Application of machine learning-based exosome-related gene profiles in precision diagnosis and treatment of osteoarthritis Authors: Chuanfei You, Furen Dai, Bingzhu Dai, Weijun Wu, Le Fang, Weimin Jia, Xu Han, Zhi Su, Jian Li Published: 2025 ABSTRACT AbstractOsteoarthritis (OA) involves a complex pathogenesis encompassing inflammation, metabolic dysregulation, and aberrant intercellular communication. Despite their crucial role as mediators of intercellular signaling, exosomes remain largely underexplored in OA. This study aims to investigate exosome-related genes (ERGs) and their roles in OA pathogenesis. Four OA-related gene expression datasets retrieved from GEO were harmonized using the ComBat algorithm to mitigate batch effects. Differentially expressed genes (DEGs) were identified through differential expression analysis and weighted gene co-expression network analysis (WGCNA). ERGs were screened utilizing the ExoCarta and Vesiclepedia databases. Core ERGs were prioritized using LASSO, random forest, and XGBoost algorithms. Predictive models were constructed and subsequently evaluated using SHAP analysis to ascertain feature importance. Additionally, pathway enrichment, immune infiltration, and molecular subtype identification were performed, followed by validation of core ERG expression via RT-qPCR in clinical samples. Integration of the four GEO datasets yielded 231 DEGs significantly enriched within OA-associated pathways (e.g., inflammation, immune cell migration, extracellular matrix remodeling). From a pool of 79 candidate ERGs, 10 core ERGs (EPB41L2, ISLR, HLA-DRB1, HLA-DRA, PGLYRP1, PTEN, TKT, CTNNB1, THSD4, ATP9A) were identified. Random forest models achieved impressive AUCs of 0.991, 1.0, and 0.935 in the training, validation, and external validation sets, respectively, demonstrating substantial clinical net benefit. SHAP analysis underscored CTNNB1 and PGLYRP1 as pivotal predictors. Core ERGs were intricately linked to immune regulation (e.g., M1 macrophage infiltration) and metabolic perturbations (e.g., fatty acid metabolism). Distinct molecular subtypes of OA were delineated based on ERG profiles, thereby revealing the inherent disease heterogeneity. RT-qPCR further corroborated the differential expression of core ERGs in clinical samples. This study comprehensively integrates exosome-related genomic data with advanced machine learning techniques to identify and validate 10 core ERGs associated with OA, thereby elucidating their pivotal roles in immunometabolic regulation. These seminal findings illuminate the intricate molecular heterogeneity of OA, concurrently offering promising novel biomarkers and therapeutic targets for early diagnosis and precision treatment.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33127-y.Keywords: Osteoarthritis, Exosome-related genes, Machine learning, Diagnostic biomarkers, ExosomesSubject terms: Biomarkers, Computational biology and bioinformatics, Diseases, Genetics, Immunology Introduction Osteoarthritis (OA), a pervasive degenerative inflammatory joint disorder, manifests through clinical symptoms such as articular pain, edema, and structural deformity1,2. Globally, over 500 million individuals are afflicted by OA, with predisposing factors encompassing advanced age, obesity, and genetic susceptibility3–5. Presently, joint arthroplasty remains a cornerstone of effective OA treatment. Complementary non-surgical interventions, including pharmacotherapy, intra-articular injections, tailored exercise regimens, and prudent dietary habits, constitute integral facets of OA management6. Nevertheless, the multifaceted pathogenesis of OA, driven by the interplay of articular damage, genetic influences, and metabolic dysregulation, remains only partially elucidated7,8. Consequently, unraveling these underlying mechanisms to facilitate early prevention and intervention is of utmost significance1. In recent years, heightened scrutiny has been devoted to the molecular underpinnings of OA, particularly the dynamic interplay between host tissues and their microenvironment. Within this complex milieu, exosomes have emerged as pivotal mediators of intercellular communication9. These nanoscale, membrane-enclosed vesicles, secreted by cells, encapsulate a diverse repertoire of biomolecules, including lipids, proteins, DNA, and various RNA species10–12. Exosomes exert profound influence on intercellular signaling and are implicated in the progression of myriad diseases13. For example, in oncology, exosomes derived from hepatocellular carcinoma can enhance PD-L1 expression in tumor-associated macrophages through the STAT3 signaling pathway, thereby facilitating tumor immune evasion14. Moreover, dysregulated exosomal miRNAs, as potential diagnostic and prognostic biomarkers, are intricately linked to breast cancer progression, modulating processes such as cellular proliferation, differentiation, metastasis, and therapeutic resistance15,16. These findings underscore the critical regulatory roles of exosomes in disease initiation and progression. Yet, the specific contributions and mechanisms of exosomes in OA remain underexplored, necessitating rigorous investigation to harness their therapeutic potential. Currently, microarray technology and integrative bioinformatics analyses are extensively utilized to pinpoint potential biomarkers and delineate their roles across various pathologies17. The Gene Expression Omnibus (GEO), an international public repository, curates and organizes high-throughput sequencing data. This study endeavors to comprehensively evaluate the expression profiles of exosome-related genes during the onset, progression, and management of OA. By employing machine learning methodologies, we aim to identify and validate pivotal exosome-related genes, thereby offering novel biomarkers and therapeutic targets to advance the clinical diagnosis and treatment of OA. Materials and methods Data acquisition and preprocessing Six datasets (GSE12021, GSE51588, GSE55235, GSE55457, GSE169077, and GSE82107), encompassing gene expression profiles from osteoarthritis (OA) patients and healthy controls, were procured from the Gene Expression Omnibus (GEO) database. Platform annotation files were utilized to generate the corresponding gene expression matrices. To rectify batch effects and harmonize the data structure, the “ComBat” algorithm within the “sva” R package was applied to amalgamate GSE12021, GSE51588, GSE55235, and GSE55457 into a unified training cohort, while the remaining datasets (GSE169077 and GSE82107) were reserved as independent validation sets. Comprehensive details regarding the GEO datasets are delineated in Table 1. Table 1Microarray information for gene expression Datasets.TypeDatasetSourceSamplePlatformGroupControlOAGEOGSE12021synovium1319GPL96Training setGSE51588synovium520GPL13497GSE55235synovium2026GPL96GSE55457synovium1010GPL96GSE169077synovium56GPL96External validation setGSE82107synovium710GPL570 Microarray information for gene expression Datasets. To curate a panel of Exosome-Related Genes (ERGs) implicated in intercellular communication and inflammatory cascades within OA pathogenesis, we primarily interrogated the ExoCarta database, a manually curated repository of exosomal molecular constituents. Complementary data were assimilated from Vesiclepedia to ensure broad coverage of extracellular vesicle cargo18–20. Following the exclusion of duplicates and intersection analysis with human protein-coding genes, a total of 922 unique ERGs were identified. The complete inventory of these ERGs is tabulated in Supplementary Table 1. Screening of differentially expressed genes (DEGs) Differentially expressed genes (DEGs) between osteoarthritis (OA) and healthy control cohorts were identified using the “limma” package in R, with thresholds established at |log₂ fold change| > 0.585 and p < 0.05. Volcano plots and heatmaps were generated using the “ggplot2” and “pheatmap” packages, respectively, to delineate differential expression patterns. Weighted gene co-expression network analysis (WGCNA) Weighted Gene Co-Expression Network Analysis (WGCNA), a robust systems biology framework, constructs gene co-expression networks from high-throughput genomic data21. In WGCNA, genes are represented as nodes, with edges reflecting the strength of co-expression across samples. In this study, WGCNA was utilized to identify mRNAs associated with OA. Using the “WGCNA” R package, co-expression networks for mRNAs were constructed from GEO datasets. An optimal soft-thresholding parameter β was selected to ensure a scale-free network topology. Weighted adjacency and topological overlap matrices were generated to establish the co-expression network. Dynamic tree cutting was applied to delineate correlated modules, with target modules selected based on their significant correlation with mRNAs. Hierarchical clustering was subsequently performed to identify modules comprising at least 60 genes, and similar modules were merged. Key modules were identified through correlation analysis with clinical phenotypes, and the intersection of DEGs and WGCNA module genes was leveraged to refine the selection of key exosome-related genes (ERGs). Machine learning-based screening of ERGs To identify pivotal exosome-related genes (ERGs), three machine learning algorithms were integrated for feature selection: Extreme Gradient Boosting (XGBoost), Least Absolute Shrinkage and Selection Operator (LASSO) regression, and Random Forest (RF). XGBoost captures intricate nonlinear relationships among genes through ensemble tree models, delivering robust predictive performance and feature importance rankings. LASSO regression, with its inherent sparsity penalty, eliminates redundant or less relevant genes, facilitating variable selection and dimensionality reduction. Random Forest constructs multiple decision trees and aggregates their outputs, ensuring stable gene importance assessments and mitigating overfitting risks. The synergistic integration of these methodologies enhances the robustness and precision of key gene identification, pinpointing gene markers central to exosome-related biological processes. Multi-model construction and SHAP interpretation analysis Based on the intersected feature genes identified from multiple approaches (LASSO, random forest, and XGBoost), five commonly used machine learning algorithms were selected to construct predictive models for osteoarthritis classification: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine with Radial Basis Function Kernel (SVM-Radial), Neural Network (NNET), and Naive Bayes. These models were implemented in R (version 4.3.2) primarily using the “caret” package, which provides a unified interface for model training, hyperparameter tuning, and evaluation. Additional packages included “randomForest,” “xgboost,” “e1071,” and “nnet” where necessary. To enhance methodological robustness and address potential bias from a single data split (as highlighted in reviewer comments), we adopted a repeated cross-validation and bootstrap resampling strategy. Specifically, the dataset was subjected to 100 iterations of bootstrap resampling (with replacement). In each iteration, samples were randomly partitioned into training (70%) and testing (30%) sets. The training set was used for hyperparameter optimization via 10-fold cross-validation (repeated 3 times), and the final model was evaluated on the held-out testing set. Performance metrics (including accuracy, AUC-ROC, sensitivity, specificity, precision, and F1-score) were recorded for each iteration, with median values and 95% confidence intervals (percentile method) reported as the final results to account for variability across resamples. To interpret the contributions of individual features to model predictions, SHAP (SHapley Additive exPlanations) analysis was performed on the best-performing model (e.g., the top-ranked model based on median AUC-ROC). SHAP values were calculated using the “kernelshap” and “shapviz” packages in R, which support model-agnostic explanations compatible with caret-trained models. Global feature importance was derived from mean absolute SHAP values, visualized via summary plots and bar plots. Local interpretations included SHAP force plots for individual samples and dependence plots to reveal feature interactions and contribution directions (positive or negative impact on predictions). Validation of hub genes and clinical model construction The precision of hub genes was validated using the integrated dataset. Nomograms were developed using hub genes via the “rms” package in R. Their predictive accuracy was assessed through calibration curves. The model’s precision and clinical utility were rigorously evaluated using calibration curves, Decision Curve Analysis (DCA), and clinical impact curves. Gene set variation analysis (GSVA) To elucidate functional differences associated with the upregulation and downregulation of core exosome-related genes (ERGs), Gene Set Variation Analysis (GSVA) was conducted22. GSVA quantifies expression activity scores for gene sets to delineate their associations with specific biological processes, functions, or pathways. Utilizing the “GSVA” R package, enrichment analysis of core ERGs was performed based on KEGG and GO gene sets to illuminate their potential biological roles during upregulation and downregulation. Immune infiltration analysis To investigate immune infiltration disparities between osteoarthritis (OA) patients and healthy controls, the relative abundance of immune cells was quantified using CIBERSORT, with results visualized via the “ggplot2” package. Additionally, routine hematological data from 89 healthy individuals and 150 OA patients, encompassing monocytes, neutrophils, lymphocytes, and eosinophils, were analyzed to corroborate the immune infiltration findings. Identification of molecular subtypes in osteoarthritis based on core ERGs Using expression profiles of core exosome-related genes (ERGs), unsupervised hierarchical clustering was performed on 54 osteoarthritis (OA) samples with the “ConsensusClusterPlus” R package23. Parameters were configured as follows: maxK = 6, reps = 50, pItem = 0.8, pFeature = 1, clusterAlg = “pam,” and distance = “euclidean.” The optimal cluster number was determined using the Calinski-Harabasz criterion and inter-subgroup correlation analysis. Subtype distribution was subsequently visualized using the “Rtsne” package, and differential expression of key regulatory ERGs across subtypes was evaluated using the Kruskal-Wallis test. Validation of core ERGs by real-time quantitative PCR To substantiate the expression of core ERGs, synovial tissue samples were obtained from 12 OA patients and 10 healthy volunteers with informed consent. All methods were performed in accordance with the relevant guidelines and regulations. Primers were designed using the Primer3Plus tool and validated against the NCBI database; primer sequences are provided in Table 2. Total RNA was extracted, reverse-transcribed into cDNA, and real-time quantitative PCR (RT-qPCR) experiments were conducted per the manufacturer’s protocol. Relative gene expression was quantified using the 2^-ΔΔCt method to ensure precision and reproducibility. Table 2Primers sequence.GenePrimer (5′−3′)β-actinF: GGCTGTATTCCCCTCCATCGR: CCAGTTGGTAACAATGCCATGTEPB41L2F: TCTTTTGTGACTCATGCTCTCCR: CACCATACATGGAAAGCCTCTTHLA-DRB1F: CGGGGTTGGTGAGAGCTTCR: AACCACCTGACTTCAATGCTGHLA-DRAF: AGTCCCTGTGCTAGGATTTTTCAR: ACATAAACTCGCCTGATTGGTCPGLYRP1F: GCCTGCCCTTACGCTATGTGR: CAGGAAGTTGTAGCCCACGTCPTENF: TGGATTCGACTTAGACTTGACCTR: GGTGGGTTATGGTCTTCAAAAGGTKTF: TCCACACCATGCGCTACAAGR: CAAGTCGGAGCTGATCTTCCTCTNNB1F: AAAGCGGCTGTTAGTCACTGGR: CGAGTCATTGCATACTGTCCATTHSD4F: CCTGGCAAGTATGGCTATGGTR: CCATGCCTAGCAGAAGATGCATP9AF: AGAGGTATCCTCGGAATGTCATCR: CTCATTTCGGGAACAAACTGAGAISLRF:TGGCATGAGAGAGAAGTGCCR:CTTACGAGAGCCACCTACAG Primers sequence. Statistical analysis Differences in continuous variables between healthy control (HC) and osteoarthritis (OA) cohorts were analyzed using the nonparametric Wilcoxon rank-sum test, with a threshold of p < 0.05 denoting statistical significance. To ensure robustness, all experimental data were derived from at least three independent replicates. Data preprocessing involved outlier detection and normality assessment via the Shapiro-Wilk test to validate the use of nonparametric methods. Statistical analyses were conducted using R software (version 4.4.1) for data processing and hypothesis testing, with visualizations (e.g., bar graphs, scatter plots) generated using Prism 10 (GraphPad Software, USA) to compute means and standard errors. Significance levels were denoted as: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, to clearly delineate inter-group differences. Results Identification of differentially expressed genes To elucidate gene expression alterations in osteoarthritis (OA) tissues, four publicly available gene expression datasets (GSE12021, GSE51588, GSE55235, and GSE55457) were integrated, with batch effect correction performed using the ComBat algorithm. Prior to correction, sample expression distributions across datasets exhibited marked variability (Fig. 1A), and principal component analysis (PCA) revealed distinct batch-driven clustering (Fig. 1C), underscoring pronounced batch effects. Post-correction, expression distributions achieved uniformity (Fig. 1B), and PCA demonstrated that samples clustered by biological variability rather than batch (Fig. 1D), validating the robustness of the data integration. Batch Effect Correction and Differential Expression Analysis of RNA-seq Data. (A) Boxplot illustrating expression distribution across samples from datasets (GSE12021, GSE51588, GSE55235, GSE55457) pre-correction, with distinct colors denoting each dataset, revealing evident batch effects. (B) Boxplot depicting expression distribution post-correction, demonstrating effective mitigation of batch effects and uniform expression across datasets. (C-D) PCA clustering plots pre- and post-correction, illustrating reduced batch-driven clustering and enhanced biological variability post-correction. (E) Heatmap of the merged dataset (GSE12021, GSE51588, GSE55235, and GSE55457) after batch effect correction. Samples were ordered by group (healthy controls followed by OA patients) and dataset, without hierarchical clustering, to illustrate expression patterns post-correction. (F) Volcano plot of differential expression analysis. Significantly upregulated genes (green), downregulated genes (orange), and non-significant genes (gray) are shown, with selected genes labeled. (G) GO enrichment analysis results for DEGs, sorted by pathway significance (-log₁₀ p-value), with point size indicating gene count and color intensity reflecting enrichment significance. Differential expression analysis, conducted using the “limma” package, identified 231 significantly differentially expressed genes (DEGs) (adjusted p-value < 0.05, |log₂ fold change| > 0.585, corresponding to a minimum 1.5-fold expression change). These DEGs included established OA-associated genes such as MMP13 (matrix metalloproteinase 13), COL1A1 (collagen type I alpha 1), IL6 (interleukin 6), and TNF (tumor necrosis factor). A heatmap delineated distinct expression profiles between OA and control cohorts (Fig. 1E), while a volcano plot confirmed significant upregulation or downregulation trends for these genes (Fig. 1F). Gene Ontology (GO) functional enrichment analysis, executed via the clusterProfiler package, elucidated that Differentially Expressed Genes (DEGs) were predominantly enriched in biological processes pertaining to inflammation and immune regulation—specifically leukocyte migration, chemotaxis, cytokine activity, and the regulation of inflammatory responses. Additionally, terms associated with Extracellular Matrix (ECM) remodeling, such as ECM organization and ossification, were significantly overrepresented. All identified pathways exhibited statistical significance with adjusted P-values < 0.05 (Fig. 1G). Collectively, these findings underscore inflammation-mediated immune cell infiltration, ECM metabolic dysregulation, and aberrant angiogenesis as pivotal drivers of Osteoarthritis (OA) progression. This interpretation is highly concordant with the established pathophysiology of OA, characterized by synovitis, cartilage degeneration, and pathological tissue remodeling. Construction of OA co-expression network and identification of key ERGs To investigate key gene modules associated with osteoarthritis (OA) and their functional roles, weighted gene co-expression network analysis (WGCNA) was conducted on the integrated expression data. Sample hierarchical clustering confirmed high-quality data with no significant outliers and revealed distinct expression patterns between cohorts (Fig. 2A). Gene hierarchical clustering and dynamic tree cutting identified 16 co-expression modules, reflecting intrinsic correlations and synergistic expression patterns (Fig. 2B). A soft-thresholding power of 6 was selected to satisfy scale-free topology criteria, ensuring the biological robustness of the co-expression network (Fig. 2C). Clustering and Gene Module Analysis Based on Weighted Gene Co-Expression Network Analysis (WGCNA). (A) Sample clustering dendrogram and phenotype heatmap illustrating clustering relationships among samples and their phenotypic distribution. (B) Gene clustering dendrogram with module color assignments, delineating the classification of gene modules. (C) Plot of network scale-free topology and mean connectivity as a function of soft-threshold power. (D) Module-trait relationships in WGCNA analysis.The number in each cell represents the correlation coefficient between the module (row) and the trait (column). The color intensity indicates the strength of the correlation (see color scale). (E) Module-trait correlation heatmap and Venn diagram showing gene overlap between WGCNA modules and differentially expressed genes (DEGs). Modular-phenotype correlation analysis revealed a significant positive correlation between the MEgrey module and the OA phenotype (correlation coefficient: 0.29, p < 0.01), indicating that genes within this module may play a pivotal role in OA pathogenesis (Fig. 2D). Subsequent integration of WGCNA-derived modules, differentially expressed gene sets (DEGs), and exosome-related gene sets via Venn diagram analysis identified 79 overlapping genes. These genes exhibit multifaceted attributes, encompassing differential expression, functional relevance, and exosomal regulation, thereby positioning them as core candidate molecules in osteoarthritis research (Fig. 2E). Machine learning-based screening of key osteoarthritis genes This study integrated three machine learning methodologies—LASSO regression, Random Forest, and XGBoost—to systematically evaluate and prioritize differentially expressed genes associated with osteoarthritis (OA). LASSO regression, through its regularization mechanism, effectively identified a subset of genes with optimal predictive value, enhancing model performance (Fig. 3A). Subsequently, the XGBoost model quantified the contribution of each gene to disease classification, highlighting TKT, RAB35, and PGLYRP1 as top contributors (Fig. 3B). Random Forest analysis corroborated these findings, with consistent gene importance rankings based on “MeanDecreaseAccuracy” and “MeanDecreaseGini” metrics (Fig. 3C-D). Screening of Key Genes and Expression Analysis Using Multiple Machine Learning Methods. (A) Coefficient path plot and cross-validation for the optimal error parameter in the LASSO regression model. (B) Feature importance by gain in the XGBoost model. (C) Feature importance plots for random forest based on MeanDecreaseAccuracy (left) and MeanDecreaseGini (right). (D) Top 20 features ranked by MeanDecreaseGini from the random forest model. (E) Venn diagram showing the overlap of genes selected by LASSO, random forest (RF), and XGBoost models. (F) Correlation heatmap of the top selected features. (G) Boxplot of key gene expression in control versus OA cohorts, with statistical significance markers (*p < 0.05, **p < 0.01, ***p < 0.001). (H) Boxplot comparing key gene expression in control versus OA cohorts in an external validation set, with statistical significance markers (*p < 0.05, **p < 0.01, ***p < 0.001). Intersection analysis of these machine learning models yielded 10 candidate core genes: EPB41L2, ISLR, HLA-DRB1, HLA-DRA, PGLYRP1, PTEN, TKT, CTNNB1, THSD4, and ATP9A (Fig. 3E). These genes exhibited robust expression correlations, suggesting synergistic roles in OA pathogenesis (Fig. 3F). Expression analysis revealed significant differential expression of these 10 exosome-related genes (ERGs) in OA tissues compared to controls (Fig. 3G), with consistent patterns observed in an external validation cohort (Fig. 3H). Gene Set Variation Analysis (GSVA) demonstrated that the expression profiles of these 10 genes significantly modulated multiple OA-related KEGG pathways (Figure SF1A-J). Notably, the low-expression group of ATP9A was markedly enriched in fatty acid metabolism pathways, indicating a potential role in regulating energy metabolism and maintaining joint tissue homeostasis. The high-expression group of CTNNB1 was enriched in tumor- and immune response-related signaling pathways, suggesting its involvement in inflammatory processes and tissue remodeling in OA. Immune-regulatory genes HLA-DRA and HLA-DRB1, in their high-expression groups, significantly activated immune pathways, underscoring their pivotal role in immune-mediated inflammation in OA pathogenesis. The high-expression group of PTEN was enriched in protein degradation and cellular repair pathways, potentially contributing to cartilage repair and cellular homeostasis. Collectively, these genes influence OA onset and progression by modulating metabolic dysregulation, inflammatory responses, and cellular repair mechanisms, positioning them as promising biomarkers and therapeutic targets for OA diagnosis and management. Machine learning model construction This study developed and systematically evaluated the diagnostic efficacy of multiple machine learning algorithms for osteoarthritis (OA). Within the training cohort, the Random Forest and Gradient Boosting models exhibited superior performance, yielding AUCs of 0.991 (0.978–1.000) and 0.981 (0.960–1.000), respectively, indicative of exceptional classification precision (Fig. 4A and Supplementary Table 2). Analyses of the independent validation set further substantiated the generalizability of the models: the Random Forest model attained perfect accuracy (AUC = 1.0), while other classifiers such as SVM and XGBoost preserved high AUC metrics (Fig. 4B and Supplementary Table 2). Moreover, in the external independent validation dataset, the Random Forest model sustained a robust AUC of 0.935 (0.930–0.940), underscoring its robustness across heterogeneous data environments (Fig. 4C and Supplementary Table 2). Performance Evaluation of Multiple Machine Learning Models Across Training, Validation, and External Validation Sets. (A-C) ROC curves for models in the training (A), validation (B), and external validation (C) sets, with higher AUC values indicating superior performance. (D-F) Decision Curve Analysis (DCA) for models in the training (D), validation (E), and external validation (F) sets, evaluating clinical utility and net benefit. (G-I) Confusion matrices for models in the training (G), validation (H), and external validation (I) sets, illustrating prediction accuracy and misclassification rates. Comprehensive evaluation of classification performance revealed that, in the training set, all models exhibited robust learning capabilities, with accuracy, F1 score, negative predictive value (NPV), and positive predictive value (PPV) approaching 1 (Figure SF2A), reflecting excellent fit to the training data. In the validation set, most models sustained high accuracy and sensitivity, with strong F1 scores indicating reliable positive detection (Figure SF2B). However, variability in specificity and Youden’s index suggested potential overfitting in certain models, particularly in negative sample identification. In the external validation dataset, ensemble tree models (e.g., Random Forest, Gradient Boosting) and neural network models demonstrated stable performance, with high accuracy and F1 scores, underscoring their generalizability (Figure SF2C). Decision Curve Analysis (DCA) indicated that these models provided substantial clinical net benefits across a broad range of risk thresholds, surpassing both “full intervention” and “no intervention” strategies, thus affirming their clinical utility (Fig. 4D-F). Confusion matrix analysis revealed accuracies of 96.1%, 93.8%, and 90.9% in the training, validation, and external validation sets, respectively, with minimal misclassification rates, confirming robust classification stability (Fig. 4G-H, J). Feature importance and performance analysis This study employed SHAP (Shapley Additive Explanations) to elucidate the significance of features in machine learning models and their impact on predictive outcomes. Results revealed that CTNNB1 (mean absolute SHAP value: 0.079), PGLYRP1 (0.069), and THSD4 (0.059) were the most influential features, underscoring their pivotal roles in model predictions (Fig. 5A). A beeswarm plot delineated the relationship between feature values and SHAP values, demonstrating that elevated CTNNB1 and EPB41L2 expression levels were associated with positive predictive contributions, while lower levels exhibited neutral or negative effects (Fig. 5B). Scatter plots for individual features confirmed significant correlations between high expression of CTNNB1, PGLYRP1, and EPB41L2 and positive SHAP values, with SHAP value ranges of −0.3 to 0.1, −0.2 to 0.1, and − 0.2 to 0.2, respectively (Fig. 5C). SHAP-Based Analysis of Key Gene Feature Importance and Impact. (A) Bar plot illustrating mean absolute SHAP values for key genes, highlighting the magnitude of each feature’s contribution to model predictions, with larger values denoting greater impact. (B) Beeswarm plot depicting SHAP value distribution, where each dot represents a sample’s feature value contribution to model output; color reflects feature value magnitude (red for high, blue for low), and the x-axis indicates the direction and magnitude of SHAP values’ impact on predictions. (C) Scatter plots of SHAP values for six key genes, with colors denoting the expression levels of a related gene, elucidating gene interactions and their combined effects on model output. Disease risk prediction model performance and feature contribution analysis The multi-gene risk scoring model developed in this study exhibited robust predictive accuracy and clinical utility for osteoarthritis (OA) risk assessment. Incorporating key genes—EPB41L2, ISLR, HLA-DRB1, CTNNB1, and THSD4—the scoring system assigned individual risk scores ranging from 0 to 450, corresponding to disease risk probabilities from 0.0001 to nearly 100% (Fig. 6A). Calibration curve analysis demonstrated strong concordance between predicted and observed risks, with the bias-corrected curve closely aligning with the ideal diagonal, indicating superior calibration performance (Fig. 6B). Decision Curve Analysis (DCA) revealed that, across a threshold probability range of 1% to 50%, the selected gene model yielded significantly greater net benefits compared to a full-gene model or no-model strategy, with a cost-benefit ratio spanning 1:100 to 100:1, underscoring its robust clinical applicability (Fig. 6C). In risk stratification, increasing threshold probabilities reduced the number of high-risk individuals from approximately 350, with a corresponding decrease in true high-risk individuals experiencing events, affirming the model’s efficacy in identifying high-risk patients (Fig. 6D). Construction and Performance Evaluation of the Key Gene Risk Model. (A) Nomogram based on 10 key genes for calculating individual disease risk scores and corresponding risk probabilities. (B) Calibration curve illustrating the alignment between model-predicted probabilities and actual event rates, with the ideal curve (dashed line) closely matching the bias-corrected curve (solid line), indicating accurate predictions. (C) Decision Curve Analysis (DCA) depicting the net benefit of the risk model across various thresholds, with the red line highlighting the superior net benefit of the key gene model compared to “full intervention” or “no intervention” strategies. (D) Plot illustrating changes in the number of high-risk patients and true high-risk patients with events as risk thresholds are adjusted, facilitating risk stratification and clinical decision-making. Immune infiltration This study meticulously elucidated disparities in the immune microenvironment between osteoarthritis (OA) patients and healthy controls (HC) through comprehensive immune cell composition analysis. A stacked bar plot of immune cell subtypes (Fig. 7A) revealed the predominance of macrophages, particularly M0, M1, and M2 subtypes, across all samples, underscoring their pivotal role in local immune regulation within joints. Boxplot comparisons (Fig. 7B) demonstrated a marked increase in pro-inflammatory M1 macrophages in OA patients (p < 0.001), while anti-inflammatory M2 macrophages exhibited no significant differences. Additionally, memory B cells (p < 0.01), activated dendritic cells (p < 0.05), and helper T cells (p < 0.05) displayed significant alterations in the OA cohort, reflecting a multifaceted state of immune activation and regulation in OA tissues. Association Analysis of Key Gene Expression and Immune Cell Infiltration. (A) Stacked bar plot illustrating the estimated proportion of immune cell types across samples, with distinct colors denoting different immune cell subtypes. (B) Boxplot comparing immune cell composition between healthy control (HC) and osteoarthritis (OA) cohorts, with significance markers: ns (not significant), *p < 0.05, **p < 0.01, ***p < 0.001. (C) Spearman correlation heatmap of key gene expression and immune cell proportions, with red and blue gradients indicating positive and negative correlations, respectively; correlation coefficients are displayed, with significant correlations highlighted in bold. Correlation analysis (Fig. 7C) revealed robust associations between key gene expression and immune cell infiltration. Notably, HLA-DRB1 expression exhibited a strong positive correlation with M1 macrophages (r = 0.67, p < 0.001), suggesting its role in promoting immune cell recruitment and activation in inflammatory responses. Conversely, CTNNB1 displayed a negative correlation with M2 macrophages (r = −0.45, p < 0.01), indicating a potential suppressive role in anti-inflammatory immune regulation. Other genes, including PTEN, EPB41L2, and ISLR, showed varying degrees of correlation with multiple immune cell subtypes, highlighting intricate interactions within the immune network. Identification of ERG-related subgroups in OA Consensus clustering analysis (Fig. 8A) categorized samples into subgroups, revealing that k = 2 achieved optimal internal consistency with distinct classification boundaries (Fig. 8B), confirming two subgroups as the most robust division. Differential gene expression analysis (Fig. 8C, D) indicated that key genes, including EPB41L2, ISLR, HLA-DRB1, and THSD4, exhibited statistically significant differences between treatment and control cohorts, underscoring the regulatory impact of disease state or interventions on their expression. These genes likely modulate disease-related pathways and influence pathological processes. Immune cell infiltration analysis (Fig. 8E) revealed that, while most immune cell subtypes showed no significant differences between subgroups, the pronounced variation in M1 macrophages highlighted their critical role in host immune responses and disease progression. These findings suggest that disease status not only alters gene expression profiles but also modulates the dynamic distribution of specific immune cell subtypes, reflecting complex regulatory mechanisms in the immune microenvironment. Sample Subtype Clustering Analysis and Immune Cell Composition Differences Based on Key Genes. (A) Consensus clustering matrix heatmap depicting sample clustering at k = 2, with deep blue indicating high consistency in clustering relationships. (B) Cumulative Distribution Function (CDF) curves for varying cluster numbers (k), demonstrating clustering stability, with k = 2 showing optimal gain. (C) Boxplot of key gene expression in control versus treatment cohorts, illustrating expression differences, with significance markers: *p < 0.05, **p < 0.01, ***p < 0.001. (D) Heatmap of key gene expression, with colors reflecting expression levels, delineating heterogeneity across sample subtypes. (E) Boxplot comparing immune cell proportions between the two subtype groups, with significant differences in specific cell types marked (ns for not significant). Clinical sample validation of core ERGs expression To validate the expression of the 10 core exosome-related genes (ERGs), real-time quantitative PCR (RT-qPCR) was performed on clinical samples from 10 OA patients and 12 healthy controls. Most genes exhibited significant expression changes in OA patients. Specifically, EPB41L2, ISLR, HLA-related genes, and THSD4 were upregulated, suggesting their potential role in promoting or responding to OA pathogenesis. Conversely, PGLYRP1, PTEN, and CTNNB1 were downregulated, possibly indicating diminished protective functions (Fig. 9A). These findings highlight the significant regulation of multiple genes in OA tissues, positioning them as promising diagnostic biomarkers or therapeutic targets. Gene Expression Levels of Key Genes in Healthy Control (HC, Yellow) and Osteoarthritis (OA, Green) Cohorts. Bar plots depict mean values, with error bars representing standard error, and dot plots illustrate individual sample data. All experiments were conducted in at least triplicate, with significance markers: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Discussion Aging, obesity, mechanical stress, and congenital joint anomalies are well-established predisposing factors for osteoarthritis (OA)24,25. A pivotal pathological signaling cascade in OA pertains to cell adhesion, characterized by dysregulated interactions between cells and the extracellular matrix, which precipitate cartilage degradation and inflammatory responses25. In recent years, exosomes, as quintessential mediators of intercellular communication, have attracted burgeoning interest for their instrumental role in OA pathogenesis27. These extracellular vesicles ferry bioactive molecules, such as microRNAs and proteins, modulating chondrocyte metabolism, inflammatory cascades, and cell adhesion pathways, notably integrin signaling28. Nevertheless, no studies have hitherto employed bioinformatics methodologies to elucidate the nexus between OA and exosome-related mechanisms. To advance our comprehension of OA’s etiology from a genomic vantage point, investigations integrating bioinformatics with exosome-related genes are imperative. This study trailblazes the integration of gene expression datasets, employing meticulous batch effect correction to delineate the distinctive expression profiles of exosome-related genes (ERGs) in OA synovial tissues. By synergizing weighted gene co-expression network analysis (WGCNA) with state-of-the-art machine learning algorithms, we successfully identified and validated a cadre of 10 core ERGs. This methodical approach not only marks a seminal advancement in identifying potential OA biomarkers from expansive datasets but also, more crucially, enriches our understanding of exosomes’ role in the intricate pathological milieu of OA, particularly in immune modulation and metabolic dysregulation. These findings furnish actionable molecular targets for precision diagnostics and therapeutic interventions in OA. The 10 core genes delineated in this study (EPB41L2, ISLR, HLA-DRB1, HLA-DRA, PGLYRP1, PTEN, TKT, CTNNB1, THSD4, ATP9A) are substantiated by robust biological contexts and intricately linked to established OA pathological pathways. A salient innovation of this study lies in associating these genes with exosome-mediated intercellular communication, positing a novel paradigm that they may orchestrate OA progression via exosomal pathways. For instance, CTNNB1, a cardinal effector in the Wnt/β-catenin signaling pathway, is frequently reported to exhibit aberrant activation in OA, correlating with chondrocyte proliferation, differentiation, and inflammatory responses29,30. This study further conjectures that dysregulated CTNNB1 expression may modulate exosomal cargo (e.g., specific miRNAs or proteins), thereby altering their composition and functionality. These altered exosomes, laden with aberrant signals, may be internalized by target cells such as synovial fibroblasts, chondrocytes, or immune cells, potentially exacerbating the pro-inflammatory milieu of the synovial microenvironment and accelerating cartilage degeneration. Similarly, HLA-DRB1 and HLA-DRA, as integral components of the major histocompatibility complex (MHC) class II, underscore pronounced antigen presentation and T-cell-mediated immune responses within the OA synovium31. Our findings resonate with contemporary studies accentuating the centrality of immune-mediated inflammation in OA pathogenesis32. Moreover, the presence of these HLA molecules on exosomal surfaces may confer immunogenicity, potentially amplifying the activation of antigen-presenting cells and T-cell responses, thereby exacerbating localized joint inflammation. The upregulation of PGLYRP1, a pivotal regulator of innate immunity, and its robust correlation with pro-inflammatory M1 macrophage infiltration further substantiate the role of exosomes as conduits for immune activation signals in OA pathology33. Additionally, the association between diminished ATP9A expression and enriched fatty acid metabolism pathways unveils a novel paradigm for exploring the interplay between metabolic dysregulation and inflammation34,35. This dovetails with the emergent concept of meta-inflammation in OA research, positing that exosomes may modulate joint homeostasis by transporting metabolism-associated factors. The exemplary performance of machine learning models robustly corroborates the diagnostic potential of these exosome-related genes (ERGs) as a cohesive ensemble. Random Forest and Gradient Boosting Tree models exhibited consistently superior classification efficacy across internal and external independent validation cohorts. SHAP (SHapley Additive exPlanations) analysis further elucidated the pivotal contributions of core genes, such as CTNNB1 and PGLYRP1, to predictive accuracy. These results transcend the constraints of conventional single-biomarker approaches, furnishing a potent tool for early precision diagnosis and risk stratification in OA. The risk prediction model and nomogram, constructed based on these genes, empower clinicians to stratify patient risk and tailor personalized therapeutic strategies, such as identifying subgroups likely to respond optimally to targeted therapies (e.g., interventions addressing inflammatory or metabolic pathways), thus holding substantial promise for clinical translation. The immune infiltration analysis in this study further revealed a marked increase in M1 macrophages and activated dendritic cells within the OA synovium, aligning with multiple reports underscoring the inflammatory hallmarks of OA36–38. However, the distinctiveness of this study resides in the observed robust correlation between immune cell infiltration and the expression of specific exosome-related genes, compellingly suggesting that exosomes are not mere byproducts of pathological processes but serve as critical mediators of information transfer, orchestrating the activation, recruitment, and functional modulation of immune cells in OA. This finding provides cogent molecular evidence reinforcing the perspective that OA is not exclusively a degenerative condition but also exhibits pronounced immune-inflammatory characteristics. Most innovatively, this study pioneers the application of an unsupervised clustering strategy to delineate molecular subtypes of osteoarthritis (OA) based on the expression profiles of exosome-related genes (ERGs). This seminal breakthrough surmounts the constraints of conventional OA research, which frequently relies on single-tissue analyses (e.g., cartilage or synovium) or limited biomarkers for subtyping. Exosomes, as nanoscale vectors of intercellular communication, convey ERG expression profiles that not only mirror the pathological state of secreting cells but also dynamically transmit intricate biological information across tissues and, potentially, systemically. Consequently, ERG-based subtyping likely captures the inherent heterogeneity of OA joint microenvironments and systemic alterations with greater fidelity. This heterogeneity underpins the variability in OA clinical manifestations, disease progression, and therapeutic responses. While prior studies have endeavored to classify OA based on broad gene expression profiles or clinical phenotypic characteristics, such approaches often yield static molecular snapshots or superficial categorizations39–41. In contrast, this study embraces a novel paradigm by leveraging exosomes, which function as dynamic conduits of information transfer, to elucidate the profound molecular heterogeneity among OA patients. We posit that distinct ERG-based molecular subtypes may correspond to unique primary pathological drivers of OA. For instance, one subtype may be predominantly characterized by exosome-mediated dysregulation of immune-inflammatory pathways (e.g., elevated expression of HLA-related genes), while another may hinge on exosome-regulated metabolic remodeling imbalances (e.g., involving ATP9A and CTNNB1). This groundbreaking discovery furnishes a robust theoretical foundation for stratified medicine, paving the way for precise patient subgroup management and tailored therapeutic strategies. Nevertheless, this study is not without limitations. First, despite sourcing data from multiple public datasets and implementing meticulous batch effect correction, the absence of large-scale, multi-center, prospective clinical validation may constrain the generalizability and clinical applicability of the model. Second, while bioinformatics analyses inferred the roles of ERGs and their associations with exosomes, the precise molecular mechanisms—particularly the loading, release, targeting, and impact of their encoded products (e.g., mRNAs, proteins, miRNAs) within exosomes on recipient cells—necessitate further validation through in vitro cellular assays and in vivo animal models. Additionally, the heterogeneity of exosomes, their variable content and composition across different joint tissues and bodily fluids (e.g., synovial fluid, blood), and challenges associated with standardized isolation and purification methodologies remain partially unresolved in this study. Future investigations should integrate single-cell transcriptomics, exosome proteomics/metabolomics, and high-resolution imaging techniques to elucidate the dynamic roles of exosomes in OA pathology with enhanced granularity. Conclusion In summation, this study trailblazes the integration of exosome-related gene (ERG) expression profiles with cutting-edge machine learning methodologies, proposing innovative strategies for osteoarthritis (OA) diagnosis and molecular subtyping while elucidating the pivotal roles of exosomes and their core genes in immune modulation and metabolic homeostasis. These findings establish a robust foundation for early OA diagnosis, the development of novel biomarkers, and targeted therapeutic interventions, markedly advancing our comprehension of OA pathogenesis. Looking forward, the synergy of rigorous functional assays and multi-center clinical validation will amplify the potential of the identified ERGs and molecular subtypes to drive transformative applications of exosome-based diagnostic and therapeutic strategies in OA clinical management, ultimately delivering tailored and efficacious treatment options for patients. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Author contributions CFY: Conceptualization, Methodology, Supervision, Writing - Review & EditingFD, BZD, WJW, ZS: Visualization, Writing - Original DraftLF, WMJ, CFY, XU: Formal analysis, ValidationJL: Resources, Software, Data curation, ValidationAll authors have read and approved the manuscript. Data availability The direct links required to find each data set in the database are as follows: the GEO gene expression and clinical pathology data set: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE12021;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE51588;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE55235;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE55457;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE169077;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE82107;The data set downloaded by this direct link is the original data set. Declarations Competing interests The authors declare no competing interests. Ethics approval This study has been approved by the Ethics Committee of Siyang Hospital with the approval number HY2024005. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Data Availability Statement The direct links required to find each data set in the database are as follows: the GEO gene expression and clinical pathology data set: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE12021;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE51588;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE55235;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE55457;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE169077;https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE82107;The data set downloaded by this direct link is the original data set.