Risk model based on adenylate uridylate (AU)-rich elements genes for assessing esophageal cancer prognosis and immune landscape Authors: Chaofan Mao, Mingjiang Huang, Jianyang Ding, Bin Huang, Xuyang Peng, Xuhui Wu Published: 2026 ABSTRACT BackgroundDespite their established role in post-transcriptional regulation, the functions of adenylate uridylate-rich elements genes (AREGs) in esophageal cancer (EC) remain incompletely understood. Therefore, this study aimed to identify critical AREGs and evaluate their potential as prognostic biomarkers for EC.MethodsTo delineate AREGs expression in EC, we conducted differential expression analysis using public databases. A prognostic model was subsequently constructed by integrating least absolute shrinkage and selection operator and multivariate Cox regression analysis. For immune microenvironment assessment, algorithms including CIBERSORT, ESTIMATE, and single sample gene set enrichment analysis (GSEA) were utilized. For functional characterization, GSEA, gene ontology enrichment analysis, and Kyoto Encyclopedia of Genes and Genomes enrichment analysis were performed. Finally, molecular subtyping was achieved through non-negative matrix factorization (NMF) clustering.ResultsA reliable prognostic signature based on eight AREGs was constructed to stratify patients. Those in the high-risk group experienced worse survival, a phenotype associated with cell cycle activation. In contrast, the low-risk group was associated with cell adhesion processes, and patients in this group might have derived greater clinical benefit from anti-programmed cell death ligand 1 immune checkpoint inhibitor therapy. NMF categorized all samples into three molecular subtypes with unique features.ConclusionsThis research investigates the intricate tumor microenvironment and molecular functions of AREGs in EC, offering novel perspectives on the pathological mechanisms and clinical treatment strategies for this disease. Highlight box Key findings • This study developed a robust predictive model that integrates eight genes associated with adenylate uridylate-rich elements (ARE). What is known and what is new? • ARE, typically composed of tandemly repeated AUUUA core motifs, are widely distributed in the 3′ untranslated regions of human genes and play critical regulatory roles in the development and progression of various malignant tumors. • The model combined the characteristics of ARE and identified three subtypes with significantly different molecular regulatory and immune profiles. What is the implication, and what should change now? • The eight genes related to ARE showed potential as biomarkers for predicting the prognosis of esophageal cancer. Future studies were recommended to utilize prospective cohort samples to validate these findings and to conduct further in vitro and in vivo experiments for more in-depth exploration. Introduction Globally, esophageal cancer (EC) is the 11th most frequently diagnosed malignancy and the 7th primary cause of cancer-related mortality (1). The two predominant histological subtypes of EC are esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma (EAC) (2). Due to atypical early symptoms, most patients are diagnosed at an advanced stage (3). Despite neoadjuvant chemoradiotherapy and subsequent surgery, which is the standard treatment for locally advanced resectable EC, approximately half of patients still develop recurrent or metastatic disease (4,5). Targeted therapy is characterized by its mechanism of action targeting specific molecular pathways, and this precision medicine approach is increasingly being applied in the treatment of EC (6). Studies indicate that targeted therapy can effectively slow EC progression and improve prognosis (7,8). However, the clinical application of targeted drugs in EC still faces numerous challenges, such as individual variations in efficacy and the management of drug-related adverse reactions (9-11). Therefore, developing safer, more effective targeted treatment strategies and novel targeted drugs has become a key direction in current EC research. In post-transcriptional regulation of gene expression, the 3' untranslated region (3'UTR) of messenger RNA (mRNA) serves as a pivotal hub where it can be recognized and bound by various trans-acting factors such as microRNAs, long non-coding RNAs, and RNA-binding proteins (12,13). Adenosine uridine (AU)-rich elements (ARE), as essential cis-acting short sequences within the 3'UTR, not only influence mRNA stability and translation efficiency but are also closely associated with mRNA degradation (14). ARE typically consists of repeated AUUUA motifs and is widely present in 5–8% of human gene 3'UTR regions (15,16). In recent years, studies have confirmed that ARE genes (AREGs) play a key regulatory role in various malignant tumors. For example, in non-small cell lung cancer, tristetraprolin (ZFP36) can effectively counteract chemotherapy resistance by inhibiting the accumulation of lipid droplets (17); in hepatocellular carcinoma, abnormal expression of AUF1 is associated with poor prognosis in early-stage patients, and it may exert a pro-tumorigenic effect by activating the Wnt signaling pathway (18,19). In addition, urokinase-type plasminogen activator within AREGs is expressed in various cancers to stimulate angiogenesis, supplying tumor cells with abundant nutrients and oxygen (20). The aforementioned studies suggest that AREGs play a significant role in various cancers; however, research on the development of systematic prognostic models for AREGs in EC and their relationship with the immune system remains limited and warrants further investigation. To elucidate the biological functions of AREGs in EC progression, this study utilized transcriptomic data from public database. We first identified prognostic AREGs and integrated them into a risk score model for personalized survival prediction. Furthermore, EC was categorized into distinct molecular subtypes by applying non-negative matrix factorization (NMF) clustering to AREGs expression patterns. A systematic comparison of immune cell infiltration across these subtypes uncovered an association between AREGs expression and tumor immune microenvironment (TIME) heterogeneity, suggesting a potential molecular basis for the varied treatment outcomes among EC patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0054/rc). Methods Data sources and processing The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Transcriptome data for esophageal carcinoma patients [The Cancer Genome Atlas-esophageal carcinoma (TCGA-ESCA), normal =13, tumor =185] were downloaded from TCGA (https://portal.gdc.cancer.gov/) database as the training set. Specifically, samples with a survival time of less than 30 days were excluded, and patients for whom survival information was missing were also excluded. Concurrently, somatic mutation and clinical data were downloaded. The GSE67269 cohort (tumor =44) was derived from the GEO (https://www.ncbi.nlm.nih.gov/geo/) database as an independent validation set. Based on research by Mariathasan et al. (21), we acquired a dataset of bladder urothelial carcinoma (BLCA) patients who demonstrated heterogeneous responses to programmed cell death ligand 1 (PD-L1) receptor inhibitors (Mvigor210). Additionally, 4,884 AREGs were collected from the Adenylate Uridylate-Rich Element (https://brp.kfshrc.edu.sa/ared) dataset. Preliminary screening of EC prognostic genes To determine AREGs as prognostic markers for EC, we first performed differential expression analysis on normal and tumor tissue samples from the training set using the limma package. Genes meeting the threshold criteria [|log fold change (FC)| >1, false discovery rate (FDR) <0.05] were selected as differentially expressed genes (DEGs). The overlap between these DEGs and the pre-defined AREG list identified differentially expressed AREGs. Subsequently, univariate Cox regression analysis (survival package; P<0.05) was applied to identify candidate prognostic genes for EC. Building and rigorously validating a prognostic model for EC To refine the list of candidate prognostic genes for EC, we performed sequential variable selection. Least absolute shrinkage and selection operator (LASSO) regression (glmnet package) was first applied, which were subsequently narrowed down by multivariate Cox proportional hazards analysis (survival package). A stepwise regression approach was then employed to construct the final prognostic model based on these selected genes. The corresponding risk score for each patient was calculated using the following formula: The predictive efficacy of the model was rigorously validated in the training and independent validation cohorts. Kaplan-Meier (K-M) analysis and receiver operating characteristic (ROC) curve assessment were conducted in parallel. The distribution of risk scores, further illustrated against actual survival outcomes, provided an intuitive visualization of the model’s stratification power. Developing a clinical nomogram based on patient characteristics By integrating prognostic gene signatures, clinical features, and risk scores, we conducted both univariate and multivariate Cox regression analyses to construct a predictive score for assessing EC risk. Calibration curves and decision curve analysis (DCA) were then applied to validate the clinical utility of this model in forecasting 1-, 3-, and 5-year survival outcomes. Gene enrichment analysis Two analytical approaches were utilized to delineate functional enrichments. First, gene set enrichment analysis (GSEA) was employed to uncover pathways enriched in the high- versus low-risk groups. Second, to decipher the biology of different EC molecular subtypes, independent Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted on their respective sets of DEGs, systematically outlining their unique functional attributes and pathway involvements. Immune infiltration analysis This study employed three computational algorithms—ESTIMATE, single sample GSEA (ssGSEA), and CIBERSORT—to characterize immune cell infiltration and functional activity across different patient subgroups. CIBERSORT, a support vector regression-based deconvolution method, utilizes gene expression data to ESTIMATE the relative fractions of diverse immune cell types in tissue samples. Specifically, the LM22 signature gene set was applied to expression matrices to quantify the abundances of 22 immune cell subsets and compare infiltration patterns between subtypes. The ESTIMATE algorithm assesses immune and stromal components in tumor tissues based on specific gene expression features, thereby inferring tumor purity. Here, ESTIMATE was used to calculate immune scores, stromal scores, and tumor purity for each sample, enabling subtype comparisons. ssGSEA evaluates the enrichment level of predefined gene sets within individual samples. Using immune-related gene sets, ssGSEA generated enrichment scores for immune cells and immune functions, facilitating the assessment of immune activity differences across subtypes. Tumor mutational burden (TMB) and drug sensitivity analysis We examined the mutation landscape within the EC training cohort and identified the top 20 genes most frequently mutated in both high- and low-risk groups, visualizing these mutations using a waterfall plot generated with the maftools package. Drug sensitivity was evaluated by estimating the half-maximal inhibitory concentration (IC50) of chemotherapeutic agents via the pRRophetic package. Prognosis-associated anticancer drugs were retrieved from the CellMiner database, and Spearman correlation analysis was performed to determine the relationship between gene expression levels and drug sensitivity. Molecular typing of EC This study integrated EC samples from the TCGA-ESCA and GSE67269 datasets, employing batch effect correction to eliminate cross-cohort variations and ensure comparability in subsequent analyses (sva package). Utilizing the expression profiles of prognostic genes, employing NMF methods, EC samples were further classified into distinct subtypes (NMF package). The parameters were set as follows: the number of clusters (k) ranged from 2 to 10, with 50 replicates (nrun =50) for each value of k; the “brunet” algorithm was used, and a random seed (seed =1,234) was set to ensure the reproducibility of the results. To validate the clustering results, the optimal number of clusters was determined by comprehensively evaluating the trend of the cophenetic correlation coefficient as a function of k. Further validation was performed using a consensus matrix heatmap for visualization. Subsequently, to delineate the biological heterogeneity across the identified subtypes, subsequent analyses involving gene enrichment and immune cell infiltration were performed. Statistical analysis All statistical analyses and graphical visualizations were conducted using R software (version 4.3.1). Intergroup differences between two groups were evaluated with the Wilcoxon signed‑rank test, whereas comparisons among three or more groups were performed using the Kruskal-Wallis test. For K-M survival curves, statistical significance between groups was determined by the log-rank test. Correlations were assessed using both Pearson and Spearman correlation coefficients, with a threshold of P<0.05 considered statistically significant. Results Preliminary screening of EC prognostic genes Based on the training set data, differential analysis between tumor and normal samples identified 1,514 DEGs (Figure 1A). Of these, 361 were characterized as AREGs (Figure 1B). From this refined set, univariate Cox regression analysis selected 16 genes with potential prognostic value for EC (Figure 1C, Table S1). Correlation analysis confirmed associations among these genes (Figure 1D), whose expression profiles differed substantially between normal and tumor groups (Figure 1E). Additionally, somatic mutation analysis identified ERO1B and NUP210 as the most frequently mutated genes, each exhibiting a mutation frequency of 28% (Figure 1F). Preliminary screening of EC prognostic genes. (A) Volcano plot of DEGs. (B) Intersection analysis of DEGs and AREGs. (C) HR assessment of differentially expressed AREGs. (D) Correlation analysis of EC candidate prognostic genes. (E) Comparison of expression levels of EC candidate prognostic genes between tumor and normal tissues. (F) Profiles of somatic mutations within the candidate prognostic gene set. *, P<0.05; **, P<0.01; ***, P<0.001. CI, confidence interval; DEGs, differentially expressed genes; EC, esophageal cancer; ESCA, esophageal carcinoma; HR, hazard ratio; TMB, tumor mutational burden. Construction and validation of an EC prognostic model To pinpoint the optimal prognostic genes for EC, this study initially conducted LASSO regression analysis on 16 candidate prognostic genes, which resulted in a selection of 14 genes (Figure 2A,2B, Table S2). Subsequent multivariate Cox regression analysis further narrowed these down to eight EC prognostic genes (Figure 2C, Table S3). Differential expression analysis demonstrated marked differences in the expression levels of these eight genes between the Normal and Tumor groups (Figure 2D). Finally, an EC prognostic model based on this gene set was developed using the following risk score formula: Construction and validation of an EC prognostic model. LASSO coefficient profiles (A) and cross-validation curve (B) for optimal λ selection. (C) Feature selection of prognostic genes via multivariate Cox regression. (D) Comparison of prognostic gene expression levels between different groups. (E) Comparison of prognostic gene expression levels across different risk groups. (F) Evaluation of model performance using the training set (left) and validation set (right) via ROC curves, K-M survival curves (G), and risk score distribution plots (H). **, P<0.01; ***, P<0.001. AUC, area under the curve; CI, confidence interval; EC, esophageal cancer; HR, hazard ratio; K-M, Kaplan-Meier; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic. Patients were categorized into high- and low-risk groups according to the median risk score. Expression patterns of the eight prognostic genes differed distinctly between these categories (Figure 2E). The model’s robust predictive power was evidenced by ROC curve analyses in both the training and independent validation sets (Figure 2F), and further corroborated by K-M analysis, which showed a significant overall survival (OS) advantage for the low-risk group (Figure 2G). Finally, the correlation between individual risk scores and actual survival outcomes was clearly depicted in a distribution plot, further highlighting the prognostic separation achieved by the model (Figure 2H). GSEA for high- and low-risk groups GSEA revealed a marked divergence in enriched pathways between the risk strata. Pathways such as protein folding and RNA metabolism were pronounced in the high-risk group (Figure 3A), whereas the low-risk group was characterized by terms related to cellular adhesion and developmental regulation (Figure 3B). This clear dichotomy in molecular functions suggests that the risk groups operate through disparate biological programs. GSEA for high- and low-risk groups. GSEA results of the high- (A) and low-risk (B) groups. GSEA, gene set enrichment analysis. Constructing a nomogram for EC prognosis Risk scores showed significant associations with key clinical features, being elevated in younger patients (≤65 years) and in advanced-stage (II/III) disease (Figure 4A,4B). Subsequently, univariate Cox regression analysis revealed that riskScore, stage, and node (N) were all significant prognostic predictors (Figure 4C). Multivariate Cox regression analysis further confirmed the validity of riskScore as an independent prognostic predictor (Figure 4D). To predict survival probability in EC patients, we developed a nomogram model integrating riskScore with key clinical parameters (Figure 4E). The model’s performance was rigorously validated, with calibration and decision curve analyses confirming its accuracy and clinical utility for predicting 1- and 3-year survival (Figure 4F,4G). Constructing a nomogram for EC prognosis. Comparison of risk scores grouped by different clinical characteristics: age (A) and stage (B). Identification of independent prognostic factors through univariate (C) and multivariate (D) Cox regression analysis. (E) Nomogram construction based on clinical characteristics and risk scores. (F) DCA results. (G) Nomogram calibration curve. ns, not significant; *, P<0.05; **, P<0.01. DCA, decision curve analysis; EC, esophageal cancer; f, full model; M, metastasis; N, node; T, tumor. Immune infiltration analysis A pronounced contrast in immune profiles was observed between the risk groups. Analyses using ssGSEA and CIBERSORT algorithms consistently revealed greater infiltration of various immune cell types—including iDCs, Natural killer cells (NK cells), resting mast cells, and monocytes—in the low-risk group (Figure 5A-5C). In contrast, the high-risk group was characterized by elevated expression of specific immune checkpoint genes (LGALS9, TNFRSF14, TNFSF15; Figure 5D). The translational impact of this immune dichotomy was evident in the IMvigor210 cohort, where the low-risk subgroup not only had a survival advantage but also displayed a superior objective response rate to anti-PD-L1 treatment (Figure 5E,5F). Accordingly, the risk score was significantly higher among non-responders (NR) compared to responders (R) (Figure 5G). Immune infiltration analysis. (A) Heatmaps generated by ssGSEA visualize the heterogeneity of immune cell infiltration across prognostic risk groups. Immune infiltration analysis based on ssGSEA (B) and CIBERSORT algorithm (C). (D) Comparison of immune checkpoint expression levels across different risk groups. (E) K-M survival curves for patients in different risk groups. (F) Response rates to PD-L1 antibody therapy in patients from different risk groups. (G) Comparison of risk scores between the NR group and R group. *, P<0.05; **, P<0.01; ****, P<0.0001. K-M, Kaplan-Meier; NR, non-responder; PD-L1, programmed death ligand-1; R, responder; ssGSEA, single sample gene set enrichment analysis. TMB and drug sensitivity analysis A comparative analysis revealed divergent genomic and pharmacogenomic profiles between the defined risk groups. High-risk status was correlated with a greater mutational burden, including a higher frequency of TP53 mutations (90% vs. 80%; Figure 6A,6B). Patients in the low-risk group might have been more sensitive to docetaxel, dasatinib, paclitaxel, and cisplatin; however, this conclusion needed to be further confirmed by in vitro and in vivo experiments and clinical studies (Figure 6C). Furthermore, correlation analysis revealed that LAMC2 expression levels were negatively correlated with sensitivity to barasertib, pipamperone, and crizotinib, while BMS-690514 exhibited the opposite trend (Figure 6D). TMB and drug sensitivity analysis. (A,B) The top 20 mutated genes in high-risk (A) and low-risk (B) cohorts. (C) Comparison of IC50 for four commonly used therapeutic agents among EC patients in different risk stratifications. (D) Correlation analysis between prognostic gene LAMC2 expression levels and multiple drug sensitivity. EC, esophageal cancer; IC50, half maximal inhibitory concentration; TMB, tumor mutational burden. Molecular subtypes of EC This study integrated EC samples from the TCGA-ESCA cohort and the GSE67269 dataset. After merging the data, principal components analysis (PCA) was performed (Figure 7A), and batch effects were corrected to eliminate cross-cohort variations and ensure comparability for subsequent analyses (Figure 7B). Based on the expression profiles of prognostic genes, cluster analysis identified three molecular subtypes of EC: cluster 1 (74 cases), cluster 2 (54 cases), and cluster 3 (93 cases) (Figure 7C). K-M survival analysis revealed significantly poorer survival in cluster 3 patients (Figure 7D). Gene expression analysis indicated significant expression differences among subgroups for CXCL8, DUSP2, LAMC2, NUP210, and RUNX1 (Figure 7E). Further GO and KEGG enrichment analysis of upregulated DEGs within each cluster revealed distinct molecular mechanisms: cluster 1 predominantly enriched pathways related to Tyrosine metabolism; cluster 2 enriched immune-related pathways such as IL-17 signaling and NOD-like receptor signaling; while cluster 3 was associated with focal adhesion (Figure S1A-S1C). Subsequent analysis of the immune profiles across clusters revealed that cluster 3, associated with the poorest prognosis, exhibited the lowest immune score, whereas cluster 2 showed the highest immune score; in contrast, cluster 1 had significantly greater tumor purity than cluster 2 but demonstrated an opposite trend in ESTIMATE scores (Figure 7F). Based on ssGSEA and CIBERSORT algorithms, immune profiling revealed that cluster 2 exhibited a more pronounced immune infiltration, encompassing most cell types by ssGSEA (Figure 7G,7H) and specifically higher levels of activated mast and dendritic cells by CIBERSORT (Figure 7I,7J). This comprehensive analysis led to the identification of three molecularly and immunologically heterogeneous EC subtypes. Molecular subtypes of EC. (A,B) PCA plot of merged TCGA-ESCA and GSE67269 cohorts and batch-calibrated PCA plot. (C) Sample clustering analysis based on prognostic gene expression signatures. (D) K-M survival analysis evaluating survival curves across different subgroups. (E) Comparison of prognostic gene expression levels across subgroups. (F) Comparison of Stromal Score, Immune Score, ESTIMATE score, and tumor purity across different subgroups. Heatmap (G) and boxplot (H) of immune cell and functional infiltration levels across subgroups, evaluated using the ssGSEA algorithm. Heatmap (I) and boxplot (J) of immune cell subtype infiltration levels across subgroups, analyzed using the CIBERSORT algorithm. ns, not significant; *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. EC, esophageal cancer; K-M, Kaplan-Meier; PCA, principal component analysis; ssGSEA, single sample gene set enrichment analysis; TCGA-ESCA, The Cancer Genome Atlas-esophageal carcinoma. Discussion This study systematically analyzed the potential role of AREGs in EC based on transcriptomic data from the TCGA and GEO databases. By integrating multiple bioinformatics analyses, we successfully demonstrated eight AREGs associated with EC prognosis and constructed a robust and effective prognostic risk model based on these findings. Furthermore, this study systematically evaluated TIME characteristics in EC patients, revealing associations between AREGs expression and infiltration levels of specific immune cell subsets. This study identified eight AREGs as prognostic genes for EC (CXCL8, DUSP2, LAMC2, NUP210, PGK1, PTGES3, SERPINB9, and RUNX1), which may play key roles in the initiation and progression of EC. C-X-C chemokine ligand 8 (CXCL8) is a multifunctional proinflammatory chemokine that not only acts as a neutrophil chemotactic agent but also plays a significant role in tumorigenesis (21). Studies indicate significantly elevated CXCL8 expression in EC tissues (22), which promotes EC tumor proliferation by mediating neural precursor cell expressed, developmentally down-regulated 9 (NEDD9) induced Granulocyte-like myeloid derived suppressor cells (G-MDSC) recruitment (23). Laminin subunit gamma 2 (LAMC2), encoding the γ-2 chain subunit of laminin, is overexpressed in ESCC, and knockdown of LAMC2 inhibits tumor cell migration capacity (24,25). Phosphoglycerate kinase 1 (PGK1), a key metabolic enzyme in glycolysis, participates in non-metabolic regulation of tumor progression beyond its metabolic function (26). In ESCC, PGK1 enhances β-catenin-dependent c-Myc transcription, thereby promoting tumorigenesis and migration (27). RUNX family transcription factor 1 (RUNX1), highly expressed in ESCC cells, forms a core regulatory circuit with GLI family zinc finger 2 (GLI2) and tumor protein p63 (TP63) to drive tumor progression (28,29). Additionally, lincRNA-uc002yug.2 promotes the assembly of RUNX1 and alternative splicing factors in the cell nucleus, leading to increased production of RUNX1a and reduced expression of CCAAT enhancer binding protein alpha (CEBPα), thereby promoting ESCC progression (30). The downregulation of dual specificity phosphatase 2 (DUSP2) contributes to nasopharyngeal carcinoma metastasis. This process is mediated by a dysregulated activation of the ERK signaling pathway and a consequent upregulation of matrix metalloproteinase 2 (MMP2) (31,32). Depletion of nucleoporin 210 (NUP210) in a mouse model of breast cancer inhibits lung metastasis formation (33). In colorectal cancer, NUP210 knockdown suppresses cell proliferation, invasion, and metastatic capacity both in vivo and in vitro (34). Prostaglandin E synthase 3 (PTGES3), acting as an oncogene mediating prostaglandin E2 (PGE2) expression (35), promotes prostate cancer metastasis through its co-expressed chaperone protein p23 (36). Serpin family B member 9 (SERPINB9) serves as an endogenous inhibitor of granzyme B (GzmB) (37,38). In a melanoma mouse model, its knockout effectively suppressed tumor cell proliferation (39). It is worth noting that previous studies have reported that CXCL8 (22), LAMC2 (40), PGK1 (27), and RUNX1 (28) may serve as potential biomarkers for EC. In contrast, direct evidence regarding the prognostic significance of NUP210, DUSP2, PTGES3, and SERPINB9 in EC remains insufficient; however, studies have shown that these genes are involved in tumor progression and the regulation of the immune microenvironment in other malignancies, making them worthy of further validation in EC cohorts. Furthermore, all eight of the aforementioned prognosis-associated genes exhibit significant overexpression in EC, suggesting that the development or application of corresponding inhibitors targeting these genes may represent a potential therapeutic strategy for inhibiting the malignant progression of EC. To further investigate the molecular heterogeneity among EC samples classified into different risk groups based on prognostic gene risk scores, this study performed GSEA. The results showed that the high-risk group was significantly enriched in pathways such as protein folding, regulation of cyclin-dependent protein kinase activity, 5'-end processing of RNA, and RNA capping, suggesting that tumor cells in this group may be in a malignant state characterized by high proliferation and abnormally active transcription and translation processes. This phenotype is closely associated with the abnormal regulation of AREGs. For example, amphiregulin (AREG) is highly expressed in colorectal cancer cells, and its knockdown significantly inhibits DNA replication efficiency and cell proliferation (41). Furthermore, in various cancers, the RNA-binding protein embryonic lethal abnormal vision gene (ELAVL) is often upregulated or exhibits abnormal nuclear/cytoplasmic distribution, while another ARE-binding protein, ZFP36 is downregulated; ZFP36 is known to promote the degradation of target mRNAs, whereas ELAVL typically stabilizes target transcripts and promotes their translation (42). The aforementioned imbalance in the expression of AREGs and their binding proteins may serve as a key molecular basis driving post-transcriptional dysregulation and malignant proliferation phenotypes in the high-risk group. In contrast, the low-risk group showed significant enrichment in adhesion-related pathways, such as epithelial cell-cell adhesion and focal adhesion assembly. Previous studies have revealed that the RNA-binding protein pumilio RNA binding family member 2 (PUM2) is highly expressed in prostate cancer and participates in the maintenance of adhesion functions by regulating the mRNA stability of integrin subunit alpha 3 (ITGA3) and other genes associated with local adhesion pathways (43). This mechanism suggests that the enrichment of adhesion-related pathways in the low-risk group may similarly be associated with the precise regulation of target mRNA stability by specific ARE-binding proteins. In summary, differences in post-transcriptional regulatory networks mediated by AREGs and their binding proteins can partially explain the significant differences in pathway enrichment between high- and low-risk groups regarding proliferative activity and adhesion function. Immune cells, as key components of the tumor microenvironment, play a crucial role in tumor progression and treatment response (44). NK cells exhibit higher infiltration levels in low-risk patients. NK cells are immune cells with potent tumor-killing capabilities, mediating tumor cell destruction through multiple mechanisms including the release of the perforin/granzyme pathway, secretion of interferon-gamma (IFN-γ) and tumor necrosis factor α (TNF-α), and mediation of antibody-dependent cellular cytotoxicity (ADCC) (45,46). Recent studies indicate that neoadjuvant immunochemotherapy improves clinical outcomes in EC patients by mediating the antitumor immune function of CD16+ NK cells (47). Immune checkpoints, as key molecules regulating immune responses and maintaining self-tolerance, play a pivotal role in tumor immune escape (48). This study found that CD44 is highly expressed in the low-risk group. CD44 plays a critical role in tumor progression by regulating biological processes such as cell adhesion, survival, and proliferation (49,50). Studies indicate that anti-CD44 monoclonal antibodies exhibit significant antitumor activity in EC mouse xenograft models (51). Furthermore, this study observed higher response rates to PD-L1 inhibitors in low-risk group patients compared to the high-risk group. In EC, the interaction between programmed death receptor 1 (PD-1) and PD-L1 induces T-cell exhaustion and apoptosis, enabling tumor cells to evade immune surveillance (52). Currently, PD-1/PD-L1 inhibitors such as nivolumab and pembrolizumab are widely used clinically, significantly extending OS in EC patients and profoundly altering first-line treatment strategies (53). Notably, this study found that patients in the low-risk group may exhibit a superior response to immunotherapy. TMB analysis revealed a higher TP53 mutation frequency in high-risk patients (90%) than in low-risk patients (80%). As one of the most critical tumor suppressor genes, TP53 encodes the p53 protein, which plays a critical role in inhibiting cell division and promoting apoptosis (54). TP53 mutations inactivate p53, leading to uncontrolled cell proliferation and invasive capacity, thereby promoting tumor progression (55). In ESCC, TP53 mutations induce ΔNp63 (a transcription regulator) upregulation, which activates the SRC/ERK/AKT signaling pathway via the ephrin B1-EPH receptor B4 (EFNB1-EPHB4) axis. This activation is crucial for the epithelial-mesenchymal transition (EMT) biological process (56,57). On the other hand, AUF1 binds to mRNA encoding phosphoinositide dependent kinase-1 (PDK1) and enhances its stability. PDK1 subsequently phosphorylates threonine 308 of protein kinase B (AKT), thereby activating AKT signaling and inducing EMT in various cancer cells (58). Whether AREGs and TP53 mutations regulate EMT in EC by affecting key prognostic genes warrants further investigation. Although this study advances a systematic understanding of AREGs in EC, its limitations point to clear research avenues. The clinical relevance of immune-heterogeneous molecular subtypes needs prospective validation. Findings based on public cohorts also necessitate confirmation in expanded, multi-center populations. Most importantly, the mechanistic basis of EC progression, particularly the crosstalk between prognostic signatures, AREGs, and EMT, demands further elucidation. Consequently, subsequent research should focus on: functional genetic studies using experimental models; high-resolution mechanistic mapping via single-cell and spatial omics; and the translation of biomarkers into diagnostic kits for improved early detection and risk stratification. Conclusions Beyond their association with the EC tumor microenvironment, AREGs appear to contribute to the regulation of EMT. Consequently, therapeutic targeting of these genes represents a potential novel strategy for EC treatment. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0054/prf