Identification of Centrosome Duplication‐Related Biomarkers in Hypertrophic Cardiomyopathy Through Integrative Multi‐Omics, Single‐Cell Transcriptomics, and Experimental Validation

Abstract

Background

Hypertrophic cardiomyopathy (HCM) is a prevalent inherited cardiovascular disorder characterized by ventricular wall thickening and myocardial fibrosis. Centrosome duplication‐related gene play critical roles in cell cycle regulation, microtubule organization, and cellular structural homeostasis; however, their mechanistic involvement in HCM remains unclear.

Methods

An integrative multi‐omics strategy was used, incorporating differential expression analysis, weighted gene co‐expression network analysis, Mendelian randomization), and multiple machine‐learning models to identify centrosome duplication‐related genes associated with HCM. Single‐cell RNA sequencing was used to assess cell‐type‐specific expression patterns, followed by in vitro functional assays in cardiomyocytes.

Results

Two candidate biomarkers, HBEGF and VPS8, were significantly associated with HCM across multiple data sets. Single‐cell transcriptomic analysis revealed high VPS8 expression in cardiomyocytes and dendritic cells. Functional assays show that VPS8 knockdown suppressed cardiomyocyte proliferation, increased apoptosis, and reduced the expression of proteins involved in centrosome and microtubule organization, suggesting its involvement in structural maintenance and cell cycle regulation.

Conclusions

This study suggests a potential association between centrosome duplication‐related gene dysregulation and HCM pathogenesis and identifies VPS8 as a key regulator bridging endosomal–lysosomal homeostasis and immune‐related remodeling. VPS8 may represent a candidate biomarker and a potential therapeutic target for early diagnosis and intervention in HCM.

Nonstandard Abbreviations and Acronyms

ANN

Artificial Neural Network

BP

biological process

CDRGs

centrosome duplication‐related genes

DEGs

differentially expressed genes

EGFR

epidermal growth factor receptor

GO

Gene Ontology

GWAS

Genome‐Wide Association Study

HBEGF

heparin‐binding EGF‐like growth factor

HCM

hypertrophic cardiomyopathy

IVs

instrumental variables

MR

Mendelian randomization

PCA

principal component analysis

PI

propidium iodide

QC

quality control

RNA‐seq

RNA sequencing

scRNA‐seq

single‐cell RNA sequencing

shRNA

short hairpin RNA

sh‐VPS8

shRNA targeting VPS8

UMAP

uniform manifold approximation and projection

VPS8

vacuolar protein sorting 8 homolog

WGCNA

Weighted Gene Coexpression Network Analysis

Clinical Perspective.

What Is New?

• By integrating multi‐database analysis, Mendelian randomization validation, and in vitro functional experiments, we first identified that centrosome duplication‐related genes HBEGF and VPS8 are specifically upregulated in hypertrophic cardiomyopathy, and VPS8 affects cardiomyocyte proliferation and apoptosis by regulating centrosomal proteins such as γ‐tubulin, which can serve as potential diagnostic biomarkers for hypertrophic cardiomyopathy.

What Are the Clinical Implications?

• HBEGF and VPS8 provide non‐invasive and highly specific novel diagnostic targets for hypertrophic cardiomyopathy, and their mediated regulatory mechanism of centrosome duplication offers a new direction for the development of precision therapeutic drugs for hypertrophic cardiomyopathy, facilitating the improvement of clinical early diagnosis and treatment decisions.

Hypertrophic cardiomyopathy (HCM) is a hereditary cardiac disorder characterized by ventricular wall hypertrophy, narrowing of the ventricular chamber, impaired left ventricular filling, and reduced diastolic compliance. ¹ , ² , ³ The disease has a high global incidence rate and significant familial clustering. Current evidence suggests that mutations in pathogenic genes of HCM lead to changes in the structure and function of myosin, which is the primary molecular genetic basis of the disease. ³ These changes can impair cardiac function, resulting in symptoms such as difficulty breathing, fatigue, chest pain, and fainting. In severe cases, HCM may progress to heart failure and sudden death. Currently, no curative therapy exists for HCM, and management is mainly symptomatic, focusing on alleviating symptoms, preventing complications, and improving patients' quality of life. However, patients remain at risk for worsening symptoms, continuous deterioration of heart function, and the occurrence of complications. ⁴ , ⁵ Therefore, identifying more effective treatment methods to improve the quality of life and therapeutic effects of patients with HCM has become a major focus in cardiovascular research. Within this framework, identifying biomarkers associated with HCM may provide an essential reference for refining treatment strategies.

The centrosome is a membraneless organelle composed of multiple proteins, primarily responsible for organizing microtubules and establishing cell division polarity. It functions as the core microtubule‐organizing structure within cells. ⁶ , ⁷ Existing research has confirmed that abnormal centrosome replication or dysfunction is closely associated with the occurrence and development of various major diseases, such as cancer, neurodegenerative disorders, and cardiovascular conditions. ⁸ , ⁹ , ¹⁰ Within the cardiovascular system, centrosomes in cardiomyocytes not only contribute to the assembly and maintenance of the cytoskeleton but also influence cardiomyocyte proliferation and differentiation. ⁶ , ¹¹ Studies have shown that impaired centrosome reorganization may underlie infantile dilated cardiomyopathy, and preliminary centrosome‐targeted therapeutic strategies for this condition have been identified. ¹² However, the specific biological functions and molecular mechanisms of centrosome duplication‐related genes (CDRGs) in HCM remain largely unexplored, and research in this area is relatively scarce. Therefore, an in‐depth exploration of the biological roles and molecular mechanisms of CDRGs in HCM is critical for understanding disease pathogenesis and identifying potential new treatment strategies.

RNA sequencing (RNA‐seq) allows for comprehensive profiling of gene transcription within cells or tissues, revealing the molecular mechanisms underlying disease onset and progression. ¹³ , ¹⁴ Mendelian randomization (MR) leverages genetic variants, particularly single‐nucleotide polymorphisms (SNPs), as instrumental variables (IVs) to infer potential causal relationships between exposures and outcomes. The validity of MR is based on the following 3 core assumptions: (1) the IVs must be closely associated with the exposure, (2) must be independent of potential confounders, and (3) must influence outcomes solely through the exposure factors. ¹⁵ Single‐cell RNA sequencing (scRNA‐seq) enables precise measurement of gene expression at the individual cell level, facilitating the analysis of cellular heterogeneity and revealing differential expression patterns across distinct cell types or within the same cell type under varying conditions. ¹⁶ In HCM research, integrating RNA‐seq, MR, and scRNA‐seq provides a multilevel approach—from gene expression profiling to causal inference and, ultimately, to cellular heterogeneity—thereby offering a more comprehensive understanding of the molecular mechanisms underlying disease onset and progression.

This study integrated transcriptomic and scRNA‐seq data from HCM from publicly available databases and applied multiple bioinformatics analyses, including MR, to systematically identify key biomarkers associated with centrosome duplication in HCM. Subsequent in vitro experiments confirmed that these biomarkers play crucial roles in regulating cardiomyocyte proliferation, apoptosis, and centrosome duplication. These findings enhance the mechanistic understanding of HCM pathogenesis, enable early diagnostic applications, and facilitate the exploration of novel therapeutic strategies.

METHODS

In accordance with the Transparency and Openness Promotion Guidelines, we provide the following statements regarding the availability of data, methods, and materials.

Data availability: The transcriptomic data sets (GSE36961, GSE141910, GSE130036, GSE160997, GSE32453, and GSE174691) analyzed in this study are publicly accessible through the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). CDRGs were retrieved from the Molecular Signatures Database (https://www.gsea‐msigdb.org/), and GWAS (Genome‐Wide Association Study) data were obtained from the MiBioGen (https://www.mibiogen.org/) and Integrative Epidemiology Unit OpenGWAS (https://gwas.mrcieu.ac.uk/) databases, all of which are publicly available.

Method availability: Analytical procedures—including differential expression analysis, Weighted Gene Coexpression Network Analysis (WGCNA), MR, and machine‐learning modeling—are described in detail in the Materials and Methods section. The R code scripts used for data processing and statistical analyses are available from the corresponding author upon reasonable request.

Materials availability: The H9c2 cell line used in this study is commercially available, and all reagents are specified in the Materials and Methods section; all materials are accessible to researchers to enable experimental replication.

Ethics statement: All omics and GWAS data sets used in this study were obtained from open‐access databases. The original studies obtained Institutional Review Board approval and documented informed consent from participants. Secondary analysis of deidentified public data was therefore exempt from further ethical review. The H9c2 cell line was obtained from a commercial supplier. No in vivo animal experiments or human studies were conducted, and all in vitro procedures adhered to standard laboratory biosafety guidelines and ethical practices.

Data Source

Transcriptomic data sets GSE36961 (platform: GPL15389), GSE141910 (platform: GPL16791), GSE130036 (platform: GPL20795), GSE160997 (platform: GPL11154), GSE32453 (platform: GPL6104), and GSE174691 (platform: GPL21697) were obtained from the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). GSE36961 served as the training set and comprised cardiac tissue samples from 106 patients with HCM and 39 normal controls. The validation sets included GSE141910, GSE130036, GSE160997, and GSE32453. Analysis of myocardial tissue samples in the validation sets yielded HCM‐to‐control ratios of 28:166 in GSE141910, 28:9 in GSE130036, 18:5 in GSE160997, and 8:5 in GSE32453. The scRNA‐seq data set GSE174691 contained nine cardiac tissue samples from patients with HCM. Furthermore, Gene Ontology (GO) biological processes associated with centrosome duplication and negative regulation were retrieved from Molecular Signatures Database (https://www.gsea‐msigdb.org/), yielding 76 CDRGs (Table S1).

Differential Expression Analysis

To identify differentially expressed genes (DEGs) between the HCM and control groups in the GSE36961 data set, the limma package (v3.54.0) ¹⁷ was applied with thresholds of |log₂ fold change|>0.5 and uncorrected P<0.05. Considering the exploratory nature of this study and the subsequent multi‐layer validation strategy (including independent data sets, machine learning, and in vitro experiments), an uncorrected P value threshold was initially used to prioritize DEGs, which were then stringently filtered through downstream analyses to minimize false positives. Visualization of DEGs was conducted using volcano plots generated using the ggplot2 package (v3.4.1) ¹⁸ and heatmaps constructed with the ComplexHeatmap package (v2.14.0). ¹⁹

WGCNA

The single‐sample gene set enrichment analysis, implemented using the GSVA package (v1.42.0), ²⁰ was applied to compute the CDRG scores for all samples in the GSE36961 data set. Subsequently, the Wilcoxon rank‐sum test was used to assess differences in CDRG scores between the HCM and control groups (P<0.05). To identify the modular genes most strongly associated with CDRGs, WGCNA was conducted on HCM samples from GSE36961 using the WGCNA package (v1.70.3). ²¹ Following this, the samples were analyzed using clustering. A sample clustering tree was constructed, and outlier samples were excluded. To maximize the scale‐free topological fit of gene interactions, a soft threshold (power) was selected for constructing the coexpression network based on a scale‐free fit index (signed R ²) exceeding 0.85 and a mean connectivity approaching zero. The filtered expression matrix was used to construct the WGCNA network, with a minimum module size of 100 genes and a module‐merging threshold (merge cut height) set to 0.25. Coexpression modules were subsequently identified and visualized using hierarchical clustering. CDRG scores were treated as phenotypic traits, and Pearson correlation analysis was conducted to assess the correlation matrix between CDRG scores and coexpression modules (|correlation coefficients (cor)|>0.30, P<0.05). Genes exhibiting the strongest positive or negative correlations with CDRG scores within the relevant modules were identified as key module genes.

Identification and Functional Analysis of Candidate Genes

The intersection of DEGs and key module genes was implemented using the VennDiagram package (v1.7.1) ²² to identify candidate genes associated with CDRGs in HCM. Functional annotations of these candidate genes were subsequently performed using the clusterProfiler package (v4.2.2), ²³ which facilitated GO and Kyoto Encyclopedia of Genes and Genomes enrichment analyses of the candidate genes, with P<0.05 considered statistically significant. The GO terms were ranked in ascending order of P values, and the top 5 significantly enriched pathways within the biological process (BP), cellular component, and molecular function categories were visualized using the enrichplot package (v1.18.3). ²⁴ Similarly, significantly enriched Kyoto Encyclopedia of Genes and Genomes pathways were visualized using the GOplot package (v1.0.2). ²⁵ To examine interactions among candidate genes at the protein level, a protein–protein interaction network was constructed using the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (https://www.string‐db.org) with an interaction score greater than 0.15. The resulting protein–protein interaction network was visualized using Cytoscape software (v3.8.2). ²⁶

Screening of IVs

MR analysis is based on 3 core assumptions: (1) IVs are strongly associated with the exposure factors; (2) IVs are independent of confounding factors; and (3) IVs influence the outcome exclusively through exposure factors and not through alternative pathways. Expression quantitative trait loci (eQTL) data for candidate genes were obtained from the Integrative Epidemiology Unit OpenGWAS database (https://gwas.mrcieu.ac.uk/), which provides eQTL summary statistics derived from European‐ancestry populations, and were treated as exposure factors. Concurrently, the HCM‐related data set (ebi‐a‐GCST90018861) from the Integrative Epidemiology Unit OpenGWAS database, comprising 24 199 797 SNPs from 489 727 European samples (cases:controls=507:489220), was used as the outcome variable. Exposure factors were initially read, and IVs—including SNPs and insertion/deletion variants—were extracted using the extract_instruments function of the TwoSampleMR package (v0.56). ²⁷ IVs significantly associated with exposure factors were selected using a threshold of P<5×10⁻⁶. To mitigate potential bias due to linkage disequilibrium, IVs (SNPs) with linkage disequilibrium were removed using the parameters clump=TRUE, r ²=0.001, and kb=10. The strength of the IVs was assessed using F‐statistic tests, calculated as follows:

$$F=\frac{R^2\times \left(n-k-1\right)}{\mathrm{k}\times \left(1-R^2\right)},$$

where R ² is the squared correlation coefficient between the IVs and the exposure factor, n is the sample size, and k is the number of IVs. IVs with F‐statistics below 10 were excluded from the analysis. Finally, the harmonise_data function of the TwoSampleMR package (v 0.56) was used to unify the effect alleles and effect sizes, thereby ensuring proper matching between exposure factors, IVs, and outcome data.

MR Analysis

Potential causal relationships between candidate genes and HCM were assessed using the mr function of the TwoSampleMR package (v0.56) with 5 complementary algorithms—MR Egger, ²⁸ weighted median, ²⁹ inverse variance weighted, ³⁰ simple mode, ³¹ and weighted mode. ³² The inverse variance weighted method served as the primary method for evaluating the results. Candidate genes were considered causally associated with HCM if the inverse variance weighted results were significant (P<0.05). Candidate genes with odds ratio (OR)>1 were classified as risk factors for HCM, whereas those with OR<1 were regarded as protective factors for HCM. Furthermore, MR results were visualized using scatter and forest plots, and funnel plots were used to assess adherence to Mendel's second law. Meanwhile, reliability and robustness were evaluated through heterogeneity tests (P>0.05), horizontal pleiotropy tests (P>0.05), and leave‐one‐out sensitivity analyses. Additionally, the Steiger directionality test was conducted using the directionality_test function to confirm the causal relationship between the candidate genes and HCM, with a correct direction of causality=TRUE and P<0.05 indicating a unidirectional causal relationship. Candidate genes with a significant causal correlation with HCM through these analyses were subsequently designated as candidate feature genes.

Identification of Biomarkers Through Machine Learning, Expression Validation, and Receiver Operating Characteristic (ROC) Analysis

In the GSE36961 data set, 11 machine‐learning algorithms were applied to generate 101 algorithmic combinations. The algorithms included random forest, Least Absolute Shrinkage and Selection Operator, ridge regression, elastic net, stepwise generalized linear model, support vector machine, glmboost, linear discriminant analysis, gradient boosting machine, eXtreme gradient boosting, and naïve Bayes. Candidate feature genes were used to train diagnostic models across all 101 algorithm combinations with 10‐fold cross‐validation in GSE36961. The performance of each of the 101 algorithmic models was evaluated by calculating the area under the curve (AUC) in both GSE36961 and GSE141910. The optimal combined model was defined as one with an AUC ≥0.7 in both data sets and possessing the highest average AUC. Genes derived from this optimal combination model were designated as feature genes. For GSE36961 and GSE141910, the ROC curves were plotted using the pROC package (v1.18.0) ³³ to assess the diagnostic value of the optimal model. Additionally, the expression levels of feature genes between HCM and control groups in the GSE36961 and GSE141910 data sets were compared using the Wilcoxon rank‐sum test through the rstatix package (v0.7.2). ³⁴ Feature genes that exhibited significant intergroup‐differential expression (P<0.05) and consistent expression trends across both data sets were defined as candidate biomarkers. The diagnostic value of these candidate biomarkers for HCM was further evaluated using ROC analysis with the pROC package (v1.18.0), and candidate biomarkers with an AUC ≥0.7 in both GSE36961 and GSE141910 were defined as biomarkers. Subsequently, expression trends and ROC analyses for the validated biomarkers were further examined in data sets GSE130036, GSE160997, and GSE32453.

Construction and Assessment of the Artificial Neural Network (ANN) Model

To further evaluate the predictive ability of biomarkers for HCM, an ANN model was constructed using the neuralnet package (v1.44.2) ³⁵ in the GSE36961 data set. The biomarker expression data were first preprocessed and normalized. Min–max scaling was applied to transform the data into the [0,1] range before training the neural network. Before calculation, the data values were standardized by determining their maximum and minimum values, and the number of hidden layers was established as 3. While there were no fixed rules for determining the number of hidden layers and neurons, the number of neurons in each hidden layer was constrained to fall between the sizes of the input and output layers, and 3 neurons were selected for the ANN model. Subsequently, a confusion matrix was generated using GSE36961 and GSE141910 to assess the predictive accuracy of the ANN model. In parallel, the diagnostic capability of the ANN model for HCM was assessed using ROC curves generated with the pROC package (v1.18.0) in both GSE36961 and GSE141910, and the AUC was calculated to quantify the predictive performance of the model. An AUC greater than 0.7 indicates a favorable predictive ability.

scRNA‐Seq Data Processing

The 10× scRNA‐seq data set GSE174691 was processed for quality control (QC) using the Seurat package (v4.1.0). ³⁶ This data set includes nine patient samples, and detailed patient characteristics, including age and sex, are provided in Table S2. Notably, this data set lacked single‐cell samples from normal controls, which is an inherent limitation of the data. ³⁷ , ³⁸ Cells with 200 genes and genes covered in 3 cells were excluded from subsequent analyses. Furthermore, the percentage of mitochondrial gene expression (percent.mt) was calculated using the PercentageFeatureSet function. Before QC, the distribution of nFeature_RNA (number of detected genes), nCount_RNA (total RNA counts per cell), and percent.mt were visualized as violin plots using the VlnPlot function. Subsequently, QC of scRNA‐seq data was performed with the following criteria: 200<nFeature_RNA<1000, nCount_RNA<2000, and percent.mt<20%. Following QC, the data were normalized using the NormalizeData function. A total of 2000 hypervariable genes were identified using the FindVariableFeatures function. The result was visualized using the LabelPoints function, and the top 10 genes were labeled. The data were then normalized using the ScaleData function. Principal component analysis (PCA) was performed on the 2000 hypervariable genes using the RunPCA function to reduce dimensionality. The top 50 statistically significant principal components (P<0.05) were identified using the JackStrawPlot function. Next, an elbow plot was generated using the ElbowPlot function to determine the optimal number of principal components for downstream dimensionality reduction. An unsupervised clustering analysis was conducted with the FindNeighbors and FindClusters functions based on the selected principal components, using a resolution parameter of 0.4 to determine the number of cell clusters. The uniform manifold approximation and projection method was used to visualize the cell clusters. To further understand the cell types in the scRNA‐seq data from GSE174691, cell‐type annotation of the cell clusters was performed using marker genes from the SingleR package (v1.831) ³⁹ supported by relevant literature. ⁴⁰ Finally, the distribution and expression levels of candidate biomarkers across annotated cell types were further analyzed and visualized using uniform manifold approximation and projection and bubble plots in the Seurat package (v4.1.0).

Cell Culture and Transfection

In this study, the rat cardiomyocyte cell line H9c2 was used for functional validation due to its stable proliferative capacity and retention of cardiomyocyte‐like phenotypes, making it a classic model for the preliminary exploration of gene functions. ⁴¹ , ⁴² H9c2 cell lines were cultured in Dulbecco's modified Eagle medium supplemented with 10% fetal bovine serum at 37 °C, containing 5% CO₂. ⁴³ Cells were transfected with short hairpin RNA (shRNA) targeting VPS8 (sh‐VPS8) or a negative control shRNA) using Lipofectamine 2000 reagent to establish gene knockdown models. ⁴⁴

Cell Proliferation Assay

Cell proliferation was assessed using a Cell Counting Kit‐8 (CCK‐8) assay. ⁴⁵ Following transfection, treated cells were seeded into 96‐well plates at an appropriate density. At 0‐, 24‐, 48‐, and 72‐hour post‐transfection, the CCK‐8 reagent was added to each well. Absorbance was measured at 450 nm using a microplate reader to assess cell proliferative activity.

Apoptosis Analysis

Cell apoptosis was assessed using flow cytometry. ⁴⁶ At 48‐hour post‐transfection, cells were collected, washed, double‐stained with Annexin V–fluorescein isothiocyanate and propidium iodide (PI), and quantitatively analyzed using flow cytometry to determine the apoptosis rate in each group.

Immunofluorescence Staining

Centrosome‐related proteins were examined using immunofluorescence microscopy. ⁴⁷ Following transfection, cells were fixed with 4% paraformaldehyde, permeabilized with Triton X‐100, and blocked with bovine serum albumin. Cells were then sequentially incubated with primary antibodies targeting centrosomal proteins (including γ‐tubulin and Centrin‐2) and corresponding Cy3‐ or fluorescein isothiocyanate‐conjugated secondary antibodies. Nuclei were counterstained with Hoechst 33342, and images were captured using a fluorescence microscope.

RESULTS

Identification of 800 DEGs and 759 Key Module Genes

Differential expression analysis identified 800 DEGs between the HCM and control groups, including 311 upregulated genes and 489 downregulated genes in the HCM group (Figure 1A and 1B). Furthermore, CDRG scores were markedly elevated in the HCM group relative to the control group (P<0.001) (Figure 1C). Subsequently, a WGCNA network was constructed using HCM samples from the GSE36961 data set. The sample clustering tree showed no outlier samples, indicating that all HCM samples were included in the network construction (Figure 1D). A soft‐thresholding power of 7 was identified, with a signed R ² value exceeding 0.85 and a mean connectivity of approximately zero (Figure 1E). Hierarchical clustering based on this network identified 17 coexpression modules (Figure 1F). Additionally, correlation analysis between gene modules and CDRG scores revealed that the MEgreenyellow module exhibited the highest positive correlation with CDRG scores (cor=0.50, P<0.001), whereas the MEblack module showed the strongest negative correlation (cor=−0.47, P<0.001) (Figure 1G). The MEgreenyellow module contained 272 genes, and the MEblack module contained 487 genes, resulting in a total of 759 key module genes.

Figure 1. Differential gene expression and Weighted Gene Coexpression Network Analysis (WGCNA) in hypertrophic cardiomyopathy (HCM).

A and B, Volcano and heat maps of the distribution of differentially expressed genes (DEGs) between HCM and Control. C, Differences in centrosome duplication‐related genes (CDRGs) scores in the HCM and Control groups. D, Sample level clustering. E, Soft threshold filtering. F, Co‐expression modules identified. G, Heatmap of correlations between module eigengenes and CDRG scores.

Identification of 34 Candidate Genes and Exploration of their Functions

Intersecting the 800 DEGs with the 759 key module genes yielded 34 candidate genes (Figure 2A). Subsequently, functional enrichment analyses were conducted to provide preliminary insights into the signaling pathways involving these candidate genes. GO enrichment analysis identified 286 significantly enriched GO terms, including 218 BPs, 31 cellular components, and 37 molecular functions (P<0.05) (Table S3). Among the top 5 BPs, terms such as “regulation of sodium ion transport” and “sodium ion transport” were prominent (Figure 2B; Table S3), indicating that the candidate genes may be implicated in the regulation of ionic homeostasis in cells and organisms. The top enriched cellular component terms were primarily related to “bicellular tight junction” and “tight junction” (Figure 2B; Table S3). Meanwhile, the most significantly enriched molecular functions included “ATPase‐coupled cation transmembrane transporter activity” and “ATPase‐coupled transmembrane transporter activity” (Figure 2B; Table S3), suggesting that candidate genes played crucial roles in the transmembrane transport of substances and the stability of the intracellular environment. Additionally, Kyoto Encyclopedia of Genes and Genomes enrichment analysis of the 34 candidate genes revealed enrichment in six pathways, including “cytoskeleton in muscle cells” and “phagosome” (P<0.05) (Figure 2C; Table S4). The results demonstrated that candidate genes significantly contribute to muscle‐related diseases as well as immune and inflammatory responses. Additionally, a protein–protein interaction network was constructed based on the 34 candidate genes using the Search Tool for the Retrieval of Interacting Genes/Proteins database with an interaction score threshold >0.15. Of these, 21 candidate genes with known interactions formed a network comprising 22 edges (Figure 2D). Within this network, SEC61A1, SOCS2, and XBP1 showed multiple interactions with other candidate genes. The remaining 13 candidate genes had no reported interactions at the selected confidence threshold and were therefore excluded from the network.

Figure 2. Identification of 34 candidate genes and exploration of their functions.

A, Candidate gene identification. B, Circle plots of Gene Ontology (GO) enrichment results for candidate genes. C, Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment results and chord diagrams of candidate genes. D, Protein–protein interaction plot of candidate genes. DEGs indicates differentially expressed genes.

Identification of 8 Candidate Feature Genes through MR Analysis

MR analysis was performed using 34 candidate genes as exposure factors and HCM as the outcome variable. Based on the inverse variance weighted method, 14 genes were identified as significantly associated with HCM (P<0.05) (Table 1), of which 6 were protective factors (OR<1, eg, HBEGF and VPS8) and 8 were risk factors (OR>1, eg, PAFAH2 and CDC42EP2). The MR analysis yielded robust causal estimates, as illustrated by scatter plots showing individual SNP effects (Figure 3A) and forest plots showing the overall causal estimates (Figure 3B). Funnel plots showed symmetrical SNP distributions, supporting the validity of the MR analysis (Figure 3C). Sensitivity analyses further refined the candidate set. Cochran's Q test revealed heterogeneity for CDC42EP2 (P < 0.05), leaving 13 genes without significant heterogeneity (Table 2). The horizontal pleiotropy test excluded 5 genes (MYH7B, KLHL24, PAFAH2, CDC42EP2, and B4GALT5) due to significant pleiotropy (P<0.05), resulting in 9 genes with no evidence of horizontal pleiotropy (Table 3). Leave‐one‐out analysis confirmed the robustness of the results, as no single SNP disproportionately influenced the overall estimates (Figure 3D). The Steiger directionality test further confirmed unidirectional causal relationships for 8 genes (correct_causal_direction=TRUE, P<0.05), while XBP1 was excluded due to an indeterminate result (Table 4). Collectively, the 8 genes that passed all sensitivity analyses—WDR62, PLA2G15, HBEGF, MFN2, AZIN1, VPS8, SLMAP, and RNF103—were designated as candidate feature genes.

Table 1.

Mendelian randomization (MR) estimates for the causal effects of candidate genes on Hypertrophic cardiomyopathy (HCM)

id.exposure	id.outcome	Outcome	Exposure	Method	nsnp	b	SE	pval	lo_ci	up_ci	or	or_lci95	or_uci95	SYMBOL	ENSEMBL
eqtl‐a‐ENSG00000075702	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000075702	Inverse variance weighted	10	0.3103	0.1267535	0.0143628	0.0618631	0.558737	1.3638343	1.0638167	1.7484627	WDR62	ENSG00000075702
eqtl‐a‐ENSG00000078814	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000078814	Inverse variance weighted	104	0.3121858	0.0492985	2.41E‐10	0.2155607	0.408811	1.3664086	1.2405573	1.5050272	MYH7B	ENSG00000078814
eqtl‐a‐ENSG00000100219	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000100219	Inverse variance weighted	165	−0.062495	0.0121888	2.94E‐07	−0.086385	−0.038605	0.939418	0.9172412	0.9621311	XBP1	ENSG00000100219
eqtl‐a‐ENSG00000103066	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000103066	Inverse variance weighted	59	0.643921	0.0835306	1.27E‐14	0.4802011	0.8076409	1.9039315	1.6163994	2.2426112	PLA2G15	ENSG00000103066
eqtl‐a‐ENSG00000113070	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000113070	Inverse variance weighted	33	−0.446427	0.1068324	2.93E‐05	−0.655818	−0.237035	0.6399108	0.5190173	0.7889637	HBEGF	ENSG00000113070
eqtl‐a‐ENSG00000114796	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000114796	Inverse variance weighted	48	0.304806	0.0384758	2.34E‐15	0.2293934	0.3802186	1.3563618	1.2578367	1.4626042	KLHL24	ENSG00000114796
eqtl‐a‐ENSG00000116688	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000116688	Inverse variance weighted	69	−0.123201	0.0159287	1.04E‐14	−0.154421	−0.09198	0.8840863	0.8569113	0.9121231	MFN2	ENSG00000116688
eqtl‐a‐ENSG00000149798	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000149798	Inverse variance weighted	54	0.2569964	0.0706212	0.0002736	0.1185789	0.3954139	1.2930405	1.1258957	1.4849987	CDC42EP2	ENSG00000149798
eqtl‐a‐ENSG00000155096	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000155096	Inverse variance weighted	31	−0.099318	0.0379679	0.0089007	−0.173735	−0.024901	0.9054547	0.8405195	0.9754064	AZIN1	ENSG00000155096
eqtl‐a‐ENSG00000156931	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000156931	Inverse variance weighted	86	−0.120372	0.0292375	3.84E‐05	−0.177678	−0.063067	0.8865902	0.837212	0.9388808	VPS8	ENSG00000156931
eqtl‐a‐ENSG00000158006	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158006	Inverse variance weighted	66	0.1832746	0.0372575	8.69E‐07	0.1102499	0.2562993	1.2011442	1.116557	1.2921394	PAFAH2	ENSG00000158006
eqtl‐a‐ENSG00000158470	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158470	Inverse variance weighted	32	−0.387009	0.0945264	4.24E‐05	−0.57228	−0.201737	0.6790853	0.5642373	0.81731	B4GALT5	ENSG00000158470
eqtl‐a‐ENSG00000163681	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000163681	Inverse variance weighted	52	0.36483	0.0703139	2.12E‐07	0.2270149	0.5026452	1.4402692	1.2548485	1.6530882	SLMAP	ENSG00000163681
eqtl‐a‐ENSG00000239305	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000239305	Inverse variance weighted	42	0.4768677	0.0751263	2.19E‐10	0.3296201	0.6241153	1.6110203	1.3904398	1.8665939	RNF103	ENSG00000239305

The inverse variance weighted method was used as the primary analysis. b represents the estimated effect size (log odds ratio); se, standard error; Pval, P value; lo_ci and up_ci, lower and upper bounds of the 95% CI for b; or_lci95 and or_uci95, lower and upper bounds of the 95% CI for the odds ratio.ENSEMBL indicates Ensembl gene identifier; OR, odds ratio; and SYMBOL, gene symbol.

Figure 3. Mendelian randomization (MR) analysis identifies candidate feature genes associated with HCM.

A, Scatter plots illustrating the causal effects of individual single‐nucleotide polymorphisms (SNPs) on HCM. B, Forest plots display the overall causal estimates from the inverse variance weighted method for each candidate feature gene. C, Funnel plots show symmetrical distribution of SNPs, supporting the validity of the MR analysis. D, Leave‐one‐out sensitivity analysis demonstrates that no single SNP disproportionately influences the overall MR estimate.

Table 2.

Heterogeneity tests for instrumental variables (IVs) in MR analysis

	id.exposure	id.outcome	Outcome	Exposure	Method	Q	Q_df	Q_pval	SYMBOL	ENSEMBL
1	eqtl‐a‐ENSG00000075702	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000075702	Inverse variance weighted	4.1486517	9	0.9013589	WDR62	ENSG00000075702
2	eqtl‐a‐ENSG00000078814	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000078814	Inverse variance weighted	93.816786	103	0.730085	MYH7B	ENSG00000078814
3	eqtl‐a‐ENSG00000100219	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000100219	Inverse variance weighted	63.029984	164	1	XBP1	ENSG00000100219
4	eqtl‐a‐ENSG00000103066	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000103066	Inverse variance weighted	24.773409	58	0.9999607	PLA2G15	ENSG00000103066
5	eqtl‐a‐ENSG00000113070	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000113070	Inverse variance weighted	15.045177	32	0.9952626	HBEGF	ENSG00000113070
6	eqtl‐a‐ENSG00000114796	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000114796	Inverse variance weighted	48.010171	47	0.4316548	KLHL24	ENSG00000114796
7	eqtl‐a‐ENSG00000116688	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000116688	Inverse variance weighted	65.843425	68	0.5515421	MFN2	ENSG00000116688
8	eqtl‐a‐ENSG00000149798	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000149798	Inverse variance weighted	71.217295	53	0.0481812	CDC42EP2	ENSG00000149798
9	eqtl‐a‐ENSG00000155096	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000155096	Inverse variance weighted	7.1241553	30	0.9999948	AZIN1	ENSG00000155096
10	eqtl‐a‐ENSG00000156931	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000156931	Inverse variance weighted	70.745825	85	0.8663327	VPS8	ENSG00000156931
11	eqtl‐a‐ENSG00000158006	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158006	Inverse variance weighted	61.930927	65	0.5850332	PAFAH2	ENSG00000158006
12	eqtl‐a‐ENSG00000158470	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158470	Inverse variance weighted	26.898423	31	0.6772661	B4GALT5	ENSG00000158470
13	eqtl‐a‐ENSG00000163681	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000163681	Inverse variance weighted	31.138122	51	0.9872554	SLMAP	ENSG00000163681
14	eqtl‐a‐ENSG00000239305	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000239305	Inverse variance weighted	8.56499	41	1	RNF103	ENSG00000239305

Q indicates Cochran's Q statistic for heterogeneity; Q_df, degrees of freedom for the Q statistic; and Q_pval, P value for the Q statistic (values >0.05 indicate no significant heterogeneity).

Table 3.

Horizontal pleiotropy tests using the MR‐Egger method

	id.exposure	id.outcome	Outcome	Exposure	egger_intercept	SE	pval	SYMBOL	ENSEMBL
1	eqtl‐a‐ENSG00000075702	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000075702	0.337712906	0.240325126	0.197574532	WDR62	ENSG00000075702
2	eqtl‐a‐ENSG00000078814	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000078814	−0.053100326	0.022910669	0.022462184	MYH7B	ENSG00000078814
3	eqtl‐a‐ENSG00000100219	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000100219	−0.004618279	0.014568274	0.751642172	XBP1	ENSG00000100219
4	eqtl‐a‐ENSG00000103066	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000103066	−0.011042171	0.049228653	0.823323059	PLA2G15	ENSG00000103066
5	eqtl‐a‐ENSG00000113070	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000113070	0.018727205	0.044955742	0.679863433	HBEGF	ENSG00000113070
6	eqtl‐a‐ENSG00000114796	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000114796	0.067135497	0.029682293	0.02847977	KLHL24	ENSG00000114796
7	eqtl‐a‐ENSG00000116688	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000116688	−0.001058094	0.016150434	0.947959119	MFN2	ENSG00000116688
8	eqtl‐a‐ENSG00000155096	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000155096	0.053097046	0.033335476	0.122045888	AZIN1	ENSG00000155096
9	eqtl‐a‐ENSG00000156931	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000156931	0.019783982	0.015401109	0.202469816	VPS8	ENSG00000156931
10	eqtl‐a‐ENSG00000158006	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158006	−0.044875701	0.016598395	0.00877508	PAFAH2	ENSG00000158006
11	eqtl‐a‐ENSG00000158470	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000158470	0.112223982	0.041093202	0.010471987	B4GALT5	ENSG00000158470
12	eqtl‐a‐ENSG00000163681	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000163681	−0.05152955	0.029255673	0.084292802	SLMAP	ENSG00000163681
13	eqtl‐a‐ENSG00000239305	ebi‐a‐GCST90018861	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000239305	−0.036370046	0.033980526	0.290891334	RNF103	ENSG00000239305

egger_intercept, the intercept term from the MR‐Egger regression (a non‐zero intercept indicates directional horizontal pleiotropy); SE, standard error of the intercept; Pval, P value for the intercept (values >0.05 indicate no significant horizontal pleiotropy).

Table 4.

Steiger Directionality Test for Causal Direction Between Candidate Gene Expression and HCM

	id.exposure	id.outcome	Exposure	Outcome	snp_r2.exposure	snp_r2.outcome	correct_causal_direction	steiger_pval	SYMBOL	ENSEMBL
1	eqtl‐a‐ENSG00000075702	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000075702	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.062042992	2.07E‐05	TRUE	1.30E‐189	WDR62	ENSG00000075702
2	eqtl‐a‐ENSG00000100219	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000100219	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	2.863003094	0.000182376	TRUE	NA	XBP1	ENSG00000100219
3	eqtl‐a‐ENSG00000103066	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000103066	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.104132151	0.000171906	TRUE	0	PLA2G15	ENSG00000103066
4	eqtl‐a‐ENSG00000113070	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000113070	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.049404088	6.64E‐05	TRUE	8.29E‐301	HBEGF	ENSG00000113070
5	eqtl‐a‐ENSG00000116688	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000116688	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.884906562	0.000256753	TRUE	0	MFN2	ENSG00000116688
6	eqtl‐a‐ENSG00000155096	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000155096	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.374094892	2.85E‐05	TRUE	0	AZIN1	ENSG00000155096
7	eqtl‐a‐ENSG00000156931	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000156931	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.871290996	0.0001798	TRUE	0	VPS8	ENSG00000156931
8	eqtl‐a‐ENSG00000163681	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000163681	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.113132417	0.000118549	TRUE	0	SLMAP	ENSG00000163681
9	eqtl‐a‐ENSG00000239305	ebi‐a‐GCST90018861	|| id:eqtl‐a‐ENSG00000239305	Hypertrophic cardiomyopathy || id:ebi‐a‐GCST90018861	0.117478409	9.98E‐05	TRUE	0	RNF103	ENSG00000239305

snp_r2.exposure, the proportion of variance in the exposure (gene expression) explained by the IVs (R²); snp_r2.outcome, the proportion of variance in the outcome (HCM) explained by the IVs; correct_causal_direction, logical indicator (TRUE/FALSE) of whether the assumed causal direction (gene expression→HCM) is correct according to the Steiger test; steiger_pval, P value from the Steiger test (values <0.05 indicate the causal direction is correctly specified).

HBEGF and VPS8 Were Identified as Biomarkers with Diagnostic Potential

To further identify biomarkers associated with HCM and centrosome duplication, the 8 candidate feature genes were incorporated into an integrative machine‐learning model. Across 101 predictive model combinations in GSE36961 and GSE141910, the Lasso+glmBoost model emerged as the optimal combined model, achieving an average AUC of 0.868 in both data sets (Figure 4A). This model identified WDR62, PLA2G15, HBEGF, MFN2, AZIN1, VPS8, and RNF103 as feature genes, while SLMAP was not included. Specifically, the AUC values of the Lasso+glmBoost model in the ROC curve were 0.9821 (95% CI: 0.9578–0.9971) in GSE36961 and 0.7537 (95% CI: 0.6396–0.8524) in GSE141910 (Figure 4B and 4C). Moreover, the findings revealed a notable elevation in the expression levels of PLA2G15, HBEGF, MFN2, AZIN1, VPS8, and RNF103 in the HCM group (P<0.05), with a consistent pattern of expression observed across GSE36961 and GSE141910 (Figure 4D and 4E). Notably, HBEGF and VPS8 individually showed AUC values greater than 0.7 (Figure 4F and 4G), indicating their satisfactory diagnostic potential for HCM. Validation in additional cohorts confirmed the diagnostic potential of HBEGF and VPS8. In GSE130036, HBEGF expression was significantly higher in the HCM group than in controls (P<0.0001), with an AUC>0.9 (Figure S1A and S1D). In GSE160997, HBEGF showed a non‐significant upward trend in the HCM group, while VPS8 was significantly higher in the HCM group (P<0.05); both achieved AUC>0.7 (Figure S1B and S1E). In the small exploratory cohort GSE32453 (n=13), HBEGF remained significantly higher in the HCM group than in controls (P<0.05), whereas VPS8 showed a non‐significant trend toward higher expression in the HCM group, likely due to limited statistical power; nevertheless, both biomarkers yielded AUC>0.7, though these results warrant cautious interpretation (Figure S1C and S1F). Collectively, these findings suggest that HBEGF and VPS8 exhibit consistent expression patterns and may have diagnostic value across multiple cohorts, warranting further investigation as biomarkers for HCM.

Figure 4. Machine learning–based identification of centrosome duplication–related biomarkers in HCM.

A, Area under the curve (AUC) values of 101 machine learning prediction models evaluated in the GSE36961 and GSE141910 data sets. B and C, Receiver operating characteristic (ROC) curves of the optimal Lasso+glmBoost model in the training cohort GSE36961 (B) and the validation cohort GSE141910 (C). D and E, Expression levels of feature genes identified by the Lasso+glmBoost model in HCM and control samples from GSE36961 (D) and GSE141910 (E). F and G, ROC curves evaluating the diagnostic performance of HBEGF and VPS8 in GSE36961 (F) and GSE141910 (G). MR indicates Mendelian randomization.

ANN Model Shows Favorable Diagnostic Performance for HCM

The ANN model was constructed to further evaluate the diagnostic value of biomarkers for HCM. The model comprised one input layer, one hidden layer, and one output layer, as illustrated in Figure 5A. Specifically, the input layer had two neurons, the hidden layer had 3 neurons, and the output layer had two neurons. The confusion matrix indicated that the ANN model achieved high prediction accuracy for both the HCM and control groups in both the GSE36961 and GSE141910 data sets (Figure 5B and 5C). Moreover, the AUC values of the ANN model in GSE36961 and GSE141910 were 0.9320271 and 0.7917384, respectively (Figure 5D and 5E), indicating that the ANN model exhibited favorable diagnostic performance in HCM.

Figure 5. Artificial Neural Network (ANN) model architecture and diagnostic performance for HCM.

A, Visualization of ANN Models, the leftmost layer is the input layer, the middle layers are the hidden layers, and the rightmost layer is the output layer. B and C, Confusion Matrix Heatmap: The Horizontal Axis Represents True Values, and the Vertical Axis Represents Predicted Values. D and E, ROC curves (Training Set on the Left, Validation Set on the Right). AUC indicates area under the curve.

VPS8 Is Highly Expressed in Cardiomyocytes and Dendritic Cells

Before QC of the scRNA‐seq data, the data set comprised 241 382 cells and 23 411 genes. After QC, 196783 cells were retained, while the gene count remained at 223411 (Figure S2A and S2B). Following standard data processing, a subset of 2000 highly variable genes was identified (Figure 6A). PCA was performed, and the top 20 statistically significant principal components (P<0.05) were selected for subsequent analysis (Figure 6B). Using uniform manifold approximation and projection clustering, 11 distinct cell clusters were identified (Figure 6C), and the expression patterns of key marker genes for each cell cluster were visualized through bubble plots (Figure S2C). Marker genes for each cell cluster are listed in Table 5. Annotation of cell clusters revealed six cell types, including cardiomyocytes, dendritic cells, endothelial cells, fibroblasts, pericytes, and smooth muscle cells (Figure 6D). uniform manifold approximation and projection and bubble plots showed that VPS8 was highly expressed in cardiomyocytes and dendritic cells (Figure 6E and 6F). Moreover, the results indicated that cardiomyocytes and dendritic cells were involved in regulating centrosome duplication in HCM, thereby identifying them as key cell types.

Figure 6. VPS8 was highly expressed in cardiomyocyte and dendritic cell.

A, Highly Variable Gene Screening. B, Screenfi Plot of Principal Components. C, Uniform Manifold Approximation and Projection (UMAP) Dimensionality Reduction and Clustering. D, UMAP Diagram for Cell Type Annotation. E and F, UMAP Diagram and Bubble Chart of Key Gene Expression Levels.

Table 5.

Marker Genes Used for Annotation of Distinct Cell Types Identified by Single‐Cell RNA Sequencing (scRNA‐seq)

Cell type	Cell clusters	Marker gene
Cardiomyocyte	0,3,7,8,9	MYBPC3, MYH7B, MYH7, TNNT2, ACTC1
Dendritic	6	HLA‐DRB1, ITGAX
Endothelial cells	4	AQP1, PECAM1, VWF
Fibroblast	1,2	COL1A1, COL6A3, FBN1, IGF1
Pericyte	5	ABCC9, KCNJ8, NOTCH3, STEAP4
Smooth muscle cells	10	SingleR package (v 1.831) 39

Validation of VPS8 Expression in H9c2 Cells

Western blot analysis confirmed a significant reduction in VPS8 protein levels in H9c2 cells following shRNA‐mediated knockdown (Figure 7A). β‐actin served as an internal control, with no apparent differences among groups, confirming successful VPS8 silencing and providing a reliable basis for subsequent functional assays.

Figure 7. Validation and functional analysis of VPS8 in H9c2 cardiomyocytes.

A, Western blot analysis of VPS8 protein expression in H9c2 cells transfected with control, Negative control shRNA (sh‐NC), or shRNA targeting VPS8 (sh‐VPS8) plasmids, with β‐Actin as the internal control. B and C, Cell Counting Kit‐8 (CCK‐8) assay performed to measure cell proliferation of H9c2 cells at 24, 48, and 72 hours after transfection, and line charts generated to compare cell viability among groups. D through G, Flow cytometry analysis using Annexin V‐fluorescein isothiocyanate/propidium iodide (PI) dual staining to evaluate apoptotic cell populations in each group, accompanied by a quantitative summary of apoptosis percentages. H through K, Immunofluorescence staining of centrosome‐ and microtubule‐associated proteins, including γ‐tubulin, CEP135, PLK4 and Centrin‐2, showing their cellular distribution in H9c2 cells. L, Western blot detection of the same centrosome‐related proteins to verify their expression patterns at the protein level.

VPS8 Knockdown Suppresses Cell Proliferation

CCK‐8 assay results indicated that VPS8 knockdown suppressed cell proliferation, with the most notable reduction observed from 24 hours onwards and becoming more pronounced at 48 and 72 hours (Figure 7B). The line graph further illustrates that the viability of cells in the sh‐VPS8 group was consistently lower than that of the control and negative control shRNA groups (Figure 7C), suggesting the role of VPS8 in promoting cell growth.

VPS8 Knockdown Promotes Apoptosis

Flow cytometry analysis using PI and Annexin V–fluorescein isothiocyanate staining revealed a considerable increase in the proportion of apoptotic cells in the sh‐VPS8 group compared with the control and negative control shRNA groups (Figure 7D through 7F). Quantitative bar charts showed a significant elevation in the apoptosis rate following VPS8 knockdown (Figure 7G), indicating that VPS8 may exert a protective effect against apoptosis.

VPS8 Regulates Centrosome Proteins and Microtubule Architecture

Immunofluorescence staining revealed that VPS8 knockdown led to marked reductions in centrosomal and microtubule‐associated proteins, including γ‐tubulin, CEP135, Centrin‐2, and PLK4 (Figure 7H through 7K), with particularly notable decreases in γ‐tubulin and PLK4 signals. Western blot analysis supported these findings, showing decreased levels of these proteins in the sh‐VPS8 group compared with the controls, with γ‐tubulin exhibiting the most significant decline (Figure 7L). Collectively, these results suggest that VPS8 may regulate centrosome integrity and microtubule organization by modulating the expression and assembly of key centrosomal proteins, thereby influencing cell cycle progression and structural stability in cardiomyocytes.

DISCUSSION

This study, focused on the role of CDRGs in HCM, constructed a multilayered chain of evidence spanning transcriptomic analysis, coexpression networks, and causal inference through MR, further integrated through machine learning, single‐cell atlas resolution, and functional validation of a key candidate gene at the cellular level. HCM is an inherited cardiovascular disorder characterized by abnormal thickening of the ventricular wall. Patients are often asymptomatic in the early stages; however, as myocardial fibrosis progresses, heart failure frequently develops, for which pharmacological interventions have limited efficacy. The centrosome, which functions as the microtubule‐organizing center within cells, plays a crucial role in regulating cell division and signaling, and its duplication is closely associated with cardiomyocyte proliferation and maturation. Previous research has indicated that inhibiting centrosome duplication promotes the differentiation of immature cardiomyocytes, whereas the Hippo pathway facilitates cardiomyocyte maturation by regulating centrosome disassembly. ⁶ These findings suggest that aberrations in centrosome replication may contribute to the transition of myocardial growth from hyperplastic to hypertrophic, thereby influencing the pathogenesis and progression of HCM. Based on this framework, this study systematically analyzed the expression profiles and potential mechanistic roles of CDRGs in HCM.

Through differential expression analysis, WGCNA‐based coexpression network construction, MR causal inference, and multimodel integrative screening, HBEGF and VPS8 were identified as two candidate genes with stable diagnostic values across multiple cohorts. WGCNA revealed that modules significantly correlated with CDRG activity scores in HCM samples were enriched in pathways involving ion transmembrane transport, cell junction organization, and phagosome/muscle cell cytoskeleton regulation, suggesting a coupling imbalance between cell structure, membrane transport, metabolism, and immune signaling, which may constitute a molecular hallmark of HCM. By applying stringent criteria for IV selection and controlling for heterogeneity and pleiotropy, MR analysis identified consistent causal links between several genes and HCM risk, reinforcing the causal plausibility of CDRG dysregulation in HCM pathology. Further single‐cell transcriptomic analysis revealed that both HBEGF and VPS8 were upregulated in HCM myocardial tissues but exhibited distinct cell‐type distributions and functional associations.

Under physiological homeostasis, heparin‐binding EGF‐like growth factor (HBEGF), as a key ligand of epidermal growth factor receptor (EGFR), can regulate BPs such as cell migration, angiogenesis, and anti‐apoptosis by activating the EGFR/PI3K/Akt and MAPK signaling pathways. ⁴⁸ , ⁴⁹ , ⁵⁰ Its high expression in cardiomyocytes and smooth muscle cells supports normal cardiac development and exerts a fundamental cardioprotective effect by inhibiting cardiomyocyte apoptosis, ⁵¹ , ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ which is consistent with our MR analysis, which identified HBEGF as a protective genetic factor in HCM. When the myocardium is exposed to pathological stresses, such as sarcomere remodeling and abnormal pressure overload induced by HCM‐associated mutations, the body can initiate an adaptive compensatory response to upregulate HBEGF expression, attempting to further antagonize cardiomyocyte apoptosis, improve myocardial blood supply, and delay fibrosis by enhancing the activity of the EGFR/PI3K/Akt pathway. This regulatory logic is similar to the regulatory mode of the HBEGF/EGFR/PI3K/Akt axis in Tetralogy of Fallot, in which loss‐of‐function mutations in ADAM17 impede the normal shedding of HBEGF, resulting in the failure of effective EGFR activation, which in turn triggers abnormalities in the PI3K/Akt pathway and pathological phenotypes, including cardiomyocyte hypertrophy and disorganized sarcomeres. ⁵⁸ However, the high HBEGF expression in the myocardial tissues of patients with HCM included in this study likely reflects the middle and advanced stages of the disease. At these stages, the upregulation amplitude of HBEGF may no longer match the extent of myocardial injury due to the long‐term pathological microenvironment, resulting in insufficient activation of its downstream protective pathways. Eventually, even sustained high HBEGF expression cannot reverse established irreversible pathological changes, such as myocardial hypertrophy and diastolic dysfunction, presenting a decompensated state characterized by “compensatory upregulation yet failure to block disease progression.” However, the extent of its enrichment in cardiomyocytes and its direct coupling with the centrosome/microtubule system remain insufficiently characterized, limiting its current feasibility for experimental validation.

Conversely, vacuolar protein sorting 8 homolog (VPS8) plays a pivotal role in intracellular trafficking and lysosomal function. As a core subunit of the CORVET complex, VPS8 interacts with the Rab5 homolog Vps21 to precisely regulate the maturation and fusion of early endosomes, thereby maintaining the dynamic equilibrium of the endosomal–lysosomal system. ⁵⁹ , ⁶⁰ , ⁶¹ , ⁶² , ⁶³ , ⁶⁴ , ⁶⁵ Previous studies have established a close association between lysosomal dysfunction and cardiovascular diseases such as HCM and Fabry disease. In Fabry disease, abnormal accumulation of the substrate Gb3 induces cardiomyocyte inflammation and fibrosis, ultimately leading to left ventricular hypertrophy and cardiac dysfunction. ⁶⁶ Our finding that VPS8 knockdown impairs cardiomyocyte proliferation, increases apoptosis, and dysregulates centrosomal protein suggests that its functions may extend to maintaining cell cycle progression and cytoskeletal stability. Previous studies have shown a close association between lysosome‐mediated autophagy and HCM. As the core organelle of the autophagic pathway responsible for degrading and recycling damaged or obsolete intracellular components, lysosomes can help cardiomyocytes resist external injuries such as hypoxia and oxidative stress through this process, while autophagy is closely linked to the progression of myocardial hypertrophy in HCM. ⁶⁷ Therefore, the observed upregulation of VPS8 in HCM tissues may represent a compensatory response of the body to pathological stresses induced by the disease. On the one hand, it can enhance lysosomal function, promoting the efficient degradation of damaged components and nutrient recycling to counteract injuries such as hypoxia and oxidative stress induced by HCM‐causing mutations and antagonizing cardiomyocyte apoptosis. On the other hand, it can maintain cell cycle progression and cytoskeletal stability, ameliorating abnormalities such as impaired cardiomyocyte proliferation and centrosomal protein disorders induced by VPS8 knockdown, thereby delaying the progression of myocardial hypertrophy. However, this protective effect has inherent limits: once the disease reaches a decompensated stage, VPS8 upregulation alone cannot reverse the established pathological myocardial changes.

Notably, our experiments confirmed that VPS8 knockdown significantly reduces the expression of γ‐tubulin and PLK4, two key centrosomal proteins. The downregulation of these proteins may regulate the progression of HCM through multiple mechanisms. From a histological perspective, PLK4 deficiency directly impairs centriole duplication, triggering a p53‐dependent cell cycle surveillance mechanism. This mechanism prevents the continuous proliferation of cells with abnormal centriole duplication, ultimately resulting in irreversible cell cycle arrest after several cell divisions. ⁶⁸ Combined with the cell cycle regulatory principle that the completion of centriole duplication is a critical prerequisite for cells to transition from the G2 phase to the M phase, this arrest manifests as G2/M phase arrest. At this stage, cell division is blocked while intracellular protein synthesis and organelle biogenesis continue. Consequently, cardiomyocytes accumulate contraction‐related proteins and organelles, such as mitochondria, which directly promote cardiomyocyte hypertrophy. However, these hypertrophic cardiomyocytes cannot properly organize myofibrils due to abnormal proliferation, which results in cellular disorganization. For instance, in diabetic cardiomyopathy, abnormal centriole duplication has been implicated as a key molecular mechanism driving cardiomyopathy development through the ANXA11‐mediated mitochondrial dysfunction pathway. ⁶⁹ γ‐Tubulin is a ubiquitous and highly conserved member of the eukaryotic tubulin family. Studies have shown that targeted regulation of γ‐tubulin‐mediated microtubule network tyrosination and interventions in microtubule network function can improve cardiac function in HCM, providing a novel direction for HCM treatment through modulation of the nonsarcomeric cytoskeleton. ⁷⁰ , ⁷¹ , ⁷² Although existing studies have not clarified the mechanistic association between γ‐tubulin and HCM, the literature indicates that cardiac myosin‐binding protein C (cMyBP‐C) haploinsufficiency can induce α‐tubulin acetylation and upregulate related heat shock proteins in HCM myocardial tissue, promoting abnormal proliferation of the microtubule network. This alteration of the microtubule network represents a novel pathomechanism for the development of cMyBP‐C haploinsufficiency–mediated HCM. Meanwhile, α‐tubulin can interact with the protein QC system to maintain cardiomyocyte proteostasis, and its abnormal expression and modification can exacerbate HCM‐related proteostasis imbalance. ⁷³ Ultimately, these two factors synergistically exacerbate characteristic pathological phenotypes of HCM, including myocardial hypertrophy and cellular disorganization.

This study, integrated with findings from Garmany et al and Liang et al, advances the research framework for understanding HCM pathogenesis and its clinical translation. Garmany et al ⁷⁴ identified upregulation of the Ras‐MAPK signaling cascade as a key driver of myocardial hypertrophy in HCM through multi‐omics analysis, whereas Liang et al ⁷⁵ confirmed that heart transplantation and other approaches are effective therapeutic strategies for end‐stage HCM. Conversely, our research focuses on the previously underemphasized centrosome duplication–related pathway, providing unique complementary insights. Through integrated multidatabase analysis and in vitro functional validation, we are the first to identify HBEGF and VPS8 as HCM‐specific diagnostic biomarkers, with their diagnostic efficacy validated across multiple cohorts. Functional experiments revealed that VPS8 regulates cardiomyocyte proliferation, apoptosis, and centrosome integrity by modulating key proteins, including γ‐tubulin. Notably, this mechanism is independent of the Ras‐MAPK pathway reported by Garmany et al, ⁷⁴ yet it collectively contributes to the pathological progression of HCM. Furthermore, the high expression of VPS8 in cardiomyocytes and dendritic cells suggests a potential association among aberrant centrosome regulation, myocardial dysfunction, and immune modulation, offering a novel target for precise intervention in HCM. This mechanistic insight establishes a translational bridge to the clinical treatment strategies highlighted by Liang et al, ⁷⁵ further enhancing the feasibility of translating basic HCM research into clinical practice. It is noteworthy that, while this study builds on existing literature, it offers clear, innovative extensions. Garmany et al ⁷⁶ confirmed the involvement of the Ras‐MAPK pathway in the pathophysiological process of obstructive HCM through multi‐omics analysis; this study further identified HBEGF as a potential upstream regulator of this pathway, addressing a knowledge gap regarding its upstream regulatory mechanisms. Similarly, Liang et al ⁷⁷ proposed the core hypothesis that immune dysregulation contributes to HCM progression. In this study, we revealed that VPS8‐mediated centrosome activation can regulate the immune system, suggesting a novel pathway through which immune dysregulation may influence HCM pathophysiology. These findings further strengthen the immune‐related hypothesis, providing new experimental evidence for HCM pathogenesis by linking centrosome regulation and immune disorders.

The MR analysis revealed that genetically high‐expression variants of VPS8 and HBEGF act as protective factors against HCM (OR<1), reflecting their inherent protective roles in maintaining cardiomyocyte function under physiological homeostasis. This finding is consistent with the phenotypic observation that knockdown of VPS8 in vitro inhibits cardiomyocyte proliferation and promotes apoptosis. Notably, the cardioprotective effect of HBEGF has been confirmed to maintain cardiomyocyte survival by activating the EGFR signaling pathway, ⁷⁸ whereas dysregulation of VPS8 is potentially associated with the pathological process of myocardial injury. ⁷⁹ The observed upregulation of these 2 genes in myocardial tissue from patients with HCM may represent a compensatory or adaptive response by the body to counteract the pathological processes of the disease. In the early stages of the heart failure pathological cascade, such compensatory upregulation can effectively antagonize myocardial fibrosis, suppress cardiomyocyte apoptosis, and preserve cardiac function. ⁸⁰ Based on this, we speculate that the HCM samples in this study may represent a decompensated stage, in which the extent of myocardial injury exceeds the threshold for genetic compensation. Consequently, even sustained high expression of VPS8 and HBEGF cannot reverse established, irreversible pathological changes, such as myocardial hypertrophy and diastolic dysfunction. This spatiotemporal discrepancy between the intrinsic protective potential of these genes and the adaptive responses within the pathological microenvironment underscores the complex regulation of gene function in HCM pathogenesis.

Our single‐cell analysis revealed significant upregulation of VPS8 in both cardiomyocytes and dendritic cells. In HCM, cardiomyocytes exhibit characteristic alterations, including hypertrophy, disrupted myosin and calcium cycling, and polyploidization. ¹ , ² , ⁸¹ , ⁸² Conversely, dendritic cells may mediate cardiac immune–inflammatory responses and fibrotic processes through antigen presentation. ⁸³ The high expression of VPS8 in these cell types suggests its potential involvement in HCM pathogenesis through 2 pathways: first, by regulating centrosome duplication and microtubule homeostasis in cardiomyocytes, thereby affecting cell cycle progression and survival; and second, by mediating immune signaling in dendritic cells, potentially promoting inflammation and fibrosis. Based on this cellular localization and mechanistic rationale, VPS8 was prioritized for functional validation. Experimental results showed that VPS8 knockdown significantly inhibited cardiomyocyte proliferation, promoted apoptosis, and reduced the expression of key centrosomal and microtubule‐associated proteins, including γ‐tubulin, CEP135, Centrin‐2, and PLK4, further validating its crucial role in maintaining cardiomyocyte structural homeostasis. Collectively, these findings establish a complete evidence chain—from bioinformatic screening and single‐cell localization to experimental validation—highlighting the central role of VPS8 in HCM pathological remodeling.

This study has several methodological limitations, and caution is required when defining its scope of application and interpreting its conclusions. For in vitro functional validation, the H9c2 rat cardiomyoblast cell line was used to examine the biological effects of VPS8. As embryonic‐derived rat cardiac precursor cells that have not undergone terminal differentiation, H9c2 cells differ significantly from mature human cardiomyocytes in gene expression profiles, cellular structures (eg, sarcomere organization), and electrophysiological contractile functions. Additionally, evolutionary species divergence may result in species‐specific regulatory pathways and functional effects of genes, including VPS8. Therefore, extrapolating in vitro experimental results to human HCM pathophysiology should be approached with caution. Due to experimental constraints and ethical considerations, validation was not performed in primary human cardiomyocytes or in mature human cardiomyocytes derived from induced pluripotent stem cells. Future research should use human‐derived cardiomyocyte models to further verify the functions of target genes and enhance the clinical translational value of these findings. Regarding data sources and analyses, the core discoveries of this study were primarily based on secondary analyses of Gene Expression Omnibus data sets. Although multiple independent data sets were used for model training and validation to enhance robustness, these data sets may have inherent demographic imbalances (eg, age and sex) or limited ethnic diversity. Such potential selection bias could limit the generalizability of biomarker models and related conclusions across broader populations. ⁸⁴ Subsequent studies should validate these findings in larger, prospectively collected, and more ethnically diverse cohorts. Additionally, some validation data sets (eg, GSE141910) showed imbalanced case–control sample sizes. Although robust statistical methods and machine‐learning algorithms, including Least Absolute Shrinkage and Selection Operator with inherent tolerance for imbalance, were applied and cross‐validation was performed to evaluate model performance, sample size disparities may still affect the statistical power of differential expression analyses and model calibration. The convergence of evidence across multiple data sets, however, mitigates the impact of biases in individual data sets. Finally, validation in cohort GSE32453 was limited by its very small sample size (n=13). While the observed trends were consistent with larger cohorts, statistical estimates from such a small sample are inherently unstable and should be interpreted as preliminary or supportive rather than definitive. The primary evidence supporting our conclusions derives from larger, more robust data sets. At the single‐cell level, limitations include the absence of multiple‐testing correction, as the analysis was primarily focused on phenotypic characterization in a relatively small sample. Furthermore, the data set lacked normal control samples, thus limiting the comparative interpretation. Future studies should validate these single‐cell findings in larger cohorts with paired normal and tumor samples.

Traditionally, genetic studies of HCM have predominantly focused on sarcomeric genes. Conversely, our multi‐omics framework—encompassing transcriptome‐based coexpression network analysis, MR‐based causal inference, single‐cell resolution localization, and functional validation—highlights a previously underrecognized group of CDRGs as potential contributors to HCM pathogenesis. These findings indicate that the pathological mechanisms of HCM may involve defects in the contractile apparatus and broader dysregulation of cell cycle regulation, intracellular trafficking, and structural maintenance systems. Within this regulatory network, HBEGF and VPS8 emerge as central nodes representing 2 complementary pathological pathways: growth factor–mediated survival signaling and endomembrane system–mediated maintenance of cellular homeostasis, respectively. Their concurrent dysregulation may collectively drive the complex phenotypic manifestations of hypertrophic cardiomyocyte growth, abnormalities in electromechanical coupling, and interstitial fibrosis in HCM.

In summary, this study proposes a novel molecular framework in which dysregulation of CDRGs may, through VPS8‐mediated imbalance in vesicular trafficking and the endosomal–lysosomal pathway, disrupt the structural and metabolic homeostasis of cardiomyocytes. This process may act synergistically with immune cell–mediated inflammatory and fibrotic processes, thereby promoting the electromechanical–immune coupled remodeling observed in HCM. The coordinated upregulation of HBEGF and VPS8 reflects the involvement of 2 complementary pathological pathways: growth factor signaling and maintenance of cellular structural homeostasis. Building on the functional validation of VPS8, this study provides new mechanistic insights into HCM pathogenesis and suggests potential translational directions for early diagnosis and targeted therapeutic interventions. Future studies employing cardiomyocyte‐specific conditional knockdown or overexpression of Vps8 in animal models will be essential to further validate its critical role in myocardial hypertrophy and fibrosis at the histological, molecular, and functional levels.

Sources of Funding

This research was supported by the National Natural Science Foundation of China (Grant No. 82560082) and the Natural Science Foundation of Jiangxi Province (Grant No. 20212BAB206042).

Disclosures

None.

Supporting information

Data S1 Supporting Information

Tables S1–S4

Figures S1–S2