Programmed cell death-related gene S100A9 promotes macrophage M1 polarization and chondrocyte apoptosis in rheumatoid arthritis Authors: Qingyuan Xu, Jinfu Liu, Qiang Ding, Canbin Zhao, Weiwei Wang, Hao Li, Chicheng Niu, Wei Chen, Ping Zeng, Donghui Guan, Ronghua Zhang Published: 2026 ABSTRACT AbstractBackgroundRheumatoid arthritis (RA) is a heterogeneous chronic autoimmune disease. Its high disability rate has a serious impact on individuals and society. Programmed cell death (PCD) patterns play a key role in several diseases. However, the significance of the interplay between PCD and RA remains underexplored.MethodsIn total, 18 PCD patterns were analyzed for the model construction. Single-cell RNA-seq transcriptome (scRNA-seq) and bulk RNA-seq data were collected from the GSE200815, GSE1919, GSE77298, GSE206848, GSE89408, GSE12021, GSE55235, and GSE55457 cohorts to validate the model. In vivo and in vitro experiments were performed to determine the role of S100A9 in RA.ResultsWe developed a programmed cell death-related (PCDR) model for RA using 113 combinations of 12 machine learning algorithms and significant PCD signatures; 2 RA clusters were identified. A significant difference was noted in the macrophage numbers between the two groups. Macrophages were identified as key effector cells that play a central role in RA pathogenesis through cellular communication and the transition of cell states. S100A9 was identified as a key gene in the PCDR model, and its knockdown significantly slowed RA progression by reducing joint synovitis and cartilage damage. M1 macrophage polarization was accompanied by the overexpression of S100A9 in the synovial tissues of RA model mice. Compared with RA mice, AAV-shRNA-mediated S100A9 knockdown mice showed decreased M1 macrophage polarization, attenuated severity of synovitis, and elevated expression of the cartilage phenotype proteins—collagen II and BCL-2. Additionally, S100A9 knockdown inhibited M1 macrophage polarization in vitro. Hence, S100A9 inhibition may be a promising therapeutic strategy for RA treatment.ConclusionWe established a novel PCDR model by comprehensively analyzing diverse cell death patterns. S100A9 inhibition may be a promising therapeutic strategy for RA treatment.Supplementary informationThe online version contains supplementary material available at https://doi.org/10.1186/s13062-026-00866-5.Keywords: Rheumatoid arthritis, Programmed cell death, S100A9, Macrophage polarization, Machine learning Introduction Rheumatoid arthritis (RA) is a chronic inflammatory condition resulting from an autoimmune reaction. It usually manifests as symmetrical polyarthritis, marked by joint pain, swelling, and rigidity. The initial symptoms frequently begin with pain in minor articulations of hands and feet and can impact almost every facet of daily activities [1]. Recent data from the World Health Organization indicates that approximately 1% of the popula tion is affected by RA. With increasing life expectancy, the prevalence of RA is projected to more than double by 2030, with a particularly high incidence among the elderly [2]. Although significant progress has been achieved with anti-rheumatic drugs, many patients remain unresponsive to treatment. This subset of “refractory” patients requires new therapeutic targets and strategies. Therefore, elucidating the molecular mechanisms that contribute to the pathogenesis of RA is crucial for the advancement of effective and personalized treatment strategies. Programmed cell death (PCD) is particularly important for organisms because it helps the body to maintain homeostasis in tissues and removes cells that are damaged or redundant. PCD primarily includes apoptosis, autophagy, alkaliptosis, oxeiptosis, netotic cell death, NETosis, necroptosis, parthanatos, pyroptosis, ferroptosis, lysosome-dependent cell death, entosis, anoikis, entotic cell death, immunogenic cell death, methuosis, cuproptosis, and paraptosis [3]. In the development of RA, owing to certain pathological factors, some abnormalities occur in PCD. These abnormalities cause immune cells to release many inflammatory substances and signaling molecules uncontrollably. This abnormal release process directly causes continuous inflammation of the synovium and damage to bones and joints. For instance, cell shrinkage and apoptotic body formation primarily characterize apoptosis. Tumor necrosis factor (TNF)-α can inhibit apoptosis in fibroblast-like synoviocytes (FLS) in RA through multiple mechanisms, thereby promoting the inflammatory response and pathological progression of RA [4]. Pyroptosis is a specialized form of cell death that falls under the category of inflammatory PCD. This type of cell demise is characterized by plasma membrane permeabilization, leading to cellular swelling, rupture, and the subsequent release of inflammatory substances. Given that RA-FLS demonstrate the capacity to express various gasdermins and generate IL-1β, Demarco et al. [5] proposed that these cells might undergo pyroptosis, consequently amplifying inflammatory responses and strengthening the cellular communication between FLS and macrophages. In addition to the previously mentioned content, the functions of other types of PCD (autophagy, ferroptosis, and necroptosis) in RA are widely reported [6–8]. However, studies investigating the specific functional aspects of PCD in RA remain limited. During RA, M1 polarization of macrophages persists throughout the disease course [9]. In the preclinical stage of RA, patients exhibit metabolic abnormalities that initiate a pro-inflammatory program, accompanied by the production of autoantibodies and synovial microinflammation. Macrophages undergo differentiation to adopt the M1 phenotype, during which, they secrete inflammatory mediators including IL-1β and CXCL8, thereby further intensifying the inflammatory reaction [10, 11]. As the disease progresses, M1 macrophage polarization becomes markedly enhanced. The substantial accumulation of M1-type macrophages leads to the release of matrix metalloproteinases (MMPs) and TNF-α, which act synergistically with interleukin-17 (IL-17). This cascade directly contributes to cartilage degradation and bone erosion, thereby exacerbating RA [12]. Consequently, M1 macrophage polarization is crucial in RA pathogenesis and advancement, and regulating this process presents significant promise for both preventing and treating the disease. S100A9, belonging to the S100 protein family, represents a significant protein with calcium-binding capabilities that serves a crucial function in PCD mechanisms. This protein is primarily produced in neutrophils and monocytes and is important for controlling the development of inflammatory responses and inflammatory-related diseases [13]. Over the past few years, S100A9 has attracted considerable attention regarding the equilibrium of the immune system [14], as elevated S100A9 protein levels can serve as a marker of inflammation across multiple autoimmune conditions—such as inflammatory bowel disease, systemic lupus erythematosus, various forms of vasculitis, and systemic sclerosis [15–18]. A similar pattern is observed in individuals with RA, and their serum S100A9 levels and inflammation are remarkably higher than those in healthy individuals [19, 20]. S100A9 and calprotectin are expressed by monocytes and neutrophils that infiltrate the synovium [21]. These two proteins are particularly abundant at the junction of cartilage and synovium—a precise location of the most critical cartilage damage and bone wear [19, 22]. Additionally, research by Sunahori et al. [23] established that macrophage-derived S100A8/A9 proteins function to stimulate NF-κB signaling pathways. This activation, in turn, results in more robust responses from pro-inflammatory cytokines. These findings indicate that S100A9 is involved in mechanisms related to innate immunity, which are observed in RA, and this is partly due to its local production within inflamed joints. We aimed to establish a programmed cell death-related (PCDR) model to predict RA onset and develop immunotherapeutic strategies. In parallel, we investigated the regulatory effect of the gene S100A9, a key member of the PCDR family, on M1 macrophage polarization and its therapeutic potential in RA. Materials and methods Collection and pre-treatment of raw data We obtained bulk RNA-sequencing data from the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/). This study used seven published synovial datasets: GSE1919 (RA = 5, control = 5), GSE77298 (RA = 16, control = 7), GSE206848 (RA = 2, control = 7), GSE89408 (RA = 152, control = 28), GSE12021 (RA = 12, control = 9), GSE55235 (RA = 10, control = 10), and GSE55457 (RA = 13, control = 10). Four datasets, GSE1919, GSE77298, GSE206848, and GSE89408, were combined in the test set, which included 47 normal and 175 RA synovial membranes. Three datasets, GSE12021, GSE55235, and GSE55457, were combined to form an external validation cohort that included 29 normal and 35 RA synovial membranes. The GSE89408 dataset was selected as the internal validation cohort. The complete gene lists for each PCD pattern are provided in Table S1. We used Limma (version 3.4.0) for normalization/correction of all expression microarrays and gene-name annotation. The “ComBat” function within the “sva” package was utilized to rectify batch effects across different datasets [24]. Weighted gene co-expression network analysis (WGCNA) The WGCNA algorithm was applied using the “WGCNA” package to identify gene modules significantly linked to RA. First, samples were examined to detect any apparent outliers. The correlation matrix was subsequently built using the “WGCNA” package, and it was converted into an adjacency matrix through the application of an optimal soft threshold. From this adjacency matrix, a topological overlap matrix was generated. In the third step, hierarchical clustering combined with dynamic tree cutting was performed to define modules. In the fourth step, gene significance (GS) and module membership (MM) were calculated to associate modules with clinical characteristics. For further analysis, data for the genes within each module were retrieved. Modules with the strongest correlation with RA were designated as key modules, and the genes in these modules were identified as significant PCD regulators for subsequent analysis. Analysis of functional enrichment Gene Ontology (GO) analysis serves as a conventional approach for extensive functional enrichment investigations, encompassing three main categories: biological processes (BP), molecular functions (MF), and cellular components (CC) [25]. Kyoto Encyclopedia of Genes and Genomes (KEGG) database is commonly utilized for its information on genomes, diseases, drugs, and biological pathways [26]. Enrichment analyses were conducted using clusterProfiler for the identified genes in terms of GO and KEGG pathways [27]. Results with a false discovery rate (FDR) lower than 0.05 were regarded as statistically meaningful. Establishment of PCDR model Using 12 machine-learning methods assembled into 113 combinations, we screened for significant PCDS with robust accuracy and stability. The machine-learning algorithms employed were Ridge, stepwise GLM (StepGLM), Random Forest, SVM, XGBoost, Lasso, Naive Bayes, Elastic Net (Enet), GBM, linear discriminant analysis (LDA), GLM boosting (glmBoost), and PLS regression for GLMs (plsRglm). The signature construction proceeded as follows: (a) Overall, 113 algorithm combinations among the significant PCD regulators were evaluated to calibrate diagnostic models, with calibration using 10-fold cross-validation in the test set. (b) All models were validated across three independent datasets: GSE12021, GSE55235, and GSE55457. (c) Each model’s area under the receiver operating characteristic (ROC) curve (AUC) was computed across all validation sets, and the model with the highest mean AUC was selected as optimal. A PCDS-based nomogram was then constructed to predict RA. In addition, a calibration curve was generated to assess the model’s accuracy. Decision curve analysis (DCA) and clinical impact curve were also developed to evaluate whether the model-based decisions would be beneficial for patients. Consensus clustering analysis Unsupervised clustering analysis was employed from the “ConsensusClusterPlus” package to determine the optimal number, testing values ranging from two to nine with parameters “reps = 50, pItem = 0.8, pFeature = 1”, and employing the k-means algorithm [28]. We selected 2 as the optimal clustering solution since the delta area and consensus cumulative distribution function showed minor increases beyond k = 2. Gene set variation analysis (GSVA) GSVA assigns gene-set scores to individual samples by estimating, in an unsupervised manner, the variation in gene-set enrichment across samples in an expression dataset. The “GSVA” package was employed to compute the GSVA score, which signified the combined levels of PCD-related gene expression. Subsequently, the samples were computationally segregated into two categories (Groups A and B) according to the median GSVA score. The GSVA algorithm groupings were quantified by performing dimensionality reduction using principal component analysis (PCA) to examine the reliability of the groupings and further determine the accuracy of the grouping types. Immune cell infiltration analysis The infiltration of distinct immune cells in the different groups was quantified using four independent algorithms, including the xCell [29], QuanTISeq [30], single-sample gene-set enrichment analysis (ssGSEA) [31], and CIBERSORT [32]. Heatmaps and boxplots were used to visualize immune cell abundance. scRNA-seq data processing The scRNA-seq dataset GSE200815, comprising four patients with RA, was selected for this study. A Seurat object was created from the dataset using the Seurat package. For the analysis that followed, it was necessary for cells and genes to fulfill quality-control standards: (1) A gene should be identified in a minimum of three cells; (2) a minimum of 200 genes must be expressed per cell; and (3) no more than 25% of detected genes may be mitochondrial. SCTransform was employed for data normalization and correction, which identified 3,000 highly variable genes for subsequent analyses. Dimensionality reduction was performed using PCA, with the top 30 principal components selected for subsequent analysis. Harmony was utilized to correct batch effects. Cell clustering was performed with the “FindNeighbors” and “FindClusters” functions, with the resolution set to 0.8. Data visualization was conducted using uniform manifold approximation and projection (UMAP). Subsequently, data specific to macrophages were obtained through the “SubsetData” function within the “Seurat” package. The “FindClusters” function was then applied to these data with the resolution parameter set to 0.05 for cluster identification. The “FindAllMarkers” function from Seurat was utilized to detect genes that are expressed differently in each cluster, and subsequently, cluster annotation was performed according to the expression of classical markers. All cells were categorized into 9 major cell types, and macrophages were separately clustered and annotated to investigate their heterogeneity. Spatial transcriptomics data analysis To investigate the spatially specific immune microenvironment and macrophage polarization in RA, we conducted spatial transcriptomic analysis with the SDY2213 dataset from the immport database (https://www.immport.org). Two synovial tissue samples were analyzed with quality control results using SpaceRanger software. SCT was employed to convert, normalize, and select highly variable genes from the data. PCA using the top 30 components was applied for both dimensionality reduction and clustering. M1 macrophages were identified based on CD86 and CD74 markers, and M2 macrophages were identified based on CD163 and MSR1 markers. The “AddModuleScore” function was used to further quantify the spatial distribution pattern of M1 and M2 macrophages, and the polarization score was calculated. Cell–cell communication analysis In the present study, we used the R package “CellChat” [33]. Quantitative inference and analysis of intercellular communication networks can be achieved through this tool using scRNA-seq data. Ligand–receptor pairs with a significance level of p < 0.05 were chosen to assess the relationships among different cell types. Analysis of pseudotime trajectories and identification of dynamic genes The “Monocle2” package was applied to perform single-cell trajectory analysis [34]. All macrophages were extracted, and the “DDRTree” approach was employed to perform dimensionality reduction on the cells. Subsequently, a dimensionality reduction function was applied to ascertain the cell differentiation status. Finally, the cells in the state of S1 are the root. The results were visualized using the functions “plot_cell_trajectory,” “netVisual_aggregate,” and “netAnalysis_dot.” Complete Freund’s adjuvant (CFA)-induced RA models and treatments C57BL/6 mice (6–8 weeks old, both sexes) were obtained from Hunan Slack Jingda Laboratory Animal Co., Ltd. (Hunan, China; license SCXK (Xiang) 2019-0004; Quality Certificate No. 430727241101090254). We randomized the mice into three groups (n = 3) after 1 week of adaptive feeding: control, RA, and RA+sh-S100A9. To induce CFA-induced RA, mice were anesthetized with isoflurane inhalation followed by the administration of an injection of 20 μL of 5 mg/mL of CFA (MedChemExpress, USA) into the ankle joint. Mice that were administered an equal volume of saline using the same treatment regimen served as healthy controls. Five days prior to CFA injection, the ankle joints were intra-articularly injected with 9.28 × 1012 vector genomes (vg)/ml of AAV (adeno-associated virus)-S100A9 (GenePharma, Suzhou, China). Arthritis severity scores were determined at different time points after RA model construction. Ankle thickness was monitored using a finely calibrated caliper, and changes from baseline were recorded. The paws underwent examination, and the extent and site of the swelling were evaluated using a 5-point scale. A score of 0 indicated the absence of swelling or erythema. Scores of 1–4 corresponded to slight swelling and/or erythema, mild to moderate edema, substantial edema that restricted joint movement, and excessive edema with joint stiffness, respectively. The summed score per mouse served as the arthritis index, with a maximum score of eight [35]. Humane euthanasia was performed on mice after 27 days, with ankle collection for subsequent analysis. Histological analysis Following 48 h of fixation in 4% paraformaldehyde, ankle joints were decalcified in EDTA and paraffin-embedded. The specimens of the joints were cut into 5-μm-thick slices by employing a microtome. These slices were subjected to staining with both hematoxylin and eosin (H&E) and Safranin O/Fast Green. Each histological section was evaluated by two independent observers in a double-blind manner and scored separately for inflammation, synovial hyperplasia, and pannus formation. Inflammatory cell infiltration served as the indicator of inflammation. Synovial hyperplasia was scored based on synovial cell proliferation and thickening. Pannus formation was scored according to the distribution of small blood vessels in the synovium. Each parameter was graded on a scale from 0 to 3: 0 indicating none, 1 for mild, 2 representing moderate, and 3 denoting severe. The individual scores were combined to determine the overall pathological score. Immunofluorescence staining Following a 20-min fixation in 4% paraformaldehyde, tissue or cell slides underwent permeabilization using 0.2% Triton X-100 at room temperature. Nonspecific binding sites were blocked using 5% bovine serum albumin. Overnight incubation at 4 °C was performed using primary antibodies against S100A9, F4/80, Col2a1, Bcl-2, and iNOS (all sourced from Proteintech, Wuhan, China). Subsequently, Alexa Fluor 488- or 594-conjugated secondary antibodies (ZSGB-Bio, China) were applied. Nuclei were counterstained with DAPI (Beyotime) for 5 min. Images were captured via fluorescence microscopy and laser-scanning confocal microscopy (ZEISS, Germany). Cellular culture and transfection methods RAW264.7 cells were obtained from the Servicebio Cell Bank located in Wuhan, China. The cells were cultivated at 37 °C in a 5% CO₂ environment using standard DMEM (Servicebio, Wuhan, China), with the addition of 10% FBS (Servicebio, Wuhan, China) and 1% penicillin/streptomycin (Beyotime, Shanghai, China). Cells were divided into sh-NC, LPS+sh-NC, and LPS+sh-S100A9 groups. The S100A9 short hairpin RNA (shRNA) and lentiviral vector were designed and synthesized by Saisofi Biotech (Wuxi, China), and the cells were transfected according to the manufacturer’s instructions. To evaluate the function of S100A9 in the polarization of M1 macrophages, lipopolysaccharides (LPS) (MedChemExpress, United States) were employed at a concentration of 100 ng per mL to induce the polarization [36]. CD86 (Proteintech, Wuhan, China), an M1 macrophage surface marker, was evaluated by flow cytometry (BD Accuri C6 Plus). Western blot The grouping and intervention methods were consistent with those described in the previous step. Following the intervention, proteins were extracted from each group. To quantify the protein concentrations, the BCA assay was employed. Each sample was adjusted to contain 25 μg of protein for SDS-PAGE, followed by membrane transfer and blocking. Primary antibodies against iNOS, COX2, and β-actin (all from Proteintech, Wuhan, China) were applied, with overnight incubation at 4 °C. Following the incubation with the primary antibody, secondary antibodies were introduced and allowed to incubate for 1 h at room temperature prior to imaging. β-Actin served as the loading control to standardize protein expression levels. The intensities of the bands were measured using GelPro32 software (version 4.0). Quantitative real-time polymerase chain reaction (qRT-PCR) The grouping and intervention methods were consistent with those described in the previous step. Following the induction process, total RNA was isolated from cells via the Trizol approach. Subsequently, reverse transcription was carried out, and the cDNA was then amplified. The reverse transcription conditions were as follows: synthesis of single-stranded cDNA at 37 °C for 25 min; synthesis of double-stranded cDNA at 85 °C for 10 s; and storage at 4 °C. The qRT-PCR was run for 40 cycles with: 95 °C, 15 min pre-denaturation; 95 °C, 10 s denaturation; 60 °C, 30 s (annealing); 72 °C, 1 min extension; and a 72 °C, 2 min final extension. β-Actin was used as the internal reference gene for normalization of qRT-PCR data. The calculation of gene expression was performed utilizing the 2-ΔΔct method. A summary of the primer sequences and product lengths for qRT-PCR is provided in Table S2. Statistical analysis A paired or unpaired Student’s t-test (parametric) was utilized to evaluate the differences between two groups. When making comparisons across three or four groups, a one-way analysis of variance (which is parametric) was employed. Nonparametric data were analyzed using the Kruskal–Wallis test with multiple comparisons. Prism 10 software (GraphPad Software Inc.) was used to generate all graphs. Data are presented as the mean ± standard deviation (SD). Statistical significance was defined as p < 0.05. Results Identification of significant PCD regulators in RA Figure 1 presents a flowchart of the entire study. We first isolated genes in the test set associated with PCD and used them to construct a co-expression network. For sample clustering, we applied the Pearson correlation coefficient, with results presented in Fig. 2A. To meet the requirements of scale-free topology, we set the soft threshold value to 4, corresponding to an R2 of 0.8 (Fig. 2B). The MEDissThres parameter was set to 0.25, yielding a total of five modules (Fig. 2C). The analysis identified a “yellow” module containing 110 genes as the most clinically significant module in RA (Cor = 0.66, p = 2 × 10–28) (Fig. 2D). A strong correlation between GS and MM is demonstrated in Figure 2E for the “yellow” module. Study workflow depicting the comprehensive analysis of programmed cell death patterns in RA. S100A9 was identified as a substitute marker for the PCDR model, and its roles in promoting macrophage M1 polarization and chondrocyte apoptosis were demonstrated through in vivo and in vitro experiments Construction of the co-expression network. (A) By employing the average clustering method for the hierarchical classification of samples according to the Euclidean distance, a dendrogram representing the samples and a heatmap of features were created. (B) Top: soft-thresholding power; bottom: mean connectivity for the WGCNA network. (C) Genetic hierarchical tree clustering chart. Top panel: gene dendrogram; bottom panel: gene modules colored by module. (D) Module–trait associations: rows = modules, columns = traits; each cell reports the correlation coefficient and p-value. (E) Scatterplot of GS versus MM in the yellow module. (F) GO enrichment results were grouped into BP, CC, and MF; in each category, the 10 terms with the lowest p-values (most significant) were selected. (G) The top 30 significant KEGG signaling pathways GO enrichment analysis of 110 significant PCD regulators (Fig. 2F) showed significant enrichment in BP—including regulation of apoptotic, extrinsic apoptotic, and other signaling pathways—CC—such as secretory granule lumen, and vesicle lumen terms—and MF—including growth factor activity, cytokine activity, and transcription co-regulator binding. Furthermore, the KEGG enrichment analysis illustrated that these hub genes were mainly enriched in the “HIF-1,” “Cyto-CytoR Interaction,” “Ferroptosis,” “Rheumatoid arthritis,” and “IL-17” pathways (Fig. 2G). Construction of a diagnostic model for RA Overall, 113 prediction models were developed, and the AUC for each model was computed across all datasets, which included the test set as well as GSE12021, GSE55235, and GSE55457. In these models, we found that the combination of the two methods, Lasso and Naive Bayes, achieved optimal performance, with an AUC of 0.919, higher than those of the other models (Fig. 3A). This model, referred to as the PCDR model, included 12 significant PCD regulators: BRCA2, CRH, DDIT3, GATA4, PF4, S100A9, TFRC, NPC1, AGA, GNS, MANBA, and PRDX4. Of them, six genes (BRCA2, CRH, DDIT3, GATA4, PF4, and S100A9) were associated with apoptosis; four (NPC1, AGA, GNS, and MANBA) with lysosome-dependent cell death; three (DDIT3, GATA4, and NPC1) with autophagy; two (BRCA2 and PRDX4) with anoikis; one (TFRC) with ferroptosis; and one (DDIT3) with paraptosis. Compared with that in healthy individuals, the expression of BRCA2, PF4, S100A9, TFRC, NPC1, AGA, GNS, MANBA, and PRDX4 was upregulated in patients with RA, whereas the expression of CRH, DDIT3, and GATA4 was downregulated (Fig. 3B). Correlation analysis of the 12 PCD regulators revealed that BRCA2, AGA, PRDX4, S100A9, TFRC, NPC1, GNS, and MANBA exhibited strong positive correlations, whereas PF4 was negatively correlated with most other genes (Fig. 3C). These findings suggest that crosstalk among multiple forms of PCD could have a significant impact on RA. Construction and validation of the programmed cell death-related (PCDR) model. (A) Across four cohorts, the C-index was computed for 113 prediction models derived from 12 machine learning algorithms. (B) Boxplot showing differential expression of 12 significant PCD regulators between the control and RA groups. (C) Heatmap showing positive (red) and negative (blue) correlations among the 12 significant PCD regulators. (D) Nomogram based on 12 significant PCD regulators for RA risk prediction. (E–H) ROC curves evaluating diagnostic potential in the test set, GSE12021, GSE55235, and GSE55457. *p < 0.05, **p < 0.01, ***p < 0.001 A nomogram was constructed to predict RA progression and further evaluate the predictive capacity of the PCDR model (Fig. 3D). ROC analysis demonstrated the diagnostic performance of the model, with AUC values of 0.950 for the test set (Fig. 3E), 0.861 for GSE12021 (Fig. 3F), 0.980 for GSE55235 (Fig. 3G), and 0.885 for GSE55457 (Fig. 3H). High accuracy of the nomogram for RA diagnosis was validated through calibration curves (Figure S1A). DCA plot showed that the gray and black lines remained consistently below the red line across the range of 0–1, indicating that the nomogram model is a valuable decision-making tool for RA assessment (Figure S1B). Similarly, the clinical impact curve lent credence to the model’s predictive capacity (Figure S1C), and this trend was consistently observed in additional validation cohorts (Figure S1D–L). Collectively, these results highlight the excellent diagnostic utility of the 12 PCD regulators. Analysis of consensus clustering of 12 significant PCD regulators To investigate PCD patterns in RA, consensus clustering was conducted with the 12 significant PCD regulators. When k was set to 2, the differences between subgroups were most pronounced, indicating that RA samples could be grouped into two distinct subtypes (Clusters 1 and 2) (Fig. 4A). The xCell and QuanTISeq algorithms were employed to assess immune cell infiltration between the two clusters. Consensus clustering analysis of significant PCD regulators. (A) According to the PCD model, when k was defined as 2, individuals with RA were classified into two distinct molecular clusters. (B–C) Differences in immune-cell infiltration between Cluster 1 and Cluster 2 were revealed by xCell and QuanTIseq analyses. (D) The boxplot illustrates the differences in the expression of macrophage marker genes between the two clusters. *p < 0.05, **p < 0.01, ***p < 0.001 In the xCell analysis, Cluster 2 showed a marked increase in macrophages, M2 macrophages, and CD4+ Th2 T cells, compared with Cluster 1. In contrast, the levels of CD4+ memory T cells, neutrophils, endothelial cells, and hematopoietic stem cells were significantly reduced (Fig. 4B). In the QuanTISeq analysis, the abundance of M1 macrophages was notably higher in Cluster 2, whereas neutrophil and NK cell levels were lower, compared with those in Cluster 1 (Fig. 4C). Notably, both methods consistently demonstrated higher macrophage levels in Cluster 2 than in Cluster 1. Expression of macrophage marker genes (C1QB, RNASE1, LGMN, MAF, and SLCO2B1) was also higher in Cluster 2 than in Cluster 1 (Fig. 4D). This trend was validated in both internal and external cohorts, highlighting the pivotal role of macrophages in the RA immune microenvironment. To further investigate the differences in macrophage composition between the clusters, CIBERSORT analysis was performed. The results showed higher levels of M2 macrophages in Cluster 2 than in Cluster 1 (Figure S2A). Additionally, the 12 significant PCD regulators were strongly associated with immune cell infiltration. For example, S100A9 was significantly and positively correlated with M1 macrophages (Figure S2B). Differential expression analysis between the two PCD clusters revealed that PF4 and S100A9 were more highly expressed in Cluster 1, whereas CRH, DDIT3, AGA, GNS, and MANBA were relatively upregulated in Cluster 2 (Figure S2C). These PCD regulators may serve as genetic markers to distinguish between PCD-based RA subtypes. Identification of two molecular subtypes using the GSVA algorithm To confirm the PCD patterns, we employed clustering analysis utilizing the GSVA algorithm to compute PCD scores and categorize RA into different molecular subtypes based on 12 significant PCD regulator expression patterns. Similar to the above-mentioned results, two distinct clusters, Groups A and B, were identified (Fig. 5A). To quantify these patterns, PCA scores of the 12 significant PCD regulators for each sample in the different groups were calculated. The results demonstrated that Group A had higher scores than Group B (Fig. 5B). ssGSEA indicated that Group A exhibited notably higher immune cell infiltration than Group B (Fig. 5C). Group A also exhibited higher expression of macrophage marker genes (C1QB, RNASE1, LGMN, PLTP, and SLCO2B1) (Figure S3A–C). Additionally, the alluvial diagrams indicated that Cluster 1 was mainly connected to Group B, while Cluster 2 was largely associated with Group A (Fig. 5D). These findings validate the accuracy of PCD patterns and suggest that macrophages have a crucial function in RA. Clustering analysis characteristics of the GSVA algorithm. (A) Heatmap displaying the division of RA samples into two groups based on PCD scores in the test set, external validation set, and internal validation set. (B) PCD score differences between groups A and B. (C) A violin plot shows the ssGSEA scores between the two clusters. (D) Alluvial diagram showing the interrelationship among molecular clusters, GSVA groups, and PCD scores in RA. *p < 0.05, **p < 0.01, ***p < 0.001 Examination of RA microenvironment based on PCD using single-cell RNA-seq (scRNA-seq) Analysis of the scRNA-seq dataset revealed excellent integration across the four samples, with no significant batch effects, thereby enabling reliable downstream analysis (Fig. 6A). We divided all cells into 19 different groups, identified gene expression patterns within each group, and finally identified 9 major cell types (Fig. 6A): fibroblasts (DCN, COL1A2, and COL1A1), endothelial cells (PECAM1 and VWF), macrophages (C1QA, C1QB, and C1QC), T cells (PTPRC, CD3D, and CD3E), plasma cells (IRF4 and JCHAIN), NKTs (CCL5 and NKG7), DCs (HLA-DQA1 and HLA-DQB1), B cells (MS4A1 and CD79A), and pericytes (NOTCH3, RGS5, and TAGLN) (Fig. 6B). Figure 6C shows that macrophages had the highest PCDR scores among all cell types in the RA microenvironment. Using median PCD scores as the cutoff, we stratified cells into high- and low-scoring groups to evaluate cell type proportions. Figure 6D demonstrates that macrophage proportions were significantly elevated in the high-scoring group than in the low-scoring group. Notably, macrophages in the high-scoring group displayed higher outgoing interaction strength (Fig. 6E). Figure 6F demonstrates that macrophages influence the RA microenvironment through both autocrine and paracrine mechanisms. Key communication signals among different cell types are presented in Fig. 6G. Our analysis revealed that macrophages interact with other cell types primarily via CXCL, VISFATIN, SPP1, and VEGF signaling pathways. Overall, ligand–receptor interactions across various cell types in RA highlight the critical importance of macrophages in disease advancement. Clustering and cell–cell communication analysis of scRNA-seq data in GSE200815. (A) UMAP of cells, from left to right: colored by patients, clusters, and cell types. (B) Marker genes across nine cell types, displayed as violin plots. (C) Distribution of PCD scores among different cell types. (D) Proportions of cell types in low- and high-score groups. (E) Outgoing and incoming signaling strengths across cell types contrasted between high and low PCD score groups. (F) Interaction counts and weights among RA cell types. The color and breadth of the lines indicate the quantity of interacting pairs among cell types. (G) Key outgoing and incoming signaling patterns by cell type, shown as bubble plots. Dot size reflects the contribution fraction in cellular communication; higher values denote more abundant signaling pathways Identification and spatial distribution of dynamic PCDR genes during the transition of the macrophage state To investigate whether PCD signatures (PCDS) contribute to cell fate decisions in macrophages, trajectory analysis was performed. First, macrophages were divided into M1 and M2 subtypes using marker genes (Fig. 7A–B). The high proportion of M1 macrophage polarization during RA progression may be attributed to the shift from M2 to M1 phenotype (Fig. 7C). The expression of S100A9 in macrophages gradually increased over time, aligning with its distribution in M1 macrophages, whereas TFRC expression gradually decreased over time, indicating potential regulatory roles for these genes in macrophage differentiation pathways (Fig. 7D). S100A9 expression was significantly elevated in M1 than in M2 macrophages, whereas the opposite pattern was observed for TFRC (Fig. 7E). Pseudotime analysis and spatial distribution of macrophages and PCDR model genes. (A) UMAP showing the main subclusters of macrophages. (B) Expression of M1/M2 markers displayed as violin plots. (C) Pseudotime analysis of macrophage subtypes. (D) Dynamic expression of PCD model genes during macrophage differentiation. (E) Boxplot showing differential expression of S100A9 and TFRC between M1 and M2 macrophages. (F) Spatial expression maps of M1 and M2 macrophage scores and S100A9 and TFRC To investigate the relationship between spatial distribution and macrophage polarization, spatial transcriptomic data obtained from synovial tissues of patients with RA underwent comprehensive analysis. The findings indicated that M1 macrophages were extensively dispersed within the synovial tissues. The spatial distribution and expression pattern of S100A9 resembled that of M1 macrophages (Fig. 7F). Therefore, S100A9 may be a key regulator of macrophage cellular states during RA progression. S100A9 knockdown alleviated synovitis in RA by inhibiting M1 macrophage polarization To examine S100A9‘s regulatory function in M1 macrophage polarization throughout RA pathogenesis, mice with S100A9 knockdown were generated. This was achieved through intra-articular injection of adeno-associated virus serotype 9 (AAV9) expressing shRNA specific to S100A9. Arthritis scores were recorded every 3 days for 27 days (Fig. 8A). As expected, S100A9 knockdown effectively alleviated ankle swelling in RA mice (Fig. 8B, upper images). Histopathological evaluation demonstrated that S100A9 knockdown markedly reversed synovial hyperplasia and inflammatory cell infiltration (Fig. 8B). Both arthritis scores and paw thickness were significantly reduced (Fig. 8C). S100A9 knockdown inhibited M1 polarization and alleviated synovial inflammatory infiltration in CFA-treated mice. (A) Diagram illustrating the animal experiment. (B) Representative ankle joint images and H&E staining of synovial tissues. (C) Severity scores for arthritis were documented at various time points, with statistical evaluation performed on day 27 (N = 3). (D) Representative immunofluorescence images showing synovial S100A9 and F4/80 expression and their colocalization in control and RA mice (scale bar: 50 μm; N = 3). (E) In synovium from RA and RA+sh-S100A9 mice, representative immunofluorescence images show F4/80 and iNOS (scale bar 50 μm; N = 3). *p < 0.05, **p < 0.01, ***p < 0.001 To determine whether S100A9 is involved in RA synovitis, its expression was examined in synovial tissue and its colocalization with the macrophage marker F4/80. Synovial tissues from RA mice demonstrated significantly increased S100A9 expression relative to control mice, with predominant localization observed within accumulated macrophage populations (Fig. 8D). Moreover, S100A9 knockdown significantly reduced macrophage aggregation in synovial tissue of mice with RA, accompanied by a decrease in the M1 macrophage marker iNOS (Fig. 8E). These findings suggest that S100A9 knockdown alleviates synovial inflammation in RA by inhibiting M1 macrophage polarization. S100A9 knockdown alleviated cartilage destruction and bone erosion in RA by inhibiting apoptosis Histopathological evaluation showed that S100A9 knockdown effectively mitigated ankle cartilage injury and preserved cartilage thickness as well as collagen amounts in mice with RA (Fig. 9A). Immunofluorescence examination demonstrated significant reductions in both Col2a1 and anti-apoptotic factor Bcl-2 expression within the articular cartilage of RA mice compared with control mice. Following S100A9 knockdown, Col2a1 and Bcl-2 expression increased (Fig. 9B). In addition, colocalization analysis showed that S100A9 and Bcl-2 were co-localized in articular cartilage. After S100A9 knockdown, Bcl-2 expression was significantly upregulated in the articular cartilage of RA mice (Figure S4). These results suggest that S100A9 knockdown alleviates cartilage destruction and bone erosion during RA progression by inhibiting chondrocyte apoptosis. S100A9 knockdown attenuated CFA-induced mouse cartilage destruction and bone erosion. (A) Representative images of cartilage stained with H&E and Safranin O/Fast Green (N = 3). (B) Immunostaining of COL2A1 and BCL-2 in cartilage (scale bar 50 μm; N = 3) S100A9 knockdown inhibited LPS-induced M1 macrophage polarization To investigate how S100A9 influences the polarization of M1 macrophages, additional experiments utilizing LPS-stimulated RAW264.7 cells were performed. Stimulation of RAW264.7 cells with LPS notably increased M1 macrophage-related inflammatory gene expression, including IL-18, iNOS, IL-1β, and TNF-α, whereas S100A9 knockdown effectively reduced their expression (Fig. 10A). The reduction of S100A9 levels significantly diminished M1 macrophage-related pro-inflammatory protein levels, including iNOS and COX2 (Fig. 10B). An analysis using immunofluorescence was subsequently conducted to evaluate the expression levels and colocalization of iNOS and S100A9 across various groups. As expected, iNOS and S100A9 were highly expressed and co-localized in the cytoplasm of LPS-induced M1 macrophages, and S100A9 knockdown substantially reduced iNOS and S100A9 expression (Fig. 10C). Furthermore, analysis using flow cytometry revealed that the knockdown of S100A9 led to a notable reduction in the percentage of cells that were positive for CD86, suggesting a lower expression of markers associated with M1-type macrophages (Fig. 10D). Overall, these results indicate that S100A9 is essential in reducing the severity of RA and serves as a significant modulator of M1 macrophage polarization. S100A9 inhibited M1 macrophage polarization in RAW264.7 cells. (A) M1 macrophage marker transcripts were quantified by qPCR in cells exposed to LPS for 24 h, with or without sh-S100A9. (B) Quantitative assessment and western blot analysis were conducted examining iNOS and COX2 proteins in cells that were exposed to LPS for 48 h, with or without the presence of sh-S100A9. (C) Representative immunofluorescence images showing iNOS and S100A9 expression in cells treated with LPS with or without sh-S100A9 for 48 h. (D) Flow cytometry and quantitative analysis of CD86+ cells after treatment with LPS with or without S100A9 for 48 h Discussion Under normal physiological conditions, PCD maintains cellular homeostasis within organisms, ensuring stability of the intracellular microenvironment and promoting overall organismal balance. However, this equilibrium is disturbed during the pathological development of RA. Multiple PCD pathways may influence the onset and development of RA by promoting chronic synovial inflammation, joint erosion, and angiogenic imbalance through complex interaction mechanisms [37]. While most studies have focused on a single type of PCD, the mutual relationships and functions of multiple PCD types in RA remain poorly understood. In this study, we revealed global alterations in 18 types of PCD at the transcriptional and genetic levels, along with their intercorrelations in RA. We also developed a PCDR model to predict RA, which was validated across multiple cohorts. Disruption of synovial macrophage homeostasis occurs prior to RA clinical manifestation [38], suggesting that macrophages are essential in disease pathogenesis. These cells can differentiate into various phenotypes to orchestrate immune and inflammatory responses, as well as tissue repair [9]. Synovial macrophages are strongly associated with synovial inflammation, clinical severity/outcomes, and therapeutic responsiveness, highlighting the need to further define macrophage heterogeneity to develop new RA treatments [39]. In this study, based on the 12 significant PCD regulators identified in the previous analysis, patients with RA were classified based on 2 distinct PCD patterns via consensus clustering analysis and GSVA. Immune cell infiltration analysis demonstrated a clear difference between groups with low PCD scores and groups with high scores. In the high-scoring group, macrophages demonstrated increased infiltration, with elevated expression levels of macrophage-specific marker genes. Meanwhile, macrophages exhibited the highest PCD scores in the RA microenvironment. In the high-scoring cell population, the proportion of macrophages increased, and they exhibited a high outgoing interaction strength. These findings align with previous reports, further confirming the pivotal role of macrophages in RA. In the pseudo-time analysis, we found that S100A9 expression gradually increased over time during M1 macrophage development and differentiation, showing a positive correlation. In addition, spatial transcriptome analysis revealed that M1 macrophages were widely distributed in the synovial tissues of patients with RA, and the spatial location and expression level of S100A9 resembled the distribution pattern of M1 macrophages. Therefore, S100A9 may be a key target in regulating the cellular state of M1 macrophages during RA progression. In vivo study revealed a marked increase in S100A9 levels within the synovial tissue of RA mice, primarily found in macrophages. S100A9 knockdown significantly alleviated ankle swelling and inflammation in RA mice. Furthermore, silencing S100A9 markedly reduced macrophage clustering and the proportion of M1 macrophages within synoviocytes from RA mice. In addition, S100A9 knockdown increased Col2a1 and Bcl-2 expression in articular cartilage, accompanied by attenuation of cartilage destruction and bone erosion. Finally, S100A9 knockdown downregulated M1 macrophage-associated inflammatory factors (iNOS, COX2, IL-18, IL-1β, and TNF-α) and inhibited LPS-induced M1 macrophage polarization in vitro. Overall, these results indicate that S100A9 may be crucial in delaying RA progression and that this mechanism is closely linked to M1 macrophage polarization. One significant drawback of this research is the lack of sample data sourced from various cohorts to confirm the model’s accuracy, potentially restricting the comprehensiveness and generalizability of the results. Moreover, small validation cohort sizes may inflate performance (AUC) metrics. Additionally, the scRNA-seq analysis included only four RA patients, which may limit the characterization of macrophage heterogeneity and PCD-related cell states. Despite extensive bioinformatics analyses and preliminary experimental validation, the precise mechanism of S100A9 requires further investigation. Future studies should include larger patient populations and expanded experimental cohorts, and the results should be confirmed in multiple independent cohorts to ensure that the conclusions are robust and clinically relevant. The potential of targeting S100A9 for RA treatment warrants comprehensive functional studies and clinical trials. Conclusion In conclusion, this study developed a PCDR model for RA using machine learning algorithms and identified distinct PCD-related molecular patterns in RA, with macrophages confirmed as key effector cells in RA pathogenesis. Moreover, inhibition of the key gene S100A9 significantly alleviated RA progression by reducing M1 macrophage polarization, synovitis, and cartilage damage. However, further validation through expanded functional studies and large-scale clinical trials is necessary to establish the broader clinical applicability of targeting S100A9 for RA therapy. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Acknowledgements The authors sincerely thank the publicly available databases that provided data for this study. Abbreviations Adeno-associated virus Area under the receiver operating characteristic curve Biological processes Cellular components Decision curve analysis Dendritic cells False discovery rate Fibroblast-like synoviocytes Gene Ontology Gene significance Gene set variation analysis Interleukin-17 Kyoto Encyclopedia of Genes and Genomes Lipopolysaccharides Molecular functions Module membership Matrix metalloproteinases Natural Killer T cells Principal component analysis Programmed cell death Quantitative real-time polymerase chain reaction Receiver operating characteristic Standard deviation Sodium dodecyl sulfate-polyacrylamide gel electrophoresis Short hairpin RNA Single-sample gene-set enrichment analysis Tumor necrosis factor Uniform manifold approximation and projection Weighted gene co-expression network analysis Author contributions All authors contributed to the study conception and design. Qingyuan Xu: Writing–original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation. Jinfu Liu: Writing–original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Funding acquisition. Qiang Ding: Writing–original draft, Validation, Methodology, Investigation, Formal analysis, Data curation. Canbin Zhao: Writing–original draft, Validation, Methodology, Investigation, Formal analysis, Data curation. Weiwei Wang: Methodology, Investigation. Hao Li: Investigation, Formal analysis. Chicheng Niu: Investigation, Formal analysis. Wei Chen: Visualization, Validation. Ping Zeng: Writing–review & editing, Supervision, Project administration, Funding acquisition. Donghui Guan: Writing–review & editing, Supervision, Project administration, Funding acquisition. Ronghua Zhang: Writing–review & editing, Supervision, Project administration, Methodology, Funding acquisition. Funding This study was supported by the Guangdong Provincial Key Laboratory of Traditional Chinese Medicine Informatization (2021B1212040007); The Joint Funds of the National Natural Science Foundation of China (U24A6013); Shandong Natural Science Foundation General Project (ZR2022MH147); Natural Science Foundation of Guangxi Province of China (2025GXNSFBA069262); National Key Research and Development Program of China (2018YFC2002500); The National Natural Science Foundation of China (82560944). Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate All the study designs and test procedures were performed in accordance with the Helsinki Declaration. The study and experimental protocols were approved by the Institutional Animal Ethics Committee of Guangxi University of Traditional Chinese Medicine, following the Guide for Care and Use of Laboratory Animals (No. DW20240919-241). Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Qingyuan Xu, Jinfu Liu, Qiang Ding, and Canbin Zhao have contributed equally to this work. Contributor Information Ping Zeng, Email: zengp@gxtcmu.edu.cn. Donghui Guan, Email: guanyisheng0720@163.com. Ronghua Zhang, Email: tzrh@jnu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Supplementary Material 4 Data Availability Statement No datasets were generated or analysed during the current study.