A total of two LUAD-related datasets, GSE31210 and GSE68465(10), were retrieved from the GEO database (11), in which the control samples comprised non-tumorous normal lung tissues rather than patient-matched adjacent tissues (GSE31210: 20 normal, 226 LUAD; GSE68465: 19 normal, 443 LUAD); these datasets were chosen for their large sample sizes, availability of normal controls, and rich clinical annotation (including overall survival), making them widely used benchmark cohorts for discovery (GSE31210) and external validation (GSE68465). DEGs were identified using the ‘limma’ R package (v3.60.6, Bioconductor), implemented in R v4.4.1 (www.r-project.org/foundation/) (12), with screening criteria set as |log₂fold change|>1 and P<0.05. Volcano plots were generated using the ‘EnhancedVolcano’ package for visualization (13).

WGCNA was performed on the GSE31210 dataset using the ‘WGCNA’ R package (14), because this dataset contains both tumor and normal samples with comprehensive clinical annotation and a sufficiently large sample size, making it suitable for constructing reliable co-expression networks to identify key gene modules associated with LUAD. A soft-thresholding power β=6 was chosen based on the scale-free topology criterion (scale-free R²>0.85). The minimum module size was set to 30 to ensure module robustness. The resulting sample dendrogram and trait heatmap are presented in Fig. S1. Modules with a Pearson correlation coefficient of r>0.92 and P<0.05 were considered statistically significant, with correlations evaluated between module eigengenes and clinical traits (such as tumor vs. normal status, stage), measured from the sample annotation of the GSE31210 dataset.

GSEA was performed using GSEA software (v4.3.2) (15) on the entire set of DEGs, whereas Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the ‘clusterProfiler’ R package (v4.12.0; www.r-project.org/foundation/) (16,17) on the same DEG set. GO analysis included biological process, cellular components and molecular functions. KEGG analysis was used to identify significantly enriched signaling pathways. P<0.05 and q<0.05 were considered to indicate a statistically significant difference.

Identified DEGs were submitted to the STRING database to construct a PPI network (18), with a confidence score threshold of >0.9 (high confidence). Interaction data were imported into Cytoscape (v3.10.0; https://cytoscape.org/) for visualization. Network topology was analyzed using two centrality algorithms: Degree Centrality and Betweenness Centrality (19).

A total of two machine learning algorithms were applied: Random Forest (RF) and Support Vector Machine-Recursive Feature Elimination (SVM-RFE) (20,21). RF analysis was performed using the ‘randomForest’ package (v4.7-1.2) in R v4.4.1 (www.r-project.org/foundation/), and feature importance was ranked accordingly according to the default feature importance measures implemented in the ‘randomForest’ R package (Mean Decrease Accuracy and Mean Decrease Gini, as recommended by the package developers). SVM-RFE analysis was performed using the ‘e1071’ package (v1.7-14) in R v4.4.1 (www.r-project.org/foundation/) to further select optimal features based on gene importance. A linear kernel was applied with the default cost parameter (C=1), and feature ranking was performed using 10-fold cross-validation. The gene subset corresponding to the lowest cross-validated classification error was selected as the optimal feature set. P<0.05 was considered to indicate a statistically significant difference.

RT-qPCR (22) was used to validate the expression levels of caveolin-1 (CAV1) and cadherin 5 (CDH5) in tumor and adjacent normal tissues from 50 patients with LUAD, with adjacent tissues collected at least 5 mm away from the tumor margin. All tissue specimens were collected during surgical resection at Hebei Provincial Hospital of Traditional Chinese Medicine (Shijiazhuang, China) between May 2021 and June 2023, immediately snap-frozen in liquid nitrogen, and stored at -80˚C until RNA extraction. Among the patients, 28 were male and 22 were female. Inclusion criteria were histopathologically confirmed LUAD and availability of paired tumor and adjacent normal tissues; exclusion criteria included prior chemotherapy, radiotherapy, or other malignancies. The median age was 63 years (range, 45-78 years). The study protocol was approved by the Ethics Committee of Hebei Provincial Hospital of Traditional Chinese Medicine (Shijiazhuang, China; approval no. HBZY2023-KY-076-01) and followed the ethical principles of The Declaration of Helsinki. Written informed consent was obtained from all participants.

Total RNA was extracted using TRIzol^® reagent (Thermo Fisher Scientific, Inc.), and RNA purity and concentration were assessed using a NanoDrop 2000 spectrophotometer. High-quality RNA was reverse transcribed into cDNA, and qPCR was performed using SYBR Green PCR Master Mix on a QuantStudio 6 Real-Time PCR System (Thermo Fisher Scientific, Inc.). Relative gene expression was calculated using the 2^-ΔΔCq method (22), with GAPDH used as the internal control. Primers were designed and validated using Primer-BLAST. The following primer sequences were used: GAPDH forward, 5'-AATGGACAACTGGTCGTGGAC-3' and reverse, 5'-CCCTCCAGGGGATCTGTTTG-3'; CAV1 forward, 5'-GCAGAACCAGAAGGGACACAG-3' and reverse, 5'-CCAAAGAGGGCAGACAGCAAGC-3'; and CDH5 forward, 5'-AGGTGCTAACCCTGCCCAAC-3' and reverse, 5'-CGGAAGACCTTGCCCACATA-3'.

ROC curve analysis was performed using the ‘pROC’ package (v1.18.5) in R v4.4.1 (www.r-project.org/foundation/) (23). An area under the curve (AUC) of >0.9 was considered to indicate notable diagnostic performance. P<0.05 was considered to indicate a statistically significant difference.

scRNA-seq data related to LUAD were obtained from the Tumor Immune Single Cell Hub (TISCH) database (http://tisch.comp-genomics.org) (24). The expression levels of key genes across different cell types were visualized using heatmaps and feature plots, which were generated directly from the TISCH database portal based on preprocessed single-cell RNA-seq data provided by the database.

Statistical analysis was performed using GraphPad Prism 9.0 (Dotmatics). Differences between tumor and normal tissues in The Cancer Genome Atlas (TCGA) and GEO datasets were assessed using unpaired Student's t-test, while differences between matched tumor and adjacent normal tissues in the clinical validation cohort and paired comparisons within the TCGA cohort (Fig. 3D) were assessed using paired Student's t-test. For comparisons among multiple subgroups, one-way ANOVA followed by Tukey's post hoc test was applied. Kaplan-Meier survival analysis was performed using the Kaplan-Meier Plotter database (http://kmplot.com/analysis/) based on TCGA-LUAD data; patients were stratified into high- and low-expression groups using the median expression value as the cut-off, and statistical significance was evaluated using the log-rank test. P<0.05 was considered to indicate a statistically significant difference.

Based on the GSE31210 dataset, a total of 9,292 DEGs were identified, including 5,399 upregulated and 3,893 downregulated genes in LUAD tumor tissues compared with controls, as this dataset contains both tumor and non-tumor lung tissues with comprehensive clinical annotation, making it suitable for DEG identification. A volcano plot was generated to visualize these DEGs (Fig. 1A). To further identify key genes, a WGCNA was performed using the same dataset, resulting in the identification of 25 co-expression modules (Fig. 1B).

Correlation analysis between these modules and LUAD phenotypes revealed that 17 modules were significantly associated with LUAD, among which the brown module showed the strongest correlation (Fig. 1C). This suggested that this module may serve a critical role in the pathogenesis and progression of LUAD. Further scatter plot analysis of module membership and gene significance demonstrated a strong positive correlation within the brown module (Pearson's r=0.92, P=1x10^-1,000), indicating its relevance as a key module (Fig. 1D).

Subsequently, a Venn diagram was used to identify overlapping genes between DEGs and those in the key module identified by WGCNA, resulting in 1,059 common genes (Fig. 1E). GO and KEGG enrichment analyses were then performed on these overlapping genes. GO analysis revealed enrichment in signaling-related molecular functions such as ‘signaling receptor binding’, ‘kinase activity’ and ‘protein-containing complex binding’, cellular components such as ‘vesicle’, ‘endomembrane system’ and ‘extracellular region part’, as well as biological processes such as ‘regulation of cell communication’, ‘cellular developmental process’ and ‘negative regulation of cellular process’ (Fig. 1F). KEGG analysis revealed significant enrichment in ‘TGF-β signaling pathway’, ‘TNF signaling pathway’, ‘PI3K-Akt signaling pathway’ and ‘pathways in cancer’, suggesting that these genes may serve regulatory roles in LUAD initiation and progression (Fig. 1G).

A PPI network was constructed using the 1,059 overlapping genes (Fig. 2A). Topological analysis in Cytoscape using the Degree Centrality and Betweenness Centrality algorithms identified the top 20 hub genes (Fig. 2B and C, respectively). The intersection of these two analyses yielded 13 shared hub genes (Fig. 2D).

Subsequently, two machine learning approaches, RF and SVM-RFE, were employed to further refine key gene selection. The RF algorithm identified the top five genes based on feature importance scores (Fig. 2E), whereas SVM-RFE selected another top five genes through iterative optimization (Fig. 2F). A Venn diagram analysis of both methods identified two overlapping candidate genes: CAV1 and CDH5 (Fig. 2G).

Furthermore, GSEA indicated that both CAV1 and CDH5 were significantly enriched in signaling pathways such as MAPK, Wnt and TGF-β, suggesting their involvement in LUAD development, progression and modulation of the tumor immune microenvironment (Fig. 2H and I).

To explore the roles of CAV1 and CDH5 in LUAD, their expression levels were analyzed in the GSE31210 dataset. Both genes were significantly downregulated in LUAD tissues compared with those in normal tissues (Fig. 3A). Moreover, consistent downregulation trends were observed in the GSE68465 and TCGA datasets for CAV1 (Fig. 3B and C). Although CDH5 expression was upregulated in the LUAD tissues in the GSE68465 dataset, it was downregulated in the TCGA cohort. To further investigate whether tumor stage contributes to this inter-dataset discrepancy, a subgroup analysis was performed using the GSE68465 dataset, stratifying patients by N stage (N0, N1 and N2). As shown in Fig. S2, CDH5 expression was significantly higher in N0-stage samples compared with that in N1 (P<0.001) and a significant difference was also observed between N1 and N2 (P<0.05). This stage-associated decline supports a biological trend of progressive CDH5 downregulation during LUAD advancement.

Pairwise analyses in the TCGA dataset demonstrated significantly lower expression of CAV1 and CDH5 in LUAD tissues compared with available adjacent normal lung tissues, eliminating potential interference from inter-individual biological variation (Fig. 3D). Similarly, RT-qPCR results from 50 LUAD samples demonstrated significantly decreased expression levels of CAV1 and CDH5 in tumor tissues compared with those in matched adjacent normal tissues (P<0.0001; Fig. 3E). Notably, although the difference in CDH5 expression between tumor and control tissues was relatively small in magnitude, it remained statistically significant. This expanded validation supports the reliability and reproducibility of the bioinformatics-driven identification of these genes.

Furthermore, Kaplan-Meier survival analysis revealed that high expression of CAV1 and CDH5 was significantly associated with improved overall survival, suggesting that CAV1 and CDH5 have potential tumor suppressor functions (Fig. 3F). A diagnostic nomogram model based on the expression of CAV1 and CDH5 was then constructed (Fig. 3G), with calibration curves indicating strong agreement between predicted and actual outcomes, as the calibration curves closely overlapped with the ideal 45˚ reference line (Fig. 3H). In addition, ROC curve analysis demonstrated notable diagnostic performance for CAV1 (AUC=0.979) and CDH5 (AUC=0.969) in GSE31210, and this was validated in both GSE68465 and TCGA datasets (Fig. 3I), supporting the potential of CAV1 and CDH5 as diagnostic biomarkers for LUAD.

To precisely characterize the cell-type-specific expression and potential functions of CAV1 and CDH5, three LUAD-related scRNA-seq datasets (EMTAB6149, GSE127465 and GSE131907) from the TISCH database were analyzed: EMTAB6149 includes ~20,000 cells in total from the primary LUAD tissues of 4 patients; GSE127465 comprises ~14,000 cells in total from surgically resected LUAD tumors and adjacent tissues from 7 patients; and GSE131907 includes ~208,506 cells in total derived from 11 patients with LUAD. These datasets represent a diverse range of LUAD microenvironments and provide robust resolution of immune and stromal cell populations. Both genes were revealed to be highly expressed in endothelial cells (Fig. 4A), suggesting potential roles in tumor-associated angiogenesis, vascular barrier maintenance or microenvironment regulation. To visualize their spatial expression across immune cell subpopulations, t-distributed stochastic neighbor embedding plots were generated for CAV1 and CDH5, which showed that both genes were predominantly expressed in endothelial cell clusters, with lower expression observed in other immune cell subtypes (Fig. 4B). Additionally, violin plots were used to quantify expression levels across major cell types, revealing marked heterogeneity in gene expression patterns, with CAV1 and CDH5 showing higher expression in endothelial cells but minimal expression in most immune cell subsets, indicating substantial variability across cell populations (Fig. 4C). These findings not only corroborate the high expression observed in heatmaps, particularly in endothelial cells, but also highlight the complexity of gene regulation in the tumor microenvironment, providing a theoretical basis for further functional studies.