Kinase-based molecular subtyping and a prognostic risk model in colorectal cancer with functional validation of PLK4

Abstract

Background

Colorectal cancer (CRC) exhibits marked molecular heterogeneity, yet prognostic tools centered specifically on protein kinases remain scarce. This study aimed to identify kinase-related molecular subtypes of CRC and to develop and validate a kinase-based prognostic risk model integrating clinical and immune features.

Methods

Transcriptomic data from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO) were integrated to identify prognosis-related kinase genes. Unsupervised consensus clustering was used to define molecular subtypes, and differences in survival, immune infiltration, and immune checkpoint expression between subtypes were assessed. A nine-gene prognostic model was constructed using least absolute shrinkage and selection operator (LASSO) and multivariate Cox regression and validated in an independent external cohort. Associations with the consensus molecular subtypes (CMS), tumor mutational burden (TMB), drug sensitivity, and immunotherapy response were analyzed. The expression of core genes was verified by quantitative real-time polymerase chain reaction (qRT-PCR) in CRC cell lines, and the function of PLK4 was examined by knockdown experiments.

Results

Consensus clustering revealed two distinct kinase-based subtypes with significantly different survival; the poorer-prognosis subtype overlapped predominantly with CMS4 and showed altered immune infiltration. The nine-gene risk model provided reproducible prognostic stratification, remained an independent predictor of survival after adjustment for tumor-node-metastasis (TNM) stage, and was validated in the external validation cohort. Compared with high-risk patients, low-risk patients showed higher TMB and greater predicted responsiveness to immunotherapy and selected chemotherapeutic agents. qRT-PCR confirmed that nine core genes exhibited expression trends consistent with the bioinformatics predictions, and PLK4 knockdown consistently inhibited cell proliferation and migration and promoted apoptosis in both HT-29 and HCT116 cells, supporting an oncogenic role.

Conclusions

This kinase-based subtyping and prognostic model offers a complementary tool for risk stratification and individualized treatment decisions in CRC. Further multi-omics and in vivo studies are warranted to facilitate clinical translation.

Highlight box.

Key findings

• We identified two kinase-related molecular subtypes of colorectal cancer (CRC) and developed a nine-gene prognostic risk model that was validated in an independent external cohort and remained an independent predictor of survival after adjustment for tumor-node-metastasis (TNM) stage; functional assays confirmed that PLK4 promotes CRC cell proliferation and migration and suppresses apoptosis.

What is known and what is new?

• Protein kinases play central roles in CRC proliferation, invasion, drug resistance, and immune evasion, yet most existing prognostic models focus on processes such as the cell cycle, metabolism, or cell death, and kinase-centered models integrating clinical and immune features remain scarce.

• This study provides a kinase-based subtyping scheme and prognostic model that links risk stratification to the consensus molecular subtypes, the immune microenvironment, tumor mutational burden, drug sensitivity, and immunotherapy response, together with in vitro validation identifying PLK4 as a candidate oncogenic driver.

What is the implication, and what should change now?

• The model offers a complementary tool—rather than a replacement for conventional TNM staging—for individualized prognosis assessment and treatment selection in CRC, and highlights PLK4 as a potential therapeutic target. Further multi-omics and in vivo studies are warranted before clinical translation.

Introduction

Colorectal cancer (CRC) is among the most common malignant tumors of the digestive system worldwide, with persistently high incidence and mortality rates. Despite advances in early screening and therapeutic strategies, significant prognostic heterogeneity remains in advanced patients, with markedly different survival outcomes observed even among individuals at similar clinical stages (1). Such heterogeneity greatly hampers the implementation of precision medicine, highlighting the urgent need to develop molecular biomarkers and predictive tools that more accurately reflect tumor biology. The primary cause of this variability lies in the molecular heterogeneity of CRC, including diverse gene expression profiles, mutational landscapes, and patterns of signaling pathway activation (2). In recent years, multiple studies have attempted to identify prognostic biomarkers through molecular subtyping and bioinformatics-based models, aiming to improve individualized prognosis assessment and clinical decision-making for CRC patients (3).

Among various molecular mechanisms, protein kinases have attracted considerable attention due to their pivotal roles in regulating cell proliferation, differentiation, apoptosis, and migration. As central components of intracellular signaling, aberrant activation of kinases is tightly linked to tumorigenesis and progression in multiple cancers. Accumulating evidence indicates that certain kinases are abnormally overexpressed or functionally dysregulated in CRC, closely associated with metastasis, drug resistance, and poor prognosis. For example, PRKCI promotes epithelial-mesenchymal transition (EMT) and enhances invasiveness of CRC cells by stabilizing TGFBR1 and activating the TGF-β pathway during metastasis (4); moreover, its overexpression is strongly linked to oxaliplatin resistance and reduced survival (5). Similar findings suggest that kinase hyperactivation exerts critical effects on multiple pathways, such as Wnt/β-catenin, PI3K/AKT, and MAPK, which not only drive tumorigenesis but also profoundly influence therapeutic responses and immune evasion (6).

Recent proteogenomic studies have further highlighted striking differences in kinase activities and signaling networks between primary and metastatic CRC, suggesting that kinase-based molecular subtyping may have important clinical implications (7). In addition, CRISPR-Cas9 screening experiments have demonstrated that specific CRC subtypes [e.g., consensus molecular subtypes (CMS)4] depend on key kinases such as PAK2, providing opportunities for subtype-specific targeted therapy (8). However, systematic investigations of kinase-related subtypes in CRC remain limited, particularly those integrating kinase expression, clinical prognosis, and immune microenvironmental features into comprehensive prognostic models. Most existing models have focused on aspects such as cell cycle, metabolism, or cell death pathways, while prognostic risk models centered specifically on kinases are still relatively rare (9).

Notably, integrating multi-omics data with machine learning methods has yielded significant advancements in prognostic research for other cancers. For instance, in renal cancer, researchers developed a Treg-related programmed cell death signature model by integrating Mendelian randomization and machine learning methods, which significantly improved the accuracy of prognostic prediction (10). In bladder cancer, a prognostic risk stratification model with clinical guidance significance was established by using various machine learning algorithms to analyze mitotic catastrophe heterogeneity (11). In lung adenocarcinoma, a comprehensive analysis of the genetic interaction between atherosclerosis and lung cancer led to the construction of an integrated model with both diagnostic and prognostic value based on interacting genes (12). These studies offer methodological references for constructing pan-cancer prognostic models and suggest the necessity and feasibility of conducting similar research in CRC.

Therefore, identifying kinase-related molecular subtypes and constructing prognostic risk models that integrate clinical data with immunological analysis can not only provide a more comprehensive understanding of CRC molecular heterogeneity but also support the development of individualized therapeutic strategies. This study aimed to systematically analyze kinase-related gene (KRG) expression profiles, identify subtypes closely associated with prognosis, and establish a robust risk assessment tool to assist in prognosis prediction and therapeutic decision-making in CRC. The overall design and workflow of this study are summarized in Figure 1. We present this article in accordance with the TRIPOD and MDAR reporting checklists (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1446/rc).

Figure 1.

Workflow of the study. KRGs were analyzed in the TCGA-COAD/READ training cohort to identify kinase-related molecular subtypes and to construct a nine-gene prognostic risk model using univariate Cox, LASSO, and multivariate Cox regression; the model was validated in the GSE87211 cohort. Downstream bioinformatics analyses and in vitro experiments (including PLK4 knockdown) were then performed. CCK-8, Cell Counting Kit-8; COAD, colon adenocarcinoma; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; KRGs, kinase-related genes; LASSO, least absolute shrinkage and selection operator; miRNA, microRNA; qRT-PCR, quantitative real-time polymerase chain reaction; READ, rectal adenocarcinoma; RNA-Seq, RNA sequencing; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas; TF, transcription factor; TMB, tumor mutational burden.

Methods

Data collection

RNA sequencing (RNA-seq) data of CRC were obtained from UCSC Xena (https://xena.ucsc.edu/), including The Cancer Genome Atlas colon adenocarcinoma (TCGA-COAD) (normal: 41; tumor: 473) and rectal adenocarcinoma (TCGA-READ) (normal: 10; tumor: 167). These datasets were merged to generate the CRC training cohort (normal: 51; tumor: 640). The latest copy number variation data and corresponding clinical information were downloaded from TCGA (https://portal.gdc.cancer.gov/), with patients lacking survival information excluded. Microarray data (GSE87211, normal: 160; tumor: 203) were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) as the validation cohort, containing 203 tumor and 160 control samples. Data preprocessing employed the exact same limma method as the TCGA training set, specifically utilizing the normalizeBetweenArrays function for normalization. This approach aims to eliminate technical variations between platforms, ensuring consistency in cross-platform validation. Probe IDs were mapped to gene symbols via the GPL13497 platform annotation file. A total of 538 KRGs were collected from published literature (13). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Analysis of KRGs

Principal component analysis (PCA) was applied to evaluate batch effects between TCGA-COAD and TCGA-READ datasets. Differential expression analysis of tumor versus normal tissues in CRC was conducted using the limma package [|log₂fold change (FC)| >1, adj.P.value <0.05]. Intersecting the differentially expressed genes (DEGs) with KRGs yielded differentially expressed kinase-related genes (DEKRGs, n=145, Table S1). Functional enrichment of DEKRGs was performed via clusterProfiler for Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis. After excluding patients with survival time <30 days, univariate Cox regression was performed using the survival package to identify prognosis-related DEKRGs (P<0.05, n=15, Table S2). Expression levels of prognosis-related DEKRGs were compared between tumor and control samples, and Pearson correlation coefficients were calculated. Protein-protein interactions (PPIs) were predicted using the STRING database (https://string-db.org/) with a confidence threshold of 0.15, and the resulting PPI network was visualized with Cytoscape v3.10.2. GO enrichment analysis was further performed on prognosis-related DEKRGs.

Identification of kinase-related molecular subtypes and immune analysis in CRC

The expression matrix obtained from limma-normalized analysis based on prognosis-related DEKRGs was subjected to unsupervised clustering of tumor samples using the ConsensusClusterPlus algorithm. Specific parameter settings were: maximum number of clusters maxK=9, resampling iterations reps =100, sample proportion pItem =0.8, and feature proportion pFeature =1. The clustering algorithm employed k-means clustering with Euclidean distance as the metric, and a random seed was set to ensure reproducibility. The optimal number of clusters (K=2) was determined by the proportion of ambiguous clustering (PAC), dividing patients into two molecular subtypes. Survival analysis of subtypes was conducted using the survival package. Differential expression of prognosis-related DEKRGs between subtypes was evaluated. The single-sample gene set enrichment analysis (ssGSEA) algorithm was applied to assess differences in immune functions and immune cell infiltration, while the CIBERSORT algorithm was used to estimate immune cell infiltration. Expression of immune checkpoint-related genes between subtypes was also compared. Finally, differential expression analysis between subtypes was performed (|log₂FC| > 0.585, adj.P.value <0.05), and KEGG pathway enrichment was conducted for the identified DEGs. In addition, we used the CMScaller package to perform CMS typing on the samples.

Construction of prognostic model

Candidate genes (n=15, Table S2) identified from univariate Cox regression (P<0.05) were used to build the prognostic model. To avoid overfitting, least absolute shrinkage and selection operator (LASSO) regression was performed with the glmnet package, and the optimal penalty parameter (lambda.min) was selected by 10-fold cross-validation to exclude highly correlated genes, yielding 9 genes (Table S3). Multivariate Cox regression was then conducted on these 9 genes to construct the prognostic risk score model (Table S4). A prognostic risk score was constructed for each patient by linearly combining the expression levels of the 9 genes, weighted by their multivariate Cox regression coefficients, using the following formula: risk score = 0.045 × CAMK2B + 0.210 × DAPK1 + 0.008 × PLK4 − 0.260 × ULK4 + 0.277 × FJX1 − 0.028 × MAP2K6 + 0.149 × EPHA6 − 0.157 × BMX + 0.076 × NTRK2.

Where the coefficient represented the regression coefficient and gene expression denoted the normalized expression level of each gene (i = 1 to 9). Patient risk scores were calculated by summing expression levels of model genes weighted by their Cox regression coefficients, and patients were stratified into high- and low-risk groups based on the median risk score. Kaplan-Meier survival curves were plotted, while time-dependent receiver operating characteristic (ROC) curves [1-, 3-, and 5-year area under the curve (AUC)] were generated using the timeROC package. Risk score distribution and survival status plots were generated, and the model was validated in the independent GEO cohort. The specific method for model validation was as follows: first, model feature genes were matched in the validation set. Subsequently, based on the expression of model feature genes and utilizing the TCGA Cox model object (fixed coefficients, Table S4), the risk scores for the validation set were calculated. Kaplan-Meier survival curves were also generated for each feature gene, and their expression levels were compared between tumor vs. control and high- vs. low-risk groups.

Enrichment analysis

Samples were stratified into high- and low-risk groups by the median risk score, and gene set enrichment analysis (GSEA) v4.3.2 was performed. DEGs between risk groups were also subjected to GO and KEGG enrichment analysis.

Construction of nomogram

Univariate and multivariate Cox regression analyses were conducted incorporating clinical features and risk scores. A nomogram was constructed using the rms package to predict 1-, 3-, and 5-year overall survival. Calibration curves were plotted to evaluate concordance between predicted and observed outcomes. Decision curve analysis (DCA) was employed to assess clinical utility.

Subgroup analysis based on clinicopathological characteristics

Baseline clinical characteristics were summarized in relation to the prognostic risk score in the TCGA training cohort. Distribution of risk scores across clinical subgroups was compared. Patients were stratified by age (≤65 vs. >65 years), sex (female vs. male), stage (I + II vs. III + IV), T stage (T1+2 vs. T3+4), N stage (N0 vs. N1–2), and M stage (M0 vs. M1), and Kaplan-Meier survival curves were plotted for each subgroup.

Immune infiltration analysis and prediction of immunotherapy response

Immune infiltration scores for 29 immune cell types/functions were computed using ssGSEA in high- vs. low-risk groups. CIBERSORT was applied to evaluate immune infiltration differences, and correlations between model feature genes and immune cell infiltration were assessed by Pearson correlation. Expression of immune checkpoint-related genes was compared between risk groups. To assess immunotherapy response prediction, immunophenoscore (IPS) data were downloaded from The Cancer Immunome Atlas (TCIA, https://tcia.at) for TCGA-COAD and TCGA-READ cohorts. Tumor Immune Dysfunction and Exclusion (TIDE) (http://tide.dfci.harvard.edu/) was used to predict immunotherapy outcomes, and violin plots were generated for TIDE scores between risk groups. Furthermore, the model was applied to the IMvigor210 cohort of patients treated with anti-programmed death-ligand 1 (PD-L1) therapy, using the IMvigor210CoreBiologies package, to evaluate associations between risk stratification and immunotherapy response.

Tumor mutational burden (TMB) analysis

Mutation data from the training cohort were used to summarize the top 20 mutated genes in high- vs. low-risk groups. TMB scores were calculated per sample, and Wilcoxon tests were used to compare TMB levels between groups.

Drug sensitivity analysis

To identify novel therapeutic opportunities, drug sensitivity was predicted using the pRRophetic package by calculating the half-maximal inhibitory concentration (IC50) for drugs in high- vs. low-risk groups. Lower IC50 indicated higher sensitivity. Additionally, the CellMiner database (https://discover.nci.nih.gov/cellminer/) was used to identify drugs significantly correlated with model feature genes (Table S5).

Construction of transcription factor (TF) and microRNA (miRNA) regulatory networks

The miRNet database (https://www.mirnet.ca/) was used to identify miRNAs potentially targeting the feature genes of the model. NetworkAnalyst (https://www.networkanalyst.ca/) was applied to predict upstream TFs regulating the feature genes. Both networks were visualized using Cytoscape.

Cells and culture conditions

Human colonic epithelial NCM460 cells (CL-0749) and human CRC cell lines, including HCT116 (CL-0096), HT-29 (CL-0118), and SW480 (CL-0223B) (all from Procell, Wuhan, China), were cultured in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10 fetal bovine serum (FBS; A5256701, ThermoFisher, Waltham, MA, USA) and 1 penicillin-streptomycin (BL505A, Beijing Lanjieker Technology Co., Ltd., Beijing, China) at 37 ℃ in a humidified atmosphere containing 5% CO₂. The CRC cell lines were authenticated by single-nucleotide polymorphism (SNP) array and short tandem repeat (STR) analyses.

Cell transfection

PLK4 was downregulated using a lentiviral vector carrying sh-PLK4, with a sh-negative control (sh-NC) (GeneChem, Shanghai, China). HT-29 and HCT116 in the logarithmic growth phase were trypsinized, resuspended, and seeded in 12-well plates at a density of 1×10⁵ cells/mL. After overnight incubation, the medium was replaced with fresh DMEM, and cells were infected with lentivirus diluted in phosphate-buffered saline (PBS). Following 24 h of incubation, the medium was refreshed, and successfully infected cells were selected with puromycin (concentration not specified) for 48 h. PLK4 knockdown was confirmed by quantitative real-time polymerase chain reaction (qRT-PCR) and western blot analyses.

qRT-PCR analysis

Total RNA was extracted from cells using TRIzol reagent (RC122-01, Nanjing Nuoweizan Biotechnology Co., Ltd., Nanjing, China). For each sample, 1 μg of RNA was reverse transcribed into cDNA using oligo(dT) primers according to the manufacturer’s instructions (R323-01, Nanjing Nuoweizan Biotechnology Co., Ltd., Nanjing, China). qRT-PCR was performed using SYBR Green qPCR Master Mix (Q711-03, Nanjing Nuoweizan Biotechnology Co., Ltd., Nanjing, China) with the following cycling program: initial denaturation at 95 ℃ for 10 min, followed by 40 cycles of 95 ℃ for 15 s and 60 ℃ for 45 s. Glyceraldehyde 3-phosphate dehydrogenase (GAPDH) was used as the internal control, and relative expression levels were calculated using the 2^−ΔΔCT method. All reactions were performed in triplicate. Primer sequences are listed in Table S6.

Western blot

HT-29 and HCT116 were lysed with radioimmunoprecipitation assay (RIPA) buffer (Beyotime, Shanghai, China) and centrifuged at 12,000 g for 15 min at 4 ℃. Supernatants were denatured at 98 ℃ for 20 min. Protein samples (40 μg) were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE; Bio-Rad, Hercules, CA, USA) and transferred onto 0.22 μm polyvinylidene fluoride (PVDF) membranes (Bio-Rad) for 30 min. Membranes were blocked with 5% non-fat milk in PBS for 1 h and incubated overnight at 4 ℃ with primary antibodies against PLK4 (sc-518222, Santa Cruz Biotechnology, Dallas, TX, USA) and GAPDH (sc-47724, Santa Cruz Biotechnology). Membranes were then incubated with horseradish peroxidase (HRP)-conjugated secondary antibody (1:50,000) for 1 h at room temperature. Protein bands were visualized using enhanced chemiluminescence.

Cell Counting Kit-8 (CCK-8) assay

After transfection, HT-29 and HCT116 cells were seeded in 96-well plates and cultured at 37 ℃. At 0 and 24 h, 10 μL of CCK-8 reagent (E606335, Sangon Biotech, Shanghai, China) was added to each well, and the optical density (OD) at 450 nm was measured using a microplate reader.

Apoptosis assay [Annexin V-FITC/propidium iodide (PI) staining]

HT-29 and HCT116 cells were harvested 24 h post-transfection and stained using an Annexin V-FITC/PI Apoptosis Detection Kit (AT101, LianKe Biotechnology, Hangzhou, China) according to the manufacturer’s instructions. Cells were incubated with Annexin V-FITC for 10 min and PI for 5 min, and apoptosis was analyzed by flow cytometry (C2060, LongCyte, Shanghai, China). Experiments were independently repeated three times.

Wound healing assay

HT-29 and HCT116 cells were grown in T75 flasks until 90% confluence, then seeded in 96-well plates at 5×10⁴ cells/well in DMEM. After 24 h, a scratch was made in the cell monolayer using a pipette tip. Detached cells were removed by washing with cold PBS, and the medium was replaced with serum-free DMEM. Cells were allowed to migrate for 24 h, and wound closure was observed under a light microscope (Zeiss GmbH, Jena, Germany). The rate of wound healing = [(the wound width of 0 h − 24 h)/0 h wound width] × 100%.

Transwell assays

HT-29 and HCT116 cells were re-suspended in serum-free DMEM and seeded on a transwell apical chamber with a membrane (8.0 μm, Corning Glass Works, Corning, NY, USA). The basolateral chamber (96-well plate) was filled with 10% FBS-DMEM. The apical chamber was loaded with 1×10⁵ cells after pre-coating with Matrigel before performing the invasion assay. Pre-coating was not required for the migration assay. After 24 h of incubation at 37 ℃, the cells that had invaded or migrated were fixed with 4% paraformaldehyde, stained with 0.5% crystal violet (Sangon) for 30 min, and then evaluated.

Statistical analysis

All analyses were conducted in R (v4.4.1). Wilcoxon tests were used for two-group comparisons. Kaplan-Meier survival curves were compared using the log-rank test. ROC curves were generated with the timeROC package. Visualization was performed using ggplot2. Correlations between variables were assessed by Pearson or Spearman methods. Median values were used for dichotomization. All in vitro experiments were performed in at least three independent biological replicates. Statistical significance was defined as P<0.05.

Results

Differential analysis of KRGs in CRC

We combined TCGA-COAD and TCGA-READ datasets as the training cohort and conducted PCA analysis. The results showed no obvious batch effect between the two datasets (Figure 2A), indicating that the data integration was of high quality and suitable for subsequent analyses. Differential expression analysis between tumor and normal samples in the training cohort identified 4,922 DEGs, including 2,363 upregulated and 2,559 downregulated genes, visualized in a volcano plot (Figure 2B). The large number of DEGs reflected significant transcriptomic differences between CRC and normal tissues. Intersecting DEGs with KRGs yielded 145 DEKRGs (Figure 2C). This provided the foundation for subsequent kinase-focused analyses. GO and KEGG enrichment analysis of these DEGs revealed biological processes and pathways of interest. As shown in Figure 2D, DEGs were mainly enriched in the biological process of protein serine/threonine kinase activity, suggesting that serine/threonine kinase activity may play a key role in CRC development. KEGG results (Figure 2E) indicated enrichment in pathways such as circadian entrainment and long-term potentiation. These KEGG terms are largely driven by shared calcium-, GPCR-, and neuroactive ligand-receptor signaling genes, so their appearance most likely reflects dysregulation of these common signaling modules in CRC rather than a direct role of the named physiological processes. Consistent with this, several enriched genes (e.g., CAMK2B, MAP2K6) converge on calcium and MAPK signaling, which are well-established contributors to colorectal tumorigenesis.

Figure 2.

Identification of prognosis-related kinase genes in CRC. (A) PCA analysis for batch effects in the training cohort; (B) volcano plot of DEGs; (C) intersection of DEGs and KRGs; (D) GO enrichment of DEGs; (E) KEGG enrichment of DEGs; (F) forest plot of prognosis-related DEKRGs; (G) expression of 15 prognosis-related DEKRGs in tumor vs. control; (H) correlation heatmap among 15 prognosis-related DEKRGs; (I) PPI network of 15 prognosis-related DEKRGs; (J) GO enrichment circle plot of the 15 prognosis-related DEKRGs, showing the z-scores of enriched GO biological processes and the logFC values of associated genes, with corresponding GO term descriptions. ***, P<0.001. BP, biological process; CC, cellular component; CRC, colorectal cancer; DEG, differentially expressed gene; DEKRG, differentially expressed kinase-related gene; FC, fold change; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; KRG, kinase-related gene; MF, molecular function; PCA, principal component analysis; PPI, protein-protein interaction.

Univariate Cox regression identified 15 prognosis-related DEKRGs with P<0.05 (Figure 2F, Table S2), representing a core set of kinase genes significantly associated with overall survival. Expression profiling revealed that RPS6KA5, CAMK2B, DAPK1, SPEG, NPR1, MAP2K6, BMX, and NTRK2 were significantly downregulated in CRC tumors, while CHEK1, BUB1, HASPIN, PLK4, ULK4, and FJX1 were significantly upregulated (Figure 2G). These patterns suggest their potential roles as tumor suppressors or oncogenes in CRC. A correlation heatmap among the 15 genes (Figure 2H) highlighted possible synergistic or antagonistic relationships. PPI network analysis (Figure 2I) further revealed interconnections among these genes, implying that prognosis-related kinases may function within shared or interwoven signaling pathways. GO enrichment of the 15 candidate genes demonstrated significant enrichment in processes such as negative regulation of mitotic nuclear division, protein autophosphorylation, and negative regulation of nuclear division (Figure 2J), which are closely associated with cell cycle regulation and protein modification, further underscoring their potential roles in tumor cell proliferation and survival.

Molecular subtyping of CRC based on KRGs and immune characteristics

Using the 15 prognosis-related DEKRGs, unsupervised consensus clustering was performed on CRC samples from the training cohort. Evaluation of clustering stability by the consensus matrix, cumulative distribution function, and delta area supported two robust molecular subtypes (Figure 3A-3C). PCA visualization confirmed distinct clustering between the two subtypes (Figure 3D). Kaplan-Meier survival analysis revealed significantly worse survival in subtype 1 (Figure 3E), suggesting a link between kinase-based subtypes and patient outcomes. Differential expression analysis between subtypes identified top-ranked DEGs, visualized in a heatmap (Figure 3F). KEGG enrichment of these DEGs revealed significant enrichment in pathways including chemokine signaling, viral protein interaction with cytokine and cytokine receptor, and cytokine-cytokine receptor interaction (Figure 3G), all highly relevant to immune regulation and inflammatory response, suggesting that subtype-specific differences involve tumor microenvironmental alterations. Expression boxplots of the 15 genes showed distinct expression patterns across subtypes (Figure 3H), forming the molecular basis of subtype classification.

Figure 3.

Molecular subtyping of CRC based on kinase-related gene signatures and immune characteristics. (A-C) Consensus clustering results; (D) PCA plots of subtypes; (E) Kaplan-Meier survival curves of subtypes; (F) heatmap of top DEGs between subtypes; (G) KEGG enrichment of DEGs; (H) expression of 15 prognosis-related genes across subtypes; (I) immune cell infiltration scores (ssGSEA); (J) immune function scores (ssGSEA); (K) expression of immune checkpoint-related genes across subtypes. ***, P<0.001; **, P<0.01; *, P<0.05; ns, P>0.05. CDF, cumulative distribution function; CRC, colorectal cancer; DEG, differentially expressed gene; KEGG, Kyoto Encyclopedia of Genes and Genomes; PCA, principal component analysis; ssGSEA, single-sample gene set enrichment analysis.

Immune infiltration analyses using ssGSEA demonstrated higher immune cell infiltration in subtype 1 compared with subtype 2 (Figure 3I). Subtype 1 also showed elevated immune-related functional scores (Figure 3J). Although high immune infiltration is often linked to complex immune microenvironments, its prognostic impact depends on immune cell types and functional status. Boxplot analysis of immune checkpoint gene expression revealed significant subtype differences (Figure 3K), with genes such as TNFSF13B, CCL5, and CCL19 being more highly expressed in subtype 1. These differences suggest distinct responses to immune checkpoint inhibitors (ICIs) across subtypes, with subtype 1 patients potentially more responsive to immunotherapy.

To further characterize the biological differences between the two kinase-related subtypes, we performed ssGSEA of selected kinase-associated Hallmark signaling programs. Subtype 1 showed higher enrichment of EMT, TGF-β, WNT/β-catenin, and Hedgehog signaling, whereas subtype 2 was more enriched in E2F targets, G2M checkpoint, MYC targets V1, PI3K/AKT/mTOR, and mTORC1 signaling (Figure S1A). Differential kinase-target co-expression network analysis further revealed distinct subtype-specific connectivity patterns, including altered connections involving HASPIN, NPR1, and BUB1, as well as stronger DAPK1-E2F2/EXO1 co-expression in subtype 2 (Figure S1B). Collectively, these findings indicate that the two kinase-related subtypes are characterized by distinct signaling programs and kinase-associated network architectures, providing additional biological support for the molecular classification. We further evaluated the distribution of the two kinase-related molecular subtypes across different clinicopathological and molecular backgrounds. Significant differences were observed in T and N stage distributions between the two subtypes, whereas no significant differences were found in overall stage or M stage. In addition, subtype distribution did not differ significantly according to mismatch repair (MMR) status or KRAS/BRAF status (Table S7).

Construction and validation of prognostic risk model in CRC

To assess patient prognosis, we constructed a CRC risk model. From univariate Cox regression, 15 candidate genes significantly associated with survival were identified. LASSO regression was then applied to avoid overfitting, yielding 9 feature genes (Figure 4A,4B, Table S3). Functional annotations of the nine signature genes, including their major biological roles and potential relevance to CRC, were further summarized (Table S8). Multivariate Cox regression was performed to construct the final prognostic model (Table S4).

Figure 4.

Construction and validation of prognostic risk model in CRC. (A) LASSO coefficient profiles; (B) LASSO regression coefficient plot; (C) ROC curves at 1, 3, and 5 years in training cohort; (D) Multivariate Cox forest plot; (E) Kaplan-Meier survival curves in training cohort; (F) Distribution of risk scores and survival status; (G) Expression of model feature genes in high- vs. low-risk groups; (H) PCA plot of risk groups; (I) ROC curves in validation cohort (GSE87211); (J) Kaplan-Meier curves in validation cohort; (K) risk score distribution and survival plots in validation cohort; (L) Expression of model feature genes in tumor vs. control. (M) Comparison with other prognostic models. ***, P<0.001; **, P<0.01; *, P<0.05. AIC, Akaike information criterion; AUC, area under the curve; CRC, colorectal cancer; LASSO, least absolute shrinkage and selection operator; PCA, principal component analysis; ROC, receiver operating characteristic.

The time-dependent ROC analysis in the training cohort showed AUCs of 0.675, 0.675, and 0.745 for 1-, 3-, and 5-year survival, respectively (Figure 4C), indicating moderate but stable discriminative ability that improved over longer follow-up intervals. The corresponding multivariate Cox regression results are shown in Figure 4D. Kaplan-Meier analysis revealed significantly worse survival in the high-risk group compared with the low-risk group (Figure 4E). Risk score distribution and survival status plots (Figure 4F) demonstrated a clear association between higher risk scores and poorer survival. Expression profiles of the 9 feature genes (Figure 4G) indicated that CAMK2B, DAPK1, FJX1, EPHA6, and NTRK2 were highly expressed in the high-risk group, whereas the remaining genes were enriched in the low-risk group. PCA confirmed distinct distributions of high- and low-risk patients (Figure 4H).

Validation in the external GSE87211 cohort showed higher discrimination, with AUCs of 0.904, 0.823, and 0.749 for 1-, 3-, and 5-year survival, respectively (Figure 4I). This higher discrimination relative to the training cohort likely reflects differences in cohort composition, follow-up duration, event rates, and the array-based platform of the validation set, rather than an intrinsic improvement in model performance; the consistent direction and statistical significance of risk stratification across both cohorts support the robustness of the signature. High-risk patients again demonstrated significantly worse survival (Figure 4J). Risk score distribution and survival plots (Figure 4K) confirmed robust stratification. Expression comparisons between tumor and normal samples revealed that CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, and NTRK2 were downregulated in tumors, whereas PLK4, ULK4, and FJX1 were upregulated (Figure 4L). These patterns underscore the biological relevance of the feature genes to CRC pathogenesis. Additionally, Figure 4M demonstrated the AUC values of our model compared to five other previously published models to highlight the superiority of our model (Table S9).

Kaplan-Meier analysis of individual feature genes further demonstrated that low expression of BMX, PLK4, and ULK4 predicted poor survival, while low expression of other genes indicated favorable prognosis (Figure S2). These analyses clarified the individual contributions of model genes to overall patient risk.

Functional enrichment analysis between high- and low-risk groups

Based on the prognostic risk model, patients were stratified into high- and low-risk groups by the median risk score. GSEA demonstrated that high-risk groups were enriched in biological processes such as cell differentiation (Figure 5A), suggesting greater tumor plasticity or dedifferentiated states. In contrast, low-risk groups were enriched in biological processes including cell cycle and tRNA modification (Figure 5B), indicating potentially higher proliferative activity but also greater sensitivity to conventional chemotherapy. Differential expression analysis between groups (Figure 5C) showed DEGs enriched in biological processes such as ERK1/ERK2 cascade, leukocyte migration, and connective tissue development (Figure 5D), as well as KEGG pathways including rheumatoid arthritis, protein digestion and absorption, and leishmaniasis (Figure 5E). In the KEGG database, immune/inflammatory disease terms such as rheumatoid arthritis and leishmaniasis are populated largely by cytokine, chemokine, and antigen-presentation genes; their enrichment therefore indicates differences in immune and inflammatory signaling between risk groups rather than involvement of these specific diseases. Together with the leukocyte-migration and ERK1/ERK2 cascade terms, these results point to differences primarily in immune/inflammatory signaling and tissue remodeling.

Figure 5.

Functional enrichment results between high- and low-risk groups. (A) GSEA results for high-risk group; (B) GSEA results for low-risk group; (C) volcano plot of DEGs between groups; (D) GO enrichment of DEGs; (E) KEGG enrichment of DEGs. DEG, differentially expressed gene; FC, fold change; FDR, false discovery rate; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; GSEA, gene set enrichment analysis.

Construction of a nomogram integrating clinical features and risk score

To comprehensively evaluate prognosis in CRC patients, both risk scores and clinical features were included in univariate (Figure 6A) and multivariate Cox analyses (Figure 6B). Risk score, age, stage, M stage, and N stage were all significantly associated with survival, indicating that the risk score is an independent prognostic factor beyond conventional staging. A nomogram was then constructed combining risk score and clinical factors for individualized prediction of 1-, 3-, and 5-year survival (Figure 6C). Calibration plots demonstrated strong concordance between predicted and observed survival outcomes (Figure 6D). DCA indicated substantial net clinical benefit at 1, 3, and 5 years (Figure 6E). Collectively, the nomogram effectively predicted prognosis and has potential utility for clinical decision-making. The joint analysis in this section highlighted the advantages of risk models over traditional tumor-node-metastasis (TNM) staging. Notably, the combined analysis integrating risk scores with clinical variables demonstrated that risk models provided significant incremental predictive value beyond traditional TNM staging, rather than serving as a replacement. Furthermore, direct comparison with carcinoembryonic antigen (CEA) was not feasible due to the absence of CEA data in the TCGA cohort. Previous study reported CEA 1-, 3-, and 5-year AUCs of 0.696–0.744 (14), comparable to our model's training set AUC range (0.675–0.745), confirming its robust generalizability. These findings indicated that combining the risk score with CEA could further enhance treatment monitoring sensitivity.

Figure 6.

Construction and validation of a comprehensive prognostic nomogram. (A) Forest plot of univariate Cox analysis; (B) forest plot of multivariate Cox analysis; (C) nomogram integrating risk score and clinical factors; (D) calibration plots for 1-, 3-, and 5-year OS; (E) DCA curves for 1-, 3-, and 5-year OS. DCA, decision curve analysis; CI, confidence interval; M, metastasis; N, node; OS, overall survival; T, tumor.

Performance of the risk model across clinical subgroups

We further analyzed the distribution of risk scores across different clinicopathological features, including age, sex, stage, T stage, N stage, and M stage. Violin plots revealed significant differences in risk scores across TNM stages (Figure 7A), consistent with disease progression. Subgroup Kaplan-Meier analyses stratified by age (≤65 vs. >65 years), sex (male vs. female), stage (I + II vs. III + IV), T stage (T1+2 vs. T3+4), N stage (N0 vs. N1–2), and M stage (M0 vs. M1) showed that high-risk groups consistently had significantly worse overall survival compared with low-risk groups in all clinical contexts (Figure 7B). These findings indicate that the model is robust and applicable across diverse patient subgroups. A summary of clinicopathological distributions between high- and low-risk groups is presented in Table 1.

Figure 7.

Stratification analysis of high- and low-risk groups based on clinical features. (A) Violin plots of risk scores across clinicopathological features; (B) Kaplan-Meier survival curves of clinical subgroups stratified by risk score. M, metastasis; N, node; T, tumor.

Table 1. Clinical characteristics of high- and low-risk groups in the training cohort.

Name	High (n=295)	Low (n=295)	P value
OS			<0.001
0	214 (72.5)	256 (86.8)
1	81 (27.5)	39 (13.2)
OS time, years	2.2±1.7	2.6±2.2	0.009
Age, years	66.3±12.4	65.0±12.4	0.19
Gender			0.74
Female	137 (46.4)	132 (44.7)
Male	158 (53.6)	163 (55.3)
Stage			0.050
Stage I	40 (13.6)	61 (20.7)
Stage II	102 (34.6)	111 (37.6)
Stage III	90 (30.5)	81 (27.5)
Stage IV	51 (17.3)	33 (11.2)
Unknown	12 (4.1)	9 (3.1)
T			0.002
T1	6 (2.0)	12 (4.1)
T2	40 (13.6)	63 (21.4)
T3	203 (68.8)	199 (67.5)
T4	44 (14.9)	20 (6.8)
Unknown	2 (0.7)	1 (0.3)
N			0.006
N0	152 (51.5)	181 (61.4)
N1	70 (23.7)	74 (25.1)
N2	71 (24.1)	38 (12.9)
Unknown	2 (0.7)	2 (0.7)
M			0.03
M0	218 (73.9)	219 (74.2)
M1	50 (16.9)	33 (11.2)
Unknown	27 (9.2)	43 (14.6)

Data are presented as mean ± standard deviation or n (%). M, metastasis; N, node; OS, overall survival; T, tumor.

Immune infiltration analysis and prediction of immunotherapy response

To explore immune microenvironmental differences, ssGSEA-based heatmaps showed distinct immune infiltration patterns between high- and low-risk groups (Figure 8A). CIBERSORT analysis further revealed significant differences in immune cell compositions (Figure 8B), with macrophage infiltration elevated in high-risk groups and CD4⁺ T cell infiltration higher in low-risk groups. Since macrophages, particularly M2 type, promote immunosuppression and tumor progression, whereas CD4⁺ T cells support antitumor immunity, these findings help explain prognostic differences. A correlation heatmap between feature genes and immune infiltration is shown in Figure 8C.

Figure 8.

Immune infiltration and immunotherapy response analysis. (A) ssGSEA heatmap of immune infiltration; (B) boxplots of immune infiltration by CIBERSORT; (C) correlation heatmap of feature genes and immune cells; (D) boxplots of immune checkpoint gene expression; (E) IPS scores between risk groups; (F) TIDE score violin plots; (G) Kaplan-Meier curves in IMvigor210 cohort; (H) objective response rates (R vs. NR) across risk groups; (I) risk score distribution between responders and non-responders. ns, P>0.05; **, P<0.01; ***, P<0.001. IPS, immunophenoscore; NR, non-response; R, response; ssGSEA, single-sample gene set enrichment analysis; TIDE, Tumor Immune Dysfunction and Exclusion.

Immune checkpoint expression also varied significantly (Figure 8D), with most immune checkpoint-related genes upregulated in low-risk groups, suggesting greater responsiveness to ICIs. IPS scores were significantly higher in low-risk groups (Figure 8E), supporting stronger immunotherapy sensitivity. TIDE analysis showed lower TIDE scores in low-risk groups (Figure 8F), consistent with better predicted immunotherapy outcomes.

In the IMvigor210 cohort (anti-PD-L1 therapy), low-risk patients exhibited significantly better survival (Figure 8G). Response analysis revealed that low-risk patients had higher objective response rates [complete response (CR)/partial response (PR) vs. stable disease (SD)/progress disease (PD)] (Figure 8H), with responders showing significantly lower risk scores than non-responders (Figure 8I). Collectively, these results highlight the predictive utility of the risk model for immunotherapy outcomes.

TMB and drug sensitivity differences

Mutation landscapes were compared between risk groups. The top 20 mutated genes were summarized for each group, with APC being the most frequently mutated gene in both (Figure 9A,9B), consistent with typical CRC mutational profiles. TMB levels were significantly higher in low-risk groups (Figure 9C). High TMB was associated with better prognosis (Figure 9D), consistent with increased neoantigen generation and enhanced immunotherapy response.

Figure 9.

Mutation and drug sensitivity landscape. (A) Waterfall plot of top 20 mutated genes in low-risk group; (B) waterfall plot in high-risk group; (C) violin plot of TMB between risk groups; (D) Kaplan-Meier survival by TMB status; (E-H) violin plots of IC50 for four drugs across risk groups; (I-L) correlation between feature genes and drug sensitivity in CellMiner; (M) miRNA-gene regulatory network; (N) TF-gene regulatory network. IC50, half-maximal inhibitory concentration; miRNA, microRNA; TF, transcription factor; TMB, tumor mutational burden.

Drug sensitivity analysis predicted by the pRRophetic algorithm showed that low-risk patients were more sensitive to 5-fluorouracil (Figure 9E), whereas high-risk patients exhibited greater sensitivity to docetaxel (Figure 9F) and trametinib (Figure 9G). Low-risk patients were also more sensitive to vinorelbine (Figure 9H). Correlation analysis with the CellMiner database revealed: CAMK2B negatively correlated with selumetinib (r=−0.365) (Figure 9I); DAPK1 positively correlated with erlotinib and gefitinib (r=0.354, 0.409) (Figure 9J,9K); FJX1 negatively correlated with vemurafenib (r=−0.533) (Figure 9L). These results provide potential therapeutic implications for targeted therapies.

Furthermore, regulatory network analyses identified 37 miRNAs potentially targeting model feature genes via miRNet, constructing a miRNA-gene regulatory network (Figure 9M). Upstream transcription factor prediction by NetworkAnalyst identified 40 TFs (Figure 9N), providing insights into transcriptional regulation of feature genes.

Analysis of CMS subtypes in the TCGA cohort

Based on the TCGA cohort, we conducted a CMS subtype analysis. We calculated the distribution of CMS subtypes in CRC across different subtypes and risk groups, as shown in Table 2. Significant differences in CMS subtype distribution were observed between groups. Kinase subtype 1 and the high-risk group predominantly exhibited CMS4, while kinase subtype 2 and the low-risk group primarily showed CMS2.

Table 2. Distribution of CRC CMS subtypes in the TCGA cohort.

Group	CMS1, n (%)	CMS2, n (%)	CMS3, n (%)	CMS4, n (%)	Unknown, n (%)	P value
Subtype						<0.001
1 (N=272)	28 (10.3)	47 (17.3)	49 (18)	124 (45.6)	24 (8.8)
2 (N=368)	70 (19.0)	129 (35.1)	46 (12.5)	82 (22.3)	41 (11.1)
Risk						<0.001
High (N=295)	52 (17.6)	58 (19.7)	22 (7.5)	143 (48.5)	20 (6.8)
Low (N=295)	37 (12.5)	106 (35.9)	67 (22.7)	47 (15.9)	38 (12.9)

CMS, consensus molecular subtype; CRC, colorectal cancer; TCGA, The Cancer Genome Atlas.

Validation of core genes in CRC cell lines

To validate bioinformatics predictions of nine core genes (CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, NTRK2, PLK4, ULK4, FJX1), we performed qRT-PCR in CRC cell lines HCT116, SW480, HT-29, and in normal colon epithelial cells NCM460. The results showed that CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, and NTRK2 were significantly downregulated in CRC cells, whereas PLK4, ULK4, and FJX1 were significantly upregulated (P<0.01), consistent with bioinformatics analysis (Figure 10). These findings further confirm the reliability of the predicted core genes in vitro.

Figure 10.

Validation of 9 core genes in CRC cell lines. qRT-PCR analysis of core gene expression (CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, NTRK2, PLK4, ULK4, FJX1) in normal colon epithelial cell line NCM460 and CRC cell lines HCT116, SW480, HT-29. Data are shown as mean ± SD; *, P<0.05; **, P<0.01. CRC, colorectal cancer; qRT-PCR, quantitative real-time polymerase chain reaction; SD, standard deviation.

Functional validation of PLK4 in CRC cells

To further explore the biological function of PLK4 in CRC cells, we knocked down PLK4 (sh-PLK4) in HT-29 and HCT116 cells and performed assays for proliferation, apoptosis, and migration. qRT-PCR and Western blot analysis showed that PLK4 mRNA and protein levels were significantly reduced in the sh-PLK4 group compared with the controls in both cell lines (Figure 11A,11B). CCK-8 assays demonstrated that PLK4 knockdown significantly inhibited cell proliferation in both HT-29 and HCT116 (Figure 11C), while flow cytometry analysis revealed a marked increase in apoptosis following PLK4 knockdown (Figure 11D). In addition, both wound healing and Transwell migration assays indicated that PLK4 knockdown substantially suppressed the migration of both HT-29 and HCT116 cells (Figure 11E,11F). Taken together, the consistent results in two CRC cell lines indicate that PLK4 promotes proliferation, inhibits apoptosis, and enhances migration in CRC cells, further supporting its pro-tumorigenic role in CRC.

Figure 11.

PLK4 knockdown suppresses proliferation and migration and induces apoptosis in HT-29 and HCT116 cells. (A) qRT-PCR analysis showing PLK4 mRNA knockdown efficiency; (B) western blot analysis showing PLK4 protein knockdown; GAPDH was used as a loading control; (C) CCK-8 assay showing reduced cell viability upon PLK4 knockdown; (D) flow cytometry analysis showing increased apoptosis in sh-PLK4 cells; (E) wound healing assay demonstrating reduced migration; (F) transwell migration assay confirming decreased migration after PLK4 knockdown (crystal violet staining). Data are presented as mean ± SD; *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. CCK-8, Cell Counting Kit-8; NC, negative control; qRT-PCR, quantitative real-time polymerase chain reaction; SD, standard deviation.

Discussion

In this study, we identified kinase-related molecular subtypes in CRC and successfully constructed a prognostic risk model. Our findings demonstrated that different subtypes exhibited significant differences in gene expression, signaling pathway activation, immune microenvironment composition, and clinical prognosis. Patients in the high-risk subtype showed markedly poorer survival, suggesting that aberrant kinase expression is closely associated with CRC progression. Functional annotation of the nine signature genes revealed that they encompass diverse but biologically relevant kinase-associated processes, including cell-cycle regulation, calcium-dependent signaling, apoptosis and stress responses, immune/inflammatory signaling, and receptor and non-receptor tyrosine kinase signaling (Table S8). This functional diversity supports the biological coherence of the signature and suggests that its prognostic value may reflect coordinated alterations across multiple kinase-related signaling processes. These results are consistent with recent studies on the role of PRKCI in CRC metastasis and drug resistance, further confirming the central role of kinases in tumor malignancy (15,16).

The kinase-based subtypes identified in this study showed partial but incomplete overlap with the CMS classification, with kinase subtype 1 predominantly enriched in CMS4 and subtype 2 mainly distributed in CMS2. Similar to CMS4, subtype 1 was characterized by enrichment of TGF-β signaling, EMT-related features, stromal infiltration, an immunosuppressive microenvironment, and unfavorable prognosis (2). However, the two classification systems are not equivalent: CMS reflects broader transcriptomic characteristics, whereas our classification focuses specifically on KRGs and associated signaling dysregulation, consistent with previous evidence of subtype-specific kinase dependencies in mesenchymal CRC (8). In addition, subtype 1 showed greater predicted sensitivity to the MEK inhibitor trametinib, suggesting that kinase-based classification may provide therapeutic information complementary to the CMS framework. Together, these findings indicate that the two systems capture overlapping but distinct biological dimensions of CRC heterogeneity. Beyond molecular subtyping, the risk model provided consistent and reproducible prognostic stratification across both the training and independent validation cohorts. Although its discriminative accuracy in the training cohort was moderate (AUC: 0.68–0.75), the risk score remained an independent predictor of survival after adjustment for TNM stage [hazard ratio (HR) =1.566, P<0.001], and its integration with clinical variables in the nomogram further improved predictive performance, positioning it as a complementary risk-stratification tool rather than a replacement for conventional staging. This performance is comparable to recently developed prognostic models based on cell cycle checkpoint genes or immune-related signatures (17). Compared with previous CRC prognostic and molecular classification studies based on diverse biological features (18-20), our study specifically focused on KRGs, a functionally coherent class of signaling molecules with potential therapeutic relevance in CRC (21). Beyond survival stratification, the kinase-based framework was further linked to CMS classification, immune infiltration, immune checkpoint expression, TMB, drug sensitivity, and predicted immunotherapy response. Together with independent external validation, these features support a kinase-oriented framework integrating molecular heterogeneity, prognosis, and potential therapeutic response in CRC.

To further evaluate the predictive performance of this model, we compared it with five published prognostic models for CRC. The results showed that this model had better AUC values in the independent validation set (1-year: 0.904, 3-year: 0.823, 5-year: 0.749) than existing models, indicating good predictive efficacy and generalizability. Simultaneously, multivariate Cox regression analysis showed that the risk score was independent of TNM stage (HR =1.566, P<0.001), and the predictive accuracy was further improved after the nomogram integrated the risk score and TNM stage. For CEA, a commonly used clinical indicator, we could not directly compare it because the TCGA cohort lacked relevant data. However, the reported 1- to 5-year AUC range of CEA is between 0.696 and 0.744 (14), while the AUC range of this model is 0.675 to 0.745, demonstrating comparable predictive ability, suggesting that combining the risk score with CEA in the future may further improve the sensitivity of treatment monitoring.

Further analyses revealed that the high-risk kinase subtype was associated not only with tumor invasion and metastasis but also with features of potential chemoresistance. Previous studies have shown that kinases such as PRKCI promote extracellular matrix degradation and EMT through regulation of the TGF-β and PI3K/AKT pathways, thereby reducing chemotherapy sensitivity (22,23). Our results similarly suggest that aberrant kinase activity may play a key role in CRC drug resistance, providing a rationale for exploring combined strategies using kinase inhibitors and conventional chemotherapy. In drug sensitivity analysis, this study found that the high-risk group exhibited greater sensitivity to trametinib, a MEK inhibitor. Trametinib, as a selective MEK inhibitor, has demonstrated the ability to enhance the anti-tumor effects of 5-fluorouracil in CRC and functions by inhibiting the MAPK signaling pathway (24). Existing research indicates that abnormal activation of the MAPK signaling pathway plays a crucial role in the development and progression of CRC, with the expression levels of genes related to this pathway being closely associated with patient prognosis (25,26). Our pathway enrichment analysis revealed that DEGs between the high- and low-risk groups were significantly enriched in the positive regulation of the MAPK cascade, suggesting that MAPK pathway activity may be upregulated in the high-risk group. This finding offers a potential mechanistic explanation for the higher sensitivity of the high-risk group to trametinib: tumor cells with upregulated MAPK pathway activity may be more dependent on signaling through this pathway, thus rendering them more sensitive to MEK inhibitors.

In addition, the high-risk subtype displayed a unique immune infiltration profile, with enrichment of immunosuppressive cells and elevated expression of immune checkpoint genes, while the low-risk subtype demonstrated stronger antitumor immune responses. These findings are consistent with recent reports that dysregulated signaling pathways modulate the tumor immune microenvironment, suggesting that kinases not only affect intrinsic tumor cell behavior but also participate in tumor progression via immune regulation (27,28). Of note, despite exhibiting higher overall immune infiltration, the high-risk group demonstrated poorer prognosis. This seemingly paradoxical phenomenon suggests that the functional status and composition of infiltrating immune cells may be more critical than their quantity. Consistent with the immune infiltration analysis, the high-risk group showed significant enrichment of immunosuppressive cells, such as M2 macrophages, alongside a marked reduction in the infiltration of CD4⁺ T cells, which possess antitumor activity. This indicates that the immune microenvironment in the high-risk group is skewed toward an immunosuppressive phenotype, with effector T cells potentially in a state of exhaustion, thereby promoting tumor immune escape. The kinase genes in the model may play a crucial role in shaping this immunosuppressive microenvironment: FJX1 is negatively correlated with CD8⁺ T cells, and its high expression can lead to reduced immunotherapy efficacy and drug resistance (29). DAPK1 can mediate CSN5 phosphorylation via IKKβ, further deubiquitinating PD-L1, thereby weakening NK cell-mediated killing effects (30). PLK4 expression levels are positively correlated with most immunosuppressive cytokines (31). TIDE algorithm analysis also confirmed a higher risk of resistance to immunotherapy in the high-risk group. Validation using the IMvigor210 cohort demonstrated a significantly lower objective response rate compared to the low-risk group. In-depth analysis suggests that this phenomenon may be attributed to the following mechanisms: first, while chemokine signaling pathways recruit immune cells into the tumor microenvironment, these cells are reprogrammed into pro-tumor phenotypes, forming an immunosuppression-dominant microenvironment. Second, GSEA of the high-risk group revealed active cell differentiation pathways, suggesting that tumor cells may competitively inhibit T cell function through metabolic reprogramming. Finally, high immune infiltration may reflect continuous immune cell recruitment in a chronic inflammatory state, rather than an effective anti-tumor immune response, which is consistent with the stroma-rich, EMT characteristics of the CMS4 subtype. This paradox suggests that the sheer number of immune cells is insufficient to predict prognosis; instead, the functional status of immune cells and the polarization patterns of the tumor microenvironment are more critical. Future single-cell RNA sequencing analysis will help to further elucidate the key nodes of this immunosuppressive network.

To validate the reliability of bioinformatics analysis, we performed qRT-PCR to examine the expression of nine core genes (CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, NTRK2, PLK4, ULK4, FJX1) in CRC cell lines (HCT116, SW480, HT-29) compared to normal colon epithelial cells (NCM460). The results showed that CAMK2B, DAPK1, MAP2K6, EPHA6, BMX, and NTRK2 were significantly downregulated in CRC cells, while PLK4, ULK4, and FJX1 were significantly upregulated. This expression pattern was consistent with our bioinformatics findings, confirming the role of these kinases as potential drivers of malignant behavior in CRC. Specifically, we investigated the biological function of PLK4, a key gene in our model, by constructing an sh-PLK4 knockdown model in two independent CRC cell lines, HT-29 and HCT116. Our results demonstrated that PLK4 knockdown consistently inhibited cell proliferation and migration and promoted apoptosis in both cell lines, strengthening the evidence for a pro-tumorigenic role of PLK4 in CRC. Consistent with previous studies that have highlighted the oncogenic role of PLK4 in CRC through pathways such as Wnt/β-catenin (32), our study further supports the biological relevance of PLK4 within the kinase-associated risk signature. Together with the consistent phenotypic effects observed in HT-29 and HCT116 cells, these results provide additional experimental evidence for a pro-tumorigenic role of PLK4 in CRC, although further in vivo studies are required to establish its therapeutic relevance.

Although this study demonstrates the potential clinical relevance of kinase-associated subtypes and provides functional validation of PLK4 in two CRC cell lines, several limitations should be acknowledged. First, the model was primarily constructed using TCGA data and validated in a single external cohort (GSE87211). In addition, the high proportion of missing information on MMR status and KRAS/BRAF status in TCGA precluded reliable evaluation of the prognostic performance of the kinase-based classification within these molecular subgroups. Further validation in larger independent cohorts with complete molecular annotation is required. Therefore, its generalizability requires further evaluation in prospective, multicenter studies. Second, although the functional role of PLK4 was validated in HT-29 and HCT116 cells, the other eight signature genes were mainly supported by expression validation and functional annotation, and further gene-specific functional studies and in vivo experiments are needed to clarify their roles in CRC progression and immune regulation. In addition, bioinformatics analyses were primarily based on transcriptomic data, and integrating phosphoproteomics and single-cell sequencing data would provide a more comprehensive understanding of kinase activity and its impact on the tumor microenvironment. Finally, the predicted implications for immunotherapy remain preliminary and warrant clinical validation. Addressing these limitations will be critical for translating kinase-based molecular subtyping and prognostic modeling into clinical practice.

Conclusions

This study systematically identified kinase-related molecular subtypes in CRC and constructed a prognostic risk model that was validated in an independent external cohort. Additional in vitro experiments in two CRC cell lines further supported the pro-tumorigenic role of PLK4. High-risk subtypes were associated with poor survival, increased invasive potential, an immunosuppressive tumor microenvironment, and potential chemoresistance, while low-risk subtypes showed better prognosis and potential sensitivity to immunotherapy. These findings highlight the critical role of protein kinases in CRC progression and provide a practical framework for patient stratification. Future studies integrating multi-omics data and prospective clinical validation will be essential to translate these subtypes and the risk model into clinical practice, ultimately supporting personalized treatment strategies in CRC.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data Sharing Statement

Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1446/dss

tcr-15-9-683-dss.pdf ^{(73.9KB, pdf)}

DOI: 10.21037/tcr-2026-1446