Integrative analysis of polyamine-associated genes reveals a prognostic and immunological signature in esophageal squamous cell carcinoma

Abstract

Esophageal squamous cell carcinoma (ESCA) is a highly aggressive malignancy with substantial heterogeneity and poor prognosis. Polyamine metabolism has been implicated in tumor progression and immune regulation, yet its specific role in ESCA remains unclear. Here, we performed integrative single-cell and bulk transcriptomic analyses to explore the significance of polyamine metabolism in ESCA. Using single-cell RNA-seq data from four ESCA patients (GSE188900), we identified seven major cell populations and evaluated polyamine activity via gene set scoring. B lymphocytes and myeloid cells exhibited the highest polyamine scores. Differential expression and enrichment analyses between Polyamine-High and -Low groups revealed associations with immune-related pathways, including T cell activation and cell adhesion. From these genes, we developed a prognostic model consisting of eight polyamine-associated genes (LYPD3, DHPS, MUC5B, CXCL14, SQSTM1, RUNX3, PTPRC, and KRT14) using Cox and LASSO regression. The model effectively stratified patients into high- and low-risk groups in both the GSE53624 training and GSE53622 validation cohorts, with the high-risk group showing significantly worse survival. Immune infiltration analysis using MCPcounter, xCell, and ssGSEA showed distinct immune landscapes across risk groups, with low-risk patients exhibiting higher neutrophil and type 2 helper T cell infiltration. Drug sensitivity analysis based on oncoPredict revealed compounds with differential efficacy between risk groups, and the model also predicted response to PD-1 blockade in an external immunotherapy cohort. In summary, polyamine metabolism is closely linked to the immune microenvironment and prognosis of ESCA, providing a potential biomarker for patient stratification and treatment optimization.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03915-z.

Introduction

Esophageal squamous cell carcinoma, the most common histological type of esophageal cancer (EC), constitutes nearly 80% of all EC cases globally and continues to pose a significant challenge to public health worldwide [1]. Globally, esophageal cancer is the sixth most frequent cause of cancer-related death and ranks eighth among all malignancies, with more than 572,000 newly diagnosed cases and approximately 509,000 deaths reported in 2018 [2]. The incidence of ESCA is particularly high in East Asia, especially in China, where it accounts for over 95% of EC cases and causes an estimated 350,000 deaths annually [3]. Despite advances in surgical techniques, endoscopic detection, and multimodal therapies, the prognosis of ESCA remains dismal, with a 5-year survival rate below 30% and as low as 15.3% in advanced stages [4, 5]. The insidious onset and lack of early symptoms result in late diagnosis for the majority of patients, many of whom present with unresectable or metastatic disease [6]. Although various clinical and molecular biomarkers, such as TP53, CCND1, and MDM2 mutations, have been implicated in ESCA pathogenesis, their utility in early diagnosis or accurate prognostication remains limited [7]. Moreover, current predictive models—mostly based on traditional statistical methods and limited clinical indicators—often lack robustness, reproducibility, and clinical applicability [8, 9]. These findings highlight the pressing necessity to discover novel and dependable prognostic biomarkers, as well as to establish comprehensive predictive models capable of accurately stratifying patient risk and informing personalized therapeutic approaches in ESCA.

Polyamines, including putrescine, spermidine, and spermine, are small aliphatic polycations that play essential roles in DNA stabilization, gene transcription, mRNA translation, and cell proliferation [10, 11]. In cancer, polyamine metabolism is frequently dysregulated, leading to enhanced biosynthesis, uptake, and recycling to meet the increased metabolic demands of rapidly dividing tumor cells [11, 12]. Oncogenes such as MYC upregulate key enzymes in the polyamine biosynthetic pathway, while tumor suppressors tend to downregulate these metabolic processes [12]. In addition to supporting tumorigenesis, polyamines are increasingly recognized as modulators of the tumor immune microenvironment [13]. TME is an ecosystem with a high level of organization, characterized by the intricate interactions of its diverse cellular and noncellular components [14]. Increased polyamine concentrations have been demonstrated to inhibit T cell proliferation and activation, while facilitating the expansion of immunosuppressive populations like myeloid-derived suppressor cells (MDSCs), consequently weakening antitumor immune responses [13, 15]. Targeting polyamine metabolism is thus a promising therapeutic strategy, offering dual benefits of inhibiting tumor cell growth and enhancing immune-mediated tumor clearance [10, 13]. Furthermore, studies in colorectal, pancreatic, and lung cancers have demonstrated that polyamine metabolic signatures correlate with distinct tumor microenvironment characteristics, prognosis, and responses to immunotherapy [16–20]. Although progress has been made, the function of polyamine metabolism in ESCA is still not well elucidated, necessitating further investigation to clarify its prognostic significance and potential as a therapeutic target in this cancer.

Materials and methods

Data acquisition

Gene expression profiles and corresponding clinical annotation data were retrieved from the Gene Expression Omnibus (GEO) public database (https://www.ncbi.nlm.nih.gov/geo/) to ensure robust and reproducible bioinformatic analyses. Specifically, we included the following datasets in this study:

GSE53624, which comprises bulk RNA-sequencing (RNA-seq) data from 117 ESCA patients, along with comprehensive clinical follow-up information, was used as the training cohort for model development [21]. GSE53622, consisting of bulk RNA-seq data from 60 ESCA patients, served as an independent external validation cohort to assess the robustness and generalizability of the prognostic model [21]. GSE188900, containing scRNA-seq data from four primary ESCA samples, was utilized for single-cell transcriptomic analysis to characterize the tumor microenvironment and identify cell type–specific features associated with polyamine metabolism [22]. All datasets included in this study were publicly available and did not require additional ethical approval. The data were processed and normalized according to the corresponding original publications or GEO platform guidelines before downstream analysis.

Polyamine metabolism gene set

The gene set related to polyamine metabolism was obtained from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/index.jsp), a well-established resource for curated gene sets used in functional enrichment analyses [23]. Specifically, the polyamine-related gene set was downloaded and used to assess the activity of polyamine metabolism at both bulk and single-cell transcriptomic levels. This gene set was further applied in downstream analyses, including module scoring, differential expression analysis, and pathway enrichment, to explore its biological significance and prognostic relevance in ESCA.

Single-cell RNA-seq data processing and analysis

Single-cell RNA sequencing (scRNA-seq) data were analyzed using the Seurat R package [24, 25]. Quality control involved removing cells expressing fewer than 200 or more than 10,000 unique genes via nFeature_RNA, as well as cells with mitochondrial gene content exceeding 10% of total UMI counts (percent.mt >10%), to exclude low-quality or apoptotic cells. Gene expression normalization was carried out with the NormalizeData function using default settings, scaling expression relative to each cell’s total expression.

To mitigate batch effects and integrate single-cell datasets from multiple samples, we utilized the Harmony algorithm. Following identification of highly variable genes and data scaling, principal component analysis (PCA) was performed for dimensionality reduction, with the top 15 principal components selected for further analysis. Visualization was achieved using t-distributed stochastic neighbor embedding (t-SNE). Cell clustering was conducted via the FindClusters function with a resolution of 1.5. The single-cell transcriptomic analysis included four ESCA patient samples from the GSE188900 dataset. Based on canonical cell-type marker genes, cells were annotated into seven major clusters: B lymphocytes, T lymphocytes, Fibroblasts, Endothelial cells, Epithelial cells, Myeloid cells, and Mast cells.

To evaluate polyamine metabolic pathway activity at the single-cell level, we applied the AddModuleScore function in Seurat using a polyamine-related gene set sourced from the MSigDB database. Immune cell clusters were categorized into Polyamine-High and Polyamine-Low groups based on their polyamine scores. Additionally, Gene Set Variation Analysis (GSVA) was performed for 50 hallmark tumor-associated pathways from the MSigDB Hallmark collection across seven cell clusters to examine functional heterogeneity. The results demonstrated distinct pathway enrichment patterns among the various cell populations.

Cell–cell communication analysis

Cell–cell communication analysis was performed using the CellChat R package (v0.0.2), which infers intercellular signaling networks based on known ligand–receptor interactions [26]. The CellChatDB.human database was used as the reference to identify potential signaling events between cell types. This analysis enabled the identification of key signaling pathways and interaction strengths among different cell populations.

Differential gene expression analysis

Differentially expressed genes (DEGs) distinguishing the Polyamine-High and Polyamine-Low groups were identified utilizing the FindAllMarkers function within the Seurat package. Genes meeting the thresholds of adjusted P-value < 0.05 and |log2 fold change| >0.585 were considered significant. These selected DEGs were then subjected to subsequent functional enrichment analyses to investigate the associated biological processes and signaling pathways.

Functional enrichment analysis

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed on the DEGs between the Polyamine-High and Polyamine-Low groups using the clusterProfiler R package (version 4.0.5). Enrichment results with a false discovery rate (FDR) below 0.05 were considered statistically significant.

Construction and validation of the prognostic risk model

To identify prognostic genes associated with polyamine metabolism, we first conducted univariate Cox regression on the DEGs using the tinyarray R package [27]. Genes with P-values below 0.05 were considered significant, resulting in 17 candidate prognostic genes. Next, Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to reduce dimensionality and prevent overfitting [28]. cv.glmnet (X, y, alpha = 1) was set, and λ sequence was then generated as default automatic selection, covering the range from minimum MSE to near-zero penalty. 10-fold cross-validation was performed to calculate the average error MSE under each λ, and the cross-validation curve was visualized. The plot of cvfit shows the change of MSE with λ as “U-shaped”, while the lowest point or near the lowest point of λ is selected. Finally, eight genes—LYPD3, DHPS, MUC5B, CXCL14, SQSTM1, RUNX3, PTPRC, and KRT14—were selected to construct the prognostic model. Each patient’s risk score was calculated as a weighted sum of the expression levels of these genes, with weights corresponding to the LASSO-derived coefficients, according to the following formula:

Coef represents the prognostic contribution of each gene obtained from multivariate Cox regression, while gene expression values indicate the corresponding levels of these prognostic genes. Using the median risk score as the cutoff, patients in the GSE53624 cohort were categorized into high- and low-risk groups. Kaplan–Meier survival analysis showed that the low-risk group had significantly better overall survival than the high-risk group. To evaluate the model’s stability and external applicability, it was further validated in an independent cohort (GSE53622) of 60 ESCA patients, where the risk score similarly demonstrated effective prognostic stratification.

Immune infiltration analysis

To comprehensively assess the immune cell infiltration landscape in ESCA, we employed three complementary computational algorithms: MCPcounter, single-sample Gene Set Enrichment Analysis (ssGSEA), and xCell [29–31]. MCPcounter was used to quantify the absolute abundance of ten immune and stromal cell populations based on transcriptomic profiles. ssGSEA was applied to estimate the enrichment scores of 28 predefined immune cell types for each sample using gene sets derived from previous immunological studies. xCell provided additional high-resolution estimates for a broader range of immune and stromal cell types by incorporating spillover compensation and cell-type-specific gene expression signatures.

Immune infiltration results were depicted through boxplots comparing immune cell abundances between high- and low-risk groups, heatmaps showcasing relative infiltration patterns across samples, and scatter plots illustrating correlations between immune cell populations and risk scores or prognostic genes.

Immune checkpoint inhibitor cohort analysis

To evaluate the potential predictive value of the established risk score for immunotherapy response, we applied it to an independent immunotherapy-treated cohort derived from the CheckMate study. This cohort includes patients with metastatic urothelial carcinoma who received nivolumab, a PD-1 immune checkpoint inhibitor. Patients were categorized into high- and low-risk groups according to the median risk score. Clinical outcomes following PD-1 blockade, such as complete response (CR) and partial response (PR), were compared between these groups.

Drug sensitivity analysis

Drug sensitivity was estimated using the oncoPredict R package based on gene expression profiles and the GDSC database [32]. Predicted drug response scores (IC50) were calculated for each sample to assess differences in drug sensitivity between high-risk and low-risk groups.

Cell culture

The human normal esophageal epithelial cell line Het-1 A and the esophageal squamous cell carcinoma cell line EC9706 were obtained from American Type Culture Collection (ATCC, VA, USA). Both cell lines were maintained in RPMI-1640 medium supplemented with 10% FBS and incubated at 37 °C in a humidified atmosphere containing 5% CO₂.

Real-time quantitative PCR (RT-qPCR)

Total RNA was isolated using TRIzol reagent (Takara, Japan) following the manufacturer’s protocol. Complementary DNA (cDNA) was synthesized via reverse transcription of the extracted RNA. GAPDH served as the internal reference gene. RUNX3 expression levels were quantified by RT-qPCR. Primer sequences are provided in Supplementary Table S1.

Statistical analysis

All statistical analyses were carried out in the R software environment. Univariate and multivariate Cox proportional hazards regression analyses were performed using the survival and survminer packages, with a significance cutoff of p < 0.05. Kaplan–Meier curves were generated to assess survival differences, which were evaluated using the log-rank test. Additional statistical methods, including correlation and differential expression analyses, were applied as appropriate. All tests were two-sided, and p-values below 0.05 were considered statistically significant.

Result

Single-cell transcriptomic profiling and cellular communication landscape in ESCA

As illustrated in Fig. 1, we first constructed a comprehensive analytical workflow integrating single-cell RNA sequencing (scRNA-seq), prognostic modeling, immune infiltration characterization, and drug sensitivity prediction to investigate the tumor microenvironment and molecular heterogeneity in esophageal squamous cell carcinoma (ESCA).

Fig. 1.

Overview of the experimental design and analytical workflow

We performed scRNA-seq on tumor tissues from four ESCA patients and conducted dimensionality reduction and clustering analysis. A total of seven distinct cell populations were identified and annotated, including B lymphocytes, T lymphocytes, fibroblasts, endothelial cells, epithelial cells, myeloid cells, and mast cells (Fig. 2A). Cell-type annotation was based on the expression of canonical marker genes, such as CD79A (B lymphocytes), CD3D (T lymphocytes), LUM (fibroblasts), S100A2 (epithelial cells), LYZ (myeloid cells), TPSAB1 (mast cells), and RAMP2 (endothelial cells) (Fig. 2B). We then quantified the cellular composition across the four individual samples and observed marked inter-patient heterogeneity in the distribution of cell populations (Fig. 2C). Specifically, Sample 1 was predominantly composed of T lymphocytes, fibroblasts, and myeloid cells; Sample 3 was enriched for T lymphocytes; Sample 4 was primarily constituted by fibroblasts and T lymphocytes; while Sample 5 exhibited a dominance of epithelial cells and B lymphocytes.

Fig. 2.

Single-cell atlas and cellular communication landscape of ESCA. A t-SNE visualization of ESCA single-cell clusters and representative cell type-specific marker genes. The left panel shows the t-SNE plot of all ESCA cells colored by clusters, and the right panels depict the expression patterns of representative marker genes for different cell types. B Bubble plot showing the expression of cell type-specific marker genes. Marker genes are displayed for each cluster, with dot size indicating the proportion of cells expressing the gene and color indicating average expression level. C Proportions of distinct cell types across four ESCA patients. The bar plot illustrates the cellular composition for each patient, highlighting inter-individual heterogeneity. D Cell–cell communication network among major cell types in ESCA. Line thickness reflects the total interaction strength between cell types, suggesting prominent communication hubs

To investigate intercellular communication within the tumor microenvironment, we performed cell–cell communication analysis. The results revealed that fibroblasts and endothelial cells exhibited the most extensive and robust interactions with other cell types (Fig. 2D), suggesting their central roles in the ESCA stromal network. Furthermore, we found that B lymphocytes, epithelial cells, myeloid cells, and mast cells predominantly communicated with T lymphocytes, indicating potential immunoregulatory dynamics (Fig. 3A).

Fig. 3.

Polyamine metabolism heterogeneity and pathway enrichment across ESCA single-cell populations. A Cell–cell communication network among seven major cell populations. Edge thickness represents the strength of intercellular signaling interactions, highlighting key communication nodes. B t-SNE plot showing the distribution of single-cell populations across four ESCA patients. Each color represents a distinct cell cluster, demonstrating patient-specific cellular composition. C t-SNE visualization of polyamine metabolism activity. Cells are colored based on Polyamine scores derived from gene set enrichment analysis of the polyamine metabolism pathway, classifying cells into Polyamine-High (Polyamine-H) and Polyamine-Low (Polyamine-L) groups. D Raincloud plot illustrating the distribution of Polyamine scores across cell types. E Heat maps show enrichment scores for hallmark pathways for seven cell populations. Rows represent different cell types and columns represent different pathways

Polyamine metabolic profiling and functional heterogeneity across ESCA cell subsets

To further explore inter-patient heterogeneity and metabolic features within the tumor microenvironment, we visualized the single-cell distribution from the four ESCA patients using t-SNE, which revealed distinct cellular clustering patterns across individuals (Fig. 3B). Subsequently, we evaluated polyamine metabolism activity at the single-cell level. Using the AddModuleScore function and a curated gene set related to the polyamine metabolic pathway, we calculated a Polyamine score for each cell. Based on the resulting scores, cells were stratified into Polyamine-High and Polyamine-Low groups (Fig. 3C). Notably, B lymphocytes and myeloid cells exhibited the highest polyamine metabolic activity among all cell populations, indicating that these cell types may play key roles in polyamine-associated immune regulation (Fig. 3D).

To further dissect the functional states of different cell subsets, we performed GSVA using the 50 hallmark gene sets from the MSigDB. The heatmap displayed distinct enrichment patterns across the seven major cell clusters (Fig. 3E). We observed that several cancer-related pathways were differentially activated. For instance, T lymphocytes showed strong enrichment in cell cycle–related pathways, including G2M checkpoint and E2F targets, suggesting heightened proliferative activity. In contrast, fibroblasts displayed marked activation of epithelial–mesenchymal transition (EMT) and angiogenesis pathways, highlighting their potential roles in tumor progression and stromal remodeling.

Identification of prognostic DEGs and construction of a polyamine-related risk model

To discover prognostic biomarkers linked to polyamine metabolism, we performed differential expression analysis between Polyamine-High and Polyamine-Low groups. The identified DEGs underwent GO and KEGG enrichment analyses. GO results highlighted significant enrichment in immune-related processes such as leukocyte cell–cell adhesion, regulation of cell adhesion, and T cell activation, suggesting a role for polyamine metabolism in modulating the tumor immune microenvironment (Fig. 4A). KEGG analysis showed that these genes were involved in pathways including Human T-cell leukemia virus 1 infection, Epstein–Barr virus infection, hematopoietic cell lineage, and differentiation of Th17, Th1, and Th2 cells, emphasizing their relevance in tumor immune signaling (Fig. 4B).

Fig. 4.

Functional enrichment analysis and prognostic model construction based on DEGs. A GO enrichment bar plot of differentially expressed genes (DEGs). Significantly enriched GO biological processes are shown, highlighting functional characteristics of DEGs. B KEGG enrichment bar plot of DEGs. The most relevant KEGG pathways are presented, reflecting major signaling and metabolic pathways involved. C LASSO Cox regression analysis for feature selection. Ten-fold cross-validation results showing the partial likelihood deviance and corresponding coefficient profiles plotted against log(λ), used to identify optimal predictors. D Kaplan–Meier survival curves for high- and low-risk groups in the training cohort. Patients were stratified based on the risk score, with significant survival differences observed between groups. E Scatter plot illustrating the association between risk score and patient survival status. Each dot represents a patient, colored by survival status and arranged by increasing risk score

Univariate Cox regression analysis was performed on the DEGs to identify those associated with overall survival in ESCA patients. Seventeen genes with p-values < 0.05 were selected as potential prognostic markers. These were further narrowed down by LASSO regression, yielding eight genes with the strongest prognostic value: LYPD3, DHPS, MUC5B, CXCL14, SQSTM1, RUNX3, PTPRC, and KRT14 (Fig. 4C).

A polyamine-related risk score was constructed based on the expression and regression coefficients of these eight genes using the GSE53624 training set. Patients were classified into high- and low-risk groups according to the median risk score. Kaplan–Meier survival curves showed that patients in the low-risk group had significantly better overall survival than those in the high-risk group (P = 0.007; Fig. 4D). Additionally, the distribution of survival status and risk scores indicated that patients with higher risk scores had an increased risk of death within the training cohort (Fig. 4E).

External validation of the prognostic model and genomic characterization of key genes

To evaluate the stability and external validity of the polyamine-related prognostic model, we tested it on an independent cohort (GSE53622) comprising transcriptomic and clinical data from 60 ESCA patients. Kaplan–Meier survival analysis showed a significant survival difference between high- and low-risk groups, with better outcomes in the low-risk group (P = 0.042; Fig. 5A). Consistently, the distribution of risk scores and survival status indicated that patients with higher risk scores faced increased mortality risk, supporting the model’s prognostic value (Fig. 5B). Analysis of the eight signature genes revealed that CXCL14, DHPS, KRT14, and LYPD3 were significantly upregulated in the low-risk group, whereas RUNX3 expression was notably higher in the high-risk group (Fig. 5C). These differential expression patterns indicate that these genes may exert opposing effects on ESCA progression. Correlation analysis among the eight model genes revealed several gene clusters with strong co-expression relationships. Specifically, DHPS and SQSTM1 exhibited a positive correlation; similarly, CXCL14, LYPD3, and KRT14 were positively correlated with each other. Another distinct correlation cluster was observed among RUNX3, MUC5B, and PTPRC (Fig. 5D), suggesting potential co-regulatory or pathway-based functional interactions.

Fig. 5.

Validation and visualization of the prognostic model in the testing cohort. A Kaplan–Meier survival curves for high- and low-risk groups in the validation cohort. Significant survival differences indicate the robustness of the prognostic model. B Scatter plot showing the relationship between risk scores and patient survival status in the validation set. Patients are arranged by increasing risk score and colored by survival outcome. C Box plots of the expression levels of eight model genes between high- and low-risk groups. D Correlation heatmap (corrplot) of the eight model genes. E Nomogram predicting 1-, 3-, and 5-year overall survival. F Genomic distribution of the eight model genes

To facilitate individualized survival prediction in clinical practice, we constructed a nomogram incorporating the risk score and relevant clinicopathological parameters to estimate 1-, 3-, and 5-year overall survival probabilities for ESCA patients (Fig. 5E). This model provides an intuitive and quantitative tool for clinical decision-making. Finally, to elucidate the genomic locations of the eight prognostic genes, we mapped their chromosomal distributions. RUNX3 and PTPRC were located on chromosome 1, CXCL14 and SQSTM1 on chromosome 5, MUC5B on chromosome 11, KRT14 on chromosome 17, and both LYPD3 and DHPS on chromosome 19 (Fig. 5F). These spatial arrangements may offer clues for future studies on chromosomal aberrations and co-localized regulatory mechanisms in ESCA.

Association between risk score and immune landscape reveals immunological relevance of the prognostic model

To further investigate the immunological relevance of our risk model, we examined the correlation between immune checkpoint inhibitor (ICI) target gene expression and the polyamine-related risk score. As illustrated in Fig. 6A, IL12A expression showed a positive correlation with the risk score, while IFNA2, IL13, PDCD1, and TNFSF9 were significantly negatively correlated. These results imply that patients classified as high-risk may possess an immune microenvironment with diminished responsiveness to immune activation signals. To validate the predictive potential of our risk model in the context of immunotherapy, we utilized a publicly available dataset from the CheckMate clinical trial, which includes transcriptomic and clinical response data from metastatic urothelial carcinoma patients treated with the PD-1 inhibitor nivolumab. Interestingly, patients classified into the high-risk group exhibited a higher proportion of objective responses (CR/PR) compared to the low-risk group, indicating a potentially greater sensitivity to anti-PD-1 immunotherapy (Fig. 6B).

Fig. 6.

Immune landscape and immunotherapy relevance of the prognostic model. A Correlation analysis between risk score and expression of immune checkpoint-related genes. B Bar plot of immunotherapy response in high- and low-risk groups within the CheckMate cohort. The proportions of CR/PR (complete or partial response) and SD/PD (stable or progressive disease) are compared between the two risk groups. C Box plots showing differences in immune cell infiltration between high- and low-risk groups. Immune infiltration levels were estimated using the MCPcounter algorithm. D Stacked bar plot of immune cell composition across 117 patients. Based on MCPcounter results, the cellular makeup of the tumor microenvironment is visualized for each patient. E Heat map of correlation between eight model genes and immune cell populations. Correlation matrix displays the association between key genes and immune cell populations inferred by MCPcounter

We further analyzed immune cell infiltration patterns using the MCPcounter algorithm, which estimates the relative abundance of immune and stromal cell types from transcriptomic data. The analysis revealed significantly higher neutrophil infiltration in the low-risk group compared to the high-risk group (Fig. 6C). The overall immune cell composition across all patients, as inferred by MCPcounter, is shown in the stacked bar plot (Fig. 6D), highlighting differences in immune cell subset distributions between risk groups. Additionally, correlation analysis between the eight model genes and immune cell abundance demonstrated that LYPD3 and DHPS were positively correlated with most immune cell types, whereas MUC5B, RUNX3, and PTPRC showed mainly negative correlations (Fig. 6E). These associations reinforce the involvement of polyamine-related genes in modulating the ESCA immune microenvironment.

Integrated immune infiltration analysis reveals multi-algorithmic consistency and risk-associated immune components

To further characterize the immune landscape of ESCA in the context of risk stratification, we expanded our immune infiltration analysis using both xCell and ssGSEA algorithms. These complementary approaches enabled us to comprehensively profile the tumor immune microenvironment and its association with the polyamine-based risk score. Using the xCell algorithm, we identified 11 immune and stromal cell types that exhibited significant differences in enrichment between the high-risk and low-risk groups. These included Astrocytes, CD4⁺ Tem, CD8⁺ Tem, Epithelial cells, Hematopoietic stem cells (HSC), Keratinocytes, Macrophages M2, Mast cells, Melanocytes, Plasma cells, and Sebocytes (Fig. 7A). Correlation analysis among these cell types revealed intricate intercellular relationships. For example, Sebocytes were positively correlated with Epithelial cells and Keratinocytes, suggesting potential tissue-specific co-enrichment. Conversely, Astrocytes exhibited strong negative correlations with Epithelial cells, Keratinocytes, Macrophages M2, Mast cells, Plasma cells, and Sebocytes (Fig. 7B). A ridge plot further illustrated the distributional differences in immune cell abundance across high- and low-risk groups, underscoring the immunological divergence between them (Fig. 7C). Additionally, correlation analysis revealed that LYPD3 and KRT14 were significantly and positively correlated with most immune cell types, whereas both genes showed negative correlations with Astrocytes, suggesting a potential suppressive role of astrocyte-like stromal components in modulating tumor immunity (Fig. 7D). Building upon these findings, we performed a direct correlation analysis between risk score and immune cell abundance as inferred by xCell. We identified seven immune cell types significantly associated with the risk score: Astrocytes, CD4⁺ memory T cells, and Melanocytes were positively correlated, while Macrophages M2, Mast cells, Plasma cells, and Keratinocytes exhibited negative correlations with the risk score (Fig. 8A). These results suggest that specific immune populations may play divergent roles in tumor progression and patient prognosis.

Fig. 7.

Figure 7. Immune infiltration profiling between high- and low-risk groups using xCell algorithm. A Box plots showing differential abundance of 11 immune cell types between high- and low-risk groups. Immune cell infiltration was estimated using the xCell algorithm, with significant differences observed in 11 immune populations. B Correlation heatmap (corrplot) of the 11 differentially infiltrated immune cell types. C Ridge plots comparing immune cell infiltration levels between high- and low-risk groups. Distributions of the 11 immune cell scores reveal distinct immune patterns associated with risk stratification. D Heatmap showing correlations between the eight model genes and 11 xCell-derived immune cell populations

Fig. 8.

Immune infiltration analysis and its association with risk score. A Scatter plots showing correlations between risk score and immune cell types based on xCell algorithm. B Box plots comparing immune cell infiltration between high- and low-risk groups based on ssGSEA algorithm. C Correlation heatmap (corrplot) of immune cell types inferred by ssGSEA

To confirm the immune infiltration patterns, we applied single-sample gene set enrichment analysis (ssGSEA). The results validated that the low-risk group had significantly increased infiltration of neutrophils and Type 2 T helper (Th2) cells compared to the high-risk group (Fig. 8B), suggesting enhanced immune surveillance in low-risk ESCA. Additionally, correlation analysis of immune cell types based on ssGSEA data showed mostly positive associations among immune populations, implying possible co-regulation or coordinated activation within the tumor immune microenvironment (Fig. 8C).

Drug sensitivity analysis and experimental validation highlight RUNX3 as a potential therapeutic biomarker

To further explore potential therapeutic strategies for ESCA patients stratified by polyamine-related risk score, we performed a comprehensive drug sensitivity analysis using pharmacogenomic data. In the correlation analysis, we identified several clinically relevant compounds whose sensitivity was significantly associated with the expression levels of the model-derived prognostic genes. Notably, RUNX3 expression was negatively correlated with the sensitivity to AZD5991_1720, EPZ004777_1237, GSK343_1627, Olaparib_1017, and OSI − 027_1594, suggesting that elevated RUNX3 expression may contribute to resistance against these agents. Conversely, RUNX3 expression was positively correlated with Ribociclib_1632 and SCH772984_1564, implying potential vulnerability to these targeted therapies in RUNX3-overexpressing tumors (Fig. 9A).

Fig. 9.

Drug sensitivity analysis associated with the prognostic gene signature. A Heatmap showing correlations between the eight model genes and drug sensitivity-related compounds. B Box plots comparing drug sensitivity between high- and low-risk groups. C Bar graphs illustrate the differential expression of the RUNX3 gene between normal and ESCA cells

In addition, comparative drug sensitivity analysis between high-risk and low-risk groups revealed a number of compounds with significantly different predicted responses. We highlighted six representative agents with pronounced differential sensitivity: Olaparib_1017, PFI3_1620, WIKI4_1940, Gallibiscoquinazole_1830, EPZ5676_1563, and EPZ004777_1237 (Fig. 9B). These findings suggest that polyamine-related risk stratification may inform individualized treatment choices and improve therapeutic precision.

To experimentally validate key genes from the prognostic model, we focused on RUNX3, which was found to be highly expressed in the high-risk group. RUNX3 mRNA levels were measured by RT-qPCR in the normal human esophageal epithelial cell line Het-1 A and the esophageal squamous cell carcinoma cell line EC9706. Consistent with the transcriptomic data, RUNX3 expression was significantly elevated in EC9706 cells compared to Het-1 A cells, supporting its potential involvement in ESCA tumor development and progression (Fig. 9C). Overall, these results indicate that RUNX3 may function both as a prognostic biomarker and a potential predictor of drug response, providing valuable insight for risk stratification and therapeutic intervention in ESCA patients.

Discussion

In this study, we employed an integrative framework combining single-cell transcriptomic analysis, intercellular communication inference, and functional profiling to dissect the cellular and molecular heterogeneity of ESCA. Through scRNA-seq analysis of tumor tissues from four ESCA patients, we identified seven major cell populations, including immune and stromal components, and revealed substantial inter-patient variability in their composition. Notably, fibroblasts and endothelial cells emerged as central hubs within the tumor microenvironment, displaying the highest degree of cell–cell interactions, which underscores their pivotal roles in maintaining the structural and signaling landscape of ESCA. In contrast, immune cells such as B lymphocytes, myeloid cells, and mast cells demonstrated preferential communication with T lymphocytes, suggesting complex immunoregulatory networks within the tumor niche. Furthermore, by quantifying polyamine metabolic activity at single-cell resolution, we uncovered cell type–specific metabolic heterogeneity, with B cells and myeloid cells exhibiting elevated polyamine pathway activity. This finding suggests that polyamine metabolism may contribute not only to the metabolic demands of proliferative immune cells but also to their immunomodulatory functions. Functional pathway analysis further highlighted cell type–dependent enrichment of hallmark pathways, with T cells showing strong proliferative signatures and fibroblasts exhibiting mesenchymal and angiogenic programs. These results collectively provide a high-resolution view of the ESCA microenvironment and emphasize the biological significance of polyamine metabolism in shaping cell states and interactions. Our findings lay a foundation for further investigation into the metabolic–immune crosstalk in ESCA and support the rationale for targeting polyamine pathways in therapeutic strategies.

To elucidate the functional implications of polyamine metabolism in ESCA, we first compared transcriptomic profiles between Polyamine-High and Polyamine-Low cell populations and identified a set of DEGs. Functional enrichment analysis revealed that these DEGs were predominantly involved in immune-related biological processes and pathways. GO terms such as “leukocyte cell–cell adhesion,” “regulation of cell–cell adhesion,” and “regulation of T cell activation” suggest that altered polyamine metabolism may influence immune cell recruitment, intercellular interactions, and T cell functional states within the tumor microenvironment. These findings are in line with accumulating evidence that polyamines act as modulators of immune cell function, affecting T cell proliferation, macrophage polarization, and myeloid-derived suppressor cell (MDSC) activity [33–36].

We constructed an eight-gene prognostic model based on polyamine metabolism-related differential genes that successfully divided ESCA patients into high and low risk groups and showed good survival prediction ability in both the training set and the external validation set (GSE53622). In the model, CXCL14, DHPS, KRT14, LYPD3 and other genes were highly expressed in the low-risk group and positively correlated with immune cell infiltration, while RUNX3, MUC5B, and PTPRC were associated with high-risk and immunosuppressive characteristics, suggesting that these genes may play different roles in tumor immune regulation. The immunoassay results showed that patients in the high-risk group had weak immunocompetence, accompanied by down-regulation of key immune checkpoints such as PDCD1 and TNFSF9; while in the immunotherapy dataset, high-risk patients responded better to PD-1 inhibitors, suggesting their potential immunotherapeutic sensitivity. Further combining algorithms such as MCPcounter, xCell, and ssGSEA, we found that the low-risk group was enriched for more immunocompetent cells, such as neutrophils and Th2 cells, while the high-risk group showed more cell types associated with immunosuppression. In ESCA, the levels of tumor-infiltrating neutrophils are strongly associated with tumor progression, and their high levels often indicate poor prognosis [37]. In addition, IL-17 can mediate antitumor immune responses and enhance their antitumor activity by promoting the recruitment and activation of neutrophils, suggesting that neutrophils have dual roles in specific immune contexts [38]. Pretreatment neutrophil-to-lymphocyte ratio (NLR) is an important indicator of immune status, and high NLR usually indicates immunosuppressive status and poor survival prognosis [39]. In particular, the increased ratio of neutrophils to CD8 + T cells often suggests diminished cytotoxic immune function and enhanced Th2-dominant immune responses, further associated with adverse tumor outcomes [40]. These results further support the reliability and potential for clinical application of this model in reflecting the state of the tumor immune microenvironment. At present, the TNM stage is the most crucial element in assessing the prognosis of ESCA patients [41]. It would be a novel insight to integrate TNM stage with our established model to prognose ESCA for promising efficacy. The Cox model is commonly utilized in clinical follow-up studies to explore the association between survival outcomes and covariates such as age [42]. Numerous clinical researchers assert that biases in observational data can be rectified by applying a statistical regression model, like a logistic model for response or a Cox model for survival time, to account for patient prognostic covariates [43]. In the future, we would merge deep learning approaches with the conventional Cox model to assess the nonlinear effects of clinical covariates for forecasting the clinical outcomes of ESCA patients.

On the basis of revealing that polyamine metabolism is tightly associated with the tumor immune microenvironment, we further explored potential treatment strategies for patients with ESCA stratified based on polyamine-related risk scores. Through drug sensitivity analysis, it was found that there was a significant association between model key gene expression and the response to a variety of clinically relevant drugs, especially the high expression of RUNX3 in the high-risk group was closely related to multiple drug resistance, such as AZD5991, EPZ004777, GSK343, olaparib and OSI-027; while it showed potential sensitivity to targeted drugs such as Ribociclib and SCH772984, suggesting that RUNX3 may mediate the selective response of different drugs. This differentiated drug sensitivity not only provides a basis for precise medication in patients at high and low risk, but also reflects the potential role of polyamine metabolism in the mechanism of tumor resistance. Further verified by in vitro experiments, RUNX3 was significantly upregulated in esophageal cancer cell lines, supporting its critical role in tumor development.

Interestingly, several literatures have shown that RUNX3 expression is high in normal esophageal epithelial tissues, while silencing or down-regulation often occurs in ESCA tissues, mainly related to high-frequency methylation of promoter regions [44–46]. This epigenetic alteration is thought to be one of the early events in the development of ESCA, leading to its loss of function. As a major tumor suppressor, RUNX3 is involved in gastric, colon, and several other solid tumors. It is often inactivated by mechanisms such as hemizygous deletion, promoter hypermethylation, histone modification, and protein mislocalization [47]. In addition, loss or silencing of RUNX3 is strongly associated with radioresistance and poor prognosis, suggesting that it has an important role in treatment response and tumor malignant progression [48]. RUNX3 is responsible for activating transcription in cytotoxic T cells and NK cells. Overexpression of Runx3 increased the abundance of tumor-infiltrating lymphocytes and the expression of granzyme B along with specific core tissue residency genes, while reducing the expression of core circulating genes [49]. Mechanistically, RUNX3 can inhibit the EMT process by regulating the TGF-β/Smad pathway, thereby limiting the invasion and migration of tumor cells, further supporting its role as a tumor suppressor gene [50].

It is notable to be pointed out limitations. First, the small sample size of the scRNA-seq may pose challenges to the generalizability of the findings. Additional independent validation in more diverse populations would enhance the model’s clinical applicability. In the context of ESCA, genetic variations associated with different ethnicities may influence disease progression and response to treatment. For instance, studies have shown that certain genetic polymorphisms can affect drug metabolism and therapeutic outcomes, highlighting the necessity of inclusivity in model validation. Second, this study is absence of functional experiments in animal models to validate the mechanisms underlying polyamine-associated gene functions would add biological credibility. We only examined the RUNX3 mRNA level between Het-1 A and EC9706 lines. Additional correlation analysis of polyamine metabolism-related gene expression with drug response are future study direction on these cells. Furthermore, the inclusion of diverse treatment settings in validation experiments helps assess the model’s robustness under varying therapeutic conditions.

In summary, the role of RUNX3 in esophageal squamous cell carcinoma is not a single tumor suppressor, but has bidirectional regulatory potential, and its function may be influenced by tumor heterogeneity, metabolic status, and immune environment. In this study, RUNX3 not only acts as a key marker gene in the high-risk group, but may also mediate tolerance and sensitivity to some drugs, suggesting that it has dual potential as a prognostic marker and therapeutic target. Future studies may further explore its epigenetic status, signaling pathway interactions, and specific mechanisms in immune regulation to promote the precise clinical application of RUNX3 in ESCA.

Conclusion

In this study, we constructed an eight-gene prognostic model related to polyamine metabolism, which can effectively predict the survival outcome of ESCA patients and is closely related to immune microenvironment characteristics. The results suggest that polyamine metabolism may be involved in tumor progression by affecting immune regulation and has potential clinical application value.

Author contributions

Xiuli Cao: Conceptualization (lead); Formal analysis (lead); Methodology (lead). Yuanyuan Chen: Data curation (equal); Resources (equal); Writing- original draft (supporting). Tao Li: Project administration (supporting); Writing- original draft (supporting). Jinxing Wei: Funding acquisition (lead); Project administration (lead); Writing- review & editing (lead).

Funding

This work was supported by Science and Technology Development Project of Henan Provincial Science and Technology Department (Grant Number232102310306).

Data availability

The entire RNA-seq and scRNA-seq profile data and the clinical data of ESCC patients in this study come from Gene expression omnibus (GEO, GSE53622, GSE53624, GSE188900, https://www.ncbi.nlm.nih.gov/geo/) database. GO: https://www.geneontology.org/. KEGG: https://www.genome.jp/kegg/. MSigDB: https://www.gsea-msigdb.org/gsea/msigdb. For the processing of the original data, please refer to the methodology. For the processed data, please contact the corresponding author to obtain it.

Declarations

Ethics approval and consent to participate

Since the data came from a public database, no ethical statement is required for this study. The authors participated in different projects of the article respectively, see author contribution for details.

Consent for publication

All authors agree to publish.

Competing interests

The authors declare no competing interests.