Screening and Analysis of Core Genes for Osteoporosis Based on Bioinformatics Analysis and Machine Learning Algorithms

Abstract

Objective

This study aimed to identify osteoporosis-related core genes using bioinformatics analysis and machine learning algorithms.

Methods

mRNA expression profiles of osteoporosis patients were obtained from the Gene Expression Profiles (GEO) database, with GEO35958 and GEO84500 used as training sets, and GEO35957 and GSE56116 as validation sets. Differential gene expression analysis was performed using the R software “limma” package. A weighted gene co-expression network analysis (WGCNA) was conducted to identify key modules and modular genes of osteoporosis. Kyoto Gene and Genome Encyclopedia (KEGG), Gene Ontology (GO), and gene set enrichment analysis (GSEA) were performed on the differentially expressed genes. LASSO, SVM-RFE, and RF machine learning algorithms were used to screen for core genes, which were subsequently validated in the validation set. Predicted microRNAs (miRNAs) from the core genes were also analyzed, and differential miRNAs were validated using quantitative real-time PCR (qPCR) experiments.

Results

A total of 1280 differentially expressed genes were identified. A disease key module and 215 module key genes were identified by WGCNA. Three core genes (ADAMTS5, COL10A1, KIAA0040) were screened by machine learning algorithms, and COL10A1 had high diagnostic value for osteoporosis. Four core miRNAs (has-miR-148a-3p, has-miR-195-3p, has-miR-148b-3p, has-miR-4531) were found by intersecting predicted miRNAs with differential miRNAs from the dataset (GSE64433, GSE74209). The qPCR experiments validated that the expression of has-miR-195-3p, has-miR-148b-3p, and has-miR-4531 was significantly increased in osteoporosis patients.

Conclusion

This study demonstrated the utility of bioinformatics analysis and machine learning algorithms in identifying core genes associated with osteoporosis.

Introduction

Osteoporosis (OP) is a prevalent bone disease that reduces bone mass and damages bone tissue, resulting in an increased risk of fractures [1]. As the global population ages, OP has become a serious public health concern affecting the elderly population in particular [2]. Age-related bone loss further exacerbates this issue, significantly impacting individuals’ quality of life and healthy lifespan worldwide. Therefore, precise prevention and treatment measures are crucial to effectively address this major public health problem [3].

Bone is a continuously evolving tissue that depends on two types of cells: osteoblasts from Bone Marrow Mesenchymal Stem Cells (BMSCs) and hematopoietic stem cell-derived osteoclasts. The equilibrium between these two cell types is essential for healthy bone tissue homeostasis. However, disturbances in this balance can lead to various medical conditions, including OP [4, 5]. OP is managed using anti-bone resorption or bone-forming drugs, yet these therapies elicit adverse side effects, including nephrotoxicity, multiple myeloma, and fractures [6]. Therefore, identifying novel drug targets and elucidating the pathogenesis of OP are pivotal in advancing its clinical management.

WGCNA is a widely used technique that employs weighted analysis to identify core genes and therapeutic targets for a range of diseases by characterizing gene-module-clinical feature relationships [7]. Furthermore, machine learning, as a rapidly evolving branch of artificial intelligence, has emerged as a valuable tool for identifying potential mechanisms, biomarkers, and therapeutic targets in diverse disease research [8].

In this study, gene chip data related to OP from the GEO database were utilized to identify differentially expressed genes and OP core genes through WGCNA [9, 10]. The core genes were determined by employing machine learning algorithms such as LASSO, SVM-RFE, and RF. Prediction and validation of miRNAs associated with these core genes may offer a molecular basis for investigating the pathogenesis of OP.

Materials and Methods

Data Retrieving

Two OP-related gene expression datasets, GSE35958 and GSE84500, including 9 and 24 bone marrow-derived mesenchymal stem cell (BMSC) samples respectively, were retrieved from the GEO database. GSE35958 comprised 4 nonosteoporotic (normal group) and 5 osteoporotic patient (OP group) samples, while GSE84500 contained induced osteogenesis (normal group) and induced adipogenesis (OP group) samples. Both datasets encompassed 12 samples derived from cell cultures at 1, 2, 3, and 7 days, with 6 samples per time point. Integration of the datasets was performed using “Limma” and “sva” R packages, by removing batch effects with the combat function as the training set. Validation was conducted using GSE35957 and GSE56116 datasets. Two additional miRNA expression datasets, GSE64433 and GSE74209, containing 25 osteoporotic and 23 normal samples, and 6 osteoporotic and 6 normal samples, respectively, were also downloaded from the GEO database.

Screening for DEGs (Differential Gene Expression)

The platform annotation information was used to convert probe identifiers in each dataset to “Entrez IDs”. Differential analysis of standardized gene expression profiles was conducted using the limma package. Bayesian equation was employed for multiple testing correction and DEGs were screened using cutoff criteria of |log2FC|> 1 and adjusted P < 0.05. Visualization of DEGs was carried out using the ggplot2 package to create volcano plots.

Weighted Gene Co-expression Network Analysis (WGCNA)

The GSE35958 and GSE84500 datasets were merged, and the Sva package’s Batch normalization function in R software (version 4.2.0) was used for data correction. Following this, the “WGCNA” and “limma” packages in R software were utilized to cluster all samples and remove outliers. A suitable soft threshold was selected to construct a scale-free network, and the module with the strongest co-expression correlation with OP was identified as the key module. Subsequently, using the “WGCNA” and “limma” packages, gene significance and gene-module correlations were computed for filtering, resulting in the output of each module’s genes.

Analysis of Gene Ontology (GO) and Kyoto Encyclopedia of Genomes (KEGG) Enrichment Functions

The DEGs were intersected with the key module genes using the “veen” package in R software. The intersection genes underwent GO and KEGG pathway enrichment analyses using the “Bioconductor” package in R software, with a screening criterion of P < 0.05 and the consideration of BP, CC, MF, and signaling pathways. Moreover, the Enrichment of Gene Set Analysis (GSEA) method, with c5.go.v7.4.symbols.gmt and c2.cp.kegg.v7.4.symbols.gmt as reference Gene Sets, was employed to achieve more comprehensive results. Using GSEA software, the corrected gene expression matrices of OP patients in the combined dataset were simulated 1000 times separately to obtain the GO and KEGG enrichment analysis results.

Screening Key Genes with Machine Learning Algorithms

Machine learning algorithms are widely employed in biomarker exploration to develop complex and refined models. In this study, we introduced three such algorithms: least absolute value convergence and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and random forest (RF). The LASSO algorithm implemented variable selection and complexity adjustment through fitting a generalized linear model using the “glmnet” R package. SVM-RFE is a supervised machine learning technique that recursively ranks features. RF scored classification variables through an iterative process of constructing a decision tree classifier model, which generated highly accurate classification features. Finally, core genes were jointly screened from the joint chip dataset using the RF approach.

Receiver-Operating Characteristic (ROC)

ROC curves were used to assess the ability of ADAMTS5, COL10A1, and KIAA0040 gene expression levels, measured in datasets GEO35958 and GEO84500, to distinguish disease states. The “pROC” package in R software was employed to plot both the evaluation and test ROC curves using datasets GEO35957 and GSE56116, respectively, to determine the accuracy of this method.

Clinical Blood Sample Acquisition

A total of 12 blood samples were collected in this study: 6 from elderly individuals with Normal plasma and 6 from elderly patients with osteoporotic plasma. Dual-energy X-ray absorptiometry was used to measure the bone mineral density, with a T Score ≤ 2.5 serving as the diagnostic criterion for OP.

Prediction of MicroRNA

We selected TargetScan (https://www.targetscan.org/vert_80/), miRDB (https://mirdb.org/) and miRWalk (http://mirwalk.umm.uni-heidelberg.de/) three miRNA prediction methods. The crossover results from the three databases were selected as the predicted mirnas.

RNA Isolation and Real-Time Quantitative PCR (qPCR) Assay

The RNA isolation was prepared from blood plasma. Total RNA was extracted with Trizol reagent and reverse-transcribed into cDNA using the PrimeScript RT Reagent Kit. Relative gene expression levels were evaluated using the TB Green Premix TaqII protocol and 2 − ∆∆CT formulae. Amplification of has-miR-148a-3p (5′-UCAGUGCACUACAGAACUUUGU3-3′), has-miR-195-3p (5′-CCAAUAUUGGCUGUGCUGCUCC-3′), has-miR-148b-3p (5′-GGCAC-CACACCTTCTACAAT-3′), has-miR-4531 (5′-AUGGAGAAGGCUUCUGA-3′), U6 (5′-CTCGCTTCGGCAGCACA-3′), ADAMTS5 (5′-CCTGCCCACCCAATGG-TAAATC-3′), COL10A1 (5′-ATGCTGCCACAAATACCCTTT-3′) and KIAA0040 (5′-ATCCCAAAGTCCCAGAGTGC-3′),GAPDH(5′-AGCCACATCGCTCAGACAC-3′) was performed. The PCR reaction was conducted at 95 °C for 30 s and followed by 40 cycles of 95 °C for 5 s and 60 °C for 34 s in the real-time PCR system.

Statistical Analysis

The Wilcoxon’s rank-sum test was employed to assess differences between Normal and OP groups. Two-sided statistical tests were conducted, with P values less than 0.05 considered statistically significant. R 4.2.2 software was utilized for statistical analysis, and visualization of the results was achieved through the ggplot2 software package (Wickham 2016).

Results

Screening for Differential Genes and Merging Data

The dataset was sorted using the limma package in R language, with screening conditions set at llog2(fold change) > 1 and P < 0.05. A total of 795 differentially expressed genes were found in the GSE35958 dataset, with 532 up-regulated genes and 263 down-regulated genes. The GSE55457 dataset yielded 504 differential genes, including 220 up-regulated genes and 284 down-regulated genes (Fig. 1A, B). Batch normalization of the combined GSE35958 and GSE84500 datasets was conducted using the Sva package in R software (version 4.2.0) (Fig. 1C).

Fig. 1.

A and B Volcano plots of DEGs, green dots are up-regulated expressed genes, yellow dots are down-regulated expressed genes, and black dots are expressed nondifferentially; C gene expression profiles removing batch effects; D PCA plots

Construction of Key Modules

Clustering of samples using “WGCNA” and “limma” packages in R software was followed by analysis of scale-free index and average connectivity. For B = 3, the network achieved a scale-free topological broken value of 0.9 with appropriate end node connectivity, leading to construction of scale-free network (Fig. 2A, B) and co-expression matrix network. By implementing dynamic tree cutting and calculation, five gene modules (Fig. 2D) were identified as blue (249 genes), green (68 genes), brown (215 genes), turquoise (837 genes), and yellow (148 genes). Correlation analysis between the modules and the normal/OP groups was conducted, generating a correlation heat map (Fig. 2C). The brown module, which exhibited a strong correlation with OP (r = 0.7, P = 5.6E−33), was established as the key module. Ultimately, 99 core genes within the brown module were selected based on filtering conditions of gene significance (0.5) and gene-module correlation (0.8) (Fig. 2E).

Fig. 2.

A Soft threshold parameters versus scale independence analysis; B soft threshold parameters versus mean connectivity analysis (optimal soft threshold is 3); C correlation heat map of clinical traits with gene modules, the numbers in the cells indicate the correlation coefficient r and the corresponding P value, respectively; D gene clustering tree with gene modules; E scatter plot of brown module genes

Enrichment Analysis of Intersecting Genes and Biological Functions

The intersection of 1280 DEGs and 215 key module genes produced 151 intersection genes (Fig. 3A). Using the STRING database, we constructed a PPI network which was visualized with Cytoscape software (Fig. 3B). Subsequently, using R software, GO and KEGG enrichment analysis was performed on the 215 intersection genes. Biological processes such as negative regulation of transmembrane receptor protein serine/threonine kinase signaling pathway, regulation of transmembrane receptor protein serine/threonine kinase signaling pathway, and transmembrane receptor protein serine/threonine kinase signaling pathway were enriched. Cellular components comprised collagen-containing extracellular matrix, actin cytoskeleton, and Golgi lumen. Molecular functions mainly included transforming growth factor beta receptor binding, growth factor activity, and receptor ligand activity. Among the twelve KEGG pathway enrichments, the TGF-beta signaling pathway, p53 signaling pathway, and FoxO signaling pathway were the most relevant (Fig. 3C). GSEA analysis showed significant enrichment of Ribosome, Eukaryotic Translation Initiation, Eukaryotic Translation Elongation, SRP-dependent Cotranslational Protein Targeting To Membrane, and Gpcrs Class A Rhodopsinlike pathways in the gene expression matrix of OP patients (Fig. 3D).

Fig. 3.

A Wayne plot of DEGs and brown module core genes; B PPI plot of intersecting genes; C GO and KEGG enrichment analysis of intersecting genes; D GSEA enrichment analysis plot

Machine Learning Screening for Core Genes

In this study, the LASSO regression algorithm, implemented via the R package “glmnet,” identified 18 meaningful genes among 151 intersection genes (Fig. 4A, B). Similarly, the SVM-RFE algorithm, implemented through the R package “e1071,” removed SVM-generated variables and identified 10 meaningful genes (Fig. 4C). Random forest analysis, using the “randomForest” package, selected 10 genes and accurately predicted continuous variables with enhanced sensitivity and specificity (Fig. 4D).

Fig. 4.

A LASSO regression intersection and validation process, the point with the smallest error shows the corresponding number of core genes is 10; B LASSO regression coefficient plot with different penalty parameter values; C SVM-RFE algorithm to get 10 feature genes; D RF algorithm to rank the important genes; E Venn diagram of three machine learning algorithms to screen OP core genes; F qPCR validation of mRNA expression in the normal and OP groups

The integration of results from LASSO, SVM-RFE, and RF screening led to the identification of three core genes: ADAMTS5, COL10A1, and KIAA0040 (Fig. 4E).The qPCR assays revealed a significant decrease in the expression of COL10A1 in the OP group compared to the Normal group (Fig. 4F).

Core Genetic Diagnostic Efficacy Evaluation

We conducted ROC analysis based on the expression levels of core genes in the Normal and OP groups to diagnose OP. The AUC values of the three core genes (ADAMTS5, COL10A1, KIAA0040) were >0.7, indicating high diagnostic accuracy with sensitivity and specificity (Fig. 5A). In the independent test datasets GSE35957 and GSE56116, the AUC values of COL10A1 were 0.960 and 1.00, respectively (Fig. 5E, I), which highlights its potential as a diagnostic factor for OP. Moreover, the expression level of COL10A1 was lower in the OP group than in the control group (Fig. 5C, G, K), suggesting its contribution in the pathogenesis of OP.

Fig. 5.

A ROC curves of core genes in GSE35958 and GSE84500 datasets, B, C, D: expression of core genes in GSE35958 and GSE84500 datasets. E, I ROC curves of core genes in GSE35957 and GSE56116 datasets, F, G, H, J, K, and L expression of core genes expression in the GSE35957 and GSE56116 datasets. *P < 0.05, **P < 0.01, ***P < 0.001

Prediction and Validation of OP-Related miRNAs

This study employed TargetScan, miRDB, and miRWalk databases to predict 25 miRNAs (Fig. 6A) and screened differentially expressed miRNAs from GSE64433 and GSE74209 datasets (Fig. 6C, D). Differential miRNAs were also identified from GSE74209 datasets (Fig. 6C, d). The intersection of predicted and differential miRNAs yielded four core miRNAs (has-miR-148a-3p, has-miR-195-3p, has-miR-148b-3p, and has-miR-4531) (Fig. 6E). Importantly, qPCR results indicated high expression levels of has-miR-195-3p, has-miR-148b-3p, and has-miR-4531 in the serum of OP patients (Fig. 6F).

Fig. 6.

A Venn diagram of miRNAs predicted by three methods for COL10A1; B network diagram of COL10A1 with miRNAs; C, D volcano diagram of differential miRNAs in GSE64433 and GSE74209 datasets; E Venn diagram of differential miRNAs with predicted miRNAs in GSE64433 and GSE74209 datasets. F qPCR validation of miRNA expression in the Normal and OP groups

Discussion

This study aimed to investigate the pathogenesis of OP, a degenerative systemic bone metabolic disease characterized by reduced bone mass, bone resorption, and microstructural deterioration leading to fractures and impacting health [11, 12]. To achieve this goal, bioinformatics and WGCNA were employed to identify differentially expressed genes and core genes associated with OP, which were further screened using machine learning algorithms (Table 1). Biological functions and diagnostic efficacy of the core genes were analyzed, and qPCR experiments were conducted to confirm differential mirnas in OP patients. The findings provide valuable insights into the molecular-level understanding of OP and can guide future investigations.

Table 1.

Genes selected by three machine learning methods

Bone marrow mesenchymal stem cells (BMSCs) are pluripotent cells that exhibit self-renewal, differentiation, and immune regulation. Their therapeutic potential for various diseases has been widely studied [13–18]. BMSCs have the ability to differentiate into bone, muscle, fat, tendon, cartilage, and bone marrow matrix. Age-related declines in BMSC proliferation and differentiation, along with accelerated senescence, lead to an imbalance between adipocytes and osteoblasts, reducing bone formation and increasing the risk of OP in the elderly [19–23]. Although research on OP mechanisms emphasizes balance between osteoblasts and osteoclasts, evidence increasingly suggests that as the progenitor cells of osteoblasts, BMSCs play a critical role in maintaining this balance.

Transcriptome sequencing data of BMSCs from normal and OP groups were obtained from the GEO database, yielding 1280 DEGs. Using WGCNA, a disease key module and 215 key genes were identified. Intersection of DEGs with module key genes resulted in 151 intersection genes, with KEGG pathway enrichment analysis showing the TGF-beta signaling pathway as the most significant. TGF-β is critical for regulating cell growth, promoting BMSC differentiation into osteoblasts, and modulating the TGF-β-Smad signaling pathway, which is essential for stem cell renewal, cell proliferation, differentiation, migration, and apoptosis. Indeed, TGF-β activity is important for postnatal bone stabilization, including osteoblast proliferation and differentiation and bone formation. Disrupting the TGF-β-Smad signal transduction pathway can cause bone metabolism disorders and OP. Therefore, targeting this pathway may be a useful strategy for treating OP [24–26].

A machine learning algorithm identified three OP core genes (ADAMTS5, COL10A1, KIAA0040). Diagnostic test performance was evaluated using ROC curves and AUC, with an optimal score of 1. Among the three genes, COL10A1 demonstrated good specificity and sensitivity in the GSE35958 and GSE84500 datasets, with AUC values of 0.743. In the GSE35957 and GSE56116 datasets, COL10A1 exhibited highly specific and sensitive diagnostic performance with AUC values of 0.960 and 1.00, respectively. Therefore, it may serve as a promising OP biomarker.

COL10A1, or Collagen X, is a cartilage-specific type X collagen located in the q21−q22 region of human chromosome 6. It is primarily expressed in hypertrophic chondrocytes and distributed in growth plate cartilage during embryonic and developmental stages. COL10A1 induces calcification of pericellular chondrocyte matrix and endochondral ossification. Although mutations and expression changes of COL10A1 have been linked to various bone diseases, its association with OP is scarce [27–29]. A case report of a premenopausal woman with idiopathic OP showed mutations in both COL10A1 and COL27A1 [30]. Given previous research on cartilage mast cells and COL10A1, we propose that decreased expression of COL10A1 in BMSCs may contribute to OP via the TGF-β-Smad signaling pathway.

We investigated the involvement of COL10A1 in OP by comparing its predicted miRNA targets with differentially expressed miRNAs from two datasets (GSE64433, GSE74209). We identified four key miRNAs (has-miR-148a-3p, has-miR-195-3p, has-miR-148b-3p, has-miR-4531), and observed significantly higher expression of has-miR-195-3p, has-miR-148b-3p, and has-miR-4531 in OP patients as determined by qPCR. Our results suggest that miRNA-dependent regulation of COL10A1 expression is likely to influence osteogenic differentiation of BMSCs.

RUNX2(Runt-related transcription factor 2) and NFATc1(nuclear factor of activated T cell 1) are crucial transcription factors involved in the commitment of mesenchymal stem cells to osteoblasts and hematopoietic stem cells to osteoclasts, respectively [31, 32]. The regulation of osteoblast differentiation includes the participation of various miRNA pathways, such as miR-23b, which influences the osteogenic differentiation of bone marrow-derived mesenchymal stem cells (BMSCs) by targeting RUNX2 [33]. In patients with osteoporosis, there is a notable upregulation of miR-133a in serum. Further investigations reveal that reducing miR-133a levels inhibits the translation of NFATc1, thereby suppressing the differentiation of RAW264.7 and THP-1 cells into osteoclasts mediated by RANKL [34]. In addition, qPCR experiments demonstrate elevated expression of has-miR-195-3p, has-miR-148b-3p, and has-miR-4531 in osteoporosis patients, suggesting their potential involvement in promoting osteoporosis development through the regulation of RUNX2 and NFATc1.

In this study, we employed GEO dataset, WGCNA, and three machine learning algorithms to identify three genes crucial for the development of OP. COL10A1 displayed favorable diagnostic efficacy in the validation cohort and was found to be down-regulated in osteoporotic BMSCs. Additionally, miRNAs predicted by COL10A1 exhibited differential expression in the blood of OP patients.

Our findings indicate that COL10A1 and its associated miRNAs may serve as potential biomarkers for the diagnosis and treatment monitoring of OP. Additional functional experiments are necessary to clarify the involvement of these vital genes and miRNAs in OP pathogenesis. Thus, future experimentation is warranted to validate the functions of these key genes and miRNAs in OP.

Abbreviations

BP

Biological process

CC

Cellular component

DEGs

Differentially expressed genes

GEO

Gene Expression Omnibus

GESA

Gene set enrichment analysis

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

MF

Molecular function

miRNAs

Predicted microRNAs

OP

Osteoporosis

WGCNA

Weighted gene co-expression network analysis

Funding

Chengdu Medical Research Project (2023403).

Availability of Data and Materials

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Conflict of Interest

The authors declare that they have no conflict of interest.

Ethics Approval and Consent to Participate

This study was approved by the medical ethics committee of Chengdu Seventh People’s Hospital People’s Hospital. Informed consent was obtained from all patients prior to participation.

Informed Consent

For this type of study informed consent is not required.

Consent for Publication

Not applicable.