An epithelial–mesenchymal transition-like synovial stromal-remodeling programme associated with knee osteoarthritis pain: Exploratory transcriptomic discovery and spatial analysis

Background Pain is the defining symptom of knee osteoarthritis (OA), yet its molecular correlates in synovial tissue remain incompletely understood. Methods Across publicly available synovial transcriptomic datasets, we derived an exploratory pain-associated gene program by differential expression in synovium stratified by pain severity (GSE99662, n = 10; high vs low pain), characterized its pathway enrichment, tested specificity against an OA-versus-healthy contrast (GSE89408, n = 50), related it to inflammatory and fibrotic axes (GSE283079, n = 36 OA), and examined a larger, independent patient-level whole-transcriptome spatial cohort (Philpott et al. 2025 [1]; GSE248454; 32 knee-OA patients, 16 more- vs 16 less-pain; intimal-lining, sublining and perivascular compartments, one region per compartment per patient). The spatial analysis applied a limit-of-quantification detection filter and permutation of residuals (Freedman–Lane) from models adjusting for age and region cellularity, alongside correlation-aware competitive testing (CAMERA). This analysis plan was specified post hoc. Results Discovery yielded an exploratory program of 38 genes (35 up-regulated). An epithelial–mesenchymal transition (EMT)-like stromal/remodeling signature — interpreted in synovium as matrix remodeling rather than literal epithelial transition — was the only Hallmark set of 50 significantly enriched (NES = 1.885, p adj = 0.0022), although at n = 10 it did not survive exact phenotype-label permutation (p = 0.226). In the spatial cohort only 4,584 of 18,695 targets were above the detection limit in a typical region. After filtering, the discovery-derived Hallmark-EMT programme was positively associated with more pain in sublining regions (Holm-adjusted p = 0.010 and 0.005 at ≥5% and ≥10% detection across the discovery-derived hypotheses × compartments), persisting after adjustment for age and cellularity and against gene sets matched on abundance and detection, and carried by matrix, adhesion and TGF-β-associated transcripts. Three limits are reported with equal weight: the effect on the module scale was modest and imprecise (+0.30 z-units, 95% CI −0.06 to +0.65); an undirected scan of all 50 Hallmark sets across three compartments yielded no set surviving correction (EMT q = 0.105); and a within-patient contrast did not confirm that the association differs between compartments (p = 0.196). The 35-gene signature did not replicate in any compartment; a nominal unadjusted perivascular association did not survive adjustment, attenuated primarily by age. The programme was not enriched in OA-versus-healthy synovium. Conclusions An EMT-like stromal-remodeling transcriptional programme is associated with worse patient-level knee pain in sublining synovium. The evidence is exploratory and method-dependent, the compartment difference is not established, and the specific discovery signature does not generalize.