Dissecting T-cell exhaustion heterogeneity and immune ecosystem dynamics in colorectal cancer through multi-omics machine learning Authors: Zhijing Zhang, Peng Ouyang, Kai Cui, Jianqin Lai, Yixiang Wen, Xin Deng, Wanyu Chen, Zhenhong Xian, Qi Qi, Zhen Bao, Jin Gong, Xiao He Published: 2026 ABSTRACT AbstractBackgroundImmunotherapy has shown limited efficacy in a substantial subset of CRC patients, yet the mechanisms underlying therapeutic resistance remain incompletely understood. T-cell exhaustion (TEX) in the tumor microenvironment has been identified as a pivotal driver of immune evasion and tumor progression. Dissecting its contribution to CRC is essential for the development of rational therapeutic strategies.MethodsWe integrated scRNA-seq and bulk transcriptomic data to identify CD8⁺ T-cell exhaustion core genes via hdWGCNA and ten machine learning algorithms, constructed a multivariate Cox-based TEX score model validated across independent cohorts and immunotherapy datasets, and experimentally confirmed our findings by RT-qPCR, Western blot, and quantitative multiplex immunofluorescence in clinical CRC specimens.ResultsOur single-cell analysis revealed a continuum of intra-tumoral CD8⁺ T-cell exhaustion states, identified a five-gene TEX score (KLF3, LMNA, SLC2A3, ARL4C, TIMP1) that predicted poor prognosis and an immunosuppressive microenvironment. Further experimental validation confirmed the differential expression and spatial co-localization with CD8⁺ T cells in clinical specimens.ConclusionsOur findings implicate TEX as a central mediator of immunotherapy resistance in CRC, offering a clinically actionable framework for patient stratification and therapeutic decision-making.Graphical abstract Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16343-1.Keywords: Colorectal cancer (CRC), Immunotherapy sensitivity, T cell exhaustion (TEX), Tumor microenvironment (TME), Machine learning, Predictable model Introduction Colorectal cancer (CRC) ranks as the third most prevalent malignancy globally and constitutes the second leading cause of cancer-related mortality worldwide [1]. Currently, a multidisciplinary paradigm centered on surgical resection remains the cornerstone of curative-intent therapy for CRC patients [2]. While resection with curative intent successfully eliminates the primary tumor in most patients, a notable proportion experience distant recurrence: roughly 5% of AJCC stage I, 15% of stage II, and 40% of stage III CRC ultimately develop metastatic disease in the years following surgery [3]. Concomitantly, conventional chemotherapeutics confer considerable risk of severe adverse effects and frequently engender acquired therapeutic resistance [4, 5]. The emergence of immune-checkpoint blockade (ICB) has redefined the therapeutic landscape of solid tumors, yet its impact on CRC remains circumscribed [6] due to the molecular and clinical heterogeneity in CRC [5, 7]. Only subgroups with mismatch repair-deficient or microsatellite instability-high (MSI-H) of 5 ~ 15% of all CRC can benefit from immune checkpoint inhibitors (ICIs) [8]. To extend clinical benefit to a broader patient population, it’s crucial that we elucidate the mechanism of ICI resistance in CRC patients, and sensitize the patients through combining with other treatments [9]. CD8+ T cells represent a cornerstone of cancer immunotherapy [10, 11], however, prolonged antigen exposure and a suppressive tumor microenvironment caused progressive loss of effector function, leading to the exhaustion of CD8+ T cells [12]. Exhausted T cells represent a distinct lineage characterized by this progressive, hierarchical functional impairment, and their presence critically determines ICI therapeutic outcomes [13, 14]. While diverse immunosuppressive populations contribute to T-cell dysfunction, this study specifically interrogates the intrinsic heterogeneity of CD8⁺ TEX states, which constitute the direct therapeutic targets of ICIs [15]. Here, we leverage integrated single-cell and bulk transcriptomic analyses to dissect the molecular architecture of CD8⁺ TEX in CRC. We identify core regulatory genes governing exhaustion heterogeneity and construct a multi-biomarker prognostic model that stratifies patients based on TEX signatures. This framework enables comprehensive evaluation of the TME, prediction of immunotherapy responsiveness, and refined prognostication. Collectively, our findings illuminate novel molecular regulatory mechanisms underlying CD8⁺ TEX and provide a clinically actionable foundation for precision immunotherapy. Method Data acquisition The Bulk seq dataset was derived from the TCGA and GEO databases (GSE39582); Single-cell sequencing dataset (scRNA seq) was derived from the GEO database (GSE132465). Single-cell data processing and quality control Data analysis and quality evaluation using the R package “Seurat” (version 5.3.0; https://satijalab.org/seurat/). Firstly, cells expressing ≥ 300 genes and each gene being detected in ≥ 5 cells were retained, and the proportions of mitochondrial genes, red blood cell genes, and ribosome genes in each cell were calculated. Cells with a proportion of mitochondrial genes lower than 25%, a proportion of ribosome genes higher than 3%, and a proportion of red blood cell genes lower than 1% were screened. The cells after quality control were standardized. LogNormalize (with a scaling factor of 10,000) was used to correct the differences in sequencing depth, 2000 highly variant genes were screened to capture biological heterogeneity, and Z-score conversion was used to achieve the unification of gene expression level scales. The R package “Harmony” (version 1.2.3; https://www.harmony.r-project.org/) was used to eliminate the batch effect. Visualization was carried out using t-SNE. Single-cell annotation Cluster with a resolution of 1 to obtain different cell populations. In order to identify and annotate specific cell types, marker genes of multiple cell populations were selected, including: Epithelial cells marker gene EPCAM (Epithelial Cell Adhesion Molecule); CD8+ T cells (CD8+ T cells) marker genes CD3D and CD8A; B cells (B cells) mark genes CD79A and MS4A1 (Membrane Spanning 4-Domains A1); Macrophage marker genes CD68 and CD163; Fibroblasts marker gene COL1A1 (Collagen Type I Alpha 1 Chain); endothelial cells marker gene PECAM1 (Platelet and Endothelial Cell Adhesion Molecule 1). Identification and analysis of CD8+ T cell subtypes CD8+ T cells were isolated and re-clustered using the R package “Seurat”. Subsequently, a single-cell pseudo-time trajectory was constructed using the R package “Monocle2”. Next, weighted gene-related network analysis was conducted using the R package “hdWGCNA” to identify the core gene set in the CD8+ T cell population. The R package “Cellchat” was used to explore the intercellular communication among all cell populations. The functional status of the gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways in each CD8+ T cell population was analyzed for differences using the R package “ClusterProfiler”. Furthermore, the GSEA pathway from “MSigDB” (gsea-msigdb.org) was evaluated using the R package “fgsea”. Finally, gene set variation analysis (GSVA) was conducted through the R package “GSVA” to analyze the differences among populations in the HALLMARK pathway. Gene-specific analysis (GSA) was implemented to compare T cell cluster 1 against cluster 2 within the tumor microenvironment. Differentially expressed genes were identified using stringent statistical thresholds (p < 0.05; |log2 fold change| ≥ 2). The dot plot visualization was generated to display the expression patterns of established T cell exhaustion markers, including inhibitory receptors, transcription factors, and effector molecules across defined T cell clusters. Dot color intensity represents relative score, while dot size indicates relative magnitude. Construction and validation of TEX markers Next, to further screen the core genes, we employed a variety of machine learning methods, including: Partial Least Squares (PLS) [16], Random Forest (RF) [17], Decision Tree (DTS) [18], Support Vector Machine (SVM) [19], Logistic Regression (Logistic) [20], K-Nearest Neighbor Algorithm (KNN) [21], XGBoost [22], Gradient Boosting Machine (GBM) [23], Neural Network (NeuralNet) [24], Generalized Linear Model Boosting (glmBoost) [25]. Through these methods, we screened out the core genes related to prognosis. To visualize the screened core genes, we utilized the Upset graph to display the intersections among the core genes. Finally, we constructed a prognostic model using the Cox regression model, where the training set was from the TCGA dataset and the validation set was from the GSE39582 dataset. The performance of the prognostic model was evaluated by calculating the area under the curve (AUC) and the receiver operating characteristic curve (ROC), and was visualized. The difference analysis between the two groups was processed using the R packages “survminer”, “survival”, “rms” and “timeROC”. Clinical features and nomogram establishment We employed univariate Cox regression and multivariate Cox regression analyses to assess the correlation and independence of risk values and clinical parameters in the TCGA cohort. To depict the differences among patient subgroups, a nomogram was established. This nomogram can accurately predict the probability of an individual experiencing a certain event in a clinical setting, combining independent clinical prognostic factors such as age, gender, and vascular infiltration. Subsequently, the performance of the nomogram in prognosis prediction was evaluated through calibration curves and ROC curves [26]. Evaluation of immune-related characteristics A variety of methods, such as CIBERSORT [27], EPIC [28], MCP_counter [29], Quanti-seq [30] and xCell [31], were adopted for immune infiltration analysis. Based on the CIBERSORT results, we also conducted a correlation analysis between immune cells and genes. Meanwhile, the immune capacity of the tumor microenvironment (TME) was evaluated using seven different steps of the tumor immune cycle (TIP, hrbmu.edu.cn) and various immune indicators calculated through the “easier” package. In addition, we also examined the expression levels of co-stimulatory molecules, co-inhibitory molecules and HLA molecules. We calculated the differences in parameters such as T-cell inflammatory gene expression profile (GEP), cytotoxic activity (CYT), and IFN-γ between the high-risk and low-risk groups. We also conducted an immune checkpoint correlation analysis to further explore the role of the immune system in different risk groups. Tumor Immune Dysfunction and Exclusion (TIDE) To evaluate the association between TEX scores and immunotherapy efficacy, we employed the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm (http://tide.dfci.harvard.edu/) to compare immune escape scores between high- and low-TEX scores groups. TCGA-COAD TPM expression data were normalized and submitted to the TIDE web portal for computational assessment. Subsequently, Spearman correlation analysis was performed to quantify the relationship between TIDE scores and TEX scores. Prediction of immunotherapy In this study, we conducted immunotherapy validation through the dataset of GSE78002 (HeLa cells), GSE126044 (non-small cell lung cancer), and the clinical trial cohort of immunotherapy for urothelial carcinoma led by the MD Anderson Cancer Center in the United States (IMVigor) [32, 33]. To evaluate the response to immunotherapy, the proportions of responders and non-responders in the high-risk group and the low-risk group were calculated. In addition, we evaluated the performance of the immunotherapy prediction model by calculating the ROC curve and AUC value. The ROC curve is generated by correlating the immune response with the risk-score and using the AUC value to quantify the accuracy of the predictive model [34, 35]. Patients and tissue samples After approval by the Institutional Review Board, the studies were performed in accordance with the International Code of Ethics for Biomedical Research involving humans (CIOMS). We obtained written consent from the subjects before the study was performed. Paired tumor specimens were obtained from six CRC patients treated at the First Affiliated Hospital of Gannan Medical University between 2025 and 2026, comprising three responders and three non-responders to immunotherapy. Specimens were collected from the CRC tissue bank of the First Affiliated Hospital of Gannan Medical University (Approval No. LL8C-2025333). The specimens were properly stored in 4% paraformaldehyde (biosharp, BL539A) or a -80 ℃ refrigerator after acquisition. Real-time quantitative fluorescence polymerase chain reaction (RT-qPCR) After sample collection, total RNA was extracted using the total RNA Extraction Kit (NCM Biotech, M5105). Subsequently, cDNA was obtained following the steps of the Reverse Transcription Kit (Servicebio, G3331-50). Finally, the qPCR reaction was carried out using Universal Blue SYBR Green qPCR Master Mix (Servicebio, G3326-01). The quantification of relative gene expression levels was conducted using the 2-△△CT method. The primer sequence is shown in Supplemental Table 1. Western blot After sample collection, total protein was extracted using RIPA lysate (Thermo Fisher Scientific) containing PMSF (Servicebio, G2008-1ml) and phosphatase inhibitors (GLPBIO, GK10011, GK10012). Subsequently, protein quantification was performed using Coomas brilliant Blue (Servicebio, G2039-250ML), followed by SDS-PAGE electrophoresis and electrotransfer onto NC membranes (ShareBio, SB-WB310). After blocking with 5% skimmed milk, the primary antibody was incubated overnight at 4 ℃. Next, after incubating the secondary antibody at room temperature for 1 h, it was developed in the electrochemiluminescence imaging system (Tanon, Shanghai, China). All the antibodies involved are listed in the Supplemental Table 2. Immunofluorescence The specimens were immersed in 4% paraformaldehyde and fixed overnight. Then, the tissues were paraffin-embedded and cut into 6 μm sections. After reacting the sections with the specific primary antibody, they were incubated with the corresponding fluorescent secondary antibody, and the nuclei were stained with DAPI (4’,6-diamidino-2-phenylindole). The staining results were collected by a fluorescence sectioning scanner, and then the data were processed with CaseViewer to analyze the expression levels of each marker and their spatial correlations. Fluorescence integrated density measurements, cell counting, and colocalization metrics were made in Fiji/ImageJ (v2.16.0). Statistical analysis All statistical analyses were performed using R software (version 4.1.0) and IBM SPSS Statistics (version 21.0). Correlations were assessed using Spearman’s rank correlation coefficient. Comparisons between the two risk groups were conducted using the Wilcoxon rank-sum test. A two-tailed p value < 0.05 was considered statistically significant. *p < 0.05, **p < 0.01, ***p < 0.001. Result Comprehensive single-cell atlas of the CRC immune microenvironment To explore the complete cell atlas of CRC and its potential cell interactions, we conducted an in-depth analysis of the single-cell dataset in the GEO database (GSE132465). After rigorous quality control and data curation of single-cell transcriptomes from 23 CRC patients, cells were subjected to dimensionality reduction via t-Distributed Stochastic Neighbor Embedding (t-SNE) for visualization. Unsupervised clustering analysis using the Seurat framework revealed an optimal resolution partitioning cells into 24 distinct clusters (Fig. 1A). Differential gene expression (DGEs) was identified through pairwise comparisons of each cluster against all others for initial annotation across the 24 clusters, we annotated cellular subpopulations within the colorectal tumor microenvironment, encompassing both immune cells (B cells, CD8+ T cells, and macrophages) and non-immune cells (endothelial cells, epithelial cells, and fibroblast) (Fig. 1B and C). Given that CD8+ T cells had always served as the principal effectors of anti-tumor immunity [11], represented the primary cellular targets of ICIs [36], and implicated the ICIs’ resistance directly [37], we subsequently focused our analysis on this population. After a comprehensive examination of the landscape and dynamics of CD8+ T cell, we therefore redefined CD8+ T cell subsets based on their distinct functional states (Fig. 1D). Notably, cluster 2 exhibited a terminal exhaustion signature marked by elevated expression of TOX, PD-1, TIM-3, LAG-3, TIGIT, alongside metabolic marker CD39 and tissue residency marker CD103 (Fig. 1E). Conversely, cluster 1 retained effector potential with heightened OX40, GITR, and DNAM-1 expression (Fig. 1E). These findings identify cluster 2 as terminally exhausted T cells. Comprehensive single-cell atlas of the CRC immune microenvironment. A t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of transcriptionally distinct cell clusters identified from scRNA-seq analysis of CRC samples. B Annotation of six major cell populations based on established lineage-specific marker genes. Colors denote cell types: T cells (Coral), B cells (SaddleBrown), Macrophages (Turquoise), epithelial cells (SteelBlue), fibroblasts (Red), and endothelial cells (Purple). C Dot plot showing expression profiles of canonical marker genes across the six major cell subsets. Dot size represents the percentage of cells expressing the gene; color intensity indicates average expression level. D Sub-clustering analysis of CD8⁺ T cells, revealing two distinct subpopulations: cluster 1 (early activation state, Lightblue) and cluster 2 (exhaustion state, Steelblue), as defined by differential expression of activation and exhaustion markers. E Dot plot showing the relative genes expression used to identify CD8+ exhausted T cells in different cluster Integrated analysis of CD8⁺ T cell intercellular communication networks in the CRC microenvironment To further contextualize the functional relationship between the two CD8⁺ T cell subsets, computational pseudo-time ordering was employed as an exploratory framework to visualize potential transcriptional continuums between cluster 1 and cluster 2 cells (Fig. 2A). While cross-sectional datasets preclude true temporal resolution, we leveraged intercellular communication networks to delineate the functional roles of distinct CD8⁺ T cell clusters within the TME. Malignant epithelial cells, exhibiting maximal interaction number and weights, acted as active architects, not passive bystanders (Fig. 2B and C). Notably, within the CD8+ T cell compartment, cluster 2, which exhibited a terminal exhaustion signature, displayed significantly elevated interaction frequency and strength relative to cluster 1. (Fig. 2B and C). This apparent paradox, wherein terminally exhausted T cells maintain extensive network connectivity despite effector dysfunction, prompted us to interrogate the underlying signaling pathway of these interactions. Intercellular communication analysis of CD8⁺ T cells. A Pseudo-time trajectory analysis of CD8⁺ T cell subsets constructed using Monocle2. Cells are ordered along a computational reconstruction of transcriptional similarity from cluster 1 to cluster 2, with darker colors indicating inferred pseudo-time values. This analysis serves as an exploratory visualization tool and does not represent direct temporal measurement. B Circle plot visualization of intercellular interaction number (left) and weights/strength (right) among major cell populations in the tumor microenvironment. Edge thickness and color intensity represent interaction strength; circle size indicates total interaction number. C Heatmap depicting the number (left) and weights/strength (right) between cell populations. Rows and columns represent source and target cells, respectively. D Cluster-cytokine association heatmap showing the activity patterns of cytokine signaling pathways across CD8⁺ T cell cluster 1 and cluster 2. Color scale represents normalized pathway enrichment scores or activity indices Subsequent analysis of communication signals revealed divergent signaling architectures in cluster 2. Outgoing patterns were characterized by enhanced surface adhesion molecules and inflammatory mediators, concomitant with impaired core TCR signaling (Fig. 2D). Conversely, incoming signals demonstrated potent antigen stimulation and adhesion inputs, alongside preserved TCR signaling capacity, yet exhibited attenuated co-stimulatory, metabolic, and migratory cues (Fig. 2D). This signal imbalance trapped T cells in the tumor with sustained antigen exposure but insufficient resources for cytotoxic function, establishing a terminally exhausted, functionally paralyzed state. Systematic identification of exhaustion-associated core genes through differential expression and weighted gene co-expression network analysis To systematically map the transcriptional determinants of terminal exhaustion, we performed DEGs analysis among different clusters and drew volcano plots and heat plots (Fig. 3A and B). Subsequently, we employed the high-dimensional weighted gene co-expression network analysis (hdWGCNA) algorithm to calculate the gene expression profiles of two CD8⁺ T cell subsets, and divided the core genes among these subsets into different gene modules to identify the core gene set. Ultimately, we verified the correlation between genes and modules in the network. By setting the value (β) to 6, we successfully constructed the scale-free network (Fig. 3C). The genes were segregated into respective modules via hierarchical clustering, and a gene similarity heatmap was generated based on the topological overlap matrix (TOM) (Fig. 3D). Notably, the brown module exhibited the strongest module-trait correlation (~ 47%) and was enriched for core exhaustion-related genes (Fig. 3E), whereas the remaining modules contained genes with weaker exhaustion associations. Identification of exhaustion-associated core genes via hdWGCNA. A Volcano plot showing differentially expressed genes between CD8+ T cell cluster 1 and cluster 2. Significantly upregulated genes in Cluster 2 (log₂FC > 0.5, adjusted p < 0.05) are highlighted in red. B Heatmap visualization of differentially expressed genes between CD8+ T cell cluster 1 and cluster 2, ordered by Monocle2-derived pseudo-time. Rows represent genes and columns represent individual cells, with expression values normalized and scaled. C Scale-free topology fitting index (left y-axis) and mean connectivity (right y-axis) as functions of soft-thresholding power (x-axis) for hdWGCNA network construction. A power of β = 6 was selected to achieve scale-free topology (R² > 0.8). D Hierarchical clustering dendrogram of genes based on topological overlap matrix dissimilarity measure. Colors below the dendrogram indicate gene modules identified by dynamic tree cutting. The brown module showed the highest correlation with the exhaustion phenotype. E Module–trait correlation heatmap depicting the relationship between hdWGCNA gene modules (rows) and CD8+ T cell clusters (columns). The brown module exhibited the strongest positive correlation with cluster 2 (exhaustion state, r = 0.47, p < 0.001) Machine learning-based construction and comprehensive validation of the TEX-signature prognostic model in CRC Building upon the WGCNA-derived exhaustion-associated modules, we implemented an integrative screening pipeline to identify robust exhaustion biomarkers. Candidate genes were first intersected with pseudo-time-ordered differential expression profiles, yielding 50 high-priority targets (Fig. 4A). These candidates were subsequently subjected to an ensemble machine learning framework comprising ten algorithms (DTS, GBM, glmBoost, KNN, Logistic, NeuralNet, PLS, RF, SVM, XGBoost), all demonstrating robust predictive performance (C-index > 0.8; Fig. 4B and C). Next, cross-model intersection and Cox regression analysis converged on five survival-associated genes: kruppel-like factor 3 (KLF3), lamin A/C (LMNA), glucose transporter member 3 (GLUT3/SLC2A3), ARF-like 4 C (ARL4C), and tissue inhibitor of matrix metalloproteinase 1 (TIMP1) (Fig. 4D and E). Subsequently, we derived TEX scores as a cox-weighted composite of these genes and validated its prognostic value in TCGA cohorts. Kaplan-Meier analysis revealed significantly inferior overall survival (OS) and disease-free survival (DFS) in high-risk patients (Fig. 4F and G), with time-dependent ROC curves demonstrating strong predictive accuracy (1-year AUC = 0.633, 3-year AUC = 0.609, 5-year AUC = 0.633; Fig. 4H). Machine learning-based construction of TEX-signature prognostic model. A Venn diagram showing the intersection of pseudo-time-derived differential genes and brown module core genes identified by hdWGCNA, yielding 50 candidate genes for downstream machine learning analysis. B Importance ranking of genes selected by ten machine learning algorithms (DTS, GBM, glmBoost, KNN, Logistic, NeuralNet, PLS, RF, SVM, and XGBoost) in bubble plot. Bubble size indicates relative feature importance; consistency across algorithms highlights robust candidate biomarkers. C ROC curves comparing the predictive performance of ten machine learning models. AUC values demonstrate model discrimination capability, with PLS and XGBoost showing optimal performance. D UpSet plot illustrating the intersection of biomarkers selected by each machine learning algorithm. E Forest plot of multivariate Cox regression analysis for the five core genes in the TCGA cohort. Hazard ratios (HR) and 95% confidence intervals indicate the independent prognostic value of each gene. F Kaplan-Meier survival curves comparing OS between high- and low-TEX scores patients stratified by TEX-signature in the TCGA training cohort. p-value calculated by log-rank test in TCGA cohorts. G From top to bottom, the point plot shows high- and low‐TEX scores patients groups divided by the cutoff values (top). The distribution plot of survival time and survival status of high- and low-risk patient of TCGA dataset (bottom). H Time-dependent ROC curves for 1-, 3-, and 5-year survival predictions in the TCGA cohort, demonstrating the prognostic accuracy of the TEX-signature model across different time horizons Clinically applicable nomogram for individualized survival prediction and subgroup validation To optimize clinical deployment, we integrated TEX scores with established prognostic factors (gender, age, metastatic stage) into a unified predictive nomogram in the TCGA cohorts (Fig. 5A). The model demonstrated exceptional calibration, with predicted survival probabilities closely tracking observed outcomes (Fig. 5B). Stratified validation across independent GEO cohorts confirmed consistent prognostic discrimination: TEX scores low-risk patients exhibited significantly prolonged survival compared to high-risk counterparts, irrespective of gender or age (Fig. 5C and D), validating TEX scores as a robust, context-independent biomarker for precision oncology in CRC. Clinical nomogram construction and subgroup validation. A Prognostic nomogram integrating clinical variables (gender, age, metastasis stage) with the molecular TEX-signature for predicting 1-year, 3-year, and 5-year OS probability in CRC patients. Each variable axis is scaled according to its prognostic contribution; the sum of individual points yields total points, which correspond to specific survival probabilities on the bottom scale. B Calibration curves assessing the agreement between nomogram-predicted survival probabilities and observed outcomes at 1, 3, and 5 years. The diagonal dashed line represents perfect calibration; the solid lines indicate the actual performance of the nomogram. Patient cohorts were derived from TCGA. C-D Subgroup survival analysis stratified by clinical characteristics. Kaplan-Meier curves comparing TEX-signature high- versus low-TEX scores patients in (C) age subgroups (≤ 65 years versus > 65 years) and (D) gender subgroups (female versus male) Immune landscape and modulators associated with the TEX-signature Given the pivotal role of CD8+ T cells in anti-tumor immunity, we investigated the relationship between TEX scores and immune microenvironment composition in CRC. Deconvolution of bulk transcriptomic data using five independent algorithms (CIBERSORT, EPIC, MCP-counter, Quanti-seq, xCell) revealed distinct microenvironmental phenotypes stratified by TEX scores (Fig. 6A). High-TEX clusters exhibited immune profiles dominated by regulatory populations and terminally differentiated immune cells, whereas low-TEX clusters maintained cytotoxic effector cells and antigen-presenting cells (Fig. 6A). Concordantly, high-TEX clusters exhibited inflammatory exhaustion characterized by concurrent immune checkpoint upregulation, costimulatory dysfunction, and physical T cell retention, creating a functionally paralyzed yet metabolically active state (Fig. 6B). Furthermore, correlation analysis demonstrated that the five core genes (KLF3, LMNA, SLC2A3, ARL4C, TIMP1) were significantly associated with these immunomodulatory factors and immune infiltrates, with SLC2A3 and ARL4C showing the strongest immune correlations (Fig. 6C). However, these five genes showed limited association with total CD8+ T cell abundance (Fig. 6D). These findings indicated that TEX scores reflected an immunosuppressive, exhaustion-enriched microenvironment in CRC; however, its utility for predicting ICI response requires validation in prospective immunotherapy cohorts. Immune landscape and modulator associations of the TEX-signature. A Correlation between TEX-signature risk score and immune cell infiltration levels inferred by five deconvolution algorithms, CIBERSORT, MCP-counter, xCell, EPIC, and quanTIseq. Heatmap displays Pearson correlation coefficients; asterisks indicate statistical significance. B Correlation between TEX-signature and immune modulators categorized into seven functional groups, chemokines, receptors, MHC molecules, immune-inhibitors, immune-stimulators, cell adhesion molecules, and others. Color intensity represents correlation strength; circle size indicates significance level. C Core genes specific immune modulator associations. Heatmap displays the correlation matrix between individual core genes (rows) and immune modulator categories (columns), revealing distinct immune-modulatory patterns for KLF3, LMNA, SLC2A3, ARL4C, and TIMP1. D Pearson correlation analysis between core genes and significantly differentially infiltrated immune cells. Bar length represents correlation coefficient magnitude; color indicates positive (red) or negative (blue) associations. All data were shown in mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001 TEX-signature association with cancer immunity cycles and immune-related biological processes Building upon established associations between TEX scores and tumor immunity [26], we investigated the immunobiological processes underlying this correlation. Systematic analysis of the cancer-immune cycle revealed that high-TEX scores clusters exhibited inhibition across multiple sequential steps, including release of cancer cell antigens, priming and activation, immune cell infiltration, and recruitment of diverse immune populations (T cells, CD4+ T cells, CD8+ T cells, Th1 cells, dendritic cells, Th22 cells, macrophages, neutrophils, NK T cells, eosinophils, basophils, and Th17 cells) (Fig. 7A). TEX-signature association with cancer immunity cycles. A Stepwise analysis of the cancer-immunity cycle comparing high-risk versus low-risk TEX-signature patients. The seven-step cycle includes: (1) release of cancer cell antigens, (2) cancer antigen presentation, (3) priming and activation of T cells, (4) trafficking of immune cells to tumors, (5) infiltration of immune cells into tumors, (6) recognition of cancer cells by T cells, and (7) killing of cancer cells. Bar plots display normalized activity scores with statistical comparisons between risk groups. B Comparison of immune-related indices between TEX-signature risk groups. Indices were calculated using the “easier” package, including T cell inflamed score, interferon-gamma response, and cytolytic activity. Box plots display distribution differences with statistical significance indicated. C Multi-dataset immune signature enrichment analysis. Box-and-whisker plots comparing scores for immune cell populations and functional pathways derived from five reference collections: LM22 (CIBERSORT), HALLMARK, IEGS (innate immune gene sets), Angelova (immune checkpoint), and REACTOME. Each panel represents signatures from one database; y-axis indicates normalized enrichment score. D Transcriptional regulatory network analysis on bulk RNA-seq data from TCGA. Heatmap visualization of correlations between TEX-signature risk score and activity scores of immune-related transcription factors (TFs). Color scale indicates Pearson correlation coefficients with hierarchical clustering of transcription factors. E-G Pathway-specific immune activity correlations. Scatter plots with fitted regression lines showing the relationship between TEX-signature and ssGSEA enrichment scores for: (E) GEP signature representing antigen presentation machinery, (F) IFN response pathway, and (G) CYT score reflecting granzyme and perforin expression. All data were shown in mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001 Concordantly, high-TEX clusters demonstrated reduced T cell activity and inflammatory responses, meanwhile, analysis of immune resistance programs (RESF-Down, RESF-Up, resF) indicated that high-TEX tumors exhibited elevated immune resistance, whereas low-TEX tumors maintained lower resistance levels (Fig. 7B). Subsequently, Integrated analysis of five immune signature databases (LM22, HALLMARK, IIGS, ANGELOVA, REACTOME) confirmed decreased immune cell infiltration in high-TEX clusters (Fig. 7C). Strikingly, the high-TEX scores clusters showed a synergistic activation of chronic NF-κB/STAT1 inflammatory signaling and EMT transcription program, driving T cells into the terminal differentiation/matrix retention state; while the low-TEX clusters maintained MYC/E2F-driven proliferative metabolic program, supporting the reversible activation and functional execution of effector T cells. This differential distribution of transcription factors provides a theoretical basis for reversing exhaustion by targeting specific TFs. (Fig. 7D). Finally, Single-sample gene set enrichment analysis (ssGSEA) further demonstrated that high-TEX clusters exhibited decreased inflammatory pathway activity (GEP, IFN, CYT) (Fig. 7E and G), consistent with chronic inflammation-associated exhaustion. Collectively, these findings establish that high-TEX CRC is characterized by an immunosuppressive, exhaustion-enriched microenvironment with active oncogenic signaling, whereas low-TEX tumors maintain immunological competence. Enrichment analysis of functional pathways for TEX scoring To decode the systems-level pathways underpinning TEX scores stratification, we integrated multi-omic functional annotations on differentially expressed genes identified within CD8⁺ T cell subpopulations. Gene Ontology of biological processes (GO-BP) analysis revealed that high-TEX tumors were characterized by extracellular matrix (ECM) remodeling intertwined with aberrant immune responses (Fig. 8A), wherein effector T cells are physically constrained within dense ECM networks, precluding effective antitumor surveillance. Cellular component (GO-CC) analysis indicated endoplasmic reticulum stress (ERS) activation, suggesting that chronic antigen stimulation drives proteostatic overload and triggers the unfolded protein response (UPR), predisposing T cells toward apoptotic or exhausted fates (Fig. 8B). Molecular function (GO-MF) analysis further revealed that high-TEX T cells retain antigen recognition capacity yet exhibit T cell-T cell interaction-induced exhaustion and collagen/scaffold binding-mediated physical retention within fibrotic niches (Fig. 8B). Function enrichment of the TEX-signature and core genes. A GO-BP enrichment analysis of high-risk TEX-signature patients. Bar plot shows significantly enriched biological processes ranked by normalized enrichment score (NES); bar color indicates statistical significance. B Tripartite GO landscape of high-risk patients. Cellular Component (GO-CC) showing subcellular localization of enriched genes; Biological Process (GO-BP) highlighting dysregulated pathways; Molecular Function (GO-MF) describing altered protein activities. Dot size corresponds to gene ratio; color intensity indicates -log10 (FDR). C GSVA comparing high-risk versus low-risk TEX-signature groups. Bar plot displays differential enrichment scores for hallmark gene sets; bar height indicates normalized enrichment score, with positive values (red) indicating up-regulated and negative values (blue) indicating down-regulated in high-risk group. D KEGG pathway enrichment analysis of five core genes (KLF3, LMNA, SLC2A3, ARL4C, TIMP1). Heatmap displays enrichment significance (-log10 (FDR)) across pathways; color intensity indicates statistical significance. E MSigDB HALLMARK gene set analysis of core gene functions. Heatmap of NES across 50 hallmark gene sets. Rows: individual gene sets; columns: sample groups or core genes. Red: pathway activation; blue: pathway suppression; hierarchical clustering applied to both dimensions Subsequently, GSVA analysis demonstrated that high-TEX tumors recruit abundant immune cells to fibrotic matrices; however, these cells functionally shift toward antibody production and inflammatory signaling at the expense of antitumor cytotoxicity, with concomitant damage to basal metabolic machinery (Fig. 8C). This metabolic-ionic reprogramming failure, characterized by overstimulation without nutrient/signal integration, prevents T cells from completing activation-dependent ion flux and metabolic transitions. Finally, Core genes pathway mapping in KEGG and HALLMARK identified three oncogenic axes: WNT/TGF-β/HEDGEHOG-driven stromal remodeling, MTOR/p53/NOTCH-mediated metabolic-EMT reprogramming, and IL-6-STAT3/IFN-γ-dependent immunosuppression (Fig. 8D and E). These convergent programs establish the “proliferation-invasion-immune escape” triad characteristic of high-TEX malignancy. Clinical utility of TEX-signature for immunotherapy stratification and mechanistic validation To establish TEX scores as a mechanistic predictor of immunotherapy response, we benchmarked it against the Tumor Immune Dysfunction and Exclusion (TIDE) immune escape framework. Strong positive correlation (R = 0.554, p < 0.001; Fig. 9A and B) positioned TEX scores as a robust resistance metric. Cross-cohort analysis of GSE78002, IMvigor210, and GSE126044 confirmed that elevated TEX scores predicted non-response to immune checkpoint inhibitors (Fig. 9C and D). Clinical utility of TEX-signature. A Boxplot showing the distribution of TIDE scores between low- and high-TEX scores groups. The high-TEX group exhibited significantly elevated TIDE scores compared to the low-TEX group (p < 0.001). B Scatter plot demonstrating the correlation between TEX risk score and TIDE score (Spearman R = 0.554, p < 0.001). Red and blue dots represent high- and low-TEX scores samples, respectively. C-E Immunotherapy response prediction by TEX-signature across three independent cohorts. (C) GSE78002 Hela cell cohort, (D) IMvigor urothelial cancer cohort, and (E) GSE126044 non-small lung cancer cohort. Left panels: distribution of TEX scores in responders versus non-responders; right panels: ROC curves evaluating predictive accuracy with AUC values We subsequently validated the bidirectional core gene signature in tumor biopsies from immunotherapy response (n = 3) and non-response (n = 3) CRC patients. With the help of qPCR and western blot, we revealed a divergent expression pattern that the inhibitory factor KLF3 was downregulated in non-response tumors, while the functional quartet LMNA, SLC2A3 (GLUT3), ARL4C, and TIMP1 were coordinately upregulated (Fig. 10A and C). These findings indicate that bulk tissue expression levels of core genes align with TEX scores predictions. Clinical utility and mechanistic validation of TEX-signature. A RT-qPCR analysis of KLF3, LMNA, SLC2A3, ARL4C, and TIMP1 mRNA levels normalized to GAPDH. B Western blot analysis of core protein expression in response and non-response tissue. C Quantification of the protein levels indicated in B. D Quantification of the positive area percentage for KLF3, LMNA, SLC2A3, ARL4C, and TIMP1 within the CD8⁺ T cell population. E Spatial validation by multiplex immunofluorescence. Confocal microscopy images (40× objective) demonstrated co-expression patterns of CD8⁺ T cell marker (red) with individual core genes (green) in formalin-fixed paraffin-embedded tumor sections from immunotherapy-response patients, with parallel analysis of non-response patients for comparative validation. Nuclear counterstain with DAPI (blue). Insets show magnified regions of interest. Scale bar represents 25 μm To further delineate the spatial distribution of core genes within T cells, we performed multiplex immunofluorescence staining using CD8 as a T cell marker in conjunction with gene-specific antibodies. Immunofluorescence staining confirmed enhanced co-localization of CD8⁺ T cells with all five core genes in immunotherapy-non-responsive lesions. Notably, KLF3 exhibited exhaustion-associated depletion within the tumor microenvironment, whereas LMNA, SLC2A3, ARL4C, and TIMP1 demonstrated relative enrichment (Fig. 10D and E). Further quantification analysis confirmed our conclusions from the prior analyses (Supplementary Fig. 1A-1E). These data mechanistically implicate the five-gene signature as a druggable resistance node and TEX scores as a precision oncology tool for ICI stratification. Discussion CRC remains a formidable health challenge globally, ranking as the third most common malignancy and the second leading cause of cancer-related mortality [5]. Although significant progress has been achieved in CRC treatment in recent years, chemotherapy resistance and associated toxic side effects remain important factors affecting the prognosis of patients with CRC [38]. Despite the remarkable success of immunotherapy across various cancer types, the limited responsiveness of CRC to ICIs underscores the urgent need for deeper mechanistic insights to broaden the beneficiary population [9]. CD8+ T cells, originating from CD34+ hematopoietic stem cells in the bone marrow, constitute the principal effectors of antitumor immunity, and their functional exhaustion frequently precipitates tumor immune evasion and disease progression [39]. Tumor-induced TEX is further compounded by the immunosuppressive TME [40]. Recently, the proteome of exhausted CD8+ T cell rose a mechanistic vulnerability and a new therapeutic target to improve cancer immunotherapies [41], which reminded us further refine and scrutinize exhaustion models, seeking additional insights to advance immunotherapy for CRC. Through systematic interrogation of the molecular and functional attributes of distinct CD8+ T cell subgroups in CRC, we figure out the intercellular communication and functional pathway enrichment of different clusters. It helps us better understand the microenvironment of CRC and visualize the function of different T cell cluster. Interestingly, we observed that cluster 2, the exhaustion cluster, exhibited pronounced activation across inflammatory and immune signaling pathways, alongside enhanced ligand-receptor interactions. Collectively, cluster 2 represents a distinct subset of terminally exhausted CD8⁺ T cells characterized by hyperactive intercellular communication. These cells exhibit an “inflammatory exhaustion” phenotype wherein they attempt to compensate for progressive effector dysfunction through enhanced ligand-receptor engagement and augmented outgoing inflammatory signaling [42]. However, this compensatory hyperactivation ultimately precipitates a self-reinforcing vicious cycle of overactivation followed by functional collapse. Mechanistically, cluster 2 cells are physically constrained by adhesive interactions and subjected to proximal signal blockade, yet paradoxically sustain high outgoing communication in attempt to overcome these constraints. This persistent inflammatory signaling, rather than restoring functionality, drives further exhaustion and terminal differentiation. Consequently, cluster 2 represents a therapeutically challenging state that is refractory to conventional immune checkpoint blockade targeting “silent” exhaustion. We posit that effective intervention will require combinatorial strategies encompassing anti-adhesive disruption to facilitate physical egress from immunosuppressive niches and metabolic reprogramming to reset aberrant signaling circuitry. Summing up, we propose a process of functional exhaustion in the differentiation of CD8+ T cells between the two clusters. Notably, we recognize that our focused analysis on CD8⁺ T cell intrinsic states does not fully capture the extrinsic suppressive mechanisms mediated by Tregs, MDSCs, or metabolic constraints imposed by the stromal compartment. Future multi-dimensional analyses integrating these suppressive axes will provide a more comprehensive view of immunotherapy resistance. Subsequently, leveraging machine learning algorithms and systematic identification of exhaustion-associated core genes, we identified five prognostic genes, KLF3, LMNA, SLC2A3, ARL4C, and TIMP1, and constructed the TEX-signature, which was subsequently validated using nomogram analysis. Although these five genes have been individually implicated in CRC prognosis [43–46], their collective relationship with the tumor microenvironment had not been systematically examined. By defining TEX scores as the weighted sum of their expression levels, we uncover a significant association with immune infiltration, thereby establishing a previously unrecognized link between these prognostic genes and antitumor immunity. Multiple machine learning models have confirmed that the low-TEX-signature symbolize the better survival, more immune infiltration, and stronger immunity cycle. Finally, utilizing publicly available GEO datasets and clinical cohort of immunotherapy in non-small cell lung cancer, and urothelial carcinoma, as well as cervical cancer, we validated the TEX-signature model. Furthermore, the expression of core genes in CRC validates our conjectures. Collectively, our research established a promising foundation for future research and potential advancements in the treatment of CRC. Several limitations warrant acknowledgment. First, our reliance on single-cell sequencing data from relatively limited sample sizes may not fully capture the inter-patient heterogeneity that critically influences immunotherapy efficacy in CRC. Validation in larger, prospectively collected cohorts would enhance the robustness and generalizability of our findings. Second, despite the statistical advantages afforded by publicly available datasets, the integration of multi-source scRNA-seq and bulk RNA-seq data inevitably introduced technical heterogeneity stemming from disparities in sample processing protocols, platform-specific batch effects, and variable detection sensitivities. Additionally, incomplete clinical annotations—including ethnic background, MSI/MSS status, specific chemotherapeutic regimens, and objective immunotherapy response criteria (e.g., RECIST)—in certain public cohorts constrained our capacity to delineate precise clinical correlates of TEX states. To address these limitations, future studies will incorporate prospectively collected, clinically annotated biospecimens with standardized procurement protocols and harmonized analytical workflows, ensuring robust cross-validation and clinical translatability of the TEX scores model. Third, further molecular and functional experiments are necessary to elucidate the mechanistic roles of core genes and decipher the molecular basis of TEX-signature-mediated immune regulation. It is imperative to acknowledge that the majority of our findings were predicated upon correlative bioinformatic analyses, which, while robust for delineating transcriptional regulatory networks underlying TEX, inherently preclude definitive conclusions regarding protein-level expression dynamics and spatial heterogeneity within the colorectal tumor microenvironment. Specifically, our reliance on scRNA-seq and bulk RNA-seq datasets, integrated through computational deconvolution algorithms, may not fully resolve the intricate cellular crosstalk or spatially restricted niches that govern T cell dysfunction. This limitation is particularly pertinent given our observation that cluster 2 CD8⁺ T cells exhibited enhanced intercellular communication and ligand-receptor interactions, phenomena that necessitate spatial contextualization. To circumvent these constraints, future investigations should incorporate orthogonal validation strategies, including multiparameter flow cytometry for quantitative assessment of exhaustion marker co-expression (e.g. PD-1, TIM-3, LAG-3), multiplex immunofluorescence or imaging mass cytometry for spatially resolved profiling of immune cell infiltration patterns, and RNA-scope or in situ hybridization for direct visualization of core gene transcripts (KLF3, LMNA, SLC2A3, ARL4C, TIMP1) within tissue architecture. These approaches will be instrumental in bridging the gap between transcriptional signatures and functional phenotypes, thereby strengthening the mechanistic underpinnings of the TEX scores model. At the same time, while our analyses demonstrate robust associations between the TEX-signature and TEX phenotypes, we recognize that these computational and correlative findings do not establish functional causality. The current study establishes TEX-signature as a clinically predictive biomarker with strong mechanistic grounding, but we acknowledge that definitive causal inference requires experimental manipulation of individual genes. Future studies utilizing CRISPR-mediated gene editing, overexpression systems, or conditional knockout mouse models are warranted to dissect the specific contributions of KLF3, LMNA, SLC2A3, ARL4C, and TIMP1 to exhaustion programming, thereby complementing our biomarker development with causal mechanistic insights. Finally, our clinical validation cohort comprised only six CRC patients, necessitating further prospective validation to establish the clinical utility of TEX-signature in guiding treatment decisions. Meanwhile, we recognize that bulk tissue validation cannot definitively attribute gene expression to specific cellular subsets without additional deconvolution or spatial transcriptomics. However, the strong correlation between TEX scores and TIDE immunotherapy resistance scores (Fig. 9A), combined with immunofluorescence validation (Fig. 10), supports the utility of this tissue-level proxy for clinical stratification. In conclusion, our research establishes a novel assessment framework, distinct from the conventional MSS/MSI-H stratification, for personalized evaluation of immunotherapy sensitivity in CRC. The TEX-signature enables more individualized therapeutic regimens and illuminates new research directions for overcoming TEX-mediated immunotherapy resistance. Conclusion By integrating single-cell and bulk transcriptomic data through multiple machine learning algorithms, we developed a prognostic signature that simultaneously predicts overall survival and TEX states in CRC. The TEX scores model stratifies patients based on exhaustion-associated molecular profiles, offering a robust biomarker for immunotherapy response prediction and informing precision therapeutic strategies. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Acknowledgements We extend our sincere gratitude to all researchers who have generously shared their datasets in public repositories, enabling this work. We also thank our colleagues and collaborators whose contributions were instrumental to the research cited herein. Finally, we are deeply indebted to the patients and their families who agreed to participate in this study. Abbreviations Colorectal cancer T cell exhaustion single cell RNA sequencing Kruppel-like factor 3 Lamin A/C Glucose transporter member 3 ARF-like 4 C Tissue inhibitor of matrix metalloproteinase 1 Overall survival Disease-free survival Hazard ratios Tumor microenvironment American Joint Committee on Cancer Immune checkpoint inhibitors Microsatellite instability-high The Cancer Genome Atlas Gene Expression Omnibus t-distributed Stochastic Neighbor Embedding high-dimensional weighted gene co-expression network analysis single-sample gene set enrichment analysis Gene ontology Gene ontology Cellular Component Gene ontology Biological Process Gene ontology Molecular Function Normalized enrichment score False discovery rate negative logarithm of false discovery rate Kyoto Encyclopedia of Genes and Genomes Gene set variation analysis Partial Least Squares Random Forest Decision Tree Support Vector Machine Logistic Regression K-Nearest Neighbor Algorithm Gradient Boosting Machine Neural Network Generalized Linear Model Boosting Area under the curve Receiver operating characteristic curve Tumor immune cycle Gene expression profile Cytotoxic activity Interferon-gamma International Code of Ethics for Biomedical Research involving humans Real-time quantitative polymerase chain reaction Uniform Manifold Approximation and Projection topological overlap matrix Authors’ contributions Zhijing Zhang, Peng Ouyang, Kai Cui, and Jianqin Lai contributed equally to this work. Zhijing Zhang conceived the study, designed and conducted the experiments, and wrote the initial draft of the manuscript. Peng Ouyang and Kai Cui performed data analysis, interpreted the results, and contributed to manuscript writing. Jianqin Lai contributed greatly during the revision of the manuscript and assisted with data interpretation Xin Deng assisted with data interpretation and manuscript revision and obtained funding. Yixiang Wen and Wanyu Chen assisted with experimental work and data collection. Zhenhong Xian assisted with specimen and data collection. Qi Qi, Zhen Bao, Jin Gong and Xiao He share corresponding authorship. They supervised the project, provided critical feedback, obtained funding and were responsible for the final approval of the manuscript. All authors agree to be accountable for all aspects of work ensuring integrity and accuracy. Funding This work was supported by the Guangdong Association for Science and Technology Youth Science and Technology Talent Development Program (NO. SKXRC2025117), National Natural Science Foundation of China (NO. 82504048), the Bethune Charitable Foundation (NO. 803292), Guangzhou Basic Research Program City - University (Institute) - Enterprise Joint Funding Project (NO. SL2024A03J01364), Medical Scientific Research Foundation of Guangdong Province of China (NO. A2023398), Fundamental Research Funds for the Central Universities (NO. 11624305). Data availability All data will be available on request.Publicly available datasets analyzed during this study can be found in the following repositories. [https://portal.gdc.cancer.gov/](https:/portal.gdc.cancer.gov) (TCGA Colorectal Adenocarcinoma dataset); GEO: GSE39582 (bulk RNA-seq), GSE132465 (scRNA-seq), GSE78002 and GSE126044 (immunotherapy cohorts), and IMvigor210 (urothelial cancer immunotherapy cohort). The clinical experimental data and analysis scripts used during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate Informed consent was obtained from all participants, and the study was approved by the Ethics Committee of First Affiliated Hospital of Gannan Medical University (approval number: LL8C-2025333). This study adhered to the principles of the Declaration of Helsinki. Consent for publication Not applicable. This study utilized anonymized data from public repositories (TCGA, GEO) and de-identified clinical samples. No individual patient data or identifiable information are presented. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Zhijing Zhang, Peng Ouyang, Kai Cui and Jianqin Lai contributed equally to this work. Contributor Information Qi Qi, Email: qiqikc@jnu.edu.cn. Zhen Bao, Email: baozhen@jnu.edu.cn. Jin Gong, Email: zaitushuguan@163.com. Xiao He, Email: hexiao00914@163.com. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Data Availability Statement All data will be available on request.Publicly available datasets analyzed during this study can be found in the following repositories. [https://portal.gdc.cancer.gov/](https:/portal.gdc.cancer.gov) (TCGA Colorectal Adenocarcinoma dataset); GEO: GSE39582 (bulk RNA-seq), GSE132465 (scRNA-seq), GSE78002 and GSE126044 (immunotherapy cohorts), and IMvigor210 (urothelial cancer immunotherapy cohort). The clinical experimental data and analysis scripts used during the current study are available from the corresponding author on reasonable request.