Senescent epithelial cells remodel the tumor microenvironment and drive early LUAD progression: a multiomics risk model and single-cell analysis

Abstract

Lung adenocarcinoma (LUAD) remains a major cause of cancer-related mortality, and there are currently few reliable biomarkers available for accurate prognosis and effective targeted therapy. Accumulating evidence demonstrates that senescent cells play an important role in tumor progression and immune evasion; nevertheless, their specific contribution to LUAD pathogenesis has not yet been fully elucidated. Accordingly, a five-gene senescence-related risk model comprising FGF2, GAPDH, CCNA2, ENO1, and DKK1 was established using the Least Absolute Shrinkage and Selection Operator regression applied to bulk RNA sequencing data obtained from The Cancer Genome Atlas (TCGA). Patients stratified as high risk by this model exhibited significantly poorer overall survival and progression-free survival, accompanied by marked activation of pathways associated with immune infiltration, epithelial–mesenchymal transition, and extracellular matrix remodeling. Integrative single-cell RNA sequencing analysis further revealed a distinct epithelial subpopulation (E9) defined by preferential activation of the senescence-associated gene DKK1. This subpopulation, which emerged from integrative single-cell transcriptomic analysis, exhibited pronounced senescence-associated characteristics, significantly increased cellular stemness, and extensive intercellular communication capacity, and uniquely expressed COL17A1 together with transcriptional programs that were strongly associated with epidermal development. Pseudotime analyses consistently positioned E9 cells at an early stage of tumor evolution. At the functional level, DKK1 contributed to senescence-associated phenotypic features, thereby promoting tumor cell migration and metastatic potential. From a clinical perspective, patients with concurrent high levels of COL17A1 and DKK1 expression experienced significantly worse clinical outcomes. Collectively, these findings identify a senescence-driven epithelial subpopulation that contributes to LUAD progression via DKK1-mediated activation of aging-related pathways and highlight DKK1 as a potential therapeutic target.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00018-026-06101-8.

Introduction

Lung adenocarcinoma (LUAD), the most prevalent histological subtype of lung cancer, remains one of the leading causes of cancer-related mortality worldwide, with an average five-year survival rate of only 15% [1, 2]. This unfavorable prognosis primarily arises from early metastatic dissemination and the progressive development of therapeutic resistance [3, 4]. Although major advances have been achieved in targeted therapy and immunotherapy over the past decade, intrinsic intratumoral heterogeneity and adaptive tumor survival mechanisms continue to drive treatment failure in a large proportion of patients [4]. Consequently, there is a critical and ongoing need to identify robust prognostic biomarkers that can improve risk stratification, enable more accurate outcome prediction, and ultimately enhance the quality of life of patients with LUAD [5].

Among emerging biological processes, cellular senescence has attracted considerable attention as a stable state of cell cycle arrest that exerts complex and context-dependent effects on tumor development and progression [6, 7]. Historically, senescence was viewed as a tumor-suppressive mechanism that limits the proliferation of genetically damaged cells [6, 8, 9]. However, growing evidence now demonstrates that senescent cells persisting within the tumor microenvironment (TME) can paradoxically promote malignant progression [10]. In particular, senescent cells frequently acquire a senescence-associated secretory phenotype (SASP), characterized by the sustained release of inflammatory cytokines, chemokines, growth factors, and proteases [11]. Prolonged exposure to these bioactive factors amplifies tumor invasiveness, facilitates metastatic dissemination, promotes immune evasion, and contributes to resistance to anticancer therapies [12]. Moreover, exposure to radiotherapy or chemotherapy can induce therapy-induced senescence in tumor cells, resulting in the emergence of a dormant-like phenotype [13]. Collectively, these findings underscore the importance of systematically defining the senescent landscape in LUAD to refine patient stratification and inform clinical decision-making.

Despite intensive investigation, identifying and characterizing senescent cells in vivo remains technically challenging due to the lack of unique, well-established markers [14–17]. Moreover, the molecular and phenotypic characteristics of senescent cells vary substantially depending on cell lineage, the nature of the senescence-inducing stimulus, and the duration of growth arrest [18, 19]. The rapid advancement of bulk RNA sequencing (bulk RNA-seq) and single-cell RNA sequencing (scRNA-seq) technologies has facilitated comprehensive profiling of cancer-related features and the identification of cell-type–specific signatures across varied biological contexts [20–22].

Recent methodological advances have shifted cancer research from isolated single-technology analyses toward systematic multiplatform data integration [23, 24]. For example, computational frameworks such as FigureYa efficiently integrate multiomics datasets through modular code and facilitate the generation of high-quality visualizations [24]. These innovations are driving a fundamental transformation in cancer biomarker discovery and provide a strong technical and conceptual foundation for mechanistic studies and for clinically stratified diagnosis and treatment.

In this study, we developed an Aging Risk Score (ARS) for patients with LUAD based on the AgingAtlas and SenMayo aging-related gene sets [25, 26]. High-risk patients showed significantly poorer clinical prognosis and marked aging-associated biological features, including enhanced epithelial–mesenchymal transition (EMT), increased tumor-associated fibrosis, and elevated macrophage infiltration. Integrative scRNA-seq analysis further identified Dickkopf-1 (DKK1), an aging-related prognostic gene, as being preferentially and selectively utilized by a distinct subpopulation of tumor epithelial cells. These cells exhibited robust senescence-associated characteristics, extensive intercellular communication capacity, and strong functional associations with extracellular matrix (ECM) organization and cellular structural pathways. Importantly, functional experiments confirmed that DKK1 actively promotes senescence-associated phenotypic features, thus substantially enhancing tumor cell migratory behavior and metastatic potential. Together, these findings provide compelling evidence that a senescence-driven epithelial subpopulation contributes to LUAD progression via DKK1-mediated TME remodeling.

Results

Construction and validation of an aging-related risk model in cohort studies

A total of 563 senescence-related genes (SRGs) were initially identified from the combined SenMayo gene set and AgingAtlas database. These SRGs were then intersected with the differentially expressed genes (DEGs) derived from LUAD tumor tissues and adjacent normal tissues in The Cancer Genome Atlas (TCGA) database, resulting in the identification of 159 senescence-related DEGs (Fig. 1A). Univariate Cox regression analysis demonstrated that 38 of these genes were significantly associated with patient prognosis. To reduce the risk of model overfitting, the Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was performed on the 38 prognostic genes. Based on the optimal λ value, five genes, FGF2, DKK1, ENO1, CCNA2, and GAPDH, were selected to construct the ARS model (Fig. 1B).

Fig. 1.

Prognostic model related to cellular senescence based on the Agingatlas and SenMayo gene sets. A Volcano plot of differentially expressed genes (DEGs) between tumor and normal tissues; B Key gene selection using LASSO regression analysis; C Forest plot showing the results of multivariate cox regression analysis between senescence-related prognostic genes and overall survival; D, E Kaplan-Meier survival curves for high-risk and low-risk patients in validation cohorts GSE31210 and GSE31219; F Kaplan-Meier survival curves for progression-free survival (PFS) in high-risk and low-risk patients from a lung adenocarcinoma cohort at Nanjing Medical University Affiliated Cancer Hospital

Subsequent multivariate Cox proportional hazards analysis revealed that ENO1 (HR = 1.9, 95% CI: 1.27–2.8, P = 0.001), GAPDH (HR = 1.4, 95% CI: 1.01–2.0, P = 0.046), DKK1 (HR = 1.2, 95% CI: 1.07–1.3, P < 0.001), and FGF2 (HR = 1.4, 95% CI: 1.06–1.8, P = 0.016) served as independent prognostic factors for patients with LUAD (Fig. 1C). The prognostic performance of this model was further evaluated in three independent validation cohorts (GSE31210, GSE30219, and the Nanjing Medical University Affiliated Cancer Hospital cohort) by stratifying patients into low-risk and high-risk groups according to the median ARS calculated from these five genes (Fig. 1D and F). In two of the validation cohorts, overall survival (OS) was significantly shorter in the high-risk group than in the low-risk group (Fig. 1D: P = 0.0056; Fig. 1E: P = 0.00065). In addition, the predictive value of the model for progression-free survival (PFS) was assessed, demonstrating that high-risk patients also experienced markedly shorter PFS than low-risk patients in the Nanjing Medical University Affiliated Cancer Hospital cohort (Fig. 1F: P = 0.0099). Collectively, these results indicate that the aging-related risk model reliably stratifies patients with LUAD according to their risk of poor OS and PFS. This conclusion was further supported by receiver operating characteristic curve analyses in all three cohorts (Figs.S1A–S1C).

Comparison of differences in DEGs and TME between high-risk and low-risk groups

Using the established formula, risk scores were calculated for all patients with LUAD in the TCGA cohort. Patients were then classified into high-risk and low-risk groups based on the median risk score (Table S1). To characterize the molecular differences between these groups, DEGs were identified using a one-fold change threshold and an adjusted p-value of < 0.05. This analysis revealed 1,128 genes that were upregulated and 1,936 genes that were downregulated in the high-risk group compared with the low-risk group. The resulting volcano plot illustrated these expression patterns and highlighted the top five genes exhibiting the largest log2(fold change) (logFC) values in both the upregulated and downregulated gene sets (Fig. 2A). Notably, the SRG DKK1 was not only significantly upregulated but also displayed the highest −log10(p-value) among all genes analyzed, underscoring its strong differential expression and potential biological relevance.

Fig. 2.

Comparative analysis of differentially expressed genes, tumor microenvironment characteristics, and clinical indicators between high-risk and low-risk groups. A Volcano plot depicting differentially expressed genes (DEGs) between the high-risk and low-risk groups; B Dot plot representing Gene Ontology (GO) enrichment analysis of DEGs in the high-risk group; C Bar graph showing KEGG enrichment analysis of DEGs in the high-risk group; D-G Comparative analysis of senescence-associated features, EMT markers, cancer-associated fibroblasts, and macrophage infiltration characteristics between the two subgroups; H-K Bar charts comparing differences in Aging Risk Score across various clinical indicators in our hospital cohort; L Bar charts comparing the TIDE scores of different risk groups in the TCGA-LUAD cohort; Statistical significance was assessed using the Wilcoxon test, and p-values are marked with asterisks: *: p < 0.05; **: p < 0.01; ***: p < 0.001; ****: p < 0.0001

Subsequent functional enrichment analysis of DEGs in the high-risk group revealed that Gene Ontology (GO) pathways were predominantly associated with ECM components and cellular structural organization (Fig. 2B). Meanwhile, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis demonstrated significant enrichment in pathways related to cellular processes and environmental information processing, with a particular emphasis on cell cycle regulation (Fig. 2C). These findings suggest that dysregulation of the cell cycle represents a key mechanism through which the high-risk group influences cellular senescence.

To further characterize differences in TME, the Immuno-Oncology Biological Research (IOBR) algorithm was applied. The high-risk group exhibited significantly higher scores in aging-related pathways, including DNA damage response, cell cycle regulation, and DNA repair (Fig. 2D). In addition, EMT activity and cancer-associated fibroblast features were markedly elevated in the high-risk group, potentially linking this molecular subtype to pulmonary fibrosis (Fig. 2E and F). A substantially higher macrophage score further indicated increased macrophage infiltration in the high-risk group (Fig. 2G).

Validation in the Nanjing Medical University Affiliated Cancer Hospital cohort demonstrated that patients with poorer prognostic characteristics consistently presented with higher risk scores. Compared with lepidic predominant adenocarcinoma (LPA), the pathological subtypes acinar predominant adenocarcinoma (APA), papillary predominant adenocarcinoma (PPA), and solid predominant adenocarcinoma (SPA) exhibited significantly elevated risk scores (P < 0.05), with SPA showing the most pronounced increase (Fig. 2H, P < 0.01). Furthermore, patients with positive spread through air spaces (STAS) lesions had significantly higher risk scores than those without STAS (Fig. 2I, P < 0.05). Additional analyses revealed a progressive increase in risk scores with advancing lymph node metastasis and tumor stage, suggesting a close association between molecular risk stratification and clinical disease progression (Fig. 2J and K). Application of the IOBR algorithm to sequencing data from this cohort confirmed that the high-risk group displayed trends consistent with those observed in the TCGA-LUAD cohort; notably, the high-risk subgroup demonstrated significantly increased activity of senescence-related pathways and an upward trend in macrophage and cancer-associated fibroblast infiltration (Figs. S2A–S2C).

We next investigated whether the five-gene risk score could predict patient responsiveness to immunotherapy by capturing aging-associated alterations within TME. To this end, the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm was applied [27]. High-risk patients displayed significantly higher TIDE scores than low-risk patients, indicating enhanced immune evasion and a reduced likelihood of benefiting from immune checkpoint blockade therapy (Fig. 2L, P < 0.001).

Dissecting the mechanisms of aging-related epithelial cell subpopulations via scRNA-Seq transcriptomics

Building on the findings of Philip Bischoff et al. and the GSE131907 single-cell sequencing dataset, the cellular landscape underlying LUAD progression has been comprehensively mapped. On this foundation, the present study further investigates the influence of cellular senescence on LUAD development. A total of 233,850 cells derived from normal tissue, early-stage lung cancer (tLung), advanced lung cancer with bronchial invasion (tL/B), and metastatic lymph nodes (mLN) were classified into eight distinct lineages according to the expression of canonical marker genes (Fig. 3A and B). The Uniform Manifold Approximation and Projection visualization of tLung, tL/B, and mLN samples demonstrated highly similar cellular distribution patterns across these sample types (Fig.S3A).

Fig. 3.

Identification and biological function analysis of senescence-associated epithelial cell subpopulation E9 in LUAD. A UMAP clustering analysis divides cells from normal, tumor, and metastatic lymph node samples into eight major lineages; B The average expression dot plot of typical marker genes for the eight major genealogies is shown; C Box plots demonstrate the distribution of AUCell scores for epithelial cells derived from different samples across a senescence-related gene set. Statistical significance was assessed using the Wilcoxon test, and p-values are marked with asterisks: *: p < 0.05; **: p < 0.01; ***: p < 0.001; D UMAP clustering analysis segregates epithelial cells from tumor samples into 12 cell clusters; E Box plots illustrate the distribution of AUCell scores for epithelial cell subpopulations across the senescence-related gene set; F Relative proportions of cell cycle phases for each cell subpopulation are presented; G Violin plots show the expression of senescence-associated prognostic genes in epithelial cell subpopulations; H GO pathway enrichment and KEGG pathway enrichment analyses for the E9 subpopulation. Colors represent functional categories of the pathways; I GSEA enrichment analysis of the E9 subpopulation across the hallmark gene set

To systematically evaluate senescence-related features in epithelial cells, two primary sources of biological evidence were integrated: senescence-associated prognostic genes derived from the risk model and the oncogene-induced senescence pathway (GO:0040402). Using these molecular references, the AUCell algorithm was applied to quantify senescence features in epithelial cells from each sample by calculating gene set activity scores (Fig. 3C). Notably, scores in mLN samples were significantly higher than those in tLung (P < 0.001) and tL/B (P < 0.001) samples. In addition, advanced-stage tumors exhibited higher scores than early-stage tumors, suggesting that senescence-associated features progressively enhanced during tumor progression and metastasis.

Subsequent analysis of epithelial cell subpopulations in lung tumor samples clustered the cells into 12 distinct groups (Fig. 3D). By incorporating the SenMayo gene set used for model construction, the senescence states of these subpopulations were further quantified. Among them, subpopulation E9 displayed the highest activity in senescence-related pathways (Fig. 3E). Furthermore, multiple aging-associated pathways representing distinct mechanisms of senescence were used for AUCell scoring to more comprehensively evaluate the senescence status of the E9 subpopulation and to assess the ability of the selected gene sets and models to capture the lung aging microenvironment. The E9 subpopulation exhibited significantly higher scores in aging/senescence-induced gene and SASP-related pathways than other subpopulations, indicating pronounced senescence characteristics (Figs.S4A–C). Consistent with these results, cell cycle analysis confirmed that E9 contained a significantly greater proportion of G1-phase cells than other subpopulations (Fig. 3F). Moreover, the risk gene DKK1 was specifically enriched in E9, correlating with the highest risk gene score and suggesting that this subpopulation may promote SRG expression and corresponding senescence phenotypes (Fig. 3G).

DEGs in E9 satisfying logFC of > 0.5 and adjusted p-value of < 0.05 were selected for downstream enrichment analysis. GO and KEGG pathway analyses revealed that this subpopulation was predominantly associated with cell adhesion junctions and cytoskeleton-related pathways (Fig. 3H). COL17A1, a marker gene of this cluster, was highly enriched in collagen-associated pathways (Fig.S5A). Consistent with these findings, gene set enrichment analysis (GSEA) demonstrated that the most significantly upregulated gene sets were related to EMT and apical junction pathways, both of which are closely linked to cancer invasion and metastasis (Fig. 3I).

Pseudotemporal analysis reveals senescent epithelial cell states promoting LUAD progression in early tumorigenesis

To investigate the role of the E9 subpopulation in epithelial cell differentiation, pseudotemporal analyses were conducted on epithelial cells derived from primary tumor tissues and mLNs. The inferred differentiation trajectory revealed a trifurcated developmental structure comprising five transcriptional states (Fig. 4A and B), indicating that epithelial cells undergo multiple differentiation transitions during tumorigenesis. Notably, E9 cells were predominantly localized within State 4, whose transcriptional profile closely aligned with molecular features associated with tumor progression and lymph node metastasis (Fig. 4A). Furthermore, CytoTrace analysis demonstrated that epithelial cells in States 4 and 5 exhibited higher differentiation potential than those in the remaining states, suggesting the presence of two key transcriptional programs active during early tumor development (Fig. 4C).

Fig. 4.

Pseudo-temporal analysis of epithelial cell subpopulation E9 and its association with low survival rates. A-C Monocle2 analysis of epithelial cell differentiation trajectories in tumor and metastatic lymph node samples, displayed as cell type (top), state (middle), and CytoTrace score (bottom). tLung: early-stage lung cancers; tL/B: advanced-stage lung cancers; mLN: metastatic lymph nodes; D Differential enrichment of cell states and their associated clinical parameters in stage I, II and III LUAD. UL: upper lobe; LL: lower lobe; E Box plot showing the distribution of AUCell scores for the Senmayo gene set across cells in different states. Statistical significance was assessed using the Wilcoxon test, and p-values are marked with asterisks: *: p < 0.05; **: p < 0.01; ***: p < 0.001; F, G In our hospital cohort, patients were stratified based on E9 subtype marker genes COL17A1 and hijacked DKK1. Kaplan-Meier survival curves were constructed for COL17A1(+)/DKK1(+), COL17A1(+)/DKK1(-), and COL17A1(-)/DKK1(+) groups to compare progression-free survival (PFS) among different patient subgroups

Based on these findings, patients from an early-stage lung cancer cohort were stratified into Group A, which contained substantially more State 4 cells than State 5 cells, and Group B, which comprised the remaining patients (Fig. 4D). Clinically, patients in Group A exhibited more advanced disease stages and potentially higher KRAS mutation rates than those in Group B.

Consistent with these observations, E9 cells were significantly enriched in State 4, indicating that this subpopulation may play a pivotal role during the early phases of tumorigenesis (Fig.S6A). As differentiation potential declined, two additional transcriptional states, State 1 and State 2, emerged during later tumor stages. State 1 cells were predominantly associated with cytoplasmic translation and ribosome processing, whereas State 2 cells displayed upregulated pathways governing cell motility and migration (Fig.S6B). In parallel, senescence features assessed using the SenMayo gene set revealed that State 4 cells exhibited markedly elevated senescence signatures compared with other state cells (Fig. 4E).

Given these pronounced senescence signatures, the marker gene COL17A1, specific to E9, and the senescence-associated gene DKK1 were jointly applied to stratify patient prognosis. Patients whose tumors exhibited concurrent high expression of COL17A1 and DKK1 had significantly shorter PFS than those with high COL17A1 but low DKK1 expression (Fig. 4F), and this unfavorable trend remained evident when compared with patients showing low COL17A1 but high DKK1 expression (Fig. 4G). These findings underscore the central role of cellular senescence in tumor progression and highlight its potential value as a prognostic indicator.

Overall, pseudotemporal trajectory analysis of epithelial cells from tumor and mLN samples demonstrates how the senescent E9 subpopulation may drive LUAD progression. During early tumorigenesis, a subset of epithelial cells appears to acquire enhanced senescence features that confer tumorigenic stemness, thereby further accelerating disease progression and metastatic dissemination. Collectively, these results indicate that epithelial cells possess substantial phenotypic plasticity, enabling dynamic adaptation across different stages of tumor development.

Aging epithelial cell subpopulation mediates multicellular interactions to drive TME remodeling

Intercellular communication analyses revealed that the E9 subpopulation exhibited extensive reciprocal signaling within the epithelial cell compartment, including particularly robust autocrine signaling activity (Fig. 5A and B). Compared with other epithelial subtypes, E9 displayed more pronounced outward and inward interaction patterns. Notably, E9 also demonstrated strong interaction intensity with myeloid cells in TME, indicating high biological activity and functional relevance (Fig. 5C and D).

Fig. 5.

The E9 subpopulation exhibits extensive interactions with other cell types in the tumor microenvironment. A, B The E9 subpopulation plays a significant role in the production of outgoing signals among other subpopulations of epithelial cells, exhibiting a notable autocrine phenomenon; C, D There are widespread outgoing signals between the E9 subpopulation and other cells; E Overall signaling patterns among epithelial cell subpopulations; F Ligand-receptor pairs between the E9 subpopulation and other epithelial cell subpopulations involved in signaling pathways; G Overall signaling patterns between the E9 subpopulation and other cells; H Ligand-receptor pairs between the E9 subpopulation and other cell types involved in signaling pathways

Several key protumorigenic signaling pathways mediating interactions between E9 and other cell types were identified, including laminin (LAMININ), midkine (MDK), collagen (COLLAGEN), macrophage migration inhibitory factor (MIF), and amyloid precursor protein (APP) (Fig. 5E and F). These pathways are closely associated with tumor progression and microenvironmental remodeling. Among them, the MDK pathway plays a particularly pleiotropic role in tumor evolution, participating in multiple oncogenic processes such as cell proliferation, EMT, and angiogenesis [28–34]. Through the MDK ligand–receptor signaling system, E9 may regulate both its own functional state and that of neighboring epithelial cells, thereby promoting tumor progression.

Within the cross-cell-type interaction network, COLLAGEN, LAMININ, and APP emerged as the dominant signaling pathways (Fig. 5G). Of these, the APP–CD74 signaling axis was especially prominent (Fig. 5H), exhibiting strong interactions with endothelial cells, myeloid cells, and B cells, with the most pronounced interactions observed within the myeloid cell population. Previous studies have implicated this pathway in tumor progression and metastasis [35, 36]. Accordingly, the APP–CD74 signaling axis may contribute to remodeling of the tumor immune microenvironment, thereby facilitating tumor metastasis and disease progression.

To further substantiate these findings, we conducted an in-depth analysis of the interactions among the E9 subpopulation, myeloid cells, and fibroblasts. Pathway activity analysis revealed that intercellular signaling from E9 to fibroblasts occurred predominantly through the Focal adhesion pathway, whereas signaling from E9 to myeloid cells was mediated primarily through the Cellular senescence pathway (Fig.S7A). These results are highly consistent with our earlier observations. Additionally, these complex communication networks potentially reveal tumor heterogeneity and are involved in reshaping the immune microenvironment [37].

Relationship between DKK1 expression and clinicopathological features of patients with LUAD

To investigate the role of DKK1 protein in tumor progression, immunohistochemistry was performed on 115 tissue samples obtained from 52 patients, including 49 adjacent normal tissue samples and 66 tumor tissue samples. Immunohistochemistry revealed a significant increase in DKK1 protein staining intensity in tumor tissues compared with adjacent normal tissues (Fig. 6A). Furthermore, comparative analysis of the average optical density (OD) values of positively stained areas across different tumor categories demonstrated that average OD values were significantly higher in samples from patients with advanced tumor stages than in those from patients with lower tumor stages (Fig. 6B, P < 0.01). Similarly, the average OD values were markedly elevated in samples from patients with regional lymph node metastasis compared with those without metastasis (P < 0.05). In addition, a positive association was observed between tumor size and OD values.

Fig. 6.

DKK1 is associated with poor pathology and prognosis in lung adenocarcinoma. A Representative images of adjacent normal tissue(top), primary lung adenocarcinoma tissue(middle) and metastatic lung adenocarcinoma tissue(bottom) samples are shown; B Differences in the average OD values of IHC samples under different tumor types; C Kaplan-Meier survival curves across IHC cohort, demonstrating that patients with high DKK1 expression have significantly worse progress free survival

For survival analysis, patient samples were stratified into high and low DKK1 protein expression groups based on the median average OD value. Kaplan–Meier analysis demonstrated that patients in the high OD group exhibited significantly shorter PFS than those in the low OD group (Fig. 6C, P = 0.034). Collectively, these findings further support a close association between DKK1 protein expression and tumor progression, suggesting that DKK1 may serve as a potential biomarker for disease progression.

DKK1 overexpression promotes cellular senescence and invasive phenotype

The senescence phenotype of the target cell populations was quantified using the validated human universal senescence index (hUSI) and SenCID computational programs [38, 39]. Notably, the E9 subpopulation exhibited significantly higher senescence scores than all other subpopulations in both assessments (Fig. 7A, Figs.S8A–C). According to the SenCID scoring system, the E9 subpopulation was further classified within the SID2 model, consistent with the reported senescence identity of pulmonary epithelial cells assigned to the SID2 category.

Fig. 7.

DKK1 promotes aging and metastasis of LUAD. A The UMAP plot shows the differences in hUSI scores among different subgroups; B The GSEA analysis was conducted to compare cells overexpressing DKK1 with control cells; C SA-β-gal staining and quantification of DKK1-overexpressing cells and control cells;

D Gene expression analysis of senescence-associated inflammatory factors in DKK1-overexpressing cells, detected by qRT-PCR; E The influence of DKK1 overexpression on cell migration and invasion in experiments; F Detection of senescence-associated inflammatory factors secreted by DKK1-overexpressing cells using ELISA; G, H DKK1 overexpressing and control cells were injected into the tail veins of nude mice. The lung metastasis was monitored using an in vivo imaging system, *: p < 0.05; **: p < 0.01; ***: p < 0.001; ****: p < 0.0001

To functionally investigate the role of DKK1, an A549 lung cancer cell line stably overexpressing DKK1 was established. Transcriptomic profiling followed by GSEA showed that DKK1 overexpression was associated with significant activation of senescence-related programs, including the SenMayo signature, EMT, and TGF-β signaling pathways (Figs. 7B). Consistently, DKK1-overexpressing cells displayed significantly increased senescence-associated β-galactosidase (SA-β-gal) activity (Figs. 7C). Subsequently, quantitative reverse transcription polymerase chain reaction (qRT-PCR) was performed to assess the messenger RNA expressions of IL1A, IL1B, IL6, IL8, and MMP9, and the corresponding protein concentrations were measured using enzyme-linked immunosorbent assay (ELISA). The results demonstrated significant upregulation of these cytokines in the DKK1-overexpressing cell line, indicating that elevated DKK1 expression promotes the secretion of senescence-associated inflammatory factors (Fig. 7D and F). Cell migration and invasion assays further revealed that both capacities were significantly enhanced in the DKK1-overexpressing cells compared with the control group, in agreement with prior reports describing the pro-tumorigenic effects of DKK1 (Figs. 7E) [40]. Finally, to evaluate the impact of DKK1 overexpression on tumor progression in vivo, luciferase-labeled DKK1-overexpressing cells and control cells were injected into nude mice via tail vein injection (eight mice per group). The DKK1 overexpression group developed a significantly greater number of metastatic tumor cells than the control group (Fig. 7G and H). Collectively, these results demonstrate that the DKK1-overexpressing cell population exhibits markedly enhanced invasive and metastatic potential.

Discussion

LUAD, one of the most prevalent and lethal malignancies, is associated with a poor prognosis largely attributable to its complex biological behavior and immunosuppressive TME. In this study, a five-gene senescence-related prognostic model was developed using bulk RNA-seq data, while key senescence-associated biological pathways were further characterized through scRNA-seq analysis. These integrated approaches underscore the substantial contribution of cellular senescence to LUAD progression and highlight the potential clinical utility of the proposed prognostic model.

The genes included in the model contribute to cellular senescence and tumor biology through distinct molecular mechanisms. FGF2, a basic fibroblast growth factor secreted at elevated levels by senescent cells, functions as a survival factor and promotes polarization toward an M2-like macrophage phenotype, thereby modulating tumor immunity [41, 42]. GAPDH mediates oxidative stress responses, and its nuclear translocation can induce cell death, possibly contributing to cancer cell senescence [43]. ENO1, a glycolytic enzyme essential for energy homeostasis, is frequently implicated in tumor progression and metabolic adaptation in malignancies, including lung cancer [44, 45]. CCNA2, a cyclin that regulates cell proliferation and apoptosis, is closely associated with cancer progression and patient survival outcomes [46, 47]. DKK1 antagonizes Wnt signaling and thereby regulates cellular processes critical to senescence, apoptosis, differentiation, and metastasis across multiple cancer types [48].

Building on established knowledge regarding the high mortality of LUAD, the present study provides an integrated framework illustrating how senescence-associated gene signatures can stratify patient outcomes and remodel TME. Cohort analyses indicated that the aging-related risk score was generally higher in patients with advanced tumor stage, lymph node metastasis, and positive STAS. Notably, significant differences in risk scores were observed across distinct pathological subtypes. Compared with LPA, the risk scores for APA, PPA, and SPA were significantly elevated, with SPA exhibiting the highest score [49–52]. Moreover, LPA is associated with the most favorable prognosis, followed by APA and PPA, whereas SPA carries the poorest prognosis. These observations suggest that patients with more adverse clinical outcomes tend to exhibit higher risk scores, potentially reflecting an underlying enrichment of cellular senescence.

Notably, DKK1, one of the prognostic genes identified in the model, displayed the most significant differential expression between high-risk and low-risk patients. DKK1 can trigger early cellular senescence [53, 54], and its biological effects in tumor cells appear to be context dependent. Elevated DKK1 expression in solid tumors, including non–small-cell lung cancer, hepatocellular carcinoma, and pancreatic cancer [55–57], with positive expression correlating with increased risk of lymph node metastasis and reduced OS [58]. Importantly, DKK1 can induce the expression of multiple genes associated with macrophage-mediated inflammation and fibrosis, which is consistent with the TME characteristics observed in high-risk patients [59, 60]. In the present study, comprehensive analyses and experimental validation further demonstrated that DKK1 overexpression in lung cancer cell lines not only increased the expression of multiple senescence markers but also significantly enhanced tumor cell invasiveness in both in vitro and in vivo models.

Additionally, we identified a distinct epithelial cell subpopulation characterized by high COL17A1 expression that exhibited pronounced senescence-associated phenotypes driven by DKK1-dependent signaling. COL17A1 has been strongly associated with aggressive tumor phenotypes, including enhanced tumor-initiating capacity and increased metastatic potential, suggesting that this epithelial subset represents a functionally specialized population in which senescence and malignant traits coexist [61–64]. By integrating multiple senescence-related pathways with distinct molecular mechanisms and applying the validated hUSI and SenCID programs, we demonstrated that this cell population displays prominent senescence features. Notably, the SenCID program classified these cells into the SID2 model, consistent with previous reports describing senescent lung epithelial cells characterized by epidermal development features [39]. Furthermore, Yao et al. reported that high expression of the senescence-associated prognostic gene DKK1 enhances stem cell-like properties [65, 66].

Intercellular communication analysis further revealed that this senescence-related cluster exhibited strong autocrine and paracrine signaling activity. Within the epithelial cell interaction network, the most prominent pathways included LAMININ, MDK, COLLAGEN, MIF, and APP, all of which are intricately associated with tumor progression. Among these, MDK was particularly notable, as previous studies have demonstrated its overexpression in multiple cancer types [28–31]. MDK promotes tumor progression by positively regulating cell proliferation, anti-apoptotic signaling, metastasis, angiogenesis, and immune resistance [31–34]. In the cross-cell-type interaction network, this cluster primarily communicated with other cell populations, including fibroblasts and myeloid cells, through pathways such as COLLAGEN, LAMININ, and APP, with the APP–CD74 signaling axis emerging as especially prominent. Detailed analysis of the interactions among the E9 subpopulation, fibroblasts, and myeloid cells further revealed that the relevant signaling pathways are closely associated with cellular senescence and epidermal development processes.

Recent studies have extensively explored aging-related diseases and their therapeutic interventions [67]. The five-gene aging signature identified in this study was strongly associated with poor prognosis, and higher scores were also correlated with reduced responsiveness to immunotherapy, thereby providing valuable guidance for therapeutic decision-making. In the present study, the epithelial cell population characterized by DKK1 expression was confirmed to exhibit a senescent phenotype. Our findings suggest that these senescence-associated features substantially influence tumor progression and the composition of TME, highlighting this cell population as a promising target for cancer immunotherapy. Notably, DKK1 inhibitors have demonstrated the ability to prevent metastasis and exert antitumor effects in multiple malignancies, including gastroesophageal cancer and breast cancer bone metastasis, supporting their potential as a therapeutic strategy for metastatic disease [68, 69].

In conclusion, our integrative analyses provide a comprehensive framework for understanding how senescent cell populations shape TME and drive LUAD progression. By demonstrating the clinical relevance of a five-gene senescence-based prognostic model and identifying a distinct senescent epithelial subpopulation that promotes DKK1-driven oncogenic pathways, this study establishes a foundation for future therapeutic strategies aimed at disrupting senescence-mediated tumor evolution.

Methods

Bulk RNA-seq data collection and aging-related pathway characterization

Bulk RNA-seq data and corresponding clinical information for the TCGA-LUAD cohort were obtained from TCGA (https://portal.gdc.cancer.gov/). Microarray expression datasets for GSE31210 and GSE30219 were retrieved from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo). In addition, sequencing data from patients treated at Nanjing Medical University Affiliated Cancer Hospital were incorporated into the analysis. Collectively, these datasets included 541 tumor samples and 59 normal samples from TCGA-LUAD, 393 tumor samples used for clinical analyses, 226 samples from GSE31210, 85 samples from GSE30219, and 70 samples from Nanjing Medical University Affiliated Cancer Hospital. Furthermore, both the DKK1-overexpressing cell lines and the corresponding control cell lines constructed in this study were included in the downstream analyses. The SenMayo gene list was obtained from the GSEA website (https://www.gsea-msigdb.org/gsea/msigdb/index.jsp), and the AgingAtlas gene list was retrieved from https://ngdc.cncb.ac.cn/aging/index. To evaluate the applicability of the prognostic model and the biological relevance of DKK1 in the context of lung aging, pathways associated with cellular senescence were curated from the CellAge database, MSigDB gene sets, and relevant literature [70–72]. These pathways were categorized into three principal groups: CellAge-related pathways, aging/senescence-induced gene-related pathways, and SASP-related pathways. The ability of the model to detect cellular senescence was systematically evaluated by scoring these pathway groups.

Human tissue specimens

Clinical tumor specimens were collected from 70 patients with confirmed LUAD who underwent surgical resection at the Affiliated Cancer Hospital of Nanjing Medical University. Fresh tumor tissues were harvested intraoperatively, immediately snap-frozen in liquid nitrogen, and preserved at − 80 °C for subsequent genomic sequencing analyses. In addition, formalin-fixed, paraffin-embedded tumor specimens obtained from 52 patients were prepared for tissue microarray construction. All pathological diagnoses were independently examined and verified by two board-certified pathologists to ensure diagnostic reliability, consistency, and tissue integrity.

The inclusion criteria included histologically confirmed primary LUAD and the availability of adequate tumor tissue following surgical resection. The exclusion criteria included age younger than 18 years, diagnosis of other concurrent malignancies, or incomplete clinicopathological data. All participants provided written informed consent before sample collection. The study protocol was approved by the Ethics Committee of the Affiliated Cancer Hospital of Nanjing Medical University (Approval No. 2023092), and all procedures were conducted in accordance with the Declaration of Helsinki.

Construction and validation of prognostic age-related gene models

The R package DESeq2 was used to identify DEGs between tumor and normal samples in the TCGA-LUAD dataset, using |logFC| of > 1 and an adjusted p-value of < 0.05 as the filtering thresholds [73]. These DEGs were subsequently intersected with known aging-related genes, and univariate Cox regression analysis was performed to determine which genes were significantly associated with OS.

To minimize the risk of model overfitting, the LASSO algorithm was applied. A 10-fold cross-validation using the cv.glmnet function was conducted to identify the optimal penalty coefficient λ (lambda.min), which minimizes the partial likelihood deviance and facilitates the selection of the most informative features. Subsequently, 1,000 iterations were performed to identify prognostic genes based on the concordance index (C-index). Genes with an HR of > 1.0 and a p-value of < 0.05 were retained for model construction. Among these candidates, genes exhibiting the highest C-index values were selected as aging-related prognostic genes. Finally, multivariable Cox regression analysis was conducted to compute individual patient risk scores, resulting in the following formula for the ARS:

Using the median ARS as the cutoff value, patients were stratified into low-risk and high-risk subgroups. Kaplan–Meier survival curves were generated to compare OS between these two subgroups in both the training cohort and the external validation cohorts. In addition, the predictive performance of the model for PFS was further evaluated in the hospital cohort.

Gene functional enrichment and immune-related signature analyses

DEGs between the high-risk and low-risk groups were identified using the R package DESeq2, with log2FC of > 1.0 and a p-value of < 0.05 applied as the selection thresholds. The R package clusterProfiler was subsequently used to investigate the biological functions and pathways associated with genes that were upregulated in the high-risk group [74]. Specifically, KEGG pathway analysis, GO enrichment analysis, and GSEA were performed to characterize the underlying biological processes [75–77]. The R package ggplot2 was then used to construct heatmaps for data visualization. To further characterize differences in TME and immune cell infiltration between the two risk groups, the IOBR R package was applied [78]. Moreover, to assess differences in predicted immunotherapy responsiveness across the distinct risk groups, the TIDE algorithm was utilized, with TIDE scores obtained from the public portal (http://tide.dfci.harvard.edu/) [27]. Correction for multiple testing in this analysis was performed using the Benjamini–Hochberg procedure.

scRNA-seq analysis

scRNA-seq data generated by Philip Bischoff et al. and from the GSE131907 dataset were integrated for analysis, encompassing normal tissue samples from 21 patients, tumor tissue samples from 25 patients, and mLN samples from 7 patients. All scRNA-seq datasets analyzed in this study are publicly available through the Gene Expression Omnibus (GEO) and CodeOcean (https://codeocean.com/capsule/8321305/tree/v1) [79, 80].

Following the analytical framework described by Philip Bischoff et al., Seurat V5 was used for single-cell data processing and downstream analyses [81]. Gene expression data from all patients were merged, and quality control filtering was applied to retain only cells with 500–10,000 detected genes and 1,000–100,000 unique molecular identifiers, with fewer than 30% of reads mapping to mitochondrial genes and fewer than 5% mapping to hemoglobin genes [79]. After filtering, data normalization was performed using the SCTransform function. Batch effects were then corrected using the Harmony R package following principal component analysis [82]. The top 15 principal components were selected for Uniform Manifold Approximation and Projection clustering analysis, and the resulting embeddings were accordingly visualized. Cell type annotation was performed based on Nayoung Kim et al., and the DotPlot and ggplot2 functions were used to generate scatter plots [80].

Epithelial cells from tumor samples were extracted using the subset function, and cell clusters were identified using the FindClusters function with the resolution parameter set to 0.2. The AUCell package was applied to quantify cellular senescence feature scores [83]. Cell cycle status was determined using the CellCycleScoring function. The FindAllMarkers function (min.pct = 0.25, logFC > 0.25) was used to identify marker genes for each cluster, whereas the FindMarkers function (min.pct = 0.1, logFC > 0.25) was applied to detect DEGs between target cell populations and all other cells. Upregulated genes with log2FC of > 0.5 and an adjusted p-value of < 0.05 were subsequently subjected to KEGG, GO, and GSEA analyses. Cellular stemness was assessed using the CytoTrace package, followed by pseudotime trajectory analysis using the Monocle2 package [84–87]. Finally, cell–cell communication networks were explored using the CellChat and CellCall packages [88, 89].

Additional pseudotime analyses of epithelial cells derived from tLung, tL/B, and mLN were conducted using the CytoTrace and Monocle2 packages. The CellChat package was further applied for extended cell–cell interaction studies, and the CellCall package was used for pathway activity analyses and downstream transcription factor enrichment. Machine learning algorithms, hUSI and SenCID, were also used to identify and validate the aging status of the epithelial cell subgroups [38, 39].

Immunohistochemistry

Experiments were conducted in accordance with previously published protocols [90]. The primary antibody used for immunohistochemistry was an anti-DKK1 antibody (cat. no. ab307367; Abcam, Cambridge, UK). Immunohistochemical staining was performed to detect the protein expression of target molecules in the tumor tissues. Briefly, paraffin-embedded tissue sections were deparaffinized, rehydrated, and subjected to antigen retrieval in citrate buffer (pH = 6.0) using a microwave oven. Endogenous peroxidase activity was quenched with 3% hydrogen peroxide, and nonspecific binding was blocked with 5% goat serum. The sections were incubated with primary antibodies against the target proteins overnight at 4 °C, followed by incubation with HRP-conjugated secondary antibodies for 1 h at room temperature. The sections were then developed using 3,3′-diaminobenzidine and counterstained with hematoxylin. Images were acquired under an optical microscope. Staining intensity and average OD values were independently evaluated by two pathologists using ImageJ software.

Cell culture

The human non–small-cell lung cancer cell line A549 was obtained from the American Type Culture Collection and routinely maintained in RPMI-1640 medium supplemented with 10% fetal bovine serum (Corning, USA). The cells were cultured at 37 °C in a humidified incubator containing 5% CO₂. Before experimental use, cultures were systematically tested to exclude mycoplasma contamination, interspecies cross-contamination, and to verify cell line authenticity. All functional assays were conducted using cells within 20 passages to ensure experimental consistency and reproducibility.

Plasmid construction and transfection

Transient transfection was performed using Lipofectamine 3000 reagent (Thermo) in accordance with the manufacturer’s protocol. The cells were transfected with a DKK1 expression plasmid or the corresponding empty vector control, both obtained from GeneChem (Shanghai, China). To generate the DKK1 overexpression construct, full-length DKK1 cDNA was amplified using the following primers: forward, 5′-TGCGCTAGTCCCACCATGATGGCTCTGGGCGCA-3′; reverse, 5′-TAACAACGCTGGAATTTAGTGTCTCTGACAAGTGTGAAGC-3′. The amplified fragment was enzymatically digested with XhoI and EcoRI and subsequently subcloned into the pcDNA3.1 expression vector.

SA-β-gal staining

The SA-β-gal activity was evaluated using a chromogenic SA-β-gal Staining Kit (Beyotime Biotechnology, China). The cultured cells or tissue cryosections were fixed at room temperature for 15 min using the fixation solution supplied with the kit, followed by thorough washing with phosphate-buffered saline. The samples were then incubated with SA-β-gal staining solution at 37 °C for the specified duration. After staining, the cultured cells were rinsed with phosphate-buffered saline and examined under a light microscope. The tissue sections were subsequently washed, counterstained with eosin for 1–2 min, and mounted for imaging and analysis.

RNA extraction and qRT-PCR

Total RNA was extracted from cells using TRIzol reagent (Thermo Fisher Scientific), according to the manufacturer’s instructions. Purified RNA was reverse-transcribed into complementary DNA using an RNA-to-cDNA synthesis kit (Takara). Then, qRT-PCR was performed using SYBR Green-based detection chemistry (Vazyme) on a QuantStudio 6 Flex Real-Time PCR System (Applied Biosystems). Gene expressions were normalized to GAPDH, which served as the internal reference control. Primer sequences used for qRT-PCR analysis are provided in Supplementary Table 2.

ELISA

The concentrations of secreted cytokines and matrix metalloproteinase in cell culture supernatants were quantified using ELISA. Following the indicated treatments, conditioned media were collected, centrifuged to remove cellular debris, and stored at − 80 °C until analysis. The IL-1α, IL-1β, IL-6, IL-8, and MMP-9 levels were measured using commercially available ELISA kits, according to the manufacturers’ protocols.

Cell invasion and migration assay

Cell migratory and invasive capacities were assessed using Transwell chambers (Corning). For invasion assays, the upper surfaces of the inserts were precoated with Matrigel (BD Biosciences, USA), whereas migration assays were conducted in the absence of Matrigel. After the indicated treatments, the cells were seeded into the upper chambers, and migration or invasion toward the lower compartment was allowed to proceed for the specified duration. The cells that traversed the membrane and adhered to the lower surface were fixed, imaged, and quantitatively analyzed under a light microscope.

Animal studies

Female BALB/c nude mice (age: 4 weeks) were obtained from GemPharmatech and maintained under specific pathogen-free conditions at Nanjing Medical University. The animals were housed in a controlled environment with a 12-h light/dark cycle, a constant temperature of 22 °C, and a relative humidity of 55–70%, with unrestricted access to food and water. All animal experiments were conducted in strict accordance with institutional guidelines, and tumor burden did not exceed the maximum size permitted by the Institutional Animal Care and Use Committee (IACUC).

Tail vein metastasis models were established to assess the metastatic potential of tumor cells in vivo. Stable DKK1-overexpressing A549 cells and corresponding control cells (1 × 10⁶) were injected into the lateral tail vein of female BALB/c nude mice. Metastatic progression was monitored by in vivo bioluminescence imaging at the designated time points, and imaging data were quantitatively analyzed to determine metastatic burden. To ensure blinding during data collection, mouse handling and image acquisition were performed by independent investigators.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yujia Zhou, Chen Chen and Fengyi Zuo contributed equally to this work and were responsible for data analysis, material preparation, and manuscript drafting. Siqi Ding, Bin Zhu, Bangyu Wu, Chen Liu and Tianhao Yuan were involved in data collection and clinical sample management. Hui Wang and Xing Huang participated in material organization and manuscript writing. Xinyu Xu and Dawei Ma contributed to critical revision and editing of the manuscript. All authors reviewed and approved the final version of the manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (Grant No. 81902334), and the Medical Research Project of the Jiangsu Provincial Health Commission (Grant No. H2023007).

Data availability

All data generated in this study are available upon request from the corresponding authors.

Declarations

Competing interests

The authors declare no competing interests.