Comprehensive Analysis of scRNA-Seq and Bulk RNA-Seq Identified TIMP1 as a Prognostic Marker in Colorectal Cancer

Abstract

Colorectal cancer (CRC), one of the most prevalent and lethal malignancies of digestive system, continues to impose a substantial burden on global health due to its high morbidity and mortality. Tumor microenvironment (TME) is a critical regulator for CRC progression and therapeutic response, but the in-depth understanding on the relationship of TME with CRC remains to be elucidated. In this study, we leveraged single-cell RNA sequencing (scRNA-seq) and bulk RNA-seq data to dissect the immune heterogeneity in CRC patients. The differential expression genes analysis, functional enrichment analysis, random forest analysis and the Least Absolute Shrinkage and Selection Operator method were used to construct a molecular immune prognostic model. The molecular model demonstrated robust performance in stratifying patients based on their immune microenvironment characteristics. The experimental results showed that TIMP1 was highly expressed in CRC. Knockdown of TIMP1 gene significantly inhibited RKO cell proliferation and invasion. By integrating scRNA-seq and bulk RNA-seq data, we developed a new prognostic model that effectively predicts clinical outcomes in patients with CRC and identifies TIMP1 as a promising prognostic biomarker for CRC.

INTRODUCTION

Colorectal cancer (CRC) remains the third most prevalent malignancy worldwide and ranks as the second leading cause of cancer-related mortality, as reported by the updated global cancer statistic data (1). Despite the overall 5-year survival rate exceeding 60%, outcomes vary markedly across disease stages; for patients with distant metastases the rate drops below 15% (2). Although surgery and systemic chemotherapy remain the mainstays of care, the incremental benefit of targeted agents against VEGF receptor or epidermal growth factor receptor (EGFR) signaling is confined to molecularly selected sub-groups (3,4). Likewise, immune-checkpoint blockade produces durable responses only in a minority of mismatch-repair–deficient or highly mutated tumors (5,6,7,8). Notably, substantial interpatient variability in recurrence rates and survival outcomes persists among CRC patients with the same clinicopathological features, reflecting the underlying molecular heterogeneity of the disease (6,7). Therefore, there is an urgent need to elucidate the molecular mechanisms underlying CRC tumorigenesis and identify novel candidate molecular biomarkers to improve diagnostic accuracy, guide therapeutic decisions, and enhance prognostic prediction in clinical practice.

In recent years, advances in cancer genomics have established the bulk RNA-seq as a predominant approach for transcriptomic analysis. This technological progress has enabled the identification of numerous genetic alterations serving as effective therapeutic targets across various tumor types (9). While bulk RNA-seq provides a comprehensive overview of transcriptional profiles, this approach has inherent limitations in precisely identifying cellular subpopulations and resolving intratumoral heterogeneity at single cell resolution (10). Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for characterizing tumor heterogeneity by enabling comprehensive analysis of cell-specific transcriptomes. Besides, this technology can investigate differential gene expression patterns and their spatial distribution across diverse cellular subpopulations within tumors (11). In addition, scRNA-seq facilitates the development of personalized treatment strategies by providing critical insights into tumor evolution and mechanisms of drug resistance during cancer progression (12). Given these advantages, numerous studies have integrated bulk RNA-seq with scRNA-seq technologies to identify potential tumor biomarkers, enabling more precise patient stratification. For instance, researchers identified a robust intrinsic signature of neuroendocrine cells through comprehensive single-cell and whole-transcriptome analysis, which effectively predicts prostate cancer progression (13). Several research groups have employed integrated scRNA-seq and bulk RNA-seq approaches to elucidate fundamental molecular classifications of necroptotic apoptosis, characterize tumor microenvironment (TME) infiltration patterns, and enhance prognostic prediction in CRC patients (14). Furthermore, complementary single-cell and whole-transcriptome analyses have enabled the identification of 2 distinct epithelial tumor cell states, leading to refinement of the consensus molecular classification system for CRC (15).

TIMP1, initially identified as the first characterized collagenase inhibitor (16), exhibits broad inhibitory activity against multiple matrix metalloproteinases (MMPs) involved in tumorigenesis and metastatic processes. Consequently, TIMP1 has been predominantly studied in the context of its MMP-inhibitory function, leading to the prevailing hypothesis that its expression correlates with attenuated cancer progression (17,18,19). However, emerging clinical evidence has demonstrated that TIMP-1 exhibits consistently elevated expression levels in both circulating blood and tumor tissues, with its overexpression significantly correlating with poor clinical outcomes across multiple cancer types (20,21,22,23). TIMP1 exhibits dual biological functions, serving both as an inhibitor of MMPs and as an active signaling molecule (24). These MMP-independent functions are mediated through direct binding to specific cell surface receptors, triggering diverse cellular responses. TIMPs exhibit pleiotropic cytokine-like activities, modulating critical cellular processes including proliferation, angiogenesis, differentiation, and apoptosis (25,26,27,28). However, there are limited research on how TIMP1 regulates the microenvironment to mediate the biological behavior of CRC.

This study employed scRNA-seq to characterize major cellular subpopulations in CRC and developed a TME-based prognostic model based on CRC-related genes. Through integrated bioinformatics analysis of multiple databases and subsequent in vitro experimental validation, we demonstrated the oncogenic role of TIMP1 in CRC progression.

MATERIALS AND METHODS

Data sources and processing

scRNA-seq datasets (GSE178341 and GSE188711) were obtained from the National Center for Biotechnology Information Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). Bulk RNA-seq gene expression profiles and corresponding clinical data for CRC were downloaded from The Cancer Genome Atlas (TCGA) portal (https://portal.gdc.cancer.gov, accessed October 24, 2024).

Analysis of RNA-seq data

Single-cell data underwent quality control with 3 sequential filters: 1) cells with 300–5,500 detected genes were retained; 2) genes expressed in <2 cells were removed; and 3) cells exceeding 15% mitochondrial RNA content were excluded. Then, count values from scRNA-seq were normalized to 1e6 per cell, followed by log-transformation [log₁₀(1+CPM)]. Later, genes with high variability were detected. When performing cell clustering, the sc.pp.highly_variable_genes function (default parameters) was utilized to filter these high-variability genes. In addition, 40 principal components were used for principal component analysis. Correct for bulk effects using the Harmony integration algorithm. Next, clusters with Leiden algorithm were detected, with cluster resolution set to 0.1, and Uniform Manifold Approximation and Projection was subsequently used to visualize them. Cluster annotation was completed with marker genes reported previously and differential expression genes (DEGs) for every cluster obtained by Wilcoxon rank sum test (sc.tl.rank_genes_groups function).

InferCNV analysis

We calculated the large-scale chromosome copy number variation (CNV) scores for epithelial cells by Python package inferCNVpy (v0.3.0; BioTuring Inc., San Diego, CA, USA). Raw count matrices, annotation files and gene/chromosome position information files were prepared in accordance with data requirements by inferCNVpy (https://github.com/icbi-lab/infercnvpy). B cells and T cells were selected as reference normal cells and analyzed using default parameters. Epithelial cells with CNV scores higher than the median were defined as malignant epithelial cells, while the rest were considered as non-malignant epithelial cells.

Construction and validation of prognosis model

The top500 gene in malignant epithelial cells was used as the characteristic gene set. Single-factor Cox risk regression was performed using Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression and Cox regression for obtaining prognosis-related genes and corresponding risk scores of diverse samples. The risk score calculation formula is:

Risk Score=(Gene 1 Coefficient×Gene 1 Level)+(Gene 2 Coefficient×Gene 2 Level)+...+(Gene n Coefficient×Gene n Level)

According to median risk score, we classified samples as high- or low-risk group. The Kaplan-Meier (KM) survival analysis software was finally utilized to evaluate prognostic characteristics of diverse risk groups.

Analysis of immune cell infiltration

For assessing the correlation of risk profile with immune microenvironment, this study used CIBERSORT algorithm to conduct a comprehensive assessment of the immune microenvironment of patients with TCGA data sample (29). We used the CIBERSORT algorithm (http://cibersort.stanford.edu/) for examining 22 different immune cell type distribution and determining their proportions. Pearson correlation coefficients were calculated for analyzing the relation of risk score with the presence of tumor-infiltrating immune cells.

Gene set enrichment analysis (GSEA)

For understanding heterogeneities in biological functions between subtypes of malignant and non-malignant epithelial cells, the present work performed GSEA using Python package GSEApy (http://github.com/zqfang/GSEApy) (30). Then, 1000 GSEA analyses were conducted for genes ranked according to log-fold change of expression, targeting the canonical gene sets of KEGG_2021_Human.

Gene set variant analysis (GSVA)

GSVA was conducted to evaluate gene set enrichment within transcriptomic data (31). We obtained gene sets based on Molecular Signature Database (http://www.gsea-msigdb.org/gsea/index.jsp), and performed GSVA using limma software to accurately assess the presence of biological functions in different samples of possible gene sets.

Drug sensitivity analysis

According to gene expression patterns, we estimated drug sensitivity with “OncoPredict” R package (R Foundation for Statistical Computing, Vienna, Austria). The method helped calculate drug sensitivity levels of diverse samples, and the smaller levels indicated the greater drug efficacy. Drug sensitivity was analyzed comprehensively in every sample for determining the levels of common antitumor agents.

Cell communication analysis

CellPhoneDB is the publicly available database regarding receptors, ligands and corresponding interactions (32). In this study, CellPhoneDB (2.1.1) was used for cellular crosstalk interactions. Mean values represent the average expression of ligand and receptor in a particular cell type and are calculated from percentage of cells expressing a particular gene and the average gene expression. The p-values are calculated from the proportion of mean values equal to or higher than the actual mean and represent the likelihood that a given receptor-ligand complex is present in a particular cell type.

Random survival forest (RSF) analysis

The RSF algorithm was applied to assess the significance of associated genes via the R software “randomForestSRC” (33) through 1,000 iterations using MonteCarlo simulation (n rep=1,000). We included genes whose relative relevance was >0.3 into our final model.

Cell culture and cell lines

HCT116, HT29, LOVO and RKO human CRC cells were cultivated within DMEM that contained 10% FBS and 1% penicillin-streptomycin (all reagents were purchased from Suzhou New NCM Biotech, Suzhou, China), and later cultivated under 37°C with 5% CO₂. For generating stable TIMP1 cells, RKO cells were transfected with wild-type TIMP1-FLAG-GFP lentivirus or negative control lentivirus. At 3 days after transfection, 2 ug/ml puromycin was added to select cells for a 7–10 day period. At last, quantitative PCR (qPCR) and Western blotting were performed for verification.

Western blotting

Protein immunoblotting was performed with reference to the previously published protocol (34). The cells were grown to about 80% fusion within the 10 cm petri dish. After collection, cells were cleaved with a pre-chilled lysis buffer (150 mM NaCl, 50 mM HEPES pH 7.4, RIPA, 5% glycerol, protease/phosphatase inhibitor tablets). We then added the cell lysates into a 4×SDS sample buffer (40% glycerol, 180 mM HEPES pH 7.4, 4% β-mercaptoethanol, 4% SDS, 0.04% bromophenol blue), followed by 5 min of denaturation within the 95°C metal bath. Next, the sample is electrophoretic with the protein Marker. The proteins were then transferred to a 0.2 micron Trans-Blot^® Turbo Mini polyvinylidene fluoride membrane (Bio-Rad, Hercules, CA, USA). After 1 h of sealing using 5% skimmed milk powder within Tris buffered hydrochloride that contained 0.1% Tween-20, the membrane underwent overnight incubation under 4°C and additional 1 h of HRP-conjugated secondary antibody incubation at room temperature. Immobilon Luminata Crescendo Western HRP substrate (EMD Millipore, Billerica, MA, USA) was adopted for displaying protein bands. ImageJ software (National Institutes of Health, Bethesda, MD, USA) was used for measuring and quantifying the strength of protein bands.

RNA extraction and quantitative real-time PCR

By adopting TRIzol (Invitrogen, Carlsbad, CA, USA), total RNA extraction was completed in line with specific protocols. CDNA was then synthesized and qPCR experiments were carried out with LightCycler^®480 (Roche Diagnostics, Basel, Switzerland). Quantitative RT-PCR (qRT-PCR) assay was later conducted according to previous description (17) using appropriate primers. With beta-actin being an endogenous control, target gene levels in experimental and control groups were determined by the 2^−ΔΔCt approach. The primers for PCR were: F: ACTACCTGCAGTTTTGTGGC; R: CTGGAAGCCCTTTTCAGAGC.

5-Ethynyl-2'-deoxyuridine (EdU) assay

We assessed cell proliferation through performing the EdU assay, BeyoClick™ EdU Cell Proliferation Kit (with Alexa Fluor 647) was employed to evaluate cells of exponential growth phase in 2 groups following specific protocols. The EdU-positive cell number was counted with a microscope in 5 randomly chosen fields of view from every well.

Transwell assay

The cell suspension was added to the upper chamber of the Transwell chamber, and a medium containing 15% serum was added to the lower chamber. The Transwell plate was subject to incubation in a 37°C, 5% CO₂ incubator for 24–48 h to allow cell migration. After the chamber was cleaned, it was fixed with 4% paraformaldehyde for 10–15 min and washed with PBS prior to crystal violet staining. After that, this chamber was counted and imaged.

Statistical analysis

R software (v.4.1.1; R Foundation for Statistical Computing) and GraphPad Prism 8.0 (GraphPad Software Inc., San Diego, CA, USA) were utilized to analyze data. Data were represented by mean ± SD. Relations among variables were evaluated through calculating Spearman and Pearson correlation coefficients. Between-group differences were analyzed by 2-tailed unpaired t-test or Wilcoxon rank-sum test. Significance levels were set at p<0.05, p<0.01, and p<0.001.

RESULTS

Identification of CRC cell subtypes

First, double cells were removed by scrublet with strict parameters (total counts >30,000; pct-count-mt <15%) for quality control, and 247,559 high-quality single-cell data were used for subsequent analysis (Fig. 1A). We identified 2,006 highly variable genes (Fig. 1B) and used them for analysis, followed by batch correction with Harmony. Unsupervised clustering with the Leiden algorithm (resolution=0.1) resolved 11 transcriptionally distinct clusters (Fig. 1C). Cluster identities were assigned by canonical markers in the CellMarker database, resulting in the annotation of B cells (CD37, CD79A), plasma cells (MZB1, TNFRSF17), cycling cells (MKI67, TOP2A), T cells (CD3D, CD3E), mast cells (TPSAB1, TPSB2), dendritic cells (CD83, LAMP3), goblet cells (TFF3, LGALS4), macrophages (LYZ, SPI1), and 3 stromal lineages epithelial (KRT8, KRT18), endothelial (VWF, RAMP2) and fibroblast (SPARC, COL6A2) cells (Fig. 1D and E). Comparative analysis of paired tumor and normal mucosa revealed pronounced enrichment of epithelial cells, macrophages and several T-cell subsets in the tumor micro-environment, whereas goblet and plasma cells were depleted (Fig. 1F). Patient-level inspection confirmed that epithelial and macrophage populations dominated most tumors (Fig. 1G). Sub-clustering of the macrophage lineage showed that normal tissue was populated largely by M0-like macrophages (70.27% in normal vs. 50.30% in tumor), while pro-inflammatory CXCL8⁺/S100A8⁺ macrophages consistent with an M1-skewed state were rare in normal mucosa but expanded markedly in tumors (0.13% vs. 7.04%) (Supplementary Fig. 1). Given that CRC arises from malignant epithelial transformation and that this compartment exhibited the greatest tumor-specific expansion and transcriptional heterogeneity, subsequent analyses centered on the epithelial lineage.

Figure 1. Immune cell landscape within CRC and normal tissue microenvironment at single-cell transcriptional level. (A) Distribution of single-cell sequencing data. (B) Top genes with the greatest changes in screening. (C) UMAP plot showing 11 clusters from 247559 cells. (D) Final cell annotation with CellMarker. (E) Bubble plot showing annotated Marker genes. (F) Distribution of cell type in tumor and normal tissue. (G) Distribution of annotated cells in the sample.

UMAP, Uniform Manifold Approximation and Projection.

These results establish a comprehensive single cell atlas of CRC and provide a robust framework for downstream dissection of malignant epithelial programs and their crosstalk with immune and stromal elements of the TME.

Identification of malignant epithelial cells

Among the 50,515 epithelial cells isolated in the previous step, we inferred large scale chromosomal copy number alterations with inferCNVpy. Cells whose mean CNV score exceeded 0.04 were classified as malignant, yielding 32,024 malignant and 18,491 non malignant epithelial cells (Fig. 2A and B). To characterize the malignant compartment, we extracted the 500 most over expressed genes (Supplementary Table 1) and performed Gene Ontology biological process enrichment. The dominant terms included protein transport, endoplasmic reticulum to Golgi vesicle mediated transport and the unfolded protein response, suggesting heightened secretory activity and ER stress in tumor epithelium (Fig. 2C). GSEA analysis further demonstrated significant up regulation of complement activation, oestrogen response, mTOR and TNF-α/NF-κB signaling, coagulation cascades and androgen-response pathways in malignant relative to non-malignant cells (Fig. 2D). Together, these data indicate that CRC derived epithelial cells acquire broad transcriptional reprogramming linked to immune modulation, hormone signaling and proteostasis, consistent with multifaceted hallmarks of colorectal carcinogenesis.

Figure 2. Biological characteristics of malignant epithelial cells. (A) UMAP shows CNV score. (B) UMAP shows malignant and non-malignant epithelial cells. (C) Bar plot shows GO enrichment of genes characteristic of malignant epithelial cells; (D) GSEA enrichment analysis shows pathways for malignant epithelial cell enrichment.

UMAP, Uniform Manifold Approximation and Projection; GO, Gene Ontology.

Construction and functional enrichment analysis of nomogram prediction model

Building on the malignant epithelial signature defined above, we selected the top 500 up-regulated genes in malignant vs. non-malignant epithelial cells as the candidate pool for prognostic modelling. Univariate Cox regression in the TCGA-COAD cohort identified 30 genes significantly associated with overall survival (OS) (Fig. 3A). To prevent overfitting, these genes were subjected to 10-fold cross-validated LASSO analysis; the minimum partial-likelihood deviance indicated an optimal penalty (λ) that narrowed the list to 10 genes—RNASET2, CDH1, CTNNA1, RAB7A, TIMP1, CTSD, TMED4, IFNGR2, PTTG1IP and KRT10 (Fig. 3B). The model Cox coefficient was acquired through multivariate Cox proportional risk regressions. Then, based on coefficients and classification values of the expression level, a risk score is constructed as follows:

Figure 3. Construction and validation of prognostic models in TCGA cohort. (A) Univariate Cox regression of OS-related genes. (B) BLASSO regression of genes associated with OS. (C-F) KM and receiver operating characteristic curve analyses. (G, H) Establishment and validation of prediction models. (I, J) GSEA for GO and KEGG pathways enriched in the high- and low-risk groups. (K) GSVA of all genes in 2 risk groups was performed to determine identify enriched pathway.

GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Risk Score=(0.0733×RNASET2)−(0.0044×CDH1)−(0.0460×CTNNA1)−(0.0405×RAB7A)+(0.2765×TIMP1)+(0.0099×CTSD)+(0.0055×TMED4)+(0.0139×IFNGR2)+(0.0047×PTTG1IP)+(0.0339×KRT10)

Risk scores of tumor cases from TCGA cohort were computed, and then the patients were classified into high-risk and low-risk groups on the basis of the median risk score (Supplementary Tables 2 and 3). The results of the KM survival curve showed that individuals with high-risk scores had shorter survival than those with low-risk scores (Fig. 3C), and the area under the curve (AUC) values of the 1-year, 3-year and 5-year OS were 0.85, 0.77 and 0.71, respectively, suggesting the robust accuracy of this risk model (Fig. 3D). According to KM curve for validation set, individuals with high-risk scores had markedly reduced OS (Fig. 3E) and that the AUCs for 1-, 3- and 5-year OS were above 0.7 (Fig. 3F). For identifying independent predictors, this study analyzed risk scores along with clinical factors like age, gender, clinical stage through univariate as well as multivariate Cox analysis. Results were shown in a nomogram with a number of independent factors. As observed, risk scores well predicted OS in CRC patients, which can be used to independently predict patient prognosis (Fig. 3G and H). The related pathways that had the highest degree of enrichment between the high- and low-risk groups were identified via GSEA; in the high-risk group, these pathways included “detection of chemical stimulus involved in sensory perception” and “nucleosome assembly.” Furthermore, the high-risk group exhibited an enrichment of pathways associated with “rheumatoid arthritis,” “viral protein interaction with cytokine and cytokine receptor” were up-regulated in the high-risk group, whereas the “alcoholism,” “neutrophil extracellular trap formation” and “systemic lupus erythematosus” pathways were down regulated (Fig. 3I and J). GSVA results showed that the “PRION_DISEASES,” “CARDIAC_MUSCLE_CONTRACTION” signaling pathways were uniquely enriched in the high-risk group, and the “NON_HOMOLOGOUS_END_JOINING,” “ENDOMETRIAL_CANCER” pathways were enriched in the low-risk group (Fig. 3K).

Immunotherapy-relevant landscape of the 2 risk group

Using bulk TCGA-COAD expression profiles, we deconvoluted immune cell composition with CIBERSORT. The 22 inferred cell types displayed marked inter tumor variability (Fig. 4A) and non random inter correlations (Fig. 4B). Stratification by the 10 gene risk score revealed that low-risk tumors harbored higher proportions of resting memory CD4 T cells and resting NK cells, whereas high-risk groups were enriched for follicular helper T cells, memory B cells, cytotoxic CD8 T cells and neutrophils (Fig. 4C). Consistently, the risk score correlated positively with regulatory T cells, follicular-helper T cells and CD8 T cells, and negatively with resting memory CD4 T cells (Fig. 4D), suggesting that high-risk groups possess a more activated yet potentially exhausted immune milieu.

Figure 4. Immunotherapy-related landscape of the 2 risk group. (A) Histogram shows immune cell proportions in each sample. (B) Correlation between diverse immune cell infiltration levels. (C) Diverse infiltrating immune cell proportions between high- and low-risk patients. (D) Univariate Cox regression for risk scores of diverse immune cells. (E) Sensitivity analysis for high- and low-risk groups to common chemotherapy agents. (F, G) The tumor mutational burden levels in 2 risk groups were estimated through a risk model. (H, I) Cell–cell interactions between different cell types in CRC.

TFH, follicular helper T; ns, no significance.

^*p<0.05, ^**p<0.01, ^***p<0.001, ^****p<0.0001.

Drug response prediction with OncoPredict indicated that higher risk scores were associated with increased sensitivity to several targeted agents, including the IGF-1R/IR inhibitor BMS-754807, the Mcl-1 inhibitor MIM1, the third-generation EGFR inhibitor osimertinib, the p53 reactivator PRIMA-1 MET, the CDK4/6 inhibitor ribociclib-1,632, the BCL-XL inhibitor WEHI-539 and the β-catenin antagonist XAV-939 (Fig. 4E). Mutation profiling showed comparable frequencies of canonical CRC drivers (APC, TP53, TTN, KRAS) across risk groups, although DNAH5 alterations were more common in the high-risk subset (Fig. 4F and G).

Finally, CellPhoneDB-based ligand receptor analysis highlighted intensified intercellular crosstalk in high-risk tumors: macrophages engaged multiple lineages, epithelial cells signaled to fibroblasts via CCL20–CCR6 and JAG1–NOTCH3, and bidirectional CCL15–CCR1 interactions linked epithelial cells with macrophages (Fig. 4H and I, Supplementary Fig. 2).

Malignant epithelial cells with elevated TIMP1 interact with subsets of T and B lymphocytes through the chemokine axis CCL20–CCR6 (Supplementary Fig. 3). These findings delineate distinct immunological and genomic contexts for the 2 risk strata and point to differential vulnerabilities that could inform personalized immuno-targeted therapeutic strategies in CRC.

RSF analysis further identified key prognostic genes

To verify the relative contribution of each member of the 10 gene signature, we trained a RSF model in the TCGA-COAD cohort. Five genes—TIMP1, RNASET2, CDH1, CTNNA1 and RAB7A—exceeded the pre-specified variable-importance threshold of 3% and were retained as the core prognostic features (Fig. 5A). Differential-expression analysis confirmed that TIMP1 and RNASET2 were up-regulated in tumor tissues, whereas CDH1, CTNNA1 and RAB7A were down-regulated relative to matched normal tissues (Fig. 5B and C). Among these, TIMP1 displayed the highest importance score and yielded the greatest separation of overall-survival curves; patients with elevated TIMP1 expression had significantly poorer outcomes (Fig. 5D and E, Supplementary Fig. 4).

Figure 5. RSF analysis further identified key prognostic genes. (A) RSF and prognostic related gene expression analysis. (B) Key gene expression in different clusters. (C) Key gene expression within CRC and normal samples. (D) Receiver operating characteristic curve verifies the disease prediction ability of hub genes. (E) KM curves show the OS for CRC cases classified based on key gene expression.

ns, no significance; TPR, true positive rate; FPR, false positive rate; CI, confidence interval.

^***p<0.001; ^****p<0.0001.

Prediction of the regulatory network of hub genes

We used miRWalk to identify mRNA-miRNA pairs of 5 key genes, obtaining a total of 655 miRNA. Second, merely mRNA-miRNA pairs involving CRC-related miRNA were kept, resulting in 13 miRNA (Fig. 6A). Using mirNet to search for lncRNA, we identified 6 miRNA with 206 edges (Fig. 6B). Gene Cards database (https://www.genecards.org/) was employed to score 9,119 genes related to CRC. Those top 19 gene levels within both CRC and non-carcinoma samples were explored (Fig. 6C). The 19 gene levels were related to 5 key genes (Fig. 6D).

Figure 6. Prediction of the regulatory network of hub genes. (A) Relations between miRNAs and CRC development as well as 5 hub genes according to Human microRNA Disease Database and miRWalk databases. (B) The ceRNA network involving 5 hub genes was visualized. (C, D) Relations of the 5-gene signature with 19 CRC-related genes.

Functional validation of TIMP1 in CRC cells

Given its dominant prognostic weight, TIMP1 was selected for mechanistic interrogation. Baseline screening by qRT-PCR and immunoblotting showed that TIMP1 mRNA and protein were most abundant in RKO cells (Fig. 7A and B). Stable knock-down with sh-TIMP1 markedly reduced TIMP1 protein levels (Fig. 7C) and led to a significant decline in EdU incorporation, indicating impaired proliferation (Fig. 7D). Transwell assays demonstrated a parallel reduction in migratory capacity after TIMP1 silencing (Fig. 7E). These results corroborate the bioinformatic prediction that TIMP1 acts as a driver of malignant behavior in CRC and support its candidacy as both a prognostic biomarker and a potential therapeutic target.

Figure 7. Functional validation of TIMP1 in CRC cells. (A) TIMP1 protein expression levels in different CRC cells lines. (B) qRT-PCR analysis on transcriptional level of TIMP1 in different colorectal cell lines. (C) Western blotting analysis of TIMP1 expression in RKO cell transfected with shTIMP1. (D) The effect of TIMP1 knockdown in RKO cell proliferation was demonstrated via EDU assays. (E) The invasion of TIMP1 knockdown cell was assessed via Transwell assays.

ns, no significance.

^*p<0.05, ^****p<0.0001.

DISCUSSION

CRC is one of the most common malignancies worldwide, with its incidence rising in many countries. Despite significant advancements in CRC management, its heterogeneity and aggressiveness still limit its prognostic assessment (35,36,37). Therefore, identifying novel biomarkers remains crucial for developing individualized treatments and improving patient outcomes (38). While traditional bulk RNA-seq measures average gene expression across cell populations, scRNA-seq offers higher resolution, enabling transcriptional profiling of distinct cell subsets. This approach helps identify cell-specific biomarkers and uncover tumor heterogeneity in CRC and other cancers (36,39). Consequently, this study comprehensively integrated bulk RNA-seq with scRNA-seq for developing the risk model, which showed superb power in predicting immunotherapy response to CRC. Previous joint analyses have validated 9 cancer-associated fibroblast-related gene signatures as independent predictive markers for CRC patients who may not benefit from immunotherapy (40). However, tumor heterogeneity complicates cellular communication dynamics, while heterogeneities in TME interactions are critical for tumor occurrence and treatment resistance (37,38,41). For instance, Guo et al. (42) combined scRNA-seq and bulk RNA-seq to investigate TME and T-cell dynamics in triple-negative breast cancer, identifying key prognostic genes. Through combining scRNA-seq with RNA-seq data and employing several machine e-learning tools, a new predictive model was developed for predicting survival probabilities in bladder cancer cases. Therefore, building a novel prognostic model by utilizing advantages of scRNA-seq and bulk seq (38), taking into account tumor heterogeneity, interactions of various cell populations, immune infiltration, tumor mutational burden and clinical characteristics to the maximum extent, accurately identifying survival outcomes and immunotherapy responses of tumor patients will offer direct evidence for stratification and precision treatment of CRC patients. In this study, we performed a comprehensive analysis of bulk RNA-seq plus scRNA-seq for developing the risk model, which demonstrated superb power in forecasting immunotherapy response to CRC.

Our single-cell atlas resolved eleven major cellular lineages within the TME, of which epithelial, immune and stromal populations displayed pronounced transcriptional heterogeneity. Functional interrogation of the top 500 genes specifically up-regulated in malignant epithelial cells revealed significant enrichment of oestrogen-response and PI3K/AKT/mTOR signatures, pathways increasingly recognized as convergent drivers of colorectal carcinogenesis. Experimental and clinical studies indicate that oestrogen receptor-β modulates multiple tumor-associated programs, including hypoxia signaling and HIF-mediated chemo radio-resistance, thereby reshaping the TME and influencing therapeutic efficacy (43,44). In parallel, constitutive activation of the mTOR axis promotes cell growth, metabolic rewiring and metastatic dissemination; its dysregulation is also implicated in acquired resistance to standard cytotoxic regimens. The therapeutic relevance of these findings is underscored by ongoing trials evaluating single-agent or combination PI3K/AKT/mTOR inhibitors in advanced CRC (40). Collectively, our data place malignant epithelial cells at the nexus of steroid-hormone and nutrient-sensing networks, offering mechanistic insight into tumour aggressiveness and highlighting actionable vulnerabilities that warrant prospective validation.

The 10 risk signature stratified patients into low- and high-risk groups with marked immunobiological divergence. High-risk group were enriched for interferon, complement and cytokine receptor pathways and harbored greater fractions of follicular helper and regulatory T cells as well as cytotoxic CD8 T cells, pointing to a chronically stimulated but potentially exhausted micro-environment. In parallel, these tumors exhibited higher tumor-mutation-burden quartiles, differential expression of multiple checkpoint molecules and a shift in macrophage stromal cell crosstalk, all of which may attenuate responsiveness to conventional checkpoint blockade. Concordance between the risk score and adverse clinicopathological features advanced T and N stage, positive lymph nodes and vascular invasion—underscores the biological plausibility of the signature and reinforces its independent prognostic value. Pharmacogenomic interrogation of the GDSC and CTRP resources suggested that high-risk profiles are preferentially sensitive to several targeted agents, including the IGF-1R/IR inhibitor BMS-754,807, the Mcl-1 inhibitor MIM1, the third-generation EGFR inhibitor osimertinib, the p53 reactivator PRIMA-1 MET, the CDK4/6 inhibitor ribociclib-1,632, the BCL-XL antagonist WEHI-539 and the β-catenin inhibitor XAV-939. While many of these compounds remain in early-phase evaluation or have yet to be tested formally in CRC, their predicted efficacy highlights actionable dependencies that could be explored in biomarker-driven trials.

Five prognostic related genes were screened using RSF analysis: CTNNA1, RAB7A, TIMP1, RNASET2, and CDH1. Furthermore, the regulatory network involved has an important effect on CRC progression. Consequently, the ceRNA regulatory network that utilizes 6 key DEGs was established to elucidate the underlying mechanisms of CRC to gain insight in those unexplored regulatory networks related to CRC. Among the prognostic genes screened above, CTNNA1 encodes α-catenin and participates in a variety of physiological effects, including the production and signaling of adhesion junctions (45). CTNNA1 shows abnormal expression within leukemia and diverse solid tumors, including the digestive system, which is related to tumor genesis, development and prognostic outcome. CTNNA1 is previously suggested with a dual role in inhibition and carcinogenicity in intestinal tumors. In most cancers, its exact mechanisms are still unclear. Therefore, it is necessary to conduct other preclinical research on CTNNA1 to diagnose, predict prognosis and treat patients early. As the small GTPase belonging to RAS oncogene family (46), RAB7A is localized at lysosome and exerts an essential effect on endosomal lysosome system maturation. RAB7A is not only related to immunomodulation and hepatocellular carcinoma progression (47), but also has a key effect on different cancers like CRC and is related to regulating autophagy (48,49). Belonging to T2 ribonuclease family depicted in our species, the human RNASET2 gene was cloned from 6q27 in 1997 (50). It is noteworthy that the tumor inhibitory activity of RNASET2 is determined in diverse cancers like CRC (51,52). RNASET2 has been shown to promote tumor in the pathogenesis of renal clear cell carcinoma (53). Epithelial cadherin (CDH1 gene), as the transmembrane glycoprotein, binds to epithelial cells, usually at adhesion junctions (54). In normal cells, E-cadherin exerts its tumor-suppressing effects primarily by isolating β-catenin binding to lymphoid enhancer factor/T cell factor, which function as transcribed genes serving the proliferative Wnt signaling pathway. Although debate continues about whether the loss of E-cadherin is epithelial-mesenchymal transition (EMT), loss of E-cadherin function is usually related to dismal prognostic outcome of tumor patients (55,56). E-cadherin expression not only has a multifaceted impact on cell function in the context of carcinogenesis, but also has clinical significance in diagnosis, prognosis and treatment (57). The 7 compounds identified by OncoPredict are aligned with signaling axes that are clustered on or regulated by our signature core genes. TIMP1 has been shown to activate PI3K/AKT and to trans-activate IGF-1R via CD63; this provides a mechanistic link to the predicted sensitivity to the IGF-1R/IR inhibitor BMS-754807. TIMP1-induced AKT signaling can stabilize β-catenin and up-regulate anti-apoptotic proteins such as BCL-XL and Mcl-1, rationalizing the association with XAV-939, WEHI-539 and MIM1. RAB7A regulates endosomal trafficking and degradation of EGFR, offering a plausible connection to Osimertinib, while CDH1/CTNNA1 loss drives cyclin-D–CDK4/6 dependence, consistent with ribociclib-1632 sensitivity. Finally, RNASET2 has been reported to interact with the p53 pathway, supporting the link to PRIMA-1 MET. To determine whether these in-silico predictions translate into functional vulnerability, we are currently testing the drug sensitivity of TIMP1-knock-down and control RKO cells; results from these assays will provide priority information for compounds for future in vivo validation.

TIMP1 was originally described as an endogenous broad-spectrum inhibitor of MMPs and, through its impact on extracellular-matrix remodeling, was long regarded as a negative regulator of tumor invasion and metastasis (16,58,59,60). Paradoxically, subsequent clinical studies across multiple malignancies including our own data have shown that TIMP1 is frequently over-expressed and that its high abundance correlates with shortened survival (17,18,19,61,62,63,64,65,66). Consistent with these observations, we confirmed robust TIMP1 mRNA and protein expression in CRC cell lines and demonstrated that shRNA-mediated silencing significantly curtailed proliferation and Transwell migration in RKO cells, supporting a context-dependent pro-tumorigenic role.

Beyond its canonical MMP-inhibitory activity, TIMP1 may shape tumor behavior by modulating immune interactions. CellPhoneDB analysis revealed that malignant epithelial cells with elevated TIMP1 expression engage T and B lymphocyte subsets via the chemokine axis CCL20–CCR6. CCL20, secreted by malignant epithelium, can recruit CCR6-positive Th17 cells and B cell populations, thereby fostering an inflammatory milieu that promotes tumor progression and immune evasion. This chemokine circuit provides a mechanistic link between TIMP1 over-expression and the immunologically ‘hot-but-dysfunctional’ landscape observed in high-risk tumors, characterized by increased follicular-helper and regulatory T cells yet poor clinical outcome. Targeting TIMP1 itself or disrupting downstream CCL20–CCR6 signaling could therefore attenuate both the invasive phenotype and the pro-tumor immune crosstalk, representing a dual-pronged therapeutic strategy in CRC. Future work will focus on dissecting the molecular basis of TIMP1 driven chemokine secretion and evaluating the combinatorial efficacy of TIMP1 inhibition with immune checkpoint blockade in relevant preclinical models.

This study has several constraints that temper immediate clinical translation. (i) The cohorts, though sizeable for public data, remain modest and largely Western, limiting ethnic generalizability. (ii) The 10-gene signature was only internally validated within TCGA; external verification in independent multicenter series is under way. (iii) Retrospective design means prospective, risk-stratified trials are still required to show clinical benefit. (iv) Functional work on TIMP1 was confined to a single cell line; ongoing rescue and in vivo studies will test mechanistic relevance. Addressing these gaps will be essential to confirm robustness and real-world utility.

Abbreviations

AUC

area under the curve

CNV

copy number variation

CRC

colorectal cancer

DEG

differential expression gene

EdU

5-ethynyl-2'-deoxyuridine

EGFR

epidermal growth factor receptor

EMT

epithelial-mesenchymal transition

GSEA

gene set enrichment analysis

GSVA

gene set variant analysis

KM

Kaplan-Meier

LASSO

Least Absolute Shrinkage and Selection Operator

MMP

matrix metalloproteinase

OS

overall survival

qPCR

quantitative polymerase chain reaction

qRT-PCR

quantitative reverse transcriptase polymerase chain reaction

RSF

random survival forest

scRNA-seq

single-cell RNA sequencing

TCGA

The Cancer Genome Atlas

TME

tumor microenvironment

SUPPLEMENTARY MATERIALS

Supplementary Table 1

The top 500 characteristic genes of malignant epithelial cells

Supplementary Table 2

The median risk score of test cohort

Supplementary Table 3

The median risk score of train cohort

Supplementary Figure 1

The macrophage subsets in CRC.

Supplementary Figure 2

The cellular interactions between high-risk and low-risk groups.

Supplementary Figure 3

The cell interactions between high-expressing TIMP1 cells and immune cells.

Supplementary Figure 4

The analysis of the correlation between TIMP1 expression and clinical characteristics.