A risk scoring model for lung squamous cell carcinoma based on epithelial-mesenchymal transition-related genes: an integrative analysis of prognosis and immune infiltration characteristics

Abstract

Background

Despite expanding therapeutic options, the prognosis of lung squamous cell carcinoma (LUSC) remains poor. Immune checkpoint inhibitors benefit only a subset of patients, and epithelial–mesenchymal transition (EMT) has been implicated in invasion, metastasis, treatment resistance, and immune heterogeneity. Therefore, EMT-related biomarkers may offer improved risk stratification.

Aim

To identify differentially expressed EMT-related genes (DEEMTGs) in LUSC, construct an EMT-based prognostic signature, and evaluate its associations with the tumor microenvironment (TME), tumor mutational burden (TMB), and tissue-level expression patterns.

Methods

The Cancer Genome Atlas (TCGA) RNA-seq and clinical data were analyzed to obtain DEEMTGs. A prognostic model was built using LASSO and multivariable Cox regression. Survival performance was assessed via Kaplan–Meier, ROC, and Cox analyses. Immune infiltration (CIBERSORT), stromal/immune scores (ESTIMATE), and TMB were compared between risk groups. Exploratory immunohistochemistry (IHC; n = 8) provided orthogonal expression validation.

Results

A total of 1,651 DEEMTGs were identified, and a six-gene signature (GAB2, ALDOA, PCDHA3, TMEM92, ERH, IRS4) was established. The risk score independently predicted overall survival and corresponded to distinct TME patterns: low-risk tumors showed higher CD8⁺ T cells, activated CD4⁺ memory T cells, and naïve B cells, whereas high-risk tumors had more resting CD4⁺ memory T cells and M0 macrophages. TMB differences were nonsignificant. IHC provided directional protein-level support while acknowledging transcript-protein variability.

Conclusion

We developed a biologically interpretable EMT-based prognostic model that stratifies survival and reflects immune-microenvironment heterogeneity in LUSC. Larger, stage-balanced and immunotherapy-treated cohorts are needed to further validate its clinical utility.

Introduction

Lung cancer is a leading cause of cancer-related mortality worldwide, posing a formidable challenge to public health due to its high incidence and mortality rates. Despite advancements in diagnostic techniques and therapeutic strategies, the overall survival rate for lung cancer patients remains low, largely because most cases are diagnosed at an advanced stage (Siegel et al., 2026). Additionally, the heterogeneity of lung cancer further complicates treatment, posing significant challenges to the efficacy of personalized therapies (Gregorc et al., 2021). Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases and includes various subtypes, with lung adenocarcinoma, lung squamous cell carcinoma (LUSC), and large cell carcinoma being the most common. Although treatment options for NSCLC include surgery, radiotherapy, chemotherapy, and targeted therapy, outcomes are often unsatisfactory due to the inherent heterogeneity and complexity of the disease (Jiang et al., 2019; Xue et al., 2017). In particular, LUSC, characterized by unique histopathological features and clinical manifestations, faces numerous challenges in diagnosis, treatment, and prognosis. LUSC lacks well-defined molecular targets, which restricts the use of precision medicine compared with lung adenocarcinoma. Moreover, LUSC exhibits an inherent resistance to various existing treatments, further complicating management. Due to the absence of effective early screening methods, many LUSC patients are diagnosed at an advanced stage, missing the optimal window for surgical intervention and resulting in poor prognosis (Adachi et al., 2017; Sun et al., 2025). The molecular heterogeneity of LUSC also complicates the identification of universally effective therapeutic targets. Currently, LUSC treatment relies heavily on chemotherapy and radiotherapy, but resistance to these therapies and a high recurrence rate limit their efficacy (Lim & Ma, 2019). Therefore, exploring novel molecular biomarkers and therapeutic targets to improve the clinical management and prognosis of LUSC patients is urgently needed.

Epithelial–mesenchymal transition (EMT) is a dynamic and multifaceted biological process that plays a pivotal role in tumor progression, invasion, and metastasis (Huang, Hong & Wei, 2022; Nowak & Bednarek, 2021). During EMT, epithelial tumor cells progressively lose cell–cell adhesion and polarity while acquiring mesenchymal features that enhance motility, invasiveness, and survival under therapeutic stress (Cao et al., 2023). In LUSC, accumulating experimental and clinical evidence indicates that EMT contributes not only to local invasion and distant dissemination, but also to intrinsic and acquired resistance to conventional treatments, thereby complicating disease management (Du & Shim, 2016; Xiao et al., 2023). EMT-associated transcriptional reprogramming has been linked to altered apoptotic signaling, metabolic adaptation, and activation of alternative survival pathways, which together limit the durability of chemotherapy, radiotherapy, and emerging targeted or immunotherapeutic strategies (Xiao et al., 2023). Moreover, EMT is increasingly recognized as a key modulator of the tumor immune microenvironment, where it may promote immune evasion through suppression of antitumor immune responses and reshaping of immune cell infiltration patterns (Bergman et al., 2021; Mak et al., 2016). Emerging evidence further suggests that EMT status may influence responses to immune checkpoint blockade, although the extent and clinical relevance of this association appear to be tumor-type specific and remain to be fully clarified in LUSC (Yu et al., 2023). Collectively, these observations suggest that EMT represents a central biological axis underlying therapeutic resistance and poor clinical outcomes in LUSC, highlighting the need for systematic evaluation of EMT-related molecular signatures with prognostic and translational relevance.

Although EMT has been extensively investigated in LUSC, most existing studies have focused on individual EMT markers, specific signaling pathways, or descriptive associations with tumor aggressiveness. Comparatively fewer studies have systematically integrated EMT-related transcriptional alterations with prognostic modeling, immune microenvironment characteristics, and clinical outcome prediction in a unified analytical framework. In particular, the prognostic relevance of EMT signatures derived from tumor-adjacent normal contrasts and their interaction with tumor immunity and mutation burden in LUSC remains insufficiently characterized. Therefore, a comprehensive and integrative evaluation of EMT-related molecular signatures may provide additional insights into risk stratification and biological heterogeneity in LUSC.

Materials and Methods

Data collection

The data for this study was obtained from TCGA database. We collected RNA-seq and clinical data related to LUSC, comprising 497 LUSC samples and 51 normal control samples. To analyze EMT-related genes, we retrieved and downloaded 6,067 EMT-related genes from the GeneCards database. RNA-seq data was normalized and calculated using log2-transformed TPM (Transcripts Per Million) values. Duplicate samples were removed, retaining only one representative sample per patient. To ensure accuracy and consistency in analysis, we meticulously curated the clinical data, which included demographic information (e.g., age, gender), clinical staging of the tumor, and follow-up data (including survival time and status). Samples with missing survival time or survival time of zero days were excluded. During data integration, RNA-seq data and clinical data were matched to ensure all data used in the analysis was complete and consistent, supporting reliable bioinformatics analyses. This rigorous data preprocessing step lays a strong foundation for the reliability of the study’s conclusions.

Identification of differentially expressed genes

To identify differently expressed genes (DEGs) between LUSC samples and normal samples, we systematically analyzed RNA-seq data using the “DESeq2” package. First, a DESeq2 dataset was constructed, and the DESeq function was used to conduct differential expression analysis between LUSC and normal tissue RNA-seq data. The criteria for selecting DEGs were set as |log2(fold change)| > 0.5 and an adjusted p-value (padj) < 0.05 to ensure that the selected genes exhibited significant and biologically meaningful expression changes.

Following the identification of significantly differentially expressed genes, we further visualized these genes using the “ggplot2,” “pheatmap,” and “ggboxplot” packages. A volcano plot was used to display the overall differential expression landscape, presenting gene significance and fold-change variations intuitively. A heatmap was generated to show the expression patterns of significant DEGs between LUSC and normal tissues, enabling a clear comparison of expression differences. Additionally, box plots were used to illustrate the expression levels of certain key DEGs across different sample types, providing a direct understanding of the expression distribution of these genes in tumor and normal tissues. This series of analyses offers a comprehensive view of gene expression differences between LUSC and normal tissues, providing a solid foundation for subsequent functional research and clinical applications.

Identification of intersection genes

To identify key genes closely related to EMT, we integrated the list of DEGs with EMT-related genes to obtain intersecting genes, visually representing this intersection through a Venn diagram. First, EMT-related genes were retrieved from the GeneCards database, filtering for genes with a Relevance Score greater than 1 to ensure high relevance in EMT biological processes. Then, we conducted differential expression analysis on LUSC and normal samples using the “DESeq2” package, with selection criteria set at |log2(fold change)| > 0.5 and an adjusted p-value (padj) < 0.05 to ensure significant expression changes.

After obtaining the DEGs and EMT-related genes, we identified their intersection, yielding key genes that are significantly differentially expressed and highly associated with EMT in LUSC. These intersecting genes not only show notable expression differences but may also play crucial roles in EMT processes. To visually display the distribution of these intersecting genes, we used the “ggvenn” package to create a Venn diagram illustrating the overlap between DEGs and EMT-related genes. This analysis aids in pinpointing potential driver genes associated with EMT in LUSC, providing a basis for further research into their roles in tumor progression.

Construction and validation of DEEMTG-based risk score

To construct and validate a risk score model based on DEEMTGs, we first performed univariate Cox regression analysis to identify DEEMTGs significantly associated with patient survival, using a selection threshold of p < 0.05. This yielded 250 DEEMTGs significantly linked to survival. We then applied the Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis to further refine this gene set. LASSO regression introduces a penalty term to reduce the number of genes in the model, thereby decreasing complexity and preventing overfitting. Using cross-validation, we determined the optimal penalty parameter (λ) and selected the best set of genes for the model.

Based on the selected genes, we constructed a risk score model using a multivariate Cox regression model. The formula for calculating the risk score is as follows: Risk Score = (Expression of Gene A × Coefficient A) + (Expression of Gene B × Coefficient B) + … + (Expression of Gene N × Coefficient N). Based on the calculated risk scores, patients were classified into high-risk and low-risk groups, with the median risk score as the cutoff. To validate the predictive power and stability of the model, we conducted internal validation. Specifically, Kaplan-Meier survival analysis was used to assess survival differences between the risk groups, while receiver operating characteristic (ROC) curve analysis was employed to evaluate the predictive accuracy of the model. These methods allowed us to assess the application value of the risk score model in LUSC patients, demonstrating its effectiveness and reliability in predicting prognosis.

Survival analysis and model validation

Overall survival (OS) was the primary endpoint, measured from the date of diagnosis or surgery to death; patients alive at last follow-up were censored. The risk score (riskScore) was computed from the locked six-gene linear predictor, riskScore = ∑_iβ_i × Expr_i, and, unless otherwise stated, was treated as a continuous variable for inference and modeling. For visualization, stratification used the pre-specified median of the TCGA cohort; the external cohort (GSE30219) was scored with the fixed TCGA coefficients and the same threshold, without retraining or tuning. Within the discovery cohort, an exploratory cut point was derived via maximally selected rank statistics (survminer::surv_cutpoint) to illustrate discrimination; this cut point was not used for external validation or multivariable modeling to avoid optimism bias.

Survival curves were estimated by the Kaplan–Meier method and compared by the log-rank test. Discrimination at 1/2/3 years was assessed by time-dependent ROC (time ROC) with area under the curve (AUCs); overall discrimination was summarized by Harrell’s C-index (survival::surv Concordance). Multivariable Cox models included clinical stage (III/IV vs I/II), T category (T3/4 vs T1/2), M category (M1 vs M0), age, and sex, reporting HRs, 95% CIs, and P values; the proportional-hazards assumption was tested using Schoenfeld residuals. A nomogram (rms) combining clinical covariates with riskScore was constructed, with internal calibration via 1,000 bootstrap resamples and calibration curves at 2/3/5 years.

To evaluate incremental value, on common complete cases we compared a “clinical baseline model” (Stage/T/M/age/sex) with a “clinical + riskScore” Cox model using a likelihood-ratio test (LRT; 2ΔlogL) and the Akaike information criterion (AIC; lower indicates better fit). Where appropriate, NRI/IDI and decision-curve analysis were provided as sensitivity assessments. Unless otherwise noted, tests were two-sided with α = 0.05; for multiple comparisons (e.g., immune infiltration/TME panels), false discovery rate was controlled by the Benjamini–Hochberg procedure with q values reported. All analyses were performed in R (v4.2.3) using survival, survminer, time ROC, and rms.

Immune infiltration and immune scoring analysis

To explore the relationship between the risk score model and the tumor immune microenvironment, we applied the CIBERSORT algorithm to estimate the proportion of immune cell infiltration within tumor samples. First, we extracted gene expression data from high- and low-risk patients and used the CIBERSORT tool to perform deconvolution analysis, estimating the proportions of various immune cell types in tumor samples. By comparing immune cell infiltration between the two groups and using the Wilcoxon rank-sum test to assess the significance of differences, we examined the differences in immune cell infiltration characteristics between the high- and low-risk groups.

In addition, we computed each patient’s Immune Score and Stromal Score to further characterize the distribution of immune and stromal components within the tumor microenvironment. These scores, derived from the ESTIMATE algorithm, quantify the relative proportions of immune and stromal cells in tumor samples. Differences between high- and low-risk groups were assessed using Wilcoxon rank-sum tests. For multiple comparisons across the 22 CIBERSORT immune-cell types and the three ESTIMATE-derived scores, p values were adjusted by the Benjamini–Hochberg procedure to control the false discovery rate, with q < 0.05 (two-sided Wilcoxon) considered significant.

To visually represent immune infiltration characteristics, we processed mRNA expression data from the TCGA project by removing duplicate patient IDs and transposing the data into a data frame format. Using CIBERSORT for deconvolution analysis, we extracted and normalized the resulting immune cell infiltration proportion data. We then transposed and scaled the data and generated a heatmap to illustrate immune cell infiltration characteristics across patients in different risk groups. In the heatmap, different colors represent immune cell infiltration levels, with blue indicating low infiltration and red indicating high infiltration. Through these analyses and visualizations, we unveiled the relationship between the risk score model and the tumor immune microenvironment, providing insights into immune cell infiltration differences between the high- and low-risk groups.

Immune therapy response analysis

To explore the putative association between the EMT-based risk score and immunotherapy outcomes, we conducted a cross-tumor, hypothesis-generating analysis in the urothelial carcinoma cohort IMvigor210 (not considered external validation owing to histologic and therapeutic differences from LUSC). Fixed coefficients trained in the TCGA-LUSC cohort were applied to IMvigor210 to compute risk scores without retraining or tuning. Kaplan–Meier curves were then plotted, with descriptive log-rank comparisons, to generate hypotheses rather than confirmatory inferences.

For visualization, an internal optimal cut point was identified within IMvigor210 using surv_cutpoint, partitioning samples into high- and low-risk groups for display. We explicitly acknowledge potential leakage and optimism because the split derives from the same evaluation dataset; results are therefore exploratory and not used for extrapolative claims. All IMvigor210 findings represent cross-tumor inference and require prospective validation in real-world NSCLC/LUSC immunotherapy cohorts using preregistered, fixed coefficients and thresholds.

Mutation analysis

To delineate the mutational architecture of LUSC, we imported tumor mutation data matched to the expression cohort and stratified patients into high- and low-risk groups by the risk score. Using maftools, we generated oncoplots for each stratum, depicting the 15 most frequently mutated genes to contrast dominant mutational spectra. Tumor mutational burden (TMB) was computed per patient and visualized with violin/box plots; between-group differences were assessed using two-sided Wilcoxon rank-sum tests. No fixed-threshold dichotomization was used for inference. Proportion bar charts based on cohort quantiles were produced solely for illustrative purposes and were not subjected to significance testing.

KEGG pathway enrichment analysis of differentially expressed genes

To investigate the biological functions of DEGs in LUSC, we conducted a KEGG pathway enrichment analysis. First, significantly upregulated and downregulated DEGs were converted to ENTREZ IDs for matching with the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. Using the enrichKEGG function from the clusterProfiler package, we performed enrichment analysis and selected significantly enriched pathways with an adjusted p-value (padj) <0.05. This analysis aimed to uncover the potential roles of these DEGs in key biological pathways and to explore their mechanisms in LUSC. Significant pathways were visualized using dotplot and barplot functions, providing an intuitive display of pathway enrichment. Additionally, the pathview package was used to further visualize gene expression patterns within specific pathways, illustrating gene interactions within these pathways.

Gene ontology enrichment analysis of differentially expressed genes

To explore the biological functions of DEGs in depth, we conducted a gene ontology (GO) enrichment analysis. First, we selected significantly upregulated and downregulated genes from the DEG data and converted these genes to their corresponding ENTREZ IDs. Using the enrichGO function from the clusterProfiler package, we analyzed the enrichment of DEGs in the three GO categories: Cellular Component, Biological Process, and Molecular Function. The selection criterion was set at an adjusted p-value (padj) <0.05 to ensure the statistical significance of the identified GO terms. Visualization of the results was done with bar plots and dot plots, displaying the degree of enrichment, significance, and the number of genes involved in each GO term. Specifically, bar plots illustrated the enrichment across different GO categories, while bubble plots further displayed the significance and gene count for each GO term. Through these analyses, we gained a comprehensive understanding of the distribution of DEGs across various biological processes, cellular components, and molecular functions, providing crucial insights into the potential molecular mechanisms of LUSC.

Consensus clustering analysis

To identify tumor subgroups and assess their association with patient prognosis, we performed consensus clustering. Prognostically significant genes were first selected and then clustered using the ConsensusClusterPlus package, with the maximum number of clusters set to four. Dimensionality reduction was conducted with t-distributed stochastic neighbor embedding (t-SNE) and principal component analysis (PCA), and cluster distributions were visualized with ggplot2.

Cluster assignments were subsequently integrated with clinical data, and survival differences among subgroups were evaluated using Kaplan–Meier curves and log-rank tests. The optimal number of clusters was determined by the Proportion of Ambiguous Clustering (PAC) metric to ensure stability and robustness.

In parallel, we systematically reviewed prior EMT-based prognostic models in NSCLC. Key studies are summarized in Table S1, standardizing comparisons across cohorts, feature scales, statistical methodologies, endpoints, and validation strategies.

Construction and validation of the prognostic model

We compiled clinical data from the The Cancer Genome Atlas (TCGA) database, including age, gender, and staging information. First, we used univariate and multivariate Cox regression analyses to identify variables with independent prognostic significance, ultimately constructing a Cox regression model comprising risk score, T stage, and M stage. A nomogram was created to predict 2-year, 3-year, and 5-year survival rates. Subsequently, calibration curves were generated through repeated simulations (1,000 iterations) to validate the predictive accuracy of the model. The results indicated high concordance between predicted and actual survival rates at different time points, further demonstrating the reliability of the prognostic model.

Patients and samples

This study included patients diagnosed with LUSC who were treated at Huabei Oil Administration General Hospital between December 2022 and June 2024. Clinical and pathological information (such as age at surgery, gender, tumor size, pathology, TNM staging, Fuhrman grade, and all necessary follow-up information) was collected for all participants and subsequently evaluated. Cancer and adjacent tissue samples were obtained from eight stage III-IV LUSC patients for immunohistochemical (IHC) analysis. Patient information is provided in Table S2.

This study was approved by the Ethics Committee of the Huabei Petroleum Administration Bureau General Hospital in Hebei Province, China, with approval number: hbyyec-wz-03. Written informed consent was obtained from all patients or, where applicable, from their legally authorized representatives prior to the use of archived tissue samples for research purposes.

Immunohistochemistry

Immunohistochemical staining was performed using the streptavidin-peroxidase method. Primary antibodies used included ALDOA (1:300; Rabbit, Santa Cruz, Dallas, TX, USA), ERH (1:150; Rabbit, Sigma, St. Louis, MO, USA), PCDHA3 (1:200; Rabbit, Santa Cruz, Dallas, TX, USA), TMEM92 (1:500; Rabbit, Sigma, St. Louis, MO, USA), IRS4 (1:100; Rabbit, Invitrogen, Carlsbad, CA, USA), and GAB2 (1:100; Rabbit, Invitrogen, Carlsbad, CA, USA). Quantitative analysis of IHC staining was conducted using Image-Pro Plus (IPP) software by measuring the integrated optical density (IOD) value of each slide, which allowed for precise assessment of gene expression levels across different tissue types, ensuring the accuracy and reliability of quantitative results.

Results

Identification of DEGs and DEEMTGs in LUSC

To delineate EMT-related dysregulation in LUSC, we conducted differential-expression and intersection analyses comparing tumors with adjacent normal tissues. We identified 9,745 DEGs, and their overlap with the EMT gene set yielded 1,651 DEEMTGs after multiple-testing correction, indicating pervasive EMT-linked transcriptional abnormalities in LUSC. Representative genes showed concordant directional shifts in tumors, implicating the EMT axis in shaping LUSC phenotypes. This compendium provided a stable foundation for deriving a parsimonious risk signature and for pathway characterization (Figs. 1A–1D).

Figure 1. Workflow for identifying differentially expressed EMT-related genes.

(A) Heatmap displaying the top 20 upregulated and top 20 downregulated differentially expressed genes (DEGs) across all samples, with red indicating upregulated genes and blue indicating downregulated genes. (B) Volcano plot showing the results of differential expression analysis based on the TCGA database, identifying a total of 19,512 genes, including 4,233 upregulated genes (red), 2,972 downregulated genes (blue), and 12,307 genes with no significant differential expression (gray). (C) Venn diagram illustrating the intersection of 19,512 DEGs with 3,376 EMT-related genes (EMT-Genes), identifying 1,651 differentially expressed EMT-related genes (DEEMTGs). (D) Box plot displaying the median and interquartile range of expression for 15 specific genes in normal and tumor tissues. DEGs: differentially expressed genes; EMT_Genes: EMT-related genes. Statistical significance is denoted by asterisks, where an asterisk (*) indicates p < 0.05, two asterisks (**) indicate p < 0.01, and three asterisks (***) indicate p < 0.001.

Construction and validation of the DEEMTG risk score

Starting from DEEMTGs associated with overall survival, we applied LASSO followed by multivariable Cox regression and finalized a six-gene panel: four risk genes (GAB2, ALDOA, PCDHA3, TMEM92) and two protective genes (ERH, IRS4), each with statistically significant coefficients (Fig. 2). Median-based stratification in TCGA revealed markedly poorer OS in the high-risk group (log-rank p < 0.0001; Fig. 3A), with concordant separation observed in the external GSE30219 cohort (p = 0.029; Fig. 3B). Time-dependent ROC analyses yielded AUCs of 0.61/0.67/0.71 at 1/2/3 years in TCGA and 0.92/0.74/0.68 in GSE30219 (Figs. 3C and 3D), indicating a transferable, moderate discriminatory capacity within the same histologic context.

Figure 2. Validation of differentially expressed EMT-related genes.

(A) Univariate Cox regression analysis identified 250 differentially expressed EMT-related genes. The forest plot displays genes with p-value < 0.01, showing their hazard ratios (HR) and 95% confidence intervals (CI). (B) LASSO coefficient distribution plot based on the optimal λ value. (C) Plot showing the selection of optimal parameters (lambda) in the model, with the minimum criterion indicated at the optimal value. (D) Forest plot showing the hazard ratios (HR) and 95% confidence intervals (CI) of significant genes identified by multivariate Cox regression analysis, including GAB2, ALDOA, ERH, IRS4, PCDHA3, and TMEM92.

Figure 3. Survival analysis.

(A) Kaplan–Meier survival curves for high-risk and low-risk patients in the overall cohort, showing significantly lower survival rates in the high-risk group (p < 0.0001). (B) Kaplan–Meier survival curves for high-risk and low-risk patients in the validation cohort, with significantly lower survival rates in the high-risk group (p < 0.05). (C) ROC curves for 1-year, 2-year, and 3-year survival rates based on risk scores in the overall cohort, with AUC values of 0.61, 0.67, and 0.71, respectively. (D) ROC curves for 1-year, 2-year, and 3-year survival rates based on risk scores in the validation cohort, with AUC values of 0.92, 0.74, and 0.68. AUC: Area under the receiver operating characteristic curve.

At the individual level, the distribution of risk scores aligned monotonically with survival outcomes (Figs. 4A–4D), and expression patterns of the six genes corroborated the high- versus low-risk stratification (Figs. 4E and 4F).

Figure 4. Correlation analysis between survival time and risk score.

(A) Correlation between patient risk scores and LUSC mortality in the overall cohort, showing higher risk scores associated with greater risk. (B) Correlation between patient risk scores and LUSC mortality in the test cohort, indicating higher risk scores associated with increased risk. (C) Scatter plot showing the correlation between survival time and risk scores in LUSC patients in the overall cohort. (D) Scatter plot showing the correlation between survival time and risk scores in LUSC patients in the validation cohort. (E) Heatmap displaying the expression of six genes in the overall cohort, with IRS4 and ERH highly expressed in the low-risk group, while ALDOA, TMEM92, GAB2, and PCDHA3 show higher expression in the high-risk group. (F) Heatmap displaying gene expression in the validation cohort, with IRS4 and ERH highly expressed in the low-risk group, and ALDOA, TMEM92, GAB2, and PCDHA3 showing higher expression in the high-risk group.

Orthogonal protein-level corroboration (IHC)

Following identification of the six-gene panel from transcriptomic data, we performed IHC on an independent set of advanced LUSC tissues (n = 8, stage III–IV) to obtain cross-level histological evidence. This analysis aimed to verify whether tumor-versus-adjacent expression directions for panel genes were concordant with transcriptomic findings; it was not intended to assess prognostic performance. Overall, GAB2 and ALDOA exhibited higher IOD in tumors, PCDHA3 showed attenuated membranous staining (lower IOD) in tumors, ERH demonstrated higher nuclear positivity in tumors, IRS4 was reduced in tumors, and TMEM92 showed no significant difference (Figs. 5, 6, 7, 8, 9, and 10; raw individual values in Table S4). Given the small, late-stage cohort and potential influences of protein–RNA discordance, antibody and chromogen conditions, tumor purity, and spatial heterogeneity, the IHC findings are positioned as directional evidence of expression.

Figure 5. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of GAB2 in LUSC and adjacent normal tissues. The left panel shows GAB2 staining in tumor tissue (A2 group), while the right panel shows GAB2 staining in normal tissue (B2 group). The magnified region (bottom) highlights areas of stronger staining in tumor tissue. A bar chart displays the protein expression level of GAB2 (IOD value) in tumor and normal tissues, indicating significantly higher GAB2 expression in tumor tissues (p = 0.022).

Figure 6. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of ALDOA in LUSC and adjacent normal tissues. The left panel shows ALDOA staining in tumor tissue (A2 group), and the right panel shows ALDOA staining in normal tissue (B2 group). The magnified region (bottom) emphasizes stronger staining in tumor tissue. A bar chart displays ALDOA protein expression levels, showing significantly higher expression in tumor tissues compared to normal tissues (p < 0.001).

Figure 7. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of PCDHA3 in LUSC and adjacent normal tissues. The left panel shows PCDHA3 staining in normal tissue (B2 group), while the right panel shows its expression in tumor tissue (A2 group). The magnified region (bottom) highlights intense membrane staining of PCDHA3 in tumor cells. A bar chart shows that PCDHA3 expression is significantly lower in tumor tissue compared to normal tissue (p = 0.005).

Figure 8. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of ERH in LUSC and adjacent normal tissues. The left panel shows ERH expression in tumor tissue (A2 group), and the right panel shows expression in normal tissue (B2 group). The magnified region (bottom) highlights strong nuclear staining of ERH in tumor cells. A bar chart displays ERH protein expression, with significantly higher levels in tumor tissue (p = 0.002).

Figure 9. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of IRS4 in LUSC and adjacent normal tissues. The left panel shows IRS4 expression in tumor tissue (A2 group), and the right panel shows its expression in normal tissue (B2 group). The magnified region (bottom) highlights strong perinuclear staining of IRS4 in tumor tissue. A bar chart shows that IRS4 expression is significantly lower in tumor tissue than in normal tissue (p = 0.006).

Figure 10. Expression analysis of GAB2, ALDOA, PCDHA3, TMEM92, IRS4, and ERH in LUSC and adjacent normal tissues.

Expression of TMEM92 in non-small cell lung cancer (NSCLC) and corresponding normal tissues. The left panel shows TMEM92 expression in tumor tissue (A2 group), and the right panel shows expression in normal tissue (B2 group). The bar chart indicates no significant difference in TMEM92 expression between tumor and normal tissues (p = 0.798).

Role of the risk score in the immune microenvironment and prognosis

In the TCGA cohort, CIBERSORT-derived immune composition revealed risk score–aligned differences in the tumor microenvironment: low-risk tumors harbored higher infiltration of naïve B cells, CD8⁺ T cells, and activated CD4⁺ memory T cells, whereas high-risk tumors exhibited higher proportions of resting CD4⁺ memory T cells and M0 macrophages (Fig. 11). Across 22 immune-cell types and 4 TME metrics (26 features total), six remained significant after Benjamini–Hochberg correction (q < 0.05). ESTIMATE further showed higher StromalScore, ImmuneScore, and ESTIMATEScore in the low-risk group, indicating a more immune-activated/enriched microenvironment, while the high-risk group displayed a relatively suppressed or non-activated milieu—directionally concordant with the survival separation.

Figure 11. Immune Infiltration and tumor microenvironment scoring analysis based on risk score.

(A) Heatmap displaying the infiltration levels of different immune cells in high-risk (pink) and low-risk (blue) groups. Colors range from blue (low expression) to red (high expression), indicating variations in infiltration levels. (B) Box plot showing the distribution of different immune cell infiltration levels in high-risk (orange) and low-risk (blue) groups. Statistical significance is denoted by asterisks, where an asterisk (*) indicates p < 0.05, two asterisks (**) indicates p < 0.01, and three asterisks (***) indicates p < 0.001. (C) Comparison of stromal score (StromalScore), immune score (ImmuneScore), and ESTIMATE score (ESTIMATEScore) between high-risk and low-risk groups, showing significant differences across all three scores with p < 0.001.

For clinically practical stratification, an optimal risk-score cut point of 0.07 was identified in the training set using maximally selected rank statistics, partitioning patients into high- and low-risk groups. Under this threshold, Kaplan–Meier curves demonstrated a clear separation in overall survival (log-rank p = 0.00026; Fig. 12), supporting stable prognostic stratification at the population level. This threshold was subsequently carried forward as a pre-specified strategy, while discrimination reporting primarily relied on time-dependent AUCs from the continuous risk score to minimize optimism from threshold selection.

Figure 12. Risk score distribution and Kaplan–Meier survival analysis.

(A) Risk score distribution plot showing patient risk scores, with the low-risk group represented by the blue curve and the high-risk group by the red curve. The optimal cutoff point (Cutpoint = 0.07) is used to stratify patients into high- and low-risk groups. (B) Kaplan–Meier survival curve showing the survival probability of patients based on risk score groups. Patients in the high-risk group (red) have significantly lower survival rates compared to those in the low-risk group (blue), with a log-rank test p-value of 0.00026, indicating statistically significant survival differences between the two groups.

Mutation characteristics analysis by risk group

In the TCGA-LUSC cohort, ≥96% of samples in both groups harbored somatic mutations (Figs. 13A and 13B). The overall mutational spectra were similar across strata: TTN and TP53 were the most frequently mutated genes, followed by MUC16, CSMD3, and RYR2; missense variants predominated, with frameshift deletions and nonsense mutations next most common. No clear rank-order shifts in top-gene mutation frequencies were observed between groups, suggesting that risk stratification does not reflect a fundamentally different “driver” composition.

Figure 13. Mutation gene analysis and tumor mutation burden distribution by risk group.

(A, B) The most common mutation types in both low-risk and high-risk groups were missense mutations, frame shift deletions, and nonsense mutations. The top five mutated genes in both groups were TTN, TP53, MUC16, CSMD3, and RYR2. (C) Violin plot showing the TMB distribution in high-risk and low-risk groups (p = 0.052). (D) Bar plot displaying the proportion of high and low TMB within high-risk and low-risk groups, revealing TMB heterogeneity between different risk groups.

When TMB was compared on a continuous scale, the median was modestly higher in the high-risk group but did not reach statistical significance (two-sided Wilcoxon, p = 0.052; Fig. 13C). For illustration, we present proportion plots based on cohort quantiles (Fig. 13D); these are purely schematic and are not used for dichotomous testing or inference. Overall, the prognostic information carried by the risk score appears not to be driven by total mutational burden, but more plausibly by upstream transcriptional and immune programs—consistent with the aforementioned TME differences. Prospective re-evaluation of the TMB–riskScore relationship in independent cohorts with harmonized sequencing and normalization pipelines is warranted.

All statistical comparisons employed two-sided Wilcoxon tests. No fixed-threshold dichotomization of TMB was performed in this study.

Functional enrichment analysis

KEGG enrichment of DEEMTGs revealed significant overrepresentation of pathways implicated in tumor motility and immune regulation (BH-FDR, q < 0.05), with leading terms including neuroactive ligand–receptor interaction, cytokine–cytokine receptor interaction, hematopoietic cell lineage, regulation of the actin cytoskeleton, and calcium signaling (Figs. 14A and 14B). Collectively, these pathways converge on receptor–ligand signaling, cell adhesion/cytoskeletal remodeling, and heightened immune responses—biologically consistent with the immune-infiltration differences observed across risk strata.

Figure 14. Functional pathway and enrichment analysis in LUSC.

(A) Bar plot displaying the top 25 enriched pathways in the Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis. (B) Bubble plot showing the top 25 enriched pathways in KEGG analysis. (C) Bar plot displaying the top 10 enriched pathways in the Gene Ontology (GO) analysis. (D) Bubble plot showing the top 10 enriched pathways in GO analysis.

GO analysis further substantiated these trends from a functional perspective (Figs. 14C and 14D). Significant terms were enriched for positive regulation of the MAPK cascade (BP), receptor–ligand activity and signal receptor activator activity (MF), and passive transmembrane transporter activity related to membrane receptors/transport (q < 0.05). Overall, the functional profile of DEEMTGs indicates that signal transduction and cytoskeletal/adhesive processes are key nodes of EMT-related dysregulation in LUSC.

To visualize the principal signaling axes, we mapped differentially expressed genes onto KEGG pathway diagrams (Figs. 15, 16, and 17). TGF-β, Wnt, and PI3K–Akt pathways were prominently enriched—each tightly linked to EMT induction and tumor progression—and aligned with the putative functions of the six-gene panel: for example, GAB2/IRS4 associate with the PI3K–Akt/ERK axis, PCDHA3 with cell adhesion, and ALDOA with glycolysis/metabolic rewiring. These findings provide independent support for the model’s biological interpretability and generate testable mechanistic hypotheses.

Figure 15. Role of key signaling pathways in EMT and LUSC prognosis.

Illustration of the TGF-beta signaling pathway, showing its regulation of gene expression through Smad proteins, influencing cell proliferation, differentiation, and EMT.

Figure 16. Role of key signaling pathways in EMT and LUSC prognosis.

Illustration of the Wnt signaling pathway, demonstrating how it mediates gene expression through β-catenin, impacting cell polarity, proliferation, and cancer invasion.

Figure 17. Role of key signaling pathways in EMT and LUSC prognosis.

Illustration of the PI3K-Akt signaling pathway, highlighting its role in promoting cell survival and proliferation and its importance in cancer progression and drug resistance.

Through pathway visualization in the KEGG database, we further illustrated the expression changes of these genes within enriched pathways. These aberrant expressions may significantly impact the EMT process and prognosis of LUSC patients. Notably, we focused on signaling pathways closely associated with EMT and LUSC prognosis, such as the TGF-β signaling pathway, Wnt signaling pathway, and PI3K-Akt signaling pathway (Figs. 14A–14C). These pathways play crucial roles in regulating cell proliferation, differentiation, survival, and invasiveness, thereby affecting tumor development and progression. These findings provide essential clues for further research on the specific mechanisms of these pathways in LUSC and offer directions for subsequent experimental validation.

Clustering of gene expression subtypes and survival analysis

To delineate expression-level heterogeneity and its prognostic implications, we performed consensus clustering on survival-associated genes. Based on resampling stability metrics (e.g., PAC), k = 2 was selected as the optimal cluster number. The two expression subtypes were well separated in reduced-dimensionality space (Figs. 18A and 18B) and exhibited significantly different overall survival (Kaplan–Meier, p = 0.012629; Fig. 18C). These findings indicate that transcriptional heterogeneity in LUSC carries clear prognostic information, providing a robust framework for subsequent stratified functional and clinical analyses.

Figure 18. Visualization and survival analysis of clustering subtypes.

(A) t-SNE plot showing the distribution of two clustering subtypes in low-dimensional space. Red and blue represent the two different subtypes, with clear separation indicating effective clustering. (B) PCA plot showing the separation of clusters. (C) Kaplan–Meier survival curve displaying survival differences between the two subtypes (p = 0.012629).

Gene set enrichment analysis of key gene sets in LUSC

Using a rank-ordered transcriptome based on the high/low-risk phenotype, gene set enrichment analysis (GSEA) revealed that, at the GO level, terms such as negative regulation of epithelial-to-mesenchymal transition, negative regulation of immune system process, and regulation of B-cell proliferation were significantly positively enriched (Fig. 19A; BH-FDR, q < 0.05). At the KEGG level, the B-cell receptor signaling pathway was positively enriched, whereas the p53 signaling pathway was inversely enriched (Fig. 19B; exemplar q ≈ 0.01 and 0.023). These patterns echo the TME findings: the low-risk group aligns with a B-cell/adaptive immune–activated phenotype, while the high-risk group exhibits enhancement of p53 stress/damage pathways. Concordant enrichment of “negative regulation of EMT” with better prognosis further suggests that an EMT suppression–immune activation axis may underlie the observed stratification.

Figure 19. GSEA enrichment analysis of key gene sets in LUSC.

(A) GSEA enrichment curves for three significant GO biological processes in LUSC. The peak of each curve indicates the enrichment level of each gene set in the ranked gene list, with higher ES reflecting significant enrichment. (B) GSEA enrichment curves for two important KEGG signaling pathways in LUSC, with upward slopes indicating enrichment in highly expressed genes, and downward slopes indicating enrichment in lowly expressed genes.

Independent prognostic analysis and correlation with clinical characteristics based on risk score

In independent prognostic analyses, univariable Cox models identified several clinical variables significantly associated with overall survival, with the risk score exhibiting the largest effect size (HR = 2.72, 95% CI [2.06–3.59], p < 0.0001). Clinical stage (HR = 1.57, 95% CI [1.14–2.16], p = 0.0060), T category (HR = 1.70, 95% CI [1.23–2.36], p = 0.0010), and M category (HR = 3.11, 95% CI [1.27–7.61], p = 0.0130) were likewise adverse factors, whereas N category and sex were not significant (Fig. 20). In a multivariable model including stage (III/IV), T3/4, M1, age, and sex, the risk score remained independently prognostic, indicating incremental stratification beyond conventional clinical variables.

Figure 20. Independent prognostic analysis of the EMT-related risk score and clinicopathological factors in LUSC.

Forest plots showing hazard ratios (HRs) and 95% confidence intervals derived from univariate and multivariate Cox regression analyses. The EMT-related risk score and selected clinicopathological variables were evaluated for their independent associations with overall survival in patients with lung squamous cell carcinoma.

A nomogram integrating the risk score with T/M categories, age, and sex provided individualized 2-, 3-, and 5-year OS estimates (Fig. 21). Internal validation with 1,000 bootstrap resamples showed calibration curves at 2, 3, and 5 years closely tracking the ideal line (Figs. 22A–22C), indicating good agreement between predicted probabilities and observed outcomes. Collectively, the six-gene risk score is independently associated with survival and, when combined with baseline staging, enhances individualized risk assessment and clinical utility in LUSC.

Figure 21. Nomogram prediction model integrating the EMT-related risk score and clinicopathological variables.

A nomogram combining clinicopathological factors and the EMT-related risk score to predict 2-, 3-, and 5-year overall survival probabilities in patients with lung squamous cell carcinoma. The model provides an individualized quantitative tool for prognostic assessment and clinical decision-making.

Figure 22. Calibration curves.

(A) Calibration curve showing the predicted vs. actual 2-year survival rates. The blue line represents the observed survival rate, and the gray line represents the ideal prediction line, indicating good calibration of the 2-year survival rate prediction model. (B) Calibration curve for 3-year survival rates, showing alignment between observed and predicted survival rates. (C) Calibration curve for 5-year survival rates, indicating good model accuracy for predicting 5-year survival.

On the common complete-case set (n = 484), we compared a clinical baseline model with a “clinical + six-gene risk score” model. Adding the risk score significantly improved model fit (LRT χ² = 52.45, df = 1, p = 4.42 × 10⁻¹³) and reduced AIC from 2,213.592 to 2,163.145 (ΔAIC = 50.45), demonstrating substantial incremental predictive value beyond standard clinical factors (see Table S5).

Discussion

LUSC, a major subtype of NSCLC, accounts for ~30% of all lung cancers. Owing to late-stage diagnosis in most cases and the paucity of actionable molecular targets, overall outcomes remain suboptimal (Lau et al., 2022). The pronounced heterogeneity of LUSC further increases therapeutic complexity, and the effectiveness of surgery, radiotherapy, and chemotherapy in advanced disease is limited (Villanueva, 2021). Beginning with tumor–adjacent normal contrasts, we systematically identified DEEMTGs in LUSC, constructed a six-gene, LUSC-specific prognostic model, and integrated TME features for comprehensive evaluation.

EMT plays a central role in invasion, metastasis, and treatment resistance (Huang, Hong & Wei, 2022; Yuan et al., 2014). In LUSC, EMT can diminish cell–cell adhesion, remodel the extracellular matrix, and activate migratory programs, thereby promoting malignant phenotypes; it may also upregulate drug-efflux pathways and suppress apoptosis, enhancing chemoresistance (Ebrahimi et al., 2024; Rivas et al., 2021; Vafaeinik et al., 2022). These properties align with the EMT-related transcriptional abnormalities observed here, underscoring a pivotal role for EMT in LUSC progression.

We further conducted a structured comparison between our six-gene EMT signature and representative NSCLC/EMT studies. Prior models were frequently derived from pan-NSCLC or LUAD cohorts, often included larger gene sets, and in some cases lacked systematic TME analyses or pathological corroboration; a four-gene model specific to LUSC has been reported, but with limited TME integration. By contrast, our six-gene panel (GAB2, ALDOA, PCDHA3, TMEM92, ERH, IRS4), developed and validated exclusively in LUSC, balances parsimony with biological interpretability, demonstrates prognostic value independent of clinical stage (III/IV), T3/4, and M1, and shows coherent associations with CIBERSORT/ESTIMATE immune–stromal features. Exploratory IHC provides tissue-level directional support, collectively indicating incremental clinical utility and biological consistency.

In TCGA, model discrimination was moderate; in the external GSE30219 cohort, the 1-year AUC was comparatively higher, highlighting time- and cohort-dependence of time-dependent ROC performance. Because the external evaluation applied fixed TCGA coefficients without retraining or tuning, we regard the 1-year results as supportive rather than definitive; broader temporal generalizability requires further independent cohorts. Multivariable Cox analysis showed that the risk score is independent of clinical stage (III/IV), T3/4, and M1, indicating incremental stratification beyond conventional factors and laying the groundwork for combined models with staging/imaging features. Improvement in the likelihood-ratio test and AIC further demonstrates that adding the six-gene risk score to a clinical baseline model significantly enhances fit and informational value (LRT χ² = 52.45, df = 1, p = 4.42 × 10⁻¹³; AIC: 2,213.592 → 2,163.145), evidencing substantive incremental utility.

After Benjamini–Hochberg FDR control, several immune-cell subsets and ESTIMATE metrics remained significantly different between risk groups (q < 0.05), supporting a population-level association between EMT-related stratification and TME composition. Given the sensitivity of immune deconvolution to tumor purity, batch effects, and expression normalization, these findings should be interpreted as population statistics rather than individual diagnostics. On a continuous scale, TMB was slightly higher in the high-risk group but not statistically significant; recognizing right-skew and threshold sensitivity, we did not perform fixed-threshold dichotomization. Any proportion plots are illustrative only. Considering platform and purity effects on TMB estimation, these results are exploratory and warrant validation in same-histology immunotherapy cohorts under harmonized sequencing/normalization.

Regarding IMvigor210, we treat the analysis as hypothesis-generating, not external validation: the cohort differs from LUSC in histology and treatment context (cross-tumor extrapolation), and the cut point was derived within the same dataset via maximally selected rank statistics, introducing potential leakage and optimism. Hence, the findings merely suggest that “the EMT risk score may relate to immunotherapy response,” and require prospective testing in NSCLC/LUSC immunotherapy cohorts using preregistered, fixed coefficients and thresholds.

As a signaling adaptor, GAB2 integrates multiple receptor inputs to activate PI3K/AKT and MAPK/ERK pathways, promoting proliferation and inhibiting apoptosis (Chen et al., 2016; Ding et al., 2015; Glaviano et al., 2023). The expression changes observed in LUSC are directionally concordant with GAB2’s pro-proliferative/anti-apoptotic roles across cancers, implicating it in EMT-associated malignant phenotypes. In breast cancer, GAB2 overexpression correlates with aberrant HER2 pathway activation, driving rapid proliferation and apoptosis resistance (Zhang et al., 2018); in gastric cancer, GAB2 enhances PI3K/AKT signaling and increases drug resistance (Lee et al., 2007); and in leukemia, high GAB2 expression augments anti-apoptotic signaling, supporting leukemic cell survival and disease progression (Gong et al., 2022).

ALDOA is a key glycolytic enzyme frequently linked to an accentuated Warburg effect, metabolic reprogramming, and invasive–metastatic behavior (Alberghina, 2023; Fu et al., 2018; Tang & Cui, 2024; Zhong et al., 2022). In this study, ALDOA followed the expected directionality, suggesting participation in a metabolism–EMT axis that promotes malignant phenotypes. For instance, in gastric cancer, ALDOA overexpression correlates with heightened migratory and invasive capacity and poorer prognosis (Gu et al., 2022; Jiang et al., 2018); similarly, in hepatocellular carcinoma, elevated ALDOA is closely associated with tumor progression and metastasis (Song et al., 2023). We stress that these are mechanistic inferences requiring functional experiments and causal validation; no therapeutic claims are drawn here.

Regarding IRS4 and ERH: although both carry “protective” coefficients in multivariable survival models, tumor tissues in our IHC cohort (n = 8, stage III–IV) showed higher protein levels than adjacent normals. We do not view this as a true contradiction for three reasons. First, Cox coefficients reflect the adjusted direction of the expression–risk association and are not equivalent to tumor–normal contrasts. Second, transcript–protein discordance is common—shaped by post-transcriptional regulation, protein turnover, and antibody specificity/batch effects—precluding linear correspondence (Guijarro et al., 2023; Hao et al., 2021; Hoxhaj, Dissanayake & MacKintosh, 2013; Zhang et al., 2022). Third, our IHC series is small and late-stage; sampling, tumor purity, and spatial heterogeneity can bias estimates. Biologically, IRS4 modulates the PI3K/AKT axis governing proliferation/survival, and ERH participates in mRNA splicing, cell-cycle control, and genome stability (Pang et al., 2022, 2019; Weng & Luo, 2013). Increased protein abundance may therefore reflect responses to replicative stress or genomic instability rather than a simple uni-directional oncogenic or tumor-suppressive effect, with strong context dependence across tumor types and microenvironments (Onyango & Feinberg, 2011). Accordingly, we position the IHC results as directional clues, not population-level conclusions; larger, stage-balanced cohorts with standardized H-scores, blinded dual reads, and digital quantification are warranted.

PCDHA3 is an adhesion-related gene. In tumor–normal comparisons, its protein level was lower, yet within-cohort survival modeling associated higher expression with risk. This points to dual layers of action: across tissues, diminished adhesion aligns with EMT/migration (tumor < normal) (Sisto, Ribatti & Lisi, 2021); within tumors, relatively higher PCDHA3 may mark specific lineages or microenvironmental coupling linked to poorer survival (Imai et al., 2008). Disentangling cellular origin and network context will require purity-adjusted analyses and single-cell/spatial multi-omics.

TMEM92 showed no significant IHC difference but was risk-associated in survival models. As a transmembrane protein, functional–expression decoupling may arise from subcellular localization, cell-subset specificity, or post-translational modification (Herrera-Quiterio & Encarnación-Guevara, 2023); conventional IHC may undercapture its active state. Co-localization immunofluorescence and perturbation assays (knockdown/overexpression) are therefore advisable.

Taken together, the “directional inconsistencies” between IHC and bioinformatics largely reflect differences in metric meaning (survival association vs tumor–normal contrast), cross-level molecular mismatch, and sample/technical factors. We interpret these data as histologic complements to the model genes; future work with larger independent cohorts and functional studies will clarify the context dependence of IRS4, ERH, PCDHA3, and TMEM92 in LUSC.

Within the LUSC tumor microenvironment, the composition and distribution of immune infiltrates strongly influence disease course and prognosis. Our analyses further illuminate the relationship between immune infiltration and the risk score, revealing distinct patterns in high- versus low-risk patients and exploring their biological implications. After Benjamini–Hochberg FDR control across ESTIMATE and CIBERSORT comparisons, a subset of differences remained significant (q < 0.05), with others showing trends, indicating that these observations primarily reflect population-level associations.

Notably, naïve B cells, CD8⁺ T cells, and activated CD4⁺ memory T cells were more abundant in low-risk patients—immune populations typically linked to robust antitumor responses (Sommermeyer et al., 2016). CD8⁺ T cells directly recognize and eliminate tumor cells (Koh et al., 2023; Schreiber, Old & Smyth, 2011); their higher infiltration in the low-risk group suggests stronger immune surveillance that may delay or prevent progression (Fridman et al., 2012). Activated CD4⁺ memory T cells potentiate CD8⁺ function via cytokine support and facilitate B-cell maturation and antibody production, reinforcing antitumor immunity (Borst et al., 2018; Sage & Sharpe, 2015; Speiser et al., 2023; Zhu et al., 2017). Although naïve B cells can play context-dependent roles across cancers, higher infiltration is often associated with favorable outcomes, implying a beneficial contribution to immune surveillance.

Conversely, resting CD4⁺ memory T cells and M0 macrophages were relatively increased in high-risk patients. Enrichment of resting memory T cells may indicate a quiescent, ineffectual state that enables immune escape, while elevated M0 macrophages may presage an immunosuppressive milieu. Under specific cues, M0 macrophages can polarize toward an M2-like phenotype that suppresses antitumor immunity and promotes tumor growth (Sica & Mantovani, 2012), and M2 abundance is frequently linked to poorer prognosis (Ogiya et al., 2016). Macrophage plasticity, however, allows therapeutic repolarization toward M1 states that restrain tumor growth (Cassetta & Pollard, 2018). Given the sensitivity of deconvolution to purity, batch effects, and normalization, these results should be interpreted as population-level observations.

Collectively, immune-infiltration profiling reveals TME heterogeneity aligned with risk stratification: low-risk patients harbor a more immune-activated landscape (e.g., CD8⁺ T cells, activated CD4⁺ T cells), whereas high-risk patients display a more suppressed or non-activated phenotype (resting CD4⁺ T cells, M0 macrophages). These findings generate testable hypotheses for immunophenotypic characterization; their therapeutic implications require validation in independent cohorts under harmonized workflows.

In the treatment of LUSC, immunotherapy—particularly checkpoint blockade—has shown encouraging efficacy, yet responses vary markedly among patients. This heterogeneity partly reflects differences in the TME and in tumor mutational burden (TMB) (2023; Goodman et al., 2017). TMB, the number of mutations within a tumor genome, is often considered a putative predictor of immunotherapy benefit because higher TMB can increase neoantigen load, enhance immune visibility, and trigger stronger antitumor responses (Chalmers et al., 2017).

Here, we assessed the relationship between TMB and risk stratification using two pre-specified statistical perspectives, with thresholds and tests defined in the Methods. First, on a continuous scale (Wilcoxon rank-sum), violin and bar plots indicated slightly higher TMB in the high-risk group, without statistical significance (p = 0.052). Second, dichotomous analyses employed two sensitivity thresholds: the commonly cited absolute cut-off of 10 muts/Mb and the cohort upper quartile (Q3), comparing “high TMB” proportions by Fisher’s exact test. Under dichotomization, a greater proportion of “high TMB” samples was observed in the low-risk group.

These discordant directions are readily explained by TMB’s right-skewed distribution, threshold sensitivity, and potential outliers; platform/normalization differences and tumor purity further affect TMB stability. In terms of mutational spectra, frequent TP53 mutations align with adverse survival (Olivier, Hollstein & Hainaut, 2010). Although TTN is commonly mutated across cancers—partly due to gene length—its functional significance in LUSC remains to be clarified, with some studies suggesting links to increased neoantigen load (Ying et al., 2024). In light of these uncertainties, we classify our TMB findings as exploratory and refrain from confirmatory claims.

Crucially, TMB is not the sole determinant of immunotherapy response. Even with high TMB, immunosuppressive signaling within the TME can blunt neoantigen-driven immunity and limit efficacy (Topalian et al., 2016). A more prudent strategy is to integrate TMB with PD-L1 expression, the composition and activation state of infiltrating immune cells, and tumor metabolic status, among other dimensions (Sharma et al., 2017).

Building on these findings and limitations, we propose a pragmatic, operational pathway for clinical translation, with key standardization points summarized in Table S4. Step 1: Analytical validation and assay standardization. Implement an FFPE-compatible six-gene RT–qPCR assay, normalizing ΔCt to the geometric mean of 3–4 housekeeping genes; establish pre-analytical SOPs (sampling, fixation time, section thickness, tumor-cell fraction, RNA quality thresholds such as DV200/RIN); set repeatability/reproducibility criteria (technical CV ≤ 10%, inter-laboratory ICC ≥ 0.90); conduct blinded ring trials; and perform a bridging study to map RNA-seq coefficients onto the qPCR scale, permitting intercept calibration only without retraining coefficients. Step 2: Multicenter prospective validation. Use a preregistered protocol with locked coefficients and thresholds (illustrative threshold 0.13); enroll pretreatment biopsies; designate OS/PFS as primary endpoints, with AUC, C-index, calibration plots, and decision curve analysis (DCA) as secondary metrics; harmonize sampling/submission SOPs and predefine missing-data handling; during external evaluation, do not retrain the model—allow only minimal intercept adjustments for scale alignment when necessary. Step 3: Clinical utility assessment. Through biomarker-stratified or enrichment designs, quantify impact on therapeutic decision-making (immunotherapy/chemotherapy) and health economics (turnaround ≤5 days, per-test cost, net benefit), and benchmark against joint models that combine the score with stage or routine predictors (see Table S6). This pathway is intended to ensure assay robustness, generalizability, and demonstrable clinical benefit prior to real-world deployment.

Future studies should reassess the discriminative value of TMB within same-histology immunotherapy cohorts using harmonized sequencing/normalization pipelines and purity correction, alongside threshold-sensitivity and robustness analyses. Cross-validation with orthogonal indicators—such as neoantigen load and IPS/TIDE—will help establish a reproducible, translatable chain of evidence.

Conclusions

This LUSC-specific study developed a prognostic model comprising six EMT-related genes—four risk-associated (GAB2, ALDOA, PCDHA3, TMEM92) and two protective (IRS4, ERH)—which achieved moderate stratification performance in the TCGA cohort and remained independent of conventional clinical factors. The risk strata delineated by the model showed population-level associations with TME composition. Analyses of TMB yielded directionally discordant findings when viewed on continuous versus thresholded scales and are therefore interpreted as exploratory. A small IHC series provided directional support for the panel genes: ALDOA and/or GAB2 were elevated in tumors, PCDHA3 was reduced, TMEM92 showed no clear difference, and IRS4/ERH displayed protein–transcript discordance, suggesting post-transcriptional or proteostatic regulation.

Overall, the EMT-informed signature offers potential value for risk stratification and immunophenotypic characterization in LUSC; however, core conclusions require validation in larger, stage-balanced independent cohorts—especially same-histology immunotherapy populations with locked thresholds and external calibration. Pathologically, standardized H-scores, blinded dual reads, and digital quantification should be employed, alongside functional assays and spatial/single-cell multi-omics. Orthogonal metrics such as IPS/TIDE and neoantigen load will be essential to elucidate mechanisms and assess translational readiness.

Supplemental Information

Supplemental Information 3. Quantitative IHC of six genes in tumor (A2) versus adjacent (B2) tissues.

Mean ± SD of IOD (n = 8) and two-sided P values for between-group comparisons.

Supplemental Information 4. Group comparisons of CIBERSORT immune-cell fractions and ESTIMATE scores between risk strata.

Raw p values (Wilcoxon) and Benjamini–Hochberg–adjusted q values; features with q < 0.05 indicate FDR-significant differences.

Supplemental Information 5. Model fit and incremental value comparing Clinical-only versus Clinical + Signature on the same complete-case set (n = 484).

Log-likelihood, AIC, ΔAIC, likelihood-ratio χ², df, and p value. Smaller AIC indicates better fit; LRT tests added value of the six-gene score.

Supplemental Information 9. RNA-seq expression and clinical data used in LUSC analysis.

Normalized RNA-seq TPM values and matched clinical data from TCGA-LUSC samples, used for identifying EMT-related genes, constructing a prognostic risk score, and analyzing correlations with survival and immune infiltration.

Supplemental Information 10. Immunohistochemistry raw data of six EMT-related genes.

Raw immunohistochemistry data (image-derived optical density values and clinical annotations) from tumor and adjacent tissues of 8 LUSC patients, used to validate the expression of GAB2, ALDOA, ERH, IRS4, PCDHA3, and TMEM92 at the protein level.

Supplemental Information 11. R scripts for differential gene expression and prognostic model construction in LUSC.

This archive contains all R scripts used for data preprocessing, DESeq2 differential expression analysis, LASSO and Cox regression modeling, immune infiltration estimation using CIBERSORT, mutation analysis, and visualization of results in the study of EMT-related gene signatures in lung squamous cell carcinoma (LUSC).

Supplemental Information 14. R script for parsing TCGA metadata JSON files and matching sample identifiers.

The complete metadata for all TCGA-LUSC samples

Supplemental Information 15. Comprehensive EMT-related gene list with relevance scores exported from GeneCards.

The full set of EMT-associated genes exported from the GeneCards database, including gene symbols and relevance scores

Funding Statement

This work was supported by the Medical Science Research Project of Hebei (Grant No. 20250271). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Additional Information and Declarations

Competing Interests

The authors declare that they have no competing interests.

Author Contributions

Anqi Zhang conceived and designed the experiments, performed the experiments, analyzed the data, prepared figures and/or tables, authored or reviewed drafts of the article, and approved the final draft.

Jinping He conceived and designed the experiments, performed the experiments, authored or reviewed drafts of the article, and approved the final draft.

Qiang Lin conceived and designed the experiments, analyzed the data, prepared figures and/or tables, authored or reviewed drafts of the article, and approved the final draft.

Human Ethics

The following information was supplied relating to ethical approvals (i.e., approving body and any reference numbers):

Ethics Committee of Huabei Petroleum Administration Bureau General Hospital. Approval No. hbyyec-wz-03.

Data Availability

The following information was supplied regarding data availability:

The raw data is available in the Supplemental Files.