Comprehensive machine learning for identifying platelet-associated diagnostic biomarkers and immune landscape in Kawasaki disease

Abstract

Background

Kawasaki disease (KD) is an acute, self-limited febrile illness primarily affecting children under 5 years of age. Platelets play a crucial dual role in both hemostasis and inflammatory/immune responses, contributing to vascular damage in diseases such as KD, making their study essential for the diagnosis and treatment of KD. Therefore, this study aimed to identify platelet-related diagnostic biomarkers, construct a machine learning-based diagnostic model, and computationally characterize their associated immune features, pathway activities, and regulatory networks in KD.

Methods

This study integrated transcriptomic datasets from 4 KD cohorts in the Gene Expression Omnibus database (training set: GSE68004/GSE73461; validation set: GSE100154/GSE63881). Platelet-associated differentially expressed genes were screened through differential analysis and Weighted Gene Co-expression Network Analysis. Four diagnostic biomarkers were identified via cross-validation using least absolute shrinkage and selection operator, random forest, and eXtreme Gradient Boosting algorithms, with model efficacy assessed through Receiver operating characteristic curve analysis. Immune infiltration characteristics were analyzed using single sample Gene Set Enrichment Analysis (GSEA) and CIBERSORT, pathway enrichment was performed via GSEA, molecular subtypes were classified using the non-negative Matrix Factorization algorithm, and miRNA and transcription factor (TF) regulatory networks were predicted through miRNet/NetworkAnalyst.

Results

We identified 4 platelet-associated biomarkers (CD63, F5, STXBP2, and SERPINA1) with exceptional diagnostic power. These genes orchestrate KD pathogenesis through dysregulated coagulation (F5), immune hyperactivation (CD63/STXBP2), and vascular injury (SERPINA1), further validated by their strong correlations with neutrophil infiltration and ribosome pathway suppression. Molecular subtyping revealed distinct immune-endotypes (e.g., neutrophil-dominant vs. natural killer-cell-enriched clusters), while regulatory network analysis uncovered has-miR-155-5p and TF hubs (SMAD5/SAP30/PHF8) as potential master regulators.

Conclusions

This study utilized platelet-associated biomarkers to establish a machine learning-based diagnostic model for KD and elucidated their pathogenic roles in immune dysregulation and vascular injury, thereby laying the foundation for improved diagnosis and targeted therapy in KD.

Highlight box.

Key findings

• Four platelet-associated biomarkers—CD63, F5, STXBP2, and SERPINA1—were identified and validated in multiple Kawasaki disease (KD) cohorts.

• A machine learning-based diagnostic model integrating least absolute shrinkage and selection operator, random forest, and eXtreme Gradient Boosting showed high diagnostic performance (area under the curve ≥0.80).

What is known and what is new?

• Platelets are involved in immune and vascular injury in KD, but platelet-based diagnostic markers are limited.

• This study establishes a platelet-focused machine learning diagnostic framework and identifies four robust biomarkers with inferred immune and regulatory features.

What is the implication, and what should change now?

• This framework may support improved molecular diagnosis of KD.

• The identified biomarkers and regulatory hubs may serve as candidates for future experimental validation and precision-oriented therapeutic exploration.

Introduction

Kawasaki disease (KD) is an acute, self-limited, systemic vasculitis syndrome predominantly affecting small and medium-sized blood vessels. It was first systematically described by Japanese physician Dr. Tomisaku Kawasaki in 1967 and was previously termed “mucocutaneous lymph node syndrome” (1,2). The disease primarily affects infants and young children aged 6 months to 5 years and represents the most common preventable cause of acquired heart disease in children from developed countries and regions (3). KD typically presents as a febrile, acute, non-life-threatening disease; however, it can progress to severe and potentially life-threatening complications, including atypical KD, KD Shock Syndrome, or secondary Hemophagocytic Lymphohistiocytosis, with the most severe cases leading to coronary artery aneurysms (4). Characteristic clinical features include persistent high fever lasting ≥5 days, accompanied by bilateral non-purulent conjunctivitis, lip fissuring/“strawberry tongue”, polymorphous rash, indurative edema of the hands and feet followed by periungual desquamation, and unilateral non-suppurative cervical lymphadenopathy (5,6). Epidemiological studies indicate that while KD occurs globally, its incidence is highest in East Asia, particularly in Japan, Korea, Taiwan (China), and coastal mainland cities of China (7). Notably, the CT genotype at CASP3 rs113420705 was significantly associated with KD incidence in children from northern India, suggesting that genetic susceptibility factors, such as CASP3, together with environmental triggers, may jointly contribute to disease development (8). Although its exact etiology remains incompletely understood, the prevailing hypothesis posits that genetically susceptible individuals mount an exaggerated immune response to one or more unidentified pathogens, leading to acute and potentially chronic vasculitis and vascular abnormalities, with the most severe impact on the coronary arteries (9). Consequently, early standard treatment with intravenous immunoglobulin (IVIG) combined with aspirin significantly reduces the incidence of coronary artery aneurysms and improves long-term prognosis (10,11).

Platelets not only maintain hemostasis and thrombus formation through their adhesion, aggregation, and release reactions but also function as “inflammatory-immune” cells (12,13). Their quantity, volume, and activation status—such as mean platelet volume, platelet distribution width, and plateletcrit—have been demonstrated to predict risks in diverse inflammatory diseases and cardiovascular events (14,15). Upon recognizing pathogens or endothelial injury signals via receptors like CD62P, TLR4, and CD40L, platelets immediately release cytokines, antimicrobial peptides, and microvesicles (16). They form “platelet-leukocyte aggregates” with neutrophils, monocytes, and dendritic cells, thereby amplifying inflammatory cascades and triggering thrombo-inflammation. This dual mechanism orchestrates both immune defense and vascular damage in vasculitic diseases such as sepsis, atherosclerosis, and KD (17,18). Platelet-related genes (PRGs) have been confirmed as diagnostic biomarkers for multiple diseases, such as ischemic stroke (19). Studies also indicate that PRGs can serve as prognostic marker genes for diseases (20). However, studies on PRGs in KD remain insufficient. Therefore, investigating platelet-associated diagnostic biomarkers and immune features is crucial for improving the treatment of KD.

This study systematically constructed a platelet biomarker-based diagnostic framework for KD by integrating machine learning algorithms with multi-omics data, revealing the central regulatory role of platelets in immune dysregulation and vascular injury during KD pathogenesis. Molecular subtyping was employed to resolve disease heterogeneity, while regulatory network analysis identified potential therapeutic targets. This research provides novel insights for the precise diagnosis and personalized treatment of KD. Not only has it established a robust biomarker screening system, but has also deepened the understanding of KD pathological processes through multi-dimensional mechanistic exploration, laying a theoretical foundation for future development of targeted interventions and clinical translation. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-aw-800/rc).

Methods

Data collection

RNA-Seq and clinical data from datasets GSE68004 (37 controls, 76 KD), GSE73461 (55 controls, 77 KD), GSE100154 (23 controls, 21 KD), and GSE63881 (170 controls, 171 KD) were retrieved from the Gene Expression Omnibus (GEO) database (https://portal.gdc.cancer.gov/). All transcriptomic datasets were obtained from publicly available sources. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Batch effects in GSE68004 and GSE73461 were corrected using the “sva” package, with these merged datasets forming the training cohort. GSE100154 and GSE63881 were designated as the validation cohort. A list of 300 PRGs was obtained from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb) (Table S1).

Differential analysis and Weighted Gene Co-expression Network Analysis (WGCNA)

Differential analysis between control and disease groups was conducted using the “limma” package (thresholds: |logFC| >1, adjusted P value <0.05). Subsequently, WGCNA was applied to the top 50% most variable genes. Hierarchical clustering with a soft thresholding power (β=24) grouped genes into co-expression modules based on expression variation patterns. Module-trait correlations were evaluated via Pearson analysis to identify disease-associated modules. Finally, intersections among differentially expressed genes (DEGs), PRGs, and WGCNA module genes yielded differentially expressed platelet-related genes (DEPRGs).

Screening diagnostic signature genes to construct a diagnostic model

The intersection genes served as candidate features, and to prevent model overfitting, least absolute shrinkage and selection operator (LASSO) regression analysis was implemented using the “glmnet” package with penalty parameter lambda selected via cross-validation to eliminate highly correlated genes and reduce model complexity. Feature selection was subsequently performed via both “randomForest” and “xgboost” packages. The intersection gene set derived from eXtreme Gradient Boosting (xgboost), LASSO, and randomForest was employed as the signature genes for the diagnostic model. The diagnostic performance was assessed using the pROC package, with the area under the curve (AUC) calculated to quantify discriminative power.

Immune infiltration analysis and its correlation with signature genes

Immune infiltration scores for cell types and functional signatures were computed using single sample Gene Set Enrichment Analysis (ssGSEA). CIBERSORT analysis further delineated immune infiltration disparities. Statistical differences in immune infiltration levels between the control and disease groups were compared. Additionally, correlation analysis was performed using the corrplot package to calculate Pearson correlation coefficients both among distinct immune cell types and between the identified diagnostic signature genes and immune cells.

Gene Set Enrichment Analysis (GSEA) and pathway analysis

Pathway enrichment analysis was performed on the diagnostic genes using GSEA (version 4.3.2). Concurrently, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were conducted on the identified DEGs employing the “clusterProfiler” R package. These analyses enabled comprehensive assessment of enrichment patterns across biological processes, cellular components, molecular functions, and key signaling pathways.

Constructing molecular subtypes of PRGs

Non-negative Matrix Factorization (NMF) algorithm was employed to cluster KD samples into different subtypes (K=2) based on the expression matrix of Diagnostic -related DEPRGs. Principal component analysis (PCA) was employed to examine the clustering patterns of molecular subtypes, and the expression levels of the four genes across these subtypes were assessed.

Prediction of potential transcription factor (TF) and construction miRNA-target gene regulatory networks

The miRNet database (https://www.mirnet.ca/) was utilized to identify potential miRNAs targeting the model’s signature genes, while NetworkAnalyst (https://www.networkanalyst.ca/) was employed to predict upstream TFs regulating these characteristic genes. All resulting networks were visualized using Cytoscape v3.10.2.

Statistical analysis

All statistical analyses were performed using R (version 4.4.1). The Wilcoxon test was employed to determine statistically significant differences between groups. Associations between variables were assessed using both Pearson and Spearman correlation methods. Statistical significance was defined as P<0.05.

Results

Analysis of DEPRGs in KD patients

The training datasets (GSE68004 and GSE73461) were integrated with batch correction (Figure S1), and differential expression analysis identified 507 DEGs, including 457 upregulated and 50 downregulated genes (Figure 1A); however, residual heterogeneity related to disease severity, age, or sample type may influence these results. We will use independent GEO validation dataset to verify the gene expression patterns and diagnostic efficiency of the identified feature genes. WGCNA was subsequently performed on these DEGs. Based on the scale-free topology criterion, the soft-thresholding power was calculated and set to 24 (Figure 1B). Using this power value, a hierarchical clustering tree was constructed (Figure 1C). Genes within the same module represent highly interconnected clusters, ultimately resulting in the identification of nine distinct co-expression modules, each color-coded. These modules were correlated with the KD group and the control group. The MEblue module demonstrated a significant positive correlation with the KD group (P<0.001, r=0.85) and a significant negative correlation with the control group (P<0.001, r=−0.85) (Figure 1D). Genes within the MEblue module were selected for further correlation analysis, revealing a significant positive correlation between module membership and gene significance (cor=0.79, P<0.001) (Figure 1E). The intersection of the DEGs, hub genes from WGCNA, and PRGs yielded eight key DEPRGs (Figure 1F). GO enrichment analysis indicated that the DEPRGs are primarily involved in biological processes such as wound healing, hemostasis, and blood coagulation. They are associated with cellular components including platelet alpha granules and the endoplasmic reticulum lumen, and exhibit molecular functions related to peptidase regulator activity and peptidase inhibitor activity (Figure 1G).

Figure 1.

DEPRGs identification and functional analysis. (A) Volcano plot of DEGs between the KD group and control group in the training set. (B) WGCNA soft-thresholding power analysis. (C) WGCNA clustering dendrogram. (D) Correlation between modules and traits. (E) Scatter plot of module membership in the MEblue module. (F) Venn diagram of the intersection among PRGs, DEGs, and hub genes derived from WGCNA. (G) Functional enrichment analysis of DEPRGs. BP, biological processes; CC, cellular components; DEGs, differentially expressed genes; DEPRGs, differentially expressed platelet-related genes; FDR, false discovery rate; KD, Kawasaki disease; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular functions; PRGs, platelet-related genes; WGCNA, Weighted Gene Co-expression Network Analysis.

Selection of signature genes

To prevent overfitting, we performed LASSO regression analysis on the 8 DEPRGs and identified seven candidate genes (Figure 2A). Random forest analysis revealed that the CD63 gene exhibited the highest feature importance in the model, and eight signature genes were ultimately identified (Figure 2B). XGBoost modeling results demonstrated that the four genes CD63, F5, STXBP2, and SERPINA1 exhibited the highest gain values in the feature importance evaluation (Figure 2C). The intersection across three algorithms ultimately yielded four signature genes (Figure 2D). All four signature genes demonstrated pairwise positive correlations, with the strongest correlation observed between CD63 and STXBP2 (cor=0.70, P<0.001). This pattern suggests potential cooperative functions in shared biological processes (Figure 2E).

Figure 2.

Signature gene screening. (A) Cross-validation curve for selecting the optimal λ value, and the coefficient distribution in the LASSO regression model. (B) Random forest importance ranking. (C) XGBoost score plot. (D) Venn diagram comparing three algorithms. (E) Heatmap showing correlations among the 4 candidate genes. ***, P<0.001. RF, random forest; LASSO, least absolute shrinkage and selection operator; XGBoost, eXtreme Gradient Boosting.

Diagnostic model construction and validation

The diagnostic performance of the 4 biomarkers was evaluated using receiver operating characteristic curve (ROC) analysis in both the training set and an independent validation cohort. In the training set, the integrated model demonstrated outstanding diagnostic performance with an AUC of 0.938, while each individual signature gene exhibited exceptional discriminatory power with all four genes achieving AUC values exceeding 0.966 (Figure 3A). These robust findings were consistently replicated across independent validation cohorts: the model maintained strong predictive accuracy in the GSE63881 validation set (AUC =0.934), with all signature genes showing AUC values above 0.853 (Figure 3B), and similarly achieved reliable performance in the GSE100154 validation set (model AUC =0.876), where all genes retained AUC values greater than 0.749 (Figure 3C). The nomogram demonstrates the contributions of four genes (STXBP2, F5, CD63, and SERPINA1) to the diagnostic prediction model (Figure 3D). In addition, calibration curves and clinical impact curve were presented to assess the performance of the nomograms, which showed that the predicted curves of the model are close to the ideal ones (Figure 3E). In the training set, significantly elevated expression levels of all 4 diagnostic signature genes were consistently observed in the KD group compared to control group (Figure 3F). This expression pattern was consistently replicated in the validation cohorts GSE63881 (Figure 3G) and GSE100154 (Figure 3H), with significantly higher expression levels observed in the KD group compared to control group. We expanded our analysis to evaluate the diagnostic specificity of the gene panel by comparing KD with other inflammatory diseases that share clinical features, including systemic juvenile idiopathic arthritis (JIA), group A streptococcal infections (GAS), and human adenovirus (HAdV) infections. Differential expression analysis (Figure 4A,4B) and ROC curve evaluations were performed across these conditions (Figure 4C). Although our model successfully distinguished KD from healthy controls, its ability to differentiate KD from these other diseases was more limited, with lower specificity observed in certain comparisons. The platelet-associated gene signature achieved an AUC of 0.938, demonstrating higher or comparable diagnostic performance relative to previously reported KD biomarkers, including N-terminal prohormone of brain natriuretic peptide (NT-proBNP), platelet count, erythrocyte sedimentation rate (ESR), IL-6, Neutrophil to lymphocyte ratio (NLR), and PK2, whose reported AUCs ranged from 0.678 to 0.923 based on published studies (Figure 5). These findings indicate that the platelet-associated gene model shows strong diagnostic performance when benchmarked against established and emerging KD diagnostic markers.

Figure 3.

Diagnostic model development and validation. (A) ROC curve of the training set. (B) ROC curve of GSE63881. (C) ROC curve of GSE100154. (D) A nomogram was established to predict the diagnostic capability of KD. (E) Calibration curves and clinical impact curve. The expression differences between the KD group and the control group in the training set, (F) GSE63881 (G) and GSE100154 (H). **, P<0.01; ***, P<0.001; ****, P<0.0001. KD, Kawasaki disease; ROC, receiver operating characteristic curve.

Figure 4.

Diagnostic specificity of the platelet-associated gene panel in distinguishing KD from other inflammatory diseases. (A) Box plot of platelet-associated genes in KD, GAS, HAdV infections, healthy and inKD in GSE68004. (B) Box plot of platelet-associated genes in KD, bacterial infections, systemic JIA, HSP, viral infections, control and uncertain infections in GSE73461. (C) ROC curve analysis comparing the diagnostic performance of the gene panel across different inflammatory conditions. Significance markers (*) indicating comparisons to KD. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001; GAS, group A streptococcal disease; HAdV, adenovirus; HSP, Henoch-Schönlein purpura; inKD, incomplete KD; JIA, juvenile idiopathic arthritis; KD, Kawasaki disease; ns, non-significance.

Figure 5.

Histogram of AUC values for multiple biomarkers. AUC, area under the curve.

Immune profiling analysis of disease and normal group

The distribution of immune cell types shows noticeable differences between the control and KD groups. In the KD group, there is a higher proportion of dendritic cells resting, monocytes and neutrophils, while the control group exhibits a more balanced distribution across various immune cell types, including a higher presence of T cells CD4 naive and T cells CD8 (Figure 6A). CIBERSORT analysis further demonstrated significantly increased infiltration in the KD group for the following cell subsets: T cells gamma delta, natural killer (NK) cells resting, T cells regulatory (Tregs), T cells CD4 memory activated, monocytes, neutrophils, and macrophages M0. Conversely, the control group exhibited elevated infiltration of B cells memory, T cells CD4 naive, T cells CD8, NK cells activated, and T cells CD4 memory resting (Figure 6B). ssGSEA revealed significantly higher enrichment scores in the KD group versus normal controls for APC co-inhibition, macrophages, neutrophils, Type II IFN response, and Tregs (Figure 6C). All 4 genes show a strong positive correlation with neutrophils, macrophages (M0), dendritic cells activated and monocytes; while they exhibit a strong negative correlation with T cells (CD8, CD4 naive, and CD4 memory resting) and B cells (memory) (Figure 6D).

Figure 6.

Immune microenvironment characterization. (A) Distribution of immune cell types in control group versus KD group. (B) CIBERSORT immune cell fractions in KD and control groups. (C) ssGSEA immune cell for KD and control groups. (D) Correlation between the four signature genes and immune infiltration levels. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. KD, Kawasaki disease; ns, non-significance; ssGSEA, single sample Gene Set Enrichment Analysis.

Enrichment analysis of KD and control groups

Further investigation into the potential biological functions of KD diagnostic biomarkers revealed that CD63, F5, SERPINA1, and STXBP2 were significantly enriched in pathways related to ribosome-related processes (including structural constituent of ribosome, ribosomal subunit biogenesis, ribosome biogenesis, and rRNA processing), granule-related components (such as azurophil granule, tertiary granule, and specific granule membrane), as well as mitochondrial and vesicle processes (encompassing mitochondrial translation, phagocytic vesicle, and vesicle lumen) (Figure 7A). In single-gene GSEA enrichment analysis, the enrichment scores of all four biomarkers collectively demonstrated positive correlations with the tertiary granule pathway (Figure 7B). KEGG enrichment analysis of the 4 genes revealed that the Ribosome pathway was significantly enriched across all genes (Figure 7C). Moreover, in single-gene GSEA enrichment analysis, the enrichment scores of all 4 genes universally exhibited negative correlations with the Ribosome pathway (Figure 7D).

Figure 7.

Enrichment analysis of four genes. GO enrichment analysis results (A) and single-gene GSEA enrichment analysis (B) for CD63, F5, SERPINA1, and STXBP2. KEGG enrichment analysis results (C) and single-gene GSEA enrichment analysis (D) for CD63, F5, SERPINA1, and STXBP2. GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes

Molecular subtyping based on DEPRGs

Based on 4 diagnosis-related DEPRGs, NMF clustering analysis was performed, and the optimal cluster number K=2 was selected. The clustering matrix at K=2 clearly divided into two distinct blocks. Two related disease subtypes were identified: cluster1 comprising 70 samples and Cluster2 comprising 83 samples (Figure 8A,8B). The bar plot depicts the distribution of various immune cell types across two clusters, with cluster 1 and cluster 2 showing similar compositions dominated by monocytes, macrophages M0, and neutrophils, while T cells (CD8 and CD4 memory resting) and B cells memory are present in smaller proportions (Figure 8C). CIBERSORT analysis further revealed that infiltration of neutrophils was significantly elevated in cluster1 compared to cluster2, whereas cluster2 exhibited significantly higher infiltration in resting NK cells and memory B cells (Figure 8D). ssGSEA demonstrated that both immune cell scores and immune infiltration subtype scores were significantly higher in cluster2 compared to cluster1 (Figure 8E,8F). Moreover, the two clusters could be clearly distinguished through PCA analysis (Figure 8G). There is a statistically significant difference in the expression of CD63 and F5 between cluster 1 and cluster 2, with CD63 being higher in cluster 2 and F5 being higher in cluster 1 (Figure 8H).

Figure 8.

Cluster identification and immune profiling. (A) Cophenetic coefficient plot. (B) NMF consensus clustering plot. (C) Distribution of immune cell types in different clusters. (D) CIBERSORT immune cell fractions in different clusters. (E) ssGSEA immune cell for different clusters. (F) ssGSEA immune cell and immune infiltration subtype heatmap. (G) PCA analysis. (H) Expression profiles of 4 candidate genes across different subtypes. *, P<0.05; **, P<0.01; ***, P<0.001. NMF, non-negative Matrix Factorization; ns, non-significance; PCA, principal component analysis; ssGSEA, single sample Gene Set Enrichment Analysis.

Construction of potential TF and miRNA regulatory networks

The three target genes (CD63, F5, and STXBP2) are subject to complex miRNA-mediated regulation, with certain miRNAs such as hsa-miR-155-5p and hsa-miR-6890-3p exhibiting extensive interactions with these target genes (Figure 9A). In the TF network, multiple TFs including SMAD5, SAP30, and PHF8 were identified as key regulators interacting with several diagnostic genes, suggesting their central role in transcriptional regulation (Figure 9B).

Figure 9.

Construction of regulatory networks based on signature genes. (A) miRNA network of signature genes. (B) TF-gene interaction network. TF, transcription factor.

Discussion

This study systematically constructed and validated a novel molecular diagnostic model for KD through the integration of multi-omics data and machine learning algorithms. Core diagnostic biomarkers with pivotal regulatory functions were screened based on differential gene expression and co-expression network analyses, with the model’s robustness and generalizability confirmed via multi-dimensional validation strategies. By combining immune microenvironment characteristics with gene functional enrichment analyses, the research elucidated the potential biological significance of these diagnostic biomarkers in KD pathophysiology. Simultaneously, disease heterogeneity was explored through molecular subtyping. The established integrated diagnostic framework not only provides a reliable tool for precise KD identification but also lays a theoretical foundation for in-depth mechanistic investigation and targeted intervention development through its multi-tiered analytical approach.

These 4 signature genes also play significant roles in disease pathogenesis. Their functions involve key biological processes such as cellular signaling, coagulation system regulation, protease inhibition, and immune modulation. CD63, as a member of the tetraspanin superfamily (21), is specifically enriched on platelet endosomes and alpha-granule membranes. It participates in tumor microenvironment remodeling by regulating exosome-mediated intercellular communication (22). This study indicates that CD63 demonstrates significantly elevated expression levels in KD patients. ROC curve analysis reveals that CD63 exhibits high diagnostic accuracy. The F5 gene encodes coagulation factor V (FV), a central regulator of the coagulation cascade that possesses both pro- and anticoagulant functions (23,24). Mutations in this gene can lead to diametrically opposed clinical phenotypes. For instance, the F5 c.3481C>T stop codon mutation causes complete FV deficiency manifesting as recurrent bleeding in homozygous states, while heterozygous mutations are associated with severe thrombotic events, demonstrating that gene dosage effects dictate phenotypic switching (25). Our study reveals significant upregulation of F5 in KD patients, with a potential mechanism involving F5 binding to the prothrombinase complex, culminating in the characteristic hypercoagulable state of KD. This elevation aligns with increased levels of platelet activation markers CD62P/CD63, collectively promoting microvascular occlusion (26). SERPINA1 serves as the primary serine protease inhibitor, neutralizing neutrophil elastase and complement components (27). Studies have demonstrated that plasma exosome analysis reveals significant downregulation of SERPINA1 in KD patients with coronary artery aneurysms. This deficiency drives protease/antiprotease imbalance, accelerating vascular matrix degradation and aneurysm formation (28). STXBP2, as a member of the Munc18 protein family, maintains cytotoxic T lymphocyte function by regulating cytolytic granule exocytosis (29). Studies have demonstrated that STXBP2 mutations are genetically linked to susceptibility of KD complicated by macrophage activation syndrome (MAS) (30). Collectively, these four genes orchestrate complex pathogenic networks in metabolic disorders, neoplasms, and coagulopathies through their synergistic regulation of immune responses (via CD63/SERPINA1), coagulation homeostasis (governed by F5), epigenetic modifications (mediated by STXBP2/SERPINA1), and cellular migration (modulated by CD63/F5), thereby establishing interconnected molecular pathways that drive disease progression across multiple physiological systems.

Our study demonstrates significant enrichment of the ribosome pathway across all four genes, with their enrichment scores universally exhibiting negative correlations with this pathway. Disruption of ribosome biogenesis induces nucleolar stress through the binding of ribosomal proteins (e.g., RPL5, RPL11) to MDM2, thereby activating the p53 signaling pathway and triggering cell cycle arrest or apoptosis (31,32). In KD, this stress response may accelerate vascular endothelial cell and immune cell damage, promoting coronary artery aneurysm formation (33). Research confirms significant activation of the p53 pathway in vascular endothelial apoptosis and inflammatory infiltration among KD patients (34). Ribosomes in platelets perform specialized functions. In the resting state of platelets, ribosomes bind to specific mRNAs to form protective complexes, preventing their degradation (35). Studies demonstrate that the ribosome rescue factor Pelota modulates ribosome-mediated mRNA decay, thereby maintaining the stability of mRNAs encoding platelet activation-related genes (e.g., coagulation factors, inflammatory factors) and enabling rapid protein translation upon platelet activation (36,37).

Immunological profiling demonstrated significantly elevated neutrophil infiltration in KD patients, a pattern consistently observed in cluster 1, whereas increased NK cells and resting NK cells in healthy controls mirrored the immune signature of cluster 2. As rapid-response effectors of innate immunity, neutrophils drive acute-phase inflammatory storms by releasing neutrophil extracellular traps (NETs), reactive oxygen species (ROS), and proteases (e.g., MMP-9), directly damaging vascular endothelium (38,39). Studies confirm that peripheral blood neutrophils in KD children exhibit a hyperactivated phenotype characterized by significant upregulation of chemotaxis-, NETs-, and ROS-related genes, correlating positively with inflammatory markers and coronary lesions (40). Conversely, the accumulation of resting NK cells may reflect a tendency toward immune equilibrium restoration. Evidence indicates that resting NK cells help rebuild immune homeostasis by suppressing residual inflammatory cytokines (e.g., TNF-α, IL-6) and directly eliminating activated CD4⁺ T cells, thereby constraining excessive immune activation and providing essential “braking signals” for long-term immunological recovery (41). Although we did not directly assess Th17 cell populations or IL-17 expression, the observed correlations between platelet-associated biomarkers and neutrophil infiltration may suggest that Th17/IL-17-mediated pathways contribute to immune activation in KD. Future studies integrating T cell subset profiling and cytokine measurements with platelet-associated molecular signatures could further elucidate these mechanisms.

This study reveals that the key target genes of KD (CD63, F5, and STXBP2) are precisely controlled by multi-tiered regulatory networks. At the post-transcriptional regulation level, hsa-miR-155-5p and hsa-miR-6890-3p form a core regulatory axis through their extensive actions on target genes. miR-155-5p is an microRNA that functions in multiple organs; for example, high miR-155-5p levels are associated with poor prognosis in breast cancer (42). In this study, hsa-miR-155-5p occupies a central regulatory position. Given its well-established roles in immune responses, inflammation, and cancer, this suggests these targets genes may be closely linked to related pathological or physiological pathways (43). At the transcriptional regulation level, SMAD5, SAP30, and PHF8 constitute a hub-like TF network. SMAD5, a component of the TGF-β signaling pathway, is known to regulate cell proliferation and differentiation, while PHF8 and SAP30 further highlight the epigenetic and transcriptional complexity involved in regulating these target genes (44,45). Previous studies have shown that IVIG treatment can markedly reshape the methylation landscape and expression of specific genes in KD patients (46). Epigenetic mechanisms can modulate gene expression and may provide a link between genetic predisposition and environmental susceptibility in KD. Analyses of epigenetic factors have also been proposed as a strategy to identify diagnostic and prognostic biomarkers for KD (47,48). In addition, single-cell transcriptomic approaches offer opportunities to dissect immune cell dynamics and treatment responses, with recent studies revealing systemic immune dysregulation in IVIG-resistant KD patients (49). Together, these observations suggest that future studies integrating multi-omic techniques—including single-cell transcriptomics, genomic variation, and epigenetic profiling—could provide a more comprehensive understanding of KD pathogenesis and clarify the functional roles of candidate genes in the disease.

Our findings underscore the exploratory nature of the proposed diagnostic model, highlighting its potential to distinguish KD from healthy controls but also its limitations in differentiating KD from diseases with similar clinical presentations. A key limitation is that external validation was limited by dataset availability and did not capture the full spectrum of KD-mimicking pediatric infections. The absence of detailed sampling-time metadata limited our ability to account for temporal heterogeneity in gene expression across the dynamic course of KD. And this an mRNA-based analysis, this study cannot distinguish causal mechanisms from secondary inflammatory responses nor directly infer protein-level or functional effects. These results indicate the need for further validation across a broader set of diseases and clinical contexts. Future studies should integrate animal models to investigate the causal roles of CD63/F5 in platelet activation and endothelial injury, while advancing the development of targeted nanotherapeutics against SMAD5 or has-miR-155-5p. As transcriptomic and clinical data were derived from independent cohorts, our comparison represents an indirect benchmark. Future multi-center studies integrating multi-omics data within the same cohort will be necessary for direct validation.

Conclusions

This study developed a robust diagnostic model for KD using machine learning and platelet-associated biomarkers. We identified key regulatory genes and elucidated their pathogenic roles in immune dysregulation and vascular injury. Molecular subtyping revealed distinct disease endotypes, and regulatory network analysis pinpointed critical therapeutic targets, including miRNAs and TFs. This integrated framework provides a foundation for improved diagnosis and targeted therapy development in KD.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.