Single-cell RNA sequencing reveals palmitoylation-driven cellular heterogeneity and prognostic biomarkers in lung adenocarcinoma

Highlights

• •Single-cell RNA sequencing identifies six discrete tumor cell subsets in lung adenocarcinoma, revealing intratumoral heterogeneity linked to palmitoylation activity.
• •The C1 subset, marked by elevated protein palmitoylation, exhibits unique copy-number alterations and extensive crosstalk with stromal and immune compartments.
• •Integrated ligand–receptor, pseudotime, and transcription factor network analyses map palmitoylation-driven signaling circuits across the tumor microenvironment.
• •A 12-gene prognostic signature derived from C1 markers robustly stratifies patient survival in TCGA and independent GEO cohorts.
• •Functional assays validate ASPH as a driver of proliferation, apoptosis resistance, epithelial–mesenchymal transition, and invasiveness in lung adenocarcinoma cells.

Abstract

Background

Lung adenocarcinoma (LUAD) is marked by significant variation within tumor cells and continues to be a major global cause of cancer deaths. Palmitoylation is a dynamic lipid-based modification that occurs after protein synthesis and influences the behavior and lifespan of various cancer-related proteins. However, its role in shaping cellular complexity and predicting outcomes in LUAD patients is not yet fully clarified.

Methods

We examined single-cell RNA sequencing datasets from LUAD samples to identify distinct malignant cell groups. Palmitoylation-related gene activity was estimated using GSVA and ssGSEA techniques. To further define cellular characteristics, we applied copy number variation mapping, pseudotime progression modeling, transcription factor regulatory scoring, and cell–cell interaction analyses. A 12-gene risk model was developed using marker genes from the cluster (C1) with the most prominent palmitoylation pattern. This model was trained on The Cancer Genome Atlas (TCGA) dataset and confirmed using separate GEO datasets. To evaluate tumor immune context, we analyzed immune cell presence and tumor mutational burden across different risk levels. Laboratory experiments involving both upregulation and silencing of aspartate beta-hydroxylase (ASPH) in LUAD cell cultures were conducted to validate its biological significance.

Results

We identified six tumor cell subsets (C0 to C5), with cluster C1 showing peak palmitoylation levels, distinct genomic alterations, and stronger communication with stromal and immune cells. The 12-gene model effectively categorized LUAD patients into high- and low-risk profiles, showing marked survival differences (p < 0.001) and strong performance in time-dependent ROC analysis. Patients in the high-risk group had increased tumor mutational burden and a more immunosuppressive tumor environment. Laboratory findings revealed that raising ASPH expression promoted cell growth, motility, and epithelial–mesenchymal transition. In contrast, reducing ASPH levels triggered cell death and decreased invasiveness.

Conclusions

Our single-cell analysis focused on palmitoylation reveals new dimensions of tumor diversity in LUAD and establishes a validated 12-gene risk signature. Functional studies highlight ASPH as a promising candidate for therapeutic targeting. These results deepen our understanding of palmitoylation-associated pathways and present a foundation for both outcome prediction and precision-based treatment strategies in LUAD.

Graphical abstract

Introduction

Lung adenocarcinoma (LUAD) constitutes the most prevalent form of non-small cell lung cancer (NSCLC), responsible for nearly 40 % of all lung cancer diagnoses [1,2]. This subtype poses a considerable clinical burden due to its frequent occurrence, rapid progression, and early metastatic potential [3,4]. Although modern treatment modalities—such as surgical resection, chemotherapy, radiotherapy, and molecularly targeted agents—have improved outcomes to some extent, long-term survival rates for LUAD patients remain discouragingly poor [5,6]. This scenario underscores the pressing need for deeper molecular insights that could drive the advancement of more effective diagnostic and therapeutic options.

The development of LUAD is governed by a multifaceted interplay of genetic and epigenetic disturbances. Several oncogenic drivers, including alterations in EGFR, KRAS, and ALK genes, have been identified and serve as therapeutic targets [[7], [8], [9]]. However, intratumoral heterogeneity—both molecular and cellular—remains a major obstacle in disease management [10]. LUAD tumors comprise various cell types, such as malignant epithelial cells, immune components, and stromal elements, each contributing differently to tumor behavior [11,12]. Conventional bulk transcriptomic analyses often obscure this complexity, pointing to the need for high-resolution techniques like single-cell profiling to dissect tumor architecture [13].

Single-cell RNA sequencing (scRNA-seq) has emerged as a robust technique to study cellular heterogeneity in tumors at single-cell resolution [14,15]. This method allows detailed mapping of transcriptional activity across individual cells, enabling the identification of distinct cell types, differentiation hierarchies, and communication pathways [16,17]. The application of scRNA-seq in cancer research has unveiled previously unknown subpopulations and clarified their functional roles in disease progression [18]. In LUAD, this technology offers a promising approach to examine tumor evolution, microenvironmental dynamics, and resistance mechanisms at an unprecedented scale.

The tumor microenvironment (TME) is pivotal in influencing cancer behavior and therapeutic response [19]. It comprises a diverse and interactive milieu of tumor cells, immune infiltrates, fibroblasts, endothelial structures, and extracellular matrix components [20]. These cellular interactions create a supportive environment for tumor growth, immune evasion, and metastatic spread [21,22]. A comprehensive understanding of TME composition and function is essential for optimizing immunotherapeutic strategies and enhancing treatment outcomes.

The heterogeneity of LUAD tumor cells necessitates detailed classification based on molecular and phenotypic traits [23]. Distinguishing specific tumor subtypes and their lineage trajectories can reveal mechanisms of disease initiation, progression, and treatment resistance [24,25]. Such subclassification is crucial for tailoring targeted therapies and advancing personalized medicine initiatives. Furthermore, understanding cellular differentiation states may provide insights into tumor adaptability and plasticity.

Identifying reliable prognostic indicators is vital for stratifying patients and refining therapeutic interventions. In LUAD, correlations between gene expression patterns, immune cell infiltration, and tumor mutation burden (TMB) can guide clinical decision-making. Although immune checkpoint inhibitors have shown efficacy in LUAD, variability in patient responses limits their universal success [26,27]. Investigating the connections among immune landscape, prognostic gene signatures, and mutational load may aid in predicting treatment responses and developing more effective combinatorial approaches.

Despite significant progress in LUAD research, several key questions remain unresolved. The dynamic interactions within the TME, the contribution of copy number variations (CNVs) to tumor evolution, and the molecular basis of tumor cell differentiation require further exploration [28,29]. The present study seeks to address these challenges through comprehensive single-cell transcriptomic analysis.

This research aims to characterize the cellular diversity of LUAD by profiling individual cells to identify unique populations and their roles in carcinogenesis. It will evaluate CNVs at the single-cell level to understand their influence on tumor variability and potential as therapeutic targets. The study also focuses on classifying tumor subtypes based on their differentiation patterns and lineage relationships. Furthermore, it investigates intercellular communication networks to elucidate signaling mechanisms driving tumor progression and immune escape. Special attention will be paid to palmitoylation processes and their functional implications across tumor subgroups. Transcription factor (TF) networks will be mapped to identify key regulatory elements in tumor biology. Finally, the study will assess gene signatures for their prognostic relevance and explore their association with immune infiltration and TMB.

By integrating single-cell analysis with robust computational pipelines, this work aims to advance our understanding of LUAD heterogeneity and uncover novel molecular targets for improved therapeutic and prognostic strategies.

Method

Cell culture and quantitative real-time PCR

Normal human bronchial epithelial cells (BEAS-2B) were obtained from ATCC and grown in DMEM. Two lung cancer cell lines, A549 (from ATCC) and PC-9 (from Cobioer, China), were cultured in RPMI-1640. All media were supplemented with 10 % fetal bovine serum, 100 U/mL penicillin, and 100 µg/mL streptomycin. Cells were kept at 37 °C in a humid 5 % CO₂ incubator. For RNA extraction, TRIzol™ reagent was used as per the instructions. cDNA was synthesized using ReverTra Ace® qPCR RT Master Mix with gDNA Remover. Gene expression was measured using qRT-PCR with SYBR® Premix Ex Taq™ II on an Mx3005P Real-Time PCR System. GAPDH was used as the reference gene. PCR involved initial heating at 95 °C for 10 min, then 45 cycles of 95 °C for 5 s and 60 °C for 30 s. Each gene was tested in separate wells and repeated three times. Gene expression levels were calculated using the 2^(-ΔΔCt) method. Primer details are in Supplementary Table 1.

Collection and preparation of transcriptome data

To train our prediction model, we collected RNA sequencing data and clinical information for 513 lung adenocarcinoma (LUAD) patients from The Cancer Genome Atlas (TCGA). A separate dataset was used to test the model’s performance. Before analysis, all gene expression values were converted into TPM format and log2-transformed to ensure consistent data distribution. To ensure robustness of the prognostic model, we selected TCGA-LUAD as the training cohort due to its large sample size, comprehensive clinical annotations, and uniform RNA-sequencing protocols. For external validation, GEO datasets were chosen based on availability of survival data, similar sequencing platforms, and minimal batch effects after normalization. These datasets allowed for independent assessment of model generalizability. In feature selection, we applied univariate Cox regression followed by LASSO regression to avoid overfitting and improve model interpretability.

Collection and processing of single-cell RNA sequencing data

We obtained single-cell RNA sequencing (scRNA-seq) data from GEO (ID: GSE189357), which included nine primary tumor samples. All data analysis was done using the Seurat package in R (version 4.1.3). To ensure quality, cells were filtered based on three criteria: Mitochondrial genes < 20 %, UMIs between 200 and 15,000, Detected genes per cell between 200 and 5000. After filtering, data were normalized, and 2000 highly variable genes were selected using Seurat’s standard functions. We removed the effect of cell cycle variation by adjusting for S.Score and G2M.Score during scaling. Batch effects were corrected using Harmony. We then reduced the data’s dimensions using UMAP and t-SNE, and grouped similar cells using the Louvain clustering method. To identify key genes in each group, we used the FindAllMarkers function, applying filters of p-value < 0.05, log2 fold change > 0.25, and at least 10 % expression in the group.

Identification of palmitoylation-associated genes in cancer cells

To explore genes involved in protein palmitoylation within tumor environments, we retrieved gene sets from the Molecular Signatures Database (MsigDB). Specifically, we extracted the gene list under the Gene Ontology Biological Process term: "GOBP_PEPTIDYL_L_CYSTEINE_S_PALMITOYLATION". This curated dataset includes genes known to participate in the enzymatic addition of palmitoyl groups to cysteine residues, which may play functional roles in oncogenesis and tumor progression.

Cell type annotation using canonical marker genes

Cellular annotation was performed by referencing established marker genes specific to major cell types commonly found in the tumor microenvironment. The following markers were used: Epithelial cells: EPCAM, KRT18, KRT19, CDH1, Fibroblasts: DCN, THY1, COL1A1, COL1A2, Endothelial cells: PECAM1, CLDN5, FLT1, RAMP2, T lymphocytes: CD3D, CD3E, CD3G, TRAC, Natural Killer (NK) cells: NKG7, GNLY, NCAM1, KLRD1, B lymphocytes: CD79A, IGHM, IGHG3, IGHA2, Myeloid lineage cells: LYZ, MARCO, CD68, FCGR3A, Mast cells: KIT, MS4A2, GATA2, Neutrophils: FCGR3B, CSF3R, and after assigning cell identities based on these gene expression profiles, we visualized the distribution of cell types using t-distributed stochastic neighbor embedding (t-SNE) plots. Additionally, violin plots were generated to display expression levels of selected marker genes across annotated cell clusters, enhancing the interpretability of cell-type-specific expression patterns.

Classification of tumor cell subgroups

To explore the diversity within tumor cells, we first extracted the clusters identified as tumor cell populations. These cells were then subjected to a fresh round of dimensionality reduction and clustering. For each newly formed subcluster, characteristic marker genes were determined. These markers served as the basis for classifying and labeling the subgroups according to known gene expression profiles.

Copy number variation (CNV) profiling in single cells

We analyzed copy number variations (CNVs) across different single-cell subtypes using the InferCNV tool. Stromal cells were selected as the baseline reference group for comparison. This analysis was crucial in pinpointing malignant cell populations and calculating CNV scores for each distinct cellular subgroup.

Trajectory mapping of tumor cell differentiation

To understand how tumor cells evolve over time, we carried out pseudotime trajectory analysis using the monocle2 platform. Dimensionality reduction was executed with the DDRTree algorithm, while all other settings were kept at their default values. This analysis helped reveal potential paths of differentiation among tumor subpopulations.

Identification of transcription factor activity in tumor cells

Transcriptional regulatory activity in tumor cell subtypes was examined using the SCENIC framework. The analysis incorporated both the RcisTarget and GRNBoost motif databases under default conditions. RcisTarget was employed to identify enriched transcription factor binding motifs in specific gene sets, while AUCell was used to quantify the activity levels of individual regulons across various cell types.

Intercellular communication analysis

To study how different cell types interact, the CellChat tool in R was utilized. After importing the normalized gene expression matrix into a CellChat object, key functions including identifyOverExpressedGenes, identifyOverExpressedInteraction, and ProjectData were used for preprocessing, all with default settings. Likely ligand-receptor relationships were then examined using computeCommunProb, filterCommunication, and computeCommunProbPathway. The complete communication network was built using the aggregateNet function.

Palmitoylation signature scoring

A curated set of genes associated with palmitoylation was used to compute enrichment scores at the single-cell level. This was implemented using both the GSVA and ssGSEA algorithms within the GSVA package in R.

Immune cell infiltration profiling

To assess immune cell distribution among different risk-based patient groups, the IOBR package was applied. Estimations were generated using a combination of ESTIMATE, CIBERSORT, and xCell algorithms for broader immune context. The ESTIMATE, CIBERSORT, and xCell algorithms were selected to provide complementary perspectives on immune infiltration. ESTIMATE infers overall stromal and immune cell proportions, CIBERSORT quantifies 22 immune cell subsets with high resolution, and xCell enhances specificity by incorporating gene signatures. This integrative approach minimizes biases from any single tool and ensures consistency across immune profiling methods.

Functional gene enrichment

We used the clusterProfiler package in R to explore biological pathways through Gene Ontology (GO) and KEGG analysis. Only pathways with an adjusted p-value (Benjamini–Hochberg method) below 0.05 were considered significant. The results were visualized using the ggplot2 package.

Genomic mutation comparison

Genomic mutation characteristics were analyzed using the maftools package. This allowed for comparing mutation frequencies between groups and evaluating the relationship between Tumor Mutation Burden (TMB) and risk score. The influence of TMB on survival was also assessed.

Prognostic model development based on C1 tumour cell markers

Genes highly expressed in C1 tumor cells, known for elevated palmitoylation, were screened for survival relevance using univariate Cox regression (p < 0.01). These candidate genes were input into a LASSO—Cox model using the glmnet package. Model accuracy was validated through the timeROC package, which provided AUC values for 1-, 3-, and 5-year survival predictions. To ensure model robustness and reduce overfitting, we implemented a 10-fold cross-validation procedure during the LASSO—Cox regression process using the cv.glmnet function in the glmnet R package. In this procedure, the TCGA-LUAD dataset was randomly partitioned into 10 equal-sized subsets. In each iteration, nine subsets were used for model training and the remaining one for validation. This process was repeated 10 times, with the mean cross-validated partial likelihood deviance used to determine the optimal regularization parameter (lambda). The final model corresponding to the lambda value with the minimum mean cross-validated error was selected. Risk scores were then computed based on this optimized model and subsequently validated using independent GEO cohorts to assess generalizability.

Statistical analyses

All statistical work was completed in R version 4.1.3. Pearson's correlation measured associations between continuous variables. Categorical variables were tested with the Chi-square test, while continuous group differences were evaluated via Wilcoxon rank-sum or T-tests. The survminer package was used to define optimal cutoffs. Survival trends and hazard ratios were analyzed using Kaplan-Meier plots and Cox regression models via the survival package.

Results

Single-cell expression profiling of LUAD

After standard quality control and dimensionality reduction, 113,097 single cells were retained for analysis. Cells were annotated using known lineage markers, classifying them into nine major populations: epithelial cells, fibroblasts, endothelial cells, T cells, NK cells, B cells, myeloid cells, neutrophils, and mast cells. Their spatial distribution is shown by UMAP in Fig. 1A.

Fig. 1.

Single-cell classification results of LUAD. (A) tSNE plot showing the cell type classification of single-cell data. (B) Bar plot showing the composition of cell types in each patient. (C) Bubble plot displaying the expression of marker genes for each cell type.

To examine inter-patient variability, the relative abundance of each cell type per patient was calculated and is visualized in Fig. 1B, revealing differences in cellular composition across the LUAD cohort. Additionally, to validate the integrity of cell classification, we examined the expression levels of canonical marker genes across the identified cell types. The results, presented in Fig. 1C, confirm the expected marker enrichment in corresponding clusters, supporting the reliability of the cell identity assignments.

CNV analysis of single cells

To investigate genomic alterations at the single-cell level, we employed the inferCNV algorithm, using stromal cells as a baseline reference population. The inferred CNV landscape across all cells is depicted in Fig. 2A. As anticipated, epithelial cells displayed prominent CNV deviations, suggesting widespread genomic instability in this compartment. In contrast, non-epithelial lineages, including immune and stromal cells, exhibited minimal CNV signals, consistent with their non-malignant origin.

Fig. 2.

CNV analysis results of single cells. (A) Denoised CNV heatmap using stromal cells as reference.

Subclassification analysis of tumour cells

To further investigate tumour cell heterogeneity, we performed an independent clustering analysis exclusively on tumour-derived epithelial cells. This analysis revealed six distinct subpopulations, designated based on their dominant marker gene expression profiles: C0 (Epi S100A9+), C1 (Epi SFTPC+), C2 (Epi CAPS+), C3 (Epi AGER+), C4 (Epi SPINK1+), and C5 (Epi SCGB1A1+). The distribution of these six clusters was quantified across individual patients, histological subtypes, and cell cycle phases. Cluster-wise proportions are illustrated in Figs. 3A–F, highlighting inter-patient variability and subtype-specific dominance patterns. Each cluster's relative enrichment (Ro/e ratio) in relation to histological classification and cell cycle status was also evaluated and visualized (Figs. 3G–I). Further analysis focused on assessing quantitative differences across clusters by comparing four key parameters: inferred copy number variation (CNV) scores, unique molecular identifier (UMI) counts, G2M phase scores, and S phase scores. The results, presented in Figs. 3J–M, demonstrate distinct trends in genomic instability and proliferative activity among the clusters. These measures were also stratified by histological category and cell cycle phase to identify context-specific patterns. No statistical interpretation is provided at this stage; subsequent sections address their biological implications.

Fig. 3.

Subclassification analysis results of tumour cells. (A-C) tSNE plots showing the proportion of each cluster in patients, histological types, and cell cycles. (D-H) Bar plots showing the proportion of each cluster in patients, histological types, and cell cycles. (I) The ratio of observed to expected cell numbers (Ro/e) to evaluate tissue preference of each cell subtype. (J-M) tSNE plots and violin plots showing CNV scores, nCount_RNA, G2M score, and S score across different clusters and histological types.

Cell communication analysis

Cellular interactions within the tumour microenvironment were examined to map signaling exchanges between tumour epithelial subtypes and surrounding cell types. Outgoing signaling from tumour subpopulations to non-tumour cells, including immune and stromal populations, is visualized in Fig. 4A. In contrast, Fig. 4B depicts the reverse communication—signaling received by tumour subtypes from these non-malignant cells. To further illustrate the specificity of these interactions, Figs. 4C and 4D display bubble plots summarizing key ligand–receptor pairs involved in the bidirectional signaling. These plots emphasize the most prominent molecular interactions underlying the intercellular crosstalk. This analysis outlines the structural framework of tumour-related communication but does not interpret the functional consequences, which are addressed in subsequent sections.

Fig. 4.

Cell communication analysis results. (A) Chord diagram showing ligand-receptor pairs between tumour cell subtypes and other cells (such as immune and stromal cells). (B) Chord diagram showing ligand-receptor pairs from other cells (such as immune and stromal cells) to tumour cell subtypes. (C) Bubble plot showing ligand-receptor pairs between tumour cell subtypes and other cells. (D) Bubble plot showing ligand-receptor pairs from other cells to tumour cell subtypes.

CytoTRACE and cell trajectory analysis

The CytoTRACE algorithm was used to assess the differentiation status of tumour epithelial subtypes. The results indicated that clusters C4 and C1 exhibited the highest levels of differentiation potential, whereas clusters C3 and C5 displayed the lowest (Figs. 5A–B). Following this, we conducted pseudotime trajectory analysis using the Monocle framework to reconstruct the dynamic progression of tumour cell states. The inferred trajectory positioned clusters C4, C1, and C0 at the early stages of pseudotime, suggesting a more progenitor-like or differentiated state. In contrast, clusters C2, C3, and C5 were located at the terminal ends of the trajectory, reflecting a relatively undifferentiated or mature phenotype (Figs. 5C–G).

Fig. 5.

CytoTRACE and cell trajectory analysis results. (A, B) CytoTRACE analysis showing stemness of tumour cell subtypes and their differentiation aspects. (C-G) Monocle results showing that C2, C3, and C5 tumour cells have the lowest differentiation compared to other subtypes. (H) Heatmap showing the expression of tumour cell subtype markers (S100A9, SFTPC, CAPS, AGER, SPINK1, SCGB1A1) along the pseudotime trajectory.

To support these findings, we also visualized the expression dynamics of representative marker genes across clusters and along the pseudotime continuum. These gene expression trends are shown in Fig. 5H, providing molecular insight into the differentiation gradients observed. This analysis presents a temporal framework for understanding tumour cell plasticity, without inferring functional implications at this stage.

Functional scoring of palmitoylation in tumor cell subtypes

Palmitoylation scores in tumor cells were calculated using GSVA and ssGSEA, with both methods showing consistent results. Subtype C1 had the highest score, significantly higher than other subtypes (Figs. 6A–D). We then identified differentially expressed genes between C1 and other subtypes and performed KEGG and GO enrichment analyses. These genes were mainly involved in ferroptosis, fatty acid metabolism, and related pathways (Figs. 6E–F). Differences in Hallmark gene sets were also analyzed (Figs. 6G–H).

Fig. 6.

Functional scoring of palmitoylation in tumour cell subtypes. (A, B) GSVA analysis showing differences in palmitoylation-related gene activity scores among tumour cell subtypes. (C, D) ssGSEA analysis showing differences in palmitoylation-related gene activity scores among tumour cell subtypes. (E) KEGG enrichment analysis between C1 tumour cells and other subtypes. (F) GO enrichment analysis between C1 tumour cells and other subtypes. (G) GSVA enrichment analysis between C1 tumour cells and other subtypes. (H) GSEA enrichment analysis between C1 tumour cells and other subtypes.

Transcription factor analysis

To explore regulatory heterogeneity among tumour epithelial subtypes, we analyzed transcription factor activity using a CSI (cell-specificity index) matrix. Based on similarity patterns, transcription factors were grouped into three distinct modules: M1, M2, and M3. Each module was defined by its unique set of representative transcription factors, their DNA-binding motifs, and the dominant cell types where they were active. These features are summarized in Fig. 7A. Additionally, regulatory specificity scores (RSS) were computed to quantify the enrichment of each transcription factor across tumour cell clusters. The resulting distribution of RSS values is visualized in Fig. 7B, highlighting subtype-specific regulatory signatures. These results provide a classification framework for transcriptional regulation within tumour cells, laying the groundwork for further functional investigation.

Fig. 7.

Transcription factor analysis results. (A) Regulation modules of tumour cell subtypes determined by CSI matrix, showing representative transcription factors, binding motifs, and associated subtypes. (B) Ranking of regulatory scores (RSS) for transcription factors in tumour cell subtypes, highlighted in t-SNE (red dots). Binary activity scores (RAS) of transcription factors in tumour cell subtypes shown in t-SNE plots.

Prognostic model construction using C1 tumour cell markers

To identify tumour cell-specific genes associated with clinical outcomes, we focused on marker genes derived from the C1 tumour cell cluster, which exhibited the highest palmitoylation activity. Univariate Cox regression analysis identified genes significantly associated with overall survival (P < 0.01; Fig. 8A). These genes were then used to build a prognostic model using LASSO—Cox regression (Fig. 8B). The resulting model included a subset of prognostic genes, and their corresponding regression coefficients are illustrated in Fig. 8C. We also visualized the expression profiles of these genes across patient samples (Fig. 8D). Survival analysis was conducted using Kaplan–Meier curves, along with time-dependent ROC analysis, demonstrated the model's capacity to stratify patients by risk (Figs. 8E–F). Additionally, individual survival curves were plotted for each gene included in the model to evaluate their standalone prognostic value (Fig. 8G). To validate the biological relevance of these candidate markers, we measured the expression levels of three representative genes-aspartate beta-hydroxylase (ASPH), CKS2, and IRX3— In normal bronchial epithelial cells (BEAS-2B) and lung adenocarcinoma cell lines (A549 and PC-9), qRT-PCR analysis showed that all three genes were significantly upregulated in cancer cells compared to normal controls (Fig. 9A). This suggests their potential roles in tumour progression and their utility as biomarkers or therapeutic targets in LUAD.

Fig. 8.

Prognostic model construction using C1 tumour cell markers. (A) Forest plot showing univariate Cox analysis results for marker genes of C1 tumour cells. (B) LASSO regression analysis identifying 12 genes associated with prognosis. (C) Dot plot showing coefficient (Coef) values of genes constituting the risk score. (D) Heatmap showing expression of these genes across different groups. (E) ROC curves for risk scores at 1, 3, and 5 years. (F) Kaplan-Meier survival curves for high-risk and low-risk groups. (G) Survival analysis results for the 12 prognostic genes.

Fig. 9.

Validation of prognostic gene expression in LUAD cell lines. (A) qRT-PCR analysis of aspartate beta-hydroxylase (ASPH), CKS2, and IRX3 expression in BEAS-2B (normal bronchial epithelial cells), A549, and PC-9 (LUAD cell lines). Data are presented as mean ± SD, P < 0.05.

Immune infiltration and TMB analysis

To assess the immune landscape linked to the prognostic model, we analysed the expression of model-related genes in lung adenocarcinoma and evaluated immune cell infiltration using ESTIMATE, CIBERSORT, and xCell. The combined results are presented as a heatmap in Fig. 10A, illustrating immune infiltration patterns across patient samples. Correlation analysis was then performed to explore relationships among risk scores, prognostic gene expression, and immune checkpoint molecules. Additionally, We examined the differential expression of immune checkpoint genes between high-risk and low-risk groups (Figs. 10B–C). The proportions of specific immune cell types, as predicted by the CIBERSORT algorithm, were displayed using both bar plots and box plots to compare distributions across the two risk groups (Figs. 10D–E). Pairwise correlations among immune cell types were also calculated and visualized (Figs. 10F–G), revealing potential co-infiltration trends. Finally, tumour mutation burden (TMB) was analysed to assess its relationship with the risk model. Differences in TMB between high- and low-risk groups were evaluated, and its correlation with risk scores was calculated (Figs. 10H–I). Kaplan–Meier analysis assessed the prognostic significance of TMB levels (Fig. 10J). These analyses provide a comprehensive overview of the immune and genomic features associated with the risk stratification model, without interpreting their clinical implications at this stage.

Fig. 10.

Immune infiltration and TMB analysis results. (A) Heatmap showing results of ESTIMATE, CIBERSORT, and Xcell analyses in LUAD. (B) Bubble plot showing correlation analysis between immune checkpoint genes, risk scores, and prognostic genes. (C) Differential expression of immune checkpoint genes between high-risk and low-risk groups. (D) Cibersort analysis of high-risk and low-risk groups. (E) Results of immune cell infiltration in LUAD. (F, G) Correlation analysis between immune infiltration cells and risk scores. (H-J) TMB results for low-risk and high-risk groups.

ASPH overexpression drives LUAD cell proliferation in vitro

To validate transcriptomic findings, we measured ASPH mRNA expression using quantitative RT-PCR. As shown in Figs. 11A–B, ASPH was significantly upregulated in primary LUAD tumour samples compared to matched adjacent normal tissue and showed elevated expression in multiple LUAD cell lines relative to the non-cancerous BEAS-2B bronchial epithelial cell line. To assess the functional relevance of ASPH, we performed siRNA-mediated knockdown in LUAD cell lines HCC827 and NCI-H2228. Efficient silencing (>80 % reduction in ASPH expression; Fig. 11C) led to a marked decrease in cellular proliferation over a 4-day period, as quantified by CCK-8 assays (Figs. 11D–E). These experimental results support a proliferative role for ASPH in LUAD cells under in vitro conditions. Although consistent with its proposed involvement in oncogenic pathways such as Notch signaling and EGF-like domain regulation, no mechanistic interpretations are drawn here. Further validation in vivo and analysis of downstream signaling cascades are required.

Fig. 11.

ASPH is overexpressed in lung adenocarcinoma tissues and cell lines, and its silencing impairs tumour‐cell proliferation. (A) qRT-PCR analysis of ASPH mRNA levels in paired adjacent normal (“Adjacent”) and tumour (“Cancer”) lung specimens. **P < 0.01 versus adjacent. (B) Basal ASPH expression measured by qRT-PCR in the normal bronchial epithelial cell line BEAS-2B and five LUAD cell lines (NCI-H1975, HCC827, A549, Calu-3, NCI-H2228). Data are mean ± SD; ns, not significant; **P < 0.01; ***P < 0.001; ****P < 0.0001 versus BEAS-2B (C) Confirmation of ASPH knockdown in HCC827 and NCI-H2228 cells transfected with si-ASPH or non-targeting control (si-NC). Relative ASPH mRNA levels were quantified by qRT-PCR. Data are mean ± SD; **P < 0.01; ***P < 0.001 versus si-NC. (D, E) CCK-8 assays showing growth curves of HCC827 (D) and NCI-H2228 (E) cells over 4 days post-transfection. OD₄₅₀ readings are plotted as mean ± SD from three independent experiments; **P < 0.01; ***P < 0.001 versus si-NC at day 4.

ASPH silencing promotes apoptosis and inhibits EMT, migration, and invasion in LUAD cells

To explore the broader effects of ASPH knockdown, we assessed apoptosis and epithelial–mesenchymal transition markers in LUAD cells. Flow cytometry showed that silencing ASPH in NCI-H2228 cells significantly increased early and late apoptosis compared to controls (Figs. 12A–B). Western blot analysis confirmed these results. ASPH knockdown increased cleaved caspase-3 and E-cadherin levels while reducing Bcl-2 and vimentin expression (Figs. 12C and 12E), indicating enhanced apoptosis and a shift towards an epithelial phenotype. Functional assessment using transwell assays demonstrated that ASPH silencing significantly impaired both cell migration and invasion capabilities, as indicated by a reduction in the number of cells traversing the membrane (Fig. 12D). Together, these findings show that ASPH contributes to LUAD cell survival, EMT maintenance, and motility. While suggestive of its multifaceted role in tumour progression, further mechanistic exploration is needed to confirm its therapeutic potential.

Fig. 12.

ASPH knockdown induces apoptosis and reverses EMT to impair migratory and invasive behavior in NCI-H2228 cells. (A) Representative flow-cytometry dot plots of annexin V–FITC/PI staining in si-NC and si-ASPH cells; Q1-LR and Q1-UR indicate early and late apoptotic populations. (B) Quantification of total apoptotic cells ( %); mean ± SD, ***P < 0.001 versus si-NC. (C) Western blot analysis of cleaved caspase-3, Bcl-2, E-cadherin, vimentin, and β-actin in si-NC and si-ASPH groups. (D) Transwell migration and invasion assays: representative crystal-violet images (left) and quantification of invaded cells per field (right); mean ± SD, *P < 0.05, **P < 0.01 versus si-NC. (E) Densitometric quantification of protein expression normalized to β-actin; mean ± SD, ***P < 0.001 versus si-NC.

Discussion

In this study, we used single-cell RNA sequencing to map the complex cellular and molecular landscape of lung adenocarcinoma (LUAD).Our results elucidate multiple facets of LUAD biology, offering insights that align with, refine, and extend previous findings while identifying potential avenues for targeted therapeutic interventions.

We identified nine principal cell types within LUAD tissues, including epithelial, fibroblast, endothelial, and diverse immune cell populations, thereby confirming earlier scRNA-seq studies by Zheng et al. and Lambrechts et al. [30,31], which underscored the heterogeneous tumour microenvironment (TME). The observed extensive copy number variations (CNVs) within epithelial cells support the established role of CNVs in LUAD progression, mirroring findings that link CNV-induced oncogene amplification and tumour suppressor deletions to cancer development.

A critical contribution of our study lies in the subclassification of tumour epithelial cells into six distinct clusters (C0-C5), each marked by unique gene expression signatures, such as S100A9 in C0 and SFTPC in C1. This detailed subclassification extends the work of Kim et al., who also demonstrated LUAD epithelial diversity, and suggests potential subtype-specific therapeutic vulnerabilities.

Our trajectory and differentiation analyses using CytoTRACE and Monocle highlighted that C4 and C1 represent the most differentiated states, whereas C3 and C5 remain at less differentiated, possibly stem-like stages. These observations are consistent with Guo et al. [32] and Puram et al. [33], who described LUAD as comprising a continuum from stem-like to differentiated phenotypes, influencing treatment resistance and relapse potential.

Furthermore, our investigation into cell-cell communication revealed complex ligand-receptor interactions within the TME, supporting Lambrechts et al.’s conclusions [30] on the importance of tumour-stromal crosstalk in modulating immune responses. The identification of these communication networks provides mechanistic insight and highlights candidate targets for disrupting tumour-supportive interactions.

A novel aspect of our study is the assessment of palmitoylation activity across tumour subtypes, where C1 cells displayed significantly elevated palmitoylation scores. This finding builds on prior research by Resh and Blaskovic et al. [34,35], implicating palmitoylation in oncogenic signalling. The enrichment of palmitoylation in C1 cells suggests subtype-specific metabolic dependencies that could be exploited therapeutically.

Moreover, our transcription factor (TF) network analysis delineated three major regulatory modules (M1-M3), reinforcing the concept that TF circuits underpin LUAD oncogenesis, as highlighted by Finlay et al. [36]. By mapping key TFs to specific cellular subtypes, we provide actionable targets to disrupt these oncogenic networks.

Our prognostic model, derived from C1 marker genes using univariate Cox and LASSO+Cox analyses, identified genes significantly associated with survival outcomes, paralleling biomarker discovery studies such as those by Thorsson et al. [[37], [38]]. The integration of immune infiltration analysis revealed that these prognostic genes correlate with immune checkpoint expression and immune cell composition, underscoring the interplay between tumour-intrinsic programmes and immune evasion strategies.

Consistent with the findings of Rizvi et al. [39], our observation that tumour mutational burden (TMB) correlates with both risk scores and overall survival reaffirms its utility as a predictive biomarker for immunotherapy responsiveness in LUAD.

While our study primarily focused on transcriptomic heterogeneity, it is important to acknowledge potential demographic influences. The TCGA-LUAD cohort used for model training predominantly consists of patients of European ancestry, with a near-equal distribution of males and females, and a broad age range (mean age ∼65 years). These demographic characteristics may influence gene expression profiles, immune infiltration patterns, and treatment responses. Future studies incorporating more diverse cohorts—including East Asian, African, and Hispanic populations—are needed to validate the model’s generalizability and assess ethnicity-specific palmitoylation signatures. Additionally, factors such as smoking history and comorbidities, which may affect LUAD progression, were not fully integrated and warrant further exploration.

A major strength of this study lies in its integration of scRNA-seq with multi-layered computational analyses to generate a comprehensive atlas of LUAD heterogeneity and regulatory networks. However, limitations include the lack of functional validation for identified targets and reliance on publicly available datasets, which may introduce sampling bias. To improve clarity, we acknowledge that the limitations section refers primarily to the downstream mechanistic investigation of ASPH (aspartate beta-hydroxylase). While our study demonstrated that ASPH overexpression promotes LUAD cell proliferation, migration, and EMT, and that its knockdown impairs these malignant phenotypes, we did not fully delineate the specific molecular pathways mediating these effects. Furthermore, other candidate genes identified in the 12-gene risk model, such as CKS2 and IRX3, were not subjected to functional validation, which represents another limitation. Future studies incorporating in vivo models and signaling pathway analyses (e.g., Notch or EGF-related pathways) are warranted to fully understand the oncogenic role of ASPH and the broader regulatory landscape of the model genes.

The findings of this study hold promising translational implications for the clinical management of lung adenocarcinoma. First, the 12-gene risk model derived from high-palmitoylation tumor cells demonstrates strong prognostic value and could be adapted for use in clinical risk stratification. Integration of this model into clinical practice may facilitate early identification of high-risk LUAD patients and guide therapeutic decision-making. Second, our results highlight ASPH as a potential therapeutic target; its overexpression promotes tumor proliferation, epithelial–mesenchymal transition, and immune evasion, while its silencing impairs tumor growth and invasion. These features suggest that ASPH inhibitors or palmitoylation modulators may serve as novel treatment avenues. Furthermore, the correlation between the risk signature, tumor mutational burden (TMB), and immune checkpoint expression provides a rationale for combining targeted inhibition of palmitoylation-related pathways with immune checkpoint blockade. Future translational research, including in vivo validation and clinical trials, will be necessary to assess the safety, efficacy, and applicability of these findings in real-world settings.

These findings collectively advance our understanding of LUAD pathogenesis by elucidating cellular heterogeneity, differentiation trajectories, and molecular dependencies. Future research should focus on experimental validation of subtype-specific therapeutic targets, particularly palmitoylation inhibitors and TF network disruptors, and assess their efficacy in overcoming resistance mechanisms.

In summary, our study provides a high-resolution map of LUAD at the single-cell level, revealing critical insights into tumour biology and potential strategies for precision oncology.

Ethics approval and consent to participate

This study involving human participants was reviewed and approved by the Ethics Committee of Henan Provincial People’s Hospital (Ethics approval number: 2024–180–02; serial number: AF/SC-08/05.0). All procedures were conducted in accordance with the Declaration of Helsinki and relevant national and institutional guidelines. Written informed consent was obtained from all participants prior to inclusion in the study, including consent for the collection, analysis, and anonymized publication of their clinical data.

Consent for publication

Not applicable to this work.

Availability of data and materials

All findings and data supporting this study are included within the article. For any further information, the corresponding author can be contacted.

Funding

This research received no specific grant from any funding agency.

CRediT authorship contribution statement

Taibo Huang: Writing – review & editing, Writing – original draft, Software, Resources, Data curation, Conceptualization. Lijie Kou: Supervision, Software, Resources, Formal analysis, Conceptualization. Qianqian Zhang: Visualization, Validation, Supervision, Methodology, Data curation. Xueya Liu: Visualization, Supervision, Formal analysis, Data curation. Xingang Hu: Writing – review & editing, Writing – original draft, Visualization, Resources, Project administration, Methodology, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.