Reactive oxygen species-associated molecular landscape uncovers diagnostic and immune features in osteoarthritis Authors: Jiawei Zhou, Ping Qiu, Tianyu Liang, Binger Xue Published: 2025 ABSTRACT BackgroundOsteoarthritis (OA) is a prevalent degenerative joint disease characterized by joint inflammation and degradation. Reactive oxygen species (ROS) are implicated in the pathogenesis of OA, yet the molecular pathways linking ROS-related genes to the immune microenvironment remain poorly understood.MethodsWe integrated publicly available OA gene expression datasets and performed batch correction. ROS-related pathways were curated from the MSigDB, and pathway activity was assessed using ssGSEA. We identified differentially expressed genes (DEGs) in both training and validation cohorts, followed by GO and KEGG enrichment analyses. An elastic net regression model was constructed to calculate a risk score based on critical ROS-related DEGs. Model performance was evaluated using ROC, PR curves, calibration, and decision curve analysis. We also explored microRNA and transcription factor networks alongside genetic associations via the GWAS Catalog. Correlations among risk scores, inflammatory cytokines, and immune/stromal cell infiltration were systematically analyzed. An OA model was induced in rat fibroblast-like synoviocyte (FLS) cells via LPS treatment, and a corresponding rat OA model was established. The influence of ROS on FLS cell migration was validated through wound healing assays and Transwell assays. The expression of inflammatory markers and potential ROS-related biomarkers was evaluated using Western blotting and qRT-PCR. Additionally, the fluorescent probe DCFH-DA was used to detect the ROS levels.ResultsTranscriptomic analysis indicated significant dysregulation of ROS-related pathways in OA. Upregulated DEGs were associated with extracellular matrix and immune response processes, while downregulated genes were primarily linked to metabolic functions. The risk score comprising seven genes demonstrated high diagnostic accuracy. Network analyses revealed intricate regulatory mechanisms and genetic associations. The risk score positively correlated with inflammatory cytokines and immune cell infiltration, particularly fibroblasts. In vitro and in vivo experiments confirmed that LPS treatment significantly enhances FLS cell migration, upregulates OA-related inflammatory markers and ROS levels.ConclusionsThis study establishes a significant ROS-related gene signature with diagnostic and immunological relevance in OA, offering potential pathways for the development of novel biomarkers and targeted therapies focusing on ROS and immune responses in OA.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13018-025-06571-9. Introduction Osteoarthritis (OA) affects an estimated 528 million individuals worldwide and is a leading cause of chronic pain and disability, especially among older adults [1]. OA is primarily characterized by cartilage degradation, osteophyte formation, subchondral bone remodeling, and synovial inflammation, imposing a substantial socioeconomic burden and significantly reducing patients’ quality of life [2, 3]. Despite its high prevalence, the precise pathogenesis of OA remains incompletely understood, and effective biomarkers for early diagnosis and targeted therapeutic strategies are still lacking. Accumulating evidence indicates that oxidative stress, primarily driven by excessive reactive oxygen species (ROS), plays a central role in OA pathogenesis. Elevated ROS disrupts the redox balance in articular tissues, inducing chondrocyte apoptosis, extracellular matrix degradation, and inflammatory responses [4, 5]. Clinical studies have reported increased levels of ROS and oxidative damage markers in OA cartilage and synovial fluid. Moreover, key ROS-related signaling pathways have been implicated in cartilage catabolism and joint inflammation [6, 7]. Nevertheless, a comprehensive characterization of ROS-associated molecular alterations within the OA immune microenvironment remains lacking. Accumulating transcriptomic and spatial omics studies have unveiled significant heterogeneity and immune microenvironment remodeling in OA. For instance, single-cell and bulk RNA-sequencing have revealed distinct immune and stromal compartments in OA synovium, with macrophage and T-cell infiltration as key contributors to inflammation and tissue degeneration [8, 9]. Multi-algorithm deconvolution analyses of subchondral bone and cartilage have similarly revealed dysregulated infiltration patterns, including increased macrophages and endothelial cells, in OA tissues [10]. Moreover, integrated transcriptomic profiling has demonstrated that oxidative stress–related pathways closely interact with immune and stromal cell networks, yet few studies have systematically linked ROS-related molecular alterations with shifts in the OA microenvironment [11]. Therefore, a comprehensive investigation integrating ROS pathway activity, immune cell infiltration, and molecular signatures remains an unmet need in understanding OA pathology. In this study (Figure S1), we conducted an integrative analysis of multiple publicly available OA transcriptomic datasets to systematically characterize ROS-related pathway activity and its association with immune and stromal cell infiltration. We identified ROS-associated differentially expressed genes and developed a diagnostic risk score using elastic net regression. Additionally, we constructed upstream regulatory networks, investigated genetic associations, and assessed relationships between the risk score, inflammatory mediators, and immune cell populations. These findings provide new insights into OA molecular mechanisms and highlight ROS-related molecular signatures as promising biomarkers and potential therapeutic targets for OA. Materials and methods Data acquisition and preprocessing Publicly available gene expression datasets relevant to OA were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/gds). GSE51588 and GSE89408 were designated as the training cohort (Table S1), while GSE114007, GSE12021, GSE1919, GSE55235, and GSE55457 comprised the validation cohort (Table S2). Raw expression data were processed and normalized using R (version 4.2.2). To mitigate batch effects across different datasets, the ComBat algorithm within the ‘sva’ package was employed, ensuring comparability of gene expression profiles [12]. Selection of ROS-related pathways Gene sets associated with ROS were curated from the Molecular Signatures Database (MSigDB) [13]. Five representative ROS-related pathways, closely linked to oxidative stress and OA pathogenesis, were selected for further analysis (HALLMARK_REACTIVE_OXYGEN_SPECIES_PATHWAY, GOBP_CELLULAR_RESPONSE_TO_REACTIVE_OXYGEN_SPECIES, GOBP_REACTIVE_OXYGEN_SPECIES_BIOSYNTHETIC_PROCESS, GOBP_REACTIVE_OXYGEN_SPECIES_METABOLIC_PROCESS, and GOBP_RESPONSE_TO_REACTIVE_OXYGEN_SPECIES). Single-sample gene set enrichment analysis (ssGSEA) To evaluate the enrichment of ROS-related pathways in individual samples, ssGSEA was performed in both the training and validation cohorts. Heatmaps were generated to visualize pathway activity patterns in OA and normal samples. Differential pathway activity analysis ROS pathway activities were compared between OA and normal groups using the Wilcoxon rank-sum test. Differential expression analysis was performed to identify genes with significant changes between OA and normal samples using the ‘limma’ package, with thresholds set for adjusted p-values (FDR < 0.05) and fold change. Functional enrichment analysis of OA-related differentially expressed genes Gene Ontology (GO) enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis were conducted on the differentially expressed genes (DEGs) to identify biological roles and significantly enriched pathways. The significance threshold was set at an adjusted p-value < 0.05. The DEG lists of the training cohort and validation cohort were presented in Table S3 and Table S4. Correlation analysis and construction of an elastic net regression model Eight genes intersecting between differentially expressed genes and ROS-related genes were selected for analysis. The correlation of gene expression was assessed using Pearson correlation analysis in both the training and validation cohorts. An elastic net regression model was constructed to predict OA, using standardized gene expression data from the training cohort. The elastic net regression model was fitted using the ‘glmnet’ package in R, with model parameters optimized by ten-fold cross-validation and the nfolds set to 10, family set to binomial, alpha parameter set to 0.5. Genes with non-zero coefficients at the optimal lambda value were retained as model variables. Model performance was evaluated by calculating prediction probabilities and visualizing the results with ggplot2. Model performance evaluation Model performance was assessed using ROC curve analysis to calculate the area under the curve (AUC) for both the risk score and individual genes. Precision-recall (PR) curves were generated to evaluate the discriminative ability of the model. Calibration was assessed with calibration curves comparing predicted and observed outcomes, and decision curve analysis (DCA) was performed to quantify the clinical net benefit of the risk score. Regulatory network construction and GWAS signal analysis The miRNA-mRNA and transcription factor (TF)-gene regulatory networks were constructed using the miRNet 2.0 platform (https://www.mirnet.ca/). For genetic association analysis, GWAS signals for candidate genes were retrieved from the GWAS Catalog and visualized using Manhattan plots (https://www.ebi.ac.uk/gwas/). Evaluation of inflammatory markers and immune microenvironment The association between risk score and the expression levels of inflammatory cytokines (TNF, IL1B, IL8) was evaluated using Spearman correlation analysis. Disease inference analysis was conducted using the Comparative Toxicogenomics Database (CTD, https://ctdbase.org/) to determine the inference scores reflecting the association between each key gene and osteoarthritis. Immune and stromal cell infiltration was estimated using MCPcounter [14], quanTIseq [15], and xCell [16] algorithms. These computational tools were applied to transcriptomic data to quantify the relative abundance of diverse immune and stromal cell populations in OA and normal samples. Spearman correlation analysis was further performed to examine the relationship between the expression of key OA-related genes and fibroblast infiltration scores in both OA and normal groups. Cell culture Rat primary FLSs (CP-R329, Procell, China) were cultured in F12 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin. The cultures were maintained in a humidified incubator at 37 °C with 5% CO₂ to ensure optimal growth conditions. Wound healing assay Cells were seeded at a density of 2 × 10^5 cells/well in a 6-well plate. When the cell density reached over 90%, mitomycin (5 µg/mL) was added and incubated for 4 h before wounding. A uniform scratch was then made in the cell monolayer using pipette tips. After carefully washing the cells three times with PBS, 2 mL of growth medium was added to each well. Images of the scratched areas were captured using an inverted microscope. The migration ratio was calculated by subtracting the width of the wound at 0 h from the width at 24 h, and then dividing this difference by the original wound width at 0 h. Transwell assay The Transwell chamber (Corning) was utilized to evaluate the invasive capacity of cells. Target cells (typically 1 × 10⁵ cells per well) were seeded into the upper chamber of the Transwell, using serum-free medium to minimize migration of 500 µL. The lower chamber was filled with complete culture medium supplemented with 10% serum, serving as a chemoattractant to promote cell migration of 1 mL. The cells were incubated at 37 °C in a 5% CO₂ incubator for a duration of 24 h, allowing them to invade through the membrane into the lower chamber. After the incubation period, non-invading cells in the upper chamber were carefully removed using a cotton swab. The remaining cells in the lower chamber were then washed with PBS, fixed with methanol, and stained with a 0.1% crystal violet solution. Finally, the stained cells in the lower chamber were observed and counted under a microscope to assess the cell invasion ability. Western blot Total protein was extracted from both cells and tissue samples using RIPA lysis buffer supplemented with 1% protease inhibitor cocktail. For tissue samples, an initial mechanical homogenization step was performed, followed by incubation with the lysis buffer, while cultured cells were directly lysed using the same buffer. The protein concentration was quantified using the BCA assay, with samples incubated at 37 °C for 30 min before measurement. A total of 30 µg of extracted protein samples was denatured at 95 °C for 10 min and then subjected to electrophoretic separation on a 10–12% SDS-PAGE gel. After separation, the proteins were transferred to a polyvinylidene fluoride (PVDF) membrane using the wet transfer method at 100 V for 90 min. The membrane was then blocked for 2 h at room temperature using 5% non-fat dry milk in TBST (Tris-buffered saline with 0.1% Tween-20) to reduce non-specific binding, followed by three 10-minute washes with TBST. The membrane was incubated overnight at 4 °C with primary antibodies (diluted in TBST with 5% BSA) including IL-6 (1:1000, A0286, Abclonal), TNF-α (1:1000, 17590-1-AP, Proteintech), COX-2 (1:2000, 12375-1-AP, Proteintech), MMP-1 (1:1000, A22080, Abclonal), MMP-3 (1:1000, A11418, Abclonal), MMP-13 (1:1000, A11755, Abclonal), and β-Actin (1:10000, AC038, Abclonal) as an internal control. After primary antibody incubation, the membrane was washed three times for 10 min with TBST and incubated with appropriate secondary antibodies (1:2000 dilution in TBST with 5% BSA) at room temperature for 2 h. The secondary antibodies used were goat anti-rabbit IgG (1:2000, A120-101P, Abclonal). Following secondary antibody incubation, the membrane was washed again three times with TBST, and the blots were visualized using an ECL chemiluminescent detection system (Bio-Rad, ChemiDoc). Densitometric analysis was performed using ImageJ software to quantify the expression levels of the target proteins, with β-actin used as a loading control for normalization. RNA extraction and quantitative reverse transcription polymerase chain reaction (qRT-PCR) Total RNA was extracted from cells and tissues using TRIzol reagent. Following the extraction, 1 µg of RNA was reverse transcribed to remove genomic DNA, under the manufacturer’s instructions provided with the reverse transcription kit. This process converted the RNA into complementary DNA (cDNA). Subsequently, cDNA amplification was performed using Fast Start Universal SYBR Green Master in a real-time PCR setup. The relative quantification of the target gene was conducted using the 2-ΔΔCt method, with GAPDH serving as an internal control to normalize the expression levels. The primer sequences utilized in the qRT-PCR assays are provided in Table S5. ROS detection Detection of ROS levels was performed using the fluorescent probe DCFH-DA (Abcam, UK). Prepared cell slides or tissue sections were loaded with 20 µM DCFH-DA in serum-free medium at 37 °C in the dark for 45 min. After washing with PBS to remove excess probe, images was observed and captured using a fluorescence microscope (IX35, Olympus, Japan). Relative ROS levels across groups were quantitatively compared based on image fluorescence intensity. Construction of OA models The animal experiments were approved by the Institutional Animal Care and Use Committee of the Hubei Provincial Center for Disease Control and Prevention. Eighteen male Sprague-Dawley rats (aged 6 weeks, weighing 200–250 g) were used in this study. Among them, 9 rats received destabilization of the medial meniscus (DMM) surgery to establish the OA model, and the others served in the Sham group (n = 9) without surgery. The rats were anesthetized with isoflurane, and surgeries were performed following standard sterile procedures. Briefly, the medial meniscotibial ligament in the DMM model was transected. After 8 weeks, all animals were euthanized, and samples were collected for analysis. Euthanasia was carried out using pentobarbital sodium, following standard ethical guidelines. Hematoxylin and Eosin (HE) staining HE staining was used to evaluate the histological features of joint tissues in OA rat models. Both normal and OA tissues were fixed in 10% formalin, embedded in paraffin blocks, and sectioned into 5-micron-thick slices, which were then placed on microscope slides. Following deparaffinization and rehydration, the sections were stained with hematoxylin to visualize cell nuclei and subsequently counterstained with eosin to highlight the cytoplasm. The stained sections were examined under a microscope, and images were captured for further analysis of immune cell infiltration and cartilage degeneration in the joint tissue. Statistical analysis All statistical analyses of the bioinformatics data were performed using R software (version 4.1.2). Continuous data were evaluated using the Mann-Whitney U test. A significance threshold of p < 0.05 was established to denote statistical significance. For the analysis of cell experimental data, GraphPad Prism 9.0 software (GraphPad, San Diego, CA) was utilized. All experiments were conducted in triplicate, and results are presented as mean ± standard deviation (SD). Statistical comparisons between two groups were performed using a t-test, with a p-value of less than 0.05 considered statistically significant. To ensure the reliability and reproducibility of the findings, all experiments were repeated a minimum of three times. Results Distinct ROS-Related pathway activities in OA revealed by integrative transcriptomic analysis Principal component analysis (PCA) was first performed to evaluate batch effects and sample distribution across datasets. In the training cohort, raw PCA revealed a pronounced separation between the two datasets, GSE51588 and GSE89408, while batch correction using the ComBat algorithm resulted in a substantial reduction of batch effects and improved clustering by biological condition (Fig. 1A). A similar trend was observed in the validation cohort, where initial PCA demonstrated clear batch effects among the five datasets, which were effectively minimized following batch correction (Fig. 1A). To explore the biological relevance of ROS in OA, five representative ROS-related pathways were curated from the MSigDB database and analyzed using ssGSEA. Heatmap visualization of pathway enrichment scores showed distinct clustering of OA and normal samples, with OA samples exhibiting notably different expression patterns of ROS-related pathways in both the training and validation cohorts (Fig. 1B). Quantitative comparison of ssGSEA enrichment scores demonstrated that four ROS-associated pathways were significantly upregulated in OA samples compared to normal controls (Fig. 1C). Differential gene expression analysis further identified a large number of genes with significant expression changes between OA and normal samples in both the training and validation cohorts, as visualized in volcano plots (Fig. 1D), highlighting a robust transcriptomic alteration associated with OA pathogenesis. Collectively, these results indicate that ROS-related pathway activity is markedly dysregulated in OA, underscoring the potential contribution of oxidative stress to OA development. Fig. 1Integrative analysis of ROS-related pathways in OA. A PCA of training and validation cohorts before and after batch effect correction with ComBat. B Heatmaps of ssGSEA enrichment scores for five ROS-related pathways in normal and OA samples from both cohorts. C Violin plots showing ssGSEA scores of four significantly altered ROS pathways between normal and OA groups. D Volcano plots of differential gene expression between OA and normal samples in the training and validation cohorts Integrative analysis of ROS-related pathways in OA. A PCA of training and validation cohorts before and after batch effect correction with ComBat. B Heatmaps of ssGSEA enrichment scores for five ROS-related pathways in normal and OA samples from both cohorts. C Violin plots showing ssGSEA scores of four significantly altered ROS pathways between normal and OA groups. D Volcano plots of differential gene expression between OA and normal samples in the training and validation cohorts Functional characterization of differentially expressed genes in OA To identify OA-associated molecular signatures, we first examined the overlap of differentially expressed genes (DEGs) between the training and validation cohorts. A total of 226 upregulated and 32 downregulated genes were found to be shared between both cohorts (Fig. 2A). Functional enrichment analysis was subsequently performed to explore the biological roles of these intersecting DEGs. GO enrichment analysis of upregulated genes revealed significant enrichment in terms related to extracellular matrix organization, leukocyte chemotaxis, glycosaminoglycan binding, and receptor ligand activity, spanning across BP, CC, and MF categories (Fig. 2B, left). In contrast, the downregulated genes were primarily associated with processes such as fatty acid metabolic process, regulation of vascular permeability, and amide binding (Fig. 2B, right). KEGG pathway analysis further demonstrated that upregulated genes were mainly involved in pathways such as rheumatoid arthritis, viral myocarditis, and hematopoietic cell lineage, as well as pathways related to cell adhesion molecules and asthma (Fig. 2C, left). Conversely, downregulated genes were significantly enriched in pathways related to fatty acid metabolism, AMPK signaling, insulin signaling, adipocytokine signaling, and fatty acid biosynthesis (Fig. 2C, right). Collectively, these results highlight distinct biological functions and pathways associated with OA-related gene expression changes. Fig. 2Functional enrichment analysis of intersecting DEGs in OA. A Venn diagrams showing the overlap of upregulated and downregulated DEGs between the training and validation cohorts. B GO enrichment analysis of intersecting upregulated (left) and downregulated (right) DEGs. The top GO terms in BP, CC, and MF categories are shown. C KEGG pathway enrichment analysis of intersecting upregulated (left) and downregulated (right) DEGs, with the top significantly enriched pathways displayed Functional enrichment analysis of intersecting DEGs in OA. A Venn diagrams showing the overlap of upregulated and downregulated DEGs between the training and validation cohorts. B GO enrichment analysis of intersecting upregulated (left) and downregulated (right) DEGs. The top GO terms in BP, CC, and MF categories are shown. C KEGG pathway enrichment analysis of intersecting upregulated (left) and downregulated (right) DEGs, with the top significantly enriched pathways displayed Elastic net regression identifies a robust OA predictive gene signature Correlation analysis of the eight intersecting genes between differentially expressed genes and ROS-related genes demonstrated strong and consistent gene-gene relationships in both the training and validation cohorts (Fig. 3A). These genes, including FER, MMP9, MPV17, and MMP3, have been implicated in metabolic processes and pathways relevant to OA. For example, FER and MMP9 are involved in extracellular matrix remodeling, while MPV17 and MMP3 play roles in oxidative stress and tissue degeneration, key features of OA progression. Using these candidate genes, an elastic net regression model was constructed in the training cohort. Cross-validation identified the optimal lambda value, resulting in a model that incorporated seven genes with non-zero coefficients (Fig. 3B). The coefficients for each gene in normal and OA samples were visualized, revealing distinct contributions of individual genes to OA risk prediction (Fig. 3C). Notably, genes such as FER and MMP9 showed stronger associations with OA in both normal and OA samples, suggesting their significant role in OA pathogenesis. Application of the model in both the training and validation cohorts showed that the predicted probabilities could effectively distinguish OA from normal samples, supporting the robustness and generalizability of the gene signature (Fig. 3D). The model’s ability to predict OA in both cohorts highlights the potential of these identified genes as biomarkers for OA diagnosis and prognosis, underscoring their involvement in metabolic pathways and disease-related processes associated with OA. Fig. 3Construction of an elastic net regression model for OA prediction. A Correlation matrix showing pairwise expression relationships among the eight intersecting ROS-related genes in the training and validation cohorts. B Elastic net regression model selection using ten-fold cross-validation. The top panel shows binomial deviance versus log(lambda); the bottom panel displays coefficient profiles for candidate genes. C Bar plots of model coefficients for each gene in normal and OA samples. D Predicted probabilities of OA and normal samples based on the elastic net model in the training and validation cohorts Construction of an elastic net regression model for OA prediction. A Correlation matrix showing pairwise expression relationships among the eight intersecting ROS-related genes in the training and validation cohorts. B Elastic net regression model selection using ten-fold cross-validation. The top panel shows binomial deviance versus log(lambda); the bottom panel displays coefficient profiles for candidate genes. C Bar plots of model coefficients for each gene in normal and OA samples. D Predicted probabilities of OA and normal samples based on the elastic net model in the training and validation cohorts Comprehensive evaluation of the predictive model for OA diagnosis In the training cohort, the risk score derived from the elastic net regression model demonstrated superior diagnostic performance compared to individual genes, with an AUC of 0.854 (95% CI: 0.774–0.933), as shown by the ROC curves. Calibration curve assessment indicated that the predicted probabilities were well-aligned with observed outcomes, suggesting good model calibration (Fig. 4A). Consistent results were observed in the validation cohort, where the risk score achieved an AUC of 0.823 (95% CI: 0.737–0.909) and maintained robust discriminative and calibration performance (Fig. 4B). DCA demonstrated that the risk score provided a greater net clinical benefit across a wide range of threshold probabilities compared to individual genes in both the training and validation cohorts (Fig. 4C). Collectively, these findings confirm that the established risk score exhibits strong diagnostic accuracy, reliable calibration, and favorable clinical utility for OA prediction. Fig. 4Performance evaluation of the elastic net-based risk score and individual genes for OA diagnosis. A ROC curves (left), PR curves (middle), and calibration curves (right) for the risk score and individual genes in the training cohort. B ROC curves (left), PR curves (middle), and calibration curves (right) for the risk score and individual genes in the validation cohort. C Decision curve analysis (DCA) showing the net clinical benefit of the risk score and individual genes at different risk thresholds in the training (left) and validation (right) cohorts Performance evaluation of the elastic net-based risk score and individual genes for OA diagnosis. A ROC curves (left), PR curves (middle), and calibration curves (right) for the risk score and individual genes in the training cohort. B ROC curves (left), PR curves (middle), and calibration curves (right) for the risk score and individual genes in the validation cohort. C Decision curve analysis (DCA) showing the net clinical benefit of the risk score and individual genes at different risk thresholds in the training (left) and validation (right) cohorts Regulatory network and genetic association analysis of OA-Related genes To explore the upstream regulatory mechanisms of key OA-associated genes, miRNA-mRNA and TF-gene regulatory networks were constructed using the miRNet platform. The miRNA-mRNA network analysis revealed that the candidate genes (such as SDC1, MMP9, MPV17, FER, MMP3, and APOD) were potentially targeted by a wide range of miRNAs, indicating a complex post-transcriptional regulatory landscape (Fig. 5A). However, as no data are available to validate the miRNome analysis, the results should be interpreted cautiously, and further experimental validation is needed to confirm these miRNA-mRNA interactions. Additionally, several transcription factors were identified as potential upstream regulators of these genes (Fig. 5B), suggesting the involvement of diverse transcriptional control pathways in OA pathogenesis. However, since these interactions were not experimentally validated, we recognize this as a limitation of the study. To further investigate the genetic basis of these genes, GWAS summary statistics for each candidate gene were extracted from the GWAS Catalog. Locus-specific Manhattan plots showed several SNPs with significant association signals within or near these genes (Fig. 5C), supporting a possible genetic contribution of these loci to OA susceptibility. Collectively, these results suggest that the identified OA-related genes are subject to extensive regulatory control at both the transcriptional and post-transcriptional levels, and may also harbor genetic variants associated with disease risk. Collectively, these results suggest that the identified OA-related genes are subject to extensive regulatory control at both the transcriptional and post-transcriptional levels, and may also harbor genetic variants associated with disease risk. However, further experimental validation is needed, particularly for the miRNA-mRNA network analysis, which remains speculative in the absence of experimental data. Fig. 5Regulatory network and genetic association analysis of OA-related genes. A miRNA-mRNA regulatory network showing predicted and validated upstream miRNAs targeting the key OA-related genes identified in this study. B TF-gene regulatory network displaying the potential upstream TFs regulating the expression of these candidate genes. C Locus-specific Manhattan plots for SDC1, MMP9, MPV17, FER, MMP3, and APOD, showing the GWAS association signals (– log10(p-value)) for SNPs located within or near each gene Regulatory network and genetic association analysis of OA-related genes. A miRNA-mRNA regulatory network showing predicted and validated upstream miRNAs targeting the key OA-related genes identified in this study. B TF-gene regulatory network displaying the potential upstream TFs regulating the expression of these candidate genes. C Locus-specific Manhattan plots for SDC1, MMP9, MPV17, FER, MMP3, and APOD, showing the GWAS association signals (– log10(p-value)) for SNPs located within or near each gene Characterization of inflammation and immune microenvironment in OA The OA risk score showed significant positive correlations with the expression levels of inflammatory cytokines, including TNF, IL1B, and IL8 (Fig. 6A), suggesting that higher risk scores are associated with increased inflammatory activity. Using the CTD, inference analysis revealed that several key genes, such as MMP3, MMP9, SDC1, and FER, exhibited high inference scores for osteoarthritis, indicating strong disease relevance (Fig. 6B). Immune and stromal cell infiltration was estimated with MCPcounter, quanTIseq, and xCell algorithms. The results demonstrated significant differences in the abundance of various immune and stromal cell populations between OA and normal samples, with fibroblasts, macrophages, and other cell types showing altered infiltration levels in OA tissues (Fig. 6C). Further correlation analysis indicated that the expression of key OA-related genes was significantly associated with fibroblast infiltration scores, both in OA and normal samples, highlighting the potential involvement of these genes in stromal cell regulation (Fig. 6D). Collectively, these findings indicate that OA-related gene expression patterns are closely linked to inflammation and distinct cellular microenvironment features in OA. Fig. 6Inflammation and immune microenvironment analysis in OA. A Spearman correlation analysis between the OA risk score and the expression of inflammatory cytokines (TNF, IL1B, IL8). B Disease inference scores for each key gene in osteoarthritis. C Heatmaps and differential analysis of immune and stromal cell infiltration estimated by MCPcounter, quanTIseq, and xCell algorithms in OA and normal samples. D Spearman correlation between the expression of key OA-related genes and fibroblast infiltration scores in OA and normal groups Inflammation and immune microenvironment analysis in OA. A Spearman correlation analysis between the OA risk score and the expression of inflammatory cytokines (TNF, IL1B, IL8). B Disease inference scores for each key gene in osteoarthritis. C Heatmaps and differential analysis of immune and stromal cell infiltration estimated by MCPcounter, quanTIseq, and xCell algorithms in OA and normal samples. D Spearman correlation between the expression of key OA-related genes and fibroblast infiltration scores in OA and normal groups Accumulation of ROS promotes the inflammation associated with OA in vitro To investigate the influence of ROS on OA in vitro, we induced ROS by adding LPS to FLS cells to simulate the onset of OA. LPS treatment is a well-established method for activating inflammation, and in our study, it helped to confirm the role of stromal infiltration in the synovium, which is commonly seen in OA. Initially, we conducted wound healing assays, which revealed that LPS treatment significantly enhanced the migratory capacity of FLS cells (Fig. 7A). In subsequent Transwell assays, we also observed that LPS substantially promoted the metastatic and invasive abilities of FLS cells (Fig. 7B). These results suggest a phenotypic change in FLS cells that is associated with inflammation. Additionally, we assessed the expression changes of inflammatory markers using Western blotting and qRT-PCR. The results indicated that following LPS treatment, the expressions of IL-6, TNF-α, COX-2, MMP-1, MMP-3, and MMP-13 were significantly upregulated (Fig. 7C). Notably, consistent with our prior bioinformatics analysis, qRT-PCR results showed that the expression levels of FER, MMP9, MPV17, MMP3, and SDC1 were significantly increased, while the expression levels of APOD and DDIT4 were significantly decreased (Fig. 7D). Through the DCFH-DA fluorescent probe, we observed that LPS would induce an increase in ROS levels in FLS cells (Fig. 7E). This further validated the accuracy of our initial screening results. While these findings support a link between ROS accumulation and the inflammatory progression of OA, we acknowledge that the observed phenotypic behaviors in FLS cells could be influenced by the target genes (such as FER and SDC1), but we have not directly assessed the involvement of these genes in the phenotypic changes. Future studies will aim to investigate the direct role of these genes in mediating the observed changes in cell behavior. In summary, our findings indicate that ROS is closely related to the inflammatory progression of OA, and the accumulation of ROS significantly promotes the inflammatory storm observed in OA. While we recognize the limitations of using FLS cells, this model provides valuable insights into the synovial response in OA, and future studies will aim to incorporate primary cells for a more comprehensive validation. Fig. 7Accumulation of ROS promotes the inflammation associated with OA in vitro. A Experimental results of wound healing in FLS cells induced by LPS (n = 3 replicates per condition). B Experimental results of the transwell assay in FLS cells induced by LPS (n = 3 replicates per condition). C After LPS treatment, the expression of inflammatory markers in FLS cells was significantly elevated, including IL-6, TNF-α, COX-2, MMP-1, MMP-3, and MMP-13 (n = 3 replicates per condition). D In the LPS-induced FLS cells, qRT-PCR was performed to assess the expression levels of potential biomarkers (n = 3 replicates per condition). E The intracellular ROS levels of FLS cells (n = 3 replicates per condition, magnification: ×200, Bar = 100 μm). **p < 0.01, ***p < 0.001, ****p < 0.0001 Accumulation of ROS promotes the inflammation associated with OA in vitro. A Experimental results of wound healing in FLS cells induced by LPS (n = 3 replicates per condition). B Experimental results of the transwell assay in FLS cells induced by LPS (n = 3 replicates per condition). C After LPS treatment, the expression of inflammatory markers in FLS cells was significantly elevated, including IL-6, TNF-α, COX-2, MMP-1, MMP-3, and MMP-13 (n = 3 replicates per condition). D In the LPS-induced FLS cells, qRT-PCR was performed to assess the expression levels of potential biomarkers (n = 3 replicates per condition). E The intracellular ROS levels of FLS cells (n = 3 replicates per condition, magnification: ×200, Bar = 100 μm). **p < 0.01, ***p < 0.001, ****p < 0.0001 Accumulation of ROS promotes the inflammation associated with OA in vivo To investigate the relationship between ROS and OA in vivo, we established a rat OA model. HE staining results indicated significant immune cell infiltration and structural alterations in the affected tissues, confirming the successful establishment of the rat OA model (Fig. 8A). We clarified that the in vivo tissue analyzed was joint tissue, and the primary focus was on immune cell infiltration and cartilage damage in the synovial joint. This provided a more holistic view of the disease progression, including both inflammatory and structural changes in the joint, compensating for the lack of multiple cellular models used in vitro. To assess the inflammatory status in rats, we measured the expression levels of IL-6, TNF-α, COX-2, as well as MMP-1, MMP-3, and MMP-13 using Western blotting and qRT-PCR. The results showed that all inflammatory markers were significantly elevated in the OA rat group compared to the control group (Fig. 8B), indicating the presence of an inflammatory storm within the rats. Additionally, we assessed the expression of bioinformatics-predicted biomarkers in rat joint tissues through qRT-PCR. Consistent with the in vitro findings, the expression levels of FER, MMP9, MPV17, MMP3, and SDC1 were significantly increased in the OA group, while the expression levels of APOD and DDIT4 were significantly decreased (Fig. 8C). Meanwhile, compared with the Sham group, the ROS levels in the OA group were increased (Fig. 8D). In summary, our results from both in vitro and in vivo experiments indicate that the accumulation of ROS significantly exacerbates the inflammatory progression of OA. Fig. 8Accumulation of ROS promotes the inflammation associated with OA in vivo. A The HE staining results of the OA animal model demonstrated significant immune cell infiltration and structural alterations in the affected tissues, confirming the presence of osteoarthritis pathology (n = 6 animals per group). B The expression of inflammatory markers in the OA rat model, including IL-6, TNF-α, and COX-2, MMP-1, MMP-3, and MMP-13, was significantly elevated (n = 6 animals per group). C qRT-PCR was performed to assess the expression levels of potential biomarkers in the rat model (n = 6 animals per group). D The ROS levels for the Sham and OA groups (n = 6 animals per group, magnification: ×200, Bar = 100 μm). **p < 0.01, ***p < 0.001, ****p < 0.0001 Accumulation of ROS promotes the inflammation associated with OA in vivo. A The HE staining results of the OA animal model demonstrated significant immune cell infiltration and structural alterations in the affected tissues, confirming the presence of osteoarthritis pathology (n = 6 animals per group). B The expression of inflammatory markers in the OA rat model, including IL-6, TNF-α, and COX-2, MMP-1, MMP-3, and MMP-13, was significantly elevated (n = 6 animals per group). C qRT-PCR was performed to assess the expression levels of potential biomarkers in the rat model (n = 6 animals per group). D The ROS levels for the Sham and OA groups (n = 6 animals per group, magnification: ×200, Bar = 100 μm). **p < 0.01, ***p < 0.001, ****p < 0.0001 Discussion This study integrates multiple OA transcriptomic datasets to clarify the role of ROS-related pathways in disease pathology and their interaction with the immune microenvironment. Our findings demonstrated significant dysregulation of ROS signaling pathways and identified a clear association between these pathways and immune-stromal infiltration, particularly involving macrophages and fibroblasts. Consistent with recent reports, we observed elevated ROS-related pathway activity in OA, paralleling earlier studies that linked oxidative stress to chondrocyte apoptosis and extracellular matrix breakdown [17]. Notably, ROS-driven M1 macrophage polarization has been shown to exacerbate synovial inflammation and cartilage degradation in OA models [18]. Our data extend these findings by showing a positive correlation between a ROS-based risk score and inflammatory cytokines, supporting the concept that ROS and macrophage-centric inflammation act in concert to drive disease progression. Single-cell and spatial transcriptomic studies have underscored the importance of immune and stromal components, including macrophages, T cells, and fibroblasts, in OA synovial pathology [19, 20]. Consistent with these findings, our deconvolution analysis identified increased infiltration of macrophages and fibroblasts in OA tissues, with significant associations between ROS-related gene expression and fibroblast abundance. This suggests a feed-forward mechanism where ROS promotes inflammation and stromal remodeling, further intensifying joint degeneration. The seven-gene risk score developed in this study achieved robust diagnostic performance (AUC > 0.82) across training and validation cohorts, outperforming individual markers finding in line with recent multi-omics approaches aimed at improving OA diagnosis accuracy. Integrating ROS signaling and immune cell patterns into the model improves its clinical relevance, offering a tool for early detection and disease stratification. Moreover, the interplay among ROS production, macrophage-mediated inflammation, and fibroblast activation highlights potential therapeutic avenues targeting oxidative stress and immune reprogramming. Recent evidence suggests that reducing ROS levels can shift macrophages from the pro-inflammatory M1 phenotype to the anti-inflammatory M2 phenotype, thereby mitigating cartilage damage in OA models [21]. Additionally, modulation of fibroblast phenotypes has been proposed as a strategy to alleviate pain and inflammation. These findings support the translational potential of combining antioxidant therapies with immune/stromal modulation in OA treatment. While our findings are informative, certain limitations should be acknowledged. First, the analyses were based entirely on publicly available transcriptomic datasets, which may introduce bias due to sample heterogeneity in clinical features and experimental conditions. Second, immune and stromal cell infiltration was estimated through computational algorithms rather than direct experimental validation, which may not fully capture cellular complexity. Third, although the diagnostic model showed strong performance across datasets, prospective validation in independent clinical cohorts is needed to confirm its utility. Fourth, while gene expression changes after LPS stimulation in FLS cells support a link to the phenotypic behavior observed, the connection between the observed phenotypic behavior in FLS cells and the target genes (FER, SDC1, etc.) was not included in this study. These assays would provide more direct mechanistic evidence of ROS accumulation and its relationship to inflammation in OA. Due to time and resource constraints, these experiments were not conducted, but they are essential for future work to establish a clear causal link between ROS and disease progression. Without these functional experiments, the involvement of these genes remains speculative. These experiments will be crucial for confirming the role of these genes in OA progression and inflammation. In conclusion, our integrative analysis provides new insights into the intricate relationship between ROS-related molecular signatures and the immune microenvironment in OA. The identified diagnostic risk score demonstrates promise for early OA detection, while the highlighted regulatory and inflammatory networks offer novel therapeutic directions. Continued validation, particularly through direct ROS assays and functional validation experiments (such as gene silencing/overexpression), and deeper mechanistic investigations are essential for leveraging these findings towards improved clinical management of OA. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 6