A Compendium of Mucosal Molecular Characteristics Provides Novel Perspectives on the Treatment of Ulcerative Colitis

Abstract

Background and Aims

Ulcerative colitis [UC] is a complex heterogeneous disease. This study aims to reveal the underlying molecular features of UC using genome-scale transcriptomes of patients with UC, and to develop and validate a novel stratification scheme.

Methods

A normalised compendium was created using colon tissue samples (455 patients with UC and 147 healthy controls [HCs]), covering genes from 10 microarray datasets. Upregulated differentially expressed genes [DEGs] were subjected to functional network analysis, wherein samples were grouped using unsupervised clustering. Additionally, the robustness of subclustering was further assessed by two RNA sequencing datasets [100 patients with UC and 16 HCs]. Finally, the Xgboost classifier was applied to the independent datasets to evaluate the efficacy of different biologics in patients with UC.

Results

Based on 267 upregulated DEGs of the transcript profiles, UC patients were classified into three subtypes [subtypes A–C] with distinct molecular and cellular signatures. Epithelial activation-related pathways were significantly enriched in subtype A [named epithelial proliferation], whereas subtype C was characterised as the immune activation subtype with prominent immune cells and proinflammatory signatures. Subtype B [named mixed] was modestly activated in all the signalling pathways. Notably, subtype A showed a stronger association with the superior response of biologics such as golimumab, infliximab, vedolizumab, and ustekinumab compared with subtype C.

Conclusions

We conducted a deep stratification of mucosal tissue using the most comprehensive microarray and RNA sequencing data, providing critical insights into pathophysiological features of UC, which could serve as a template for stratified treatment approaches.

1. Introduction

Ulcerative colitis [UC] is an inflammatory bowel disease characterised by idiopathic, relapsing mucosal inflammation, starting from the rectum and extending to the proximal segments of the colon.¹ In recent decades, the incidence rate of UC has been gradually increasing worldwide. However, the exact aetiology remains unclear and the pathogenesis is considered multifactorial, involving genetic susceptibility, epithelial barrier defects, dysregulated immune response, and environmental factors.² Currently, traditional drugs such as 5-aminosalicylic acid, glucocorticoids, and immunosuppressive agents are the standard treatment regimens. However, for patients with moderate to severe UC or steroid-dependent UC, biological agents such as anti-TNF-α agents [ie, infliximab and golimumab] and anti-adhesion molecule inhibitors, including vedolizumab [blocks the gut-homing α4β7 integrin], are commonly used in their treatment.^3–5 Additionally another novel agent, ustekinumab [anti-IL-12/IL-23] showed promising results in inducing clinical remission when anti-TNF therapy failed.⁶ Moreover, owing to its pro-inflammatory cytokine inhibitory properties, ustekinumab is considered the best choice for patients with extraintestinal manifestations.⁷ Although multiple treatment options are available for patients with UC, the current standard treatment regimen reports various side effects, inadequate response to the medication, and drug failure over time. Therefore, an increasing number of studies have sought to provide stratification recommendations by identifying and characterising patients with similar target states.

Thus far, the most commonly used method for classification according to disease severity and site is the Montreal classification, which divides UC into three subtypes: ulcerative proctitis, left-sided UC, and extensive UC. However, the Montreal classification only provides a clinically-based classification of UC.⁸ The recent transcriptome-wide advances have revealed various pathogenic pathways associated with UC.⁹ Lijie et al. stratified patients with UC according to transcriptome-wide gene expression data into four subgroups, of which six genes were identified as potential classification biomarkers.¹⁰ Despite the efforts of the above two studies in characterising disease subgroups and addressing disease heterogeneity in patients with UC, neither could identify subpopulations of patients who would best respond to specific medications.

Therefore, this study used the largest publicly available UC transcriptome datasets to type UC, using an unsupervised machine learning method. This study aims to gain deeper mechanistic insights into the divergent and shared features of the subtypes, by exploring their molecular and clinical characteristics. Moreover, the UC subclassification was applied to independent groups of patients with UC to evaluate the treatment effect of various drugs, providing new ideas for the treatment of drug-resistant patients from the perspective of disease heterogeneity.

2. Materials and Methods

2.1. Data processing and acquisition

Microarray gene expression datasets of mucosal biopsy samples (455 patients with UC and 147 healthy controls [HCs]) and two RNA sequencing datasets [100 patients with UC and 16 HCs] were collected from the Gene Expression Omnibus [GEO] database. The details of the study design and data pre-processing of the 10 microarray datasets [GSE87466,¹¹ GSE107499, GSE59071,¹² GSE48958,¹³ GSE47908,¹⁴ GSE36807,¹⁵GSE38713,¹⁶ GSE75214,¹⁷ GSE48634,¹⁸ and GSE13367¹⁹] and two RNA sequencing datasets [GSE130038²⁰ and GSE128682²¹] were presented in Supplementary Table S1. Several biologic agents were included, namely: anti-TNF–α (golimumab [GSE92415²²] and infliximab [GSE12251,²³ GSE16879,²⁴ and GSE23597²⁵], anti-α4β7-integrin (vedolizumab [GSE73661²⁶]), and IL-12–IL-23 blockers (ustekinumab [GSE206285²⁷]). Patients with a clear improvement in endoscopic mucosal healing [decreased to Mayo endoscopic subscore of 0 or 1 and a decrease to grade 0 or 1 on the histological score] were considered responders.

Moreover, microarray and RNA sequencing datasets were treated as training datasets and test datasets, respectively. For microarray data from Affymetrix®, we downloaded raw ‘CEL’ files and adopted a robust multiarray averaging approach with 'Affy' and 'Simpleaffy' packages to perform background adjustment and quantile normalisation. The normalised matrix files were directly downloaded for microarray data from the Illumina chip. For RNA sequencing data, the expression profile of RNA sequencing datasets was transformed into transcripts per kilobase millions [TPMs] equivalent to those generated from microarrays, to make the samples more comparable. The ‘ComBat’ algorithm of ‘sva’²⁸ R package was applied to remove batch effects in different datasets from training datasets and test datasets, separately. Before the adjustment, the clustering of samples was primarily driven by the batch effect, which is due to merging the data from different datasets [Supplementary Figure S1A, C]. After applying the ‘ComBat’ algorithm, the batch effect was mitigated [ie, samples from different datasets were mixed] [Supplementary Figure S1B, D]. After that, the meta-dataset removing the batch effect was applied for further analysis.

2.2. Differentially expressed genes: screening and function and pathway enrichment analysis

We filtered the differentially expressed genes [DEGs] between UC and HC samples using the ‘limma’ R package.²⁹ To adjust for false-positive results, the false-discovery rate [FDR] correction was used. DEGs were defined as having an adjusted p-value <0.05 and log fold change [logFC] >0.58. Pathway enrichment analyses of upregulated DEGs, namely gene ontology [GO] annotation, KEGG and Reactome pathway enrichment, were performed using Metascape³⁰; p <0.05 was considered the significantly enriched functional pathways of cut-off criterion.

2.3. Construction of protein-protein interaction network and module analysis

The search tool for the Retrieval of Interacting Genes Database³¹ was used to elucidate the complex regulatory relationships between proteins via the construction of a protein-protein interaction [PPI] network. Subsequently, a co-expression network was visualized using the Cytoscape software.^30,32 Furthermore, the molecular complex detection [MCODE] algorithm³³ was used to identify the most important modules.

2.4. Gene set enrichment analysis

To identify the potential enrichment pathways of the upregulated DEGs between UC and HC, gene set enrichment analysis [GSEA]³⁴ was conducted to assess UC-related over-representation. Gene sets were obtained from the KEGG and Reactome databases. Terms with FDR <0.05 were identified.

2.5. Classification of gene expression-driven subgroups in UC

To dissect molecular subtype heterogeneity defined by upregulated DEGs profiles related to UC, we performed hierarchical agglomerative clustering using the R package ‘ConsensuClusterPlus’.³⁵ Specifically, the ‘pam’ method based on Euclidean and Ward-D2 linkage was used in our analysis and performed 1000 times to ensure classification stability. The optimal cluster allocation was determined by the cumulative distribution function [CDF]. Furthermore, principal component analysis [PCA] was performed to confirm unsupervised clustering results. Upregulated DEGs were identified between the epithelial- or immune-enriched subtype with the other subtypes [A vs B+C, C vs A+B], and pathway enrichment was performed.

2.6. Evaluation of the cellular, molecular, and clinical characteristics of UC subtypes

Immune cell infiltration scores of the intestinal mucosa samples of patients with UC were quantified using the ‘Xcell’ R package, which evaluated the enrichment analysis of 10,800 genes and calculated the enrichment of 64 immune and stromal cell types.³⁶ The population of the activity of specific pathways was estimated using single-sample gene set enrichment analysis [ssGSEA]³⁷ which defines an enrichment score as representing the degree of absolute enrichment of a gene set in each sample within a given dataset. UC-associated pathways were curated from publicly available literature and GSEA results,³⁷ and the gene sets were from the KEGG and Reactome Databases. Additionally, the enrichment scores representing specific cell types and pathway activities among the three subtypes were estimated using the Wilcoxon test, and p <0.05 was considered statistically significant.

2.7. Construction of Xgboost classifier for subtype prediction

The Xgboost [extreme gradient boosting] algorithm was an efficient and scalable machine learning method for integrated learning classifiers, which extremely enhanced the model performance including the efficiency of the optimal solution and computational speed. The Xgboost-tree approach with a softmax objective function in a multiclassification model was used to predict subtype A, subtype B, and subtype C, based on 267 gene features, and to construct a decision tree of the model. We evaluated the performances of the predictive models using the area under curve [AUC] of the receiver operating characteristic curves [ROC]. A total of 455 patients with UC were categorised into training [n = 320] and testing [n = 135] at a 70% and 30% proportion, respectively, using the ‘caret’ R package. To assess the robustness of the patient’s molecular subtype characteristics, we further validated the classification in RNA sequencing independent datasets [n = 100]. The expression values of upregulated DEGs and subtype labels of the unsupervised clustering process results were used. We controlled overfitting with each training unit using 10-fold cross-validation and applied the fitted model to assign subtypes to the test sets.

2.8. Statistical analysis

All statistical analyses were conducted using R software [version 4.0.3]. The Kruskal–Wallis test was used to compare more than two groups, and the Wilcoxon test was used to compare two groups. Statistical significance values were set as two-tailed and p <0.05 was designated as statistically significant.

2.9. Ethical statement

This study was approved by the Ethics Committee of the Second Hospital of Shanxi Medical University [2019-YX-107].

3. Results

3.1. Acquisition and functional enrichment of DEGs and PPI network construction

A total of 267 upregulated DEGs were screened from the UC and HC samples [Figure 1A, B]. GO BP analysis revealed that these upregulated DEGs were primarily enriched in the inflammatory response, extracellular matrix, and positive regulation of immune response [Figure 1C]. Moreover, enrichment of the IL-17 signalling pathway and the extracellular matrix organisation was observed to be significant via KEGG and Reactome analyses [Figure 1D]. These terminologies were consistent with the current concept of UC pathophysiology. Using the 267 upregulated DEGs, PPI network analysis constructed an interaction network with 202 nodes and 612 edges, wherein nodes represented genes and edges represented gene interactions. Sixty genes were classified as hub genes and grouped into nine highly related protein clusters using the MCODE algorithm [Figure 1E].

Figure 1.

Identification of differentially expressed genes between patients with ulcerative colitis and healthy controls. [A, B] The heatmap and volcano plot of differentially expressed genes between patients with UC and HCs. [C, D] GO enrichment, KEGG and Reactome analyses of 267 upregulated differentially expressed genes. [E] Nine highly related protein clusters were identified using the MCODE algorithm. UC, ulcerative colitis; HC, healthy control; GO, gene ontology.

3.2. Classification of colonic mucosal gene expression-driven subgroups

To evaluate the optimal number of clusters from k = 2 to 6, we applied the ‘ConsensusClusterPlus’ R package with 1000 iterations. Subsequently, k = 3 was identified as the optimal number of clusters based on the CDF values and delta area [Figure 2A–C], ensuring the robustness of clustering results. Segregation of UC colonic mucosal subgroups was performed using PCA, and the different upregulated active genes in the three isoforms were visualised using a heatmap [Figure 2D, E]. To investigate differences in molecular processes and biological functions between subtypes, we analysed the upregulated genes that were significantly different between the three subtypes [Figure 3A, B]. A total of 135 genes were significantly upregulated in subtype A and 324 genes were substantially upregulated in subtype C, of which only one was shared [Figure 3C]. Further validation of the robustness of our subtyping could be attributed to the independent characteristics of the subtypes A and C. To explore the potential function of DEGs and each subtype’s most significantly dysregulated biological processes and signalling pathways between the three subtypes, enrichment analyses of Gene Ontology Biological Process [GO-BP], KEGG and Reactome databases were performed using Metascape. Notably, subtype A showed the strongest association towards cellular proliferation and was also highly active in transport functions, for example, sodium ion transport, transport of small molecules, and solute-carrier [SLC]-mediated transmembrane transport, whereas subtype C was significantly enriched in immune-inflammatory pathways, including inflammatory response, cell activation, signalling by interleukins, and neutrophil degranulation. Furthermore, significant differences [p < 0.001] were observed between molecular subclassification and disease status. Inactive patients were predominant in subtype A [85.7% in A vs 37.7% in B vs 20.3% in C], whereas active patients were more prevalent in subtype C compared with the other subtypes [14.3% in A vs 62.3% in B vs 79.7% in C] [Figure 3D].

Figure 2.

Consensus clustering of ulcerative colitis cohorts. [A] The consensus score matrix for ulcerative colitis samples when k = 3. A higher consensus score between two clusters indicated that they were more likely to be assigned to the same cluster in different iterations. [B] Consensus clustering for the cumulative distribution function for k = 2–6. [C] Relative changes in the area under the cumulative distribution function curve for k = 2–6. [D] Principal components analysis for the upregulated DEGs expression profiles showing the stability and reliability of the classification. [E] A heatmap of 455 UC patients with a red gradient illustrating the distribution of gene transcripts for three subtypes. In each column, patients are grouped based on cluster assignment. DEGs, differentially expressed genes; UC, ulcerative colitis.

Figure 3.

Gene expression characteristics of the three colonic mucosa subtypes. [A, B] Enrichment scores of functionally annotated clusters of the epithelial proliferation subtype [subtype A] and immune activation type [subtype C]. The top 20 most significant biological processes in the GO BP database and the top 5 most significant signalling pathways in the Reactome databases for each subtype compared with healthy controls. [C] A Venn diagram showing upregulated DEGs in subtypes A and C. [D] The disease activity among the three different subtypes presented using the boxplot. GO BP, gene ontology biological process; UC, ulcerative colitis.

3.3. Molecular and cellular characterisation of the three subtypes

The clustered subtypes were designated as subtype A [n = 118], subtype B [n = 233] and subtype C [n = 104]. To characterise the differences in immuno-inflammatory and cellular proliferation among these subtypes, we contrasted the enrichment scores of UC-related immune-related pathways and cellular components. Eight pathways or processes associated with UC were obtained from the literature and the KEGG and Reactome databases [Figure 4A]. The result revealed that subtype C was primarily enriched in immune-related pathways, such as interleukin [IL]-17, tumour necrosis factor [TNF], T cell receptor [TCR], and B cell receptor [BCR] signalling. However, subtype A was enriched less than the other subtypes for most pathways but was conspicuous for the peroxisome proliferator-activated receptor [PPAR] signalling pathway. Moreover, the ‘Xcell’ algorithm was used to estimate the enrichment of 64 different cell types. Consistent with previous pathway enrichment characteristics, immune and epithelial cells were differentially activated in all three subgroups. Epithelial cells were more activated in subtype A than in subtypes B and C; however, monocytes declined rapidly. Furthermore, immune cells, including B cells, CD4+ T cells, dendritic cells [DCs], and natural killer cells [NKs], were substantially infiltrated in subtype C. Additionally, subtype B was modestly moderately activated in all immuno-inflammatory and cellular proliferation pathways [Figure 4B].

Figure 4.

Pathway and cell subset-driven characterisation of ulcerative colitis subtypes. [A] Enrichment scores for pathways for each ulcerative colitis subtype. Box plots revealed pathway activation scores across the ulcerative colitis subgroups. [B] Cell subset enrichment scores according to ulcerative colitis colonic mucosa subgroups. Differences across the three subgroups were analysed using the Wilcoxon test; ns, not significant; *p <0.05; **p <0.01; ***p 0.001.

3.4. Validation of classification by RNA sequencing datasets

The robustness of classification results was confirmed by integrating two publicly available UC mucosal biopsy RNA sequencing datasets. These patients were categorised into three subtypes using the gene expression profiles of 258 upregulated DEGs (designated as subtype A [n = 31], subtype B [n = 35], and subtype C [n = 34]) [Supplementary Figure S2A–E]. By comparing the enrichment scores of ulcerative colitis-related pathways and cell subpopulations in the three subtypes [Supplementary Figure S3A, B], we obtained the same conclusion that subtype A was characterised by epithelial proliferation, subtype B was described as mixed, and C was described as an immune activation subtype.

3.5. Construction of a 267-gene classifier and variation between subgroups in response to treatment

A 267-gene classifier was developed using an Xgboost machine learning algorithm to validate the UC subtyping scheme. This classifier was tested on the training set [containing 320 UC samples] and applied to the validation Set I [containing 135 UC samples] and validation Set II [containing 100 UC samples] to demonstrate its efficiency. The 267-gene classifier reliably categorised the training set into three subtypes with an average AUC value of 97.18%. In addition, the classifier had an average classification performance of 94.81%, with an average AUC value of 100% in the validation Set I and an average classification performance of 79.00% with an average AUC value of 87.06% in the validation Set II, verifying the robustness of the classification model. Therefore, the classifier proved to be applicable and effective in the UC cohort, providing an effective classification strategy for the evaluation of UC colonic mucosal tissue subtypes in clinical trials.

Biologic agents may have different efficacies in patients with UC, depending on the pathological specificity of individual colonic mucosal tissues and their molecular activity. We evaluated the response to treatment with golimumab, infliximab, vedolizumab, and ustekinumab of different subtypes. In this study, subtype A and subtype B had consistently higher proportions of good responses than the C subtype. Additionally, all biologics had the most frequent response rates in subtype A. Regarding the TNF-α inhibitors [golimumab and infliximab], subtypes A [80% and 100%] and B [58.6% and 76.7%] exhibited good responses. Furthermore, subtype C was completely unresponsive to integrin α4β7 [ie, vedolizumab]. Moreover, ustekinumab, an IL-12/23 inhibitor, elicited an inadequate response in all three subtypes; however, subtype A [22.4%] displayed a relatively good response [Figure 5]. However, it was possible that these differences were not statistically significant due to insufficient sample size. In conclusion, we believed that the different molecular subtypes of colonic mucosa in ulcerative colitis may influence the efficacy of drug therapy and should be considered for future clinical use of drugs.

Figure 5.

Multiple biologic treatments respond to the ulcerative colitis subtypes. Response: responded to the biologics; non-response: did not respond to the biologics. [A] Response/non-response to golimumab: 80%/20% [4/1] in subtype A, 58.6%/41.4% [17/12] in subtype B, and 37.5%/62.5% [3/5] in subtype C. [B] Response/non-response to vedolizumab: 33.3%/66.6% [1/2] in subtype A, 12.5%/87.5% [2/14] in subtype B, and 0.0%/100.0% [0/5] in subtype C. [C] Response/non-response to infliximab: 100.0%/0.0% [5/0] in subtype A, 76.7%/23.3% [23/7] in subtype B, and 16.7%/83.3% [2/10] in subtype C. [D] Response/non-response to ustekinumab: 22.4%/77.6% [15/52] in subtype A, 14.1%/85.9% [32/195] in subtype B, and 3.0%/97.0% [2/65] in subtype C.

4. Discussion

In the present study, we classified the colonic mucosal tissues of UC based on the upregulated DEGs characterised by distinct features in terms of cellular components and immune-related pathways. Specifically, subtype A [epithelial proliferation subtype] had a transcriptomic signature in endothelial cells and tissue proliferative-related pathways, whereas subtype C [immune activation subtype] exhibited high enrichment in immune cells and proinflammatory activation-related pathways. Notably, subtype B was considered a mixed subtype. Furthermore, the three subgroups showed different responses in various clinical characteristics, such as disease activity and treatment response to biologics [Supplementary Table S2].

Studies on the molecular characterisation of UC colonic mucosa have provided new insights into the pathophysiological aspects of the disease, explaining the variability in clinical outcomes. Previous studies have mostly used the classical Montreal clinical classification, which has benefits in assessing disease prognosis and selecting the most appropriate therapy. It also divides UC into three subtypes according to the extent of spread: E1 [ulcerative proctitis], E2 [left-sided UC], and E3 [extensive UC]. Additionally, it also separates UC based on the disease severity into four subtypes: S0 [UC in clinical remission], S1 [mild UC], S2 [moderate UC], and S3 [severe UC].^8,38 Gu et al. reported two molecular subtypes, clusters A/B, which were characterised in UC by unsupervised clustering using 12 m6A RNA methylation regulators. Additionally, the above study also identified nine central genes, including BBS10 and WRN. These analyses can be a valuable indicator for developing new personalised therapeutic techniques for patients with UC; however, they do not provide an in-depth analysis of treatment measures.³⁹

In our study, the three subtypes responded differently to drug treatment; however, it remains unclear whether this difference was present from disease initiation or developed during disease progression. The data we collect on patients include different levels of disease activity. However, the impact of disease activity on genetically based differential classification is significant and hence the current results: patients with active subtype C are more common than patients with other subtypes and respond less well to biologic agents. Although the pathogenesis of UC is unclear, defects in mucosal barrier function and cytokine imbalance have been commonly hypothesised to play a crucial role in the aetiology of the disease.⁴⁰ According to our analysis, the epithelial proliferation subtype had the highest abundance of epithelial cells and responded well to various drugs. However, the immune activation subtype, which is characterised by immune cell infiltration and high disease activity, showed contrasting results. Moreover, its response to multiple drugs was also not ideal. The differences in disease activity and drug response of the different subtypes have potentially good clinical implications for the treatment of patients with UC. The formation of the mucosal barrier is an essential defence against external attack, which is essential for intestinal homeostasis and metabolite homeostasis. PPARγ, the master switch for the mucosal barrier, activates goblet cells which facilitate the formation of the mucosal barrier and provide protection to the colon.⁴¹ A potential implication of PPAR in inflammation was due to the antagonism between PPARγ and the pro-inflammatory cytokine TNF-α.⁴² Additionally, IL-22 ameliorated colitis-associated mucus layer disruption by enhancing the production of membrane-bound mucins [Muc1, -3, -10, and -13] that were associated with goblet cell depletion. Moreover, the ability of IL-22 to promote intestinal wound healing and intestinal epithelial cell proliferation in mice and humans has been reproducibly demonstrated by independent groups using different experimental approaches.⁴³ Furthermore, the blocking effect of anti-TNF-α treatment on the endogenous inhibitor of IL-22 has been reported to reduce the failure of goblet cells and the damage to the mucosal layer in UC caused by insufficient IL-22 secretion.⁴⁴ Thus, these findings suggest a mutually supportive relationship between the enrichment of the PPAR signalling pathway and IL-22 expression, highlighting the better outcomes of anti-TNFα drugs in the epithelial proliferation subtype.

As the immune activation subtype exhibited a less-than-ideal response to various drugs, we inferred that immune-active patients with highly immune infiltrative features are of the refractory UC type. Although vedolizumab can selectively block the transport of intestinal lymphocytes, studies have shown that there is no significant difference in T cell number or phenotype after treatment with a single dose of vedolizumab, confirming that vedolizumab does not interfere with lymphocyte migration into the central nervous system.^45,46 Meanwhile, the long biological half-life and low immune response rate of ustekinumab could lead to longer dosing intervals and drug onset times than the other drugs used in this study.^47–49 Therefore, an emerging strategy is to select drug combinations or small molecule inhibitors for patients with refractory disease.⁵⁰ Tofacitinib, a JAK inhibitor, induces and maintains relief of inflammatory symptoms primarily by inhibiting JAK1 and JAK3 and blocking the downstream effects of various pro-inflammatory cytokines, including IL-2, IL-3, IL-4, IL-5, IL-6, and IL-12.⁵¹ In a randomised, double-blind trial, patients administered a combination of infliximab and azathioprine achieved significantly higher rates of corticosteroid-free remission at 16 weeks [39.7%] than those administered infliximab or azathioprine alone [22.1%/23.7%]. Additionally, two patients who received treatment independently developed severe infections.⁵² The combination of a biologic agent and tofacitinib has also been widely reported; however, further evaluation of safety and efficacy is needed. Therefore, the effectiveness of small molecule inhibitors and drug combinations for refractory UC needs to be further explored, which could provide new insights into the treatment of patients with immune activation subtypes.

This study has several limitations. First, more meta-data would be ideal; however, obtaining data from the same clinical environment with identical features is difficult. Also, the complete annotation of clinical information for each UC sample is lacking. Second, despite collecting data from the largest publicly available dataset to date, our data are limited due to the insufficient duration of drug therapy and a lack of traditional drug categories. Thus more drug datasets need to be selected to explore immune-activating treatment methods. Furthermore, applying advanced machine learning techniques and standardising for this emerging technology could improve our understanding of UC at a systems level.

In conclusion, in this study, the most comprehensive colonic mucosal tissue transcriptomic data were used to segregate patients with UC into pathobiologically discrete subtypes, explaining three clustering subtypes in terms of disease heterogeneity and treatment response, to gain insight into the mechanisms and to design stratified treatments for patients. In the future, we should focus more on the choice of medication for patients with refractory ulcerative colitis, to avoid ineffective treatment.

Data Availability

The data underlying this article are available in the article and in its online Supplementary material.

Funding

This work was supported by the National Social Science Fund of China [21BTQ050] and the Key R&D Project of Shanxi Province [202102130501003].

Conflict of Interest

The authors have declared no conflict of interest.

Authors Contributions

Study design and manuscript writing: MJC and JWH. Data extraction, quality assessment, analysis and interpretation of data: MJC and JQ. All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be published. JWH had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.