Immune‐Like Malignant Epithelial Programs Shape Tumor–Immune Interactions and Inform Prognostic Stratification in Lung Adenocarcinoma

ABSTRACT

Lung adenocarcinoma (LUAD) is characterized by marked cellular heterogeneity, yet how malignant epithelial states contribute to immune regulation and clinical outcomes remains incompletely defined. We integrated single‐cell RNA‐sequencing data to map the cellular landscape of LUAD and identify malignant epithelial cells based on inferred copy‐number alterations. Epithelial states were further examined through trajectory inference, transcription factor analysis, and cell–cell communication profiling. Single‐cell‐derived genes were subsequently integrated with TCGA and independent GEO cohorts to construct and validate a machine learning‐based prognostic signature. Malignant epithelial cells displayed distinct functional programs, including an immune‐like state associated with genomic instability, immune‐related transcriptional activity, tumor–immune communication, and patient outcomes. The resulting immune‐like malignant epithelial cell signature (IMEC‐Sig) consistently stratified survival across multiple cohorts. Low IMEC‐Sig scores were accompanied by greater immune infiltration, higher immune checkpoint expression, and increased immunophenoscore, whereas high scores were linked to a comparatively immunosuppressive phenotype. Pan‐cancer analyses further identified KRT8 as a gene associated with unfavorable prognosis, and functional experiments showed that KRT8 silencing suppressed proliferation, migration, invasion, and colony formation in LUAD cells. Together, these findings connect malignant epithelial heterogeneity with the immune context and clinical outcomes, support IMEC‐Sig as a biologically informed prognostic tool, and nominate KRT8 as a potential therapeutic target in LUAD.

Immune‐like malignant epithelial cells are linked to tumor–immune remodeling and LUAD prognosis. The IMEC‐derived signature enables risk stratification, while KRT8 contributes to aggressive tumor behavior.

1. Introduction

Lung adenocarcinoma (LUAD) represents the predominant histological subtype of nonsmall cell lung cancer and remains a major contributor to cancer‐related mortality worldwide [1, 2]. Although therapeutic strategies, including surgical resection, chemotherapy, targeted therapy, and more recently immunotherapy, have advanced considerably, the clinical outcomes for patients with LUAD remain suboptimal, particularly in advanced‐stage disease where treatment options are constrained and responses are heterogeneous [3, 4]. The introduction of immune checkpoint blockade has led to meaningful clinical benefits in a fraction of patients [5, 6]; however, durable responses are limited and inter‐individual variability remains substantial [7]. Current biomarkers, such as programmed death‐ligand 1 expression and genomic indicators like tumor mutational load, exhibit restricted reliability and fail to adequately capture the underlying biological complexity of tumors [8]. Consequently, a deeper characterization of tumor heterogeneity at the cellular level is essential to enable the development of more precise prognostic models and therapeutic strategies.

The rapid development of scRNA‐seq has enabled high‐resolution characterization of tumor cellular composition and evolutionary dynamics [9, 10, 11, 12]. Previous studies have demonstrated that malignant epithelial cells exhibit substantial heterogeneity in genomic alterations, transcriptional states, and functional phenotypes [13, 14]. Nevertheless, most existing research has primarily focused on immune components or the global tumor microenvironment, while the functional diversity within malignant epithelial subpopulations and their roles in tumor progression remain insufficiently explored [12]. Recent multiomics analysis has also identified a high‐risk stem‐like malignant cell subgroup associated with adverse prognosis and an immunosuppressive microenvironment, further highlighting the clinical relevance of resolving tumor‐cell heterogeneity at the cellular level [15]. In addition, single‐cell analysis typically relies on Python‐based frameworks (e.g., Scanpy) [16, 17] for efficient preprocessing, dimensionality reduction, and clustering, whereas downstream statistical modeling and clinical association analyses are often performed using R‐based tools (e.g., survival and limma). The integration of Python and R analytical ecosystems thus provides a complementary strategy, enabling a more comprehensive transition from cellular‐level insights to clinically relevant interpretations [18].

Within the tumor microenvironment, interactions among different cell types critically influence tumor development and progression [19, 20]. In particular, myeloid populations, such as tumor‐associated macrophages, play essential roles in modulating immune responses, promoting tumor growth, and facilitating immune evasion [21]. Recently, ligand–receptor‐based communication analysis has emerged as a powerful approach to dissect intercellular signaling networks [22, 23]. However, how distinct malignant epithelial states participate in and reshape these communication networks, and how they contribute to immune microenvironment remodeling, remains largely unclear [24, 25].

From a clinical perspective, gene expression‐based prognostic models have become important tools for risk stratification and outcome prediction [26, 27]. However, conventional models are generally derived from bulk transcriptomic data and often lack the resolution to capture tumor cell heterogeneity, thereby limiting their predictive performance [28]. Integrating single‐cell‐derived cellular features with machine learning approaches may therefore offer a more biologically informative and clinically relevant framework for prognostic modeling [29].

In this study, single‐cell and bulk transcriptomic datasets were integrated to comprehensively characterize epithelial heterogeneity in LUAD. Malignant epithelial populations were defined according to inferred copy number alterations and further subdivided, with particular attention to immune‐like malignant states. Through the integration of trajectory analysis, transcription factor regulatory network inference, and cell–cell interaction analysis, their roles in tumor progression and immune modulation were elucidated. A prognostic signature (IMEC‐Sig) was subsequently established based on genes derived from immune‐like malignant epithelial cells and demonstrated stable predictive capacity across independent validation cohorts. Furthermore, KRT8 was identified as a key gene and subjected to additional evaluation through pan‐cancer analysis and in vitro functional experiments.

Overall, this work presents a single‐cell‐guided framework that connects tumor heterogeneity with the immune microenvironment and clinical outcomes in LUAD, providing a basis for improved prognostic assessment and therapeutic strategies.

2. Methods

2.1. Collection and Filtering of Multicohort Transcriptomic Data

Single‐cell transcriptomic data used in this study were derived from a publicly available Code Ocean resource (https://codeocean.com/capsule/8321305/tree/v1), which contains expression profiles from primary LUAD tumors together with matched adjacent nontumor tissues. For bulk transcriptome analysis, RNA expression data and associated clinicopathological information were gathered from the TCGA–LUAD cohort and additional GEO datasets. Cases with incomplete survival records were removed before further analysis. After data curation, the TCGA–LUAD cohort was used as the training set, whereas external GEO cohorts served as independent validation datasets. To account for nonbiological differences introduced by distinct platforms and study sources, cross‐cohort batch effects were handled using the sva package [30] or the Harmony [31] method as appropriate for the corresponding analytical workflow.

2.2. scRNA‐Seq Data Processing and Clustering Analysis

Single‐cell RNA‐sequencing data were processed using Scanpy in Python. Genes detected in fewer than three cells were removed. Cells with fewer than 200 or more than 6000 detected genes, more than 40,000 total counts, or a mitochondrial transcript proportion above 20% were excluded. After quality control, total counts were normalized to 10,000 per cell and log‐transformed. The top 2000 highly variable genes were selected, followed by principal component analysis. The first 30 principal components were retained for downstream analysis. Batch effects were corrected using Harmony with sample identity as the batch variable. A neighborhood graph was constructed using 15 nearest neighbors, followed by Leiden clustering at a resolution of 0.5 and UMAP visualization. Cell type annotation was conducted using the CellTypist [32] tool based on reference expression profiles. Additionally, differentially expressed genes among clusters were identified to support annotation accuracy. This analytical pipeline enabled subsequent investigation of tumor microenvironment characteristics, cellular subpopulation delineation, and functional state assessment.

2.3. Copy Number Variation Analysis Using Infercnvpy

Copy number variation (CNV) profiles of epithelial cells were inferred using infercnvpy [33] in Python. Epithelial cells derived from adjacent nontumor tissues were used as the reference population. Gene expression values were ordered according to their genomic positions and smoothed across chromosomal windows to generate relative CNV signals. CNV‐based clustering was subsequently performed to identify epithelial populations with similar alteration patterns. CNV scores were calculated using infercnvpy.tl.cnv_score as the mean absolute value of the inferred CNV signals within each CNV cluster, and the corresponding cluster score was assigned to its member cells. The 90th percentile of the CNV scores observed in reference normal epithelial cells was used as the cutoff. Epithelial cells with CNV scores above this threshold were classified as malignant, whereas the remaining cells were classified as nonmalignant.

2.4. Pseudotime Trajectory Analysis

To investigate dynamic changes in cellular states, pseudotime trajectory analyses were performed using both Monocle2 [34] and graph‐based approaches implemented in the Scanpy framework.

First, trajectory inference was conducted in R using Monocle2. Differentially expressed genes were selected as ordering features, followed by manifold learning and the construction of a minimum spanning tree to reconstruct cellular progression. Cells were then ordered along the inferred trajectory to represent their potential transitional states and branching structures. In parallel, trajectory analysis was performed in Python using Scanpy. A neighborhood graph was constructed to capture cell–cell relationships, and partition‐based graph abstraction (PAGA) [35] was applied to describe the connectivity between cell clusters and the global topology of the dataset. Based on this graph structure, diffusion pseudotime (DPT) [36] was further utilized to assign pseudotime values to individual cells, enabling the visualization of continuous state transitions in low‐dimensional space.

2.5. Transcription Factor Regulatory Network Analysis (pySCENIC)

To explore transcriptional regulatory mechanisms across different cell populations, gene regulatory network analysis was performed using the pySCENIC [37] framework in Python. Initially, co‐expression analysis was conducted based on the gene expression matrix to identify potential relationships between transcription factors and their target genes. These co‐expression modules were then refined by incorporating transcription factor binding motif information to obtain high‐confidence regulons. Subsequently, the activity of each regulon was quantified at the single‐cell level, and a regulon activity matrix was constructed to compare regulatory differences among cell populations. By integrating cell subtype information, key transcription factors were further identified, and their activity patterns across distinct cellular states were characterized to elucidate potential regulatory mechanisms.

2.6. Inference of Cell–Cell Communication Networks

To systematically investigate intercellular interactions among different cell populations, the CellPhoneDB [38] framework was applied to infer cell–cell communication. Briefly, input data were prepared based on cell type annotations, and normalized expression matrices were used to identify ligands and receptors expressed in each cell population. Statistical significance of ligand–receptor interactions was assessed through permutation testing, allowing the identification of significantly enriched interaction pairs. These significant interactions were further integrated to construct a cell–cell communication network, enabling the characterization of signaling patterns between different cell types and their potential biological functions. To facilitate interpretation, significant interactions were visualized to illustrate interaction strength and overall communication patterns among cell populations, providing insights into the complex cellular crosstalk within the tumor microenvironment.

2.7. Prognostic Analysis of Cell Population‐Associated Signatures

The prognostic relevance of distinct cell populations was investigated by first deriving cell type‐specific marker genes from single‐cell RNA sequencing data based on differential expression analysis. These marker sets were subsequently integrated with bulk transcriptomic datasets to enable enrichment‐based quantification. Specifically, single‐sample gene set enrichment analysis (ssGSEA) [39] was applied to the TCGA‐LUAD cohort to generate cell population‐associated scores for each patient, representing the relative abundance of corresponding cell subsets. Patients were then stratified according to these scores, and survival analyses were conducted by incorporating clinical outcome data to evaluate the relationship between cell population‐related features and patient prognosis.

2.8. Screening of Candidate Genes and Development of a Prognostic Model

Differential expression analysis was conducted in the TCGA–LUAD cohort to screen genes showing significant differences between tumor and normal tissues. The identified genes were subsequently integrated with marker genes derived from the target tumor cell population to generate a refined set of candidate genes. Feature selection was performed using a random forest algorithm, and genes with higher importance scores were retained for further analysis. These candidates were further evaluated for their prognostic relevance through univariate Cox regression, and those exhibiting significant associations with survival were incorporated into model development. Based on these selected genes, multiple machine learning approaches were applied to construct prognostic models. Model performance was systematically assessed in the training cohort, and the optimal model was determined according to the concordance index (C‐index) for downstream analyses. The optimal IMEC‐Sig model was generated using the combination of LASSO and SuperPC and comprised 10 genes: NUPR1, TM6SF1, DNAJB4, RHOV, NGEF, KRT8, CENPM, PTTG1, GAPDH, and LDHA. Individual IMEC‐Sig scores were calculated for each patient using the fitted model.

2.9. Survival Analysis and Model Performance Evaluation

Following model establishment, an IMEC‐Sig score was calculated for each patient. In the TCGA–LUAD training cohort, patients were divided into high‐ and low‐risk groups using the median IMEC‐Sig score of the training cohort as the cutoff. In each external validation cohort, the cohort‐specific median IMEC‐Sig score was used for risk stratification rather than the cutoff derived from TCGA–LUAD. This approach was adopted to account for differences in expression platforms and score distributions across cohorts. Differences in overall survival between these subgroups were examined using Kaplan–Meier estimation, and statistical significance was determined by the log‐rank test. To further assess the discriminative capability of the model, time‐dependent receiver operating characteristic (ROC) analysis was conducted at 1, 3, and 5 years. The corresponding area under the curve (AUC) values were calculated to quantify the predictive performance of the model for both short‐term and long‐term survival outcomes.

2.10. Characterization of the Tumor Immune Microenvironment and Immunotherapy‐Associated Features

TME and immunotherapy‐related characteristics were systematically profiled across different risk subgroups. Bulk transcriptomic data were analyzed using multiple immune infiltration estimation methods to quantify the composition of immune cell populations within each sample. Variations in immune cell abundance between high‐ and low‐risk groups were further compared to delineate distinct immune landscapes. In addition, the expression patterns of immune checkpoint‐associated genes were evaluated across risk groups to characterize immunoregulatory differences. Correlation analyses were subsequently performed to explore the relationships between risk scores and checkpoint gene expression levels. IPS was also evaluated as an indirect indicator of tumor immunogenicity and immunotherapy‐associated immune features. Differences in IPS between risk groups were compared to characterize variations in immune status rather than to directly predict clinical response to immunotherapy.

2.11. Pan‐Cancer Evaluation of Gene Expression and Prognostic Significance

A pan‐cancer analysis was performed to investigate the expression profile and clinical relevance of the target gene across multiple tumor types. Transcriptomic data derived from The Cancer Genome Atlas (TCGA) were utilized to assess gene expression patterns across diverse cancer tissues. Comparative analyses were conducted to evaluate differential expression across tissue origins. To further explore its prognostic implications, survival analyses were carried out in different cancer types using multiple clinical endpoints. These analyses were designed to comprehensively characterize the association between gene expression levels and patient outcomes across cancers.

2.12. Cell Culture

The human LUAD cell lines A549 and H1299 were obtained from a certified cell repository and cultured in RPMI‐1640 or DMEM supplemented with 10% fetal bovine serum (FBS). Cells were maintained in a humidified incubator at 37°C with 5% CO₂ and routinely passaged upon reaching 70%–80% confluence for subsequent experiments.

2.13. RNA Isolation and Quantitative Real‐Time PCR Analysis

Total RNA was isolated from cultured cells following standard extraction procedures. The extracted RNA was subsequently reverse‐transcribed into complementary DNA (cDNA) using a commercial reverse transcription kit. Quantitative real‐time PCR (qRT‐PCR) was performed to assess gene expression levels according to the manufacturer's instructions. GAPDH was used as the internal control, and relative expression levels were calculated using the 2^−ΔΔCt method.

The primer sequences used for KRT8 were as follows: forward primer, 5′‐CGAGGATATTGCCAACCGCAG‐3′; reverse primer, 5′‐CCTCAATCTCAGCCTGGAGCC‐3′.

Small interfering RNAs targeting KRT8 were as follows: si‐KRT8‐1, sense 5′‐CUGAGAUGAACCGGAACAU‐3′; and si‐KRT8‐2, sense 5′‐GGGUGACCCAGAAGUCCUA‐3′.

2.14. Cell Viability Assessment Using CCK‐8 Assay

Cell viability was evaluated using the cell counting kit‐8 (CCK‐8). Cells were plated into 96‐well plates and cultured under standard conditions. At designated time points, CCK‐8 reagent was added to each well followed by incubation. Absorbance at 450 nm was then measured with a microplate reader to determine cell proliferation and viability.

2.15. Assessment of DNA Synthesis by EdU Incorporation

DNA synthesis activity was assessed using an EdU incorporation assay. Cells were treated with EdU reagent and subsequently subjected to fixation, permeabilization, and staining procedures following the manufacturer's protocol. EdU‐positive cells were visualized under a fluorescence microscope, and the proportion of proliferating cells was quantified.

2.16. Evaluation of Cell Migration and Invasion Using Transwell Assays

Cell migratory and invasive capacities were examined using Transwell chamber systems. For migration assays, cells suspended in serum‐free medium were seeded into the upper chamber, while medium supplemented with serum was placed in the lower chamber to serve as a chemoattractant. For invasion assays, the upper chamber was precoated with Matrigel prior to cell seeding. After incubation, nonmigrated cells on the upper surface were removed, whereas cells that traversed the membrane were fixed, stained, and counted under a microscope.

2.17. Colony Formation Assay for Clonogenic Capacity

Cells were plated at low density in culture dishes and maintained for approximately 10–14 days, with medium replaced periodically. Upon visible colony formation, cells were fixed and stained, and the number of colonies was counted to evaluate clonogenic potential.

2.18. Statistical Analysis

All in vitro experiments were performed in three independent biological replicates, and quantitative data are presented as the mean ± standard deviation (SD). Differences among the si‐NC, si‐KRT8‐1, and si‐KRT8‐2 groups were assessed using one‐way analysis of variance followed by Tukey's multiple‐comparison test. CCK‐8 time‐course data were analyzed using two‐way analysis of variance. All statistical analyses were conducted using R software and Python. Comparisons between two groups were performed using Student's t‐test or the Wilcoxon rank‐sum test, while comparisons among multiple groups were conducted using one‐way analysis of variance (ANOVA) or the Kruskal–Wallis test, depending on data distribution. Categorical variables were compared using the chi‐square test or Fisher's exact test, as appropriate. Survival curves were estimated using the Kaplan–Meier method, and differences between groups were assessed using the log‐rank test. Univariate and multivariate Cox proportional hazards regression analyses were performed to evaluate the association between variables and survival outcomes. The predictive performance of the models was primarily assessed using the concordance index (C‐index). Correlation analyses were conducted using Pearson or Spearman correlation coefficients, depending on the data characteristics. When applicable, multiple testing correction was performed using the false discovery rate (FDR) method. All statistical tests were two‐sided, and a p < 0.05 was considered statistically significant.

3. Results

3.1. Mapping Cellular Composition of LUAD at Single‐Cell Resolution

Single‐cell transcriptomic data from LUAD and matched normal tissues were integrated, and a unified expression matrix was generated after quality control for downstream analyses. Dimensionality reduction was initially performed using principal component analysis, followed by batch effect correction with the Harmony algorithm. Comparison of cell distributions before and after integration indicated effective removal of technical variation across samples, ensuring data consistency for subsequent analyses (Figure 1A). UMAP was then applied for nonlinear dimensionality reduction. Cells from different patients exhibited partial overlap in the embedding space (Figure 1B). When stratified by tissue origin, tumor and normal cells displayed distinct distribution patterns (Figure 1C). Unsupervised clustering identified 19 transcriptionally distinct cell populations (Figure 1D). To assign biological identities, cell types were first annotated using the CellTypist framework, followed by manual validation based on established marker gene expression. Cells were subsequently classified into major cell types, including epithelial cells, T cells, B cells, myeloid cells, NK cells, endothelial cells, and fibroblasts (Figures 1E and S1A). The expression profiles of representative marker genes were consistent with the assigned annotations (Figures 1G and S1B), supporting the reliability of cell type identification. Analysis of cellular composition revealed that epithelial cells, T cells, and B cells were relatively enriched in tumor tissues, whereas myeloid cells, NK cells, and endothelial cells were comparatively reduced (Figure 1F). Epithelial cells, T cells, and B cells were relatively enriched in tumor tissues, whereas myeloid cells, NK cells, and endothelial cells showed lower relative proportions. These findings describe differences in cellular composition between tumor and normal tissues but do not directly indicate changes in the functional activity of these populations. Furthermore, notable variability in cellular composition was observed across individual patients (Figure 1F), highlighting substantial inter‐patient heterogeneity in LUAD.

FIGURE 1.

Construction of the single‐cell transcriptomic landscape and cell type annotation. (A) Comparison of cell distribution in low‐dimensional space before and after batch effect correction using the Harmony algorithm. (B) UMAP visualization of cells colored by patient origin. (C) UMAP visualization of cells grouped by tissue type (tumor vs. normal). (D) UMAP projection showing unsupervised clustering of cell populations. (E) UMAP visualization of annotated cell types. (F) Proportion of different cell types across tissue types (left) and across individual samples (right). (G) Expression patterns of representative marker genes across cell populations projected onto UMAP.

3.2. CNV‐Based Classification of Malignant Epithelial Cell Populations

To investigate malignant features within epithelial cell populations, CNV profiles were inferred using infercnvpy, with normal epithelial cells from adjacent nontumor tissues serving as the reference population. Normal epithelial cells served as the reference to establish baseline chromosomal variation patterns. Distinct CNV patterns were observed between different epithelial cell populations, with tumor‐derived cells exhibiting more extensive copy number alterations (Figure 2A,B). When visualized according to tissue origin, tumor and normal epithelial cells displayed distinct distributions in the UMAP space (Figure 2C). Clustering based on CNV features identified multiple epithelial subpopulations with distinct CNV patterns (Figure 2D). In addition, CNV scores exhibited a continuous distribution across epithelial cells (Figure 2E). Based on the CNV score distribution in reference normal epithelial cells, the 90th percentile was selected as the cutoff. Epithelial cells with scores above this threshold were classified as malignant, whereas those with scores at or below the threshold were classified as nonmalignant (Figure 2F). To further explore intratumoral heterogeneity, malignant epithelial cells were re‐clustered, revealing multiple subpopulations with distinct transcriptional characteristics (Figure 2G). According to marker gene expression patterns, malignant epithelial cells were annotated into several functional states, including Differentiated tumor, Immune‐like tumor, Remodeling tumor, and Stress‐like tumor (Figure 2H,I). Further analysis of transcriptional regulatory features revealed distinct transcription factor activity patterns across these subpopulations. Specifically, differentiated tumor cells showed higher activity of transcription factors such as NR2F2 and FOXA1; remodeling tumor cells were characterized by enrichment of FOSL1 and ETS1; stress‐like tumor cells exhibited elevated activity of PAX7 and HOXA1, while immune‐like tumor cells were associated with transcription factors such as BARX1 (Figure 2J).

FIGURE 2.

Identification of malignant epithelial cells and characterization of their transcriptional features based on CNV analysis. (A) Heatmap of inferred chromosomal copy number variations (CNVs) in epithelial cells, grouped by tissue type (normal vs. tumor). (B) Genome‐wide CNV profiles inferred from single‐cell transcriptomic data, displayed according to cnv_leiden clusters. (C) UMAP visualization of epithelial cells grouped by tissue type. (D) UMAP projection of epithelial cell clusters (cnv_leiden) identified based on CNV features. (E) UMAP projection of epithelial cells colored by CNV score. (F) Classification of epithelial cells into malignant and nonmalignant populations based on CNV scores. (G) Unsupervised clustering of malignant epithelial cells. (H) Annotation of malignant epithelial cell subpopulations. (I) Expression patterns of representative marker genes across malignant epithelial subpopulations. (J) Transcription factor activity analysis across malignant epithelial subpopulations.

3.3. Pseudotime Trajectory Analysis Reveals Differentiation Hierarchy of Normal Epithelial Cells

To further characterize differentiation states and potential transitional relationships among normal epithelial cells, nonmalignant epithelial cells were extracted according to the CNV‐based classification generated using infercnvpy and were subsequently subjected to subpopulation clustering and pseudotime trajectory analysis. Unsupervised clustering identified multiple transcriptionally distinct subpopulations (Figure 3A). Based on the expression patterns of established epithelial markers, these clusters were categorized into four major epithelial cell types, including AT1, AT2, ciliated, and club cells (Figure 3B). Distinct marker gene expression profiles were observed across subpopulations, with AGER predominantly enriched in AT1 cells, SFTPA1 mainly expressed in AT2 cells, SCGB3A1 highly expressed in club cells, and TPPP3 primarily detected in ciliated cells (Figure 3C,D). To further explore the relationships among epithelial subpopulations, PAGA was applied to infer the topological connectivity. Strong connections were observed between AT2 and AT1 cells, as well as between AT2 and club cells, while club cells also exhibited connectivity with ciliated cells (Figure 3E), suggesting that AT2 cells may occupy an intermediate position along multiple differentiation trajectories. Trajectory embedding revealed that different epithelial cell types were distributed along a continuous structure, with AT2 cells primarily located in the intermediate region of the trajectory, whereas AT1 and club cells were positioned at the terminal ends of distinct branches. Ciliated cells were mainly enriched in another branch of the trajectory (Figure 3F). When colored by pseudotime values, cells exhibited a gradual transition from low to high pseudotime, with AT2 cells predominantly distributed in early‐to‐intermediate stages, while AT1 and ciliated cells were mainly located at later stages. Club cells followed an alternative branch trajectory (Figure 3G). Further comparison of pseudotime values across subpopulations showed that AT2 cells were generally positioned at earlier stages, whereas AT1 and ciliated cells displayed relatively higher pseudotime values. Club cells exhibited a distinct distribution pattern, indicating divergent differentiation trajectories among normal epithelial subtypes (Figure 3H).

FIGURE 3.

Characterization of normal epithelial cell subsets and trajectory inference based on CNV classification using infercnvpy. (A) UMAP projection of epithelial cells classified as nonmalignant according to the infercnvpy‐derived CNV scores, with cells colored by cluster identity. (B) Distribution of annotated normal epithelial cell subtypes (AT1, AT2, Ciliated, and Club). (C) Expression patterns of representative marker genes across normal epithelial cell subpopulations. (D) Expression levels of typical marker genes in different normal epithelial cell subtypes. (E) Connectivity among normal epithelial cell subpopulations inferred by partition‐based graph abstraction (PAGA). (F) Distribution of normal epithelial cells in the trajectory embedding space, colored by cell subtypes. (G) Pseudotime ordering of normal epithelial cells, colored by pseudotime values. (H) Comparison of pseudotime values across different normal epithelial cell subpopulations.

3.4. Transcriptional Continuity Between AT2 and Malignant Epithelial Cells Inferred by Pseudotime Analysis

Given that AT2 cells occupied an early‐to‐intermediate position in the inferred trajectory of normal epithelial cells and showed connectivity with multiple epithelial subtypes, AT2 cells were jointly analyzed with malignant epithelial cells to examine their transcriptional relationships and potential state continuity. Monocle2 generated a branched pseudotime structure containing both AT2 and malignant epithelial cells (Figure 4A). Cells were distributed along several connected branches, with a gradual change in inferred pseudotime values (Figure 4B). AT2 cells were mainly enriched in the lower‐pseudotime region, whereas malignant epithelial subpopulations occupied intermediate, higher‐pseudotime, or branch‐specific regions (Figure 4C–E). These relative positions indicate transcriptional differences and potential continuity among the analyzed epithelial states but do not establish a direct lineage transition from AT2 cells to malignant cells. Monocle2 further identified two major trajectory branches, along which branch‐associated genes displayed distinct expression patterns (Figure 4F). Functional enrichment analysis suggested that these gene modules represented different transcriptional programs associated with the malignant epithelial states.

FIGURE 4.

Pseudotime analysis of transcriptional relationships between AT2 and malignant epithelial cells. (A) Pseudotime trajectory of AT2 and malignant epithelial cells constructed using Monocle2 and colored by inferred cell state. (B) Distribution of cells along the trajectory, colored by pseudotime values. (C) Distribution of AT2 cells and malignant epithelial subpopulations within the inferred trajectory. (D) Cell‐type composition across the trajectory. (E) Comparison of pseudotime distributions among the epithelial populations. (F) Expression patterns of branch‐associated genes and their corresponding functional enrichment across the inferred trajectory branches.

3.5. Characterization of Myeloid Cell Subpopulations and Monocyte–Macrophage Dynamics

To systematically characterize the composition and state transitions of myeloid cells within the tumor microenvironment, myeloid cells were re‐clustered for further analysis. Unsupervised clustering identified multiple transcriptionally distinct subpopulations (Figure 5A). Based on CellTypist annotation and expression features, these cells were further classified into major myeloid cell types, including monocytes, macrophages, and dendritic cells (Figure 5B,C). Distinct expression patterns of marker genes were observed across different subpopulations (Figure 5D), supporting the reliability of cell type annotation. In terms of cellular composition, notable variability in myeloid cell proportions was observed across patients (Figure 5E). Compared with normal tissues, tumor samples exhibited a higher relative proportion of macrophages and a lower relative proportion of monocytes, indicating differences in myeloid cell composition between the two tissue types. Topological relationships among myeloid subpopulations were further assessed using PAGA analysis, revealing strong connectivity between monocytes and macrophages (Figure 5F). Based on this, pseudotime analysis was performed focusing on monocytes and macrophages. Trajectory embedding showed that monocytes were primarily located at the early stage of the trajectory, whereas macrophages were progressively distributed toward intermediate and late stages (Figure 5G). When colored by pseudotime values, cells exhibited a continuous transition from low to high pseudotime along the trajectory (Figure 5H). Dynamic gene expression analysis along pseudotime revealed that subsets of genes displayed stage‐specific expression patterns, with certain genes highly expressed at early stages and others gradually upregulated at later stages (Figure 5I), reflecting transcriptional reprogramming during the transition from monocytes to macrophages. In addition, transcription factor activity analysis showed distinct patterns across myeloid subpopulations (Figure 5J), indicating differences in transcriptional regulatory programs associated with different cellular states.

FIGURE 5.

Identification, composition, and trajectory analysis of myeloid cell subpopulations. (A) UMAP visualization of myeloid cell clusters identified by unsupervised clustering (myeloid_leiden). (B) Distribution of myeloid cell subpopulations annotated using the CellTypist framework (majority_voting). (C) Distribution of myeloid cell subpopulations after simplified annotation (majority_simplified). (D) Expression patterns of representative marker genes across myeloid cell subpopulations. (E) Proportion of myeloid cell subpopulations across individual samples (left) and between normal and tumor tissues (right). (F) Connectivity among myeloid cell subpopulations inferred by partition‐based graph abstraction (PAGA). (G) Distribution of myeloid cells in the trajectory embedding space, colored by cell subpopulations. (H) Pseudotime ordering of myeloid cells, colored by pseudotime values. (I) Dynamic expression patterns of representative genes along pseudotime. (J) Transcription factor activity analysis across myeloid cell subpopulations.

3.6. CellPhoneDB‐Based Analysis of Intercellular Communication in LUAD

To further characterize intercellular interactions within the LUAD microenvironment, CellPhoneDB was applied to systematically infer ligand–receptor‐mediated communication. The number and inferred strength of ligand–receptor interactions varied across cell populations, with a greater number of predicted interactions observed among several malignant epithelial subpopulations, particularly remodeling tumor, stress‐like tumor, and differentiated tumor cells (Figure 6A). Comparison between normal and tumor tissues revealed distinct communication patterns (Figure 6B,C). In normal tissues, interactions were mainly observed between epithelial cells and a subset of myeloid populations. In contrast, tumor tissues exhibited a larger and more complex network of predicted ligand–receptor interactions, characterized by enhanced communication among malignant epithelial subpopulations and increased interactions between tumor cells and myeloid cells.

FIGURE 6.

Inferred intercellular communication among malignant epithelial and tumor microenvironmental cell populations in LUAD. (A) Overview of the predicted ligand–receptor interaction network among the annotated cell populations. (B) Predicted intercellular communication network in adjacent nontumor tissues. (C) Predicted intercellular communication network in LUAD tumor tissues. (D) Significant ligand–receptor pairs predicted between myeloid subpopulations and malignant epithelial states in tumor tissues. The displayed interactions include FN1‐, SPP1‐, APP/APLP2‐, chemokine‐, AREG‐, ANXA1‐, PLAUR‐, prostaglandin‐, and semaphorin‐associated pairs.

Analysis of significant ligand–receptor pairs in tumor tissues identified multiple predicted interactions between myeloid and malignant epithelial populations (Figure 6D). Macrophage‐related populations showed frequent predicted interactions with malignant epithelial cells, including FN1‐ and SPP1‐associated adhesion pairs. APP–CD74, APLP2‐associated pairs, and CXCL12‐ or CXCL14‐related chemokine interactions were also detected across several myeloid–tumor cell combinations. In addition, predicted AREG–EGFR/ERBB, ANXA1–FPR, PLAUR‐associated, prostaglandin, and semaphorin interactions were observed in selected cell pairs. These results identify transcript‐based ligand–receptor associations within the tumor microenvironment but do not directly demonstrate physical ligand–receptor binding, downstream pathway activation, or specific cellular functions.

3.7. Identification of Immune‐Like Malignant Cell‐Associated Genes and Prognostic Candidates

To further explore the prognostic relevance of different malignant epithelial cell subpopulations, Kaplan–Meier survival analysis was conducted in the TCGA–LUAD cohort based on subpopulation‐specific scores. Distinct survival patterns were observed across subpopulations, with the immune‐like tumor score showing a strong association with patient outcomes (Figure 7A). Further comparison of CNV scores among malignant epithelial subpopulations revealed that the immune‐like tumor subgroup exhibited relatively elevated CNV levels (Figure 7B), suggesting increased genomic instability. Based on its higher CNV score and its association with prognosis, this subpopulation was considered a potential key target for further investigation. Transcriptomic comparison between tumor and normal tissues in the TCGA–LUAD dataset identified 1240 upregulated genes and 1927 downregulated genes (Figure 7C). These genes were subsequently overlapped with marker genes derived from the immune‐like tumor subpopulation to generate a refined candidate gene set (Figure 7D). Functional enrichment analysis indicated that these genes were predominantly involved in tumor‐associated biological processes, including extracellular matrix organization, cell adhesion, and cell migration (Figure 7E). Additional enrichment was observed in immune regulation, inflammatory responses, and cytokine‐mediated signaling pathways. Several genes were also linked to cell proliferation and stress response‐related processes. A random forest algorithm was then applied for feature selection. The model error decreased and gradually stabilized as the number of trees increased (Figure 7F), indicating good model convergence. Based on variable importance scores, the top 60 genes with the greatest contributions were selected (Figure 7G). Univariate Cox regression analysis was subsequently performed to screen genes significantly associated with patient survival (Figure 7H), providing candidate genes for subsequent prognostic model construction. Among these candidates, KRT8 was retained through the successive feature‐selection and prognostic‐screening procedures and was subsequently included in the final IMEC‐Sig model. Thus, KRT8 originated from the immune‐like tumor‐associated candidate set but was not considered a subtype‐specific marker.

FIGURE 7.

Identification of immune‐like malignant cell‐associated genes and prognostic candidate genes. (A) Kaplan–Meier survival analysis of four malignant epithelial cell subpopulation scores in the TCGA–LUAD cohort. (B) Comparison of CNV scores among four malignant epithelial cell subpopulations. (C) Volcano plot of differentially expressed genes between tumor and normal tissues in the TCGA–LUAD cohort. (D) Intersection analysis between marker genes of the immune‐like tumor subpopulation and differentially expressed genes in TCGA–LUAD. (E) Functional enrichment analysis of the intersected genes. (F) Error curve of the random forest model. (G) Top 60 important genes identified by the random forest model. (H) Univariate Cox regression forest plot of prognostically significant genes.

3.8. Establishment of an Immune‐Like Malignant Epithelial Cell‐Derived Prognostic Model

A panel of prognostic genes identified from immune‐like malignant epithelial cells was used to construct predictive models through the integration of multiple machine learning algorithms. Model performance was systematically compared using the concordance index. Among the evaluated approaches, the combination of LASSO and SuperPC achieved the highest predictive performance and was selected to establish IMEC‐Sig (Figure 8A). The final model comprised 10 genes: NUPR1, TM6SF1, DNAJB4, RHOV, NGEF, KRT8, CENPM, PTTG1, GAPDH, and LDHA. In the TCGA‐LUAD training cohort, patients were classified into high‐ and low‐risk groups according to the median IMEC‐Sig score. For each external validation cohort, including GSE13213, GSE26939, GSE29016, GSE31210, and GSE50081, the cohort‐specific median score was used for risk stratification. Survival analysis showed clear separation between the two risk groups in the training cohort (Figure 8B). Consistent survival patterns were also observed across the external validation cohorts (Figure 8C–G). Time‐dependent ROC analyses further supported the prognostic performance of IMEC‐Sig at 1, 3, and 5 years.

FIGURE 8.

Overview of model establishment and validation for IMEC‐Sig. (A) Evaluation of different machine learning approaches using the concordance index within the training dataset. (B–G) Survival analysis and ROC curve assessment of IMEC‐Sig across several independent cohorts, including TCGA–LUAD and multiple GEO datasets (GSE13213, GSE26939, GSE29016, GSE31210, and GSE50081).

3.9. Clinical Relevance and Predictive Performance of IMEC‐Sig

Associations between IMEC‐Sig and clinical features were systematically examined to evaluate its clinical relevance. Univariate Cox regression identified both IMEC‐Sig scores and tumor stage as significant predictors of overall survival (OS), whereas no significant associations were observed for age or gender (Figure S2A). In multivariate analysis, IMEC‐Sig retained its independent prognostic value after adjustment for clinical variables (Figure S2B). Tumor stage also remained significant, while age and gender did not contribute meaningfully to outcome prediction. A predictive nomogram combining IMEC‐Sig with clinical stage was subsequently developed to estimate survival probabilities at 1, 3, and 5 years (Figure S2C), enabling individualized prognostic assessment. The clinical utility of the model was further supported by decision curve analysis, which demonstrated consistent net benefit across a range of threshold probabilities (Figure S2D). Calibration plots additionally confirmed good agreement between predicted and actual survival outcomes over time (Figure S2E), indicating reliable model performance.

3.10. Immune Landscape and Immunotherapy‐Associated Features of IMEC‐Sig

Differences in the tumor immune microenvironment were assessed between IMEC‐Sig‐defined risk groups. The low‐risk group exhibited a higher degree of immune cell infiltration (Figure 9A), indicating a more immunologically active tumor microenvironment, whereas reduced infiltration was observed in the high‐risk group. Analysis of immune checkpoint‐related genes revealed distinct expression patterns between the two groups (Figure 9B). Correlation analysis further demonstrated significant associations between IMEC‐Sig scores and the expression levels of multiple immune checkpoint genes (Figure 9C), suggesting that the model captures variations in immune regulatory status. IPS, an integrated score reflecting tumor immunogenicity, was also evaluated as an indirect immunotherapy‐associated indicator. IPS values differed between the risk groups, with higher scores observed in the low‐risk group (Figure 9D). Together, these findings characterize distinct immune states between the IMEC‐Sig groups but do not directly demonstrate differences in clinical response to immunotherapy.

FIGURE 9.

Characterization of the immune microenvironment and immunotherapy‐associated features in relation to IMEC‐Sig. (A) Immune cell composition across groups defined by IMEC‐Sig levels. (B) Expression patterns of immune checkpoint genes in different IMEC‐Sig categories. (C) Associations between IMEC‐Sig scores and checkpoint‐related gene expression. (D) Comparison of IPS among IMEC‐Sig‐based subgroups.

3.11. Pan‐Cancer Analysis and Functional Validation of KRT8 in Tumor Biology

KRT8 was selected for functional validation based on convergent evidence across the analytical workflow rather than on subtype‐specific expression alone. It originated from the candidate gene set generated by intersecting immune‐like tumor markers with differentially expressed genes in TCGA–LUAD, remained after random forest and univariate Cox screening, and was retained as one of the 10 genes in the final IMEC‐Sig model. Patients in the TCGA–LUAD cohort were ranked according to their IMEC‐Sig scores and divided into low‐ and high‐risk groups using the median score as the cutoff (Figure S3A). Death events and shorter survival durations were more frequently observed toward the high‐risk end of the score distribution (Figure S3B). The expression heatmap of the 10 IMEC‐Sig genes showed that KRT8 was relatively upregulated in the high‐risk group (Figure S3C). In addition, higher KRT8 expression was associated with unfavorable survival outcomes in LUAD. Together, these findings provided the rationale for selecting KRT8 for subsequent pan‐cancer analysis and functional validation.

Pan‐cancer expression analysis confirmed that KRT8 exhibited aberrant expression across multiple tumor types. In the TCGA dataset, KRT8 was broadly upregulated across multiple tumor types, exhibiting higher expression levels in tumor tissues compared with normal counterparts (Figure 10A). In contrast, relatively lower or opposite trends were observed in certain cancer types, including GBM, PCPG, and SARC. When integrating TCGA with GTEx data, the differential expression patterns of KRT8 became more pronounced, particularly in cancer types with limited normal samples, such as ACC, DLBC, GBM, LGG, OV, TGCT, and UCS, where clearer tumor–normal differences were observed (Figure 10B). In pan‐cancer survival analysis, KRT8 exhibited heterogeneous prognostic associations across different endpoints (Figure 10C). For OS, KRT8 functioned predominantly as a risk factor in GBM, HNSC, LUAD, and PAAD, whereas it showed a protective effect in KICH. In disease‐free survival (DFS) analysis, KRT8 was associated with unfavorable outcomes in CESC, COAD, LGG, PAAD, and PCPG, while demonstrating a protective role in UCEC. For disease‐specific survival (DSS), KRT8 was linked to poorer prognosis in GBM, LUAD, and PAAD, but showed opposite associations in BLCA and KICH. Similarly, in progression‐free survival (PFS), KRT8 acted as a risk factor in GBM, LGG, LUAD, PAAD, and PCPG. Overall, KRT8 was consistently associated with adverse outcomes in several tumor types, particularly in gliomas, LUAD, and pancreatic cancer.

FIGURE 10.

Pan‐cancer expression patterns and functional validation of KRT8. (A) Expression levels of KRT8 across different cancer types in the TCGA dataset. (B) Expression levels of KRT8 across different cancer types based on combined TCGA and GTEx datasets. (C) Pan‐cancer prognostic analysis of KRT8, shown as forest plots for OS, DFS, DSS, and PFS. (D) Validation of KRT8 knockdown efficiency in LUAD cell lines. (E) Effects of KRT8 knockdown on cell proliferation assessed by CCK‐8 assay.

Subsequent in vitro experiments were conducted to investigate the functional role of KRT8. Knockdown efficiency was confirmed in A549 and H1299 cells using two independent siRNAs (Figure 10D). CCK‐8 assays indicated that KRT8 silencing reduced cell proliferation (Figure 10E). Further functional assays demonstrated that KRT8 depletion impaired malignant phenotypes. Transwell assays showed reduced migratory and invasive capacities (Figure 11A). EdU incorporation assays indicated decreased DNA synthesis activity (Figure 11B). Consistently, colony formation assays revealed reduced clonogenic potential in both cell lines (Figure 11C).

FIGURE 11.

Effects of KRT8 knockdown on migration, invasion, proliferation, and colony formation. (A) Transwell assays showing migration and invasion of A549 and H1299 cells after KRT8 knockdown, with corresponding quantification. (B) EdU incorporation assays evaluating proliferative activity in A549 and H1299 cells following KRT8 knockdown, with quantification of EdU‐positive cells. (C) Colony formation assays showing clonogenic capacity of A549 and H1299 cells after KRT8 knockdown, with corresponding quantification.

Collectively, these findings indicate that KRT8 is upregulated in the high‐risk group defined by IMEC‐Sig and exhibits widespread dysregulation and prognostic relevance across multiple cancer types. Functional experiments further demonstrate that KRT8 knockdown suppresses proliferation, migration, invasion, and colony formation in LUAD cells, suggesting its potential role in maintaining malignant cellular behaviors.

4. Discussion

In this study, we integrated single‐cell transcriptomic data with bulk transcriptomic cohorts to comprehensively characterize LUAD from the perspectives of cellular heterogeneity, immune microenvironment, and clinical outcomes. Malignant epithelial cells were identified based on inferred copy number variations and further stratified into distinct subpopulations with diverse transcriptional profiles, including a subset with immune‐related features. By combining trajectory inference, transcriptional regulation analysis, and cell–cell communication profiling, we delineated their potential roles in tumor evolution and immune modulation. On this basis, we established a prognostic signature (IMEC‐Sig) derived from immune‐like malignant epithelial cells, linking single‐cell insights with clinical applications.

Functional heterogeneity within malignant epithelial cells represents a critical determinant of tumor progression and therapeutic response [40, 41]. While tumor epithelial cells have traditionally been considered a relatively homogeneous population, recent single‐cell studies have revealed substantial diversity in their transcriptional programs and functional states [42, 43]. In the present study, the identified immune‐like malignant epithelial cells exhibited both immune‐related transcriptional features and elevated genomic instability, suggesting that they may represent a more active tumor cell state [14]. These findings support the concept that tumor cells are not merely passive targets of immune regulation but may actively participate in shaping the immune microenvironment.

Trajectory analysis placed AT2 and malignant epithelial cells along a connected transcriptional manifold, with AT2 cells predominantly located at lower pseudotime values and malignant subpopulations distributed across later or branch‐specific regions. This pattern is consistent with potential transcriptional continuity between AT2‐like and malignant epithelial states but does not establish a direct lineage relationship. Although previous studies have proposed AT2 cells as a candidate cell of origin for LUAD [44], the present analysis was based on cross‐sectional transcriptomic data and therefore cannot confirm that malignant cells arose from AT2 cells. The observed branch structure instead reflects heterogeneity among malignant epithelial states and suggests possible state transitions or divergent transcriptional programs. Confirmation of their developmental relationships will require lineage‐tracing experiments, longitudinal sampling, or appropriate experimental models.

At the level of the tumor microenvironment, our analysis revealed extensive communication between myeloid cells and malignant epithelial cells. In particular, adhesion‐related signals such as FN1 and SPP1, as well as chemokine pathways, were broadly involved in intercellular interactions. These signaling axes are likely to contribute not only to extracellular matrix remodeling but also to immune cell recruitment and spatial organization within the tumor niche [45, 46, 47, 48]. Notably, the observed interaction patterns suggest a bidirectional regulatory relationship, in which tumor cells both respond to and actively modulate the immune microenvironment. Similarly, COL1A2⁺MMP1^+/− cancer‐associated fibroblast subsets in ER‐positive breast cancer were reported to interact extensively with tumor‐associated macrophages and contribute to an immunosuppressive microenvironment [49].

Building upon these findings, the IMEC‐Sig model demonstrated stable prognostic performance across multiple independent cohorts and separated patients with distinct immune states. The low‐score group exhibited greater immune infiltration and higher immune activity, whereas the high‐score group showed a relatively immunosuppressive profile. These observations suggest that IMEC‐Sig captures biologically relevant variation in the tumor immune landscape. However, immune infiltration, immune checkpoint expression, and IPS are surrogate transcriptomic indicators rather than direct measures of treatment response. Therefore, the relationship between IMEC‐Sig and immunotherapy benefit should be considered exploratory and requires validation in cohorts of patients receiving immune checkpoint blockade.

Emerging evidence has highlighted the role of epithelial cell states in immune regulation; however, systematic characterization of immune‐associated malignant epithelial subpopulations remains limited. By leveraging single‐cell data, our study provides a refined view of these cell states and their potential contributions to tumor–immune interactions.

The selection of KRT8 for functional validation was based on its continuity across the analytical framework rather than on its exclusive expression in a single malignant epithelial subtype. KRT8 was derived from the immune‐like tumor‐associated candidate set, retained in the final IMEC‐Sig model, relatively elevated in the high‐risk group, and associated with unfavorable survival outcomes in LUAD. These convergent findings connected the single‐cell‐informed candidate screening with patient‐level risk stratification and provided the basis for subsequent experimental investigation. At the molecular level, KRT8, a key intermediate filament protein in epithelial cells, plays an essential role in maintaining cellular integrity and responding to stress [50, 51]. Previous studies have reported aberrant KRT8 expression in various cancers and linked it to proliferative and invasive phenotypes [52, 53]. Here, we observed that KRT8 was upregulated in high‐risk malignant epithelial subpopulations and exhibited consistent associations with survival outcomes across multiple cancer types. Functional assays further demonstrated that KRT8 silencing suppressed proliferation, migration, invasion, and clonogenic potential in LUAD cells, indicating its involvement in maintaining malignant cellular behaviors.

Several limitations should be acknowledged. First, the single‐cell datasets were obtained from public resources and included a relatively limited number of samples. Because these datasets were cross‐sectional, the pseudotime trajectories reflect inferred transcriptional relationships rather than definitive lineage transitions. Similarly, the cell–cell communication analysis was based on ligand and receptor transcript expression and does not directly demonstrate physical interactions, downstream signaling activation, or causal effects between cell populations. Second, functional validation was primarily conducted in vitro and focused on the tumor cell‐intrinsic effects of KRT8 on proliferation, migration, invasion, and clonogenic capacity. These experiments do not directly establish that KRT8 regulates immune cells or influences immunotherapy responsiveness, and further in vivo and tumor–immune co‐culture studies are required. Third, the immunotherapy‐related analyses were based on indirect indicators, including immune cell infiltration, immune checkpoint expression, and IPS. No cohort of patients treated with immune checkpoint inhibitors was available for direct validation. Therefore, IMEC‐Sig should not currently be considered a clinically validated predictor of immunotherapy response. Future studies incorporating larger multicenter datasets, longitudinal or spatial analyses, immunotherapy‐treated cohorts, and additional experimental validation are warranted.

In summary, this study provides a single‐cell‐informed framework that connects tumor heterogeneity, immune microenvironment, and clinical outcomes in LUAD. The IMEC‐Sig model offers potential utility for risk stratification, while KRT8 emerges as a candidate regulator of tumor progression, together contributing to a better understanding of LUAD biology and precision oncology.

Author Contributions

Conceptualization and study design were carried out by Haiying Peng. Wei Zhang, Hongmei Sheng, and Yazheng Zhang conducted the bioinformatics analyses, performed the experimental work, and prepared the initial draft of the manuscript. Qihong Yu participated in data acquisition, data interpretation, and manuscript revision. Haiying Peng supervised the study and critically revised the manuscript. All authors reviewed, edited, and approved the final version of the manuscript.

Funding

This work was supported by the Tianjin Key Medical Discipline Project (TJYXZDXK‐3‐032C) and the Scientific Research Program of Hebei Provincial Administration of Traditional Chinese Medicine (T2026027).

Ethics Statement

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Validation of cell type annotation.

Figure S2: Comprehensive assessment of IMEC‐Sig in survival prediction and clinical application.

Figure S3: Distribution of IMEC‐Sig scores, survival outcomes, and model gene expression in the TCGA–LUAD cohort.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.