Integrated multi-omics analysis identifies lipid metabolism biomarkers in ONFH and reveals therapeutic potential of retinoic acid

Abstract

Osteonecrosis of the femoral head (ONFH) is a debilitating condition frequently associated with dysregulation in lipid metabolism. The objective of this study was to identify potential biomarkers for ONFH through various analytical methods and experimental verification, thereby providing a robust reference for disease treatment. Utilizing public datasets, differential analysis was conducted and integrated with weighted gene co-expression network analysis to pinpoint key genes. Subsequently, biomarkers related to lipid metabolism in ONFH were identified by combining machine learning techniques with receiver operating characteristic curves. Furthermore, based on these biomarkers, a nomogram model was developed for diagnostic prediction. Immune infiltration was assessed, and drug predictions involved molecular docking studies. Finally, the relationship between biomarkers and immune cells was investigated through single-cell analysis. Then the system experiment is carried out to verify. Thirty key genes were identified, primarily enriched in "phospholipid metabolic processes" and “lysosome”. A total of three biomarkers—CREBBP, GLB1, and PSAP—were recognized as exhibiting strong diagnostic efficacy. The nomogram incorporating these biomarkers demonstrated accurate predictive capabilities. Immune infiltration analysis revealed altered levels of neutrophils and activated dendritic cells in ONFH that correlated with biomarker expression. Additionally, molecular docking studies indicated stable interactions between the identified biomarkers and potential therapeutic agents such as estradiol and retinoic acid. Single-cell sequencing analysis uncovered significant differences in biomarker expression across various immune cell types including T cells. Finally, it was verified that retinoic acid improved ONFH by hsa-miR-320a/GLB1 in vitro and in vivo. This study identifies CREBBP, GLB1, and PSAP as promising biomarkers for ONFH while underscoring their potential roles in lipid metabolism and immune modulation. Moreover, we confirmed the important conclusion that retinoic acid improves ONFH through hsa-miR-320a/GLB1. Future investigations into the therapeutic implications of these findings may offer new avenues for ONFH management.

Introduction

Osteonecrosis of the femoral head (ONFH) is a localized bone metabolic disorder caused by abnormal blood supply to the femoral head¹. With complete collapse of the femoral head in 80% of untreated patients, ONFH has become one of the most serious challenges faced by orthopedic surgeons². Globally, there are over 30 million individuals with ONFH, with approximately 22,000 new cases annually in the United States³. The incidence of ONFH in adults in China is relatively high, predominantly affecting middle-aged and young individuals. The majority of patients present at a stage where the bone necrosis has advanced to the mid-to-late stages, with collapse of the femoral head. Despite numerous studies on the pathogenesis of ONFH, a consensus has not yet been reached⁴. Joint replacement surgery may offer the best therapeutic outcome. However, artificial prostheses have a finite lifespan, and young patients who undergo joint replacement may require one or more revisions later in life. Exploring early non-surgical treatment methods to delay or avoid joint replacement is currently a research direction for ONFH. Therefore, the development of diagnostic biomarkers for ONFH is crucial.

The effective regulation and maintenance of lipid homeostasis is crucial for the normal function of cells and tissues. Patients with ONFH often have severe lipid metabolism disorders and fat accumulation in the damaged bone marrow cavity⁵. Therefore, lipid metabolism disorders are often recognized as a key factor in the pathogenesis of ONFH⁶. In addition, dysregulation of lipid metabolism may lead to ischemia by increasing intraosseous pressure and reducing blood flow, ultimately resulting in non-traumatic ONFH (NONFH)⁷. Decreased levels of adiponectin (APN) have been detected in patients with ONFH, and lower APN levels may reflect the severity of NONFH⁸. In animal experimental models of ONFH, the first change observed is an increase in blood lipid levels⁹. Long-term and high-dose use of glucocorticoids can lead to lipid metabolism disorders. These disorders can result in the appearance of small fat droplets in the blood, which are prone to form thrombi within the small vessels in the femoral head area, thereby affecting the microvascular circulation of the femoral head. Concurrently, hormones also have a proliferative effect on adipose tissue in the body, which leads to significant proliferation and hypertrophy of intramedullary fat cells, and causes an increase in intramedullary pressureand ultimately ONFH. Given the critical role of lipid metabolism in ONFH, investigating its potential as a biomarker of ONFH is essential for the development of innovative therapeutic strategies.

Building on previous findings and leveraging data from multiple public databases, this study aims to utilize transcriptomic data in conjunction with advanced machine learning techniques to identify biomarkers associated with lipid metabolism in ONFH. The characteristics of these biomarkers at the cellular level were further analyzed by single-cell RNA sequencing (scRNA-seq). Subsequently, we will experimentally validate the key genes that regulate lipid metabolism and further elucidate the critical molecular pathways involved in the pathogenesis of ONFH. The flowchart of this study visualizing multiple analytical methods was shown as Fig. 1. This comprehensive approach will aid in a better understanding of the disease and provide more nuanced perspective for the clinical diagnosis of ONFH.

Fig. 1.

The flowchart of this study.

Materials and methods

Data collection

This study utilized several datasets, including mRNA transcriptome datasets (GSE123568 and GSE74089), and the microRNA (miRNA) transcriptome dataset GSE89587, all obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo). The GSE123568 dataset was based on the GPL15207 platform and contained 30 samples of ONFH and 10 control samples. The GSE74089 cohort, using the GPL13497 platform, comprised 4 samples of ONFH and 4 control samples. The GSE89587 cohort, employing the GPL21439 platform, included 10 samples of ONFH alongside 10 controls. Additionally, we extracted 720 lipid metabolism-related genes (LMRGs) from the Reactome database (https://reactome.org/) (Table S1). The scRNA-seq dataset SRP361778, sourced from the SRA database (https://www.ncbi.nlm.nih.gov/sra), included 6 samples of ONFH and 5 controls, providing a robust foundation for further analysis. The comprehensive table summarizing the key metadata for each dataset was exhibited in Table 1.

Table 1.

The specific information of the four datasets used in this study.

GEO accession numbers	platform information	sample sizes per group	dataset publication dates
GSE123568	GPL15207	30 SONFH patients and 10 non-SONFH patients (following steroid administration)	Public on Dec 31, 2019
GSE74089	GPL13497	12 NFH patients and 12 healthy controls	Public on Jun 01, 2016
GSE89587	GPL21439	10 SONFH patients and 10 non-SONFH patients	Public on Sep 14, 2017
SRP361778	ILLUMINA (Illumina NovaSeq 6000)	Control: ONFH = 5: 6	unknown

Differential expression analysis

Differentially expressed genes (DEGs) between the ONFH and control groups were identified in the GSE123568 and GSE74089 datasets using the limma package (v 3.54.0)¹⁰. The selection criteria was set at |log2-fold change (FC)|> 0.5 and FDR-adjusted P-value < 0.05. The results from both datasets were then integrated to obtain common DEGs¹¹. For visualization, volcano plots and heatmaps were generated using the ggplot2 (v 3.3.3)¹² and ComplexHeatmap (v 2.14.0)¹⁰ packages, facilitating an intuitive representation of gene expression changes.

Weighted gene co-expression network analysis (WGCNA)

Co-expression network construction using the WGCNA package (v 1.70.3)^13, with ONFH as a trait, was performed for the whole gene expression profile of the GSE123568 cohort. First, the samples were clustered and outlier samples were removed. The scale-free R² is close to 0.8 and the average connectivity is close to 0. The optimal soft threshold is chosen, at which point the network approaches a scale-free distribution. The minimum number of genes per gene module was set to 200 to divide the gene modules. Pearson correlation analysis was performed on the modules from the WGCNA results with ONFH, selecting modules with |cor| greater than 0.5 and P value less than 0.05, thereby extracting key module genes associated with ONFH.

Identification of key genes and functional enrichment analysis

The common DEGs were intersected with LMRGs and key module genes from the WGCNA to define the key genes in ONFH. Subsequently, Gene Ontology (GO) (FDR-adjusted P-value < 0.05) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (P value < 0.05) functional enrichment analysis were conducted using the clusterProfiler package (v 3.18.0)¹⁴. The KEGG pathway database is copyrighted by Kanehisa laboratories and we have obtained formal permission from them to publish this material commercially under an Open Access license^15–17.The results were visualized with ggplot2 and GOplot (v 3.6.1)¹⁸, allowing for an in-depth understanding of these genes’potential roles in biological processes.

Biomarker selection, chromosomal localization, and tissue expression analysis

The key genes were analyzed in the GSE123568 cohort using the Glmnet package (v 3.6.0)¹⁹, the Boruta package (v 7.0.0)²⁰, and the e1071 package (v1.7.0)²¹ for least absolute shrinkage and selection operator (LASSO) regression analysis, Boruta, and support vector machine-recursive feature elimination (SVM-RFE) machine learning algorithms, respectively. The ggvenn package (v 0.1.9)²² was utilized to extract the intersection of genes identified by the three algorithms as candidate biomarkers. Receiver operating characteristic (ROC) curves were plotted in the GSE123568 and GSE74089 cohorts using the pROC package (v 1.13.0)²¹, and area under the curve (AUC) values were calculated to select candidate biomarkers with AUC values greater than 0.7 in both cohorts as biomarkers. The RCircos package (v 4.1.1)^23,24 facilitated chromosomal localization analysis of these biomarkers, illustrating their positions and expression changes on chromosomes. To explore the tissue specificity of these feature genes, data from the Human Protein Atlas (http://www.proteinatlas.org/), Bgee (https://bgee.org/), and BioGPS (http://biogps.org/#goto=welcome) were analyzed for biomarker distribution in human bone marrow, adipose tissue, and blood.

Construction of nomogram

Based on the expression levels of the biomarkers, a nomogram model was constructed in the GSE123568 cohort using the rms package (v 3.5.0)²⁵ to predict the probability of ONFH diagnosis. Each biomarker’s expression corresponded to a score, when summed, a total point was produced which used to estimate disease diagnostic probability. The accuracy of the nomogram model was evaluated by plotting a calibration curve. Additionally, ROC curves were generated using the pROC package to assess diagnostic capability, while decision curve analysis (DCA) was performed with the rmda package (v 3.4.4)²⁶ to evaluate the clinical value of the nomogram model.

Immune infiltration analysis

Utilizing the CIBERSORT algorithm (v 1.03)²⁷ and LM22 gene set²⁸, we calculated the proportions of 22 immune cell types in the GSE123568 cohort. Bar plot illustrating the abundance percentages of immune cell infiltration was generated using ggplot2, and comparisons between ONFH and control groups were conducted through Wilcoxon test, as depicted in box plot. Spearman correlation analysis was conducted using the psych package (v 2.1.6)²⁹ to assess the relationship between biomarkers and 22 immune cells, with results visualized in heatmap via the ggplot2 package.

Construction of regulatory network and prediction of transcription factors (TFs)

Differentially expressed miRNAs were identified in the GSE89587 dataset using the same methods as for GSE123568 and GSE74089 cohort. Interactions between biomarkers and miRNAs were predicted via the starBase database (https://rnasysu.com/encori/), followed by predictions of relationships between miRNAs and long non-coding RNAs (lncRNAs)³⁰. TFs involved in biomarkers regulation were also predicted using the ChEA3 database (https://maayanlab.cloud/chea3/), selecting TFs supported by ChIP-seq data from the ENCODE database (https://www.encodeproject.org/). Network diagrams were visualized using the Cytoscape software (v3.9.1). Combined with the CytoHubba plugin, centrality indices such as Degree, Betweenness, and Eccentricity were calculated to identify key nodes³¹.

Drug prediction and molecular docking

Potential targeted drugs for biomarkers were predicted through the DSigDB database (https://dsigdb.tanlab.org/DSigDBv1.0/geneSearch.html) to explore possible therapeutic strategies. To validate the binding stability of predicted drugs with biomarkers, the highest-scoring drug was selected for molecular docking. Protein structures corresponding to biomarkers were obtained from the PDB database (https://www.rcsb.org/), while the 3D molecular structures of corresponding compounds were sourced from PubChem (https://pubchem.ncbi.nlm.nih.gov/) and AlphaFold (https://alphafold.ebi.ac.uk/). Molecular docking was performed using Autodock software to calculate the binding energy between ligands and receptors, with binding energies ≤ − 5.0 kcal/mol considered stable³².

Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA)

Gene sets (c2.cp.kegg_legacy.v2024.1.Hs.symbols.gmt, h.all.v2023.2.Hs.symbols.gmt) were utilized for Spearman correlation analysis between each biomarker and all genes. GSEA based on the KEGG and Hallmark databases was conducted using the clusterProfiler package, with criteria set at |NES|> 1 and P-adjust < 0.05. To further understand the differences in pathways during disease progression between groups, the ONFH and control groups were differentiated by median based on the height of the nomogram scores, and then GSVA analysis was performed using GSVA (v1.42.0)³³ and the limma package, which utilizes thresholds |t|> 2 and P value < 0.05 to show the top 10 up- and down-regulated pathways.

ScRNA-seq analysis

Quality control was first performed on the scRNA-seq dataset SRP361778 using the Seurat package (v4.1.1)³⁴, filtering out genes covered by fewer than 200 cells and those with over 10% mitochondrial gene content. Cells with ≤ 200 and ≥ 8000 genes in the cell were excluded, and genes with count numbers ≤ 200 and ≥ 80,000 were removed. Data normalization was conducted with the NormalizeData function, followed by extraction of the top 2000 highly variable genes (HVGs) using the FindVariableFeatures function. Data scaling was performed using ScaleData, and significant principal components were determined with the JackStrawPlot function. Neighbor identification and clustering of small cell groups were conducted via FindNeighbors and FindClusters, employing a resolution of 0.2 and utilizing the t-SNE clustering and UMAP method for visualization with RunTSNE and RunUMAP, respectively. The FindAllMarkers function was used to identify the top 100–200 HVGs in each cluster, which were cross-referenced with the Cell Marker (http://bio-bigdata.hrbmu.edu.cn/CellMarker/index.html) and PanglaoDB (https://panglaodb.se/) databases (accessed on 25 June 2025) for cell annotation³⁵. Finally, biomarker expression in cells was analyzed using FindMarkers and Wilcox.test, applying thresholds of |logFC|> 2 and Bonferroni-corrected P-values < 0.05.

Ethical Approval

The experimental animals, specifically SD rats, sourced from the Experimental Animal Center of Guizhou Medical University, where the animal experiments were conducted. The facility holds a certificate numbered SYXK (Qian) 2018–0001. All procedures involving the animals adhered strictly to the guidelines outlined in the "Guide for the Care and Use of Laboratory Animals" published by the National Institutes of Health (NIH; No.85–23, revised 1996). The experimental protocol underwent rigorous review and received approval from the Ethical Committee of Guizhou Medical University, under reference number 2304177. Compliance with the International Council for Laboratory Animal Science (ICLAS) standards for the protection of laboratory animals used for scientific purposes was ensured, as well as adherence to the ARRIVE guidelines (Animal Research: Reporting In Vivo Experiments). Every endeavor was made to minimize both the suffering of the animals and the number of animals utilized in this study.

Early ONFH model

The SD rats that remained were utilized to establish the model of ONFH, while an additional 20 healthy adult male rats were randomly selected to serve as the healthy control group. Subsequently, the rats were stratified into three distinct groups: normal, glucocorticoid (GC) treated, and retinoic acid treated, with each group comprising 20 SD rats. Prior to the daily administration of methylprednisolone (MP; 60 mg/kg; Pfizer, USA) into the gluteal muscle for a duration of 10 days³⁶, the buttocks of the rats were meticulously shaved and disinfected with alcohol. The body weight of each rat was recorded before each injection. Four weeks following the completion of MP treatment, samples were taken for subsequent experiments.

Anesthesia of experimental animals and euthanasia

All experimental animals were anesthetized via intraperitoneal injection prior to any surgical procedures to mitigate pain and discomfort. The anesthesia dosage was standardized at 0.25 ml per 100 g of body weight for rats. Specifically, the anesthesia cocktail consisted of a mixture of pentobarbital sodium (50 mg/kg) and ketamine (25 mg/kg), administered through intraperitoneal injection [i.p.]. Following the completion of experiments, the rats were euthanized through an overdose of sevoflurane.

Cellular immunofluorescence

Bone marrow mesenchymal stem cells (BMSCs) treated with glucocorticoids were rinsed in phosphate-buffered saline (PBS) and subsequently fixed with 10% neutral-buffered formalin for a duration at 21 °C for 30 min. To facilitate membrane permeabilization, the cells were incubated with Triton X-100 (sourced from Beyotime Biotechnology, Beijing, China) for 10 min. Blocking of nonspecific binding sites was achieved through a 1-h incubation with donkey serum (also from Beyotime Biotechnology). The cells were then incubated overnight at 4 °C with the primary antibody GLB1 (ab305174; Abcam; diluted 1:100). After three washes with PBS, the cells were incubated for 1 h at 21 °C with Alexa Fluor 594-conjugated goat anti-rabbit secondary antibody (A-11012, Thermo Fisher Scientific; diluted 2 µg/mL). Nuclei staining was accomplished with DAPI (Beyotime Institute of Biotechnology, Jiangsu, China) for 10 min. The samples were then rinsed with PBS and visualized using a confocal microscope (Olympus Life Science, Tokyo, Japan).

RNA extraction and real-time quantitative reverse transcription polymerase chain reaction (qRT-PCR)

Primary cultured BMSCs were subjected to total RNA extraction using TRIzol reagent sourced from Invitrogen (Carlsbad, CA, USA). Subsequently, 2 µg of purified RNA underwent reverse transcription utilizing the RevertAid First Strand cDNA Synthesis Kit provided by Thermo Fisher Scientific (Boston, MA, USA). The qRT-PCR was executed on the Applied Biosystems StepOnePlus Real-Time PCR System (Foster City, CA, USA), employing SYBR Green PCR Master Mix obtained from Toyobo (Japan). The expression levels of the target mRNAs were normalized against the expression of 18S ribosomal RNA (Rn18s). The relative gene expression levels were calculated using the 2^− ΔΔCT method and presented as fold-changes. Each qRT-PCR experiment was conducted in triplicate, with a minimum of three distinct biological replicates included. The primer sequences utilized in this study are detailed in Table 2.

Table 2.

The qRT-PCR primer sequences used in this study.

Gene		Sequence
GLB1	F	GCACGGCATCTATAATGTCACC
	R	GTATCGGAATGGCTGTCCATC
CREBBP	F	AAGAGCCCTCTGAACCAAGGA
	R	CTGTGATGTGGCAGGACTACT
PSAP	F	TGCTGAAAGATAATGCTACGCA
	R	GCAGGTAAGAGTCAACCACCTC
hsa-miR-320a	F	CGGGGAGAGGGCGAAA
	R	AGTGCAGGGTCCGAGGTATT
GAPDH	F	CTCCTCCACCTTTGACGC
	R	CCACCACCCTGTTGCTGT

H&E staining

The bone tissue underwent decalcification using 10% EDTA supplied by Solarbio. Following this, dehydration was achieved through an alcohol gradient sourced from Sinopharm (China). Transparency was induced using xylene, also from Sinopharm, and the tissue was subsequently embedded in paraffin, also sourced from Sinopharm. The bone tissue was sectioned using a Leica RM2016 pathological slicer (Leica, Germany) and baked at 60 °C for 3 h. The slices were dewaxed sequentially in xylene and a gradient of alcohols. The dewaxed sections were stained with Mayer’s hematoxylin and 1% water-soluble eosin solution, both obtained from Solarbio. The stained slices were then dehydrated in a graded series of alcohols, rendered transparent in xylene, and finally mounted with neutral gum sourced from Sinopharm. Images were captured using a BX53 microscope manufactured by Olympus (Japan).

Tissue immunofluorescence

Bone tissue was decalcified, dehydrated, rendered transparent, waxed, embedded, sliced, baked, and dewaxed (the procedure was the same as that outlined for H&E staining). Antigen retrieval was performed using the microwave method, and bone tissue antigen retrieval solution was purchased from BioShun (Shanghai, China). Bone tissues were blocked with goat serum (Solarbio) at 25 °C for 30 min. Anti-GLB1 antibody (ab305174; Abcam; diluted 1:100) was used for the primary antibody reaction (incubated overnight at 4 °C). After three washes with PBS, the cells were incubated for 1 h at 21 °C with either Alexa Fluor 594-conjugated goat anti-rabbit secondary antibody (A-11012, Thermo Fisher Scientific; diluted 2 µg/mL). Nuclei staining was accomplished with DAPI (Beyotime Institute of Biotechnology, Jiangsu, China) for 10 min. The samples were then rinsed with PBS and visualized using a confocal microscope (Olympus Life Science, Tokyo, Japan).

Statistical analysis

All statistical analyses were performed using R version 4.4.1, inter-group differences were evaluated via the Wilcoxon test, with statistical significance set at P < 0.05. Data analysis was further facilitated by SPSS statistical software version 22.0, while statistical graphs were generated using GraphPad Prism 6.0. For quantitative data, normality was assessed using the Kolmogorov–Smirnov test, and variance homogeneity between groups was tested using analysis of variance. Data that exhibited normal distribution and homogeneous variance were reported as the mean ± standard deviation (SD). When comparing only two groups, a two-tailed unpaired Student’s t-test was employed. For analysis involving more than two groups, one-way ANOVA with Tukey’s post hoc test was utilized. Data that did not conform to normal distribution were described by their median (P25, P75). For comparisons between such groups, the Kruskal–Wallis rank-sum test with the DSCF method was applied. P value < 0.05 was considered indicative of a statistically significant difference.

Result

Exploring the potential functions of key genes

In our study on the lipid metabolism in ONFH, we conducted a comprehensive differential expression analysis between the ONFH and control group across two datasets. The analysis of GSE123568 identified 1,977 DEGs, with 1,343 showing upregulation and 634 downregulation (Table SCite ESM.2). Meanwhile, GSE74089 cohort revealed a total of 8,717 DEGs (Table S3), including 5,481 upregulated and 3,236 downregulated genes (Fig. 2A). By integrating the differential analysis results from both datasets, we obtained 596 common DEGs as shown in Fig. 2B, among which 530 were upregulated and 66 downregulated. To identify genes with a strong correlation to ONFH, we performed a module gene selection process. Cluster analysis confirmed the integrity of our sample statuses, and eliminated the need to remove any outliers (Fig. 2C). We established an optimal soft threshold of 18 for achieving a scale-free distribution and then set the minimum number of genes per module to 200 thereby identifying 10 modules (Fig. 2D-E). Notably, the MEbrown (cor = 0.68, P < 0.001), MEgreen (cor = −0.75, P < 0.001), and MEyellow modules (cor = 0.57, P < 0.001) emerged as critical, collectively encompassing 2,510 key module genes (Fig. 2F, Table S4). Ultimately, we intersected the identified common DEGs, key module genes from WGCNA, and LMRGs, resulting in 30 key genes (Fig. 3A).

Fig. 2.

Differential expression analysis and weighted gene co-expression network analysis (WGCNA). (A) volcano plot of differentially expressed genes (DEGs) and heatmap of the expression of the top 5 DEGs in datasets GSE123568 and GSE74089. Red dots represent up-regulated DEGs, blue dots represent down-regulated DEGs, and gray dots are genes that are not statistically significant. (B) the Venn diagram visualizing the common items between these two datasets. (C) sample clustering graph. (D) screening for soft thresholds to ensure the network conforms to a scale-free topology. (E) module dynamic shear tree. F: heatmap of the correlation of the module with osteonecrosis of the ONFH.

Fig. 3.

Acquisition and enrichment analysis of key genes. (A) DEGs, key module genes from the WGCNA and 720 lipid metabolism-related genes (LMRGs) common to both GSE123568 and GSE74089 datasets intersected in 30 key genes. (B) Gene Ontology (GO) enrichment analysis of key genes. MF, Molecular Function; BP, Biological Process; CC, Cellular Component. (C) Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis of key genes.

To delve into the functional roles of these key genes, we conducted enrichment analyses. A total of 284 significant entries were enriched in GO enrichment, of which 172 entries were significantly enriched in biological processes, 12 in cellular components, and 100 in molecular functions. Specifically, it was primarily enriched in the"regulation of lipid metabolic processes","phospholipid metabolic processes", and“lipid droplets”(Fig. 3B, Table S5). These pathways align closely with our research focus. Furthermore, the KEGG analysis highlighted 17 significant pathways, including“Sphingolipid metabolism”,“Lysosome”, and“Glycerophospholipid metabolism”(Fig. 3C, Table S6). These findings indicate potential functional mechanisms underlying ONFH in patients.

Three biomarkers (CREBBP, GLB1, and PSAP) were identified

In a bid to identify potential biomarkers, we employed three machine learning techniques. LASSO regression analysis yielded 3 candidate genes (CREBBP, GLB1, and PSAP) (Fig. 4A), while Boruta analysis identified 14 genes (such as LPCAT1, MID1IP1, PLIN3, EP300, MFSD2B, and so on) (Fig. 4B). SVM-RFE further refined our selection to 24 genes (including PSAP, ORMDA3, OSBPL8, MFSD2B, GPS2, and etc.) (Fig. 4C). By intersecting the results from these analyses, we identified 3 candidate biomarkers: CREBBP, GLB1, and PSAP (Fig. 4D). To evaluate the diagnostic potential of these candidate biomarkers, we generated ROC curves across both cohorts. Notably, all candidate biomarkers demonstrated the AUC values exceeding 0.9 (Fig. 4E), indicating strong diagnostic capacity, thus qualifying them as potential biomarkers. Further analysis of the chromosomal locations revealed that CREBBP is located on chromosome 16, GLB1 on chromosome 3, and PSAP on chromosome 10 (Fig. 4F). We also examined the expression of these biomarkers in human bone marrow, adipose tissue, and blood, highlighting their tissue-specific expression (Fig. 4G). Particularly, PSAP exhibited high expression across all three databases, while GLB1 showed elevated levels in two of the databases. These findings suggest that these biomarkers may play significant roles in tissues such as bone marrow and adipose tissue.

Fig. 4.

Acquisition of biomarkers, chromosomal localization, and tissue expression analysis. (A) least absolute shrinkage and selection operator (LASSO) regression analysis. (B) boruta analysis. In the left panel, the box-and-line plot indicate the minimum, average, and maximum Z-values of the traits. The right panel shows the evolution of the Z-Score of the attributes during the Boruta run. (C) support vector machine-recursive feature elimination (SVM-RFE) screening of candidate genes. The horizontal coordinate indicates the number of genes and the vertical coordinate indicates the error rate. (D) feature genes are obtained after 3 types of machine learning on key genes, and the intersection was taken to obtain the candidate biomarkers. (E) receiver operating characteristic (ROC) curves were plotted for the candidate genesin two datasets, GSE123568 and GSE74089, and the candidate genes with high diagnostic ability [area under the curve (AUC) values greater than 0.7] were selected as biomarkers. (F) localization of biomarkers on chromosomes. (G) expression of biomarkers in tissues was analyzed using three databases [Human Protein Atlas Project (HPA), Bgee, and BioGPS].

A nomogram model for predicting ONFH risk was constructed

The nomogram is a visual statistical tool designed to provide quantitative disease risk assessment for individual patients by combining information from multiple clinical and biomarkers. To synthesize our findings, we constructed a nomogram model utilizing the three identified biomarkers as independent diagnostic factors to predict the risk of ONFH in patients (Fig. 5A). The calibration curve corresponding to the nomogram illustrated a slope close to 1, indicating accurate predictions (Fig. 5B). Additionally, the ROC curve demonstrated that the combined AUC value for the three biomarkers was 0.95, which was greater than the value of each of the three biomarkers [CREBBP (0.940), GLB1 (0.913), PSAP (0.923)], further validated the robustness of our diagnostic model (Fig. 5C). The DCA indicated the clinical applicability of our diagnostic model (Fig. 5D). These results affirm that the biomarkers can effectively predict the risk of ONFH, providing a theoretical foundation for early disease intervention.

Fig. 5.

Constructing the nomogram. (A) construction of the nomogram. (B) calibration chart of the nomogram. (C) ROC curve of the nomogram. (D) decision curve analysis (DCA) analysis of the nomogram.

Immune microenvironment in ONFH

In ONFH, infiltration of immune cells not only affects the local inflammatory response, but may also be involved in bone metabolism and remodeling processes, thus influencing the pathological process. To investigate the immune cell infiltration between the control and ONFH groups, we depicted the abundance percentage of 22 immune cell types (Fig. 6A). Notably, “Neutrophils” were highly infiltrated in ONFH, while "Dendritic cells activated" and "B cells memory" were less infiltrated in ONFH (Fig. 6B). We further explored the relationships between immune cells and biomarkers, revealing a strong negative correlation between PSAP and activated dendritic cells (cor = −0.59, P < 0.001), as well as between GLB1 and activated Dendritic cells (cor = −0.58, P < 0.001), while CREBBP showed a robust positive correlation with neutrophils (cor = 0.68, P < 0.001) (Fig. 6C). We can observe a high degree of CREBBP infiltration in neutrophils. By analyzing the microenvironment in which ONFH’s immune cells exist and exploring the role that biomarkers play in it, we can help predict the prognosis of the disease and provide early defense.

Fig. 6.

Immune infiltration analysis. (A) stacked bar graph of immune cell scores. (B) box line graph of immune cell scores in control and ONFH groups. *P < 0.05; **P < 0.01. (C) bubble plot of the correlation of biomarkers and immune cells. The size of the dots represents the significance of the correlation, and the color of the dots as well as the shade represents the direction and size of the correlation.

Exploring the regulatory relationship between biomarkers and miRNAs

To further elucidate the regulatory mechanisms of our identified biomarkers, we conducted differential expression analysis of miRNAs between ONFH and control samples (Table S7). This analysis yielded 6 differentially expressed miRNAs, with 4 showing upregulation and 2 downregulation (Fig. 7A-B). We explored potential interactions between the 3 biomarkers and the 6 differentially expressed miRNAs, identifying 3 pairs of mRNA-miRNA interactions. Subsequently, we predicted interactions involving 17 miRNAs and lncRNAs, suggesting multiple regulatory pathways, such as MALAT1 and NEAT1 indirectly influencing GLB1 via hsa-miR-320a (Table S8). Moreover, we predicted TFs corresponding to our 3 biomarkers. Notably, TCF7L2, CTCF, and EGR1 were predicted to be common across all 3 biomarkers (Table S9). A comprehensive ceRNA network was constructed by integrating miRNA-lncRNA/mRNA and mRNA-TF regulatory interactions, and comprehensive scoring-based network analysis revealed that hub nodes with comprehensive scores greater than 1 were identified, such as GLB1, PSAP, hsa-miR-93-5p, hsa-let-7i-5p, CTCF, EGR1, TCF7L2, CREBBP, hsa-miR-320a, PBX3, which were consistent with the above findings (Fig. 7C, Table S10). Moreover, it was found that 10 TFs including PBX3, REST, and IRF1 not only regulate the expression of biomarkers but are also regulated by hsa-miR-93-5p and hsa-let-7i-5p (Table S11). These results provide new understanding of the complex regulatory mechanisms of biomarkers.

Fig. 7.

Prediction of the regulatory network and drug prediction. (A) volcano plot of differentially expressed microRNAs (miRNAs). (B) heatmap of the expression of the top 20 differentially expressed miRNAs. (C) the ceRNA and transcription factors (TFs) interaction network of biomarkers. From red to blue represents the comprehensive score from high to low. (D) drug target action network of biomarkers. Red diamonds represent drugs and blue circles represent biomarkers.

Potential drug treatment for ONFH

To further explore potential targeted therapies for ONFH, we conducted a drug prediction analysis involving our three identified biomarkers. This analysis yielded a total of 42 potential drugs with 56 interaction relationships identified (Fig. 7D). Notably, CREBBP was linked to the highest number of drugs. Strikingly, both estradiol and retinoic acid were predicted to interact with all three biomarkers: CREBBP, GLB1, and PSAP. Following this, we selected the top-scoring drugs for molecular docking studies. Generally, a molecular binding energy below −5.0 kcal/mol indicates a favorable binding affinity. It was not difficult to see that the molecular binding energies of CREBBP, GLB1 and PSAP to estradiol and retinoic acid, respectively, were all less than −7 kcal/mol (Table S12), indicating strong binding potential. Specifically, the docking results revealed hydrogen bonds forming between the residues GLN-411 and TYR-413 of GLB1 with estradiol and retinoic acid, while CREBBP showed hydrogen bonding between residue MET-1160 and retinoic acid (Fig. 8). These findings suggest that the predicted drugs may form stable complexes with the biomarkers which highlight their potential as therapeutic targets.

Fig. 8.

Diagram of molecular docking binding sites. In the diagram cyan is the protein molecular structure, green is the compound and yellow dashed line is the hydrogen bonding junction.

Potential functional exploration

To delve deeper into the functional roles associated with our three biomarkers, we conducted GSEA. In the KEGG context, both GLB1 and PSAP were enriched in similar pathways, including “lysosome”,“proteasome”, and “spliceosome”. In the Hallmark context, all 3 biomarkers were activated in "interferon gamma (IFN-γ) response", while GLB1 and PSAP were activated in "MYC targets V1" and “oxidative phosphorylation” (Fig. 9A-C). To further understand the differential pathway between ONFH and control groups, we identified 21 pathways that exhibited significant differences in the KEGG and Hallmark backgrounds. Notably, the “inflammatory response”, "fatty acid metabolism", and “spliceosome” were activated in the ONFH group, while other pathways, including "VEGF signaling pathway" was suppressed (Fig. 9D). Based on the analysis results, we can identify key biological pathways and functional modules, explore biomarker-related functions, and further understand the mechanism of the disease.

Fig. 9.

Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA). (A) GSEA of CREBBP, GLB1, and PSAP in KEGG background, respectively. (B) GSEA of CREBBP, GLB1, and PSAP in Hallmark background, respectively. (C) GSVA between ONFH and control groups in Hallmark background. (D) GSVA between ONFH and control groups in KEGG background.

Characterization of biomarkers by scRNA-seq analysis

Utilizing PCA permutation testing and elbow plots, we selected the top 50 principal components for subsequent analysis (Fig. 10A, S1). Clustering revealed 17 distinct cell clusters, which were characterized against marker genes to identify their respective cell types (Fig. 10B-C). Ultimately, we categorized the clusters into 15 cell types (Fig. 10D and Table S13), including T cells, myeloid cells, fibroblasts, B cells, and others. On the basis of t-SNE clustering, UMAP has validated the favorable spatial distribution distinction of these cell types (Figure S2A). We further analyzed the expression levels and distribution of our biomarkers across different cell types (Fig. 10E-F, S2B). The results indicated that all 3 biomarkers were widely distributed among most cell types, with PSAP predominantly expressed in myeloid cells. Additionally, significant differential expression of the biomarkers was observed across various cell classifications, co-expressed in multiple cells, e.g. T cells, myeloid cells, plasma cells, etc. This reinforces their relevance in the context of ONFH.

Fig. 10.

Single-cell analysis in SRP361778. (A) principal component analysis (PCA) dimensionality analysis. The left graph demonstrates the degree of explanation of each principal component on the variance of the data; the right graph demonstrates the significance of different PCs in explaining the dataset compared to the random data. (B) cell clusters for cluster analysis. Each point in the figure represents a single cell. (C) plot of top100 ~ 200 of highly variable genes compared with marker genes of each cell type. (D) cell annotation plot using t-SNE. (E) distribution of biomarkers in cell types (t-SNE). (F) expression of biomarkers in cell types. *P < 0.05; **P < 0.01; *** P < 0.001; **** P < 0.0001.

Retinoic acid could regulate hsa-miR-320a/GLB1 to improve ONFH

Based on the results of the above multiple analyses, we first verified the expression of GLB1, CREBBP and PSAP in the normal femoral head and ONFH group, and the results showed that the expression of these three genes was different, among which GLB1 showed the most significant trend of high expression in the ONFH group (Fig. 11A). Therefore, we selected GLB1 for immunofluorescence detection in BMSCs and femoral head tissues, and showed a tendency of high expression in the ONFH group, consistent with qRT-PCR results (Fig. 11B,E). After treatment with retinoic acid, it was found that GLB1 expression in the ONFH group could be significantly reduced, while ONFH was significantly improved (Fig. 11C-F). In order to further explore the mechanism, we continued to verify the relationship between GLB1 and hsa-miR-320a. High expression of GLB1 could significantly inhibit the expression of hsa-miR-320a, while treatment with retinoic acid could significantly increase the expression of hsa-miR-320a (Fig. 11G). Subsequently, we sorted out the genome sequence and found that GLB1 and hsa-miR-320a were perfectly matched at multiple sites (Figure S3A), so we confirmed that retinoic acid could regulate hsa-miR-320a/GLB1 to improve ONFH.

Fig. 11.

Retinoic acid improves ONFH by hsa-miR-320a/GLB1. (A) qRT-PCR was used to verify the expression of GLB1, CREBBP and PSAP in the healthy femoral head group and the ONFH group. (B,E) Immunofluorescence confirmed the expression of GLB1 in normal femoral head and ONFH bone marrow mesenchymal stem cells and conducted quantitative analysis. (C,F) Immunofluorescence method confirmed the expression and quantitative analysis of GLB1 in femoral head tissue; (D) HE staining suggested that retinoic acid treatment can improve the pathological progression of ONFH. (G) The expression of hsa-miR-320a in the healthy femoral head group, ONFH group and retinoic acid treatment group was verified by qRT-PCR.^nsP > 0.05;*P < 0.05; **P < 0.01; *** P < 0.001; **** P < 0.0001.

Discussion

The relationship between ONFH and lipid metabolism disorders has garnered increasing attention. Relevant research has shown a significant correlation between lipid metabolism disorders and NONFH, with abnormal lipid metabolic states being considered one of the risk factors for the progression of ONFH³⁷. It has also been discovered that certain gene polymorphisms related to lipid metabolism are associated with an increased risk of ONFH^38,39. These genes may induce or exacerbate ischemic damage to the femoral head by affecting lipid metabolic pathways. These findings underscore the critical importance of identifying lipid metabolism biomarkers in ONFH for enhancing disease management and facilitating early diagnosis. In our study, we employed a comprehensive suite of analytical techniques to achieve this goal. By conducting differential expression analysis, WGCNA, machine learning, and ROC, we successfully identified three lipid metabolism biomarkers associated with ONFH: CREBBP, GLB1, and PSAP. Combining network pharmacology analysis, we have screened out drugs such as estradiol and retinoic acid that regulate the CREBBP, GLB1, and PSAP , this drugs have regulatory effects on genes and on ONFH. Gaining a deeper understanding of their roles in ONFH can shed light on the underlying mechanisms of the disease, potentially leading to more earlier detection strategies.

Galactosidase beta 1 (GLB1) encodes a member of the 35-protein family of glycosyl hydrolases. It was shown that excess bioactive sphingolipids are also affected by GLB1 and B4GALT1 proteins, which are up-regulated to affect the balance of osteoblast and osteoclast activity^40,41. In some cases, CREBBP mutations may lead to abnormal bone metabolism, thereby increasing the risk of ONFH⁴².Notably, the higher incidence of ONFH in patients with Rubinstein-Taybi syndrome further emphasizes the importance of CREBBP in the development of ONFH⁴³. β−1,4-galactosyltransferase 1 (B4GALT1) is an important gene in the human genome, primarily involved in the glycosylation process. B4GALT1 plays a significant role in the glycosylation process, and abnormalities in glycosylation may lead to reduced development and repair capabilities of bone tissue, thereby increasing the risk of ONFH⁴⁴. Glycosylation modifications can affect intercellular signaling, disrupt normal bone metabolic processes, and cause bone cell dysfunction, which may act as a catalyst in the occurrence of ONFH. These findings suggest that the selected biomarkers may play key contributing roles in the onset and development of ONFH. To gain a deeper understanding of the biological processes or signaling pathways that these biomarkers may influence, we performed GSEA. This analysis aims to reveal how these biomarkers affect the pathological process of ONFH through specific molecular pathways.

GSEA is capable of identifying the most significant pathways or functions under specific biological conditions, thereby revealing the overall patterns of biological processes, not just local changes. Through GSEA, we had uncovered three biomarkers were all involved in the activation process of IFN-γ response. IFN-γ plays a central role in immune regulation: it not only enhances the bactericidal capacity of macrophages and promotes the activation of T cells but also regulates antibody production, these functions are crucial for maintaining immune balance⁴⁵. Furthermore, IFN-γ has a direct impact on bone metabolism by inhibiting the function of osteoblasts and reducing bone formation, while simultaneously promoting the activity of osteoclasts and increasing bone resorption, thus playing a role in the remodeling process of bone tissue⁴⁶. The role of IFN-γ in inflammatory responses is also not to be overlooked: it can both promote the production of inflammatory factors and suppress inflammation by regulating the expression of other cytokines, a bidirectional regulatory effect that is vital for controlling inflammatory responses and maintaining tissue homeostasis⁴⁷. Additionally, we have found that GLB1 and PSAP are activated in lysosomes, which is closely related to the process of autophagy. Autophagy is an essential mechanism for cells to degrade dysfunctional cellular components through lysosomes, maintaining cellular homeostasis and ensuring cell survival under stress conditions. Dysfunction of autophagy plays a pathogenic role in a variety of human diseases, including steroid-induced osteonecrosis of the femoral head⁴⁸. Moreover, autophagy also affects the interaction between osteoblasts and osteoclasts in ONFH, which has profound implications for the balance of bone tissue and the progression of the disease. These findings provide new ideas for understanding the pathways involved in biomarkers.

The combination of bone biology and immunology has recently become an important focus of ONFH research in the field of bone immunology⁴⁹. The proper functioning of the immune system is crucial for maintaining the health of bone tissue. We found that activated dendritic cells and memory B cells infiltrated less in ONFH by immune infiltration analysis. In a healthy state, there is a balance between osteoblasts and osteoclasts, with immune cells playing a regulatory role in this process. However, when the immune system is imbalanced, this equilibrium is disrupted, leading to the death of bone cells and the degeneration of bone tissue. Research has found that specific types of immune cells exhibit abnormal expression patterns in ONFH. For example, memory B cells, activated dendritic cells, M1 macrophages, monocytes, and neutrophils are all considered to play significant roles in ONFH. These cells interact by secreting various cytokines and regulating immune responses, a process that plays a key role in the pathophysiology of ONFH. Cytokines secreted by immune cells have a significant impact on the health of bone tissue⁵⁰. For instance, certain cytokines may promote bone resorption, inhibit bone formation, leading to the destruction and necrosis of bone tissue. The dysregulation of these cytokines is a key factor in the development of ONFH. In our study, we observed significant differences in neutrophil infiltration between the ONFH and control groups. ScRNA-seq analysis revealed that these three biomarkers were highly infiltrated in ONFH, suggesting the crucial role of neutrophils in the pathogenesis of ONFH. Neutrophils, which are produced in the bone marrow, play a significant role in immune functions within the systemic circulation⁵¹. Recent studies have indicated that the functions of neutrophils are heterogeneous and can be influenced by lipid metabolism⁵². Following early infiltration in ONFH, neutrophils participate in bone remodeling, and the rate of necrosis may be associated with the active immune defense and necrotic tissue clearance by neutrophils in the early stages of bone necrosis⁵³. Furthermore, the percentage of neutrophils is significantly correlated with the incidence of ONFH⁵³, indicating that neutrophil levels may play an important role in the pathogenesis of ONFH.

Deciphering the potential relationships between biomarkers and other miRNAs, etc. may provide new insights into biological functions as well as the pathogenesis of complex diseases. Hsa-miR-320a is a miRNA that plays a regulatory role in various physiological and pathological processes. It binds to the mRNA of specific target genes, inhibiting their expression and thereby affecting multiple cellular functions. In a study on ONFH, hsa-miR-320a levels were elevated in blood samples, and it was discovered that it may block osteoblast differentiation by targeting RUNX2 while promoting adipocyte differentiation of BMSCs⁵⁴. Our research has found that hsa-miR-320a might inhibit glucocorticoid-induced ONFH by targeting GLB1. GLB1 is an important bioenzyme that may be associated with the metabolism and repair processes of bone tissue. With the over-expression of GLB1, the level of hsa-miR-320a also increased significantly, which may have an effect on bone necrosis-related pathways or mechanisms as previously discovered. Following retinoic acid treatment, the level of hsa-miR-320a decreased significantly, suggesting that hsa-miR-320a may serve as a potential therapeutic target, influencing the progression of the disease by modulating the expression of GLB1. On the other hand, our study found that the key gene GLB1 is enriched in biological processes such as “lysosome”, “proteasome”, and "IFN-γ response", which are closely related to inflammation and immune responses-key factors in the pathogenesis of ONFH^4,55. Existing evidence also shows that hsa-miR-320a is involved in these processes, such as by regulating Arf1 and PBX3 to influence inflammation and cartilage degradation^56,57. Based on this, we propose that hsa-miR-320a may regulate GLB1 expression and participate in the inflammatory or immune pathways in ONFH. Meantime, gain/loss-of-function experiments, ChIP-qPCR, and dual-luciferase reporter gene assays, might be utilized to further clarify the synergistic or temporal relationship between the transcriptional (TFs) and post-transcriptional (hsa-miR-320a) levels in GLB1 regulation so as to define their biological significance in ONFH.

Conclusion and limitation

Summarizing our current findings, CREBBP, GLB1 and PSAP play key roles in regulating the development of ONFH and can be used as current biomarkers for the diagnosis of ONFH. These findings not only provided us with an in-depth understanding of the disease mechanism, but also enabled the successful construction of an efficient diagnostic model based on these biomarkers. In addition, we further analyzed in depth the influence pathways of these biomarkers in the development of ONFH, which helped us to grasp the developmental dynamics of the disease more comprehensively. We also explored the specific features and roles of the biomarkers at the cellular level.

However, there are limitations to our study, The datasets used in this study were derived from a single source, which may not fully reflect the model performance in real-world clinical settings. In the future, we plan to collect more clinical samples, expand the study cohort, and validate the model with multi-center data to further improve its clinical applicability and reliability. For the follow-up, we will also detect the protein levels of the three biomarkers in serum by ELISA to explore the feasibility of non-invasive detection. Besides, in the KEGG analysis, uncorrected P-values (P < 0.05) were used as the significance criterion, which may increase the risk of false-positive results. However, given that we focused on pathways highly relevant to the research focus and these pathways also showed significant enrichment in the GO analysis, these KEGG pathways can be considered to have high biological significance. If more pathway data are mined in the future, we will supplement additional experiments (such as qPCR and WB) for further independent verification. These studies have enhanced our understanding of the pathological process of ONFH, and they provide potential targets and biomarkers for future therapeutic strategies.

Author contributions

All authors contributed to the study conception and design. Research design: Fei Zhang and Wuxun Peng. Data analysis: Chuan Wang, Huachuan Su and ChaoDe Cen. Resources: Yuhui Peng, Tao Wang, Jian Li. Funding acquisition and supervision: Chuan Wang, Fei Zhang. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This study were funded by the Science foundation of Guizhou Provincial Health Commission(gzwkj2023–177) and the Cultivation Project of National Natural Science Foundation of Guizhou Medical University (Grant No. gyfynsfc[2023]-02).

Data availability

This study utilized several datasets, including mRNA transcriptome datasets (GSE123568 and GSE74089), and the microRNA (miRNA) transcriptome dataset GSE89587, all obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo).  The lipid metabolism-related genes (LMRGs) from the Reactome database (https://reactome.org/).  The scRNA-seq dataset SRP361778, sourced from the SRA database (https://www.ncbi.nlm.nih.gov/sra).

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-13703-y.