Identification and Experimental Validation of Key Biomarkers for Rheumatoid Arthritis Based on Bioinformatics Analysis and Machine Learning

Abstract

Objective

Rheumatoid arthritis (RA) is a chronic autoimmune joint disease driven by dysregulated immune cells and transcription factors. Despite known molecular alterations, systematic screening of key biomarkers and their link to the immune microenvironment remains lacking, particularly regarding extensive multi-algorithm cross-validation across multiple independent cohorts. This study employs bioinformatics and machine learning to identify potential RA biomarkers, aiming to support diagnosis and targeted therapy.

Methods

Multiple RA-related transcriptomic datasets derived from synovial tissue were integrated from the GEO database to screen differentially expressed genes (DEGs). Weighted gene co-expression network analysis (WGCNA) was performed to identify RA-associated modules. A total of 107 parameter and algorithm permutations from 11 distinct machine learning approaches were employed to screen key feature genes. The optimal model was selected based on average AUC values across training and validation sets, and the final three genes were identified by integrating individual diagnostic performance, biological relevance, and experimental validation. Diagnostic performance was evaluated using receiver operating characteristic (ROC) curves, while decision curve analysis (DCA) and confusion matrices were applied to validate the clinical net benefit and classification performance of the model. Immune infiltration analysis was used to assess alterations in immune cell composition within the RA microenvironment. Collagen-induced arthritis (CIA) was used to establish rat models of RA in Sprague-Dawley (SD) rats with a modest sample size (control n = 4, CIA n = 6). Ankle joint tissues were harvested for pathological examination, and the key targets were further validated by immunohistochemistry, serving as a preliminary biological corroboration of the computational findings.

Results

Through differential expression analysis and WGCNA, a set of RA-related candidate genes was identified. After combined screening using 107 parameter and algorithm permutations and ROC curve evaluation, FOSL2, JUN, and EGR1 were ultimately determined as potential biomarkers for RA. These genes demonstrated good individual diagnostic accuracy (AUC > 0.8). Immune infiltration analysis consistently revealed significant enrichment of mast cells in the RA microenvironment. The CIA model rats were successfully established, and immunohistochemistry results showed significantly high expression of FOSL2, JUN, and EGR1 in the synovial tissue.

Conclusion

This study identifies FOSL2, JUN, and EGR1 as potential markers for RA, supporting their potential roles in RA pathogenesis and clinical application.

Introduction

Rheumatoid arthritis (RA) is a chronic systemic autoimmune disease characterized by persistent synovitis and progressive destruction of articular cartilage and bone.^1,2 It typically presents with persistent, symmetric polyarticular swelling and pain accompanied by morning stiffness.³ The prolonged disease course often results in irreversible joint deformities and functional disability.⁴ Epidemiological data indicate that the global prevalence of RA ranges from approximately 0.5% to 1.0%, with both incidence and disease burden showing an increasing trend over time.^5,6 However, due to the high heterogeneity of RA pathogenesis and the accompanying complex systemic comorbidities, early accurate diagnosis and treat-to-target management remain major challenges.^7,8 Therefore, in-depth exploration of the pathogenic network of RA and the development of novel assessment tools hold significant translational medicine value for optimizing clinical decision making and improving long-term patient outcomes.^9,10

In recent years, the development of high throughput sequencing and bioinformatics has facilitated the mining of key molecules in complex diseases.¹¹ In RA, bioinformatics enables efficient screening of differentially expressed genes, identification of pathways, and characterization of the immune microenvironment.^12,13 However, traditional analyses relying on a single method are insufficient to fully capture the heterogeneity of RA. Machine learning, owing to its feature selection and classification prediction capabilities, has been introduced into RA biomarker research. Common algorithms such as LASSO, SVM, random forest, and GBM can construct robust models.¹⁴ In addition to diagnostic modeling, characterizing the immune microenvironment is essential because RA is fundamentally driven by immune dysregulation, and understanding the cellular context of candidate biomarkers may provide insights into their pathogenic roles. Previous studies have identified key diagnostic genes in metabolic syndrome-associated RA by integrating bioinformatics analysis, machine learning, and molecular docking.¹⁵ The combination of immune infiltration analysis can further reveal the regulatory relationships between genes and immune cells.¹⁶ Most current studies remain limited to single datasets or a small number of algorithm combinations, lacking multi-center and multi-algorithm cross validation, which leads to limited generalizability of biomarkers.

The activator protein-1 (AP-1) family, comprising FOS and JUN proteins, is critically involved in RA pathogenesis through regulating inflammatory and destructive pathways.¹⁷ Nevertheless, whether specific AP-1 members or functionally related transcription factors could serve as diagnostic biomarkers for RA remains to be systematically explored. Based on this background, the present study integrates multiple independent RA transcriptomic datasets derived from synovial tissue with machine learning approaches to systematically identify key differentially expressed genes and hub genes. SHAP analysis was employed to interpret the predictions generated by the machine learning model, addressing the “black box” nature of complex algorithms and enhancing clinical interpretability. These findings are then validated at the tissue level to explore the potential roles of these genes in RA pathogenesis and to provide new theoretical insights and candidate targets for early diagnosis and targeted therapy.

Materials and Methods

Reagents and Antibodies

Collagen Type II (20022, Chondrex, USA), Incomplete Freund’s adjuvant (7002, Chondrex, USA), Polysine Microscope Adhesion Slides (YA0170, Solarbio, China), Hematoxylin-Eosin (HE) Stain Kit (G1126, Solarbio, China), the Modified Saffron-O and Fast Green Stain Kit (G1371, Solarbio, China), Neutral Balsam (G8590, Solarbio, China), Anti-JUN antibody (22114-1-AP, Proteintech, USA), Anti-TPSAB1 antibody (13343-1-AP, Proteintech, USA), Anti-Egr1 antibody (ab300449, Abcam, UK), Anti-FOSL2 antibody (AF5345, Affinity Biosciences, China), HRP-conjugated goat anti-rabbit IgG secondary antibody (G1302, Servicebio, China), DAB chromogen (G1212, Servicebio, China).

Data Acquisition and Collection

Transcriptome data from seven independent datasets in the Gene Expression Omnibus¹⁸ (GEO, http://www.ncbi.nlm.nih.gov/geo) database, all derived from synovial tissue samples, comprising 79 patients with RA and 63 healthy controls, were analyzed. The dataset accession numbers involved are GSE12021,¹⁹ GSE55235,²⁰ GSE55457,²¹ GSE77298,²¹ GSE206848, GSE1919,²² and GSE29746.²³ The “combat” algorithm²⁴ from the “sva” package²⁵ was applied to normalize GSE12021, GSE206848, GSE55235, GSE55457 and GSE77298. These datasets were then merged into a training set. The “normalizeBetweenArrays” algorithm²⁶ in the “limma” package²⁷ was applied for data normalization. GSE1919 and GSE29746 served as two independent validation cohorts. Further quality control validation was performed using principal component analysis (PCA) and boxplots.²⁸ Detailed information regarding the platform, sample characteristics, and GSE accession numbers for each dataset is shown in Table 1.

Table 1.

Summary of GEO Datasets Utilized in This Study

GSE Series	Samples	Platform	Group	Tissue Source
GSE12021	24 RA and 13 controls	GPL96	Training cohort	Synovium
GSE206848	2 RA and 7 controls	GPL570	Training cohort	Synovium
GSE55235	10 RA and 10 controls	GPL96	Training cohort	Synovium
GSE55457	13 RA and 10 controls	GPL96	Training cohort	Synovium
GSE77298	16 RA and 7 controls	GPL570	Training cohort	Synovium
GSE1919	5 RA and 5 controls	GPL91	Validation cohort	Synovium
GSE29746	9 RA and 11 controls	GPL4133	Validation cohort	Synovium

Abbreviations: RA, Rheumatoid Arthritis; GSE, Gene Expression Omnibus series.

Identification of Candidate Key Targets for RA

The “limma” package was used to screen differentially expressed genes (DEGs) in the training set. The screening criteria were |log₂ fold change| > 0.585 and an adjusted P-value < 0.05. Weighted gene co-expression network analysis (WGCNA) was performed to construct a co-expression network for RA. Genes with a standard deviation greater than 0.5 were screened for inclusion in the analysis. Based on the scale-free topology criterion, the soft-thresholding power that made the model fit index R² first reach 0.8 was selected. The optimal power value was automatically estimated as 6 using the pickSoftThreshold function. The topological overlap matrix was used to calculate distances between genes, and modules were identified by the dynamic tree-cut algorithm with parameters set as minModuleSize = 60 and deepSplit = 2. Similar modules were merged at a merge height of 0.25 to obtain the final modules. The module exhibiting the highest absolute correlation with RA status was chosen as the RA-associated module for downstream gene extraction and analysis. Genes associated with RA were retrieved from the Genecard database (https://www.genecards.org/).²⁹ The screening criterion was a Relevance score ≥ 5. Finally, the overlapping genes among the DEGs, the module genes identified by WGCNA, and the genes retrieved from the Genecard database were defined as RA-related candidate key targets.

Enrichment Analysis of Candidate Key Targets in RA

The “clusterProfiler” package was used for Gene Ontology (GO) enrichment analysis.^30,31 This analysis identified the enrichment of RA-related DEGs in biological processes, cellular components, and molecular functions. Subsequently, GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were performed to further explore the molecular mechanisms of these candidate key targets.³²

Machine Learning Algorithm

To construct the optimal model, 107 parameter and algorithm permutations from 11 distinct machine learning approaches were integrated.³³ These included LASSO,³⁴ ridge regression,³⁵ elastic network (Enet),³⁶ stepwise generalized linear model (Stepglm),³⁷ random forest (RF),³⁸ generalized boosted regression modeling (GBM),³⁹ generalized linear model boost (glmBoost),⁴⁰ partial least squares regression for generalized linear models (plsRglm),⁴¹ linear discriminant analysis (LDA),⁴² naïve Bayes,⁴³ and support vector machine (SVM).⁴⁴ The model was constructed on the training set and its performance was evaluated using two independent validation sets. AUC values and confusion matrices were calculated for the training set and each validation set. The optimal model was selected based on its AUC values and the number of genes included, which helped ensure generalizability and reduce overfitting. A stepwise generalized linear model (backward method) was used for feature selection. On the selected feature subset, a generalized boosted regression model was constructed. The shrinkage parameter was set to 0.001.

SHAP Model for RA Diagnosis

SHAP (SHapley Additive exPlanations) is a methodology designed to interpret predictions generated by machine learning models. Its primary objective is to quantify the contribution of each input feature to a specific prediction outcome, referred to as the SHAP value. In the present study, a repeated five fold cross validation strategy was adopted, wherein the training dataset was partitioned into five subsets of equal size. During each iteration of the cross validation process, four folds were utilized as the training set, while the remaining fold served as the validation set for evaluating model performance. The SHAP values provide clear insight into which features exert the most significant influence on a given prediction and whether their effects are positive or negative. The advantage of SHAP is that it provides both local and global interpretability. The former clarifies individual predictions, while the latter reveals the overall decision making mechanism.

Immune Cell Infiltration Analysis

The CIBERSORT algorithm was used to estimate the relative abundance of 22 immune cell types in each sample.⁴⁵ Boxplots were generated with the “reshape2” and “ggpubr” packages to compare immune infiltration differences between control and experimental groups. Spearman correlation analysis of immune cells in the experimental group was performed using the “corrplot” package, and a heatmap was drawn. The “ggpubr” and “ggExtra” packages were employed to visualize correlations between key gene expression levels and immune cell abundances. Scatter plots combined with marginal density plots were used to characterize the immune microenvironment of RA.

RA Rat Model Establishment

Ten 6-week-old male SD rats were housed under controlled conditions (23 ± 2 °C, 45 ± 5% humidity, 12-h light/dark cycle). After one week of acclimatization, four rats were assigned to the control group. The remaining six rats were used for CIA model induction (expected success rate 80%). For CIA induction, bovine type II collagen was emulsified with an equal volume of IFA on ice. The final emulsion concentration was 1 mg/mL. On day 0, each rat received an intradermal injection of 100 μL emulsion at the base of the tail. A booster injection of the same dose was given at the contralateral tail base on day 7. Control rats received an equal volume of sterile saline following the same injection schedule. One month after booster injection,⁴⁶ all rats were deeply anesthetized with isoflurane (4% for induction and 2% for maintenance in 100% oxygen at a flow rate of 1 L/min). Under deep anesthesia, the rats were euthanized via abdominal aortic exsanguination in accordance with the American Veterinary Medical Association (AVMA) Guidelines for the Euthanasia of Animals. Subsequently, ankle joint tissues were harvested for histological analysis. To minimise potential confounders, we arranged cages in a randomised block design on the racks. All injections and tissue collections were performed between 9:00 AM and 10:00 AM to avoid circadian effects. The investigator who performed CIA induction and tissue collection was aware of group allocation. Histopathological scoring and immunohistochemistry quantification were performed by an operator blinded to group assignment.

Preparation of Ankle Joint Paraffin Sections

After the rats were sacrificed, the entire ankle joints were dissected and fixed in 4% paraformaldehyde solution at room temperature for 48 hours. After fixation, the specimens were decalcified in 10% ethylenediaminetetraacetic acid solution with pH 7.4 for 4 weeks. The decalcifying solution was replaced every 3 days. The endpoint of decalcification was confirmed by a needle puncture test. A needle was used to pierce through the bone. If the needle passed through easily without obvious resistance, decalcification was considered complete. After confirming decalcification, the ankle joints were removed and cut into left and right halves along the sagittal plane using a sharp scalpel. After sectioning, the internal bone tissue appeared soft without hard nodules. Subsequently, the tissues were dehydrated through a graded ethanol series of 70%, 80%, 90%, 95%, and 100%, each for 1 hour. The tissues were then cleared in xylene twice, 30 minutes each time, and infiltrated with paraffin at 60°C for 2 hours. The specimens were then embedded in paraffin blocks. Serial coronal sections of 4 µm thickness were cut using a rotary microtome. The sections were floated in a 40°C water bath, mounted on polysine-coated adhesion slides, and dried overnight at 37°C. Before staining, the sections were dewaxed in xylene twice, 10 minutes each time, and rehydrated through a graded ethanol series of 100%, 95%, 90%, 80%, and 70% to distilled water. The processed sections were subsequently used for HE staining, SO/FG staining, and immunohistochemistry.

Hematoxylin and Eosin (HE) Staining of Rat Ankle Joints

Paraffin-embedded ankle joint sections were dewaxed and rehydrated. They were then stained with hematoxylin and eosin following standard procedures. After dehydration and clearing, sections were mounted with neutral balsam. Images were captured using a white-light microscope. Nuclei appeared blue and cytoplasm appeared red. The decalcification procedure for ankle joints was performed using the same method as described in our previous study.⁴⁷ The severity of synovial inflammation was evaluated using the Krenn synovitis scoring system.⁴⁸

Safranin O-Fast Green Staining of Rat Ankle Joints

Paraffin-embedded ankle joint sections were dewaxed and rehydrated. They were then stained with bone tissue fast green and safranin O solutions according to the manufacturer’s instructions. After rapid dehydration and clearing, sections were mounted with neutral balsam. Images were captured using a white-light microscope. Cartilage appeared red or orange-red, while bone appeared green. Cartilage degeneration was assessed using the OARSI (Osteoarthritis Research Society International) scoring system.⁴⁹

Assessment of Model Success

The success of CIA model induction was determined by histopathological evaluation of the ankle joint in a blinded manner at the experimental endpoint. Rats with a Krenn synovitis score (0–9 scale) of < 5 were considered induction failures and were excluded from further data analysis. All six CIA-induced rats met the criterion of a Krenn synovitis score ≥ 5, indicating high-grade established synovitis, and were therefore included in the final histopathological and immunohistochemical analyses.

Immunohistochemistry

Paraffin-embedded ankle joint sections were dewaxed and rehydrated. Antigen retrieval was performed according to the requirements of each primary antibody. Endogenous peroxidase activity was blocked, followed by blocking of non-specific binding. Sections were then incubated overnight at 4 °C with primary antibodies separately against JUN, TPSAB1, Egr1, and FOSL2. After washing, sections were incubated with the secondary antibody. DAB was used for visualization. Sections were counterstained with hematoxylin, then dehydrated, cleared, and mounted. All captured images were processed using ImageJ software. All images were converted to 8-bit grayscale format. Background brightness and contrast were uniformly normalized across all groups using consistent parameters to minimize inter-image variability. A fixed threshold was manually applied to discriminate brown DAB-positive signals from the blue hematoxylin counterstain, generating binary masks of the positive areas. The software then automatically calculated both the positive-stained area and the total synovial tissue area within each field. The positive area proportion for each field was subsequently calculated as (positive area/total tissue area) × 100%. Statistical analysis was performed using GraphPad Prism. For negative controls, serial sections were processed identically but with the primary antibody omitted and replaced with phosphate-buffered saline (PBS) at the same dilution.

Data Analysis

Data analysis was performed using SPSS software. Continuous variables are expressed as mean ± standard deviation (SD). The Mann–Whitney U-test was used for between-group comparisons. A p-value of less than 0.05 was considered statistically significant.

Results

Identification of Candidate Key Targets for RA

The training set consisted of five datasets (GSE12021, GSE206848, GSE55235, GSE55457, GSE77298) with 65 RA and 47 control samples. Two independent datasets (GSE1919, GSE29746) containing 14 RA and 16 control samples served as validation cohorts. Batch effects were effectively removed, as shown by boxplots and PCA plots (Figures 1A–D). Differential expression analysis identified 1149 DEGs (695 up-regulated, 454 down-regulated) (Figure 1E). WGCNA was performed to determine the optimal soft-thresholding power and generate the gene dendrogram with module color partitioning (Figure 1F and G). WGCNA revealed nine co-expression modules, among which the magenta module showed a negative correlation with RA (correlation =−0.17, P= 0.09) (Figures 1H and I).

Figure 1.

Recognition of RA related differentially expressed genes. (A) Pre-batch correction boxplot of training set. (B) Post-batch correction boxplot of training set. (C) PCA plot before normalization. (D) PCA plot after normalization. (E) Volcano plot displaying DEGs between normal and RA samples. (F) Gene dendrogram and module colors. (G) Selection of the soft-thresholding power. (H) Correlations between gene modules and clinical traits. (I) Screening of the modules most strongly associated with RA progression.

Abbreviations: RA, rheumatoid arthritis; PCA, principal component analysis; DEGs, differentially expressed genes.

Comprehensive Enrichment Analysis of Candidate Key Genes in RA

To identify candidate key targets associated with RA, the gene sets from WGCNA, DEGs, and the GeneCards database were integrated, and 31 overlapping genes were obtained by taking the intersection of the three datasets (Figure 2A). GO enrichment analysis revealed overrepresentation of biological processes including response to steroid hormone, response to oxygen levels, fat cell differentiation, and positive regulation of miRNA metabolic transcription. Enriched cellular components included RNA polymerase II transcription regulator complex, euchromatin, transcription repressor complex and organelle outer membrane. Overrepresented molecular functions comprised DNA-binding transcription activator activity, DNA-binding transcription factor binding and DNA-binding transcription repressor activity (Figure 2B and C). KEGG pathway analysis further mapped these candidate key targets to several core RA-associated signaling pathways. The most significantly enriched pathways included osteoclast differentiation, MAPK signaling pathway, TNF signaling pathway, and IL-17 signaling pathway (Figure 2D).

Figure 2.

Integrated Analysis of Candidate Key Targets and Pathways in RA. (A) The venn diagram showing overlapping genes from WGCNA, DEGs, and GeneCards for candidate key gene screening. (B) GO bubble plot of key candidate genes. (C) GO enrichment circle plot. (D) KEGG bubble plot of key candidate genes.

Abbreviations: RA, rheumatoid arthritis; WGCNA, Weighted gene co-expression network analysis; DEGs, differentially expressed genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Machine Learning-Based Identification and Validation of Key Genes for RA

To construct a diagnostic prediction model for RA, this study integrated 107 parameter and algorithm permutations from 11 distinct machine learning approaches. Performance was ranked based on the average AUC values from the training set and two independent validation sets. The Stepglm(backward)+GBM combination ranked first and was selected as the optimal model. In the training set, most top-ranked algorithms had AUC values above 0.85 (Figure 3A). Confusion matrix analysis for the training set and two validation datasets further confirmed stable classification performance. The true positive and true negative rates were balanced (Figure 3B). ROC curve analysis showed that the selected model had an AUC of 0.987 (95% CI: 0.965–1.000) in the training set. In the GSE1919 validation set, the AUC was 0.960 (95% CI: 0.840–1.000). In the GSE29746 validation set, the AUC was 0.808 (95% CI: 0.576–0.980) (Figure 3C). These values represent the performance of the combined diagnostic model integrating all 13 feature genes.

Figure 3.

Machine Learning Model Construction and Validation in RA. (A) ROC curves for 107 parameter and algorithm permutations from 11 machine learning approaches. (B) Confusion matrices of the optimal model in the training and validation datasets. (C) ROC curves of the optimal model in the training and validation datasets. (D) ROC curves of individual hub genes. (E) DCA of the hub genes. (F) Nomogram of the diagnostic model constructed with hub genes.

Abbreviations: RA, rheumatoid arthritis; ROC, receiver operating characteristic; DCA, decision curve analysis.

Based on the feature weights of the optimal model, 13 candidate genes were selected: KLF4, JUNB, FOSL2, PPP1R15A, NFIL3, NR4A1, CDKN1A, MAFF, SOCS3, JUN, PTGS2, EGR1, NR4A2. We further evaluated the individual diagnostic performance of each candidate gene. ROC analysis of each gene in the training set showed that FOSL2 had the highest diagnostic value (AUC=0.821), followed by EGR1 (AUC=0.809) and JUN (AUC=0.808) (Figure 3D). Decision curve analysis (DCA) indicated that, over a wide range of threshold probabilities, the combined feature model based on these genes provided higher clinical net benefit than single gene models (Figure 3E). In addition, a nomogram model integrating the machine learning selected genes was constructed to quantify individual RA risk scores (Figure 3F). The three gene signature model suggests potential for high diagnostic accuracy and clinical applicability in RA.

SHAP Model Interpretation for the RA Diagnostic Model

The decision making mechanism of the machine learning diagnostic model for RA needed further elucidation. The SHAP algorithm was introduced for this purpose. The contribution and mode of action of the 13 candidate genes in the model were interpreted using this algorithm. The overall feature importance of each gene was ranked based on the mean absolute SHAP value (Figure 4A). FOSL2 was the top contributor with a mean SHAP of 0.0552, followed by JUNB with 0.0490, NFIL3 with 0.0433, JUN with 0.0397, and NR4A1 with 0.0354. The SHAP beeswarm plot revealed the association between gene expression levels and disease prediction direction (Figure 4B).

Figure 4.

SHAP analysis of the hub genes in RA diagnostic model. (A) Bar plot of mean absolute SHAP values showing the importance of core genes. (B) SHAP summary plot showing the distributional relationship between gene expression levels and SHAP values. (C) Force plot showing the combined contribution of all features to the model prediction. (D) Waterfall plot showing the individual contribution of each core gene to the model prediction. (E) SHAP dependence plot showing the relationship between gene expression and corresponding SHAP values.

Abbreviations: RA, rheumatoid arthritis; SHAP, SHapley Additive exPlanations.

High expression of FOSL2, JUNB, NFIL3, and JUN was significantly correlated with positive SHAP values. Local interpretation for a single RA sample was provided by the force directed plot and waterfall plot (Figure 4C and D). The baseline prediction value was 0.546 and the final prediction value was 0.872. JUNB was a major positive contributor with an expression level of 8.82 and a SHAP value of +0.0871. FOSL2 contributed positively with an expression level of 6.77 and a SHAP value of +0.0799. CDKN1A showed a negative effect with an expression level of 8.09 and a SHAP value of 0.0391. The SHAP dependence plot illustrated potential interactions among key genes, showing a complex regulatory network of synergy or antagonism (Figure 4E).

Immune Cell Infiltration and Correlation Analysis in RA

The CIBERSORT algorithm was used to analyze immune cell infiltration in the RA group and the control group. The relative abundance distribution of 22 immune cell subsets across all samples in the control and treatment groups is shown as a stacked bar plot (Figure 5A). A lollipop plot presents the Spearman correlation coefficients and corresponding p values between each immune cell subset and RA status (Figure 5B). The results showed that CD8⁺T cells, gamma delta T cells, and activated CD4⁺ memory T cells were positively correlated with RA. Resting CD4⁺ memory T cells and resting dendritic cells were negatively correlated with RA. Additionally, resting mast cells and activated mast cells also showed significant correlations with RA status. Boxplots comparing the fractions of each immune cell subset between the two groups are displayed (Figure 5C). The fraction of M2 macrophages was significantly increased in the treatment group. A heatmap of pairwise Spearman correlation coefficients among the 22 immune cell subsets is shown (Figure 5D). Red indicates positive correlations, and blue indicates negative correlations. A correlation network plot illustrates the relationships between 13 hub genes and the 22 immune cell subsets (Figure 5E). Green lines represent positive correlations, and orange lines represent negative correlations. The hub genes FOSL2, JUN, and EGR1 were correlated with multiple immune cell subsets, especially activated mast cells.

Figure 5.

Immune cell infiltration profiles and correlation analysis in RA. (A) Relative proportions of 22 immune cell types in Control and RA groups. (B) Correlation of immune cell abundance with RA status. Dot size represents the absolute correlation coefficient. Red indicates P < 0.05. (C) Comparison of immune cell infiltration fractions between groups. (D) Pairwise correlation heatmap of 22 immune cell types. (E) Correlation network between key RA-related genes and immune cell types.

Abbreviation: RA, rheumatoid arthritis.

Histopathological Changes in CIA Rat Model Joints

All six rats subjected to CIA induction met the predefined histopathological criterion (Krenn synovitis score ≥ 5) and were therefore included in the subsequent histopathological and immunohistochemical analyses. Histopathological examination of ankle joints was performed to verify the successful establishment of the CIA rat model. Representative HE staining of joint tissues from the control and CIA groups is shown (Figure 6A). Control rats exhibited normal joint structure with smooth synovial lining, intact bone and cartilage, and no inflammatory cell infiltration. In contrast, CIA rats showed severe synovial hyperplasia, massive inflammatory cell infiltration, pannus formation, and destruction of bone and cartilage structures. Safranin O-fast green staining results are displayed (Figure 6B). The control group showed bright red cartilage and clear green bone tissue, while the CIA group presented with faded red cartilage, significant cartilage erosion, and loss of the normal cartilage layer structure. The OARSI cartilage degeneration score was significantly higher in the CIA group than in the control group (P < 0.0001) (Figure 6C). The Krenn synovitis score also showed a marked increase in the CIA group compared with the control group (P < 0.0001) (Figure 6D).

Figure 6.

Histopathological evaluation of joint injury in CIA rats. (A) HE staining. (B) Safranin O-fast green staining. (C) OARSI score of cartilage degeneration. (D) Krenn synovitis score of synovial inflammation. Control (n=4), RA (n=6). Significance levels:****P < 0.0001.

Abbreviations: OARSI, Osteoarthritis Research Society International; HE, Hematoxylin and eosin; CIA, collagen-induced arthritis.

Immunohistochemical Validation of Key Biomarkers in CIA Rat Model

No positive staining was observed in these controls, confirming the specificity of the immunostaining (Supplementary Figure S1). Compared with the control group, CIA rat synovial tissues exhibited significantly enhanced positive staining for TPSAB1, FOSL2, JUN, and EGR1 (Figure 7A–D). The positive signals were mainly localized in the synovial lining and sublining layers. Quantitative analysis confirmed that the CIA group had a significantly higher proportion of positive area for TPSAB1, FOSL2, JUN, and EGR1 compared with the control group (P < 0.001) (Figure 7E).

Figure 7.

IHC analysis of key biomarkers in CIA rat synovium. (A) IHC staining of TPSAB1. (B) IHC staining of FOSL2. (C) IHC staining of JUN. (D) IHC staining of EGR1. (E) Statistical analysis of the proportion of positive area (%). Control (n=4), RA (n=6). Significance levels: ***P < 0.001, ****P < 0.0001.

Abbreviations: IHC, Immunohistochemical; CIA, collagen-induced arthritis.

Discussion

This study aimed to systematically identify key biomarkers associated with the pathogenesis of RA and to investigate their relationship with the immune microenvironment. Key candidate targets were obtained through DEGs, WGCNA and GeneCard. A combination of 107 machine learning algorithms was then used to successfully identify three key biomarkers, including FOSL2, JUN and EGR1. Further immune infiltration analysis revealed a close association between these key targets and mast cells. These findings suggest that FOSL2, JUN and EGR1 have diagnostic potential in RA, and that mast cells may serve as key participants in the RA immune microenvironment.

FOSL2 is a transcription factor belonging to the Fos family. It forms the activator protein 1 (AP-1) complex with Jun family members and is widely involved in the regulation of cell proliferation, differentiation, and inflammatory responses. Previous studies have shown that the AP-1 signaling axis plays a central role in the invasive phenotype of RA synovial fibroblasts. Inhibition of c-Fos/AP-1 activity significantly reduces cartilage erosion and bone destruction in collagen-induced arthritis models. Additionally, suppression of the AP-1 pathway has been shown to decrease the production of matrix metalloproteinases (MMP) induced by TNF-α in RA synovial fibroblasts⁵⁰ In the joint microenvironment, abnormal activation of inflammatory signaling pathways and the release of inflammatory factors promote the secretion of large amounts of MMPs from tissue cells, which in turn hydrolyze the extracellular matrix and cause joint damage.^50,51 In the present study, FOSL2 was identified as the feature gene with the highest contribution by the machine learning model. Its SHAP value indicated that high expression strongly points to the diagnosis of RA. This finding is consistent with previous reports showing that FOSL2, along with EGR1, serves as a hub biomarker with high diagnostic value for RA.¹⁶ However, our study substantially extends these observations through multi-dataset integration, 107-algorithm screening, identification of JUN as a novel biomarker, SHAP-based interpretability, and in vivo protein-level validation in the CIA model. Compared with the well studied c-Fos and Fra-1, the specific function of FOSL2 in RA remains rarely explored.⁵² Differential expression analysis and animal experiments have verified the abnormally high expression of FOSL2 in RA synovium. This suggests that FOSL2 may be involved in the pathological activation of synovial fibroblasts or the inflammatory cytokine cascade. While the SHAP analysis identified JUNB as the second most important contributor to the model’s predictive output, the final selection of FOSL2, JUN, and EGR1 was based on an integrative strategy combining individual diagnostic performance, biological relevance to RA pathogenesis, and experimental validation in the CIA model. Notably, SHAP importance reflects feature contributions to the ensemble model rather than individual diagnostic performance, which explains why JUNB, despite its high SHAP contribution, was not prioritized, particularly given its functional overlap with JUN and its relatively lower individual AUC value.

JUN is a core component of the AP-1 complex. Its encoded protein, c-Jun, has been identified as a key regulator of inflammatory signaling pathways in various autoimmune diseases.⁵³ In RA, inflammatory factors such as TNF-α, IL-6, and IL-1β activate JNK via the MAPK pathway. JNK subsequently promotes the phosphorylation and stabilization of c-Jun, leading to the upregulation of MMP and chemokines.⁵⁴ The KEGG enrichment analysis in this study revealed that the identified key genes were significantly enriched in pathways related to osteoclast differentiation, the MAPK signaling pathway, and the TNF signaling pathway. JUN can form heterodimers with FOSL2, a member of the same family, to coordinately regulate the expression of downstream target genes. This suggests the existence of a complex network regulation among AP-1 family members in RA.⁵⁵ Furthermore, the immune infiltration results in this study indicated a trend of correlation between JUN expression levels and the infiltration of M2 macrophages and activated mast cells in RA synovium. This implies that JUN may participate in the formation of the local inflammatory microenvironment by regulating immune cell chemotaxis or polarization. However, at the single cell level, the expression profile of JUN in mast cells and its effects on mast cell activation, degranulation, and cytokine release remain poorly understood. This may represent an important direction for future research into the regulation of mast cell function in the RA immune microenvironment.^56,57

EGR1 is a zinc finger transcription factor that can be rapidly induced by inflammatory stimuli. Existing studies have shown that EGR1 participates in immune cell activation, such as regulating T cell differentiation and macrophage inflammatory responses.⁵⁸ In terms of angiogenesis, EGR1 responds to pro-angiogenic factors such as VEGF and bFGF and plays a key role in regulating neovascularization.⁵⁹ In addition, EGR1 promotes the expression of MMP and directly participates in extracellular matrix degradation and tissue remodeling.⁶⁰ Multiple studies have reported that EGR1 expression is elevated in the synovial tissues and cartilage of RA patients, and it can upregulate the expression of effector molecules such as MMP-1, MMP-3, and IL-8, thereby promoting joint inflammation and tissue destruction.^61,62 In the present study, EGR1 was identified as a candidate biomarker for RA, and its expression was validated in the synovium of CIA rats, further supporting the key role of this factor in inflammatory arthritis. Previous evidence has indicated that EGR1 mediates cytokine release following mast cell activation, for example by regulating the production of IL-13 and TNF-α.⁶³ Nevertheless, the specific expression of EGR1 in different immune cell subsets of RA and its functional heterogeneity remain poorly understood. Future studies are needed to elucidate these issues using single cell multiomics technologies and cell lineage analysis.⁶⁴

Mast cells are tissue-resident innate immune cells distributed in the synovium. In RA, mast cells significantly accumulate in the synovium, and their granules contain bioactive mediators such as histamine and tryptase.⁶⁵ Upon activation, these cells release mediators that contribute to synovial hyperplasia and cartilage destruction.⁶⁶ The transcription factors FOSL2 and JUN can both serve as subunits of the AP-1 complex and are involved in the regulation of cytokine expression.⁶⁷ AP-1 is one of the most important signaling-responsive transcription factor families and serves as a major mediator of inflammatory cytokine action.⁶⁸ In RA, inflammatory factors such as TNF-α activate JNK via the MAPK pathway, which subsequently promotes the phosphorylation and stabilization of c-Jun.⁶⁹ EGR1 participates in mast cell activation mediated by FcεRI and c-Kit and is essential for mast cell production of TNF-α and IL-13.⁶³ Tryptase, encoded by TPSAB1, is the major neutral protease in mast cells and is released upon activation-induced degranulation, serving as a specific marker of mast cells.⁷⁰

Immune infiltration analysis revealed a close association between these candidate genes and mast cells. This finding suggests that mast cells play a more critical role in the pathogenesis of RA than other immune cell types. In the CIA rat synovium, immunohistochemistry confirmed high expression of TPSAB1 along with elevated expression of FOSL2, JUN, and EGR1. These results collectively indicate a potential link between these transcription factors and mast cell activation status in the RA synovial microenvironment. However, the precise regulatory relationships and causal mechanisms require further experimental validation.

In terms of methodology, a cross validation strategy combining 107 parameter and algorithm permutations was adopted. This approach helped reduce the risk of overfitting and selection bias that may arise from using a single algorithm. Similar integrative approaches combining bioinformatics and network pharmacology have been successfully applied in other diseases, such as methotrexate-induced liver injury.⁷¹ AUC values and confusion matrices were compared across multiple independent datasets and external validation sets. The final Stepglm(backward)+GBM combined model showed good robustness and generalizability. Previous studies using machine learning algorithms with cross validation in RA have also shown that ensemble strategies can improve model performance and stability.⁷² The relatively lower AUC observed in the GSE29746 validation cohort likely reflects the small sample size of this dataset, which inherently increases the variability of performance estimates. Despite this, the AUC of 0.808 remains clinically acceptable, and the DCA demonstrated net benefit across a wide range of threshold probabilities, supporting the model’s potential utility.

The magenta module showed the strongest correlation with RA status (r = −0.17) with a marginal P value (P = 0.09). This borderline significance likely arises from the inherent biological heterogeneity of RA and weak disease-specific transcriptional signals masked by confounding factors. In standard WGCNA workflows, module-trait correlation magnitude, rather than P value, serves as the primary criterion for module selection, as it directly reflects the overall expression association between module eigengenes and disease phenotypes, while P values only indicate statistical reliability secondarily. Prior WGCNA studies have validated marginally significant modules (0.05<P≤0.1) as biologically meaningful for disease traits, representing faint but genuine disease-related transcriptomic changes.⁷³

However, this study has several limitations. Although GEO datasets were used for validation, the inherent bias of retrospective data cannot be completely eliminated. Many bioinformatics studies of RA face similar limitations due to the retrospective nature of publicly available transcriptomic data.⁷⁴ Prospective clinical cohorts are needed to evaluate the real-world diagnostic performance of these biomarkers. In addition, protein expression was validated only at the histopathological level in a relatively small animal cohort, which limits statistical power to some extent. Furthermore, synovial tissue biopsy is invasive and not readily suitable for routine clinical screening. Whether these biomarkers or their protein products can be detected in peripheral blood as non-invasive diagnostic tools warrants future investigation. The specific regulatory hierarchy and molecular mechanisms among the three genes remain to be elucidated in future studies. Recent advances in single-cell multiomics technologies have enabled more detailed characterization of cell-type-specific gene expression patterns in autoimmune diseases, which could help address such questions in future studies.⁷⁵ In addition, although the CIA rat model is a well-established animal model for RA, translating findings from this model to human clinical practice requires further validation. The relatively small sample size of the animal cohort limits statistical power to some extent, and future studies with larger animal cohorts are warranted.

In conclusion, this study integrated multiple transcriptomic datasets and machine learning algorithms and successfully identified FOSL2, JUN, and EGR1 as diagnostic biomarkers for RA, each demonstrating high individual diagnostic performance. The significant upregulation of these genes in RA synovium was further validated at the protein level in the CIA model. Immune infiltration analysis demonstrated a significant correlation between these genes and mast cells in the RA microenvironment. These findings establish FOSL2, JUN, and EGR1 as robust biomarker candidates for RA diagnosis, though prospective clinical validation and functional studies are still required to translate these findings into clinical practice. The integrated strategy combining computational screening with animal pathological validation employed in this study can serve as a framework model for discovering biomarkers in other heterogeneous autoimmune diseases.

Funding Statement

Supported by the National Natural Science Foundation of China (No.82274435 and 82074223), the Key project at the central government level: The establishment of sustainable use for valuable Chinese medicine resources (No. 2060302). The Fifth Batch of National Training Programme for Clinical Excellence in Chinese Medicine in 2022 (No. Chinese medicine official letter of instruction (2022) 178).

Data Sharing Statement

The datasets supporting the findings of this study are available in the GEO repository, [https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE12021/GSE206848/GSE55235/GSE55457/GSE77298/GSE1919/GSE29746]. All data used in this study are publicly accessible and are also available from the corresponding authors, J.M. Wang and Y.Z. Zhang, upon reasonable request. The complete computational code and parameter settings for the 107 machine learning iterations are available at [https://github.com/doctorxiaogan/107-machine-learning] to ensure full reproducibility.

Ethics Statement

This study involving publicly available human data from the GEO database was conducted in accordance with the ethical standards of the Declaration of Helsinki. As all data are de-identified and publicly accessible, this retrospective analysis qualified for ethical exemption under the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects (promulgated on February 18, 2023, in China). All animal experiments were approved by the Animal Ethics Committee of China-Japan Friendship Hospital (approval No. ZRDWLL260009). All animal procedures adhered strictly to the 3R principles (Replacement, Reduction, and Refinement) for the ethical treatment of laboratory animals and were carried out in compliance with the national standard GB/T 35892-2018 (Laboratory animal—Guideline for ethical review of animal welfare).

Author Contributions

Yuxin Han: conceptualization, data curation, formal analysis, writing-original draft. Pengrui Wang: methodology, software, validation, writing-original draft. Yifei Wang: investigation, visualization, writing-original draft. Guangyao Chen: software, visualization, writing-original draft. Meiqi Lan: formal analysis, data curation, writing-original draft. Fei Teng: formal analysis, writing-review & editing. Xirui Liu: validation, investigation, writing-review & editing. Yuting Bian: software, visualization, writing-review & editing. Huilan Yang: investigation, data curation, validation, writing-review & editing. Liangjie Ma: formal analysis, methodology, writing-review & editing. Yi Liu: formal analysis, methodology, writing-review & editing. Jianming Wang: conceptualization, supervision, funding acquisition, writing-review & editing. Yanzhen Zhang: conceptualization, supervision, funding acquisition, writing-review & editing. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

Dr Yuxin Han reports Support for the manuscript from The Fifth Batch of National Training Programme for Clinical Excellence in Chinese Medicine in 2022, Key project at the central government level: The establishment of sustainable use for valuable Chinese medicine resources, National Natural Science Foundation of China, during the conduct of the study. The other authors report no conflicts of interest in this work.