Machine learning identifies a NETs-related four-gene diagnostic signature for abdominal aortic aneurysm

Abstract

Abdominal aortic aneurysm (AAA) is a chronic degenerative disease characterized by localized aortic dilation and persistent inflammation. While neutrophil extracellular traps (NETs) are increasingly recognized as key drivers of vascular inflammation and aneurysm progression, the transcriptomic landscape of NETs-related genes (NRGs) in AAA remains inadequately characterized. This study aimed to identify reliable diagnostic biomarkers and explore the immune heterogeneity of AAA to facilitate early risk stratification. We integrated transcriptomic datasets from the Gene Expression Omnibus to elucidate the role of dysregulated NRGs. A comprehensive bioinformatics pipeline was employed, combining weighted gene co-expression network analysis with differential expression profiling to screen for AAA-specific NRGs. Rigorous feature selection was conducted through the intersection of 3 machine learning algorithms – least absolute shrinkage and selection operator, support vector machine-recursive feature elimination, and random forest – to derive a diagnostic signature. The model was constructed using the GSE232911 cohort and validated in an independent external cohort. A robust 4-gene diagnostic signature comprising CXCR4, GZMB, ITGA6, and CD47 was identified. This signature demonstrated a favorable diagnostic performance, achieving an area under the curve of 0.920 in the training cohort. The model maintained consistent discriminatory ability in the external validation cohort, primarily driven by the high discriminative ability of CXCR4 and GZMB. Consensus clustering based on these hub genes revealed 2 distinct molecular subtypes, with Cluster 2 characterized by significant enrichment of neutrophils and innate immune pathways, suggesting intense NETosis activity. Furthermore, drug prediction analyses identified candidate therapeutic compounds, including Eugenol and Tretinoin, offering potential avenues for targeting the NETs-associated molecular landscape. Our findings underscore the pivotal role of NETs-mediated inflammation in AAA pathogenesis and validate a robust 4-gene signature for early diagnosis. By delineating immune-related molecular subtypes and identifying potential drug candidates, this study provides a foundational framework for precision risk stratification and the development of targeted nonsurgical therapies for aneurysm management.

1. Introduction

Abdominal aortic aneurysm (AAA) is a chronic, often asymptomatic, degenerative disease characterized by localized aortic dilation (>30 mm or >50% of normal diameter), carrying a mortality rate exceeding 80% upon rupture.^[1–3] With surgical interventions (such as endovascular aneurysm repair) reserved for aneurysms larger than 5.5 cm and standard imaging limited to assessing anatomical size, effective treatments and early risk stratification tools for small- to medium-sized aneurysms remain severely limited.^[4,5] Consequently, elucidating the molecular drivers of AAA and identifying reliable diagnostic biomarkers are essential for advancing nonsurgical therapeutic approaches.

AAA pathogenesis is driven by a multifactorial process involving persistent inflammation, extracellular matrix (ECM) degradation, and vascular smooth muscle cell (VSMC) apoptosis.^[6] Neutrophils are early participants in this vascular inflammation, significantly contributing through the release of neutrophil extracellular traps (NETs) – web-like chromatin structures decorated with granular proteins such as myeloperoxidase, neutrophil elastase, and citrullinated histone H3.^[7,8] While physiologically defensive, dysregulated NETosis aggravates sterile inflammation in cardiovascular diseases.^[9] In human AAA, elevated levels of NETs-related markers stimulate macrophage-mediated cytokine release (such as interleukin-1β) and enhance matrix metalloproteinase activity, creating a feed-forward cycle of tissue injury.^[10,11] Nevertheless, the NETs-associated transcriptomic features of AAA and their relevance as diagnostic biomarkers or therapeutic targets have not yet been systematically characterized.

To overcome the limitations of conventional differential expression analyses in capturing complex gene–gene interactions, integrating transcriptomic data with machine learning (ML) provides a robust alternative. By discerning nonlinear hierarchical patterns, ML architectures already deliver recognized analytical advantages across cardiovascular medicine. These models facilitate critical clinical predictions – such as short-term mortality in pulmonary embolism,^[12] obstructive coronary disease identification,^[13] perioperative myocardial injury surveillance,^[14] and arrhythmic risk stratification^[15] – with higher precision than traditional scoring systems. Leveraging these established computational principles, specific high-dimensional algorithms, including least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and random forest (RF), are highly effective for transcriptomic feature selection.^[16] Consequently, intersecting multiple ML strategies mitigates algorithm-specific biases, ensuring the stability and reliability of the derived AAA diagnostic signature.

Therefore, we hypothesized that dysregulated NETs-related genes (NRGs) intricately modulate the immune microenvironment in AAA, and that capturing these transcriptomic changes could yield a highly accurate diagnostic biomarker signature. The primary objective of this study was to systematically identify and validate a NETs-related diagnostic model using weighted gene co-expression network analysis (WGCNA) integrated with 3 ML algorithms (LASSO, SVM-RFE, and RF). Secondary objectives included elucidating the immune heterogeneity of NETs-associated molecular subtypes and utilizing drug–gene interaction networks to predict potential nonsurgical therapeutic compounds.

2. Materials and methods

2.1. Data acquisition and preprocessing

Because this study analyzed publicly available, de-identified datasets from the Gene Expression Omnibus, institutional review board approval and informed consent were waived.

To reflect the localized aneurysmal microenvironment, dataset inclusion was strictly limited to primary human abdominal aortic tissues. GSE232911 (platform: GPL17586) served as the training cohort. Its large sample size (n = 246; tissues from 76 AAA patients and 13 donor controls) and detailed spatial metadata (tunica media and adventitia) provide adequate statistical power for ML feature selection. The independent GSE7084 dataset served as the external validation cohort. To prevent sample duplication within this series, we exclusively analyzed data generated by the GPL2507 platform, yielding a final sub-cohort of n = 15 samples (7 primary AAA cases and 8 autopsy-derived normal controls), and completely omitting overlapping samples from the GPL570 platform.

Raw expression data were background-corrected via the robust multi-array average algorithm. To further mitigate technical variance, between-array normalization was applied using the normalizeBetweenArrays function in the limma R package.^[17] The complete elimination of technical variance was subsequently verified via principal component analysis (PCA) using the prcomp function, with 95% confidence ellipses utilized to confirm the removal of non-biological clustering prior to downstream analysis. Probes were mapped to official gene symbols, averaging expression values when multiple probes corresponded to a single gene. Finally, a target list of NRGs was curated from the GeneCards database and prior literature.^[18,19]

2.2. Identification of differentially expressed genes

Differential expression analysis between AAA and normal control samples was conducted utilizing the limma R package.^[17] To rigorously control for false positives, P values were adjusted using the false discovery rate method. Genes meeting strict statistical thresholds of false discovery rate-adjusted P < .05 and an absolute log₂ fold change (|logFC|) > 0.585 were considered differentially expressed genes (DEGs). Differential expression patterns were visualized via volcano plots (ggplot2) and hierarchical clustering heatmaps (pheatmap).

2.3. Functional enrichment analysis

Gene Ontology (GO) functional annotation and Kyoto Encyclopedia of Genes and Genomes pathway enrichment analyses were performed on the identified DEGs using the clusterProfiler R package.^[20] GO analysis included biological process, cellular component, and molecular function categories. Enrichment results with adjusted P values <.05 were considered statistically significant.

2.4. WGCNA and identification of intersection genes

A gene co-expression network was constructed using the WGCNA R package after filtering low-variance genes (SD ≤ 0.5) and outlier samples.^[21] The optimal soft-thresholding power (β) was automatically selected via the pickSoftThreshold function to ensure a scale-free topology. An adjacency matrix was generated and converted into a topological overlap matrix. Modules were identified using the dynamic tree cut algorithm (minimum size = 50 genes), and highly similar modules were merged (eigengene dissimilarity threshold = 0.3). Finally, Pearson correlation between module eigengenes and clinical traits identified the most AAA-associated module.

2.5. Characterization of AAA-specific NETs-related genes

The chromosomal distribution of the identified AAA-specific NRGs was visualized using the RCircos R package.^[22] A Manhattan plot illustrating the statistical significance of differential expression was generated using the CMplot package.^[23] Gene expression levels were compared between AAA and control samples using boxplots. Pairwise gene–gene correlations were calculated using Pearson correlation analysis and visualized as a heatmap with the corrplot R package.^[24] To assess immunological relevance, correlations between gene expression levels and immune cell infiltration scores, quantified by single-sample gene set enrichment analysis (ssGSEA), were evaluated using Spearman correlation analysis and visualized using heatmaps and scatter plots.

2.6. Screening of diagnostic biomarkers via machine learning

To establish robust feature selection thresholds and prevent overfitting, 3 independent ML algorithms were deployed. First, LASSO regression was executed via the glmnet package,^[25] utilizing 10-fold cross-validation to penalize variables based on binomial deviance, extracting features exclusively at the minimal penalty parameter (λ_min). Second, SVM-RFE was performed using the e1071 package adopting a linear kernel (C = 10). A 10-fold cross-validation strategy was employed to recursively eliminate features and rank the optimal diagnostic subset minimizing classification error. Finally, RF array training was conducted via the randomForest package,^[26] initialized with 500 trees. The optimal tree count minimizing the out-of-bag error rate was automatically selected, and the top 30 essential genes were ranked utilizing the Mean Decrease Gini index. The final diagnostic hub genes were derived exclusively from the intersection of these 3 distinct algorithmic panels.

2.7. Construction and validation of the diagnostic model

A diagnostic nomogram incorporating the hub genes was constructed using the rms R package.^[27] Model discrimination was evaluated using receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC) was calculated for both the training cohort (GSE232911) and the validation cohort (GSE7084).^[28] Calibration curves were generated to assess agreement between predicted and observed outcomes, and decision curve analysis was performed to evaluate the clinical utility of the model.^[29]

2.8. Identification of molecular subtypes and immune infiltration analysis

Consensus clustering was executed via the ConsensusClusterPlus package to define novel AAA molecular subtypes based on hub gene expression.^[30] The algorithm utilized k-means clustering coupled with Euclidean distance evaluation. Clustering stability was validated evaluating up to k = 9 clusters, utilizing 50 iterations with an 80% target resampling rate. Subtype separation was mapped via PCA. Pathway activation variations between distinct subtypes were evaluated via gene set variation analysis.^[31] Furthermore, immune cell infiltration abundance was robustly quantified utilizing the ssGSEA algorithm.^[32] Infiltration differences between subtypes were assessed via the Wilcoxon rank-sum test, while molecular correlations between hub genes and specific immune subsets were evaluated using Spearman correlation analysis.

2.9. Drug prediction analysis

Potential therapeutic compounds targeting the NRG signature were explored using the Drug Signatures Database (DSigDB)^[33] via the clusterProfiler enricher function . Compounds showing significant interaction or enrichment scores were considered candidate drugs for further investigation.

2.10. Statistical analysis

All statistical analyses were conducted using R software (version 4.4.2; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were compared using Student t test or the Wilcoxon rank-sum test, as appropriate. Correlation analyses were performed using Pearson or Spearman correlation coefficients. A two-sided P value <.05 was considered statistically significant.

3. Results

3.1. Identification and functional enrichment analysis of DEGs

The overall analytical workflow is summarized in Figure 1. Differential expression analysis of the GSE232911 dataset (n = 246) identified 630 genes that were differentially expressed between AAA samples and normal controls, including 417 upregulated and 213 downregulated genes. The expression patterns of these genes are presented in a volcano plot and heatmap (Fig. 2A, B).

Figure 1.

Flowchart of the study design. The schematic illustrates the systematic analytical pipeline utilized in this study, encompassing raw transcriptomic data preprocessing, feature selection via machine learning algorithms, multidimensional construction of the diagnostic model, and downstream immune-molecular investigations. AUC = area under the curve, DCA = decision curve analysis, DEGs = differentially expressed genes, GSVA = gene set variation analysis, LASOO = least absolute shrinkage and selection operator, NETs = neutrophil extracellular traps, NRGs = NETs-related genes, ROC = receiver operating characteristic, SVM-RFE = support vector machine-recursive feature elimination, WGCNA = weighted gene co-expression network analysis.

Figure 2.

Transcriptomic profiling and functional enrichment analysis of DEGs in AAA. (A) Volcano plot of global transcriptomic alterations between AAA and normal control samples in the GSE232911 dataset (n = 246 total tissue samples, derived from 76 AAA patients and 13 donor controls). Thresholds were strictly set at FDR-adjusted P < .05 and |log₂ (FC)| >0.585. Red and blue nodes denote significantly upregulated and downregulated genes, respectively. (B) Hierarchical clustering heatmap visualizing the expression patterns of the identified DEGs. (C) Bar plot detailing GO enrichment across biological processes, cellular components, and molecular functions. (D) Bubble plot mapping the top KEGG signaling pathways. Significance in functional annotations was defined as adjusted P < .05. AAA = abdominal aortic aneurysm, DEGs = differentially expressed genes, FDR = false discovery rate, GO = Gene Ontology, KEGG = Kyoto Encyclopedia of Genes and Genomes.

Functional enrichment analysis was performed to characterize the biological relevance of the identified DEGs. GO analysis indicated that these genes were mainly enriched in biological processes related to immune regulation, including leukocyte proliferation, positive regulation of cell adhesion, and leukocyte-mediated immunity (Fig. 2C). In the cellular component and molecular function categories, enrichment was observed for terms such as external side of the plasma membrane and cytokine binding. Kyoto Encyclopedia of Genes and Genomes pathway analysis further demonstrated significant enrichment in inflammation-associated pathways, including cytokine–cytokine receptor interaction, chemokine signaling pathway, and NF‑κB signaling pathway (Fig. 2D).

3.2. Identification of AAA-specific NRGs by WGCNA

WGCNA was performed to construct a scale-free gene co-expression network and identify modules associated with AAA. After filtering low-variance genes and outlier samples, the automated pickSoftThreshold algorithm estimated an optimal soft-thresholding power of β = 11 to achieve scale-free topology (Fig. 3A). Hierarchical clustering coupled with dynamic tree cutting identified several distinct gene modules. Following the merging of highly similar modules (dissimilarity threshold = 0.3), the brown module demonstrated the strongest positive correlation with the AAA phenotype (correlation = 0.42, P = 1e-11) and contained 1252 genes (Fig. 3B).

Figure 3.

WGCNA identifying AAA-specific NETs-related genes. (A) Evaluation of the scale-free fit index and mean connectivity, establishing the optimal soft-thresholding power (β = 11). (B) Module-trait relationship heatmap mapping the Pearson correlation between identified co-expression modules and clinical status (AAA vs normal). The brown module exhibited the strongest positive correlation with the AAA phenotype. (C) Venn diagram mapping the precise intersection of DEGs, brown module components, and reference NRGs, yielding 22 diagnostic candidates. (D) Boxplot cross-examining the relative expression of the 22 candidate genes between the AAA and control cohorts. (E) Circos plot mapping the genomic chromosomal locations of the candidate genes. Statistical significance was determined via the Wilcoxon rank-sum test (*P < .05, **P < .01, ***P < .001). AAA = abdominal aortic aneurysm, DEGs = differentially expressed genes, NETs = neutrophil extracellular traps, NRGs = NETs-related genes, WGCNA = weighted gene co-expression network analysis.

To obtain AAA-specific NRGs, the intersection of DEGs, genes within the brown module, and the predefined NRG set was calculated. This approach yielded 22 candidate genes for subsequent analyses (Fig. 3C).

3.3. Expression patterns and immune relevance of AAA-specific NRGs

The expression profiles of the 22 candidate genes were examined in the training cohort. Most genes, including CXCR4, GZMB, and ITGAL, showed significantly higher expression levels in AAA samples compared with controls (P < .001; Fig. 3D). Their chromosomal locations are shown in a Circos plot, illustrating a broad genomic distribution (Fig. 3E), and a Manhattan plot highlights the statistical significance of these genes (Fig. 4A).

Figure 4.

Genomic landscape and immune crosstalk of the NETs-related candidate genes. (A) Manhattan plot charting chromosomal distributions and corresponding differential expression significance (−log₁₀ P value) of the 22 candidates. (B) Heatmap correlating candidate gene expression with defined immune cell infiltration subpopulations (quantified via ssGSEA). (C) Correlogram illustrating pairwise co-expression dynamics among the candidate genes evaluated via Pearson analysis. (D) Heatmap defining baseline differences in immune cell infiltration microenvironments between the AAA and control datasets (n = 246 total samples). (E) Violin plots confirming the differential abundance of respective infiltrating innate and adaptive immune cells (*P < .05, **P < .01, ***P < .001). AAA = abdominal aortic aneurysm, NETs = neutrophil extracellular traps, ssGSEA = single-sample gene set enrichment analysis.

ssGSEA demonstrated that the expression of these candidate genes was positively correlated with immune cell infiltration, particularly neutrophils and activated dendritic cells (Fig. 4B). Correlation analysis revealed strong positive associations among the candidate genes, suggesting coordinated regulation (Fig. 4C). Consistent with these observations, AAA samples exhibited significantly higher infiltration levels of these immune cell populations compared with controls (Fig. 4D, E).

3.4. Identification of diagnostic biomarkers using ML approaches

Three ML algorithms – LASSO, SVM-RFE, and RF – were applied to the 22 candidate genes to identify diagnostic biomarkers. LASSO regression retained 11 genes with nonzero coefficients (Fig. 5A, B). The SVM-RFE algorithm selected 16 genes that achieved optimal classification accuracy (Fig. 5C, D), while the RF algorithm identified the top 9 genes based on mean decrease accuracy (Fig. 5E). Intersection of the results from the 3 methods yielded 4 hub genes: CXCR4, GZMB, ITGA6, and CD47 (Fig. 5F).

Figure 5.

Algorithmic screening and establishment of the diagnostic hub gene signature. (A) Trajectory of independent variable coefficients evaluated by the LASSO regression model. (B) Selection of the optimal penalization parameter (λ_min) utilizing 10-fold cross-validation based on binomial deviance. (C and D) Feature ranking and classification error rate minimization tracked via SVM-RFE utilizing a 10-fold cross-validation loop. (E) Top candidate genes ranked by essentiality (Mean Decrease Gini index) prioritized via the random forest algorithm (n = 500 trees). (F) Venn diagram defining the absolute intersection of features surviving all 3 machine learning screens, ultimately yielding the 4-gene diagnostic signature (CXCR4, GZMB, ITGA6, and CD47). LASSO = least absolute shrinkage and selection operator, SVM-RFE = support vector machine-recursive feature elimination.

3.5. Construction and validation of the NETs-related diagnostic model

The diagnostic capacity of the identified signature was primarily evaluated in the training dataset (GSE232911, n = 246). Expression profiling confirmed significant dysregulation of all 4 hub genes in AAA samples compared to controls (Fig. 6A). ROC analysis demonstrated that CXCR4 yielded the highest independent discriminatory capacity (AUC = 0.890, 95% CI = 0.799–0.982), followed by GZMB (AUC = 0.818, 95% CI = 0.701–0.935), CD47 (AUC = 0.736, 95% CI = 0.633–0.840), and ITGA6 (AUC = 0.729, 95% CI = 0.615–0.842; Fig. 6B). Furthermore, a multivariate logistic regression model integrating these 4 genes achieved a substantially higher diagnostic accuracy, yielding an AUC of 0.920 (95% CI = 0.834–0.983; Fig. 6C).

Figure 6.

Performance evaluation and clinical applicability of the diagnostic model (training cohort: GSE232911). (A) Comparative boxplot confirming the differential expression of the 4 hub genes across the complete training matrix (n = 246 samples). (B) ROC array visualizing the independent discriminatory capacity and 95% confidence intervals of the individual hub genes. (C) Combined ROC curve identifying the aggregate diagnostic accuracy of the multivariate logistic regression model (AUC = 0.920). (D) A clinical nomogram integrating the 4 hub signatures to predict individual AAA risk. (E) Calibration curve mapping the stable conformity between nomogram-predicted probabilities and actual observations. (F) Decision curve analysis confirming the positive clinical net benefit of utilizing the diagnostic model across a spectrum of threshold probabilities (*P < .05, **P < .01, ***P < .001). AAA = abdominal aortic aneurysm, AUC = area under the curve, ROC = receiver operating characteristic.

To facilitate potential clinical translation, a diagnostic nomogram was constructed based on the multivariate model (Fig. 6D). Calibration plots confirmed stable agreement between the nomogram-predicted probabilities and the actual observed risks (Fig. 6E). Decision curve analysis further indicated a positive clinical net benefit across relevant threshold probabilities (Fig. 6F).

The generalizability of the signature was subsequently verified in the independent GSE7084 validation cohort. Consistent with the training set, CXCR4 and GZMB remained significantly upregulated in AAA tissues (Fig. 7A). ROC analysis confirmed the strong discriminatory stability of the signature, with both the individual CXCR4 marker and the combined 4-gene model achieving an AUC of 1.000 in this specific 15-sample localized cohort (Fig. 7B).

Figure 7.

Independent external validation and consensus clustering of molecular subtypes. (A) Validation boxplot verifying the consistent upregulation of hub genes within the independent GSE7084 cohort (AAA, n = 7; normal, n = 8). (B) ROC curve analysis confirming absolute diagnostic resolution (AUC = 1.000) within the validation series. (C) Consensus clustering matrix optimally bisecting the AAA training samples into 2 distinct molecular etiologies (Cluster 1 and Cluster 2) utilizing k-means clustering (k = 2, 50 iterations). (D) PCA scatterplot verifying the rigid spatial separation of the 2 molecular subsets (*P < .05, **P < .01, ***P < .001). AAA = abdominal aortic aneurysm, AUC = area under the curve, PCA = principal component analysis, ROC = receiver operating characteristic.

3.6. Identification of NETs-related molecular subtypes

Consensus clustering based on the expression of the 4 hub genes identified 2 molecular subtypes of AAA. The consensus matrix indicated optimal clustering at k = 2, classifying samples into Cluster 1 (C1) and Cluster 2 (C2; Fig. 7C). PCA confirmed clear separation between the 2 subtypes (Fig. 7D). Immune infiltration analysis revealed distinct immune profiles between the subtypes. C1 was characterized by increased infiltration of adaptive immune cells, including activated B cells and CD4+ T cells, whereas C2 showed significantly higher levels of neutrophils and other innate immune cells (Fig. 8A). Gene set variation analysis indicated enrichment of primary immunodeficiency and cardiac muscle contraction pathways in C2, while C1 was enriched in antigen processing and presentation as well as vascular endothelial growth factor signaling pathways (Fig. 8B). These findings suggest that C2 represents a neutrophil- and NETs-enriched AAA subtype.

Figure 8.

Deciphering the immune microenvironment of AAA subtypes and computationally repurposing therapeutic compounds. (A) Boxplot comparing ssGSEA-derived infiltrating immune cell populations between Cluster 1 and Cluster 2, mapped via the Wilcoxon rank-sum test. (B) GSVA highlighting differential activation of fundamental biological mechanisms bridging the 2 distinct clusters. (C) Bar plot displaying high-affinity candidate therapeutic compounds prioritized computationally via enrichment scores from the database. (D) Drug–gene bipartite interaction network predicting targeted pharmacological relationships between the computationally derived drug candidates and the 4 prognostic hub genes (*P < .05, **P < .01, ***P < .001). AAA = abdominal aortic aneurysm, GSVA = gene set variation analysis, ssGSEA = single-sample gene set enrichment analysis.

3.7. Prediction of potential therapeutic drugs

Potential therapeutic compounds targeting the NRG signature were explored using the DSigDB database. Several compounds showed significant negative enrichment scores, including Eugenol, Tretinoin, and 2-Nonenal (Fig. 8C). Additional candidates, such as Hydroxychloroquine and Etodolac, were also identified. A drug–gene interaction network was constructed to illustrate the potential relationships between these compounds and the hub genes (Fig. 8D).

4. Discussion

AAA remains a major clinical challenge due to its largely asymptomatic progression and the catastrophic outcome associated with rupture. Although maximal aortic diameter is currently the primary criterion guiding clinical decision-making, it does not reliably predict rupture risk at the individual level.^[34] From a pathological perspective, AAA is characterized by chronic transmural inflammation, ECM degradation, and progressive loss of VSMCs.^[1] While previous investigations predominantly evaluated macrophage and T lymphocyte dynamics, accumulating evidence suggests that NETs represent a critical, albeit underexplored, driver of AAA pathogenesis.^[35] NETs consist of decondensed chromatin decorated with granular proteins released from activated neutrophils and, beyond their antimicrobial role, have been implicated in sterile inflammation and vascular tissue injury.^[7,36] In this study, we applied an integrative bioinformatics strategy that combined WGCNA with multiple ML approaches to characterize the NETs-related transcriptomic features of AAA. Based on this framework, we identified a 4-gene diagnostic signature (CXCR4, GZMB, ITGA6, and CD47) and defined distinct molecular subtypes, providing a computationally derived perspective on the molecular heterogeneity underlying AAA.

Functional enrichment analysis demonstrated that the DEGs were mainly associated with leukocyte proliferation, positive regulation of cell adhesion, and the nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) signaling pathway. NF-κB is widely recognized as a central regulator of inflammatory responses in AAA, controlling the expression of pro-inflammatory cytokines and matrix metalloproteinase that contribute to aortic wall degradation.^[37] Mechanistically, the interplay between NF-κB and NETosis drives a self-amplifying inflammatory cycle. Pharmacological evidence confirms that functional NF-κB activation is a strict prerequisite for executing NET formation.^[38] Once extruded appropriately, the residual NET components – including cell-free DNA and histones – function as danger signals that reengage and perpetuate downstream NF-κB cascades in adjacent endothelial and VSMCs.^[39] Together, these validated sequences support our computational finding that NETs-related genomic alterations orchestrate an active, localized feed-forward inflammatory loop within the aneurysm wall.

To improve the robustness of biomarker selection, we applied a combined feature screening strategy incorporating LASSO regression, SVM-RFE, and RF algorithms. Compared with approaches based on a single algorithm, the intersection of multiple methods may reduce algorithm-specific bias and increase the stability of selected features.^[16] Using this strategy, 4 hub genes – CXCR4, GZMB, ITGA6, and CD47 – were consistently identified.

Among the 4 biomarkers, CXCR4 showed the strongest individual diagnostic performance. The CXCL12/CXCR4 signaling axis plays a central role in neutrophil trafficking from the bone marrow to sites of inflammation.^[40] In the context of AAA, dysregulated CXCR4 signaling has been associated with adventitial angiogenesis, a pathological process that facilitates inflammatory cell infiltration and weakens aortic wall structure.^[41] Consistent with this, murine studies have demonstrated that pharmacological or genetic inhibition of CXCR4 can attenuate aneurysm development.^[42,43] Taken together, these findings suggest that CXCR4 may represent a promising diagnostic marker for AAA.

GZMB is a serine protease classically associated with the cytotoxic functions of natural killer cells and cytotoxic T lymphocytes, but increasing evidence indicates its involvement in ECM degradation during aneurysm formation.^[44] Although GZMB is generally considered a downstream effector of immune activation, recent studies suggest that NETs may promote the recruitment of GZMB-expressing cytotoxic cells, thereby enhancing local proteolytic activity.^[45] In the aortic wall, extracellular GZMB has been shown to cleave key structural components, including fibrillin-1 and decorin, leading to impaired elasticity and reduced structural stability.^[46] The elevated expression of GZMB observed in our analysis is consistent with the notion that protease-driven matrix degradation plays an important role in AAA progression.

Regarding the remaining hub genes, CD47 and ITGA6 facilitate critical pathological cellular interactions. In aneurysmal lesions, increased CD47 expression on dying cells impairs efferocytosis, leading to the accumulation of apoptotic debris.^[47] Secondary necrosis of these cells can lead to the release of danger-associated molecular patterns, which are known to promote inflammatory responses and NETs formation.^[39,45] ITGA6, a member of the integrin family, is involved in cell–matrix adhesion and leukocyte–endothelial interactions. Its upregulation may facilitate sustained leukocyte adhesion and transmigration into the aneurysmal wall, thereby perpetuating local inflammation.^[48]

An important finding of this study was the identification of 2 distinct molecular subtypes of AAA. C2 exhibited a neutrophil- and NETs-enriched immune profile dominated by innate immune infiltration, whereas the 4 hub genes were more highly expressed in C1, which was characterized by increased adaptive immune cell infiltration, including activated B cells and CD4+ T cells. While seemingly paradoxical, this distribution likely reflects the temporal evolution of aneurysmal transmural inflammation. Early AAA pathogenesis is initiated by NET-mediated acute tissue injury; however, chronic maladaptive aortic remodeling ultimately necessitates the recruitment of adaptive lymphocytes. Hub genes such as CXCR4 – a critical chemokine receptor – actively facilitate both early neutrophil retention and subsequent focal lymphocyte homing. Therefore, the prominent upregulation of these hub genes in C1 may represent an advanced transitional state, wherein continuous innate inflammatory stimuli have fully engaged the local adaptive immune compartment. Recognizing this profound immunological heterogeneity precisely aligns with recent advancements utilizing integrated ML models to decode localized neutrophil subsets,^[49] ultimately suggesting that transcriptomic profiling could complement standard anatomical imaging to effectively gauge disease activity and precision immunotherapeutic responses.

To computationally validate the discriminative capacity of these biomarkers, we constructed a diagnostic nomogram. While the multigene model exhibited robust discriminatory stability in both the training (AUC = 0.920) and validation (AUC = 1.000) tissue cohorts, claims of direct clinical utility remain premature. Because our transcriptomic signature is exclusively derived from the aortic wall (tunica media/adventitia), converting these localized biological markers into a noninvasive, blood-based screening assay represents a significant biological leap. Consequently, these tissue-derived data provide a conceptual diagnostic framework that must first be rigorously cross-validated in matched PBMC or serum cohorts prior to determining their true clinical translational value.

Finally, computational pharmacogenomic screening via the DSigDB database identified Eugenol and Tretinoin as compounds with high theoretical binding affinities capable of reversing the established NETs-related genomic signature. Eugenol has been reported to exert anti-inflammatory effects through inhibition of NF-κB signaling and oxidative stress, both of which are key drivers of NETs formation.^[50] Tretinoin (all-trans retinoic acid) is widely used to induce neutrophil differentiation and has also been shown to suppress NETs formation in other inflammatory settings.^[51,52] Rather than indicating immediate therapeutic efficacy, these in silico findings solely serve as hypothesis-generating candidates. Rigorous in vivo preclinical testing, such as utilizing murine elastase-perfusion AAA models, is strictly required to ascertain whether these predicted agents can pharmacologically modify aneurysm progression.

Several limitations of this study should be acknowledged. First, the analyses were based primarily on retrospective transcriptomic datasets, and prospective validation in large, multicenter cohorts is necessary to confirm the diagnostic performance of the proposed gene signature. Second, although our results demonstrate an association between the identified genes and NETs-related pathways, the causal mechanisms by which these genes regulate NETosis in AAA remain unclear. Further in vitro and in vivo studies are required to elucidate the underlying signaling pathways and to determine whether targeting these mechanisms can effectively modify disease progression.

In summary, this in silico study systematically delineates a NETs-related transcriptomic profile strictly within the localized AAA microenvironment. The identification of a 4-gene candidate panel (CXCR4, GZMB, ITGA6, and CD47) and the corresponding stratification of immune-driven molecular subtypes expand the current biological understanding of NETosis-mediated inflammation in aneurysm pathogenesis. Importantly, these findings do not constitute an immediately applicable clinical diagnostic tool. Instead, they establish a conceptual computational framework intended to guide future experimental investigations into mechanism-targeted immunomodulation and to direct the prospective validation of these localized tissue markers in circulating biofluids.

Author contributions

Conceptualization: Feng Jiang.

Data curation: Feng Jiang, Zhen Zheng.

Formal analysis: Feng Jiang, Jiaming Lv.

Methodology: Feng Jiang, Zhen Zheng, Mengmeng Dong.

Project administration: Feng Jiang, Mengmeng Dong.

Resources: Feng Jiang, Jiaming Lv, Mengmeng Dong.

Software: Feng Jiang, Mengmeng Dong.

Writing – original draft: Feng Jiang, Jiaming Lv, Mengmeng Dong.

Writing – review & editing: Feng Jiang, Jiaming Lv, Zhen Zheng, Mengmeng Dong.

Validation: Jiaming Lv.

Investigation: Zhen Zheng, Mengmeng Dong.

Supervision: Mengmeng Dong.

Visualization: Mengmeng Dong.