Integrated multi-omics analysis of neutrophil extracellular trap-related genes in renal cell carcinoma using bioinformatics and machine learning

Abstract

Background

Renal cell carcinoma (RCC) poses high recurrence/metastasis risk with limited advanced therapy. The role of neutrophil extracellular traps (NETs) in RCC remains unclear. This study aims to identify core NET-related genes and validate their molecular subtyping and prognostic value.

Methods

Five GEO datasets were integrated to identify DEGs, subsequent functional enrichment and WGCNA extracted key modules that were cross-referenced with potential genes systematically compiled from GeneCards and literature review. Three machine learning algorithms (LASSO, SVM-RFE, RF) were refined core genes. ROC analysis validated diagnostic performance, and a nomogram was constructed. Consensus clustering defined molecular subtypes, which were subsequently characterized by immune infiltration, pathway activity, and validated in the independent TCGA-KIRC cohort for prognosis and clinicopathological correlations.

Results

We identified eight core NET-related genes with excellent diagnostic accuracy (AUCs 0.989/0.987). Based on these genes, patients were classified into two molecular subtypes: the C1 subtype exhibited high immune cell infiltration, particularly of activated CD8⁺ T cells and MDSCs, but was associated with poor prognosis and advanced tumor stage, while the C2 subtype showed low immune infiltration, was enriched in metabolic pathways, and correlated with favorable survival outcomes. Drug sensitivity analysis identified Capsaicin as a potential therapeutic agent.

Conclusion

The eight-gene NET signature demonstrates strong diagnostic accuracy and enables molecular subtyping of RCC with validated prognostic significance, highlighting its potential as a prognostic biomarker and therapeutic guide.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04965-7.

Introduction

Renal cell carcinoma is a highly aggressive malignancy with rising global incidence and mortality, posing a significant threat to urological health [1]. Globally in 2022, RCC accounted for 434,419 new cases (2.2% of all cancers) and 155,702 deaths (1.6% of cancer mortality) [2]. Pathologically, RCC is classified mainly into clear cell RCC (ccRCC, accounting for 70–80%), papillary RCC (pRCC, 10–15%), and chromophobe RCC (chRCC, 5–10%), while ccRCC exhibiting the highest malignancy and worst prognosis [3, 4]. Although targeted therapies and immunotherapies have significantly improved survival for some patients, drug resistance and limited survival benefits remain major challenges [5–7]. Therefore, there is an urgent need to explore novel biomarkers and therapeutic targets that can more accurately predict efficacy and guide treatment, in order to improve treatment outcomes for all types of renal cell carcinoma, regardless of histology.

NETs are web-like structures released by activated neutrophils, consisting of a DNA scaffold decorated with various granular proteins [8]. The formation of NETs is a unique form of regulated cell death, distinct from apoptosis, autophagy, or necroptosis. Beyond this primary role in antimicrobial defense, NETs have been increasingly implicated in the pathogenesis of various cancers [9, 10]. NETs can impact the prognosis of various cancers by modulating tumor growth, metastatic dissemination, angiogenesis, and the tumor immune microenvironment [11, 12], with their roles extensively documented in gastric [13], pancreatic [14], hepatocellular [15], and non-small cell lung cancer [16]. However, the role of NETs in RCC remains largely unexplored.

Recently, a growing body of evidence has begun to shed light on this area. A few recent studies have begun to explore the prognostic value of NET-related genes in ccRCC. For instance, Rong Li et al. constructed a prognostic model based on NET-related genes that demonstrated predictive potential for patient survival [17]. Another study developed a NET score that highlighted the critical role of NETs in the ccRCC immune microenvironment [18]. Despite these initial efforts, significant gaps remain. The existing studies predominantly focus on the independent functions of NET-related genes and lack systematic screening and integrative analysis within the broader co-expression networks of RCC. Furthermore, core gene sets identified using a single algorithm may lack robustness and reliability.

In this study, we systematically identified core NET-related genes in RCC by integrating co-expression network analysis with multiple machine learning algorithms. By constructing a molecular subtyping system, it further explores the differences among subtypes in terms of the tumor immune microenvironment, signaling pathways, and drug sensitivity. Ultimately, this work seeks to investigate the characteristics of NETs-related genes in RCC, elucidate their roles in tumor development and prognosis, and identify potential therapeutic targets.

Materials and methods

Acquisition and preprocessing of microarray data

Five RCC gene expression datasets (GSE11151, GSE15641, GSE40435, GSE53757, GSE105288) were obtained from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/).These datasets encompass diverse RCC subtypes, predominantly ccRCC, as well as other histological types. The cohorts include patients across various tumor stages and grades, with detailed clinicopathological characteristics summarized in Table S1. Datasets GSE11151, GSE15641, GSE40435, and GSE105288 were used as the training set, while GSE53757 served as the validation set. Detailed accession numbers are provided in the Data Availability Statement. To address the lack of detailed clinical progression data in GEO, the TCGA-KIRC cohort was used for independent clinical validation. Gene symbols were uniformly re-annotated using Perl (version 5.40.2.1) to generate expression matrices. The datasets were then merged, batch-corrected using the “sva” package, and normalized with “limma”.

Screening of DEGs

The “limma” and “sva” packages in R software were used to merge the training set data and remove batch effects. Boxplots and principal component analysis (PCA) plots were generated before and after batch effect correction using the “ggplot2” and “ggpubr” packages. DEGs between RCC and normal tissues were identified using thresholds of |log₂FC| > 1 and FDR < 0.05.

Functional enrichment analysis

Gene Ontology (GO) functional annotation and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of the DEGs were performed using the “clusterProfiler” package, with an adjusted P-value < 0.05 considered statistically significant. Visualization was performed using the “circlize” and “ggplot2” packages. Specifically, the top six enriched terms from each GO category—biological process (BP), cellular component (CC), and molecular function (MF)—were displayed in a circos plot, and the top 30 enriched KEGG pathways were presented in a bubble plot.

Construction of WGCNA and identification of key module genes

The “WGCNA” package in R software was used to construct a weighted gene co-expression network. The optimal soft-thresholding power was determined using the pick soft threshold function. Based on this power value, the weighted adjacency matrix and topological overlap matrix (TOM) were subsequently calculated. Hierarchical clustering with the average-linkage method was performed using 1-TOM as the distance measure. Gene modules were identified using a dynamic tree-cutting algorithm, with a minimum module size of 50, deep split sensitivity of 2, and a module merge cut height of 0.3. Module eigengenes (MEs) were then extracted, and their correlations with clinical traits were visualized in a heatmap. Key modules strongly associated with RCC were selected based on high module significance (MS) and gene significance (GS).

Screening of NET-related candidate genes

Protein-coding genes associated with the term “Neutrophil extracellular trap” were retrieved from the GeneCards database (https://www.genecards.org), applying a relevance score threshold of ≥ 10 [19], which yielded 566 genes. These were combined with 69 NET-related genes obtained from published literature [20–22]. After removing 37 duplicates, a final set of 598 NET-related genes was established. The “ggvenn” package in R software was used to identify the common genes among the DEGs, key module genes from WGCNA, and the NET-related genes, resulting in a list of NET-related candidate genes. These candidate genes were then subjected to multi-dimensional validation, which included expression analysis via boxplots (“ggpubr” package), chromosomal distribution visualization using a circos plot (“circlize” package), generation of a manhattan plot (“CMplot” package), and co-expression network analysis (“corrplot” package).

Immune cell infiltration analysis

The infiltration levels of 28 immune cell types were assessed via single-sample gene set enrichment analysis (ssGSEA), including B cells, T cell subsets (CD4⁺, CD8⁺, regulatory, helper), natural killer cells, dendritic cells, macrophages, monocytes, neutrophils, mast cells, eosinophils, and myeloid-derived suppressor cells (MDSCs). The wilcoxon rank-sum test was used to compare immune cell abundance between RCC and normal tissues. Violin plots, generated with the “vioplot” package, were used to visualize the results, with significance levels denoted as *P < 0.05, **P < 0.01, and ***P < 0.001. A heatmap illustrating the immune cell infiltration matrix was constructed using the “pheatmap” package. Furthermore, Spearman rank correlation analysis was used to quantify the association between core gene expression levels and immune scores. Significant correlations were visualized in a stratified heatmap generated with “ggplot2”.

Screening and validation of core genes

Three machine learning algorithms were employed to screen for core genes: RF was implemented using the “random Forest” package, where the optimal number of decision trees was determined and genes with a Mean Decrease Gini ≥ 5 were selected; LASSO logistic regression was performed with the “glmnet” package (α = 1), using ten-fold cross-validation to determine the optimal penalty parameter (λ.min) and retaining genes with non-zero coefficients; and SVM-RFE was conducted using the “e1071” package, in which ten-fold cross-validation was applied to evaluate feature subsets and the gene subset with the lowest cross-validation error was selected. Genes commonly identified by all three algorithms were defined as core genes. Subsequently, ROC curves for each core gene and a multi-gene logistic regression model were plotted in the training set using the “pROC” package. A nomogram, calibration curve, and decision curve were constructed to evaluate model performance. Finally, the stability of the core genes was confirmed by ROC validation in the independent validation set (GSE53757).

Consensus clustering analysis

Based on the GEO RCC dataset, the normalized expression data were preprocessed to extract a core gene expression matrix. Consensus clustering was performed using the “Consensus Cluster Plus” package (parameters: max K = 9, reps = 50, p Item = 0.8), applying the k-means algorithm with euclidean distance. The optimal number of clusters was determined by evaluating the cumulative distribution function (CDF) plot and consensus matrix heatmap. A heatmap of feature gene expression was generated, and PCA was used to visually validate the cluster segregation.

Analysis of immune microenvironment and drug sensitivity

The CIBERSORT algorithm was used to quantify immune cell infiltration levels for across subtypes, with inter-subtype differences assessed by the Kruskal–Wallis test and visualized accordingly. Gene Set Variation Analysis (GSVA) was used to assess differences in KEGG pathway enrichment between subtypes, and significantly different pathways were identified by t-test. Drug enrichment analysis was performed using the “cluster Profiler” package based on the DSigDB database (https://dsigdb.tanlab.org/), and significantly enriched drugs (P < 0.05, and FDR < 0.05) were identified. Drug enrichment bar plots and bubble plots were generated with “ggplot2”, and a drug-gene interaction network was constructed and visualized.

Independent validation using TCGA cohorts

To validate the prognostic and clinicopathological significance of the NET-related molecular subtypes, we obtained RNA-seq data and corresponding clinical annotations from the TCGA-KIRC cohort (n = 530) via the UCSC Xena platform (https://xenabrowser.net). Patients were classified into C1 and C2 subtypes using the same consensus clustering algorithm and eight-gene signature applied to the GEO cohort. Associations between subtypes and clinicopathological parameters were assessed using Spearman’s rank correlation. Survival analysis was performed using Kaplan–Meier curves with the log-rank test, and univariate and multivariate Cox proportional hazards models were used to evaluate the independent prognostic value of the subtypes. GSVA was again employed to compare pathway enrichment between subtypes in the TCGA cohort.

Statistical analysis

All statistical analyses were performed using R (version 4.5.1) software, with P < 0.05 considered statistically significant. For normally distributed data, the t-test was applied, whereas the Wilcoxon rank-sum test was used for non-normally distributed data. Comparisons among multiple groups used the Kruskal-Wallis test, and correlations were analyzed using Spearman’s rank correlation.

Results

Screening of DEGs and functional enrichment analysis

PCA of the normalized training set data showed high reproducibility among samples (Fig. 1A). A comparative analysis of 267 tumor and 138 normal samples from the database identified 604 DEGs, including 358 upregulated and 246 downregulated genes (Fig. 1B). A heatmap was used to visualize the top 50 DEGs ranked by P-value (Fig. 1C). GO enrichment analysis (Fig. 1D) revealed that BP terms were primarily associated with renal system development and kidney development; CC terms were mainly localized to the apical part of the cell, apical plasma membrane, and collagen-containing extracellular matrix; and MF terms were enriched for vitamin binding, peptide antigen binding, and carboxylic acid binding. KEGG pathway analysis (Fig. 1E) indicated that DEGs were mainly involved in Phagosome, HIF-1 signaling pathway, and PPAR signaling pathway.

Fig. 1.

Transcriptomic differential landscape and functional enrichment of RCC. A PCA plot of tumor and normal tissues. B Volcano plot of DEGs. Red indicates significantly upregulated genes, blue indicates significantly downregulated genes (threshold: |log2FC| > 1, FDR < 0.05). C Clustered heatmap of DEGs. D Bar chart of GO functional enrichment analysis for DEGs. E Bubble plot of KEGG pathway enrichment analysis for DEGs

Identification of NET-related candidate genes based on WGCNA

Following topological analysis of the sample network, a soft-thresholding power of 11 was selected according to the scale-free topology criterion to construct the weighted gene co-expression network (Fig. 2A). The constructed WGCNA network identified a total of seven distinct non-grey modules, each representing a set of highly co-expressed genes (Fig. 2B). Module-trait correlation analysis showed that seven of these modules—MEturquoise, MEred, MEblue, MEblack, MEgreen, MEmagenta, and MEyellow—exhibited significant correlations with RCC status (|Cor| > 0.5, P < 0.05). Notably, the MEturquoise module displayed the strongest positive correlation with RCC and the most statistically significant association (Fig. 2C). From this most relevant module, we extracted 1167 genes that showed strong associations with RCC pathogenesis. To further refine our candidate genes, we integrated multiple filtering approaches by taking the intersection of DEGs, WGCNA key module genes, and NET-related genes, yielding 27 high-confidence NET-related candidate genes (Fig. 2D). Chromosomal mapping of candidate genes revealed their non-random genomic distribution, with predominant localization on chromosomes 1, 3, 4, and 5 (Fig. 2E). Further visualization using a Manhattan plot highlighted chromosome 3 as containing multiple genomic regions with significant association signals, among which the KNG1 gene emerged as the most prominent locus (Fig. 2F).

Fig. 2.

Screening of NET-related candidate genes. A Selection of soft-thresholding power in WGCNA: scale-free topology fit (left) and mean connectivity (right). B Cluster dendrogram of co-expression modules. C Module–trait relationship heatmap. D Venn diagram showing DEGs, NETs-related gene sets, and key WGCNA module genes. Intersecting genes were defined as candidate genes. E Chromosomal distribution of candidate genes. F Manhattan plot of candidate genes across chromosomes. The X-axis and Y-axis represent chromosomes and -log₁₀ (P values) of genes, respectively

Candidate gene characteristics and their immune microenvironmental correlations

Analysis of the 27 candidate genes revealed significant differential expression between normal and RCC tissues (Fig. 3A). Multiple genes, including KNG1, CASR, and PLCG2, were significantly downregulated in tumor tissues, while LGALS1, VIM, and MYC were markedly upregulated. Tumor immune infiltration analysis indicated substantial alterations in the immune cell composition within the RCC microenvironment, with significant increases in both immunosuppressive cells—including regulatory T cells (Treg), MDSCs, and macrophages—and effector immune cells such as activated CD8⁺ T cells and natural killer cells, suggesting a state of high immune infiltration yet functional imbalance in the tumor microenvironment (Fig. 3B). Further correlation analysis demonstrated that upregulated candidate genes (e.g., LGALS1) exhibited positive correlations with immunosuppressive cell (e.g., Treg cells and MDSCs) infiltration, whereas downregulated genes (e.g., KNG1, CASR) showed positive associations with activated immune cell infiltration (Fig. 3C). These findings suggest that these genes may participate in RCC progression by bidirectionally regulating the immune microenvironment.

Fig. 3.

Immune infiltration-associated genomic features of candidate genes in RCC. A Comparison of candidate gene expression between tumor and normal tissues from the GEO-RCC cohort (*P < 0.05, **P < 0.01, ***P < 0.001). B Violin plots displaying the comparison of immune cell infiltration levels between tumor and normal tissues. C Heatmap of correlations between candidate gene expression and immune cell infiltration (red, positive correlation; blue, negative correlation)

Screening of core genes using machine learning

The three machine learning algorithms generated distinct sets of candidate genes. LASSO regression determined 17 key genes based on the optimal penalty coefficient derived from cross-validation (Fig. 4A). The random forest model selected 10 core feature genes according to a threshold of Mean Decrease in Gini greater than 5 (Fig. 4B). Meanwhile, the SVM-RFE algorithm identified an optimal feature set comprising 22 genes (Fig. 4C). An intersection of the results from all three algorithms ultimately identified 8 NET-related core genes: KNG1, CASR, DNASE1L3, LGALS1, PLAT, DNASE1, PDGFRA, and PLCG2 (Fig. 4D).

Fig. 4.

Identification of core gene selection using machine learning approaches. A LASSO regression analysis showing cross-validation curve with optimal lambda (left) and variable coefficient paths (right). B SVM-RFE feature selection with accuracy curve (left) and error curve highlighting optimal feature number (right). C Random forest analysis displaying error trend (left) and top 30 important genes (right). D Venn diagram illustrating overlapping hub genes from three machine learning methods (LASSO, SVM-RFE, and RF)

A diagnostic and prognostic model based on core genes in RCC

The Wilcoxon test analysis showed that the expression levels of all eight core genes (KNG1, CASR, DNASE1L3, LGALS1, PLAT, DNASE1, PDGFRA, and PLCG2) were significantly different between the normal and RCC groups (all P < 0.05). These genes exhibited consistent expression patterns in both the training and validation set (GSE53757) (Fig. 5A, B). ROC analysis further showed that each gene possessed excellent diagnostic performance (AUC > 0.85), with KNG1 showing the highest diagnostic value (AUC = 0.968 and 0.975). Subsequently, multi-gene logistic regression prediction models were constructed in the training and validation sets, achieving AUCs of 0.989 (95% CI: 0.978–0.997) and 0.987 (95% CI: 0.969–1.000), respectively, indicating strong disease discrimination ability. Based on the above prediction models, calibration curves and decision curve analysis, as shown in Fig. 5C, further confirmed the model’s favorable clinical applicability and predictive accuracy. The nomogram illustrated that the eight genes collectively constitute a risk-scoring system, in which CASR, LGALS1, PDGFRA, and KNG1 carried higher weights. This observation aligns with their stronger discriminative ability observed in univariate ROC analysis, highlighting their central role in the model.

Fig. 5.

A core gene-based model for diagnosis and prognosis prediction. A Training and (B) validation set performance of core genes. Upper sections: Box plots depict significant expression differences between control and treatment groups (***P < 0.001, **P < 0.01, *P < 0.05). Middle sections: ROC curves for each gene. Lower sections: ROC curve of the logistic regression model. C Comprehensive evaluation of the prediction model, the calibration plot (upper), the decision curve (middle), and the predictive nomogram (lower)

Molecular subtyping based on core genes and therapeutic implications

In the integrated RCC cohort, we performed consensus clustering analysis based on the eight NET-related core genes. Evaluation via the consensus heatmap and cumulative distribution function (CDF) values determined the optimal cluster number to be 2 (k = 2), thereby stratifying cancer patients into two molecular subtypes (C1 and C2) with significant differences. The consensus matrix at this point depicted the highest uniformity (Fig. 6A, B). PCA analysis further confirmed a distinct separation between the two subtypes based on the overall expression profile of cancer patients (Fig. 6C), with most core genes showing differential expression between subtypes (Fig. 6D).

Fig. 6.

Molecular subtypes of NET based on consensus clustering and its clinical correlations. A Consensus matrix heatmap at k = 2, revealing two stable molecular subtypes (C1 and C2). B Consensus cumulative distribution function analysis for k = 2–9. C PCA plot showing clear separation between subtypes C1 and C2 in gene expression space. D Heatmap of patient clustering based on differential gene expression. Red indicates high expression; blue indicates low expression. E Boxplot comparing immune cell infiltration levels between the two subtypes (*P < 0.05, **P < 0.01, ***P < 0.001). F The top 20 KEGG pathways with differential activity between subtypes identified by GSVA. G Potential therapeutic drugs predicted by drug enrichment analysis of key genes using the DSigDB database. H Interaction network between key genes and predicted drugs. Circular nodes represent genes; rectangular nodes represent drugs

Given the critical role of the tumor immune microenvironment, we systematically compared the immune infiltration patterns between subtypes. As shown in Fig. 6E, the C1 subtype exhibited an “immune-inflamed” phenotype with significantly higher infiltration of most immune cells, including effector cells (activated CD8⁺ and CD4⁺ T cells, NK cells), and immunosuppressive cells (Tregs, MDSCs). In contrast, the C2 subtype displayed an “immune-desert” phenotype with markedly lower infiltration levels of most immune cell types, including effector cells, immunosuppressive lineages, and antigen-presenting cells.

GSVA analysis revealed significant differences in the enrichment scores of multiple KEGG pathways between the C1 and C2 subgroups (Fig. 6F). The C1 subgroup featured high enrichment of immune pathways (e.g., cytokine-cytokine receptor interaction, leukocyte transendothelial migration, primary immunodeficiency), whereas the C2 subgroup showed high enrichment of metabolic pathways (e.g., aldosterone-regulated sodium reabsorption, fatty acid metabolism).

Concurrently, we evaluated the predictive value of the eight core genes for drug sensitivity. As shown in Fig. 6G, these genes were primarily associated with sensitivity to several agents, such as Capsaicin. Based on these findings, a drug-gene interaction network was constructed (Fig. 6H) to highlight the therapeutic potential of these core genes.

Independent clinical validation of the NET-related molecular subtypes in TCGA cohort

Consensus clustering analysis of the TCGA-KIRC cohort (n = 530) yielded a stable two-cluster solution (k = 2; Fig. 7A), in which most core genes were differentially expressed between the two subtypes (Fig. 7B), closely resembling the C1/C2 classification established in the GEO cohort. Immune infiltration analysis revealed that the C1 subtype exhibited a pronounced “immune-inflamed” phenotype, characterized by significantly higher infiltration levels of activated CD8⁺ T cells, activated CD4⁺ T cells, γδ T cells, and natural killer cells; in contrast, the C2 subtype displayed a typical “immune-desert” phenotype, with markedly lower infiltration levels of these immune cell populations (Fig. 7C). This immune cell distribution pattern was fully consistent with the trends observed in the GEO training cohort. GSVA analysis demonstrated that, compared with the C1 subtype, the C2 subtype showed significant downregulation of immune-related pathways, including primary immunodeficiency and intestinal immune network for IgA production, accompanied by concurrent upregulation of multiple metabolic pathways such as aldosterone-regulated sodium reabsorption, fatty acid metabolism, and propanoate metabolism (Fig. 7D). This pathway alteration profile was entirely concordant with the GEO training cohort, revealing a stable “metabolic activation, immune suppression” molecular characteristic of the C2 subtype across independent datasets. Kaplan–Meier survival curves showed that the C2 subtype had significantly better overall survival than those in the C1 subtype (Fig. 7E). Clinicopathological correlation analysis demonstrated that the C2 subtype was significantly enriched in patients with lower T stage (T1/T2) and lower tumor stage (stage1/ stage2), whereas the C1 subtype was predominantly associated with advanced T stage and tumor stage (Fig. 7F, G). Both univariate and multivariate Cox regression analyses confirmed that the C2 subtype was an independent protective factor for overall survival benefit, while the C1 subtype represented an independent risk factor for poor overall survival (Fig. 7H, I).

Fig. 7.

Independent validation of NET-related molecular subtypes in TCGA-KIRC cohort. A Consensus matrix heatmap at k = 2, revealing two stable molecular subtypes. B Heatmap of patient clustering based on differential gene expression. C Boxplot comparing immune cell infiltration levels between the C1 and C2 subtypes (*P < 0.05, **P < 0.01, ***P < 0.001). D GSVA showing differentially regulated KEGG pathways between the two subtypes. E Kaplan–Meier curves for overall survival stratified by subtype. F Distribution of pathological T stage in C1 versus C2 subtypes. G Distribution of tumor stage between the two subtypes. H Forest plot of Univariate Cox regression analysis. I Forest plot of multivariate Cox regression analysis

Discussion

Renal cell carcinoma, a highly aggressive malignancy of the urinary system, poses a growing global health burden [23]. Traditional discovery methods, often restricted to single-omics analyses, are frequently inadequate for pinpointing core functional drivers within complex gene networks [24]. Machine learning algorithms, with their capability to handle high-dimensional data, have emerged as vital tools for identifying key disease-associated genes [25, 26]. Although NETs are implicated in tumor progression [27], their gene regulatory mechanisms in RCC are not fully understood. To address this, we employed an integrative bioinformatics strategy, combining WGCNA with three machine learning algorithms, to systematically dissect NET-related genes in RCC and identify novel therapeutic targets.

Our multi-step screening identified eight core NET-related genes (KNG1, CASR, DNASE1L3, LGALS1, PLAT, DNASE1, PDGFRA, and PLCG2) as pivotal regulators of the tumor immune microenvironment in RCC. KNG1, involved in coagulation and inflammation, was downregulated in RCC and found to modulate immune infiltration [28, 29]. CASR, a NETosis factor, has a complex role: low expression correlates with poor prognosis, though some link it to bone metastasis [30–32]. DNASE1L3 and DNASE1, critical in autoimmunity and cancer, are implicated in malignancies like hepatocellular carcinoma [33, 34]. Other genes, including PLAT and PLCG2, are involved in angiogenesis [35, 36], and PDGFRA contributes to stromal remodeling [37]. Notably, LGALS1, a well-characterized immunosuppressive molecule that promotes Tregs and inhibits effector T cells [38, 39], was highly expressed in RCC and positively correlated with Treg and MDSC infiltration. MDSCs further impair T-cell function via checkpoint molecules, amino acid depletion, and reactive species, synergizing with Tregs to sustain an immunosuppressive microenvironment [40, 41].

We classified RCC patients into two subtypes (C1 and C2) via consensus clustering of eight NET-related genes, revealing distinct immune microenvironments, pathway activities, and prognoses. The C1 subtype was associated with poor prognosis and positively correlated with tumor stage. This subtype featured high infiltration of immune cells, including activated CD8⁺ T cells and MDSCs, along with significant enrichment of the primary immunodeficiency pathway. These findings indicate that high immune infiltration in the C1 subtype occurs within an immunosuppressive microenvironment, contributing to poor prognosis. This phenomenon is highly consistent with the “CD8 + Inflamed” renal cancer subtype by Clark et al., where high CD8⁺ T cell infiltration with T cell exhaustion drives immune suppression and poor outcomes [42]. Similarly, Senbabaoglu et al. reported that the ccRCC subgroup with high T-cell infiltration exhibited worse outcomes and was enriched for exhausted T-cell signatures [43]. Furthermore, single-cell transcriptomic studies by Braun et al. have elucidated the underlying mechanism, demonstrating that terminally exhausted CD8⁺ T cells (co-expressing PD-1, TIM-3, and TOX) are enriched and dominate the immune infiltrate in locally advanced and metastatic ccRCC, strongly correlating with adverse prognosis [44]. Despite the abundance of activated CD8⁺ T cells, these cells are functionally impaired and fail to mount an effective anti-tumor immune response. Consequently, the eight-gene signature can predict poor prognosis, and patients within C1 subtype may be suitable candidates for immunotherapy.

In contrast to the C1 subtype, the C2 subtype associated with good prognosis and negative correlation with tumor stage, showed relatively low immune cell infiltration and high enrichment of metabolic pathways like aldosterone-regulated sodium reabsorption and fatty acid metabolism. Similarly, Clark et al. defined the “metabolic immune-desert” subtype in renal cancer, marked by low immune/stromal scoresand oxidative phosphorylation activation, correlating with favorable prognosis [42]. This finding is also consistent with the ccA subtype defined by Brannon et al. and Brooks et al., which is characterized by angiogenesis and fatty acid metabolism signatures and associated with longer survival [45, 46]. Shahzad et al. [11]. demonstrated that RCC cells originating from renal tubular epithelium up regulate solute transport pathways (e.g., aldosterone-regulated sodium reabsorption) and undergo LDHA-mediated aerobic glycolysis, leading to lactate accumulation and extracellular acidification that inhibits T/NK cells, ultimately fostering an immune-desert tumor microenvironment [47, 48]. Wu G et al. demonstrated that dysregulated fatty acid and cholesterol metabolism drives ccRCC progression, and that LXR modulators inhibit tumor growth by reducing intracellular lipid content [49]. Consequently, patients within the C2 subtype may be more sensitive to targeted therapies and metabolic inhibitors, while exhibiting a poorer response to immunotherapy.

Drug prediction analysis identified Capsaicin, which was significantly enriched via KNG1, LGALS1, and PLAT. This finding aligns with the established multi-target characteristics of Capsaicin in RCC. In RCC models, Capsaicin induces apoptosis by activating the MAPK pathway [50] and modulates multiple cancer-related processes, including p53-mediated apoptosis, calcium signaling, AMPK/mTOR signaling and TGF-β-associated immunosuppression [51, 52]. More recently, Capsaicin has also been shown to exert immunomodulatory effects in RCC cells by downregulating PD-L1 expression through an efficient DNA damage repair response and protein carbonylation [53]. This context-dependent regulation further supports the multi-target characteristics of Capsaicin. Notably, although KNG1 and PLAT are downregulated in RCC, Capsaicin was enriched through them, suggesting its potential to modulate this NET-related gene network. This observation implies a potential role for KNG1, LGALS1, and PLAT in mediating the anti-tumor effects of Capsaicin.

Nevertheless, this study has several limitations. First, integrating multiple GEO cohorts introduced heterogeneity, which, despite batch-effect correction, may affect the precision of our findings. Second, a key limitation is the lack of detailed clinical progression data (e.g., stage, grade) in these cohorts, restricting direct assessment of the signature’s link to tumor progression. However, validation using the well-annotated TCGA-KIRC cohort supported its relevance to disease progression. Third, conclusions rely on public bioinformatics data and lack experimental validation. Fourth, findings await confirmation in an independent, prospective clinical cohort. Finally, the molecular mechanisms linking these core genes to NET formation remain unclear. Future studies should include functional experiments and large-scale prospective validation.

Conclusion

In this study, we identified an eight-gene NET-related signature with strong diagnostic performance in RCC. Based on these genes, patients were classified into two molecular subtypes. The C1 subtype exhibited high immune infiltration but poor prognosis, whereas the C2 subtype showed low immune infiltration, enrichment of metabolic pathways, and favorable survival outcomes. This classification was validated in the independent TCGA-KIRC cohort. Drug sensitivity analysis identified Capsaicin as a potential therapeutic agent. Collectively, this signature provides a promising biomarker for molecular subtyping, and personalized treatment guidance in RCC.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

C.L. and J.L. conceptualized and designed the study; C.L., T.S., Y.Y., and Z.Y. analyzed the data; J.L. and Q.Z. performed validation; C.L., T.S., and Y.Y. wrote the original draft; J.L. and Q.Z. reviewed and edited the manuscript.

Funding

This work was supported by the Dalian Municipal Health Commission under Grant [number 23Z11001] and the Education Project of Dalian Municipality under Grant [number DLUXK-2025-QN-010].

Data availability

The original datasets analyzed in this study are publicly available in the GEO database ( [https://www.ncbi.nlm.nih.gov/geo/](https:/www.ncbi.nlm.nih.gov/geo) ). Direct, persistent links to each dataset are as follows: GSE11151( [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE11151](https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE11151) ), GSE15641( [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE15641](https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE15641) ), GSE40435( [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE40435](https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE40435) ), GSE53757( [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE53757](https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE53757) ), GSE105288( [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE105288](https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE105288) ).All data generated or analyzed in this study are included in the article, and any additional relevant data can be provided by the corresponding author upon justified request.

Declarations

Ethics approval

Ethical approval was not required as this study used publicly available data.

Consent to participate

Patient consent was not applicable due to the use of de-identified data from public databases.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.