DCN, NPM3 and SULF1 are hub genes related to vasculogenic mimicry in lung adenocarcinoma

Abstract

Aim

Vasculogenic mimicry (VM), a process in which cancer cells form endothelial cell-independent vascular networks, is a hallmark of tumor aggressiveness in lung adenocarcinoma (LUAD) and supports tumor growth and metastasis. This study aims to identify and validate key genes associated with VM formation in LUAD, and to elucidate their functional roles and clinical significance.

Methods

Transcriptomic data from LUAD samples were analyzed using differential expression analysis (DEA) and weighted gene co-expression network analysis (WGCNA) to identify VM-associated genes. Machine learning algorithms were applied to refine the selection and identify hub genes. Functional enrichment and immune infiltration analyses were performed. The role of SULF1 in VM was further validated through in vitro and in vivo experiments.

Results

We identified 10,810 differentially expressed genes. WGCNA revealed 101 VM-associated genes, predominantly within the “yellow” and “brown” modules. Machine learning pinpointed three key regulators: downregulated decorin (DCN) and upregulated nucleoplasmin 3 (NPM3) and sulfatase 1 (SULF1). Functional enrichment analysis highlighted their involvement in extracellular matrix (ECM) organization and ribosomal pathways. Immune infiltration analysis indicated a positive correlation between DCN and immune cell presence, whereas NPM3 and SULF1 showed negative correlations. Critically, SULF1 overexpression promoted VM formation in vitro by enhancing cell migration and invasion, mediated through the vascular endothelial growth factor (VEGF)/transforming growth factor beta (TGF-β)/Vimentin signaling axis, and accelerated tumor growth in vivo.

Conclusion

We identified DCN, NPM3, and SULF1 as key biomarkers of VM in LUAD. SULF1, in particular, plays a central role in driving VM formation and tumor progression. These findings offer novel mechanistic insights and highlight potential therapeutic targets for LUAD treatment.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00432-025-06361-0.

Introduction

Global cancer statistics indicate that lung cancer caused approximately 2.48 million new cases and 1.81 million deaths worldwide in 2022, making it the leading cause of cancer-related mortality and imposing a substantial public health burden (Bray et al. 2024). Histologically, lung cancer is categorized into non-small cell lung cancer (NSCLC), which constitutes 80–85% of cases, and small-cell lung cancer (SCLC). NSCLC subtypes include adenocarcinoma (approximately 40%), squamous cell carcinoma (25–30%), and large cell carcinoma (10–15%) (Thai et al. 2021). Among these, lung adenocarcinoma (LUAD) is the most common subtype and is characterized by high invasiveness, frequent recurrence, and active angiogenesis. Although advances in targeted therapy and immunotherapy have markedly improved patient outcomes, drug resistance and disease relapse continue to pose significant clinical challenges (Hirsch et al. 2017). Therefore, elucidating the underlying pathological mechanisms remains a critical research priority.

Vasculogenic mimicry (VM) describes the ability of tumor cells to form endothelial cell-independent, vascular-like channels, a phenomenon first identified in human uveal melanoma (Tang et al. 2024). In contrast to conventional angiogenesis, VM involves aggressive tumor cells directly forming perfusable tubular networks through phenotypic plasticity, thereby establishing an alternative blood supply system that is closely linked to tumor malignancy (Morales-Guadarrama et al. 2021). VM has been clinically associated with poor prognosis and shortened survival in multiple aggressive cancers (Hujanen et al. 2020; Liu et al. 2022).Its presence also contributes to the limited efficacy of anti-angiogenic therapies, such as those targeting vascular endothelial growth factor (VEGF) (Huang et al. 2023). In LUAD, VM has been recognized as an indicator of unfavorable progression-free survival (He et al. 2021). Mechanistically, hypoxia-inducible factor 1-α (HIF-1α) has been shown to promote LUAD metastasis and VM formation by upregulating Neuropilin-1 (NRP1), which modulates the expression of vascular endothelial cadherin (VE-cadherin) and vimentin, thereby facilitating tumor progression (Fu et al. 2021). These findings underscore VM as a key pathogenic mechanism in LUAD and a promising target for therapeutic intervention.

This study aims to systematically characterize the expression patterns and functional contributions of VM-associated genes in LUAD using integrated bioinformatics and experimental validation. The ultimate goal is to identify novel therapeutic targets and inform personalized intervention strategies for LUAD management.

Materials and methods

Data source

Gene expression data and corresponding clinical information for LUAD were obtained from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/), which included 541 tumor tissues and 59 adjacent normal tissues, serving as the primary training cohort. For independent validation, the GSE140797 dataset, comprising 7 LUAD and 7 normal samples, was downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/). A set of 82 vasculogenic mimicry-related genes (VMRGs) was curated from the "ANGIOGENESIS.v2023.2.Hs" and "HALLMARK_ANGIOGENESIS.v2023.2.Hs" gene sets within the Molecular Signatures Database (MSigDB, https://www.gsea-msigdb.org/gsea/msigdb).

Differential expression analysis (DEA)

DEA between LUAD and normal samples in the training cohort was performed using the "limma" R package. Genes with an absolute log2-fold change (|log2FC|) greater than 1 and an adjusted p-value (Padj) of less than 0.05 were defined as statistically significant differentially expressed genes (DEGs). The results were visualized using a volcano plot. Additionally, a heatmap was generated to display the expression patterns of the top 10 most significantly up- and down-regulated DEGs, ranked by their Padj values.

Weighted gene co-expression network analysis (WGCNA)

In this study, we employed VMRGs as the gene set and conducted single-sample gene set enrichment analysis (ssGSEA) to score all disease samples in the training set. The resulting ssGSEA scores were treated as quantitative traits to construct a co-expression network using the R package “WGCNA,” aiming to identify gene modules most strongly associated with the ssGSEA scores.

Following ssGSEA scoring, hierarchical clustering based on Euclidean distance of expression levels was performed to detect potential outlier samples; no outliers were identified and thus no samples were excluded. We then selected an appropriate soft-thresholding power (β) from 1 to 20 to ensure a scale-free co-expression network. Based on the criterion of achieving a scale-free topology fit index (R²) > 0.85, the optimal soft-thresholding power was determined as β = 3. The relationship between β and both the scale-free model fit (R²) and average connectivity was evaluated accordingly.

A signed co-expression network was constructed using β = 3, with a minimum module size set to 30 genes. Module-trait associations were assessed by correlating module eigengenes with the ssGSEA scores. Modules exhibiting a strong absolute correlation with the ssGSEA scores (|cor|> 0.3, p < 0.05) were considered key modules. Among these, the yellow and brown modules showed the most significant correlations with the VM phenotype (ssGSEA score).

To further prioritize genes within these key modules, we applied thresholds for Gene Significance (GS > 0.2), which reflects the correlation between gene expression and the ssGSEA score, and Module Membership (MM > 0.7), indicating the correlation between gene expression and module eigengene. This filtering process yielded 101 high-confidence candidate genes associated with VM.

Functional enrichment analysis and protein–protein interaction network establishment

Candidate genes were identified by taking the intersection of the DEGs and the VM-associated genes derived from the WGCNA. To elucidate their functional roles, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed using the clusterProfiler R package. The GO analysis comprehensively covered biological processes (BP), molecular functions (MF), and cellular components (CC). For all enrichment results, statistical significance was determined using an adjusted p-value (FDR) < 0.05 after applying the Benjamini-Hochberg (BH) procedure to control the false discovery rate in multiple testing. Furthermore, a protein–protein interaction (PPI) network was constructed using the STRING database with a minimum interaction score of 0.4 and visualized with Cytoscape software.

Machine learning

Hub gene identification was performed using two complementary machine learning algorithms: (1) Least Absolute Shrinkage and Selection Operator (LASSO) regression was carried out with the "glmnet" R package. The optimal regularization parameter (lambda) was determined through tenfold cross-validation. To ensure model parsimony and prevent overfitting, the lambda value corresponding to one standard error above the minimum mean squared error (lambda.1se) was selected. This procedure identified 12 candidate feature genes. (2) Random Forest (RF) analysis was implemented using the "randomForest" R package with 500 trees under default settings. Genes were ranked by importance according to the %IncMSE (Percentage Increase in Mean Squared Error) metric. A threshold of MeanDecreaseGini > 1 was applied, retaining the top 26 significant genes. The final set of hub genes was defined as the intersection of the genes selected by both LASSO and RF approaches.

Expression validation

The expression levels of candidate genes were compared between tumor and normal tissues in both the training and validation cohorts using Wilcoxon rank-sum tests, with results presented as bar graphs. Protein expression was further assessed using immunohistochemistry data from the Human Protein Atlas (HPA; https://www.proteinatlas.org/). A gene was designated as a key biomarker only if it exhibited consistent differential expression across all datasets and achieved statistical significance (P < 0.05).

Molecular regulatory network analysis

To explore the transcriptional regulation of the key genes, we predicted their interacting transcription factors (TFs) using the ChEA database via the NetworkAnalyst platform (https://www.networkanalyst.ca/), and visualized the resulting mRNA-TF interaction network. Subsequently, a protein–protein interaction and co-expression network for each key gene and its associated TFs was constructed using GeneMANIA (http://genemania.org/) to identify functionally related genes and uncover predominant interaction types and biological functions within the network.

Gene set enrichment analysis (GSEA)

GSEA was performed using the "clusterProfiler" R package to identify biological pathways associated with the key genes. The analysis involved ranking all genes in the training set based on their correlation with each key gene, followed by enrichment testing against the MSigDB collections "c2.all.v2024.1.Hs.symbols" (curated pathways) and "c5.all.v2024.1.Hs.symbols" (Gene Ontology terms). Pathways meeting the dual thresholds of nominal p-value < 0.05 and FDR < 0.05 were considered statistically significant.

Immune infiltration analysis

The immune infiltration landscape in the TCGA-LUAD cohort was characterized using ssGSEA implemented in the GSVA R package, which computed sample-wise enrichment scores based on predefined gene signatures representative of each immune cell population to quantify the relative abundance of 28 immune cell types. The resulting enrichment scores were visualized in a heatmap. Differences in immune infiltration between tumor and normal tissues were assessed using Pearson correlation analysis (visualized with the corrplot package) and Wilcoxon rank-sum tests (visualized with box plots). Immune cell subtypes showing significant correlations with key genes (|r|> 0.3, p < 0.05) were considered biologically relevant.

Cell culture and transfection

The A549 cell line was obtained from the China Center for Type Culture Collection (CCTCC, Wuhan, China) and cultured in RPMI 1640 medium (Gibco, 11875093) containing 10% fetal bovine serum (Gibco, A5256701) at 37 °C with 5% CO₂. To overexpress sulfatase 1 (SULF1), lentiviral vectors carrying the SULF1 sequence (Ov-SULF1) or an empty vector control (Ov-NC) were constructed using the GV493 backbone (Ledel Biotechnology, Guangzhou, China). After 48 h of transduction, transfection efficiency was assessed. Stable polyclonal cell lines were subsequently established by selection with 6 μg/ml puromycin (Beyotime, China).

RNA extraction and quantitative real-time polymerase reaction (qRT-PCR)

Total RNA was extracted with RNAiso Plus reagent, and cDNA was synthesized using the PrimeScript RT reagent kit under the following conditions: 25 °C for 10 min, 42 °C for 30 min, and 85 °C for 5 s. qRT-PCR was performed on an ABI 7600 system with SYBR Premix DimerEraser, using the following cycling protocol: 95 °C for 5 min, then 40 cycles of 95 °C for 15 s and 60 °C for 32 s. Gene expression was quantified via the 2^ (–ΔΔCt) method, normalizing to GAPDH as the internal control. The primer sequences used were as follows: SULF1 (forward: 5′-GGCTGCTCAGGAAGTAGATAG-3′; reverse: 5′-GCCAGTGGTTGTTGTCATG-3′) and GAPDH (forward: 5′-GGGAAACTGTGGCGTGAT-3′; reverse: 5′-GAGTGGGTGTCGCTGTTGA-3′).

(Livak and Schmittgen 2001).

Transwell assay

Cell migration was evaluated using Transwell chambers (BD Falcon, 353,097). Briefly, 1 × 10⁵ treated cells in serum-free medium were seeded into the upper chamber, while the lower chamber was filled with complete medium as a chemoattractant. After incubation for 12–48 h at 37 °C, non-migratory cells on the upper surface of the membrane were carefully removed. Cells that had migrated to the lower surface were fixed, stained with 0.1% crystal violet, and counted using ImageJ software.

For the invasion assay, Transwell chambers were pre-coated with Matrigel matrix diluted 1:3 in serum-free medium and allowed to polymerize at 37 °C for 2 h, followed by overnight equilibration. Cells were then seeded following the same procedure as the migration assay. After 24–48 h, invaded cells were quantified using the same staining and counting method.

Three-dimensional cell culture

Tube formation ability was assessed with a Matrigel-based assay. In brief, 2 × 10⁴ treated A549 cells were seeded per well into 96-well plates pre-coated with growth factor-reduced Matrigel (BD Biosciences, 356,234) and cultured for 48 h under standard conditions. The number of branch points per field was quantified under a phase-contrast microscope to evaluate VM formation.

Periodic acid-Schiff (PAS) staining

Cells were seeded in 96-well plates at a density of 1 × 10⁴ cells per well and treated as indicated. After treatment, PAS staining was performed as follows: cells were washed with PBS, fixed with PAS fixative for 10 min, and rinsed with distilled water. They were then oxidized in 0.5% periodic acid for 15 min at room temperature, washed again, and incubated in Schiff’s reagent for 10 min in the dark. After a 10-min rinse under running water, nuclei were counterstained with hematoxylin for 2 min, followed by a final distilled water wash. Stained cells were visualized under a bright-field microscope.

Western blot analysis

Protein lysates were separated by 10% SDS-PAGE and transferred onto PVDF membranes. After blocking with 5% skim milk for 1 h at room temperature, the membranes were incubated overnight at 4 °C with primary antibodies against VEGF (ab46154, Abcam), TGF-β (81,746–2-RR, Proteintech), Vimentin (10,366–1-AP, Proteintech), and GAPDH (KC-5G5, Kangcheng Bio). Subsequently, membranes were incubated with appropriate HRP-conjugated secondary antibodies—Goat Anti-Rabbit IgG (1:20,000, 4050–05, SouthernBiotech) or Rabbit Anti-Mouse IgG (1:10,000, 6170–05, SouthernBiotech)—for 1 h at room temperature. Protein bands were visualized using an ECL detection system and quantified with a Bio-Rad imaging analyzer.

Animal model

Five-to-six-week-old female BALB/c nude mice (Guangdong MedKang Biotechnology Co., Ltd) were acclimatized for one week. Subsequently, 2 × 10⁶ A549 cells suspended in 200 μL PBS were subcutaneously injected into the right flank of each mouse (Day 0). Tumor size was measured every three days starting when the tumor diameter reached approximately 5 mm. For experimental procedures, mice were anesthetized with pentobarbital sodium (50 mg/kg, i.p.). Upon reaching the humane endpoint (tumor diameter ≤ 1.5 cm), mice were euthanized by CO₂ inhalation followed by cervical dislocation. Tumors were then harvested, photographed, and prepared for further analysis.

Statistical analysis

All statistical analyses were conducted using R software (version 4.1.3) and GraphPad Prism 8. Continuous variables are presented as mean ± standard deviation (SD). Differences between groups were evaluated using two-tailed Student’s t-tests, Pearson correlation analysis, and Wilcoxon rank-sum tests, with a P-value < 0.05 considered statistically significant.

Results

Screening of DEGs related to VM in LUAD

DEA identified 10,810 genes (6587 upregulated and 4,223 downregulated) that were significantly dysregulated in LUAD tissues compared to normal controls (Fig. 1A–B). WGCNA was subsequently performed, and with a soft-thresholding power of β = 3 (scale-free R² = 0.85), eight distinct co-expression modules were constructed (Fig. 1C–D). Among these, the yellow and brown modules showed the strongest correlation with VM activity (|correlation|> 0.3, p < 0.05; Fig. 1E–F). By applying dual filtering criteria (GS > 0.2 and MM > 0.7), 101 high-confidence VM-associated candidate genes were obtained from these key modules.

Fig. 1.

Screening of vasculogenic mimicry (VM)-related genes in lung adenocarcinoma (LUAD) through differential expression analysis (DEA) and weighted gene co-expression network analysis (WGCNA). A Volcano plot of differentially expressed genes (DEGs) between tumor and normal tissues (|log₂FC|> 1, adj. P < 0.05). B Heatmap of the top 10 up- and down-regulated DEGs (ranked by P-value). C Hierarchical clustering dendrogram of samples based on gene expression profiles. D Scale-free topology analysis for determining the optimal soft-thresholding power (β). E Dendrogram of co-expression modules identified by WGCNA, with colors indicating distinct modules. F Heatmap of module–trait correlations between WGCNA modules and single-sample gene set enrichment analysis (ssGSEA) scores of VM-related genes

Functional enrichment and PPI network of candidate genes

The integration of WGCNA and DEG analyses identified 30 high-confidence candidate genes (Fig. 2A). GO enrichment analysis revealed 67 significantly enriched terms (P < 0.05), including 28 biological processes (e.g., extracellular matrix (ECM) organization and translation), 26 cellular components (e.g., collagen-containing ECM and ribosome), and 13 molecular functions (e.g., rRNA binding) (Fig. 2B). KEGG pathway analysis further identified 10 significantly enriched pathways (P < 0.05), primarily related to ribosome and protein metabolism (Fig. 2C). A PPI network was constructed using the STRING database (minimum interaction score > 0.4), and the top 10 hub genes were identified using the cytoHubba plugin for subsequent analysis (Fig. 2D).

Fig. 2.

Functional enrichment and protein–protein interaction (PPI) analysis of candidate genes. A Venn diagram showing the overlap between DEGs and VM-related genes. B Kyoto encyclopedia of genes and genomes (KEGG) enrichment analysis of candidate genes. C Gene ontology (GO) enrichment analysis of candidate genes. D PPI network of candidate genes: red nodes represent the genes with the highest contribution, followed by the brown nodes, the yellow nodes are the lowest, and the connecting lines illustrate gene-pathway relationships

Identification of key genes through machine learning and expression validation

LASSO regression and RF analyses identified 12 and 26 candidate genes, respectively (Fig. 3A–B). The intersection of both gene sets yielded five overlapping genes: caveolae associated protein 1 (CAVIN1), nucleoplasmin 3 (NPM3), SULF1, decorin (DCN), and CORIN (Fig. 3C). Transcriptomic validation across multiple datasets confirmed consistent dysregulation of DCN (downregulated) and NPM3/SULF1 (upregulated) in tumor tissues (Wilcoxon test, p < 0.05; Fig. 3D). Immunohistochemistry results from the HPA further supported these expression patterns at the protein level (Fig. 3E), confirming DCN, NPM3, and SULF1 as key VM-associated genes in LUAD.

Fig. 3.

Identification and validation of key genes. A Least absolute shrinkage and selection operator (LASSO) coefficient profiles (left) and tuning parameter selection (right) for candidate genes. B Variable importance plot from random forest (RF) analysis. C Overlap of feature genes selected by LASSO and RF. D Differential expression of key genes in training and validation sets (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant; Wilcoxon test). E Representative immunohistochemical staining of Decorin (DCN), Nucleoplasmin 3 (NPM3), and Sulfatase 1 (SULF1) in normal and tumor tissues

Molecular regulatory network and GSEA of key genes

Transcriptional network analysis using NetworkAnalyst (ChEA database) identified 59 transcription factors interacting with DCN and NPM3, forming a regulatory network comprising 61 nodes and 63 edges, including 6 shared TFs (Fig. 4A). GeneMANIA analysis of DCN, NPM3, and SULF1 revealed a 20-gene interaction network predominantly involving physical interactions (77.6%) and co-expression (8.0%), with functional enrichment in proteoglycan and glycosaminoglycan metabolism (Fig. 4B). GSEA using the "c2.all.v2024.1.Hs.symbols" gene set showed significant pathway associations (P < 0.05, FDR < 0.05) for DCN (3,791 pathways), NPM3 (3,225 pathways), and SULF1 (3,289 pathways), including epidermal growth factor receptor (EGFR) signaling, hypoxia response, and translation initiation pathways. Analysis with the "c5.all.all.v2024.1.Hs.symbols" gene set further revealed broad functional involvement in mitochondrial gene expression, ribonucleoprotein complex assembly, and axoneme organization (Fig. 4C–H).

Fig. 4.

Molecular regulatory analysis of key genes. A Transcription factor (TF)–mRNA regulatory network of key genes. B Gene–gene interaction (GGI) network of key genes. C–H Gene set enrichment analysis (GSEA) of DCN, NPM3, and SULF1 using “c5.all.v2024.1.Hs.symbol” (top 5 pathways) and “c2.all.v2024.1.Hs.symbols” (top 3 pathways) gene sets, ranked by P-value

Immune infiltration analysis in LUAD

ssGSEA revealed distinct infiltration patterns of 28 immune cell types between LUAD and normal tissues (Fig. 5A). Correlation analysis indicated moderate-to-strong connectivity among these immune components, highlighting their coordinated activity within the tumor microenvironment (Fig. 5B). Comparative assessment identified 23 immune cell subsets with significantly altered infiltration levels in tumors (Fig. 5C), including activated B cells, CD4⁺ and CD8⁺ T cells, dendritic cells, and multiple myeloid cell populations. Importantly, correlation analysis showed that most of these immune subsets were negatively correlated with NPM3 and SULF1 expression, but positively correlated with DCN (Fig. 5D).

Fig. 5.

Tumor microenvironment in The Cancer Genome Atlas (TCGA)-LUAD. A Immune cell infiltration levels in LUAD vs. normal samples. B Correlation heatmap of 28 immune cell types (*P < 0.05, **P < 0.01, ***P < 0.001; Pearson correlation). C Differential immune infiltration between tumor and normal samples (P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant; Wilcoxon test). D Correlation heatmap between key genes and immune cells (P < 0.05, **P < 0.01, ***P < 0.001; Pearson correlation)

Prognostic validation and nomogram construction

As detailed in the Supplementary Materials, in-depth analysis of TCGA-LUAD data revealed distinct prognostic values of the three hub genes: NPM3 and SULF1 functioned as risk factors, while DCN exhibited a protective trend (Supplementary Fig.S1-3). Time-dependent ROC analysis demonstrated modest yet consistent predictive performance for all three genes at 1, 3, and 5 years. The Sankey diagram (Supplementary Fig.S4) further elucidated the multivariate relationships between clinicopathological characteristics—including TNM stage, smoking status, and expression levels of DCN, NPM3, and SULF1—and patient outcomes, providing a comprehensive visualization of how these variables collectively influence survival. Furthermore, a comprehensive prognostic nomogram integrating the expression of SULF1, NPM3, and DCN with clinicopathological features was developed, which showed excellent calibration and effectively predicted 1-, 3-, and 5-year overall survival probabilities (Supplementary Fig.S5).

Overexpression of SULF1 promoted VM in A549 cells

The functional role of SULF1 was investigated in A549 cells using lentiviral overexpression. Successful transfection was confirmed by qRT-PCR (Fig. 6A). SULF1 overexpression significantly promoted VM, as shown by enhanced tube formation in Matrigel-based assays (Fig. 6B). PAS staining further revealed increased deposition of basement membrane-like components (Fig. 6C), consistent with elevated VM activity.

Fig. 6.

SULF1 overexpression promotes VM in A549 cells. A qRT-PCR validation of SULF1 overexpression, (n = 3). B Left: Representative images of tube formation in Matrigel-based 3D culture (n = 3), Right: quantification of tube numbers from five randomly selected fields per sample (n = 3); Scale bar: 200 μm. C Left: Periodic acid–Schiff (PAS) staining showing basement membrane-like structures (n = 3), Right: Quantification of PAS-positive areas (n = 3); Scale bar: 50 μm. Data are presented as mean ± SD. ****P < 0.0001

Overexpression of SULF1 enhanced the migration and invasion of A549 cells

Functional assays showed that SULF1 overexpression (Ov-SULF1) significantly enhanced the migratory and invasive abilities of A549 cells compared to the negative control (NC), as demonstrated by Transwell migration and Matrigel invasion assays (Fig. 7A–B). Western blot analysis further revealed that SULF1 overexpression led to increased protein levels of VEGF, TGF-β, and Vimentin (Fig. 7C), suggesting activation of a pro-metastatic signaling axis.

Fig. 7.

SULF1 overexpression enhances the migratory and invasive capacities of A549 cells. A Left: representative images of the Transwell migration assay. Right: Quantification of migrated cells (n = 3). B Left: representative images of the Matrigel invasion assay, Right: Quantification of invaded cells (n = 3); C Western blot analysis of migration- and invasion-associated proteins (left) and corresponding densitometric quantification (right) (n = 3). Data are presented as mean ± SD. ***P < 0.001, ****P < 0.0001. Scale bar: 100 µm

Overexpression of SULF1 accelerated tumor growth in nude mice

In vivo xenograft experiments showed that tumors overexpressing SULF1 grew significantly larger than control tumors, as observed in both gross morphological appearance (Fig. 8A) and quantitative measurements of tumor size and volume (Fig. 8B–C).

Fig. 8.

SULF1 overexpression promotes tumor growth in a xenograft mouse model. A Representative photographs of tumor-bearing nude mice (n = 5). B, C Tumor size and volume measurements (n = 5). Data are presented as mean ± SD. ****P < 0.0001

Discussion

Integrated bioinformatics and experimental validation in this study systematically delineated the functional relevance of VM-associated genes in the pathogenesis of LUAD. Cross-dataset analysis identified five consistently dysregulated VM-linked genes—CAVIN1, NPM3, SULF1, DCN, and CORIN—among which DCN was significantly downregulated and NPM3 and SULF1 were markedly upregulated in tumor tissues across both training and validation cohorts. Functional enrichment analyses linked these genes to pivotal biological processes, including extracellular matrix (ECM) reorganization, ribosomal activity, and protein metabolism, suggesting their active involvement in LUAD progression. Immune infiltration profiling further revealed a negative correlation between NPM3/SULF1 expression and immune cell abundance, underscoring their potential immunomodulatory roles. Prognostic analyses substantiated the clinical relevance of these biomarkers: time-dependent ROC curves confirmed that all three genes exhibited consistent predictive capability for 1‑, 3‑, and 5‑year survival, while the constructed nomogram—integrating SULF1, NPM3, and DCN expression with key clinicopathological variables—demonstrated high calibration accuracy in predicting overall survival probabilities. Collectively, this evidence establishes DCN, NPM3, and SULF1 as clinically significant VM biomarkers and promising therapeutic targets in LUAD.

Recent studies have elucidated the complex regulatory network underlying VM, which involves multiple interconnected mechanisms. At the cellular level, VM formation is promoted by epithelial–mesenchymal transition (EMT) driven by transcription factors such as Snail, ZEB1/2, and Twist1 via TGF-β, Wnt, and Notch signaling, along with hypoxic HIF-1α–VEGF activation. These pathways cooperate with components of the tumor microenvironment—including cancer stem cells (CSCs), tumor-associated macrophages (TAMs), and cancer-associated fibroblasts (CAFs)—to facilitate VM (Luo et al. 2020). At the molecular level, key signaling axes such as EphA2–FAK and PI3K/AKT, together with matrix metalloproteinase (MMP)-mediated extracellular matrix remodeling, play essential roles (Treps et al. 2021). Additionally, non-coding RNAs—including tumor-suppressive miRNAs (e.g., miR-141, miR-27a-3p) and oncogenic lncRNAs (e.g., MALAT1, LINC00339)—further fine-tune VM dynamics (Ibarra-Sierra et al. 2025). This multilayered regulation highlights VM as a sophisticated adaptive strategy used by aggressive tumors to sustain nutrient supply and promote metastasis, underscoring the need for continued mechanistic investigation.

Pan-cancer analyses have established DCN, a small leucine-rich proteoglycan (SLRP), as a tumor suppressor in LUAD, whereas NPM3 exhibits oncogenic properties in this malignancy. Mechanistically, DCN inhibits tumor growth by suppressing cell proliferation, modulating ECM remodeling, and attenuating PI3K-Akt signaling (Yan et al. 2019). In contrast, NPM3 promotes tumor progression through regulation of ribosome biogenesis, induction of epigenetic modifications, and remodeling of the tumor immune microenvironment (Wei et al. 2023).

The heparan sulfate–modifying sulfatases SULF1 and sulfatase 1 (SULF2) dynamically regulate growth factor signaling by removing 6-O-sulfate groups from heparan sulfate proteoglycans (HSPGs), exerting context-dependent roles in cancer. While SULF2 consistently promotes tumorigenesis (Zallocco et al. 2022), SULF1 displays tissue-specific duality, acting as either a tumor suppressor or an oncogene. Early studies indicated tumor-suppressive functions of SULF1 in ovarian and breast cancer via inhibition of angiogenic signaling (Bret et al. 2011; Khurana et al. 2011). However, growing evidence shows that SULF1 drives malignancy in leukemias, gastrointestinal, and genitourinary cancers by enhancing growth factor signaling, ECM remodeling, and angiogenesis (Hammond et al. 2014). In colorectal cancer, SULF1 promotes VEGFA release and angiogenesis (Wang et al. 2024); in pancreatic cancer, SULF1/2 upregulation stimulates pro-tumorigenic signaling (Gill et al. 2014); and in cervical cancer, SULF1 activates VEGFR2/PI3K/AKT signaling to accelerate progression (Li et al. 2024). Our study aligns with these oncogenic roles, demonstrating significant SULF1 upregulation in LUAD linked to ECM remodeling and VM. Clinically, high SULF1 expression correlates with shorter overall and progression-free survival in LUAD patients (Wang et al. 2023). Functionally, SULF1 overexpression promoted VM by enhancing invasion and migration via VEGF/TGF-β/Vimentin upregulation and accelerated tumor growth in vivo, underscoring its central role in LUAD progression.

Immune profiling revealed significant correlations between the expression of NPM3, DCN, and SULF1 and dynamic shifts in immune infiltration—particularly involving macrophages and regulatory T cells (Tregs)—suggesting their potential role in modulating LUAD progression through immune regulation. TAMs are increasingly recognized as critical promoters of VM through several mechanisms: (1) secreting angiogenic factors (VEGF, TGF-β, fibroblast growth factor (FGF)); (2) releasing MMPs that remodel the ECM; (3) inducing EMT via TGF-β and Interleukin (IL)-6; and (4) establishing an immunosuppressive microenvironment through IL-10 and TGF-β (Mohamed et al. 2014; Luo et al. 2020; Teicher 2021; Treps et al. 2021). Lymphocytes also play context-dependent roles: CD4⁺/CD8⁺ T cells and natural killer (NK) cells generally suppress VM, whereas Tregs facilitate it (Song et al. 2017; Shu et al. 2018; Liu et al. 2025).

Our study reveals that the VM-associated biomarkers DCN, NPM3, and SULF1 may participate in establishing an immune-suppressive niche conducive to VM formation. This finding aligns with the growing recognition of the interplay between tumor vasculature and immune evasion. The conceptual framework is strengthened by recent multi-omics studies that systematically decode tumor microenvironment heterogeneity. For instance, in thyroid cancer, an optimized dynamic network biomarker (DNB) approach identified a critical transition point at Stage II, characterized by a distinct molecular disruption and a tipping point in disease progression (Zhang et al. 2025a, b). This methodology underscores the power of systems biology to identify key regulatory cascades and critical states during tumor evolution, which could be highly relevant for deciphering the dynamic molecular events driving VM initiation. Furthermore, the connection between specific metabolic pathways and immune regulation provides a direct mechanistic link. In LUAD, the Mitochondrial Pathway Signature (MitoPS) was developed as a robust predictor of immunotherapy response, where high MitoPS scores characterized an immune-cold phenotype and poor prognosis. Crucially, this study identified NDUFB10, a

mitochondrial complex I component, as a core gene whose knockdown enhanced T-cell infiltration and cytotoxicity, thereby improving the efficacy of PD-1 blockade (Zhang et al. 2025a, b). This evidence positions mitochondrial function and related pathways as master regulators of the tumor immune microenvironment. Furthermore, consistent with other reports, our study confirms that SULF1 promotes VM by modulating the release of VEGFA and the TGF-β signaling pathway. Wang et al. demonstrated that targeting SULF1 (e.g., using HDAC inhibitor 3) suppresses VM and enhances the response to anti-angiogenic therapy (Wang et al. 2024), supporting its potential as both a therapeutic target and a prognostic biomarker in LUAD. Taken together, the VM-associated biomarkers identified in our work not only signify aggressive tumor behavior but may also indicate the presence of an immune-evasive microenvironment and a novel mechanism of resistance to anti-angiogenesis. These findings provide a more comprehensive perspective on LUAD progression.

While this study established a VM-associated biomarker signature through integrated bioinformatics and experimental validation, several limitations warrant consideration. Primarily, the potential prognostic value of the identified genes warrants further investigation in independent, well-designed cohorts. Then, the spatial correlation between VM structures and biomarker expression in clinical tumor tissues remains to be directly validated using multiplex immunohistochemistry or immunofluorescence co-staining. Additionally, although in vitro assays provided mechanistic insights using cell line models, future work should employ patient-derived xenograft (PDX) models to better recapitulate the tumor microenvironment and verify the in vivo roles of these biomarkers in VM formation. Addressing these aspects would enhance the clinical relevance and translational potential of our findings.

Conclusions

Through integrated bioinformatics and experimental validation, this study identified and characterized a novel set of VM-associated biomarkers—DCN, NPM3, and SULF1—in LUAD. We systematically delineated their functional roles in ECM remodeling, immune modulation, and VM regulatory networks. These findings provide a robust foundation for developing targeted anti-VM therapies, with particular potential to overcome resistance to conventional anti-angiogenic treatment and immunotherapy in LUAD.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

SC and ZZW conceived and supervised this study. SC and YMF performed the bioinformatic analyses, conducted the experiments, and wrote the manuscript. CWT, ZJ, LZK, LYJ, WSQ and ZZX were involved in the study concepts and design. All authors read and approved the final version of the manuscript.

Funding

Not applicable.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and consent to participate

All animal experiments were approved by the Experimental Animal Ethics Committee of Laian Technology (Guangzhou) Co., Ltd (Approval No. G2025032) and conducted in compliance with guidelines for animal research.