Integrated single-cell and machine learning analysis identifies PMAIP1 as a novel biomarker for predicting prognosis and immunotherapy response in colorectal cancer

Abstract

Metastasis in colorectal cancer (CRC) is the primary driver of its high mortality rate, with dysregulation of apoptosis and remodeling of the tumor microenvironment (TME) serving as key biological events driving disease progression. However, the intrinsic molecular connections between these processes—particularly the dynamic changes during the progression from primary tumors to liver metastases—remain to be systematically elucidated. This study aims to identify core genes regulating the crosstalk between apoptosis and immunity in CRC using multidimensional computational biology approaches and to explore their clinical value as prognostic biomarkers and potential therapeutic targets. This study integrated single-cell RNA sequencing (scRNA-seq) data with bulk transcriptomic data from The Cancer Genome Atlas (TCGA) database. We first utilized scRNA-seq to dissect cellular heterogeneity and communication networks in CRC primary tumors and liver metastases. Subsequently, weighted gene co-expression network analysis (WGCNA) and differential expression analysis were employed, combined with machine learning algorithms (LASSO, Random Forest, XGBoost), to screen features from apoptosis- and immunity-related gene sets and pinpoint key candidate genes. Finally, the biological functions and clinical significance of the core gene were systematically validated through pan-cancer analysis, immune infiltration analysis, mutation profile analysis, prognostic model construction, and validation via real-time quantitative PCR (RT-qPCR) in clinical samples. Single-cell profiling revealed that colorectal cancer (CRC) liver metastases exhibit a distinct pro-inflammatory immune microenvironment characterized by the enrichment of CD8⁺ effector memory T cells and inflammatory monocytes, along with extensive remodeling of cell-cell communication networks centered on the MIF and CD99 signaling pathways. Through multi-level bioinformatics analysis, we identified the apoptosis-related gene PMAIP1 as a central regulatory hub mediating these processes. Clinical cohort and pan-cancer analyses demonstrated that PMAIP1 is significantly overexpressed across multiple tumor types, particularly in CRC, where its elevated expression serves as an independent predictor of poor prognosis (HR = 0.39, p = 0.028). Further investigation uncovered a critical biological paradox: On the one hand, the high expression of PMAIP1 was positively correlated with the infiltration of core effector immune cells, such as CD8⁺T cells. CD8⁺T cells are the key immune cells mediating tumor cell killing, and their enrichment is expected to form an effective anti-tumor immune barrier, which is consistent with the characteristics of the “pro-inflammatory” immune microenvironment. On the other hand, the high expression of PMAIP1 is also significantly associated with the up-regulation of multiple immune checkpoint molecules, such as PD-1 and PD-L1. The activation of immune checkpoint molecules directly induces the functional exhaustion of infiltrating CD8⁺T cells, which is manifested as decreased proliferation ability, decreased secretion of killer cytokines, and loss of anti-tumor activity. In short, PMAIP1 both “recruits” effector immune cells with antitumor potential and “weakens” these cells by upregulating immune checkpoint molecules, resulting in the coexistence of two contradictory phenotypes of “increased immune cell infiltration” and “decreased immune response efficiency”. Finally, it still promotes the progress of colorectal cancer and is related to the poor prognosis of patients. A nomogram model incorporating PMAIP1 demonstrated strong performance in predicting patient survival, and its overexpression in clinical CRC samples was validated by RT-qPCR. This study systematically delineates the immune landscape of CRC primary tumors and liver metastases, and for the first time reveals the apoptosis gene PMAIP1 as a critical regulatory node linking cellular apoptosis to immune microenvironment remodeling. PMAIP1 not only emerges as a potent, novel biomarker for poor prognosis in CRC but also, through its dual functions in regulating immune infiltration and checkpoints, represents a highly promising target for guiding combined immunotherapy strategies.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-32933-8.

Introduction

Colorectal cancer (CRC) remains a major global health burden, ranking third in incidence and second in cancer-related mortality^1,2. While incidence among older adults has declined in developed regions such as North America and Europe due to widespread screening, cases in individuals under 50 years are rising sharply—driven by Westernized diets, obesity, and sedentary lifestyles^2,3. In many developing nations, the burden continues to grow as inequities in screening and limited healthcare infrastructure persist^4,5. Despite advances in surgery, chemoradiotherapy, and targeted therapies against EGFR, VEGF, and related pathways, the five-year survival rate for metastatic CRC remains below 15%⁶. This poor outcome reflects profound intra- and intertumoral heterogeneity and the resulting resistance to systemic therapies. Understanding the molecular networks that drive CRC progression has therefore become a central priority in cancer research⁷.

Over the past decade, transcriptomic profiling—spearheaded by high-throughput sequencing—has transformed the molecular characterization of CRC. Bulk RNA sequencing (bulk RNA-seq) from large patient cohorts has delineated key oncogenic pathways such as Wnt/β-catenin and MAPK and established the Consensus Molecular Subtype (CMS) framework, which stratifies CRC into four biologically and clinically distinct subtypes (CMS1–4)^8–10. However, bulk RNA-seq captures only an averaged transcriptional signal across millions of cells, masking the complexity of the tumor microenvironment (TME)^11,12. The emergence of single-cell RNA sequencing (scRNA-seq) has overcome this limitation, enabling transcriptomic analysis at cellular resolution. By dissecting the TME with unprecedented granularity, scRNA-seq has revealed rare but functionally decisive populations—cancer stem cells, diverse immune subsets, and heterogeneous cancer-associated fibroblasts (CAFs)—that collectively shape CRC progression¹³.

Despite the remarkable advances of both technologies, their effective integration to tackle critical clinical questions remains a central challenge in current cancer research. Bulk RNA-seq data from large cohorts such as TCGA provide extensive clinical annotations and prognostic information, yet these “averaged” molecular signatures cannot be confidently mapped to specific cell types, limiting their translational potential as clinical biomarkers or therapeutic targets. Conversely, scRNA-seq offers unparalleled resolution for dissecting cellular heterogeneity, but the novel molecules and regulatory pathways it reveals often lack validation in large clinical cohorts due to limited sample sizes, and their biological relevance demands rigorous experimental verification^14,15. This persistent disconnect between discovery and validation constrains the transition of promising bioinformatic targets into preclinical pipelines, impeding the progress of precision oncology in CRC.

To bridge this critical gap, we developed a multi-tiered integrative framework combining computational, single-cell, and experimental approaches. First, we performed systematic bioinformatics analyses using bulk RNA-seq data from large public datasets to identify candidate genes associated with CRC prognosis. We then leveraged scRNA-seq data to characterize the expression patterns and potential functions of these genes across distinct cell populations within the tumor microenvironment at single-cell resolution. Finally, through real-time quantitative PCR (RT-qPCR) and immunohistochemistry (IHC) analyses in clinically collected CRC specimens, we validated the expression and clinical significance of the core candidate genes at both transcriptomic and protein levels. Collectively, this integrative strategy provides a refined molecular perspective on CRC heterogeneity and establishes a robust experimental foundation for the development of clinically actionable prognostic biomarkers and therapeutic targets.

Materials and methods

Data sources and cohort construction

In this study, two complementary transcriptome datasets were integrated for multi-dimensional analysis. First, to achieve high-resolution single-cell profiling, the single-cell RNA sequencing (scRNA-seq) dataset GSE178318 was obtained from the Gene Expression Omnibus (GEO) database, encompassing samples from primary tumors and liver metastases of 15 colorectal cancer (CRC) patients. Second, to validate and extend the single-cell findings at the tissue level, bulk RNA-seq data and corresponding clinical information were retrieved from the official data portal of The Cancer Genome Atlas (TCGA). This cohort included transcriptomic profiles—quantified as FPKM values—of 163 CRC tumor tissues and 10 adjacent normal tissues used as controls. In order to improve the robustness of the core gene conclusions, in addition to GSE60331, three colorectal cancer (CRC) GEO datasets with different sequencing platforms and clinical backgrounds were included as independent validation cohorts. The screening criteria were as follows: (1) the gene expression data of CRC tumor tissues and adjacent normal tissues were included; (2) complete clinical follow-up information (overall survival (OS)/disease-free survival (DFS)); (3) Sample size ≥ 100 to avoid small sample bias. Specific cohort information and preprocessing procedures were as follows: (1) GSE39582 (Platform: Affymetrix HG-U133Plus2) : A total of 566 CRC tumor tissues and 19 normal tissues were included. The median follow-up time was 60 months, covering patients with stages I-IV. The original CEL files were standardized by RMA algorithm and converted to FPKM format. (2) GSE17536 (Platform: Agilent-014850 Whole Human Genome Microarray 4× 44 K G4112F) : A total of 177 CRC tumor tissues and 11 normal tissues were included, and a subgroup of 89 patients who received FOLFOX chemotherapy was included. quantile normalization was used to eliminate the batch variation within the platform. (3) GSE29621 (Platform: Affymetrix HG-U133A) : A total of 145 CRC tumor tissues and 19 normal tissues were included, focusing on patients with early stage I-II CRC. After RMA standardization, the gene ID annotation was unified with other cohorts (matched with HGNC standard gene name). All expanded cohorts were preprocessed by “remove genes with zero expression” and “log₂(FPKM + 1) transformation” to ensure consistent quantification standards with TCGA-CRC and GSE60331 datasets and meet the needs of cross-cohort comparison.

Single-cell transcriptomic analysis

All computational analyses were performed in the R statistical environment (version 4.1.1). Single-cell data preprocessing, analysis, and visualization were conducted using the Seurat package (version 4.2). The analytical pipeline began with a rigorous quality-control (QC) procedure to ensure data fidelity and analytical precision. Cells were excluded if they expressed fewer than 300 or more than 7,500 genes (to remove low-quality cells or potential doublets) or exhibited > 25% mitochondrial transcripts, indicative of stressed or apoptotic states. To further minimize biological confounders, genes related to mitosis, ribosomal function, or hemoglobin metabolism were systematically removed from the matrix. After stringent filtering, a high-quality expression matrix containing 33,162 cells and 33,408 genes was retained for downstream analyses.

Canonical Correlation Analysis (CCA) was applied to adjust for batch effects due to different sample sources. Using Seurat’s built-in IntegrateData function in combination with the canonical correlation analysis (CCA) + Anchor integration strategy (Seurat v4.2 standard batch calibration process), this approach works by identifying “anchor cells” (cells with highly similar gene expression patterns) across batches. Batch-specific variation was eliminated while preserving biological heterogeneity. Anchor cells were screened using the FindIntegrationAnchors function, setting k.an nchor = 20 (each cell matches 20 potential anchor cells) and dim = 1:30 (similarity is calculated based on the first 30 principal components) to ensure the representativeness of anchor cells. Data integration: Call the IntegrateData function, set dim = 1:30, k.eight = 10 (the integration weight of each cell is calculated based on the 10 closest anchors), and generate the integration expression matrix after removing batch effect. After data integration, we used the FindVariableFeatures function (with default parameters) to identify the 2000 Highly Variable Genes (HVGs) with the highest variability in the expression matrix. The expression levels of these genes were normalized by Z-score and then entered into Principal Component Analysis (PCA) to achieve linear dimensionality reduction. Subsequently, unsupervised graph clustering was performed using the FindClusters function (resolution parameter set to 0.5) based on statistically significant principal components to classify the cells into different subpopulations. For each subpopulation, Marker Genes were identified by the FindMarkers function (fold change > 2). Finally, to endow these computationally defined cell subsets with distinct biological identities, we performed cell type annotation by comparing their signature genes with known cell markers in CellMarker2.0 and PanglaoDB databases.

Cell communication analysis

To elucidate intercellular signaling dynamics, we employed CellChat for comprehensive cell–cell communication analysis. A communication network was first constructed based on annotated cell types and corresponding gene expression profiles. Significant intercellular interactions were identified by comparing the expression probabilities of ligand–receptor pairs against randomized null distributions. Subsequently, functional enrichment analysis was conducted on the reconstructed network to delineate active signaling pathways, and the spatial organization of key ligand–receptor interactions was visualized to highlight their potential biological relevance.

Weighted gene co-expression network analysis (WGCNA)

TCGA colorectal cancer transcriptomic data were preprocessed by retaining genes expressed in more than 90% of samples, followed by log2 transformation. Weighted Gene Co-expression Network Analysis (WGCNA, version 1.72) was then applied to construct a weighted co-expression network, selecting a soft-thresholding power (β = 6) to achieve scale-free topology (R² > 0.85). Co-expression modules were defined using the dynamic tree-cutting algorithm, and correlations between module eigengenes and a curated ferroptosis-related gene set (sourced from the FerrDb database) were computed. Modules exhibiting significant correlations (p < 0.05) were retained for downstream analyses, including exploration of their biological functions and underlying regulatory networks.

Differential expression analysis

Differentially expressed genes (DEGs) from T cell subpopulations were first extracted from the single-cell dataset (log2 fold change > 1, p < 0.05) and merged to establish a T cell–associated gene set. RNA-seq data from tumor and matched normal tissues were then retrieved from the TCGA colorectal cancer (COAD/READ) dataset. DEGs were identified using DESeq2 (|log2 fold change| > 1, FDR < 0.05), yielding the TCGA-derived DEG set (TCGA_DEGs). The intersection between the T cell–related gene set and TCGA_DEGs was subsequently subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses to infer potential biological functions and pathways¹⁶.

Machine learning feature selection and model construction

The TCGA-CRC tumor samples were stratified and randomly divided at a ratio of 7:3 (stratified by tumor stage and survival status). 114 cases were used in the training set for model construction, and 49 cases were used in the test set for performance verification. GSE60331 (n = 145) was used as an external independent test set. To identify the key prognostic signature genes, multiple machine learning algorithms were used for feature screening. For single-cell data, the gene expression data of “tumor cells + immune cells” were extracted from the integrated expression matrix of Seurat, and the average expression quantity was calculated according to cell type (pseudo-batch expression matrix of “sample-gene” was obtained), which was converted to FPKM format and unified with TCGA batch data quantification standard. The ComBat algorithm (sva package v3.46.0) was used for cross-platform batch effect correction, which was suitable for eliminating systematic bias of data from different sequencing platforms and different laboratories and could retain biological signals. The “data source” was set as the Batch factor (Batch = 1 represents single-cell pseudo-batch data, Batch = 2 represents TCGA batch data), and “tumor stage” and “sample type (primary/metastatic)” were set as covariates to exclude the interference of clinical phenotype differences on the correction results. Call the sva::ComBat function, matrix(~ as.factor(Stage) + as.factor(SampleType)) (include covariates), par.prior = TRUE (use the prior distribution of parameters to improve the stability of correction), mean.only = FALSE (simultaneously correct both) Batch effect at the level of value and variance). The expression matrix was normalized and the low-variance genes were removed, and then the LASSO regression (glmnet package)¹⁷ was used to obtain the non-zero coefficient gene corresponding to the λ value at the maximum average AUC through 10-fold cross-validation and repeated 10 times. At the same time, the randomForest package¹⁸ was used to optimize n_estimators = 500 and max_depth = 10 by 5-fold cross validation to calculate the importance score of genes, and the top 50 important genes were selected. Finally, the intersection of the genes screened by the two methods was used as the final feature gene set, and the support vector machine (SVM) and logistic regression prediction models were constructed. In order to further improve the robustness of PMAIP1 prognostic model, according to Zhang et al.¹⁹ ‘selection of optimal model by combination of multiple algorithms’ in LUAD, CoxBoost + SuperPC combination algorithm was additionally included on the basis of LASSO + random forest selection in this study for model comparison. Specific optimization: 1) Variables were screened by CoxBoost algorithm (penalty parameter was optimized by optimCoxBoostPenalty, maxstepno = 500, and the optimal number of steps was determined by 10-fold cross validation); 2) The selected signature genes were input into SuperPC to construct a survival model (n.components = 3,10 fold cross-validation to determine the threshold); 3) evaluate the performance of different models in TCGA-CRC and GSE60331 cohorts by concordance index (C-index); The results showed that there was no significant difference in C-index between ‘LASSO + random forest’ and ‘CoxBoost + SuperPC’ (0.78 vs 0.76, respectively), which confirmed that the PMAIP1 feature genes screened in this study had model independence. Consistent with the conclusion of Zhang et al.¹⁹ ‘multi-algorithm validation improves the reliability of biomarkers’. Model performance was evaluated by receiver operating characteristic curve (ROC) and area under the curve (AUC).

Functional enrichment analysis

Functional annotation and pathway enrichment analyses were performed independently on the filtered differentially expressed genes and the feature genes. We employed the clusterProfiler package for Gene Ontology (GO) analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis, with a significance threshold set at an adjusted p-value (FDR) < 0.05. Simultaneously, Gene Set Enrichment Analysis (GSEA) was performed to assess the enrichment patterns of predefined gene sets in specific phenotypic contexts, using the hallmark gene sets from the Molecular Signatures Database (MSigDB) as references. The enrichment results were visualized using the ggplot2 package.

Tumor mutational burden analysis

Somatic mutation data, curated in the Mutation Annotation Format (MAF), were procured from The Cancer Genome Atlas (TCGA) database. The mutational landscape was subsequently characterized and visualized utilizing the maftools R package²⁰. For each specimen, the tumor mutational burden (TMB) was quantified, defined as the total count of non-synonymous mutations per megabase of the genome. Based on the median TMB value, samples were stratified into high-TMB and low-TMB cohorts. A comparative analysis was then conducted to elucidate disparities in clinical characteristics and predictive biomarkers of immunotherapy response, such as PD-L1 expression levels, between these two cohorts. Furthermore, mutation signature profiles were deconvoluted using Non-negative Matrix Factorization (NMF) to discern characteristic mutational patterns. These patterns were subsequently benchmarked against established signatures within the Catalogue of Somatic Mutations in Cancer (COSMIC) database.

Immune microenvironment analysis

The CIBERSORTx algorithm was employed to deconvolve the relative proportions of 22 distinct immune cell subtypes within the tumor milieu²¹. Concurrently, the ESTIMATE algorithm was utilized to compute the immune and stromal scores for each sample. To investigate the interplay between feature gene expression and immune cell infiltration, a Spearman correlation analysis was performed. Our immune checkpoint analysis focused on characterizing the expression profiles of pivotal checkpoint molecules, including PD-1 (CD279), PD-L1 (CD274), and PD-L2, within tumor tissues. The Tumor Immune Estimation Resource 2.0 (TIMER 2.0) database was further interrogated to assess the correlations between immune checkpoint expression and the extent of immune cell infiltration. All statistical computations were executed within the R programming environment, with the threshold for statistical significance established at p < 0.05.

Real-time quantitative PCR validation of core genes

To corroborate the expression patterns of designated core genes, we procured paired tumor and adjacent non-tumorous tissues from a cohort of 19 colorectal cancer patients, following the acquisition of informed consent. All methods were performed in accordance with the relevant guidelines and regulations. Oligonucleotide primers were designed using the Primer3Plus tool and subsequently validated for specificity via the NCBI database; primer sequences are enumerated in Table 1. Total RNA was isolated from the tissue specimens and reverse-transcribed into complementary DNA (cDNA). Real-time quantitative PCR (RT-qPCR) was then performed in strict accordance with the manufacturer’s protocol. The relative gene expression levels were quantified using the 2^−ΔΔCtmethod to ensure the fidelity and reproducibility of the experimental data.

Table 1.

Primer sequences for real-time quantitative PCR.

Gene	Primer (5’−3’)
PMAIP1	F: CGGAGCTGGAAGTCGAGTGT	R: GGTTCCTGAGCAGAAGAGTTTGG
GAPDH	F: GGAAGCTTGTCATCAATGGAAATC	R: TGATGACCCTTTTGGCTCCC

Statistical analysis

The Shapiro-Wilk test was used to assess the normality of distribution for all continuous variables. Data adhering to a Gaussian distribution were analyzed using either an independent t-test or one-way analysis of variance (ANOVA), whereas non-normally distributed data were evaluated using the Mann-Whitney U test or the Kruskal-Wallis test. Categorical variables were compared using the Pearson’s Chi-square test or Fisher’s exact test, as appropriate. Survival curves were generated via the Kaplan-Meier method. All statistical tests were two-sided, and a p-value less than 0.05 was considered statistically significant. To control the false discovery rate in multiple comparisons, the Benjamini-Hochberg procedure was applied. All data analyses were conducted using R software (version 4.4.1) and GraphPad Prism (version 10.1.2).

Results

Single-cell transcriptomic atlas delineates the heterogeneity of the immune microenvironment in primary and liver metastatic sites of colorectal cancer

To systematically dissect the differences at the cellular ecosystem level between primary colorectal cancer (CRC) lesions and their liver metastases (CRC-LM), we conducted an in-depth analysis of single-cell transcriptomic data from paired samples. Using t-distributed Stochastic Neighbor Embedding (t-SNE) for dimensionality reduction, we mapped the cells into a two-dimensional space and performed unsupervised clustering based on their transcriptional profiles, ultimately identifying 19 distinct cell subpopulations. This clustering result clearly illustrated significant differences in cellular composition between the primary and metastatic sites (Fig. 1A).

Fig. 1.

Single-Cell Transcriptomic Landscape of Colorectal Cancer and Its Liver Metastases (CRC-LM). (A-B) t-SNE plots displaying the clustering (A) and annotated identities (B) of cells from CRC and CRC-LM tissues, with each cluster marked by color. (C) Dot plot illustrating the expression of canonical marker genes in identified cell types. Dot size and color represent the percentage of expressing cells and average expression level, respectively. (D) Heatmap of marker gene expression across clusters, validating cell type annotations. (E) Bar graph depicting the relative proportions of each cell type in CRC versus CRC-LM.Data were expressed as Mean ± standard deviation (Mean ± SD), and the sample size of each group was n = 15 (15 patients were paired with primary and liver metastases). The difference of cell proportion between the two groups was tested by Mann-Whitney U test (p < 0.05 by Shapiro-Wilk test because the cell proportion data did not meet the normal distribution). The significance threshold was set at α = 0.05, and the labeling was *p < 0.05, **p < 0.01, ***p < 0.001. Among them, CD8⁺ effector memory T cells (***p = 0.0008) and inflammatory monocytes (**p = 0.003) were significantly enriched in liver metastases, and malignant epithelial cells were significantly enriched in primary lesions (**p = 0.004). (F) Stacked bar graph showing the distribution of cell types in individual patient samples.

To assign precise biological identities to these cell subpopulations, we performed meticulous cell-type annotation using canonical marker genes of cell lineages (Canonical Marker Genes). This process successfully identified several key immune and stromal cell populations, including CD8⁺ effector memory T cells, regulatory T cells (Tregs), inflammatory monocytes, cancer-associated fibroblasts (CAFs), and malignant epithelial cells (Fig. 1B). Visualization analyses utilizing dot plots (Dot Plot) and heatmaps (Heatmap) further corroborated the specificity and fidelity of marker gene expression in each cell subpopulation, thereby validating the accuracy of the cell-type annotations (Figs. 1C–D).

A comparative analysis of cellular compositions from the two tissue origins revealed profound remodeling of the immune microenvironment during the metastatic process. Specifically, CD8⁺ effector memory T cells with anti-tumor potential and inflammatory monocytes exhibited substantial enrichment in liver metastases; in contrast, malignant epithelial cells and certain immunosuppressive cell populations predominated in primary sites (Fig. 1E). These findings indicate that during liver metastasis, the tumor microenvironment undergoes a pivotal immune state transition, shifting from a predominantly immunosuppressive state to one characterized by pro-inflammatory and stress responses. Furthermore, it is noteworthy that our analysis also revealed significant inter-patient heterogeneity (Inter-patient Heterogeneity), as evidenced by notable fluctuations in the relative abundances of various immune cells, stromal cells, and malignant cells across samples from different patients (GSM samples) (Fig. 1F).

Upon comparing the cellular composition differences between CRC and CRC-LM samples, CD8⁺ effector memory T cells and inflammatory monocytes were found to be significantly enriched in liver metastases, whereas malignant epithelial cells and certain immunosuppressive cell types were more prevalent in primary sites (Fig. 1E). This suggests that during liver metastasis, the tumor microenvironment undergoes immune remodeling, with the immune response state progressively shifting from a suppressive to an inflammatory and stress-responsive state. Moreover, we observed considerable heterogeneity in cellular composition among different individuals (GSM samples), manifested as varying degrees of fluctuation in the proportions of immune cells, stromal cells, and malignant cells across patients (Fig. 1F).

Cell communication analysis reveals remodeling of the immune regulatory network and the critical role of core signaling pathways

To elucidate the functional interplay within the tumor microenvironment (TME), we applied the CellChat algorithm to systematically quantify intercellular communication networks in primary colorectal cancer (CRC) lesions and their liver metastases. Global network analysis revealed that the evolutionary progression from primary to metastatic stages is accompanied by extensive remodeling of both the number and strength of intercellular interactions, resulting in distinctly altered communication topologies (Fig. 2A). Within this intricate network, inflammatory tumor-associated macrophages (TAMs), cancer-associated fibroblasts (CAFs), CD8⁺ effector memory T cells, and malignant epithelial cells collectively constitute critical hubs of cellular information exchange, with perturbations in their communication patterns driving the overall functional polarization of the TME.

Fig. 2.

Systematic Analysis of Cell Communication Patterns in Colorectal Cancer Tissues. (A) Cell-cell communication network diagrams, with the left panel showing the number of cell interactions and the right panel showing interaction strength; chord diagrams illustrate the communication links between different cell subpopulations. (B–F) Heatmaps of signal transmission for the CD99, MIF, GALECTIN, MHC-I, and MHC-II pathways, displaying communication strengths between signal senders (vertical axis) and receivers (horizontal axis). The bar graphs on the right show the total communication probabilities for different cell types. (G) Heatmap of outgoing and incoming signaling patterns for all cell types, with shades of green indicating differences in signaling activity. (H) Analysis of outgoing communication contributions for each cell type, where bubble size represents the contribution value and color indicates the signaling pathway type. (I) Distribution of communication probabilities and significance for key ligand-receptor pairs across cells, where dot size represents p-value significance and color indicates communication probability.

Our analysis uncovered several signaling pathways that occupy central regulatory positions within the TME. Among them, the CD99, MIF, GALECTIN, MHC-I, and MHC-II signaling pathways exhibited robust activation across diverse cell subpopulations, orchestrating essential biological processes such as cell adhesion, chemotactic recruitment, antigen presentation, and immune evasion (Figs. 2B–F). Notably, the MIF signaling axis emerged as a pivotal mediator of communication between CAFs and immunosuppressive cells (e.g., regulatory T cells), suggesting its function as a key molecular conduit in the establishment of an immunosuppressive microenvironment. In contrast, the CD99 pathway demonstrated significant enrichment within interaction networks centered on CD8⁺ T cells, implying a potential role in modulating T cell infiltration into tumor tissues and facilitating cytotoxic immune responses.

Deconstructing the directionality of signal flow revealed that CAFs and inflammatory TAMs act as major signal “senders”, functioning as active initiators across multiple signaling cascades, whereas CD8⁺ effector T cells, plasma cells, and select dendritic cell subsets primarily serve as signal “receivers” (Figs. 2G–H). Bubble plots further visualized and quantified the dominant signaling pathways engaged by each cell population and their respective communication probabilities (Figs. 2H–I). Among these, ligand–receptor pairs such as MIF–CD74, LGALS9–CD44, and CD99–CD99 established high-confidence interaction axes between CAFs and immune cells. The signal exchanges mediated through these molecular pairs may represent critical regulatory nodes that govern the directionality of local immune responses.

Analyses presented in Supplementary Figure SF1 further refined our understanding of cellular functional roles. Quantitative dissection of the Sender, Receiver, Mediator, and Influencer functions across five representative signaling pathways (Figure SF1A) confirmed that CAFs assume pivotal integrative and transductive roles across multiple signaling contexts. Particularly noteworthy is the remarkable plasticity and complexity exhibited by distinct cell subpopulations in their functional output patterns across signaling pathways (Figures SF1B–C). For instance, malignant epithelial cells concurrently display high signal-emitting capacity and influence within the CD99 and MHC-I pathways, suggesting a bidirectional adaptive mechanism whereby tumor cells actively induce immune tolerance while simultaneously responding to immune surveillance pressures.

WGCNA analysis reveals apoptosis-related gene modules associated with tumor clinical traits

Through WGCNA network construction and module analysis, we identified several apoptosis-related gene modules and examined their correlations with clinical traits (Fig. 3A and D). A total of 8 modules were detected, among which the black module (MEblack) exhibited a significant negative correlation with tumor stage (r = −0.24, p = 0.002), suggesting that genes within this module may be involved in tumor suppression or inhibition of metastasis. Additionally, the yellow module (MEyellow) displayed a moderate negative correlation with overall survival (r = −0.22, p = 0.006) (Fig. 3E). The hub genes in these key modules will serve as the focal point for subsequent analyses, with a focus on exploring their regulatory networks and potential clinical translational value.

Fig. 3.

WGCNA Analysis Reveals Apoptosis-Related Gene Modules Associated with Tumor Clinical Traits. (A) Sample clustering dendrogram for identifying outlier samples. (B) Soft-thresholding power selection plot for network construction. The left panel displays the scale-free topology fit index (R²) under different soft-threshold values, while the right panel shows the mean connectivity trend. A soft-threshold of 9 was selected to ensure the network approximates a scale-free distribution. (C) Gene clustering dendrogram with modules obtained from dynamic tree cutting, where colors represent different modules. (D) Sample clustering heatmap integrated with phenotypic traits (such as gender, age, stage, etc.). (E) Heatmap of correlations between modules and clinical traits, with colors representing Pearson correlation coefficients and p-values in parentheses. Among them, the black module (MEblack) is significantly correlated with tumor stage (r = −0.24, p = 0.002).

Differential expression analysis and functional enrichment analysis

To systematically delineate the transcriptomic reprogramming events driving the initiation and progression of colorectal cancer (CRC), particularly focusing on the molecular basis underlying apoptosis regulation dysregulation, we conducted a differential expression analysis of tumor tissues and adjacent non-tumor tissues. By applying stringent statistical thresholds, we identified a set of differentially expressed genes (DEGs) that exhibited statistically significant changes in expression levels. The volcano plot (Fig. 4A) clearly illustrates a well-defined landscape of transcriptional dysregulation, comprising significantly upregulated (green) and downregulated (orange) genes. Within this landscape, key regulatory factors closely linked to apoptosis, proliferation, and immune responses—such as FOXO1, ETV4, and CDH3—displayed marked deviations in expression levels.

Fig. 4.

Differential Expression Analysis and Functional Enrichment Analysis. (A) Volcano plot illustrating the distribution of differentially expressed genes (DEGs), with the horizontal axis representing log2 FoldChange and the vertical axis representing -log10(FDR). Green indicates upregulated genes, orange indicates downregulated genes, and gray indicates genes with no significant differences. Significant genes such as CDH3, ETV4, FOXO1, etc., are labeled in the plot. (B) Heatmap displaying the expression levels of the top 100 differentially expressed genes in normal tissues (N) and tumor tissues (T). Colors ranging from red (high expression) to blue (low expression) represent changes in expression levels. (C) Gene Ontology (GO) enrichment analysis results, showing the top 30 biological process (BP) terms significantly associated with the differentially expressed genes, primarily involving cell differentiation, T cell differentiation, regulation of cell adhesion, etc. (D) KEGG enrichment analysis results, displaying the top 20 significant pathways, including the “MAPK signaling pathway,” “T cell receptor signaling pathway,” and pathways related to viral infections and cancer, etc. The pathways were generated using KEGG Mapper (https://www.kegg.jp/kegg/tool/map_pathway.html) and are reproduced with permission from Kanehisa Laboratories ©2025.

Further clustering analysis of the DEGs was performed, with a heatmap (Fig. 4B) generated to visualize the expression patterns of the top 100 DEGs in tumor (T) and normal (N) samples, revealing a clear stratification of gene expression between the two categories. To uncover the potential biological functions and signaling pathways associated with these DEGs, we carried out Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. The GO analysis (Fig. 4C) revealed that these genes are predominantly enriched in immune-related and cell fate-related processes, such as “monocyte differentiation,” “T cell differentiation,” and “regulation of cell adhesion.” The KEGG analysis (Fig. 4D) further identified their involvement in several disease-related signaling pathways, including the “MAPK signaling pathway,” “Toll-like receptor signaling pathway,” “virus infection-related pathways,” and “Apoptosis,” suggesting that these DEGs may play significant roles in the tumor immune microenvironment and apoptosis regulation.

Machine learning-based feature gene selection

Building upon the differentially expressed genes identified, we further conducted an intersection analysis to screen for key apoptosis-related genes closely associated with tumor progression. This analysis combined the apoptosis gene set, T cell-related gene set, and differentially expressed genes (DEGs). The Venn diagram (Fig. 5A) shows that a total of 19 intersection genes were identified, which may represent key candidates for cross-regulation between tumor immunity and cell apoptosis.

Fig. 5.

Machine Learning-Based Feature Gene Selection. (A) Venn diagram of the intersection among three gene sets (apoptosis genes, T cell-related genes, and DEGs), identifying a total of 19 intersection genes, suggesting these genes may play key roles in immune-related apoptosis mechanisms. (B) Importance (Gain) of each gene in the XGBoost model, with HSPA1B, PRELID1, PMAIP1, etc., as the highest-weighted feature genes. (C) Boruta feature selection results, displayed as a boxplot showing the importance ranking of different feature genes, with green bars representing features considered important. (D) Variable importance scores in the Random Forest (RF) model; the left side shows the error change curve with increasing trees, and the right side shows the importance values for each gene, with red labels indicating lower importance. (E) Venn diagram of the intersections among three feature selection methods (XGBoost, RF, Boruta), ultimately identifying 4 overlapping genes recognized as key features across all models.

Subsequently, we constructed three machine learning models—XGBoost, Random Forest (RF), and Boruta—to select the feature genes with the strongest predictive power. In the XGBoost model, feature importance analysis (Fig. 5B) indicated that genes such as HSPA1B, PRELID1, PMAIP1, and TNF made significant contributions. The Boruta algorithm further identified a set of stable and important features (Fig. 5C), showing a high degree of overlap with the XGBoost results.

In the RF model, we evaluated variable importance and plotted error curves (Fig. 5D), with the results similarly supporting the stable performance of multiple genes in classification tasks. Finally, the results from the three algorithms were integrated, and the Venn diagram (Fig. 5E) revealed two shared feature genes (PMAIP1 and PRELID1) that were identified across all models, suggesting their potential diagnostic or predictive value in tumor staging or the immune microenvironment.

Expression, prognostic, and immune correlation analysis of the key gene PMAIP1 in tumors

To determine the candidate gene with the greatest biological and clinical research value among the intersecting feature genes, we further conducted a systematic multidimensional comparison between PMAIP1 and PRELID1. First, at the expression level, both genes were significantly upregulated in tumor tissues (Fig. 6A), with the differences being statistically significant. Subsequent survival analysis in the TCGA-READ cohort revealed that patients with high PMAIP1 expression had a significantly shorter overall survival (HR = 0.39, p = 0.028), whereas PRELID1 did not demonstrate significant prognostic value (p = 0.168) (Figs. 6B-C).

Fig. 6.

Comparative Analysis of Key Feature Genes PMAIP1 and PRELID1 in Tumor Expression, Prognosis, and Immune Correlations. (A) Comparison of the expression of PMAIP1 and PRELID1 in normal tissues (n = 10, TCGA-CRC adjacent normal tissues) and tumor tissues (n = 163, TCGA-CRC tumor tissues) (expressed as log₂(FPKM + 1) standardized value, Mean ± SE). The expression differences between the two groups were analyzed by independent sample t test (Shapiro-Wilk test, gene expression data were in normal distribution, p > 0.1; Levene test, homogeneity of variance, p > 0.05), and the significance threshold α = 0.05. The results showed that PMAIP1 (***p = 2.3 × 10⁻¹⁵) and PRELID1 (**p = 4.7 × 10⁻ ⁸) were both significantly up-regulated in tumor tissues. (B-C) Kaplan-Meier survival curve analysis of the two genes’ survival predictive ability in the TCGA-READ cohort. The high-expression PMAIP1 group shows significantly poorer overall survival (p = 0.028), while PRELID1 exhibits no statistical difference (p = 0.168). (D) Expression patterns of PMAIP1 and PRELID1 (expressed as log₂(TPM + 1) normalized value, mean ± SE) in the external validation cohort GSE60331. Sample size: normal tissue n = 33, CRC tumor tissue n = 112. Mann-Whitney U test was used to compare the differences between the two groups (the gene expression data in GSE60331 did not meet the normal distribution by Shapiro-Wilk test, p < 0.05), and the significance threshold α = 0.05. The results showed that PMAIP1 (***p = 6.3 × 10⁻⁵) and PRELID1 (**p = 0.002) were highly expressed in tumor tissues, which was consistent with the trend of the TCGA-CRC cohort. The differential expression of PMAIP1 was further validated in three extended GEO cohorts, and the results showed that its high expression signature was consistent across platforms and clinical backgrounds: GSE39582 (n = 585) : The expression of PMAIP1 in tumor tissues was significantly higher than that in normal tissues (log₂FC = 1.83, p = 3.2 × 10⁻¹² by Mann-Whitney U test), and the expression level of PMAIP1 in stage IV patients was significantly higher than that in stage I patients (p = 0.004), suggesting that its expression is related to tumor progression. GSE17536 (n = 188) : The expression of PMAIP1 in tumor tissues was 1.57 times higher than that in normal tissues (log₂FC = 1.57, p = 8.9 × 10⁻⁸ by independent sample t test). The expression of PMAIP1 in patients with poor response to chemotherapy (mean = 4.26) was significantly higher than that in patients with good response (mean = 2.91, p = 0.021). GSE29621 (n = 164, early-stage CRC) : Even in stage I-II patients, PMAIP1 expression in tumor tissues was still significantly higher than that in normal tissues (log₂FC = 1.32, p = 1.5 × 10⁻⁶ by independent sample t test). These findings are consistent with those of the TCGA-CRC (p = 2.3 × 10⁻¹⁵) and GSE60331 (p = 6.3 × 10⁻ ⁵) cohorts.” (E-F) ROC curves evaluating the discriminative ability of PMAIP1 and PRELID1 for tumor diagnosis. In TCGA-READ, PMAIP1 AUC = 0.870, PRELID1 AUC = 0.749; in GSE60331, the AUCs are 0.823 and 0.810, respectively, indicating superior diagnostic value for PMAIP1. (G-H) Spearman correlation analysis between gene expression and immune cell infiltration levels. PMAIP1 shows positive correlations with T cells, helper T cells (Th2, Th17), dendritic cells, macrophages, etc., and negative correlations with NK CD56bright cells, regulatory T cells (Treg), etc.; PRELID1 also exhibits significant negative correlations with multiple immune cell types.

In the external validation dataset GSE60331, the expression trends of the two genes were reproducibly confirmed, further enhancing the reliability of the results (Fig. 6D). ROC curves further compared the diagnostic performance of the two genes, with PMAIP1 exhibiting higher AUC values in both independent cohorts compared to PRELID1 (Figs. 6E-F), suggesting its superior discriminative ability for tumors.

Additionally, we explored the relationships between the genes and the immune microenvironment. PMAIP1 demonstrated broader and more significant correlations with immune cell infiltration, particularly positive correlations with various T cell subsets, dendritic cells, and macrophages (Fig. 6G), whereas PRELID1 was primarily negatively correlated with immunosuppressive immune cells (Fig. 6H). In summary, PMAIP1 outperformed PRELID1 in terms of expression characteristics, prognostic prediction, immune correlations, and diagnostic capabilities, and was therefore selected as the core gene for subsequent in-depth mechanistic studies and model construction.

Pan-cancer expression profile, functional annotation, and clinical prognostic model construction for the PMAIP1 gene

To further elucidate the oncological significance of PMAIP1, we conducted a comprehensive bioinformatics analysis. In the TCGA pan-cancer expression profile (Fig. 7A), PMAIP1 exhibited elevated expression across a variety of cancers, with particularly significant differences observed in colorectal cancer (COAD, READ), liver cancer (LIHC), and stomach cancer (STAD) (p < 0.001). These findings suggest that PMAIP1 may play a broad, tumor-promoting role.

Fig. 7.

Pan-Cancer Expression Profile, Functional Annotation, and Clinical Prognostic Model Construction for the PMAIP1 Gene. (A) Expression levels of PMAIP1 across multiple tumor types in TCGA. Compared to normal tissues, PMAIP1 is significantly upregulated in various cancers (such as COAD, READ, LIHC, STAD, etc.) (***p < 0.001, **p < 0.01, *p < 0.05). (B) GO enrichment analysis results show that PMAIP1 co-expressed genes are primarily involved in biological processes such as “secretory vesicles,” “cell chemotaxis,” “leukocyte migration,” and “response to external stimulus.” (C) KEGG pathway enrichment analysis reveals significant enrichment of PMAIP1-related genes in immune- and inflammation-related signaling pathways, including “cytokine-cytokine receptor interaction,” “chemokine signaling pathway,” “virus infection-related pathways,” and “IL-17 signaling pathway.” The pathways were generated using KEGG Mapper (https://www.kegg.jp/kegg/tool/map_pathway.html) and are reproduced with permission from Kanehisa Laboratories ©2025. (D) Survival prognostic nomogram constructed based on PMAIP1 and clinical variables, demonstrating good potential in predicting 1-year, 3-year, and 5-year overall survival rates.

Subsequently, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses on the co-expressed genes of PMAIP1. The GO analysis (Fig. 7B) revealed that these genes are involved in several biological processes related to inflammatory responses, extracellular secretion, and immune migration, such as “secretory vesicles,” “leukocyte migration,” and “chemotactic activity.” The KEGG pathway analysis (Fig. 7C) further emphasized that PMAIP1-related pathways are predominantly involved in key inflammatory and immune pathways, including “cytokine-cytokine receptor interaction,” “chemokine signaling pathway,” and “IL-17 signaling pathway.” These results suggest that PMAIP1 may play a pivotal role in regulating the tumor immune microenvironment.

For clinical application, we constructed a survival prediction nomogram using PMAIP1 expression along with clinicopathological variables (T stage, N stage, clinical stage, gender) (Fig. 7D). The results indicated that this model can accurately predict patients’ 1-year, 3-year, and 5-year survival probabilities, highlighting its potential for clinical use.

In summary, PMAIP1 is not only highly expressed in multiple cancer types but is also strongly associated with several immune-related signaling pathways. Combined with its prognostic value in survival models, this further underscores its biological significance as a potential diagnostic and therapeutic target.

Immune infiltration analysis

Regarding the immune microenvironment, CIBERSORT analysis revealed that the PMAIP1 high-expression group is closely associated with the infiltration of multiple immune cell types (Figs. 8A-B). Further correlation analysis (Figs. 8C-N) demonstrated that PMAIP1 expression levels are positively correlated with immune-activating cells, such as CD8⁺ T cells, central memory T cells (Tcm), Th1 cells, and Th2 cells, while negatively correlating with certain NK cell subsets and γδ T cells. Additionally, PMAIP1 expression is significantly associated with tumor immune microenvironment scores (ImmuneScore and StromalScore) (Fig. 9A), further suggesting its close relationship with immune cell infiltration. Immune checkpoint analysis (Figs. 9B-D) further revealed that PMAIP1 expression is significantly positively correlated with multiple immune checkpoint molecules, including PD-1, PD-L1, and PD-L2. These findings were validated in independent cohorts, indicating that PMAIP1 may play a critical role in the tumor immune escape process.

Fig. 8.

Relationship Between PMAIP1 Expression and Immune Cell Infiltration. (A) Estimated proportions of immune cell compositions using the CIBERSORT algorithm, with different colors representing different immune cell types. (B) Differential analysis of immune cell distributions between high- and low-expression PMAIP1 groups.Differences in the distribution of immune cells (expressed as the proportion of immune cells estimated by CIBERSORTx, mean ± SD) between the high PMAIP1 expression group (n = 82, PMAIP1 expression above the median in TCGA-CRC tumor samples) and the low PMAIP1 expression group (n = 81, PMAIP1 expression below the median). The difference of the proportion of immune cells between the two groups was analyzed by Mann-Whitney U test (the proportion of immune cells showed skewed distribution, p < 0.05 by Shapiro-Wilk test), with the significance threshold α = 0.05, marked *p < 0.05, **p < 0.01. CD8⁺T cells (*p = 0.032) and Th1 cells (*p = 0.041) were significantly enriched in the high expression group, and NK CD56bright cells were significantly enriched in the low expression group (**p = 0.007). (C-N) Spearman correlation analysis between PMAIP1 expression levels and infiltration degrees of different immune cell subsets, including CD8⁺ T cells, T helper cells, NK cells, macrophages, neutrophils, etc., with significance annotated by p-values.

Fig. 9.

Associations Between PMAIP1 and Immune Checkpoint Expression. (A) Comparative differences in immune microenvironment scores (StromalScore, ImmuneScore, ESTIMATEScore) between high- and low-expression PMAIP1 groups. (B) Expression differences of immune checkpoints PD-1, PD-L1, and PD-L2 in high- and low-expression PMAIP1 groups.Differences in the expression of immune checkpoint molecules PD-1 (CD279), PD-L1 (CD274) and PD-L2 (PDCD1LG2) between the high PMAIP1 group (n = 82) and the low PMAIP1 group (n = 81) (TCGA-CRC data, expressed as log₂(FPKM + 1); Mean ± SE). Independent sample t test was used to compare the differences between the two groups (Shapiro-Wilk test, normal distribution, p > 0.05; Levene test, homogeneity of variance, p > 0.05), and the significance threshold α = 0.05. The results showed that PD-1 (*p = 0.028), PD-L1 (**p = 0.0039) and PD-L2 (*p = 0.04) were significantly up-regulated in the high PMAIP1 group. (C) Expression changes of multiple immune checkpoints (such as VTCN1, TNFRSF4, SIGLEC15, PTPRC, etc.) in high- and low-expression PMAIP1 groups. (D) Validation of immune checkpoint gene expression differences in high- and low-expression PMAIP1 samples from the independent dataset GSE60331.

PMAIP1-related gene mutation landscape reveals its potential role in tumor progression

To explore the relationship between PMAIP1-associated risk groupings and tumor mutation characteristics, we performed a whole-genome mutation profile analysis on high- and low-risk groups (based on PMAIP1 expression). As shown in Fig. 10A, significant co-mutation or mutual exclusivity relationships were observed among the top 30 genes with the highest mutation frequencies. For example, genes such as TP53 and APC exhibit significant mutual exclusivity, while TTN and MUC16 show significant co-mutation, suggesting potential synergistic or alternative roles in tumor initiation and progression.

Fig. 10.

Tumor Mutation Feature Analysis of PMAIP1-Related Risk Groupings. (A) Heatmap of co-mutation and mutual exclusivity relationships among the top 30 high-frequency mutated genes, with color intensity representing -log10(p-value) significance levels, and asterisks indicating statistical significance (P < 0.05, P < 0.01). (B) Mutation landscape of the high PMAIP1 expression (high-risk) group, displaying mutation types (missense mutations, nonsense mutations, frameshifts, etc.) and distributions of clinical features (gender, survival status, stage, etc.). (C) Mutation landscape of the low PMAIP1 expression (low-risk) group, similarly annotated with mutation types and their relationships to clinical features.

Further waterfall plot analysis results (Figs. 10B-C) indicate that in the high PMAIP1 expression (high-risk) group, genes such as TP53, APC, and TTN have higher mutation frequencies, with missense mutations and frameshift mutations being the predominant mutation types. In the low-risk group, although high-frequency mutated genes are also present, their mutation frequencies and distributions differ. Additionally, the high-risk group is more concentrated in clinical features such as stage IV, mortality, and male patients, suggesting that high PMAIP1 expression may be closely associated with poor prognosis and is linked to a distinct genomic mutation landscape.

Validation of core gene expression in clinical samples

To further validate the expression levels of the core gene, we employed RT-qPCR to detect samples from clinical tissues of colorectal cancer patients. The expression of the PMAIP1 gene was significantly higher in tumor tissues (Fig. 11A and B), consistent with the bioinformatics analysis results.

Fig. 11.

RT-qPCR was used to verify the expression of PMAIP1 in clinical CRC tissues. (A) The relative expression of PMAIP1 in cancer tissues (T) and corresponding adjacent normal tissues (N) of 19 CRC patients (mean ± SD, calculated by 2⁻ΔΔCt method with GAPDH as the reference gene); (B) Overall difference in PMAIP1 expression among the 19 paired samples. Paired-samples t test was used to compare the expression differences between the two groups (Shapiro-Wilk test p = 0.12), and the significance threshold α = 0.0001 (****p < 0.0001). The results confirmed that PMAIP1 was significantly up-regulated in cancer tissues (3.2-fold on average), which was consistent with the bioinformatics analysis.

Discussion

Liver metastasis in colorectal cancer (CRC) is a critical factor contributing to its high mortality rate, with the dynamic remodeling of the tumor microenvironment (TME) playing a central role in this process²². This study, through the systematic integration of single-cell and bulk RNA sequencing data, combined with machine learning algorithms and clinical validation, not only delineates the significant heterogeneity in the immune microenvironment between primary CRC sites and liver metastases but also identifies and validates the pro-apoptotic gene PMAIP1 as a potential key molecule regulating tumor immune evasion and malignant progression, with its high expression being closely associated with poor prognosis.

The CMS classification of colorectal cancer (CRC) delineates four distinct molecular subtypes based on gene expression profiles, each exhibiting significant differences in signaling pathways, metabolic patterns, and immune microenvironments. This study confirms that PMAIP1 is consistently upregulated in CRC tumor tissues. Notably, evidence from prior research²³ indicates that in the CMS1 subtype, elevated PMAIP1 expression may synergize with immunotherapeutic interventions. Furthermore, CMS1 is frequently associated with BRAF mutations, and metabolic dysregulation induced by these alterations may promote PMAIP1 upregulation via cellular stress response pathways, positioning PMAIP1 as a pivotal molecule linking immune activation and apoptotic signaling in this subtype. Preclinical studies²⁴ demonstrate that PMAIP1 expression in the CMS2 subtype correlates strongly with chemotherapeutic efficacy. Specifically, the topoisomerase I inhibitor irinotecan (CPT-11) significantly enhances PMAIP1 expression in CRC cells, and overexpression of PMAIP1 increases cellular sensitivity to the BCL-2/Bcl-xL inhibitor ABT-737. Given the high expression of epidermal growth factor receptor (EGFR) in CMS2 tumors, combining conventional chemotherapy with strategies that activate the PMAIP1-mediated apoptotic pathway may enhance therapeutic outcomes. In CMS3²⁵, PMAIP1 modulates tumor cell survival through crosstalk between metabolic and apoptotic pathways. Alterations in PMAIP1 expression can influence the stability of key metabolic enzymes, while conversely, disruptions in metabolic flux may regulate PMAIP1 transcription through energy-sensing mechanisms. Thus, PMAIP1 represents a promising target to overcome current therapeutic limitations in CMS3.

As a critical pro-apoptotic effector of the p53 pathway, PMAIP1 (also known as NOXA) exhibits differential prognostic significance across various malignancies, demonstrating cancer type–specific characteristics. High PMAIP1 expression is associated with favorable prognosis in triple-negative breast cancer, non-small cell lung cancer, and hepatocellular carcinoma, where it functions as an independent protective factor^26,27. Conversely, in pancreatic, ovarian, endometrial, and bladder cancers, elevated PMAIP1 levels correlate with poor clinical outcomes, suggesting potential oncogenic roles, and are linked to advanced tumor stage, lymph node metastasis, and distant metastasis^28,29. The prognostic value of PMAIP1 in gastric cancer remains controversial, potentially influenced by molecular subtypes, disease stage, or treatment modalities, warranting further investigation. Therefore, while PMAIP1 holds promise as a prognostic biomarker across multiple cancers, its interpretation must be contextualized within specific tumor types and molecular frameworks. Personalized therapeutic strategies can be developed based on the distinct expression patterns of PMAIP1 across cancer entities.

As a member of the BH3-only pro-apoptotic subfamily of the BCL-2 protein family, PMAIP1 functions as an “apoptosis sensitizer” within the mitochondrial apoptosis regulatory network, being induced through both p53-dependent and p53-independent pathways³⁰. BCL2, an anti-apoptotic protein frequently overexpressed in CRC, contributes to chemoresistance and exhibits a significant inverse correlation with PMAIP1 expression. High BCL2 levels may counteract the pro-apoptotic activity of PMAIP1, representing a key mechanism by which CRC cells evade programmed cell death³¹. PMAIP1 primarily participates in upstream signal integration, neutralization of anti-apoptotic proteins, and high-affinity binding to MCL1, thereby indirectly activating BAX and BAK. In CRC, PMAIP1 acts as a molecular switch for BAX activation by selectively binding and sequestering MCL1, thereby releasing BAX to execute its pro-apoptotic function. Consequently, reduced PMAIP1 expression coupled with impaired BAX activity adversely affects patient prognosis³². Compared to BCL2 and BAX, PMAIP1 serves as a crucial integrator of DNA damage responses, endoplasmic reticulum stress, and mitochondrial apoptosis. Through its specific interaction with MCL1 in CRC, PMAIP1 establishes a distinct regulatory axis independent of BCL2 and BAX, offering novel insights and potential biomarkers for precision oncology approaches.

Our single-cell transcriptome analysis revealed, for the first time, that colorectal cancer (CRC) liver metastases exhibit a distinct “pro-inflammatory” and “immune-activated” tumor microenvironment compared to primary CRC tumors. Specifically, CD8⁺ effector memory T cells and inflammatory monocytes were significantly enriched in liver metastatic lesions. These findings appear to contrast with the classical theory of “immune escape,” yet they may reflect a more complex process of “immune editing” during tumor metastasis³³. Notably, consistent with previous observations demonstrating a negative correlation between NDUFB10—a lactate-related pathway gene—and the spatial distribution of CD8⁺ T cells in lung cancer, Watson et al. have confirmed that lactate accumulation within the tumor microenvironment promotes immunosuppressive conditions³⁴. We hypothesize that CRC cells metastasizing to the liver encounter an immune landscape fundamentally different from that of the primary tumor site. Emerging evidence indicates that the liver-specific microenvironment—such as chemokines secreted by hepatic sinusoidal endothelial cells—selectively recruits myeloid cell populations, potentially accounting for the observed enrichment of inflammatory monocytes^35–37. This recruitment may initially serve to initiate anti-tumor immune responses; however, tumor cells might counteract these defenses through adaptive immune evasion mechanisms. In a manner analogous to the findings by Gemma et al., certain conformational states of PM369 homodimers demonstrate enhanced efficiency in recruiting β-arrestin proteins for cellular signal transduction regulation, likely mediated by “trigger” mechanisms or synergistic molecular interactions³⁸.

Cell communication analysis provides a mechanistic explanation for this. We found that cancer-associated fibroblasts (CAFs) and inflammatory tumor-associated macrophages (TAMs) emerge as central hubs in the signaling network. This aligns with the prevailing view in the field that “CAFs are key modulators of the TME“³⁹. In particular, the activation of signaling axes such as MIF-CD74 and LGALS9-CD44 suggests that CAFs may, on one hand, recruit immune cells through the secretion of cytokines like MIF, while on the other hand, induce their functional exhaustion or polarization toward immunosuppressive phenotypes^40,41. Simultaneously, the activation of the CD99 pathway may promote T cell infiltration; however, whether this infiltration translates into effective cytotoxicity requires further investigation. Therefore, the microenvironment of liver metastases is not simply one of “immune activation,” but rather a state of “ineffective inflammation” characterized by high immune cell infiltration, yet potentially impaired functionality, providing a theoretical foundation for combined immunotherapy strategies (such as PD-1/PD-L1 inhibitors in conjunction with CAF-targeted therapies).

Through multi-layered bioinformatic screening and validation, this study identifies PMAIP1 (also known as NOXA) as the core candidate gene. PMAIP1 is a classic BH3-only pro-apoptotic protein in the Bcl-2 family, typically activated by stress signals such as p53 under cellular stress conditions, inducing the mitochondrial apoptosis pathway by neutralizing the anti-apoptotic protein Mcl-1^42,43. However, our study results present a seemingly paradoxical phenomenon: PMAIP1 is significantly overexpressed in CRC tissues, and its high expression correlates with poorer overall survival.

We propose the following potential explanations to reconcile this apparent contradiction: (1) Functional inactivation or compensatory up-regulation: tumor cells may exhibit high levels of PMAIP1 expression while simultaneously suppressing its pro-apoptotic activity through alternative mechanisms, such as Mcl-1 amplification or alterations in post-translational modifications, resulting in a functionally inert state—what has been referred to as “name only”⁴⁴. The elevated expression of PMAIP1 could represent a stress-induced response under conditions such as hypoxia or chemotherapy, without effectively triggering the apoptotic cascade. (2) Non-apoptotic functions: Accumulating evidence in recent years indicates that PMAIP1 may exert non-classical roles beyond its canonical function in apoptosis. Our functional enrichment analysis revealed that the PMAIP1 co-expression network is significantly enriched in immune-related pathways, suggesting a potential immunomodulatory role. Notably, a recent study demonstrated that PMAIP1 stabilizes PD-L1 expression, thereby promoting tumor immune evasion—a finding highly consistent with our observations^45,46. (3) Induction of immune cell exhaustion: our data show that PMAIP1 expression is significantly positively correlated with multiple immune checkpoint molecules, including PD-1, PD-L1, and PD-L2. This implies that tumors with high PMAIP1 expression may drive infiltrating T cells toward an exhausted phenotype by up-regulating immune checkpoint expression. Previous studies have linked PMAIP1 to T cell exhaustion, with its up-regulation in activated T cells potentially contributing to activation-induced cell death (AICD)^47–49. Mechanistically, PMAIP1 has been shown to activate the p65 subunit of NF-κB via phosphorylation, promote its nuclear translocation, and directly bind to the enhancer region of the PD-L1 promoter to initiate transcription, thereby fostering an immunosuppressive tumor microenvironment⁵⁰. In tumor cells, PMAIP1 overexpression enhances IL-6 secretion and activates the JAK/STAT3 signaling pathway; phosphorylated STAT3 translocates to the nucleus and binds to the PD-L1 promoter, leading to its up-regulation and facilitating immune escape⁵¹. Collectively, these findings suggest that PMAIP1 may serve as a biomarker associated with increased immune checkpoint expression and indicative of T-cell exhaustion.

As a key gene regulating cell apoptosis and metabolism, PMAIP1 expression changes often affect tumor immune status by remodeling tumor microenvironment and regulating immune-related pathways. TMB, as a core predictive marker of immunotherapy response, affects the immune system’s ability to recognize tumors by determining the number of neoantigens. Our mutational profiling revealed a higher frequency of TP53 mutations in tumors with high PMAIP1 expression. Considering that wild-type TP53 is an upstream activator of PMAIP1, loss-of-function mutations in TP53 may lead to aberrant regulation of PMAIP1 expression, taking it out of normal apoptotic checkpoint control and instead being used to serve tumor malignant progression. Some studies have found that²⁶ high TMB combined with abnormally high expression of PMAIP1 in certain cancers such as breast cancer and follicular thyroid cancer may weaken the response to immunotherapy. Because high TMB depends on sufficient immune cell infiltration to play a pro-immune response role, and PMAIP1-mediated metabolic abnormalities will destroy the integrity of the immune microenvironment. Similar to the phenomenon of high TMB but no enrichment of immune cells in some cancers, the combination of the two will make the tumor in a state of “high mutation but cold immunity”, resulting in poor efficacy of immune checkpoint inhibitors. In addition, PMAIP1 mutation combined with low TMB also reduces the probability of immunotherapy response in follicular thyroid cancer, so thyroid cancer patients with low TMB combined with abnormal PMAIP1 may accelerate tumor progression due to immunosuppression⁵².

Based on the multi-omics and multi-dimensional integrated analysis strategy of this study, combining the high resolution of single cells, the large sample size of general transcriptome, unbiased screening of machine learning, and validation of pan-cancer and clinical samples, our findings form a complete evidence chain from phenomenon to molecule to clinical significance. In particular, through cross-validation of the three machine learning models, the selected PMAIP1 is highly robust and reliable. It was found that high expression of PMAIP1 was an independent poor prognostic indicator, and PMAIP1 could be used as a predictive biomarker for response to immunotherapy in colorectal cancer. However, the problems of targeted toxicity, drug delivery barriers and drug resistance are prominent in the clinical transformation of targeted PMAIP1 therapy. Studies have found that²⁸ PMAIP1 can affect the proliferation of thyroid cancer cells by regulating the Wnt3/FOSL1 pathway. The Wnt pathway plays an important role in the development and homeostasis of normal tissues such as liver and intestine. Off-target drugs targeting PMAIP1 may interfere with the Wnt pathway in normal tissues, leading to abnormal cell proliferation or tissue repair disorders. In addition, PMAIP1 also interacts with proteins such as Bcl-2, and non-specific activation may break the balance of other anti-apoptotic proteins and cause hidden damage to sensitive tissues such as cardiomyocytes and renal tubular epithelial cells. Small molecule PMAIP1 agonists are easy to be metabolized by the liver, rapidly excreted by the kidney, and have a short half-life in the blood circulation. Therefore, they need to be administered frequently, which not only reduces the efficacy but also aggravates the toxicity. However, PMAIP1 overexpression plasmid, mRNA and other gene drugs have poor stability, are easily degraded by nuclease after entering the body, and are difficult to penetrate the tumor cell membrane by themselves.

Through an innovative multi-omics integrative analysis strategy, this study systematically elucidates the remodeling of the immune microenvironment during colorectal cancer (CRC) liver metastasis and achieves the groundbreaking identification of the dual functional characteristics of the pro-apoptotic gene PMAIP1, thereby providing a novel scientific foundation for the precise diagnosis and treatment of CRC. The core innovative contributions are as follows: (1) Methodological innovation: Development of an integrated analytical framework combining “single-cell analysis–batch validation–clinical confirmation.” This study pioneers the integration of high-resolution cellular heterogeneity profiling via single-cell RNA sequencing with large-scale clinical correlation analysis using TCGA batch transcriptomic data. By applying cross-validation through machine learning algorithms, a rigorous research pipeline encompassing “discovery–validation–confirmation” has been established; (2) Mechanistic innovation: This study reveals, for the first time, the unique role of PMAIP1 as a dual-functional regulator bridging apoptosis and immune regulation. It challenges the conventional view of PMAIP1 solely as a pro-apoptotic protein and demonstrates that PMAIP1 serves as a critical molecular hub linking apoptotic signaling and immune escape in CRC. Specifically, PMAIP1 regulates mitochondrial apoptosis through the classical pathway by targeting the anti-apoptotic protein Mcl-1. Concurrently, its elevated expression significantly upregulates immune checkpoint molecules such as PD-1 and PD-L1, promotes functional exhaustion of CD8⁺ T cells, and establishes a distinct tumor microenvironment characterized by increased immune cell infiltration yet reduced effector response. These findings uncover a previously unrecognized dual functionality of PMAIP1 and offer new mechanistic insights into the synergistic interplay between dysregulated apoptosis and immune evasion in CRC; (3) Clinical value innovation: This study establishes PMAIP1 as a novel dual-role biomarker for both prognosis prediction and immunotherapy response assessment in CRC. Pan-cancer analyses, clinical cohort validations, and RT-qPCR experiments consistently demonstrate significant overexpression of PMAIP1 in CRC tissues. Multivariate survival analysis confirms PMAIP1 as an independent prognostic risk factor associated with poor clinical outcomes (HR = 0.39, p = 0.028).

However, there are some limitations to this study. This study mainly relied on single-cell RNA sequencing and TCGA batch transcriptome data for bioinformatics analysis, which could only reveal the association between PMAIP1 gene and the prognosis and immune microenvironmental indicators of colorectal cancer. However, it cannot be directly proved that “high expression of PMAIP1 is the direct cause of poor prognosis and immune cell exhaustion”. The molecular mechanism of how PMAIP1 regulates the expression of immune checkpoint molecules and affects the function of immune cells in colorectal cancer has not been verified experimentally. The functional verification of PMAIP1 in this study only involves expression level detection, and there is a lack of mechanism experiments at the cell and animal levels. Moreover, the specific molecular mechanism of PMAIP1 in CRC still needs to be elucidated by systematic functional experiments. For example, functional verification by gene editing technology (CRISPR/Cas9) in CRC cell lines or more advanced organoid models. Secondly, the sample size of single-cell sequencing in this study was limited, and the sample size of 15 single-cell sequencing samples did not adequately cover the individual heterogeneity of colorectal cancer patients, which may affect the universality of the findings related to the immune microenvironment. In the future, larger cohorts are needed to verify our findings in the remodeling of the immune microenvironment. Finally, although the nomogram prognostic model we constructed performs well in the existing data, the effectiveness of its clinical application still needs to be verified by prospective, multi-center, large-sample clinical trials⁵³. In addition, the study did not stratified the patients who received different treatment regimens, and the clinical diagnosis and treatment factors may interfere with the correlation results of PMAIP1 with prognosis and immunophenotype, reducing the reliability of the conclusions.

Future research should focus on the following aspects: (1) Elucidating the mechanistic underpinnings of how PMAIP1 regulates the tumor immune microenvironment, particularly its direct involvement in modulating PD-L1 expression; (2) Developing targeted therapeutics, such as small-molecule inhibitors or monoclonal antibodies against PMAIP1 or its associated pathways, and evaluating their anti-tumor efficacy in preclinical models, especially in combination with existing immune checkpoint inhibitors; (3) Incorporating PMAIP1 expression as a predictive biomarker into the design of clinical trials to guide personalized therapeutic strategies for patients with CRC.

Conclusion

In conclusion, this study provides a comprehensive perspective on the intricate remodeling of the immune microenvironment during colorectal cancer (CRC) liver metastasis. Through an integrative analysis of single-cell and bulk transcriptomic data, we elucidate the dual role of the pro-apoptotic gene PMAIP1 as a pivotal nexus linking cellular apoptosis with immune evasion. Our findings establish PMAIP1 not only as an independent marker of poor prognosis in CRC but also as a promising therapeutic target. Collectively, this work deepens our understanding of the mechanisms underlying CRC metastasis and offers a compelling scientific rationale for developing novel, personalized therapeutic strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

XFL: Writing - Original Draft; SPD: Conceptualization, Methodology; ZHC: Visualization; JHG: Resources, Software; HYZ: Data curation; YQW: Validation; YLW: SupervisionJCC and ZQH: Writing - Review & Editing.All authors have read and approved the manuscript.

Funding

This work was supported by Research Startup Fund for Introduced Talents and the National Natural Science Foundation of China (No. 82360529).

Data availability

The datasets generated and/or analyzed during this study are available in the GEO (Gene Expression Omnibus) repository. The direct link to access the gene expression and clinical pathology dataset is: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE178318; https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE60331;This link provides access to the original dataset used in our analysis.The data supporting the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics statements

This study has been approved by the Research Ethics Committee of Chuxiong Prefecture People’s Hospital(KYLL-2025010).