excessive depth of anesthesia. An H30 portable electronic endoscope (Shanghai Yuyan Instruments Co., Ltd., China) was used to directly visualize the - DSS-induced colonic mucosal injury. Before imaging, mice were anesthetized with inhaled isoflurane (RWD Life Science, China). Colitis severity was blindly assessed using the mouse endoscopic index of colitis severity (MEICS), which scores mucosal surface granularity (0–3), vascular pattern (0–3), colonic mucosal transparency (0–3), visible fibrin deposition (0–3), and stool consistency (0–3).30

Histological Assessment

After all in vivo observations were finished, cervical dislocation was performed as the primary euthanasia method prior to colon tissue collection. Complete cardiac arrest was visually confirmed via immediate anatomical dissection to verify permanent death, to ensure no residual animal pain or distress. Distal colon tissues were collected and fixed in 4% buffered formalin for 24–48 h. For hematoxylin and eosin (H&E) staining, colon tissues were dehydrated, embedded in paraffin, sectioned, and stained. Blinded pathological severity scoring was performed based on the following parameters: crypt loss and architectural disruption are considered all-or-nothing severe pathological alterations; therefore, they were scored as 0 (absent) or 3–4 (severe), whereas other parameters (inflammatory infiltration, goblet cell depletion, and mucosal thickening) were scored on a continuous 0–4 scale. This scoring approach is consistent with standard practice in DSS-induced colitis models.31 For immunohistochemistry (IHC), colon sections were stained with an anti-TIMP −1 antibody. For quantitative analysis of TIMP-1 immunohistochemistry, three non-adjacent sections per mouse were examined. From each section, five randomly selected high-power fields (×200 magnification) were captured using a digital pathology scanner (Pannoramic MIDI, 3DHistech, Hungary). The percentage of TIMP-1-positive area (brown staining) relative to the total mucosal area was quantified using ImageJ software (NIH, USA) with the color deconvolution plugin. The mean value from the five fields per section was calculated, and the average of three sections per mouse was used for final statistical analysis. All quantifications were performed by two independent observers blinded to the experimental groups.

scRNA-Seq Analysis

Single-cell RNA sequencing was performed on distal colon tissues from DSS-induced chronic colitis mice only (n = 2 mice pooled per run). The sequencing was carried out using a 10× Genomics platform (Beijing SeekGene BioSciences Co., Ltd., China). Sample demultiplexing, barcode identification, and 3′ gene counting were conducted using Cell Ranger software (version 8.0). The GRCh38/GRCm38 genome assembly was used as a reference genome. Dimensionality reduction, clustering, and data analysis were performed using Seurat (version 4.1.1). Only cells with 500–5000 expressed genes and fewer than 500 unique molecular identifier (UMI) counts were retained; cells with mitochondrial gene content of >20% were excluded. Two-dimensional visualization was performed using RunUMAP and marker genes were identified using the FindAllMarkers algorithm in Seurat. Enrichment analysis, Gene Ontology (GO) enrichment analysis, cell–-cell interaction analysis, pseudotime trajectory analysis, and other bioinformatics analyses were conducted using the SeekSoul Online platform (https://seeksoul.online/index.html#/sgvps/home).

Statistical Analysis

Statistical analyses and data visualization were performed using GraphPad Prism (v10.1.2) and R software (v4.5.0). For comparisons of hub gene expression levels between UC and control groups, Student’s t-test or the Wilcoxon rank-sum test was used as appropriate. All P values were two-sided, and statistical significance was defined as P < 0.05.

Results

Differential Gene Expression Analysis

A total of 337 upregulated and 245 downregulated DEGs were identified in the GSE38713 dataset (Supplementary Table 1), visualized by volcano plots (Figure 1A) and heatmaps (Figure 1B). In the GSE3365 dataset, 52 up-regulated and 34 down-regulated DEGs were detected (Supplementary Table 2, Figure 1C and D). In GSE24287, 37 up-regulated and 35 down-regulated DEGs were identified (Supplementary Table 3, Figure 1E and F). After gene symbol conversion and normalization of the GSE38713, GSE3365, and GSE24287 datasets, the expression profiles were obtained. Following data merging and batch effect removal, a meta-dataset was constructed from the three cohorts. This -dataset contained 10,430 gene symbols and was comprised of 83 UC samples and 80 healthy samples.

Figure 1.

Six plots showing differential gene expression via volcano plots and circular heatmaps for three datasets. Three volcano plots and circular heatmaps illustrate differential gene expression in datasets GSE38713, GSE3365 and GSE24287. A shared legend categorizes genes by Log2 fold change and P. Value thresholds. In GSE38713, the volcano plot shows a dense cluster of significant genes with Log2 fold changes from -8 to 4 and P. Values up to 12, indicating many differentially expressed genes. The circular heatmap reveals alternating high and low expression, suggesting condition-based grouping. In GSE3365, the volcano plot has fewer significant genes, with Log2 fold changes from -3 to 4 and P. Values up to 20, highlighting a few genes with strong significance. The heatmap shows moderate expression variation. In GSE24287, the volcano plot displays sparse significant genes near fold change thresholds, with Log2 fold changes from -3 to 3 and P. Values up to 12. The heatmap indicates subtle expression differences, with an expression scale from -1 to 2.

Open in a new tab

Differential gene expression results. (A, C and E) Volcano plots for GSE38713, GSE3365, and GSE24287. Red dots: genes with both |log2FC| > 1 and P < 0.05 (significant and large change). Blue dots: P < 0.05 but |log2FC| ≤ 1 (significant but small change). Green dots: |log2FC| > 1 but P ≥ 0.05 (large but non‑significant change). Gray dots: neither threshold met. (B, D and F) Heatmaps of gene expression profiles for the three datasets. Red: high expression; blue: low expression.

Construction of WGCNA and Identification of the Module of Interest

WGCNA was performed using the GSE38713, GSE3365, and GSE24287 datasets. Sample clustering was used to generate a dendrogram (Supplementary Figure 3). Soft-thresholding powers, ranging from 1 to 15, were also evaluated. The scale-free topology fit index (R2) increased with power and first exceeded 0.85 at power = 6 (R2 = 0.861), with a mean connectivity of 3.75. At powers above 13, R2 dropped sharply (R2 = 0.473 at power = 14), and the mean connectivity approached zero, indicating excessive sparseness. Therefore, a power of 6 was selected for subsequent network construction (Figure 2A). Fourteen gene modules were identified (Figure 2B), and their eigengene correlations are shown as a dendrogram (Figure 2C). Among all modules, the blue module showed the strongest positive correlation with UC (r = 0.42, P = 3 × 10−8; Figure 2D). The correlation between module membership and gene significance for UC in the blue module was 0.53 (P = 2.8 × 10−83; Figure 2E). Therefore, subsequent analyses focused on 1136 genes in the blue module. All gene symbols, corresponding gene significance values, and P-values for each module are provided in Supplementary Table 4. Intersection analysis of DEGs from the three datasets, with blue module genes, yielded 142 overlapping DEGs (Figure 2F).

Figure 2.

WGCNA analysis: 2 line graphs, 2 dendrograms, heatmap, scatter plot, Venn diagram. The image A showing two line graphs. Left graph title, Scale independence. The x axis label is Soft Threshold power, unit not shown, with ticks 2, 4, 6, 8, 10, 12, 14. The y axis label is Scale Free Topology Model Fit, R superscript 2, unit not shown, ranging about 0.0 to 1.0. A dotted horizontal reference line is near 0.85. Plotted points rise from about 0.1 at power 2 to about 0.9 by power 6, stay near 0.9 through about power 12, then drop to about 0.5 at powers 13 and 14. A point near power 6 is labeled power equals 6. Right graph title, Mean connectivity. The x axis label is Soft Threshold power, unit not shown, ticks 2 to 14 by 2. The y axis label is Mean Connectivity, unit not shown, ranging about 0 to 1200. The curve drops from about 1200 at power 2 to about 300 at power 3, about 100 at power 4, then approaches near 0 from about power 6 through 14. The image B showing a dendrogram titled Gene dendrogram and module colors. The y axis label is Height, unit not shown, ranging about 0.5 to 1.0. Below the dendrogram are two horizontal annotation rows labeled Dynamic Tree Cut and Merged dynamic, each showing many narrow vertical blocks. The image C showing a dendrogram titled Clustering of module eigengenes. The y axis label is Height, unit not shown, ranging about 0.2 to 1.2. Leaf labels include MEgreenyellow, MEturquoise, MEgreen, MEblue, MEyellow, MEcyan, MEsalmon, MEblack, MEred, MEbrown, MEtan, MEgray, MEmagenta, MEmidnightblue, MEpink, MEpurple. A horizontal cutoff line is drawn near height 0.25. The image D showing a heatmap titled Module trait relationships. The x axis categories are HC and UC, unit not shown. The y axis lists MEgreenyellow, MEturquoise, MEblue, MEgreen, MEmagenta, MEmidnightblue, MEpink, MEpurple, MEcyan, MEsalmon, MEblack, MEred, MEbrown, MEtan, MEgrey. Each cell contains a correlation value with a parenthesized P value. A vertical color scale at right is labeled 0.5, 0, negative 0.5. The image E showing a scatter plot titled Module membership vs gene significance, cor equals 0.53, p equals 2.8e minus 83. The x axis label is Module Membership in blue module, unit not shown, ranging 0.0 to 0.8. The y axis label is Gene significance for UC, unit not shown, ranging 0.0 to 0.5. Points form an upward trend from low x and low y toward higher x and higher y. The image F showing a Venn diagram with two overlapping circles labeled DIFF and WGCNA. The left-only region shows 260, the overlap shows 142 and the right-only region shows 994.

Open in a new tab

WGCNA analysis. (A) Soft-thresholding power selection (power = 6, R2 = 0.861). (B) Hierarchical clustering dendrogram of co-expression modules (14 modules). (C) Clustering dendrogram of module eigengenes. (D) Heatmap of module-trait correlations (blue module showed the strongest correlation with UC, r = 0.42, P = 3 × 10−8). (E) Correlation between module membership and gene significance in the blue module (r = 0.53, P = 2.8 × 10−83). (F) Venn diagram showing overlap between DEGs and blue module genes.

Identification of Hub Genes by Machine Learning Algorithms

For algorithms that reported feature importance, only the top 10 features were retained; for the other algorithms, all identified features were considered. LASSO analysis identified 21 features with coefficients of >0 (Figure 3A). The RF algorithm selected the top ten most important features (Figure 3B). The Boruta algorithm filtered out irrelevant features and identified 22 features (Figure 3C and D). GBM and XGBoost identified ten features (Figure 3E and F). Intersection analysis of the features from the five algorithms yielded three optimal hub genes, TIMP1, S100A12, and ALPL (Figure 3G).

Figure 3.

Multiple graphs showing gene feature selection by LASSO, Random Forest, Boruta, GBM, XGBoost and a Venn diagram. The image A showing two LASSO line graphs. First graph: x-axis label, minus Log open parenthesis lambda close parenthesis, range 1 to 6; y-axis label, Coefficients, range about minus 2.5 to 2.0. Multiple gene coefficient curves start near 0 at x about 1 and diverge by x about 6, with one curve reaching about minus 2.5 and several reaching about 1.5 to 2.0. Second graph: x-axis label, minus Log open parenthesis lambda close parenthesis, range 1 to 6; y-axis label, Binomial Deviance, range about 0.8 to 1.8. A red point curve forms a U shape with the minimum near x about 3.3 at deviance about 0.9. The image B showing a horizontal bar chart titled, Random Forest Feature Importance. x-axis label, Mean Decrease Accuracy, range 0 to 10; y-axis label, Gene. TIMP1 is highest at 10, followed by VMP1 and ZYX. The image C showing a scatter and line plot titled, Boruta Feature Selection History. x-axis label, Attributes; y-axis label, Importance, range about minus 2 to 10. Points and box style summaries rise across attributes, with the highest values near 9 for VMP1. The image D showing a horizontal bar chart titled, Boruta Feature Importance. x-axis label, Mean Importance, range 0 to 8; y-axis label, Gene. Bars are split by selection categories labeled, Tentative and Confirmed. TIMP1 is highest at 8. The image E showing a horizontal bar chart titled, GBM Feature Importance. x-axis label, Relative Influence, range 0 to 16; y-axis label, Gene. TMEM158 is highest at 16. The image F showing a horizontal bar chart titled, XGBoost Feature Importance. x-axis label, Gain, range 0.00 to 0.25; y-axis label, Gene. TIMP1 is highest at 0.25. The image G showing a five-set Venn diagram labeled GBM, Boruta, XGBoost, RandomForest and LASSO, with a central overlap value of 3. A bar chart titled, Size of each list, shows GBM 10, Boruta 22, XGBoost 10, RandomForest 10 and LASSO 21.

Open in a new tab

Screening of core feature genes by machine learning algorithms. (A) LASSO coefficient profiles of candidate optimal feature genes; the optimal lambda was determined when the partial likelihood deviance reached its minimum. (B) Top 10 feature genes ranked by importance in RF. (C and D) Feature selection process and results of the Boruta algorithm. (E) Feature genes identified by GBM. (F) Feature genes identified by XGBoost. (G) Venn diagram showing the three intersecting feature genes from the five algorithms.

External Validation and Animal Experiments

To validate the robustness of the identified hub genes, their differential expression was examined using two independent GEO datasets (GSE92415 and GSE179285; Figure 4A). Except for ALPL, which was not significantly different in GSE179285, all other genes showed consistent and significant differential expression in the UC samples. TIMP1 exhibited the most prominent differences and was therefore selected for further investigation.

Figure 4.

Graphs and images show gene expression, colitis model, colonoscopy and histology in UC vs HC. The image A shows box plots of TIMP1, S100A12 and ALPL gene expression in UC and HC groups, with significant differences noted. The image B depicts a timeline of a DSS-induced colitis model with alternating DSS and water phases over 50 days. The image C presents colon images and length measurements, showing significant differences between UC and HC groups. The image D displays colonoscopy images with scores, indicating higher scores in UC. The image E shows H&E stained colon sections with histological scores, highlighting differences between UC and HC. The image F presents TIMP-1 IHC staining images with quantitative analysis, showing increased staining in UC. Statistical significance is marked with asterisks and data are presented as mean plus or minus SEM, n = 6, with p-values less than 0.001.

Open in a new tab

External validation of hub genes and establishment of a DSS-induced chronic colitis mouse model. (A) Expression of TIMP1, S100A12, and ALPL in colonic mucosa of UC patients versus healthy controls in GSE92415 and GSE179285. (B) Schematic diagram of the DSS-induced chronic colitis experimental protocol. (C) Representative colon images and colon length measurements (mean difference = 2.28 cm, 95% CI: 1.62–2.95 cm, ***P < 0.001). (D) Mouse colonoscopic images and MEICS scores (mean difference = 10.67, 95% CI: 9.64–11.69, ***P < 0.001). (E) Representative H&E staining images and histological scores (mean difference = 12.00, 95% CI: 10.20–13.80, ***P < 0.001). (F) Representative TIMP-1 IHC staining images and quantitative analysis of positively stained area (mean difference = 19.85%, 95% CI: 16.50–23.20%, ***P < 0.001). Data are presented as mean ± SEM, n = 6, ***P < 0.001.

A chronic colitis mouse model was established by the repeated administration of 1.5% DSS (Figure 4B). Colon length was significantly shortened in the model mice (Figure 4C). Miniature endoscopic imaging revealed signs of colonic mucosal inflammation in the colitis mice (Figure 4D). Histopathological evaluation revealed substantial colonic damage, including epithelial disruption, goblet cell depletion, and extensive inflammatory cell infiltration (Figure 4E). The area of TIMP-1-positive colon tissue was increased by 2.3-fold in the DSS-treated group compared to that in the controls (Figure 4F).

Single-Cell Atlas and Cellular Heterogeneity of the Mouse Colon in UC

Colons from DSS-induced chronic colitis mice were dissociated and subjected to scRNA-seq (Figure 5A). After quality control, 67,164 cells were retained for downstream analyses. Based on previously identified cell markers,32 six major cell types were distinguished: epithelial cells (Epcam, Krt8), stromal cells (Col1a1, Col1a2, Col6a1, and Col6a2), T cells (Cd52, Cd2, Cd3d, Cd3g, Cd3e), B cells (Cd79a, Cd79b), activated B cells (Cd52, Cd79a, Cd79b, Cd83), and myeloid cells (Cd14, Fcgr3, Cd68, Cd83, Csf1r, and Fcer1g) (Figure 5B and C). The mouse homologue of human S100A12 is S100a8.33,34 Among the three genes, Timp1 showed the highest expression in stromal cells (Figure 5D and E; Supplementary Table 5). Notably, stromal cells exhibited stronger interactions with other cells, particularly epithelial cells (Figure 5F).

Figure 5.

Diagram and plots of scRNA-seq analysis in mouse colon cells, showing cell types and gene expression. The image A shows a workflow diagram illustrating the process from C57BL/6J mouse colon treatment with DSS, through cell suspension, sequence transcriptome, to data analysis and visualization. The image B shows a UMAP plot of quality-controlled cells labeled by major cell types: epithelial cells, stromal cells, T cells, activated B cells, B cells and myeloid cells. The image C shows a dot plot displaying expression of marker genes for each cell type, with dot size indicating the percentage of cells expressing the gene and color intensity indicating average expression level. The image D shows UMAP density plots for Timp1, S100a8 and Alpl expression in colon tissues. The image E shows a dot plot of Timp1, S100a8 and Alpl expression across six major cell types. The image F shows cell-cell interaction networks based on CellChat, with line thickness representing interaction strength, highlighting communication between stromal cells and epithelial cells, among other cell types.

Open in a new tab

Single-cell atlas and cellular heterogeneity of the mouse colon in UC. (A) Workflow of scRNA-seq. (B) UMAP plot of all quality-controlled cells (n = 67,164), colored by major cell types. (C) Dot plot showing expression of representative marker genes for each cell type (dot size indicates the percentage of cells expressing the gene; color intensity indicates average expression level). (D) UMAP density plots showing Timp1, S100a8, and Alpl expression in colon tissues. (E) Dot plot of Timp1, S100a8, and Alpl expression across the six major cell types. (F) Cell-cell interaction network based on CellChat; line thickness represents interaction strength, highlighting communication between stromal cells and epithelial cells as well as other cell types.

Differentiation Characteristics of Timp1⁺ Stromal Cells in UC

Stromal cells clustered based on Timp1 expression levels (Figure 6A and B). CytoTRACE trajectory analysis was performed to assess the transcriptional heterogeneity. Timp1⁺ stromal cells were mainly enriched at late pseudotime stages, whereas Timp1− stromal cells were located at early stages, indicating that Timp1⁺ stromal cells have a higher proliferative potential (Figure 6C and D). TIMP1 was highly significantly associated with branch 2 (P = 2.81 × 10−11) but not with branch 1 (P = 0.10) (Figure 6E). The top 50 pseudotime-related genes were grouped into three modules based on their expression dynamics (Supplementary Table 6). Module 1 (18 genes) was highly expressed at early stages and was enriched in extracellular matrix organization, collagen fibril assembly, and response to cadmium ions, suggesting its involvement in matrix remodeling and stress responses. Module 2 (22 genes) peaked at mid-stages and was enriched in endothelial cell differentiation, blood vessel development, and angiogenesis, indicating an endothelial-like differentiation phase. Module 3 (10 genes) increased at late stages and was enriched in complement activation, TGF-β production, and humoral immunity, indicating an immunomodulatory phase (Figure 6F). GO enrichment analysis of branch-associated genes (Figure 6G) was mainly enriched in chemotaxis, cell-substrate adhesion, and regulation of epithelial cell proliferation, consistent with the functions of module 3.

Figure 6.

A multi graph figure of Timp1 stromal cells with tSNE, dot plot, trajectories, bar chart and GO bubbles. Image A: A tSNE scatter plot of stromal cells shows two distinct clusters, Timp1 minus and Timp1 plus Stromal, with minimal overlap. Image B: A dot plot displays Timp1 expression across identities, with percent expressed and average expression values. Image C: A trajectory scatter plot illustrates a branching path with a pseudotime gradient, labeled Branch1 and Branch2. Image D: The same trajectory layout highlights group labels, showing Timp1 minus and plus Stromal identities occupying different branch areas. Image E: A bar chart compares branch association strength, with Branch1 at 1.0 and Branch2 at 10.6 on a minus log10 P value scale. Image F: Three bubble charts depict gene ontology enrichment, with GeneRatio on the x-axis and gene ontology terms on the y-axis. Bubble size indicates Count and the padjust scale shows adjusted p values. Image G: A single bubble chart of gene ontology enrichment follows the same format as Image F.

Open in a new tab

Differentiation characteristics of Timp1⁺ stromal cells in UC. (A and B) Clustering analysis of stromal cells based on Timp1 expression levels. (C and D) CytoTRACE trajectory analysis showing that Timp1⁺ stromal cells are enriched at late pseudotime stages and have lower proliferative potential. (E) TIMP1 is significantly associated with branch 2 (P = 2.81 × 10−11). (F) Top 50 pseudotime-related genes grouped into three modules enriched in matrix remodeling, angiogenesis, and immune regulation, respectively. (G) GO enrichment analysis of branch 2-associated genes, highlighting chemotaxis, cell-substrate adhesion, and regulation of epithelial cell proliferation.

Cell-Cell Communication Mechanisms Mediated by Timp1⁺ Stromal Cells

A co-expression network was constructed using the CellChat software. Compared with Timp1− stromal cells, Timp1⁺ stromal cells exhibited higher responsiveness in both signal reception and transmission, suggesting a key role for Timp1 in UC regulation (Figure 7A). Timp1⁺ stromal cells showed distinct intercellular communication pathways: HGF, CADM, ncWNT, and SLIT were the main outgoing pathways, whereas ncWNT, SLIT, TNF, ADGRA, and TWEAK were the main incoming pathways (Figure 7B). Moreover, Timp1⁺ stromal cells displayed stronger interaction strengths in both the incoming and outgoing pathways than Timp1− stromal cells (Figure 7C). Ligand-receptor analysis revealed significant signaling interactions between Timp1⁺ stromal cells and epithelial cells (Figure 7D), suggesting that Timp1⁺ stromal cells may regulate epithelial cell function through paracrine mechanisms, thereby participating in colonic mucosal repair or pathological remodeling. In addition, enrichment of the Ppia-Bsg signaling pathway indicated that Timp1⁺ stromal cells may promote and aggravate inflammation by degrading the extracellular matrix and disrupting the intestinal barrier via matrix metalloproteinases (MMPs)downstream of the Ppia-Bsg axis.35

Figure 7.

Four-panel diagram of Timp1 superscript + stromal cell interactions and signaling pathways. A four-panel diagram illustrating Timp1 superscript + stromal cell interactions and signaling pathways. A) Two network diagrams show the number of interactions and interaction weights/strengths among EpithelialCells, Bcells, ActivatedBcells, Myeloidcells, Tcells and Timp1 superscript + Stromal cells. B) A heatmap displays outgoing and incoming signaling patterns for various cell types, with labels for each signaling pathway. C) A scatter plot depicts incoming versus outgoing interaction strengths for different cell types, with bubble size indicating count. D) Two dot plots show Timp1 superscript + Stromal as a source and target for outgoing and incoming signals, respectively, with ligand-receptor pairs and interaction probabilities indicated.

Open in a new tab

Cell-cell communication network mediated by Timp1⁺ stromal cells. (A) Comparison of signal reception and transmission capacities between Timp1⁺ and Timp1− stromal cells. (B) Outgoing and incoming signaling pathways of Timp1⁺ stromal cells. (C) Interaction strengths of Timp1⁺ stromal cells in incoming and outgoing pathways are significantly higher than those of Timp1− stromal cells. (D) Ligand-receptor analysis showing significant signaling interactions between Timp1⁺ stromal cells and epithelial cells.

Discussion

In recent years, the incidence of UC has continued to rise in Asia, severely affecting patients’ quality of life. The pathogenesis of UC is multifactorial, involving genetic predisposition, environmental factors, immunity, and the gut microbiota.36 Current therapies largely focus on epithelial barrier repair and the suppression of immune-mediated inflammation; however, a substantial proportion of patients fail to achieve sustained remission. Conventional anti-inflammatory drugs have been used for decades but often yield suboptimal results, and while biological agents are effective, concerns regarding adverse reactions and safety remain.37,38 Thus, there is an urgent need for new mechanistically well-defined therapeutic targets.

In this study, we integrated three independent UC transcriptomic datasets (GSE38713, GSE3365, GSE24287) after rigorous batch correction, combined weighted gene co-expression network analysis (WGCNA) with five machine learning algorithms (LASSO, random forest, Boruta, GBM, and XGBoost), and identified three robust core genes, TIMP1, S100A12, and ALPL. External validation using two independent cohorts (GSE92415, GSE179285) and a DSS-induced chronic colitis mouse model, followed by scRNA-seq, revealed that Timp1 is not broadly expressed in the mouse colon, but is instead highly enriched in a specific stromal cell subset. Further analyses showed that this Timp1⁺ stromal subpopulation exhibited a distinct shift in the differentiation trajectory, enhanced intercellular communication, and significant interactions with epithelial cells via the Ppia-Bsg signaling axis. These findings resolve the long-standing question of the cellular origin of TIMP1 at single-cell resolution and reveal a novel mechanism involving a stromal subpopulation in UC pathogenesis.

TIMP1, an endogenous inhibitor of matrix metalloproteinases, plays an important role in tissue remodeling and inflammation.39,40 However, studies on TIMP1 in UC remain relatively limited; most reports have included it only as a candidate gene within multigene panels,41,42 and its cellular origin has been debated for years without a clear answer. Our study provides several important biological insights by localizing Timp1⁺ cells specifically to a stromal subpopulation in the colon. First, it suggests that the upregulation of TIMP1 in UC primarily reflects an activated state of stromal cells rather than being a passive byproduct of epithelial injury. Second, pseudotime trajectory analysis showed that Timp1⁺ stromal cells predominantly occupy late differentiation stages, whereas Timp1− stromal cells are found at early stages, indicating that Timp1 expression may serve as a hallmark of stromal cell maturation toward a functional end-stage phenotype. Interestingly, a similar pattern has been observed in pulmonary arterial hypertension (PAH), where a highly active endothelial subpopulation exhibits progenitor-like features and drives disease progression,43 suggesting that functional reprogramming of specific stromal or endothelial subpopulations might be a common pathogenic mechanism in chronic inflammatory diseases.

Pseudotime trajectory analysis further delineated the dynamic functional evolution of Timp1⁺ stromal cells in UC. The top 50 pseudotime-associated genes were grouped into three modules based on their expression dynamics. Module 1 (early high expression) was enriched in extracellular matrix organization and collagen fibril assembly, indicating that Timp1⁺ stromal cells participate in matrix remodeling during the initial inflammatory phase. Module 2 (peak at intermediate stages) was enriched in endothelial cell differentiation and angiogenesis, suggesting a proangiogenic role in mid-stage inflammation. Module 3 (late increase) was enriched in complement activation, TGF-β production, and humoral immunity, indicating a transition toward immune and inflammatory regulation at late stages. This continuum demonstrates that Timp1⁺ stromal cells are not a static population, but are gradually shifted from matrix remodelers to immune/inflammation regulators in response to changing inflammatory microenvironments. GO enrichment analysis of branch 2-associated genes further supported the function of the late module, with terms including chemotaxis, cell-substrate adhesion, and regulation of epithelial cell proliferation, reinforcing the idea that Timp1⁺ stromal cells may play a dominant role during the chronic inflammatory phase of UC.

Cell-cell communication analysis revealed that compared to Timp1− stromal cells, Timp1⁺ stromal cells exhibited significantly stronger interaction strength in both incoming and outgoing pathways. The main outgoing pathways included HGF, CADM, ncWNT, and SLIT, whereas the main incoming pathways were ncWNT, SLIT, TNF, ADGRA, and TWEAK. Notably, significant enrichment of the Ppia-Bsg signaling axis was observed between Timp1⁺ stromal cells and epithelial cells. Ppia (cyclophilin A) is a secreted pro-inflammatory cytokine; its receptor Bsg (CD147) can induce the expression of MMPs, leading to the degradation of the extracellular matrix and disruption of the epithelial barrier.43,44 Previous studies have shown that CD147 expression is significantly elevated in the intestinal mucosa of patients with IBD and exacerbates intestinal inflammation by activating the NF-κB pathway.43 Moreover, CD147 can induce pyroptosis in intestinal epithelial cells, directly damaging barrier function and promoting the release of inflammatory cytokines.45 Serum cyclophilin A levels are significantly elevated in patients with UC and positively correlate with tissue MMP-9 levels.46 Although research on the Ppia-Bsg pathway in IBD is limited, its importance as a ligand-receptor interaction in various inflammatory diseases is widely recognized. The present study is the first to link this pathway to stromal-epithelial crosstalk in UC, suggesting that Timp1⁺ stromal cells aggravate epithelial barrier injury through the Ppia-Bsg axis. Therefore, targeting Timp1⁺ stromal cells or the downstream Ppia-Bsg signaling axis could represent a novel therapeutic strategy for UC.

S100A12 is a calcium-binding protein expressed mainly in neutrophils and monocytes and has been investigated as a disease marker for IBD.47,48 Our study found that S100A12 was significantly upregulated in UC patients and showed robust performance in external validation datasets. Single-cell data indicated that its mouse homologue, S100a8, is expressed in both myeloid and stromal cells, suggesting that S100A12 may amplify inflammation through multi-cellular synergy in human UC. ALPL encodes alkaline phosphatase, which plays an important role in maintaining the intestinal barrier function and detoxifying bacterial endotoxins.49 Although ALPL was prioritized by the machine learning ensemble, subsequent external validation showed that its upregulation was not consistently significant across both independent cohorts (GSE179285, P > 0.05), and our scRNA-seq data revealed minimal and non-specific expression of mouse Alpl across all colonic cell types. This discrepancy may reflect potential overfitting or batch effects inherent to bulk transcriptomic data, where ALPL’s apparent association with UC could be confounded by sample composition or platform heterogeneity. Therefore, we did not pursue ALPL further, although its high feature importance in the initial screening still warrants cautious attention in future studies. Compared with the study by Yang et al,12 which identified UC biomarkers using bulk transcriptomics and machine learning, our work not only screened for core genes but also determined, at single-cell resolution, the cellular origin and functional subpopulation of TIMP1, thereby overcoming a major limitation of conventional bioinformatics approaches. Similarly, Mennillo et al used single-cell and spatial multi-omics to reveal the effects of anti-integrin therapy in UC,15 highlighting the active role of stromal cells. Our study extends this concept by specifically linking Timp1⁺ stromal cells to shifted differentiation trajectories and enhancing -cell communication. Moreover, our strategy of combining five machine learning algorithms ensures the robustness and reproducibility of core genes, providing more reliable molecular targets for future clinical translation.

This study had several limitations. First, although multiple independent datasets were used for validation, all data were derived from the GEO database and lacked independent validation in prospective clinical cohorts. Second, the scRNA‑seq data were generated from a DSS‑induced mouse model, and species differences in intestinal immunity and stromal cell phenotypes between mice and humans exist. Therefore, the precise function of Timp1⁺ stromal cells in human UC must be confirmed using human biopsy‑derived single‑cell or spatial transcriptomics in future studies. Third, the interaction via the Ppia‑Bsg axis is currently based solely on bioinformatic predictions and has not yet been functionally validated; co‑culture systems and conditional knockout mice are needed to establish its causal role. Fourth, the specific cellular origins and mechanisms of ALPL and S100A12 in UC have not yet been explored in depth. Future studies will address these limitations using multiplex immunofluorescence for human tissue validation, as well as in vitro co‑culture assays and conditional gene‑knockout mouse models for functional interrogation.

Conclusions

In summary, this study identified core molecular markers of UC at the multi-dataset level and, at single-cell resolution, revealed the functional reprogramming and cell-cell communication network of a Timp1⁺ stromal cell subpopulation in UC pathogenesis. These findings provide new candidate targets for the molecular diagnosis of UC and lay a theoretical foundation for developing precise therapeutic strategies aimed at stromal cells.

Acknowledgments

We gratefully acknowledge the research community for making invaluable publicly accessible datasets available through the GEO repository.

Funding Statement

This work was financially supported by the National Science and Technology Major Project for the Prevention and Treatment of Cancer, Cardiovascular, Respiratory, and Metabolic Diseases (2024ZD0521004), the National Administration of Traditional Chinese Medicine Science and Technology Project (GZY-KJS-2023-025), the Medical Science Research Project of Hebei (20250859), and the Research Plan Project of Hebei Provincial Administration of Traditional Chinese Medicine (2026240), the National Administration of Traditional Chinese Medicine Project for Strengthening the Evidence-Based Research System of Traditional Chinese Medicine, the S&T Program of Hebei Province (246W7701D), the Hebei Innovation Special Project for Quantitative Research on Traditional Chinese Medicine (252W7712D, 262W7710D), the Hebei Provincial Natural Science Foundation (H2024423025), the 2025 Government-Funded Clinical Medical Outstanding Talents Training Project (No. ZF2025292), and the 2025 Hebei Provincial Higher Education Young Top-Notch Talents Program (No. BJ2025177).

Abbreviations

UC, Ulcerative Colitis; WGCNA, Weighted Gene Co-Expression Network Analysis; scRNA-seq, Single-Cell RNA Sequencing; TIMP1, Tissue Inhibitors Of Metalloproteases 1; DEGs, Differentially Expressed Genes; GEO, Gene Expression Omnibus; PCA, Principal Component Analysis; TOM, Topological Overlap Matrix; GBM, Gradient Boosting Machine; XGBoost, Boruta Algorithm, Extreme Gradient Boosting; RF, Random Forest; LASSO, Least Absolute Shrinkage And Selection Operator; H&E, Hematoxylin And Eosin; IHC, Immunohistochemistry; UMI, Unique Molecular Identifier; GO, Gene Ontology; PAH, Pulmonary Arterial Hypertension; Ppia, Cyclophilin A; MMPs, Matrix Metalloproteinases; DSS, Dextran Sulfate Sodium.

Data Sharing Statement

The datasets GSE38713, GSE3365, GSE24287, GSE92415, and GSE179285 analyzed in this study are publicly available in the NCBI Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/geo/). The data supporting the findings of this study are available from the corresponding authors upon reasonable request.

Ethics Approval

This study was conducted in strict accordance with the ethical principles outlined in the Declaration of Helsinki. All human transcriptome datasets analyzed in this study (GSE38713, GSE3365, GSE24287, GSE92415, GSE179285) are publicly available, de-identified open access data stored in the NCBI GEO database. Our research team did not collect raw human tissue, blood samples, or identifiable personal clinical information. According to the Measures for Ethical Review of Life Science and Medical Research Involving Humans, research using legally obtained publicly available data, which does not cause harm to human subjects or involve sensitive personal information or commercial interests, may be exempted from ethical review. Our study falls within this exemption category. Therefore, no additional Institutional Review Board (IRB/REC) approval is required for this study.

The animal experimental protocol was approved by the Laboratory Animal Welfare and Ethics Committee of the Hebei Provincial Hospital of Traditional Chinese Medicine (IACUC-HPHCM-2026014). All animal experiments reported in this manuscript were performed in strict accordance with the ARRIVE guidelines 2.0 for animal research reporting.

Author Contributions

Xiaosong Xu: Methodology, Writing – original draft, Writing – review & editing.

Lipeng Du: Methodology, Writing – review & editing.

Ze Li: Formal analysis, Writing – review & editing.

Xuemei Jia: Formal analysis, Writing – review & editing.

Yao Du: Formal analysis, Writing – original draft.

Qian Yang: Resources, Validation, Writing – review & editing.

Yanru Cai: Resources, Validation, Writing – review & editing.

Xiaosong Xu and Lipeng Du are co-first author. They have contributed equally to this article. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare that there are no conflicts of interest in this study.

References

  • 1.Kobayashi T, Siegmund B, Le Berre C, et al. Ulcerative colitis. Nat Rev Dis Primers. 2020;6(1):74. doi: 10.1038/s41572-020-0205-x [DOI] [PubMed] [Google Scholar]
  • 2.Le Berre C, Honap S, Peyrin-Biroulet L. Ulcerative colitis. Lancet. 2023;402(10401):571–19. doi: 10.1016/S0140-6736(23)00966-2 [DOI] [PubMed] [Google Scholar]
  • 3.Ng SC, Shi HY, Hamidi N, et al. Worldwide incidence and prevalence of inflammatory bowel disease in the 21st century: a systematic review of population-based studies. Lancet. 2017;390(10114):2769–2778. doi: 10.1016/S0140-6736(17)32448-0 [DOI] [PubMed] [Google Scholar]
  • 4.Raine T, Bonovas S, Burisch J, et al. ECCO guidelines on therapeutics in ulcerative colitis: medical treatment. J Crohn’s Colitis. 2022;16(1):2–17. doi: 10.1093/ecco-jcc/jjab178 [DOI] [PubMed] [Google Scholar]
  • 5.Alsoud D, Verstockt B, Fiocchi C, Vermeire S. Breaking the therapeutic ceiling in drug development in ulcerative colitis. Lancet Gastroenterol Hepatol. 2021;6(7):589–595. doi: 10.1016/S2468-1253(21)00065-0 [DOI] [PubMed] [Google Scholar]
  • 6.Planell N, Lozano JJ, Mora-Buch R, et al. Transcriptional analysis of the intestinal mucosa of patients with ulcerative colitis in remission reveals lasting epithelial cell alterations. Gut. 2013;62(7):967–976. doi: 10.1136/gutjnl-2012-303333 [DOI] [PubMed] [Google Scholar]
  • 7.Burczynski ME, Peterson RL, Twine NC, et al. Molecular classification of Crohn’s disease and ulcerative colitis patients using transcriptional profiles in peripheral blood mononuclear cells. J Mol Diagn. 2006;8(1):51–61. doi: 10.2353/jmoldx.2006.050079 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Chen Q, Bei S, Zhang Z, Wang X, Zhu Y. Identification of diagnostic biomarks and immune cell infiltration in ulcerative colitis. Sci Rep. 2023;13(1):6081. doi: 10.1038/s41598-023-33388-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Hu W, Fang T, Zhou M, Chen X. Identification of hub genes and immune infiltration in ulcerative colitis using bioinformatics. Sci Rep. 2023;13(1):6039. doi: 10.1038/s41598-023-33292-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Zhang B, Horvath S. A general framework for weighted gene co-expression network analysis. Stat Appl Genet Mol Biol. 2005;4:Article17. doi: 10.2202/1544-6115.1128 [DOI] [PubMed] [Google Scholar]
  • 11.Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinf. 2008;9:559. doi: 10.1186/1471-2105-9-559 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Yang Y, Hua Y, Zheng H, et al. Biomarkers prediction and immune landscape in ulcerative colitis: findings based on bioinformatics and machine learning. Comput Biol Med. 2024;168:107778. doi: 10.1016/j.compbiomed.2023.107778 [DOI] [PubMed] [Google Scholar]
  • 13.Tang D, Pu B, Liu S, Li H. Identification of cuproptosis-associated subtypes and signature genes for diagnosis and risk prediction of Ulcerative colitis based on machine learning. Front Immunol. 2023;14:1142215. doi: 10.3389/fimmu.2023.1142215 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Serigado JM, Foulke-Abel J, Hines WC, Hanson JA, In J, Kovbasnjuk O. Ulcerative colitis: novel epithelial insights provided by single cell RNA sequencing. Front Med. 2022;9:868508. doi: 10.3389/fmed.2022.868508 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Mennillo E, Kim YJ, Lee G, et al. Single-cell and spatial multi-omics highlight effects of anti-integrin therapy across cellular compartments in ulcerative colitis. Nat Commun. 2024;15(1):1493. doi: 10.1038/s41467-024-45665-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Park JM, Kim J, Lee YJ, Bae SU, Lee HW. Inflammatory bowel disease-associated intestinal fibrosis. J Pathol Transl Med. 2023;57(1):60–66. doi: 10.4132/jptm.2022.11.02 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Lu J, Li F, Ye M. PANoptosis and autophagy-related molecular signature and immune landscape in ulcerative colitis: integrated analysis and experimental validation. J Inflamm Res. 2024;17:3225–3245. doi: 10.2147/JIR.S455862 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Zhang T, Song B, Zhu W, et al. An ileal Crohn’s disease gene signature based on whole human genome expression profiles of disease unaffected ileal mucosal biopsies. PLoS One. 2012;7(5):e37139. doi: 10.1371/journal.pone.0037139 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Sandborn WJ, Feagan BG, Marano C, et al. Subcutaneous golimumab induces clinical response and remission in patients with moderate-to-severe ulcerative colitis. Gastroenterology. 2014;146(1):85–e15. doi: 10.1053/j.gastro.2013.05.048 [DOI] [PubMed] [Google Scholar]
  • 20.Keir ME, Fuh F, Ichikawa R, et al. Regulation and role of αE integrin and gut homing integrins in migration and retention of intestinal lymphocytes during inflammatory bowel disease. J Immunol. 2021;207(9):2245–2254. doi: 10.4049/jimmunol.2100220 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Leek JT, Johnson WE, Parker HS, Jaffe AE, Storey JD. The sva package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics. 2012;28(6):882–883. doi: 10.1093/bioinformatics/bts034 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Luo H, Zhang Z, Liu C, et al. An advanced diagnostic framework for discriminating lung cancer tissue subtypes via the synergy of Fourier transform infrared spectroscopy and random forest. Talanta. 2026;305:11. doi: 10.1016/j.talanta.2026.129632 [DOI] [PubMed] [Google Scholar]
  • 23.Montenegro, Juan D. Gene Co-expression Network Analysis. Methods Mol Biol. 2022;2443:387–404. doi: 10.1007/978-1-0716-2067-0_19 [DOI] [PubMed] [Google Scholar]
  • 24.Shen Y, Liao Z, Tian Y, et al. Surrogate-assisted differential evolution light gradient boosting machine for mangrove aboveground biomass inversion. Swarm Evol Comput. 2025;97:15. doi: 10.1016/j.swevo.2025.102013 [DOI] [Google Scholar]
  • 25.Chao T, Kean R, Cardenas A, et al. CSF immune factor-based extreme gradient boosting algorithm helps distinguishes infectious and noninfectious CNS disorders. J Neuropathol Exp Neurol. 2023;82(6):550. [Google Scholar]
  • 26.Amiri M, Pourghasemi HR, Ghanbarian GA, Afzali SF. Assessment of the importance of gully erosion effective factors using Boruta algorithm and its spatial modeling and mapping using three machine learning algorithms. Geoderma. 2019;340:55–69. doi: 10.1016/j.geoderma.2018.12.042 [DOI] [Google Scholar]
  • 27.Xia Y, McNicholas PD. A gradient method for the monotone fused least absolute shrinkage and selection operator. Optim Meth Software. 2014;29(3):463–483. doi: 10.1080/10556788.2013.801970 [DOI] [Google Scholar]
  • 28.Ma J, Zhang J, Wang Y, et al. Modified Gegen Qinlian decoction ameliorates DSS-induced chronic colitis in mice by restoring the intestinal mucus barrier and inhibiting the activation of γδT17 cells. Phytomedicine. 2023;111:11. doi: 10.1016/j.phymed.2023.154660 [DOI] [PubMed] [Google Scholar]
  • 29.du Sert NP, Hurst V, Ahluwalia A, et al. The ARRIVE guidelines 2.0: updated guidelines for reporting animal research. BMC Vet Res. 2020;16(1):7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Becker C, Fantini MC, Neurath MF. High resolution colonoscopy in live mice. Nat Protocols. 2006;1(6):2900–2904. doi: 10.1038/nprot.2006.446 [DOI] [PubMed] [Google Scholar]
  • 31.Tang X, Shang Y, Yang H, et al. Targeted delivery of Fc-fused PD-L1 for effective management of acute and chronic colitis. Nat Commun. 2024;15(1):1673. doi: 10.1038/s41467-024-46025-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Tao Y, Ma Y, Gu L, et al. Single-cell RNA sequencing reveals Shen-Bai-Jie-Du decoction retards colorectal tumorigenesis by regulating the TMEM131-TNF signaling pathway-mediated differentiation of immunosuppressive dendritic cells. Acta Pharmaceutica Sinica B. 2025;15(7):3545–3560. doi: 10.1016/j.apsb.2025.05.013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Ansari F, Pourjafar H, Sahebi J, Jodat V. Effect of Eudragit S100 nanoparticles and alginate chitosan encapsulation on the viability of Lactobacillus acidophilus and Lactobacillus rhamnosus. Arch AMB Express. 2017;7(1):144. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Kraemer AM, Saraiva LR, Korsching SI. Structural and functional diversification in the teleost S100 family of calcium-binding proteins. BMC Evol Biol. 2008;8. doi: 10.1186/1471-2148-8-48 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Salama Y, Munakata S, Osada T, Takahashi S, Hattori K, Heissig B. Heparin-binding EGF-like growth factor via miR-126 controls tumor formation/growth and the proteolytic niche in murine models of colorectal and colitis-associated cancers. Cell Death Dis. 2024;15(10):753. doi: 10.1038/s41419-024-07126-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Agrawal M, Allin KH, Petralia F, Colombel J-F, Jess T. Multiomics to elucidate inflammatory bowel disease risk factors and pathways. Nat Rev Gastroenterol Hepatol. 2022;19(6):399–409. doi: 10.1038/s41575-022-00593-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Dubois-Camacho K, Ottum PA, Franco-Muñoz D, et al. Glucocorticosteroid therapy in inflammatory bowel diseases: from clinical practice to molecular biology. World J Gastroenterol. 2017;23(36):6628–6638. doi: 10.3748/wjg.v23.i36.6628 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Chang S, Murphy M, Malter L. A review of available medical therapies to treat moderate-to-severe inflammatory bowel disease. Am J Gastroenterol. 2024;119(1):55–80. doi: 10.14309/ajg.0000000000002485 [DOI] [PubMed] [Google Scholar]
  • 39.Cabral-Pacheco GA, Garza-Veloz I, Castruita-de la Rosa C, et al. The roles of matrix metalloproteinases and their inhibitors in human diseases. Int J Mol Sci. 2020;21(24):9739. doi: 10.3390/ijms21249739 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Schoeps B, Frädrich J, Krüger A. Cut loose TIMP-1: an emerging cytokine in inflammation. Trends Cell Biol. 2023;33(5):413–426. doi: 10.1016/j.tcb.2022.08.005 [DOI] [PubMed] [Google Scholar]
  • 41.Yu X, Tang Y, Niu J, Hu J. Integrated multidimensional bioinformatics analysis of the molecular mechanisms of ulcerative colitis-associated colorectal cancer and MMP1 as a potential therapeutic target. Cancer Genet Ther. 2025;32(9):973–984. doi: 10.1038/s41417-025-00917-5 [DOI] [PubMed] [Google Scholar]
  • 42.Eiro N, Barreiro-Alonso E, Fraile M, González LO, Altadill A, Vizoso FJ. Expression of MMP-2, MMP-7, MMP-9, and TIMP-1 by inflamed mucosa in the initial diagnosis of ulcerative colitis as a response marker for conventional medical treatment. Pathobiology. 2023;90(2):81–93. doi: 10.1159/000524978 [DOI] [PubMed] [Google Scholar]
  • 43.Liang L, Lin R, Xie Y, et al. The role of cyclophilins in inflammatory bowel disease and colorectal cancer. Int J Bio Sci. 2021;17(10):2548–2560. doi: 10.7150/ijbs.58671 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Wu S, Zhang G, Yang J, Wang S. The single-cell and spatial transcriptomics atlas of epithelial-fibroblast interactions in colorectal cancer. Dig Dis Sci. 2026;2026;1–2. [DOI] [PubMed] [Google Scholar]
  • 45.Xu Z, Liu R, Huang L, et al. CD147 aggravated inflammatory bowel disease by triggering NF-κB-mediated pyroptosis. Biomed Res Int. 2020;2020:5341247. doi: 10.1155/2020/5341247 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Piechota-Polanczyk A, Włodarczyk M, Sobolewska-Włodarczyk A, et al. Serum cyclophilin A correlates with increased tissue MMP-9 in patients with ulcerative colitis, but not with Crohn’s disease. Dig Dis Sci. 2017;62(6):1511–1517. doi: 10.1007/s10620-017-4568-0 [DOI] [PubMed] [Google Scholar]
  • 47.Kałużna A, Jura-Półtorak A, Derkacz A, Jaruszowiec J, Olczyk K, Komosinska-Vassev K. Circulating profiles of serum proguanylin, S100A12 protein and pentraxin 3 as diagnostic markers of ulcerative colitis. J Clin Med. 2023;12(13):4339. doi: 10.3390/jcm12134339 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Wacker EM, Uellendahl-Werth F, Bej S, et al. Whole blood RNA sequencing identifies transcriptional differences between primary sclerosing cholangitis and ulcerative colitis. JHEP Rep. 2023;6(2):100988. doi: 10.1016/j.jhepr.2023.100988 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49.Xu Y, Shen J, Ran Z. Emerging views of mitophagy in immunity and autoimmune diseases. Autophagy. 2020;16(1):3–17. doi: 10.1080/15548627.2019.1603547 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The datasets GSE38713, GSE3365, GSE24287, GSE92415, and GSE179285 analyzed in this study are publicly available in the NCBI Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/geo/). The data supporting the findings of this study are available from the corresponding authors upon reasonable request.

Articles from Journal of Inflammation Research are provided here courtesy of Dove Press

RESOURCES