Cross-platform transcriptomic data integration identifies an overactive neuro-immune signature in human osteoarthritis synovium

Abstract

Objective:

Chronic synovitis is associated with osteoarthritis (OA) pain, but the molecular underpinnings remain unclear. Our objective was to characterize the transcriptional phenotype of OA synovium with a focus on signaling relevant to pain.

Design:

Eight publicly-available microarray and RNA-sequencing GEO datasets from human non-OA and OA subjects underwent quality control and re-analysis for differentially-expressed genes (DEGs). Cross-platform statistical integration was performed via a weighted Z-test to combine detection power across datasets. Gene set enrichment, cell type enrichment, and regulon analyses were performed. Human single-cell RNA sequencing data was used to map gene expression to cell types. Ligand-receptor interactions were predicted via multi-omic data from human dorsal root ganglia (DRG).

Results:

Following statistical integration of individual studies (N=153 subjects), gene set enrichment analysis identified 276 differentially-activated pathway terms in OA synovium, including strong enrichment for pathways related to innate and adaptive immunity (notably MHC Class II) and fibrotic remodeling-relevant extracellular matrix organization. VEGF signaling and angiogenesis-related terms were downregulated. Enriched pain and neuronal pathways primarily related to neuro-immune interactions, including neuroinflammation, and were associated with macrophages, B and T lymphocytes, and synovial fibroblasts. A gene regulatory network comprised of STAT1, FLI1, and VDR putatively governed the expression of 27 genes driving neuro-immune signaling. An unbiased synovium-DRG interactome predicted 76 potential interactions between synovial cells and DRG nociceptors, involving 68 neuronal receptors interacting with 32 ligands overexpressed in OA synovium.

Conclusions:

End-stage OA synovium is markedly enriched for neuroinflammatory and neuro-immune signaling, putatively governed by STAT1, FLI1, and VDR.

Introduction

Synovitis is an integral component of osteoarthritis (OA) pathogenesis and is recognized as both a downstream consequence of cartilage and bone damage as well as a driver of multiple disease processes, notably cartilage breakdown and pain¹. Clinically, MRI-based effusion-synovitis scoring has been predictive of future cartilage loss², synovitis severity has been associated with overall greater risk of incident OA³, and multiple studies have demonstrated strong associations between imaging-based synovitis and pain^4–7. Despite this evidence linking synovitis to OA symptoms and disease progression, OA treatments targeting inflammatory mediators have been unsuccessful thusfar^8–10, indicating that the current understanding of the nature of synovitis in OA is insufficient for effective therapeutic development.

Synovitis in OA manifests as lining layer hyperplasia, increased immune cell content, fibrosis, increased blood vessel density and vascular damage, and increased nociceptor density and sprouting^{1, 11}. The understanding of various synovitis “endotypes” in OA, and arthritis as a whole, is rapidly evolving, and recent evidence supports the existence of arthritis subsets with varying degrees of inflammation, structural damage, and extracellular matrix (ECM) and vascular changes^{12, 13}. Philpott and colleagues recently showed that histologically-characterized synovial tissue damage score (e.g. lining erosion, fibrosis, vascular damage, and perivascular edema) correlated with patient-reported pain, whereas histological synovial tissue inflammation score (e.g. synovial lining thickness, sub-synovial infiltrate, vascularization, presence of fibrin) did not¹¹, indicating a high degree of arthritic synovial disease heterogeneity marked by varying patterns of symptoms, imaging findings, and tissue-level phenotypes.

Transcriptomic studies have significantly expanded knowledge of the genes and pathways underlying cell- and tissue-level manifestations of OA^{14, 15}. Synovial gene signatures indicative of higher fibrosis, myeloid cell content/cytokine expression, senescence, and blood vessel content have been identified^{14, 15}; however, these studies also demonstrate the heterogeneous nature of the synovial OA transcriptome arising from disease heterogeneity, biopsy inconsistencies, technical/analytical differences, and other factors.

To gain a greater understanding of the human OA synovial transcriptome, meta-analysis presents an opportunity to leverage the combined statistical power of several transcriptomic datasets. Utilizing specialized statistical approaches to integrate data across studies, OA-associated gene and pathway enrichment may be detected with greater statistical confidence and detection power than any individual study, potentially enabling the discovery of previously unidentified genes or pathways. By employing a larger, more diverse subject pool across multiple sites, meta-analyses may also generate a more clinically-representative, generalizable understanding of disease states. To date, no study has performed a cross-platform transcriptomic meta-analysis of OA synovium. Two prior studies reanalyzed microarray datasets in OA synovium: Liu et al. analyzed five microarray datasets encompassing 41 normal and 45 OA synovial tissue samples¹⁶, and seven genes (SLC2A3, CDKN1A, FOXB, TAC1, STMN2, SCRG1, NELL1) were found to be differentially-expressed across datasets. However, this study did not perform formal statistical integration, focusing only on overlapping genes across studies, which limits detection power. Xu et al. merged two microarray datasets encompassing 20 non-OA and 26 OA synovium samples, identifying 46 DEGs¹⁷, but this data represents only a fraction of the more extensive RNA-sequencing data now available from non-OA and OA synovium.

Limited information currently exists regarding the neurotrophic and proalgesic mediators overexpressed in OA synovium. Neuro-immune interactions (immune cell-derived ligands interacting with receptors expressed by nociceptors) have received increasing attention in this context^{18, 19}, and synovial fibroblasts are also emerging as important mediators of proalgesic signaling in arthritis^{20, 21}, but this has been minimally explored in the context of OA synovium. Given the strong clinical association between OA pain and synovitis, this represents an opportunity to leverage existing transcriptomics datasets to identify pain-relevant transcriptional signatures, providing a robust assessment of overall OA pathogenesis while also identifying potential targets for pain therapeutics. To this end, we conducted a cross-platform transcriptomic meta-analysis to generate a reference atlas for OA synovium and applied this dataset to unbiasedly identify OA-associated genes and pathways involved in neuro-immune and pain-relevant pathways and synovium-neuron interactions.

Methods

Database query and study inclusion criteria

Expanded methodological details are provided in our Supplemental Methods. A comprehensive review of the Gene Expression Omnibus (GEO) database was undertaken to identify transcriptomic datasets from human synovial tissue of both non-OA controls and OA subjects (Suppl.-Methods-1). In total, 8 datasets were identified for inclusion – 2 bulk RNAseq and 6 microarray datasets^22–28 (Table- 1). Non-OA synovial samples were obtained post-mortem from tissue donors involved in fatal accidents in 2/8 datasets, patients without history of OA undergoing knee surgery in 1/8, and a combination of these sources in 2/8. One study (GSE143514) only specified that synovial tissues were taken from patients without diagnosed joint disease. In 7/8 datasets, OA synovial biopsies were taken at the time of joint replacement. In one study (GSE89408), OA and non-OA samples were taken during arthroscopic examination for knee pain, with the non-OA group defined as those without previous history of OA and no evidence of cartilage damage or synovitis upon examination.

Table 1.

Summary of datasets included in integrative analysis.

Technology	GEO #	First Author	Pub. Year	Collection Site	N	n(N-OA)	n(OA)	Platform	Biopsy Location
Microarray	GSE1919	Ungethuem	2010	Berlin, Germany	10	5	5	Affymetrix HG-U95A	Unspecified
	GSE12021	Huber	2008	Hannover, Germany	14	4	10	Affymetrix HG-U133B	Unspecified
	GSE55235	Woetzel	2014	Berlin, Germany	20	10	10	Affymetrix HG-U133A	Knee
	GSE55457	Woetzel	2014	Jena, Germany	19	10	9	Affymetrix HG-U133A	Knee
	GSE82107	Broeren	2016	Nijmegen, Netherlands	14	5	9	Affymetrix HG-U133 Plus 2.0	Unspecified
	GSE206848	Bay-Jensen	2023	New York, NY, USA	13	7	6	Affymetrix HG-U133 Plus 2.0	Knee
Bulk RNAseq	GSE89408	Guo	2017	Spring House, PA, USA	41	22	19	Illumina HiSeq 2000	Knee
	GSE143514	Zhao	2021	Deyang, China	8	3	5	HiSeq X Ten	Knee

Individual dataset preprocessing and analysis

Bulk RNAseq and microarray datasets were preprocessed via platform-specific pipelines (Suppl.-Methods-2–3). Briefly, bulk RNAseq samples underwent sequencing quality control, trimming and alignment, and count matrix generation. Microarray data underwent normalization, scaling, probe annotation and summarization. Non-expressed genes, non-protein-coding genes, and sex-linked (XIST and Y-chromosome) genes were removed prior to downstream analyses. Samples were removed from analysis for poor sequencing quality (bulk RNAseq only, total alignment % < 70%) and significant contamination with muscle tissue, determined via muscle enrichment scoring (Suppl.-Methods-4).

Differential expression analysis was performed within each dataset using weighted Limma-Voom²⁹ to compare OA and non-OA subjects (Suppl.-Methods-5). Raw P-values were adjusted via the Benjamini–Hochberg correction for multiple comparisons. Significantly differentially-expressed genes (DEGs) were defined at P_adj < 0.05.

Integrative analysis and neuro-immune profiling

DE results from individual studies were statistically integrated using a two-tailed weighted Z-test (Suppl.-Methods-6), a meta-analytical P-value combination method which statistically combines detection power across multiple datasets, rather than simply searching for overlaps in findings. Because P-value combination integrates at the level of results rather than the level of subjects / expression data, this approach is amenable to synthesizing data across technologies (e.g. microarray and bulk RNAseq)³⁰. The weighted Z-test is relatively strict against incomplete association compared to other approaches (e.g. weighted Fisher, ordmeta, Lancaster tests), prioritizing findings broadly-conserved between studies though still allowing for some degree of incomplete association³¹. Importantly, the two-tailed weighted Z-test accepts both the P-value and the direction-of-effect as inputs, and expression changes in opposing directions negate one another. This approach prioritizes consistent, generalizable findings. Samples were weighted by $\sqrt{N}$, giving greater weight to larger studies with higher detection power without mitigating the contribution of smaller studies – simulations have demonstrated this produces an idealized P-value for the weighted Z-test³².

Pathway and cell-type enrichment analyses were conducted on the integrated dataset using gene set enrichment analysis (GSEA)^{33, 34}, employing the Gene Ontology (GO): Biological Processes annotation. Following unbiased analyses, significant GSEA terms were screened to extract terms related to the nervous system (Suppl.-Methods-7). Leading-edge genes from significantly upregulated neuro-immune pathways were submitted to transcription factor-binding motif analysis using RcisTarget³⁵ to identify transcription factor which putatively regulate neuro-immune-related genes (Suppl.-Methods-8). Corroborative pathway and TRUUST transcription factor binding motif analyses were conducted using Metascape³⁶ (Suppl.-Methods-9), providing a complementary, statistically-independent meta-analytical technique. Cell-type-specific gene expression was assessed with an integrated single-cell RNAseq dataset using publicly-available data from human OA synovium^{21, 37–40} (Suppl.-Methods-10).

DRG interactome analysis

To unbiasedly assess potential crosstalk mechanisms between synovium and the nervous system, we utilized a human dorsal root ganglia (DRG) ligand-receptor interactome annotation⁴¹ to derive a synovium-DRG interactome (Suppl.-Methods-11). This database was screened to extract all interactions between secreted, non-matrix constituent, OA-enriched ligands and extracellular membrane-bound DRG receptors. Potential receptors were screened for expression by human DRG neurons utilizing integrated transcript counts from Visium spatial transcriptomics ⁴² and single-nucleus RNA-sequencing data⁴³. Using the neuronal annotation described in Tavares-Ferreira et al.⁴², receptors with log-normalized read counts > 0.01 in any C-fiber neuron or Aδ high-threshold mechanoreceptor (HTMR) were defined as potential binding partners to synovial-derived factors (we excluded proprioceptors and Aβ and Aδ low-threshold mechanoreceptors (LTMRs)). The resultant interactome was manually curated into categories. Cell-type-specific expression of synovial ligands was assessed as described above.

Results

Study descriptives

After identifying 70 potential datasets, eight met our inclusion/exclusion criteria – two employed bulk RNA-seq and six employed gene microarrays, encompassing n=153 samples (n=74 non-OA vs n=79 OA) (Suppl.-Fig.-S1, Suppl.-Data). Notably, one large RNAseq dataset²⁷ was derived from a RA study in which non-OA and OA synovium were used as controls, but not directly compared; thus, our analysis represents the first publication of the OA-focused assessment of this valuable sample set. We performed rigorous quality control, excluding seven samples for sequencing quality control-related reasons and seven samples for skeletal muscle contamination (Suppl.-Methods-4, Suppl.-Fig.-1). This yielded a final dataset of n=139 patient samples (n=66 non-OA vs n=73 OA, Table- 1), 49 of which were derived from RNAseq (n=25 non-OA vs n=24 OA) and 90 from gene microarray (n=41 non-OA vs n=49 OA) (Fig.- 1A).

Figure 1.

A). Descriptive statistics of final datasets included in the meta-analysis. B). Summary of differentially-expressed genes (DEGs) identified in each dataset independently, following quality control and re-analysis. DEGs were defined as P_adj < 0.05 and |log2-foldchange| > 0.585. Only protein-coding, non-sex-linked genes were included. C). Matrix of Pearson correlation coefficients between studies, calculated from log2-foldchanges of each study’s non-OA vs OA comparison across all genes. D). Number of overlapping DEGs identified in 2 or more studies. E). Bubble plot of overlapping DEGs identified in 5 and 6 studies, showing differential gene expression between respective non-OA and OA groups across all included studies. Positive log2-foldchanges indicate upregulation in OA relative to non-OA.

Unsupervised clustering of each study demonstrated substantial heterogeneity both between and within conditions, with separation between OA and non-OA subjects ranging from high to weak within different datasets (Suppl.-Fig-S2.). Per-study differential expression analyses further demonstrated this heterogeneity (Suppl.-Data), with the number of significant OA-associated DEGs ranging from 884 to 7,165 DEGs (P_adj < 0.05) (Fig.- 1B), and relatively low correlations of overall gene fold-change patterns between datasets (Fig.- 1C).

Assessing overlap in DEG lists between all studies, we observed 1,357 genes shared between 2 studies down to just 2 genes shared between 6 studies (Fig.- 1D). No genes were shared across 7 or 8 studies, demonstrating the limitations of only focusing on overlapping DEGs in this context. Upregulated DEGs shared across 5 studies included SCRG1 (Stimulator of Chondrogenesis-1), GLT8D2 (Glycosyltransferase-8 Domain Containing 2), PRDX4 (Peroxiredoxin-4), AGMAT (Agmatinase), and B3GALNT1 (Beta-1,3-Galactosyltransferase-1) (Fig.- 1E), and downregulated DEGs shared across 5 studies included MAFF (MAF BZIP Transcription Factor-F), SPRY1 (Sprouty RTK Signaling Antagonist-1), DUSP5 (Dual Specificity Phosphatase-5), and SOCS3 (Suppressor of Cytokine Signaling-3) (Fig.- 1E). SCRG1 and MAFF were the only 2 DEGs shared across 6 studies (Fig.- 1D).

Cross-platform Integrative Analysis

To address the observed study-by-study heterogeneity in differential expression data (Fig.- 1B–D) and yield more generalizable results, we conducted a statistical meta-analysis to integrate differential expression results across datasets. This generated a direction-of-effect (i.e. upregulated/downregulated in OA) and a composite P-value based on a given gene’s direction-of-effect and P-value across all studies, weighted by sample size. Following multiple comparison adjustment, this resulted in 4,393 significant DEGs (P_adj < 0.05) detected between non-OA and OA synovium (Fig.- 2A, Suppl.-Data) – this total falls between the highest (7,165) and lowest (884) DEG counts within individual studies, demonstrating that data integration increased detection power while penalizing contradictory findings. Top upregulated DEGs included SCRG1 (stimulator of chondrogenesis 1), GLT8D2 (glycosyltransferase 8 domain containing 2), and immune-relevant genes LY86 (Lymphocyte Antigen 86), HLA-DMB (Major Histocompatibility Complex, Class II, DM Beta), and HNMT (Histamine N-Methyltransferase) (Fig.- 2B). Top downregulated DEGs included DDX54 (DEAD-Box Helicase 54), DDIT4 (DNA Damage Inducible Transcript 4), and angiogenesis-related genes MAFF (MAF BZIP Transcription Factor F), VEGFA (Vascular Endothelial Growth Factor A), and BCL6 (BCL6 Transcription Repressor) (Fig.- 2B).

Figure 2.

A). Number of differentially-expressed genes at different FDR cutoffs following weighted Z-test statistical integration of differential expression results from all eight datasets from Fig. 1B. B). Top up/down-regulated genes in OA compared to non-OA. Positive log2-foldchanges indicate upregulation in OA relative to non-OA. C). Cell type-specific expression patterns of the top DEGs, based on an integrated scRNAseq atlas of non-OA and OA synovium. D). Top differentially enriched pathways in OA relative to non-OA, based on GSEA. Positive normalized enrichment score (NES) indicates enriched in OA relative to non-OA. E). Top differentially enriched pathways in OA relative to non-OA, based on Metascape analysis, with OA-enriched pathways at the top in red and OA-de-enriched pathways in the bottom in blue. F). GSEA-based cell type enrichment analysis between OA and non-OA. Positive normalized enrichment score (NES) indicates enriched in OA relative to non-OA.

We generated a scRNAseq atlas of non-OA and OA human synovium and mapped expression of the top upregulated genes to synovial fibroblasts (SCRG1, GLT8D2, HNMT), macrophages (LY86, HLA-DMB, HNMT), B cells (LY86, HLA-DMB), and pericytes (HNMT) (Fig.- 2C). The top downregulated genes primarily mapped to pericytes, fibroblasts, and myeloid cells (DDX54, MAFF, DDIT4, VEGFA, BCL6), as well as endothelial cells (DDX54, MAFF, DDIT4), T cells (DDIT4) and B cells (DDX54) (Fig.- 2C).

Pathway analysis using GSEA revealed 276 differentially-enriched pathway terms, after computationally condensing similar terms to reduce redundancy (Suppl.-Data). The top 5 OA-enriched terms were related to inflammation and innate and adaptive immunity (antigen processing and presentation via MHC class II, leukocyte mediated cytotoxicity, positive regulation of phagocytosis), while de-enriched terms were related to vascular pathways (regulation of endothelial differentiation) and regulation of MAPK signaling (Fig.- 2D). To validate these data using a complementary approach, Metascape functional enrichment analysis was performed on the up- and down-regulated DEG lists. As in GSEA, top OA-enriched pathways included innate and adaptive immunity and cytokine signaling, while downregulated terms included VEGF signaling and vasculature development (Fig.- 2E). Metascape also implicated OA-enriched extracellular matrix organization and de-enriched adipogenesis (Fig.- 2E), consistent with synovial fibrosis observed in end-stage OA^{1, 11, 44}.

We employed cell-type enrichment analysis to infer differential abundance of cell types underpinning this signature of inflammation and immune activation⁴⁵. We observed enrichment of gene markers for dendritic cells, macrophages, monocytes, neutrophils, B cells, T cells, mast cells, stromal cells (i.e. fibroblasts), eosinophils, platelets, osteoclasts, and NK cells (Fig.- 2F). Adipocytes markers were de-enriched, corroborating clinical and preclinical studies demonstrating fibrosis of adipose tissue in OA synovium^{44, 46, 47}. Taken together, these findings indicate diverse phenotypic manifestations in OA synovium, with the primary feature conserved across datasets being a molecular signature of inflammation marked by innate and adaptive immune activation (particularly pathways related to MHC class-II antigen processing), alongside predicted enrichment of macrophages, lymphocytes, mast cells, and stromal cells.

Neuro-immune Profiling

Given the strong evidence linking synovitis to pain and OA severity¹, defining the molecular signatures relevant to pain is a critical step towards improving clinical management of OA. We performed a targeted analysis of genes/pathways related to nervous system development, neuro-immune, and neurotrophic signaling. Nervous system-related signaling accounted for 13 of all 696 (un-simplified) OA-enriched GSEA terms (Fig.- 3A), with enriched pathways relating to neuro-immune interactions (including “neuroinflammatory response”) and synaptic vesicle maturation (Fig.- 3A). No significant pathways were found related to nerve sprouting. De-enriched neuronal pathways were related to proliferation, differentiation, and development (Fig.- 3A). Given the likely advanced disease status of these tissues, this suggests any potential nerve sprouting in the synovium may occur earlier during disease progression^{48, 49}. The four significant neuro-immune signaling terms shared 27 leading-edge genes – all were significantly upregulated in OA compared to non-OA (Fig.- 3A). This gene set spans immunity and inflammation (TYROBP, AIF1, CX3CR1, C1QA, C1QB, C1QCM ITGAM, TREM2, TLR3, IL18, PTPRC, CD200R1), neuroinflammation and neurodegeneration (PSEN1, GRN, SNCA, CLU, TREM2), signal transduction (MMP3, ADCY1, PLCG2, ATM, SYT11), synaptic function (PLXNC1, IGF1, ADGRB3) and cell adhesion (ITGB2, CTSC, BPGM).

Figure 3.

A). Left: Differentially enriched neuronal-related pathways in OA relative to non-OA, based on GSEA. Positive normalized enrichment score (NES) indicates enriched in OA relative to non-OA. Right: heatmap of composite FDR following statistical integration of the 27 leading edge genes driving the neuro-immune/inflammatory pathways marked by arrows. Positive/red values indicate upregulated in OA relative to non-OA. B). Left: Rcis Target-based (top) and TRUUST-based (bottom) regulon analysis of the 27 leading edge genes derived in (A). Positive normalized enrichment or enrichment score indicates overactive transcription factor in OA relative to non-OA. Right: Differential gene expression of the three identified transcription factors by individual study and by weighted Z following statistical integration, in OA relative to non-OA. C). Putative gene regulatory network involving STAT1, VDR, and FL1. D). Left: Cell type-specific expression, based on scRNAseq, of gene modules of all 27 neuro-immune leading edge genes and all 3 predicted transcription factors. Right: Percentages of macrophages that expressed the predicted neuro-immune-related transcription factors, based on scRNAseq.

We next sought to infer potential upstream regulators of the neuro-immune pathways we identified as overactive in OA synovium. We performed transcription factor binding motif analysis of the 27 neuro-immune leading-edge genes using RcisTarget, analyzing shared binding motifs in the genes’ upstream promoter regions to identify significantly enriched motifs³⁵. This predicted 3 transcription factors – STAT1, FLI1, and VDR – which had significantly enriched binding motifs associated with the neuro-immune gene module and were also significantly upregulated in OA compared to non-OA (Fig.- 3B), suggesting these factors putatively govern the expression of the 27 upregulated neuro-immune genes (Fig.- 3B). Metascape TRUUST analysis, an untargeted analysis performed on significant DEG lists across all studies, independently corroborated all three transcription factors as significantly OA-enriched (Fig.- 3B). Network analysis of the 3 transcription factors and neuro-immune leading-edge genes revealed a highly interconnected network encompassing 17 of the 27 neuro-immune genes (Fig.- 3C). STAT1 regulated the majority (10/17) of these genes (Fig.- 3C). Macrophages were the dominant cell type expressing the neuro-immune leading-edge genes and their putative transcription factors, with lesser expression also observed in B and T cells, fibroblasts, and endothelial cells (Fig.- 3D). Taken together, these results suggest putative mediation of neuro-immune/neuroinflammatory transcriptional signature of human OA synovium by macrophages via the transcription factors STAT1, FLI1, and VDR.

Synovium-DRG Interactome

DRG neurons express a wide array of receptors, yielding a large potential interactome underpinning nociception⁴¹. While the targeted neuro-immune profiling in Figure 3 focused on genes identified from neuro-immune/neuro-inflammatory pathways significantly enriched in OA synovium, this approach only considers genes already annotated to neuro-immune/neuroinflammatory or nociception-relevant pathways. Given the incomplete knowledge about nociception-relevant mechanisms in OA, we sought to identify all potential synovial-derived factors overexpressed in OA synovium that may interact directly with a nociceptor-bound receptor. We compiled a list of possible neuronal receptors with expression data from an integrated atlas of human DRG neurons, based on spatial transcriptomic and single-nucleus RNAseq^{42, 43}. Focusing on all C-fiber neurons and Aδ HTMRs, we predicted 76 total interactions involving 32 synovial-derived ligands and 68 neuronal receptors, and interactions were categorized based on prior knowledge of a given receptor (e.g. integrins vs. toll-like receptors) (Fig.- 4). In the Neuroactive and Growth Factor and category (Fig.- 4A), HGF (hepatocyte growth factor), PDGFC (platelet-derived growth factor), SEMA3A (semaphorin-3), and VEGFC (vascular endothelial growth factor) were the most dominant ligands, predicted to interact with multiple neuronal receptors; TAC1 (substance P) interacting with TACR1 (substance P receptor) was also predicted. In the Cytokine/Chemokine category (Fig.- 4B), we identified CCL3 and CXCL16 as the only members of the CCL/CXCL chemokine families, interleukins IL10, IL15, and IL18, and TNF superfamily ligands TNFSF11 (RANK ligand), TNFSF13 (A Proliferation-Inducing Ligand), and TNFSF8 (CD30 ligand); LY86 (lymphocyte antigen 86), which was a top 5 DEG in our integrative analysis, was also predicted to bind neuronal CD180. In the Cell Adhesion category (Fig.- 4C), signaling via multiple neuronal integrin receptors dominated, with ADAM9, CD14, LYZ predicted as synovial ligands. Lastly, in the MHC/TLR/Complement category (Fig.- 4D), we identified B2M (β2 microglobulin) interacting with multiple receptors, including MHC Class I receptor HLA-F. Two alarmin genes – S100A8, S100A9, and RNASE2 (Ribonuclease A family member 2) were predicted to interact with TLR24. Two complement genes - C1QA and C1QB were further identified. This unbiased analysis reveals the broad putative interactome of OA-enriched, synovial-derived, secreted factors interacting with neuronal receptors.

Figure 4 –

Putative synovium-DRG neuron interactome, categorized into A). Neuroactive and Growth Factor Signaling, B). Cytokine and Chemokine Signaling, C). Cell Adhesion Molecules, D). MHC, Complement, and TLR Signaling, based on previously-described roles of neuronal receptors. All synovial-derived ligands were significantly upregulated (FDR < 0.05) in OA relative to non-OA, based on weighted Z-based integration. Within each category, left: alluvial plots indicating the interactions of synovial-derived ligands on left to DRG neuron-expressed receptors on right. Right: Relative gene expression of all synovial-derived ligand genes, based on weighted Z-based integration (positive values indicate upregulated in OA relative to non-OA), next to synovial cell type-specific expression of each gene, based on scRNAseq.

Discussion

To comprehensively capture the transcriptome of OA synovium, we employed a cross-platform integration of data from RNA microarrays and bulk RNA sequencing (N=138 subjects from eight distinct studies). This analysis revealed that advanced OA synovium exhibits consistent overexpression of genes associated with innate and adaptive immunity, and neuro-immune/neuro-inflammatory signaling. We identified a putative immune cell-derived regulatory network involving STAT1, FLI1, and VDR as molecular mediators of neuro-immune genes that may underpin chronic inflammatory pain in OA. An unbiased synovium-DRG interactome predicted 76 interactions, involving 32 ligands identified as overexpressed in OA synovium. Our final integration results have been provided (Suppl.-Data) to support future data mining and hypothesis generation/testing.

OA is not deemed an autoimmune disease, yet increasing evidence suggests the involvement of adaptive immune cells and the overactivation of adaptive immunity-related pathways in OA. Our cell-type enrichment analysis predicted increased T and B lymphocyte content in OA synovium, and cytokines expressed also by T and B lymphocytes underpinned neuro-immune pathway activation. Prior histopathological studies have demonstrated increased B and T cell aggregates, often surrounded by plasma cells, in advanced OA synovial tissue^{50, 51}. Infiltration of CD4+ T cells, including cells expressing IFN-γ (Th1 cells) and CD161 (Th17 cells), has also been described in synovium of patients with early post-traumatic OA⁵², suggesting this finding is not exclusive to end-stage disease. Further, T_reg dysfunction was recently associated with patient-reported pain in OA subjects⁵³, indicating a potential link between lymphocyte infiltration and proalgesic signaling. While the precise function of lymphocytes in OA remains unclear, our data and prior studies suggest that they contribute to the chronic overexpression of inflammatory cytokines characteristic to OA joints, which may also mediate pain. The mechanism underlying the lymphocyte recruitment to OA synovium is undescribed, and extensive work remains to understand the role of adaptive immune cells in OA.

We identified STAT1, FLI1, and VDR as overactive molecular regulators of neuro-immune signaling in OA synovium. STAT1, and STAT signaling per se, are well-described to regulate inflammatory target genes in myeloid cells⁵⁴, and targeting of STAT6 was recently shown to ameliorate pain behavior in a rat post-traumatic OA model⁵⁵. Our results show that STAT signaling remains overactive in end-stage OA synovium, driven by multiple immune cell populations, most notably macrophages, further supporting the targeting of synovial immune cell STAT signaling as a potential pain therapy for OA. FLI1 has not been studied in the context of OA, but prior work demonstrates its role in cytokine secretion in immunity and inflammation-related contexts^{56, 57}. Importantly, FLI1 was recently shown to bind to the STAT1 promoter to regulate inflammation-relevant genes⁵⁶, suggesting that FLI1 and STAT1 may cooperate in the regulation of neuro-immune genes implicated by our analysis. VDR, the nuclear receptor for Vitamin D, is a ligand-activated transcription factor known to mediate inflammation and immune signaling⁵⁸. VDR was shown to be expressed in RA synovial fibroblasts, and its expression was associated with arthritis-relevant proinflammatory cytokine signaling^{59, 60}.

In experimental osteoarthritis, nociceptors have been shown to sprout within the synovium in response to joint injury^{47, 48}; however, it remains unknown whether nociceptor sprouting occurs in the synovium of humans with OA. The human synovium is densely innervated, and recent studies report that fibroblasts derived from knee synovium of patients with “early OA pain” (e.g. painful OA with only moderate structural damage) enhance the growth of cultured DRG neurons to a greater extent than those from non-painful knee synovium²¹. In our data, we did not observe evidence of overactive pathways related to sprouting or migration of axons of nociceptors, though we did identify synapse adhesion molecules that could play a similar role. The datasets used for this analysis utilized synovial biopsies from established to late-stage OA patients, so these data may miss evidence of nociceptor sprouting present at earlier stages of OA. Extensive future work is required to understand the mechanisms and patterns of nociceptor sprouting and remodeling in OA, and we as a part of the Restoring Joint Health and Function to Reduce Pain (RE-JOIN) consortium are actively engaged in these studies⁶¹.

We identified a putative synovium-DRG interactome consisting of 76 ligand-receptor interactions, with interactions related to neuroactive and growth factors (including axonal guidance proteins, e.g. SEMA3A), adaptive and innate immune pathways, including TLR, Complement, and MHC-related interactions, and cell adhesion-related interactions. While cytokines such as IL18 and TNF have been classically associated with neuroinflammation, alternative pathways such as complement⁶² and TLR⁶³ signaling have also been shown to be active in neurons and associated with neuroplasticity and immune activation. These findings suggest that neuroinflammatory signaling may occur through a broader set of communication axes than previously appreciated. Cell adhesion receptors such as integrins have established roles in axon guidance and development⁶⁴, and integrin-mediated crosstalk between the fibrotic OA synovial matrix and DRG neurons may also alter nociception, but this remains unexplored. Though classically associated with immune cell interactions, MHC-I expression by neurons is established and has been shown to play a key role in synaptic development and plasticity^{65, 66}; furthermore there is experimental evidence for neuronal expression of specific MHC-I-related receptors predicted by this analysis, including HLA-F⁶⁷, CD247^{68, 69} and LILRB2⁷⁰. Similarly, chemokine receptor expression by neurons has received increasing attention, with both CXCL16-CXCR6^71–73 (predicted by our analysis) and CCL2-CCR2⁷⁴ both shown to directly mediate knee joint pain signaling in mouse models. Targeted experimental work is necessary for protein-level validation and expanded characterization of the specific interactions predicted by this in silico analysis. Lastly, multiple DRG receptors predicted by our analysis (e.g. NRP1, ITGB5, LRP1) were also predicted by Bai et al ²⁰ in a similar synovium-DRG interactome analysis on a set of pain-associated genes in low inflammatory rheumatoid arthritis. This suggests potentially shared synovium-DRG pain signaling mechanisms between OA and RA, which requires further investigation.

While our dataset limited us to modeling synovium-to-DRG communication, signaling from DRG nerve afferents towards synovial cells likely also mediates disease, which needs additional experimental evidence. For example, nociceptor-derived CGRP secretion was recently shown to mediate wound healing⁷⁵, and secretion of other neuropeptides or proteins by neurons could also contribute to this two-way communication⁷⁶. Studies analyzing paired DRG and synovia (along with other joint tissues such as articular cartilage) are currently underway as part of the RE-JOIN consortium, which will enable more robust assessment of DRG-to-synovium interactions, interactions between DRG neurons and DRG-resident immune cells, and interactions between different intra-articular joint tissues, enabling a more comprehensive assessment of the complex web of multi-tissue crosstalk involved in OA.

Our study is inherently limited by the retrospective nature of analysis from data derived across multiple studies. While this strengthens generalizability via utilization of a wide range of OA subjects, our statistical power to detect differentially-expressed genes and pathways is limited by study and subject heterogeneity. Despite rigorous quality control and statistical integration process involving study weighting, we cannot rule out false-positive or false-negative findings due to disparate sample sizes across studies. Despite this, the top DEG identified in OA synovium – SCRG1 – was identified in a previous integrative analysis including 5 of the microarray studies utilized in our meta-analysis¹⁶. As their study did not include either of the bulk RNA-seq studies we utilized, this gives confidence that our analysis was not driven solely by the outsized influence of larger datasets (e.g. GSE89408). Given the statistical methodology employed, we only considered datasets with a non-OA control group; consequently, datasets which compare early- and late-stage OA but lack non-OA controls were excluded. Comparison of OA vs. non-OA human tissues is inherently limited by the challenges of defining and procuring tissue from truly healthy joints; this was exacerbated by the meta-analysis nature of this study, as patient selection criteria varied study-by-study. We chose to refer to our control group as “non-OA” rather than “healthy,” and it should be noted that this group generally represented human subjects who have not been diagnosed with OA and have no macroscopic evidence of OA, and not necessarily completely “healthy” joints. Our interactome analysis is limited by the screening of receptor genes expressed in a reference atlas of human DRG neurons.

Conclusion

This novel cross-platform data integration comprises the largest dataset of human OA synovial transcriptomic data to-date, utilizing statistically-rigorous meta-analysis techniques and leveraging data from many OA subjects not previously used for direct non-OA vs. OA comparisons in their original work. The resultant dataset represents a generalizable transcriptomic data atlas, which we have made available to inform future research into the role of the synovium in OA pathogenesis and pain. Finally, employing this integrated dataset, we assembled a putative synovium-DRG interactome and identified a neuro-immune overactivation signature associated with STAT1, VDR, and FLI1 signaling as part of the established OA synovial phenotype.