Identification of Epigenetic Regulator‐Associated Genes in Keloid Disease Through Integrated Bulk and Single‐Cell Transcriptomics With RT‐qPCR Validation

ABSTRACT

There exists a close correlation between epigenetic factors and the progression of keloid disease (KD). The research aims to identify the key genes linked to epigenetic factors in KD, which may provide new insights into the therapeutic management of KD. The four datasets (GSE113619, GSE145725, GSE181316, and GSE293834) and genes related to epigenetic factors (ERGs) were retrieved from public repositories. Machine learning algorithms combined with mRNA expression validation via RT‐qPCR identified critical epigenetic factor‐associated genes. Follow‐up analyses included GSEA, immune cell infiltration assessment, therapeutic compound screening, and pseudotemporal trajectory analysis. RT‐qPCR assays on clinical specimens confirmed the mRNA expression patterns of the identified genes. Three feature genes were identified, among which only HR and SMYD4 showed significant and consistent differential expression (p < 0.05). GSEA indicated enrichment of these genes in metabolic and cellular pathways including alpha‐linolenic acid processing, riboflavin metabolism, and protein secretion mechanisms. Immune profiling showed elevated platelet frequencies and reduced dendritic cell populations in KD patients versus controls (p < 0.05). Computational screening identified 10 potential compounds for HR (including propylthiouracil, estradiol, and triiodothyronine) and 8 compounds for SMYD4 (including acetaminophen, bisphenol A, and pirinixic acid). Pseudotime analysis of smooth muscle cells and pericytes (SMC/PC) and melanocytes (MEL) showed that during cell differentiation, HR was predominantly expressed in MEL. Clinical validation confirmed these expression patterns. In the present investigation, two key genes associated with epigenetic factor were obtained, which might serve as potential exploratory biomarkers and warrant further epigenetic mechanistic investigation.

Step I: Screened HR/SMYD4 as key keloid epigenetic genes via bulk/scRNA‐seq and machine learning. Step II: Built a high‐performance nomogram (AUC = 0.988). Step III: Conducted multi‐omics analyses. Step IV: Experimentally verified their differential expression and pathogenic roles.

Abbreviations

cor

correlation coefficients

CTD

comparative toxicogenomics database

DEGs

differentially expressed genes

DNMT1

DNA methyltransferase 1

ECM

extracellular matrix

ERGs

epigenetic factor‐related genes

GEO

gene expression omnibus

GO

Gene Ontology

GSEA

gene set enrichment analysis

HVGs

highly variable genes

KD

keloid disease

KEGG

Kyoto Encyclopedia of Genes and Genomes

LASSO

Least Absolute Shrinkage and Selection Operator

MDSCs

myeloid‐derived suppressor cells

MSigDB

Molecular Signatures Database

ncRNAs

non‐coding RNA

NES

normalized enrichment score

PPI

protein–protein interaction

RF

random forest

ROC

receiver operating characteristic

RT‐qPCR

reverse transcription‐quantitative polymerase chain reaction

scRNA‐seq

single‐cell RNA sequencing

STRING

Search Tool for the Retrieval of Interacting Genes

SVM‐RFE

support vector machine recursive feature elimination

1. Introduction

Keloid disease (KD) is a pathological lesion of fibrous tissue with disordered collagen proliferation post‐skin healing, featuring invasive horizontal growth that extends beyond original wound margins into healthy skin [1]. Per the triad hypothesis, primary etiologies include endothelial dysfunction, inflammation, and microvascular injury; secondary factors involve genetic predisposition, infection, surgical frequency, and wound/suture site mechanical tension [2]. Available treatments include surgical and non‐surgical modalities (radiotherapy, compression, steroids, laser). Surgical excision plus adjuvant electron beam radiotherapy is a cornerstone, though recurrence rates remain ~16% [3, 4]. However, KD's molecular basis is incompletely characterized; identifying critical genes is essential to elucidate pathophysiology, offering avenues for diagnostic biomarkers and therapeutic targets.

Epigenetic mechanisms are central to KD's multifaceted pathology. They involve reversible chromatin changes modulating transcription (independent of DNA sequence) that persist through cell divisions [5, 6]. Three key epigenetic mechanisms are well‐characterized: DNA methylation (the most extensively studied), post‐translational histone modifications, and regulatory noncoding RNAs (ncRNAs). Their dysregulation, which affects cell fate and tissue homeostasis, is linked to fibrotic diseases. Emerging evidence reveals epigenetic abnormalities in KD—notably, DNMT1 (a DNA methyltransferase) is expressed in 100% of KD‐derived fibroblasts vs. ~8% of normal skin fibroblasts, indicating disrupted DNA methylation may drive KD's fibrotic phenotype [7]. However, the systemic dysregulation of epigenetic factor‐related genes (ERGs) in KD, their regulatory networks controlling fibroblast activation, abnormal proliferation, and excessive collagen deposition, and underlying functional mechanisms remain poorly understood.

For highly heterogeneous complex diseases like KD, integrating bulk and single‐cell transcriptomics is indispensable for elucidating core pathological mechanisms. Bulk transcriptomics offers global tissue gene expression but reflects an “average” of all cells, obscuring rare cell subpopulations or disease‐driving cellular states [8]. In contrast, single‐cell transcriptomics analyzes individual cells, enabling unbiased tissue cellular atlas mapping, precise identification of functionally distinct cell subtypes, dissection of cell‐microenvironment interactions and transcriptional signatures, and tracking of dynamic cellular states during KD progression [9]. Integrating these two approaches allows complementary investigation of gene functions and regulatory networks, revealing complex molecular interactions, signaling pathways, and cell subtypes in KD [10]. This combined strategy offers a robust methodological framework for systematically identifying key pathogenic genes in KD.

Utilizing open‐access KD transcriptomic repositories, we characterized genes with altered expression linked to epigenetic regulatory factors through differential expression profiling. Critical genes were further refined through integrated machine learning algorithms, expression pattern validation, and ROC curve assessment. Leveraging these identified genes, we developed a predictive nomogram and conducted comprehensive analyses including pathway enrichment evaluation, immune cell composition profiling, molecular interaction network mapping, and therapeutic compound screening. Additionally, single‐cell transcriptomic datasets were incorporated to examine the temporal expression patterns of critical genes throughout pivotal cellular differentiation stages. Our results establish potential diagnostic markers for early disease detection and identify novel molecular targets for therapeutic development.

2. Materials and Methods

2.1. Data Collection

KD‐related datasets were obtained from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/). The training set utilized GSE113619 (platform GPL21290), containing skin tissue specimens from 10 KD patients and 8 healthy controls. The GSE113619 dataset was generated based on a paired biopsy design: a 3 mm punch biopsy of non‐lesional skin was performed on Day 0 (baseline), followed by a second 4 mm punch biopsy at the same anatomic site 42 days later. Since the original study [11] aimed to characterize transcriptional changes occurring 6 weeks after wounding in KD‐prone versus healthy individuals, our analysis retained only the Day‐42 samples to capture the relatively stable fibrotic phase. Day‐0 samples were excluded because they reflect the acute, pre‐trauma homeostatic state, which does not align with the post‐trauma fibrotic microenvironment of established KD. GSE145725 (platform GPL16043) served as the validation set, comprising skin samples from 9 KD patients and 10 healthy controls. Additionally, single‐cell dataset 1 GSE181316 (GPL20301) included skin specimens from 4 KD patients and 1 healthy control (excluding 3 normal scar samples). The single‐cell dataset 2 GSE293834 (GPL11154) contains skin samples from 2 KD patients and 2 controls. A total of 720 epigenetic factor‐related genes (ERGs) were acquired from the literature (Table S1) [12]. These genes were defined as epigenetic regulator‐associated based on prior functional annotations, rather than being directly measured for epigenetic marks in our dataset.

2.2. Differential Expression Analysis

Differentially expressed genes (DEGs) between KD and control samples in the training set were identified using the “edgeR” package (v 3.42.4) [13], applying thresholds of p < 0.05 and |log₂ FC| > 0.5. Volcano plots were generated using “ggplot2” package (v 3.4.1) (https://ggplot2.tidyverse.org) to visualize DEGs, whereas heatmaps were created using “pheatmap” package (v 1.0.12) (https://CRAN.R‐project.org/package=pheatmap).

2.3. Identification and Functional Analysis of Candidate Genes

The “VennDiagram” package (v 1.7.3) [14] was applied to identify overlapping genes between Differentially Expressed Genes (DEGs) and ERGs, which were designated as candidate genes. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were executed using “clusterProfiler” package (v 4.2.2) (p < 0.05) [15]. Protein–protein interaction (PPI) networks were constructed by inputting candidate genes into the Search Tool for the Retrieval of Interacting Genes (STRING) with a confidence threshold > 0.4.

2.4. Identification and Validation of Key Genes

Candidate feature genes were screened using three machine learning algorithms to improve selection robustness. First, LASSO logistic regression was implemented “glmnet” package (v 4.1.8) (https://cran.r‐project.org/package=glmnet) with a binomial family. The optimal penalty parameter λ was determined by 10‐fold cross‐validation using the cv.glmnet function, and lambda.min (the λ value that minimized the cross‐validated binomial deviance) was selected. Second, support vector machine recursive feature elimination (SVM‐RFE) was performed using the rfe function from the caret package (v 6.0‐94) [16] with a linear kernel. Feature subset sizes from 1 to 25 were evaluated using 5‐fold cross‐validation, and the subset yielding the highest cross‐validation accuracy was retained. Third, random forest (RF) was applied using the randomForest package (v 4.7‐1.1). An RF model was built based on bootstrap aggregating and recursive feature importance scoring. Five‐fold cross‐validation was performed to determine the optimal feature subset, and the minimum cross‐validation error rate was used to identify important genes. The importance scores of all candidate genes were ranked, and the top 10 important genes were extracted. Overlapping genes from three algorithms were identified using “VennDiagram” package (v 1.7.3) [14] and defined as feature genes.

Feature gene expression levels across both datasets were compared via Wilcoxon test (p < 0.05). Genes with significant differential expression and consistent trends in both datasets were defined as key genes.

2.5. Expression Levels in Skin Tissue and Establishment and Evaluation of a Nomogram

Expression patterns of key genes across different skin tissues were examined using the Bgee database (https://bgee.org/).

A diagnostic nomogram for predicting KD prevalence was developed in the training set using “rms” package (v 6.8.1) (https://hbiostat.org/R/rms/) based on key genes. Calibration curves and receiver operating characteristic (ROC) curves were plotted using “rms” package (v 6.8.1) and “pROC” package (v 1.18.0) [17] to assess the nomogram's predictive performance.

2.6. Gene Set Enrichment Analysis (GSEA)

Functional pathways associated with key genes in KD were investigated through GSEA. The reference gene set “c2.cp.kegg.v7.4.symbols.gmt” was selected from the Molecular Signatures Database (MSigDB, https://www.gsea‐msigdb.org/gsea/msigdb). Spearman correlation analyses between each key gene and all remaining genes were conducted using the “psych” package (v 2.1.6) (https://CRAN.R‐project.org/package=psych) to determine correlation coefficients (cor). Genes were ranked in descending order based on these coefficients. GSEA was performed (|normalized enrichment score [NES]| > 1 and p < 0.05) using the “clusterProfiler” package (v 4.2.2) [15], with visualization via the “enrichplot” package (v 1.18.3) (https://bioconductor.org/packages/enrichplot).

2.7. Immune Infiltration Analysis

Immune infiltration patterns in the training set were evaluated by calculating ssGSEA scores for 67 immune cell types using the Xcell algorithm (v 1.1.0) [18]. Wilcoxon test was employed to compare ssGSEA scores between KD and control samples (p < 0.05), with visualization using “ggplot2” package (v 3.4.1). Spearman correlation analysis examined relationships between differential immune cells and key genes (|cor| > 0.3 and p < 0.05) using “psych” package (v 2.4.6.26).

2.8. RNA Binding Protein Prediction, Establishment of Molecular Regulatory Network, and Compound Prediction

The RBP2GO Database (https://rbp2go.dkfz.de/) was queried to determine RNA‐binding protein status of key genes. To elucidate regulatory mechanisms of key gene expression, TFs were predicted using TRRUST database (https://www.grnpedia.org/trrust/) and Cistrome (http://cistrome.org/db). miRNAs were predicted using StarBase (http://starbase.sysu.edu.cn/), whereas upstream lncRNAs of miRNAs were predicted using miRNet (https://www.mirnet.ca/). TF‐mRNA, mRNA‐miRNA, and miRNA‐lncRNA regulatory networks were subsequently constructed.

Potential therapeutic compounds for KD and their associations with key genes were investigated using the Drug‐Gene Interaction Database (DGIdb, https://www.dgidb.org/) and Comparative Toxicogenomics Database (CTD, https://ctdbase.org/).

2.9. Single Cell Data Quality Control and High‐Variable Gene Screening

For single‐cell dataset, comprehensive quality control (QC) was performed using “Seurat” package (v 5.1.0) (https://satijalab.org/seurat). QC parameters of GSE181316 included 200 < nFeature_RNA < 3000, nCount_RNA < 20 000, and percent_mt < 5%. QC parameters of GSE293834 included 200 < nFeature_RNA < 5000, nCount_RNA < 22 000, and percent_mt < 5%.

After quality control, the two filtered single‐cell datasets were integrated using the Seurat package. The data were standardized using LogNormalize, which normalizes the feature expression measurements of each cell relative to total expression, multiplies them by a scaling factor (default is 10 000), and applies a logarithmic transformation to the results. Using the “vst” method of the “FindVariableFeatures” function in the R package “Seurat,” the number of genes in these cells was calculated based on the relationship between the mean and variance, and the top 2000 highly variable genes (HVGs) were selected.

2.10. Cell Dimension‐Reduction Clustering, Annotation, and Cell Functional Enrichment Analysis

Data scaling was performed using ScaleData function within “Seurat” package (v 5.1.0), followed by principal component analysis via RunPCA function based on the first 2000 HVGs. Then, the RunHarmony function was used to correct the batch effect. JackStrawPlot function determined statistically significant principal components, whereas ElbowPlot function visualized dimensionality reduction results.

Unsupervised clustering was executed using FindClusters and FindNeighbors functions with resolution set to 0.7. Cell clusters were visualized using UMAP generated via RunUMAP function. After cell clustering was completed, the merged single‐cell data were annotated using published KD single‐cell annotation references [1]. Specifically, before creating annotations, first plot an expression bubble plot of the marker gene across each cell cluster. Manually classify cell types based on the expression levels of the marker gene within each cluster, and use the “RenameIdentities” function to map the number of each cluster to their corresponding cell types. Finally, according to the expression levels of key genes in cells, the cells with higher gene expression were selected as potential key cells.

Functional enrichment analysis for each cell type in the merged single‐cell dataset was conducted using the “ReactomeGSA” package (v 1.16.1) [19] (p < 0.05), with visualization via pathways function.

2.11. Cell Communication Analysis and Pseudotiming Analysis

In the merged single‐cell dataset, intercellular interactions were analyzed through ligand‐receptor pairs and molecular interactions using the “CellChat” package (v 1.6.1) [20] (p < 0.05), presenting communication networks and receptor‐ligand interaction probability plots.

Initially, the potential key cells underwent secondary dimensionality reduction and clustering with a resolution of 0.1. Cells were annotated into distinct subgroups. Pseudotime trajectory analysis was performed using the “Monocle” package (v 1.3.1) (https://bioconductor.org/packages/monocle). Single‐cell trajectory plots were constructed. Furthermore, the expression of key genes in different cell differentiation processes was also presented.

2.12. Reverse Transcription‐Quantitative Polymerase Chain Reaction (RT‐qPCR) for mRNA Expression Verification

Five KD skin tissue specimens and five control specimens were collected from the Department of Burns Surgery, The Second Affiliated Hospital of Kunming Medical University. Ethical approval was obtained from the review committee of the Second Affiliated Hospital of Kunming Medical University (PJ‐2023‐151), with written informed consent obtained from all participants. RNA concentration was measured using Nanodrop 2000, reverse transcription was performed with a cDNA synthesis kit (One‐Step gDNA Remover), and primers were synthesized commercially (Table S2). RT‐qPCR experiments used actin as the internal reference gene, with key gene expression calculated via the $2^{-∆∆C_{\mathrm{t}}}$ method. Statistical differences were determined by t‐test (p < 0.05).

2.13. Statistical Analysis

Bioinformatics analyses were executed using R programming language (v 4.2.2). Between‐group differences were evaluated by Wilcoxon test, with p < 0.05 considered statistically significant.

3. Results

3.1. There Were 25 Candidate Genes Associated With Epigenetic Factor

Analysis of the training set revealed 972 DEGs distinguishing KD from normal samples, comprising 486 upregulated and 486 downregulated genes (Figure 1a,b). To pinpoint epigenetic factor‐associated DEGs, the 972 DEGs were cross‐referenced with 720 ERGs, yielding 25 overlapping genes designated as candidates (Figure 1c). Furthermore, GO enrichment analysis was carried out for candidate genes (p < 0.05). A total of 117 results were obtained, and 15 results were shown (Figure 1d, Table S3). Among them, the biological processes included histone modification and chromatin remodeling; the cellular component included histone methyltransferase complex and histone deacetylase complex; molecular function included DNA‐binding transcription factor binding and histone demethylase activity. Moreover, it was shown that both pathways for KEGG enrichment were polycomb repressive complex and ATP‐dependent chromatin remodeling (p < 0.05) (Figure 1e, Table S4). In the PPI network, there were 25 interacting proteins. TP53 had the most interactions with other proteins, such as WAC, SMYD4, and ARID1A (Figure 1f). Overall, the comprehensive analysis has offered a wealth of information that could guide further research efforts aimed at understanding the underlying mechanisms of KD and developing innovative therapeutic strategies.

FIGURE 1.

Screening and enrichment analysis of differentially expressed genes. (a) Volcano plot of DEGs, with red representing upregulation and blue representing downregulation. (b) Heatmap of DEGs, with red indicating higher expression levels and blue indicating lower expression levels. (c) Venn diagram of candidate genes. (d) Bubble plot of GO enrichment results. The vertical axis represents the pathway, and the size of the circle represents the Gene Ratio. (e) Bubble plot of KEGG enrichment analysis results. The vertical axis represents the pathway, and the size of the circle represents the Gene Ratio. (f) PPI network diagram.

3.2. A Total of Two Key Genes That Impact KD Were Discovered

LASSO logistic regression was applied to the 25 candidate genes. The optimal λ value (lambda.min = 0.06897) determined by 10‐fold cross‐validation yielded nine non‐zero coefficient genes as potential feature genes (Figure 2a,b). Subsequently, SVM‐RFE with a linear kernel and 5‐fold cross‐validation identified an optimal subset of five features, achieving the highest cross‐validation accuracy (Figure 2c,d). Furthermore, RF with 5‐fold cross‐validation produced a ranked list of gene importance scores. The top 10 important genes were: HR, SMYD4, KDM4A, ZMYM3, UBR2, ARID1A, TP53, EZH1, HDAC7, and ACTR6 (Figure 2e). Three genes (HR, SMYD4, and KDM4A) appeared in three analyses and were designated as feature genes (Figure 2f).

FIGURE 2.

Machine learning. (a) Coefficient curve of LASSO regression analysis. (b) Cross‐validation curve of LASSO regression analysis. (c) Cross‐validation accuracy rate curve of SVM‐RFE algorithm. (d) Cross‐validation error rate curve of SVM‐RFE algorithm. (e) Gene importance plot from RF feature selection. (f) Venn diagram showing the overlap of feature genes selected by LASSO, SVM‐RFE, and random forest (RF) algorithms. Three genes (HR, SMYD4, KDM4A) were commonly identified by all three methods. Box plots of the analysis of the expression levels of candidate biomarkers in samples from the KD group and the control group in the training set (g) and the validation set (h).

Expression analysis in the training dataset revealed significant differences between KD and control samples for all three feature genes (Figure 2g). However, validation dataset analysis showed that only HR and SMYD4 exhibited statistically significant differential expression between KD and control groups, with expression patterns matching those observed in the training data (elevated expression) (Figure 2h). Consequently, HR and SMYD4 were selected as the key genes for further investigation.

3.3. The Key Genes Were Distributed in the Skin of the Abdomen and Leg Tissues, and Showed an Excellent Diagnostic Ability for KD

According to the expression scores of the key genes in skin tissues from different parts, it was found that both of the two key genes were highly expressed in the skin tissues of the abdomen and legs (Figure 3a).

FIGURE 3.

The expression of key genes in skin tissues and the construction of the Nomogram. (a) The expression of key genes HR and SMYD4 in different human skin tissues. (b) The construction of the Nomogram. Each predictor corresponds to a horizontal line, and different values correspond to different scores. (c) The calibration curve of the Nomogram. (d) The ROC curve of the Nomogram.

A nomogram model with two key genes was constructed to evaluate whether the two key genes could predict the occurrence of KD clinically (Figure 3b). The slope of the calibration curve was close to the diagonal (Figure 3c); the AUC value of the ROC curve was 0.988 (Figure 3d). These results demonstrate the nomogram's good predictive performance for KD.

3.4. Key Genes Participated in the Regulation of KD Through Diverse Pathways and They Were Associated With Immune Cells

GSEA results for the two key genes showed that HR was enriched in the TGF‐β signaling pathway, alpha linolenic acid metabolism, riboflavin metabolism, etc. (Figure 4a); SMYD4 was enriched in protein export, alpha linolenic acid metabolism, RNA degradation, etc. (Figure 4b). Additionally, the common pathways of the two key genes were alpha linolenic acid metabolism and chronic myeloid leukemia.

FIGURE 4.

GSEA enrichment analysis and immune infiltration analysis of key genes. GSEA enrichment result maps of key genes HR (a) and SMYD4 (b). The figure is divided into three parts. The top part is the enrichment score line graph. Each line represents a pathway, and the peak value of each line is the enrichment score of that pathway. The genes before the peak are the core genes under the gene set of that pathway. The second part uses lines to mark the genes located in the gene set. The third part is the rank value distribution of all genes. (c) Immunocyte abundance map of all samples in the KD group and the control group. (d) Differential box plot of the infiltration of each type of immune cell in the KD group samples and the control group samples, *p < 0.05, **p < 0.01. (e) Correlation heat map between differentially expressed immune cells in the KD group samples and the control group samples. (f) Correlation between key genes HR and SMYD4 and differentially expressed immune cells.

The proportional distribution of 67 immune cells in KD and normal samples was presented in Figure 4c. There were five differential immune cells, including dendritic cells (DC), hematopoietic stem cells (HSC), etc. The outcomes revealed that compared with the normal, the KD samples had a higher proportion of platelets and a lower proportion of DC (p < 0.05) (Figure 4d). Among them, megakaryocytes and pro B cells showed the highest positive correlation (cor = 0.41); HSC and platelets showed the most pronounced negative correlation (cor = −0.56) (Figure 4e). Furthermore, HR and SMYD4 had a positive correlation with Platelets and a negative correlation with DC (Figure 4f). These inferential results suggest a potential association between the key genes and the immune microenvironment in KD; however, which warrants further experimental validation before any clinical application.

3.5. Multiple Molecules and Compounds Might Be Related to the Key Genes in KD

The prediction results of RNA‐binding proteins for the key genes showed that SMYD4 had a stronger ability to bind to RNA, with an RBP2GO score of 4.1 (Table S5, Figure S1a). The TF‐mRNA network indicated that both HR and SMYD4 predicted 12 TFs. Among them, EGR, MYC, and EZH2 regulated the expression of HR, and SP1, MAZ, and KAT7 regulated the expression of SMYD4 (Figure S1b).

The results of miRNA‐mRNA, lncRNA‐miRNA showed that there were 225 miRNAs and 156 lncRNAs having regulatory relationships with the key genes. Among them, hsa‐miR‐99b‐5p and hsa‐miR‐98‐5p regulated the expression of HR. hsa‐miR‐98‐5p and hsa‐miR‐93‐5p regulated the expression of SMYD4 (Figure S1c). The lncRNAs ZNF436‐AS1, UBL7‐AS1, CARMN indirectly regulated the expression of the key genes by regulating the miRNAs (Figure S1d). Based on the DGIdb and CTD, the potential compounds of key genes were predicted. Among them, 10 compounds, such as propylthiouracil, estradiol, and triiodothyronine, were predicted for HR. Eight compounds, including acetaminophen, bisphenol A, and pirinixic acid, were predicted for SMYD4. The common compounds of two key genes were acetaminophen and bisphenol A (Figure S1e). These exploratory speculations based on databases may provide potential reference for the treatment research of KD.

3.6. There Were 9 Cell Types That Were Annotated Altogether in the Merged Single‐Cell Dataset

After conducting quality control on the single‐cell datasets GSE181316 and GSE293834, respectively and merging the data, the number of cells was 88 171 and the number of genes was 31 556 (Figure S2a,b). Subsequently, 2000 HVGs were extracted, with the top 10 genes highlighted, including HBB, IGKC, and S100A7 (Figure 5a). After batch correction, there is no significant batch effect among the samples (Figure S2c,d). JackStraw plot analysis revealed highly significant principal components (p < 0.05) (Figure 5b). Based on the Elbow plot inflection point, the first 15 statistically significant principal components were selected for downstream single‐cell analysis (Figure 5c). Cell clustering resulted in 33 distinct clusters (Figure 5d), which were further annotated into 9 cell types, including fibroblasts (FB), smooth muscle cells and pericytes (SMC/PC), keratinocytes (KC), endothelial cells (EC), lymphatic endothelial cells (LEC), T‐cells (TC), macrophages (MAC), schwann cells (SC), and melanocytes (MEL) (Figure 5e). Marker gene expression across cell types was illustrated in a dot plot (Figure 5f). The annotated genes are listed in Table S6. In addition, the key gene HR was more highly expressed in MEL, whereas SMYD4 was more highly expressed in SMC/PC (Figure 5g). Therefore, MEL and SMC/PC were identified as potential key cell types that warrant further investigation.

FIGURE 5.

Single‐cell analysis. (a) Screening of highly variable genes, and highly variable genes are shown in red. (b) JackStraw plot showing the statistical significance of principal components (PCs). (c) PCA scree plot. (d) UMAP plot of cell clustering. (e) Visualized UMAP plot of cell clusters after annotation. (f) Plot of the average expression of marker genes, where the darker the color, the higher the expression level. (g) Expression of key genes in different cells.

3.7. Cells Influenced the Progression of KD Through Various Functional Pathways, Communication Patterns, and Developmental States

Functional enrichment analysis of annotated cells clarified relevant biological pathways, revealing the top 10 most differentially expressed pathways (e.g., biosynthesis of maresin conjugates in tissue regeneration [MCTR] and synthesis of Hepoxilins [HX] and Trioxilins [TrX]) (Figure 6a). Cell communication analysis showed that, compared with the control group, communication between FB and other cells was enhanced in the KD group (Figure 6b,c). Analysis of receptor‐ligand interactions revealed that in KD samples, when MEL acts as the signal sender, cell‐to‐cell communication between MEL and EC may occur via CCL5‐ACKR1. When MEL acts as the signal receiver, communication between SC and MEL may occur via PTN‐NCL (Figure 6d). When SMC/PC acts as the signal sender, communication between SMC/PC and EC may occur via CCL2‐ACKR1. Communication between SMC/PC and MAC may occur via MIF − (CD74 + CXCR4) and MIF − (CD74 + CD44). When SMC/PC acts as the signal recipient, communication between SC and SMC/PC may occur via PTN − NCL and PTN − SDC2 (Figure 6e). Given that fibroblasts are the principal effector cells mediating ECM deposition and fibrosis in KD, we further interrogated the communication probabilities between the two key gene‐enriched cell types (HR‐high melanocytes and SMYD4‐high SMC/PCs) and fibroblasts. The CellChat analysis revealed that, compared to normal skin, KD samples exhibited enhanced signaling from SMC/PCs to fibroblasts predominantly via the FGF (FGF7‐FGFR1 and FGF23‐FGFR1) axes (Figure S3). These results suggest a potential paracrine route by which SMYD4‐expressing cells may indirectly modulate fibroblast behavior.

FIGURE 6.

Single‐cell enrichment analysis and cell–cell communication analysis. (a) Enrichment analysis of annotated cells. Interaction networks among all annotated cells in the healthy group (b) and the KD group (c). Ligand‐receptor interactions between MEL cells (d) and SMC/PC cells (e) and other cell types.

Subclustering of potential key cells revealed that MEL cells were grouped into two subtypes, whereas SMC/PC cells were grouped into seven subtypes (Figure 7a–b, Figure S4). Pseudo‐time series analysis showed that MEL cells follow three differentiation pathways; the deepest blue indicates the earliest differentiating cells, whereas the lightest blue represents the most recently differentiating cells (Figure 7c). Furthermore, MEL exhibits three differentiation states, with 1 representing the earliest differentiation type and 2 representing the latest (Figure 7c). SMC/PC exhibits six differentiation trajectories and eight differentiation states (Figure 7d). SMYD4 expression showed no significant changes during MEL differentiation, whereas HR expression began to rise during the late stages of MEL differentiation (Figure 7e). The expression levels of SMYD4 and HR do not show significant changes during the differentiation of SMC/PC (Figure 7f).

FIGURE 7.

Pseudo temporal analysis of key genes in different cells. (a–b) UMAP plot of dimensionality reduction and clustering analysis based on potential key cell types. (c) MEL cells exhibit pseudo‐temporal changes in subpopulations. (d) SMC/PC cells exhibit pseudo‐temporal changes in subpopulations. (e) Expression levels of key genes HR and SMYD4 at each cell differentiation stage of MELcells. (f) Expression levels of key genes HR and SMYD4 at each cell differentiation stage of SMC/PC cells.

3.8. RT‐qPCR Validation of mRNA Expression Levels of Key Genes

The RT‐qPCR results confirmed the transcriptional upregulation of HR and SMYD4 in KD samples, which is consistent with the bioinformatics predictions (p < 0.05) (Figure 8). This was in accordance with the outcomes of the bioinformatics analysis, indicating that the bioinformatics analyses were reliable.

FIGURE 8.

Analysis of the RT‐qPCR expression of key genes. Analysis of the RT‐qPCR expression levels of the key genes HR and SMYD4.

4. Discussion

KD is characterized by excessive fibroblast proliferation and aberrant accumulation of extracellular matrix (ECM), representing a significant global healthcare burden among skin disorders [21]. Epigenetic dysregulation exists in KD, with unclear molecular targets in KD‐related pathways and unknown mechanisms of epigenetic factor‐related genes [22, 23]. This study integrated KD‐related GEO datasets and existing literature via multiple analyses (differential gene expression, machine learning, functional enrichment, immune cell infiltration, single‐cell communication, pseudotemporal trajectory) to elucidate the pathological roles of two key epigenetic regulator‐associated genes (SMYD4 and HR). Their expression‐based scoring provides novel biomarkers and potential therapeutic targets for early KD detection and treatment, advancing our understanding of KD's pathogenic mechanisms.

The SMYD4 gene encodes a protein with SET and MYND domains that mediate protein–protein interactions and chromatin architectural modification. As a lysine methyltransferase of the SMYD family [24], it exerts histone methyltransferase activity to catalyze H3K4 methylation. Studies show SMYD4 facilitates MYH9 lysine monomethylation and subsequent ubiquitin‐dependent degradation, thereby inhibiting MYH9 binding to the CTNNB1 promoter [25]; it can also form transcriptional regulatory complexes with other factors to modulate gene transcription. A recent study identified SMYD4 as a potential oncogene that binds and activates the promoter of stem cell transcription factor Nanog [26, 27]. The high SMYD4 expression observed in KD in this study suggests its histone methyltransferase activity may globally remodel chromatin accessibility in KD fibroblasts, thereby activating pro‐proliferative and pro‐fibrotic gene programs. Additionally, by suppressing MYH9, SMYD4 may indirectly hyperactivate the Wnt/β‐catenin pathway—a well‐established core pro‐fibrotic pathway in KD [28]. Thus, SMYD4 may represent a novel molecular nexus linking epigenetic modification, key signaling pathways, and aberrant activation of fibroblasts in KD.

The HR gene encodes the Hairless protein, which primarily functions as a transcriptional corepressor for multiple nuclear receptors, an activity critical for maintaining normal skin homeostasis [29] (β2) pathway by downregulating the expression of microRNA‐31 in the skin [30] β signaling cascade serves as a central driver of fibrotic diseases, including KD [31] β signaling. Furthermore, HR has been shown to regulate the hair growth cycle via the Wnt/β‐catenin pathway, another key signaling axis implicated in KD pathogenesis [32]. Thus, HR overexpression may facilitate KD progression through its modulatory effects on these profibrotic pathways.

GSEA results from this study include the TGF‐β signaling pathway, alpha‐linolenic acid metabolism, riboflavin metabolism, protein export, RNA degradation, among others. The discussion will integrate relevant research linking the enriched pathways to the disease. Studies have shown the roles of TGF‐β and Eph‐ephrin signaling pathways in abnormal fibrosis and angiogenesis in KD [33]. Biological antibodies targeting cytokines such as TGF‐β may represent potential future strategies for preventing and treating such scars [34]. Sustained TGF‐β1 signaling drives organ fibrosis, and studies suggest a link between alpha‐linolenic acid metabolism and TGF‐β1 signaling [35]. Severe riboflavin deficiency clinically presents with cheilosis, angular stomatitis, glossitis, seborrheic dermatitis, and severe anemia with erythroid hypoplasia [36]. These pathways do not operate in isolation but form an intricate molecular network. SMYD4 may epigenetically regulate the transcription of key pathway genes (e.g., TGF‐β), whereas the HR may contribute to KD pathogenesis through its modulatory effects on TGF‐β and Wnt/β‐catenin signaling pathways.

In the present study, immune infiltration analysis revealed a significantly increased platelet fraction in KD tissues, whereas the proportions of dendritic cells (DCs), hematopoietic stem cells (HSCs), megakaryocytes, and pro‐B cells were decreased. Notably, the two key epigenetic regulators, HR and SMYD4, were positively correlated with platelets but negatively correlated with DCs. This pattern suggests that epigenetic dysregulation in KD may be linked to a profibrotic immune microenvironment characterized by enhanced platelet‐associated signals and impaired antigen‐presenting cell homeostasis. Given that activated platelets are a major source of profibrotic mediators, including TGF‐β1 and PDGF, their enrichment may contribute to fibroblast activation, ECM deposition, and persistent scar progression [37]. In addition, previous in vitro work has shown that platelet‐rich plasma can modulate scar‐related signaling through TGF‐β1‐dependent negative feedback on CTGF, suggesting that platelet‐derived mediators may exert context‐dependent effects during fibrotic remodeling [38]. A recent mechanistic hypothesis has further proposed that this regulatory process may be influenced by the Piezo1–YAP/TAZ mechanotransduction axis, although direct validation in KD is still lacking [39]. Collectively, our findings raise the possibility that HR and SMYD4 may participate in KD progression by affecting platelet activation‐related transcriptional programs or by indirectly modulating the megakaryocyte–platelet lineage.

The decrease in DCs observed in our analysis may reflect disruption of local immune regulation in KD. Interestingly, prior transcriptomic and immunohistochemical studies have reported increased infiltration of T cells, DCs, and mast cells in KD, together with activation of Th2‐ and JAK/STAT‐related signaling [40, 41]. Therefore, the reduced DC proportion identified here should not be interpreted as simple immune quiescence; instead, it may indicate changes in relative cellular composition, disease stage, or functional polarization of specific DC subsets within the KD microenvironment. This interpretation is also supported by recent evidence showing that Langerin+ DCs can generate latent TGF‐β1, which is locally activated by keratinocyte integrins and may subsequently promote fibroblast activation and matrix stiffening in cutaneous fibrosis [42]. In this context, the negative correlations of HR and SMYD4 with DCs suggest that these epigenetic regulators may contribute to KD pathogenesis by altering chromatin accessibility and transcriptional programs involved in DC differentiation, maturation, or functional maintenance. Although the change in megakaryocytes did not reach statistical significance, their strongest positive correlation with pro‐B cells and the significant negative correlation between HSCs and platelets together imply a potential shift from hematopoietic progenitors toward the megakaryocyte–platelet axis in KD. Overall, our results indicate that the KD immune microenvironment is characterized by platelet enrichment and DC depletion, both of which are closely associated with key epigenetic regulators, and may provide a new framework for understanding fibro‐immune crosstalk in KD.

Epigenetic regulator SMYD4, highly expressed in immune cells or immune signal‐regulated fibroblasts, may maintain this pro‐fibrotic immune microenvironment via chromatin remodeling and regulation of inflammation‐related gene expression. Additionally, immune‐suppressive subsets like myeloid‐derived suppressor cells (MDSCs) are linked to KD [43, 44], indicating KD immune abnormalities may involve both “hyperactivation” and “dysregulation.” Collectively, these results underscore the central role of immune responses in KD and offer a novel perspective for deciphering key genes' functions within the “immune‐fibroblast” pathological interaction axis.

Single‐cell analysis in this study revealed that HR expression is upregulated during the late differentiation stage of MEL, which is consistent with the profibrotic role of MEL in KD. Previous studies have shown that melanocyte activity is higher in KD tissues than in normal scars, and that melanocyte‐derived exosomes activate the TGF‐β/Smad signaling pathway to promote fibroblast proliferation and collagen synthesis [45]. SMYD4 is highly expressed in SMC/PCs, a cell population that may contribute to ECM deposition in KD through phenotypic switching [46]. As a histone methyltransferase, SMYD4 may regulate their phenotypic plasticity via H3K4 methylation [47]. Cell communication analysis revealed that SMC/PCs interact with macrophages through the MIF‐(CD74 + CXCR4) axis, a pathway that may activate inflammatory responses in fibroblasts during fibrosis [48]; moreover, SC cells send signals to MELs via the PTN‐NCL axis, which may stimulate fibroblast proliferation and collagen deposition [49]. However, the single‐cell data were derived from GSE181316 (4 KD samples + 1 control) and GSE293834 (2 KD samples + 2 control). The sample size remains insufficient to fully capture inter‐individual heterogeneity, suggesting that the current results should be considered preliminary and exploratory, warranting validation in larger cohorts. The novelty of this study lies in systematically screening and preliminarily validating, for the first time from the specific perspective of epigenetics‐related genes (ERGs), HR and SMYD4 as key genes potentially associated with epigenetic factors in KD. Although previous studies have identified immune‐related features or diagnostic biomarkers using similar integrative strategies [50, 51, 52, 53], they have not focused on the systemic expression dysregulation of epigenetic factors in KD or their cell‐type‐specific dynamics. We found that SMYD4, a histone methyltransferase, is significantly overexpressed in KD, suggesting that it may activate pro‐fibrotic gene programs through chromatin remodeling; meanwhile, aberrant expression of HR links the TGF‐β2 and Wnt/β‐catenin pathways to KD. More importantly, single‐cell pseudotime analysis revealed the expression dynamics of HR in melanocytes, providing a novel epigenetic perspective for understanding the disease mechanism of KD. Therefore, the value of this study lies not in the analytical pipeline per se, but in introducing HR and SMYD4—two previously underappreciated genes potentially associated with epigenetic factors—into the investigation of KD pathomechanisms, thereby offering new potential molecular targets for therapeutic intervention.

This study systematically identified SMYD4 and HR as pivotal epigenetic factor‐associated genes in KD by integrating bulk/single‐cell transcriptomic profiling with machine learning‐based feature selection. Functional analysis showed they may synergistically drive KD pathogenesis via histone modification/chromatin remodeling and regulation of fibrosis‐related pathways (e.g., TGF‐β). Their correlation with immune microenvironment changes links epigenetic regulation to immune dysfunction. Collectively, this work provides a novel epigenetic framework for elucidating KD pathogenesis and a theoretical basis for developing diagnostic biomarkers and targeted therapies. Several limitations of the present study should be carefully considered when interpreting the findings. First, the limited sample size of the single‐cell RNA sequencing (scRNA‐seq) dataset constrains the statistical power and generalizability of conclusions regarding cell type‐specific expression patterns. Consequently, the single‐cell level results presented herein should be regarded as hypothesis‐generating observations rather than definitive conclusions. Second, the designation of melanocytes, smooth muscle cells, and pericytes as “potentially key cell types” was based primarily on their relatively high expression levels of HR and SMYD4. This does not constitute functional evidence, nor can it completely rule out the possibility of circular reasoning. Future studies employing lineage tracing, conditional gene knockout, or cell‐specific intervention approaches are warranted to directly validate the functional contributions of HR and SMYD4 within these two cell types to the pathogenesis and progression of KD. Furthermore, the RT‐qPCR validation was based on a small sample size derived from a single population, which limits the generalizability of our conclusions. The immune infiltration analysis relied on computational deconvolution of small‐sample bulk transcriptome data, and the drug prediction was based solely on database co‐annotation; both are inferential in nature and do not constitute any clinical therapeutic claim. In future studies, we plan to expand the clinical validation cohort and employ experimental methods such as flow cytometry, immunohistochemistry, and in vitro co‐culture systems to directly validate immune cell proportions, the functions of HR and SMYD4 in specific cell types, and the actual effects of candidate compounds, thereby providing a more robust evidence base for the epigenetic regulatory mechanisms underlying KD.

Author Contributions

Yifei Wang: conceptualization, data curation, validation, visualization, writing – original draft, writing – review and editing. Shuqian Dou: validation, writing – review and editing. Wenjun Liu: visualization, writing – review and editing. Guoxun Yang: conceptualization, supervision, writing – review and editing. Mingkun Dai: conceptualization, project administration, supervision, writing – review and editing. This work currently described has not been published, is not being considered for publication elsewhere, and its publication was approved by all authors.

Funding

This project was supported by the Hospital External Cooperation Research Project of the Second Affiliated Hospital of Kunming Medical University ‐ Nan Revitalization Talent Support Program (grant number 2022dwhz04), National Natural Science Foundation of China (grant number 82460444), and Yunnan Revitalization Talent Support Program (grant number RSC2019MY018).

Ethics Statement

This study was performed in line with the principles of the Declaration of Helsinki and approved by the Ethics Committee of the Second Affiliated Hospital of Kunming Medical University. The approval number and date of approval are as follows: [PJ‐2023‐151].

Consent

All patients provided written informed consent at the time of clinical sample collection for experiments, ensuring that the research process complied with ethical norms and that patients' rights and wishes were fully respected.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Prediction of RNA‐binding proteins, analysis of molecular regulatory networks, and drug prediction. (a) Prediction maps of RNA‐binding domains of key genes HR and SMYD4. (b) Gene‐TF regulatory network. (c) Gene‐miRNA regulatory network. (d) miRNA‐lncRNA regulatory network. (e) Candidate drug‐biomarker relationship network.

Figure S2: Single‐cell data quality control. (a) Before quality control. (b) After quality control. (c) PCA scatter plot based on the original data. (d) PCA scatter plot after batch effect correction using the Harmony algorithm. The distribution of samples/cells is more uniform, and inter‐batch differences are effectively reduced.

Figure S3: Bubble plots showing the expression of marker genes in different subpopulations of MEL cells (a) and SMC/PC cells (b).

Figure S4: Bubble plots showing the expression of marker genes in different subpopulations of MEL cells (a) and SMC/PC cells (b).

Table S1: Epigenetic factor‐related genes.

Table S2: Primer list.

Table S3: GO enrichment analysis.

Table S4: KEGG enrichment analysis.

Table S5: RNA‐binding protein prediction.

Table S6: The marker gene for cell annotation.

Data Availability Statement

The datasets [GSE113619, GSE145725, GSE181316, and GSE293834] for this study can be found in the [GEO] [http://www.ncbi.nlm.nih.gov/geo/].