The gene expression profiling data for UC, encompassing GSE75214, GSE92415, and GSE16879 datasets, were retrieved from the gene expression omnibus (GEO) database.¹ The GSE75214 dataset included gene expression profiles from 97 individuals with UC and 11 controls, while the GSE92415 dataset comprised gene expression profiles from 87 pre-treatment UC patients and 21 healthy individuals. The GSE16879 dataset featured gene expression data from 24 pre-treatment UC patients and six control healthy volunteers.

Samples unrelated to this study were excluded, resulting in 246 samples included: 208 UC mucosal biopsy samples and 38 normal samples. The merged dataset was subjected to batch correction utilizing the “limma” and “sva” packages (32). Principal component analysis (PCA) was employed to evaluate the effectiveness of batch effect removal and to visualize the distribution of samples from UC and healthy individuals. Furthermore, the merged data underwent normalization utilizing the R package “preprocessCore.” Subsequently, 58 ARGs were extracted from the molecular signatures database,² as illustrated in Supplementary Table 1.

Identification of DEGs was carried out by the “limma” R package, ensuring the criteria of |log₂ fold change (FC)| > 0.5 and adjusted p-value <0.05 (33). Volcano plots were created using the R packages “limma” and “ggplot2” to show the distribution of DEGs. The “ComplexHeatmap” R package was employed to visualize the top 20 upregulated and downregulated genes. The Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were determined on the merged DEGs utilizing the R package “clusterProfiler” to explore differential biological pathways related to the signature genes. Statistically significance was defined as p < 0.05.

The overlapping genes between the upregulated and downregulated DEGs of UC and ARGs were assessed via the “limma” R package. Additionally, the GO and KEGG pathway enrichment analysis was conducted to further analyze the signaling pathways of anoikis-DEGs. Statistically significance was defined as p < 0.05.

To present a thorough overview of the differential expression of ARGs in UC and healthy individuals, the differential expression levels of ARGs were visualized in volcano plots and heatmap using the R packages “limma,” “ggplot2,” and “ComplexHeatmap.” Subsequently, we explored the expression of 21 anoikis-DEGs in UC and normal samples. Simultaneously, a protein–protein interaction (PPI) network analysis was conducted on 21 anoikis-DEGs utilizing the STRING database (https://string-db.org/) and displayed using the Cytoscape 3.9.0 software.

Three machine learning algorithms were utilized in this study to identify the key anoikis-DEGs as characteristic biomarkers in UC. The first algorithm employed was the least absolute shrinkage and selection operator (LASSO) logistic regression by using the R package “glmnet” (34). LASSO was selected for its ability to perform both variable selection and regularization, which helps identify important genes while reducing overfitting, especially in high-dimensional datasets. To ensure robustness and avoid overfitting of the model, we employed 10-fold cross-validations to determine the parameter λ. According to the optimal λ value and the corresponding coefficients, nine genes were obtained. The second algorithm was the random forest (RF) algorithm, assessed utilizing the R package “randomForest” to grow a forest consisting of 500 trees with default settings (35). RF algorithm, an ensemble learning method, was chosen for its robustness and ability to construct multiple decision trees and aggregate their predictions to improve accuracy and prevent overfitting. Using the RF algorithm, the top 10 genes with the highest importance were selected for downstream analysis. Finally, the support vector machine (SVM) algorithm was utilized, with the analysis conducted using the R package “kernlab” (36). The SVM algorithm can rank features based on recursion to avoid overfitting, and 13 genes were obtained through this approach. The five characteristic biomarkers for diagnosis were finally identified as the common core genes among the three algorithms. Furthermore, the five key anoikis-DEGs’ interactions were illustrated using the “chordDiagram” R package. Subsequently, the ROC curves and the area under the curve (AUC) were obtained utilizing the R package “pROC.”

Animal experiments were conducted in accordance with the ethical requirements and approved by the Tongji University Animal Ethics Committee. The C57BL/6 male mice were randomly allocated into normal control and colitis groups. The Control healthy mice were provided normal water. The colitis model was induced by administering 3.5% dextran sulfate sodium (DSS, Yeasen Biotechnology, 60316ES76, China) for 7 days, followed by 3 days of regular drinking water. Then, the mice colon tissues were obtained for RNA extraction to quantify the gene expression of the key anoikis-DEGs using Quantitative Real-time PCR (qPCR). Tissue total RNA was extracted using Trizol reagent (Invitrogen, 15,596,018, United States) and reverse to cDNA using an Evo M-MLV kit (Accurate Biology, AG11705, China). The Primer sequences were obtained from PrimerBank³ and synthesized at Sangon Biotech (Shanghai, China). qPCR was performed using the SYBR Green Pro Taq HS premix qPCR (Accurate Biology, AG11701, China) on a SLAN-96S Real-Time PCR System. The results were replicated three times, and the levels of target genes were normalized to GAPDH and analyzed utilizing the 2^{−(⊿⊿Ct)} method. CEACAM6 is not present in rodents, and H2-DMa is orthologous to human HLA-DMA. Thus, we analyzed the expression of four key anoikis genes: PDK4, H2-DMa, CX3CL1, and CFB. The following is a list of the primer sequences (5′-3′):

PDK4-F AGGGAGGTCGAGCTGTTCTC.

PDK4-R GGAGTGTTCACTAAGCGGTCA.

H2-DMa-F CTCGAAGCATCTACACCAGTG.

H2-DMa-R TCCGAGAGCCCTATGTTGGG.

CX3CL1-F ACGAAATGCGAAATCATGTGC.

CX3CL1-R CTGTGTCGTCTCCAGGACAA.

CFB-F GAGCGCAACTCCAGTGCTT.

CFB-R GAGGGACATAGGTACTCCAGG.

GAPDH-F TGGCCTTCCGTGTTCCTAC.

GAPDH-R GAGTTGCTGTTGAAGTCGCA.

Single sample GSEA (ssGSEA) is an extension of the GSEA method that enables the assessment of infiltrated immune cells and the activity of specific immune factors (37). ssGSEA was performed to measure the correlation coefficients between immune infiltration cells using the “GSVA” R package. Pearson’s correlation coefficient analysis was utilized to investigate the association between immune infiltrating cells. Heatmaps were generated using the R package “ggplot2.” Simultaneously, the “ggplot2” package was utilized to display immune cell infiltration as box plots in UC and healthy volunteers, with Pearson’s correlation coefficient evaluating relationships between different infiltrating immune cells. The association between the key anoikis-DEGs and immune infiltrating cells was assessed utilizing Pearson’s analysis and visualized in Lollipop plots using the R package “ggplot2.”

Next, the correlation of the key anoikis-DEGs with all UC genes was assessed, and a heatmap was utilized to display the 50 leading positively correlated genes. Additionally, Reactome pathway analysis was performed using GSEA for each diagnostic marker utilizing the R package “clusterProfiler” to explore the top 20 associated biological pathways.

To explore the functional role of ARGs in UC, unsupervised cluster analysis was performed utilizing the R package “ConsensusClusterPlus” as per the key anoikis-DEGs (38). Moreover, the differential expression of key anoikis-DEGs among different subtypes were explored. Moreover, a heatmap was generated utilizing the R package “pheatmap” to demonstrate the association between clinical features, gene expression, and subtypes.

Three gene sets (“h.all.v7.5.1.symbols.gmt,” “c2.cp.kegg.v7.5.1symbols,” and “c2.cp.reactome.v7.5.1.symbols”) were acquired from the MSigDB database⁴ as input files for gene set variation analysis (GSVA) (39). Further, pathway enrichment analysis was performed to identify distinct pathways associated with different subtypes.

To assess the reproducibility of the data, PCA was utilized to examine the overall distribution and determine the reproducibility of the data across subtypes. Additionally, a volcano plot was generated for visualizing the DEGs in distant subtypes. Enriched GO terms and KEGG pathways within the differential subtypes were analyzed. Furthermore, the correlation between the top five KEGG pathways and the DEGs in the distinct subtypes was established.

Transcription factors (TFs) and microRNAs (miRNAs) are pivotal regulators of gene regulations, functioning at the transcriptional and posttranscriptional levels, respectively. To explore TF- and miRNA-mediated gene interactions, the RegNetwork database⁵ was utilized to predict potential upstream miRNAs and TFs (40). Regulatory network of miRNAs and TFs were constructed based on key Anoikis-DEGs using Cytoscape 3.9.0 software.

Data analyses were performed utilizing R 4.3.3 and related R packages. GraphPad Prism 9.3.0 (GraphPad Software, Inc., La Jolla, CA, United States) was used for statistical analysis. All data are shown as mean ± standard deviation (SD). Differences between two groups were assessed using the t-test. The correlation between the two variables was determined by using the Pearson product–moment correlation coefficient. p < 0.05 indicated statistical significance.

To examine the biological role of ARGs in UC progression, the expression data of GSE75214, GSE92415, and GSE16879 datasets was firstly integrated. Tissues from different platforms exhibited distinct clustering patterns prior to batch effect removal (Figure 1A). The batch effects were effectively removed from the GEO datasets, resulting in an integrated dataset comprising 208 UC samples and 38 control samples (Figure 1B). Next, the data underwent normalization using the R software package “preprocessCore” (Figures 1C,D). We obtained 2,744 DEGs differentiating UC from normal tissues, comprising 1,577 upregulated and 1,167 downregulated genes, as illustrated by the volcano plot (Figure 1E) and the top 20 genes that have been upregulated and downregulated illustrated in a heatmap (Figure 1F).

To elucidate the underlying mechanisms and roles of these DEGs in UC, we carried out GO and KEGG pathway analyses. The GO enrichment analysis indicated these genes significantly involved in pathways such as cell adhesion, cytokine production, plasma membrane, and extracellular matrix (Figures 2A–C). Furthermore, the KEGG enrichment analysis demonstrated that the DEGs were predominantly enriched in multiple inflammatory response-related pathways, including the cytokine-cytokine receptor interaction, TNF signaling pathway, Th17 cell differentiation, and B cell receptor signaling pathway (Figure 2D).

To ascertain the role of ARGs in UC, the overlap between anoikis genes and DEGs in UC was initially identified. By intersecting the upregulated and downregulated DEGs with anoikis, 19 upregulated anoikis-DEGs (HLA-DMA, CFB, CD44, ITGA5, VCAM1, STAT1, PIK3R3, CCL20, NOS2, IRF1, CX3CL1, SIK1, CXCL10, CAV1, CEACAM6, MMP2, MX1, LTA, and SNAI2) and two downregulated genes (PDK4 and DAPK2) in UC were identified (Figures 3A,B). Subsequently, we conducted GO and KEGG enrichment analyses utilizing the 21 anoikis-DEGs. The GO term analysis highlighted that anoikis-DEGs exhibited enrichment in BPs, including cytokine-mediated signaling pathway, anoikis, regulation of cell–cell adhesion, and regulation of anoikis (Figure 3C). For CCs, enrichment was observed in the secretory granule membrane, focal adhesion, external side of the plasma membrane, and cell-substrate junction (Figure 3D). For MFs, the enriched functions included cytokine receptor binding, signaling receptor activator activity, receptor ligand activity, and G protein-coupled receptor binding (Figure 3E). KEGG analysis demonstrated anoikis-DEGs exhibited enrichment in the TNF signaling pathway, Chemokine signaling pathway, and Toll-like receptor signaling pathway (Figure 3F).

To evaluate the expression profiles of ARGs in UC and healthy individuals, we performed differential gene expression analysis. The volcano plot in Figure 4A illustrated the upregulation of 19 anoikis-DEGs and the downregulation of two genes in UC patients. Moreover, Figure 4B presented a heatmap displaying the differential expression of ARGs between UC groups and healthy control. Next, we compared the expression of these 21 anoikis-DEGs between UC patients and healthy individuals. Specifically, PDK4 and DAPK2 demonstrated relatively low expression in UC. In contrast, the other 19 genes (CEACAM6, CFB, HLA-DMA, CX3CL1, etc.) showed significant elevation in UC compared with normal samples (Figure 4C). Moreover, a PPI network was created to explore the relationship between anoikis-DEGs (Supplementary Figure 1). Among these, CXCL10, VCAM1, CCL20, STAT1, and CX3CL1 highlighted their interaction with other anoikis-related proteins. The PPI network highlights critical immune and inflammatory pathways, particularly involving VCAM1 and CXCL10, providing insights into the molecular mechanisms that may potentially contribute to UC pathogenesis.

Three machine-learning algorithms were employed to identify key anoikis-DEGs as characteristic biomarkers of UC from a set of 21 anoikis-DEGs. The LASSO Cox regression analysis demonstrated that nine genes were associated with UC, namely HLA-DMA, CFB, CCL20, CX3CL1, SIK1, CAV1, CEACAM6, PDK4, and DAPK2 (Figure 5A). The RF algorithm ranked the importance scores of signature genes and identified 10 genes, including CFB, HLA-DMA, CD44, CEACAM6, PDK4, VCAM1, CX3CL1, CAV1, STAT1, and IRF1 (Figure 5B). In the case of SVM outcomes, 13 signature genes were identified, namely ITGA5, HLA-DMA, CX3CL1, CD44, PDK4, VCAM1, STAT1, CFB, SIK1, CEACAM6, DAPK2, PIK3R3, and CCL20 (Figure 5C). Subsequently, after intersecting of the three machine learning results, five key anoikis-DEGs, namely PDK4, HLA-DMA, CEACAM6, CX3CL1, and CFB were identified (Figure 5D). Additionally, a correlation analysis was conducted on these five key anoikis-DEGs. Except for PDK4, which exhibited a negative correlation with the other four genes, all the other genes showed positive correlations with each other (Figure 5E). Subsequently, ROC analysis was carried out to assess the predictive value of these genes. ROC curves are commonly used to assess the performance of risk prediction models and the AUC quantifies this discriminative ability, with higher AUC values indicating better model performance. Moreover, an AUC curve that closely follows the top-left corner of the plot indicates higher accuracy, reflecting a model with strong predictive capability. In Figure 5F, the results indicated that each of the five genes exhibited relatively high predictive values (AUC values ≥0.921). Thus, these genes can serve as biomarkers for early diagnosis, disease monitoring, or evaluation the therapeutic effects. By conducting further research on the functions and mechanisms of these genes, we can gain a deeper understanding of the pathophysiological process of UC and provide clues for the development of new treatment approaches. By using the DSS-induced mouse colitis model, we aimed to substantiate the critical role of key anoikis genes in active UC. Because the CEACAM6 gene is not present in rodents, and H2-DMa is orthologous to human HLA-DMA, we ultimately analyzed the expression of four key anoikis genes: PDK4, H2-DMa, CX3CL1, and CFB in UC and normal samples (Figure 5G). The qPCR results showed an increase in H2-DMa, CX3CL1, and CFB in the DSS group, while PDK4 was decreased, consistent with the results from UC patients in Figure 4C.

To determine if UC patients exhibited altered immune system activity, an assessment of the infiltration and correlations of 23 distinct immune cell subtypes in UC was conducted. Figure 6A depicts multiple correlations among the infiltrating immune cells in UC. A significant synergistic effect was observed between activated B cell, activated CD4 cell, activated CD8 cell, activated dendritic cell, eosinophil, gamma delta T cell, macrophage, mast cell, monocyte, natural killer T (NKT) cell, natural killer (NK) cell, neutrophil, plasmacytoid dendritic cell (pDC), regulatory T cell (Treg), type 1 T helper (Th1) cell, and type 2 T helper (Th2) cell. Moreover, activated dendritic cell, macrophage, monocyte, neutrophil, pDC, and Th1 cell exhibited synergism with Th17 cell. In contrast, a competitive effect was observed between CD56^dim NK cells and most other immune cells. Subsequently, the immune infiltration analysis was performed and compared between the UC and normal groups. A significant elevation of 22 immune cell subtypes was observed in the immune cell infiltration within UC group (Figure 6B). These immune cell types included activated B cell, activated CD4 cell, activated CD8 cell, activated dendritic cell, CD56^bright NK cell, eosinophil, gamma delta T cell, immature B cell, immature dendritic cell, MDSC, macrophage, mast cell, monocyte, NKT cell, NK cell, neutrophil, plasmacytoid dendritic cell, regulatory T cell, T follicular helper cell, Th1 cell, Th17 cell, and Th2 cell. These cell types were remarkably abundant in the UC samples. Furthermore, lollipop charts present the associations between the five key anoikis-DEGs and immune cell infiltration (Figure 6C). CEACAM6, CFB, CX3CL1, and HLA-DMA exhibited positive associations with most of the immune cells. In contrast, PDK4 demonstrated a negative correlation with all immune cells.

The correlations between five key anoikis-DEGs and all merged genes were investigated, and the top 50 positively correlated genes were visualized in Supplementary Figure 2. Moreover, the Reactome pathway enrichment analyses of the five key anoikis-DEGs were conducted utilizing GSEA to identify significant biological signatures, and the top 20 pathways were presented (Supplementary Figure 3). For each key anoikis-DEGs, the top 20 pathways had the same adjusted p-value, which indicates that, statistically, they are equally important. In UC, the Reactome pathways most positively associated with CEACAM6 included the ER-phagosome pathway, antigen processing-cross presentation, asparagine N-linked glycosylation, and transport to the Golgi and subsequent modification. The Reactome pathways most positively associated with CFB were immunoregulatory interactions between a lymphoid and a non-lymphoid cell, interferon alpha/beta signaling, integrin cell surface interactions, and antigen processing-cross presentation. The top positively affected Reactome pathways of CX3CL1 included immunoregulatory interactions between lymphoid and non-lymphoid cells, interleukin-4 and interleukin-13 signaling, interferon alpha/beta signaling, and integrin cell surface interactions. The significantly enriched pathways of HLA-DMA included immunoregulatory interactions between a lymphoid and a non-lymphoid cell, cytokine signaling in immune system, signaling by interleukins, and interleukin-4 and interleukin-13 signaling. Most pathways positively related to PDK4 included the citric acid (TCA) cycle and respiratory electron transport, peroxisomal lipid metabolism, pyruvate metabolism and TCA cycle, and other pathways.

The 208 UC samples were classified as per the expression profiles of the five key anoikis DEGs utilizing an unsupervised clustering analysis. Various values of k were tested, varying from 2 to 9, and it was found that when k = 2, the consensus matrices showed the best performance, indicating that the optimal number of clusters was 2 (Figure 7A). Differences in the five key anoikis-DEGs expression between the two groups were observed. Specifically, higher expression of HLA-DMA, CEACAM6, CX3CL1, and CFB and lower expression of PDK4 were observed in cluster A (Figure 7B). Furthermore, a heatmap demonstrated the relationships among the clinical features, expression of the five key anoikis-DEGs, and subtypes (Figure 7C).

We subsequently conducted pathway enrichment analysis to investigate whether distinct pathways were perturbed in patients with different subtypes. The KEGG, Hallmark, and Reactome pathway analyses were performed to uncover the underlying pathways that differentiate cluster A from cluster B (Figure 8). Our findings revealed that these subtypes enriched in distinctive KEGG pathways. Cluster A displayed enhanced activation in antigen processing and presentation, intestinal immune network for IgA production, cytokine-cytokine receptor interaction, cell adhesion molecules, NOD-like receptor signaling pathway, Toll-like receptor signaling pathway, chemokine signaling pathway, and NK cell-mediated cytotoxicity. Furthermore, specific enrichment of tyrosine metabolism was observed in cluster B. In cluster A, Hallmark gene sets associated with the following pathways were upregulated: IL-6/JAK/STAT3 signaling, IL2/STAT5 signaling, apoptosis, TNF-α signaling via NF-κB, reactive oxygen species pathway, and mTORC1 signaling. In contrast, bile acid metabolism, oxidative phosphorylation, and adipogenesis were upregulated in cluster B. Reactome analyses indicated that antigen presentation: folding assembly and peptide loading of class I MHC, chemokine receptors bind chemokines, interferon signaling, and TNFR2 non-canonical NF-κb pathway were activated in cluster A. Taken together, these results suggest that different UC subtypes are regulated by distinct immune pathways.

The distinct distribution of subsets was visualized through PCA to account for the differences between cluster A and B subtypes (Figure 9A). Moreover, a differential gene analysis comprising 579 genes of the two subtypes was conducted, presented as a volcano plot (Figure 9B). The GO enrichment analysis highlighted that the DEGs were primarily linked to pathways such as cytokine-mediated signaling pathway, leukocyte cell–cell adhesion, cell chemotaxis in BP, external side of plasma membrane, secretory granule membrane, collagen-containing extracellular matrix in CC, receptor ligand activity, signaling receptor activator activity, immune receptor activity, cytokine activity, and cytokine receptor activity in MF (Figure 9C). KEGG analysis indicated enrichment in cytokine-cytokine receptor interaction, cell adhesion molecules, chemokine signaling pathway, NOD-like receptor signaling pathway, TNF signaling pathway, IL-17 signaling pathway, and inflammatory bowel disease (Figure 9D). Moreover, Figure 9E illustrated the relationship between the top five KEGG pathways and differential genes.

To examine the potential upstream regulators of the five key genes, the RegNetwork database was utilized to extract relationships between upstream miRNAs and TFs likely to bind to these five key genes (Supplementary Figure 4). The results revealed a multitude of miRNAs and TFs implicated in the modulation of these diagnostic genes. TP53, RARB, RXRB, and CTCF were identified as regulators of multiple genes. Especially, CTCF is critically involved in regulating CFB, CXCL3, HLA-DMA, and CEACAM6.