The research team, led by Lin Huang, Ling Cheng, and colleagues at Dianjiang People’s Hospital of Chongqing, reanalyzed three public bulk RNA-sequencing datasets rather than merging them into a single expression matrix. GSE140731, the discovery cohort, contained 22 adhesive capsulitis samples and 26 surgical controls from patients undergoing shoulder-instability repair. Two smaller datasets, GSE190023 and GSE238053, served as independent validation cohorts. The design choice was deliberate: previous studies of frozen shoulder have often relied on small, single-cohort analyses with differing comparators, making cross-study inference fragile. By keeping each cohort separate and adjusting the discovery analysis for the pronounced age and sex imbalance between cases and controls, the mean age was 52.7 years in the adhesive capsulitis group versus 23.9 years in the comparator group, and women accounted for 15 of 22 cases but only 5 of 26 controls, the team aimed to identify which transcriptional features genuinely replicate.

The technical core of the study was differential-expression analysis with PyDESeq2, a count-aware statistical framework that models raw sequencing counts directly. After filtering lowly expressed genes and standardizing age across all samples, the primary design specified the group coefficient as the adhesive capsulitis-versus-control difference after adjustment for age and sex. The adjusted discovery analysis identified 552 genes at a false-discovery rate below 0.05, including 439 upregulated and 113 downregulated genes. Representative upregulated genes included MMP9, a matrix-degrading enzyme, and TPM1, while downregulated genes included AMZ2P1 and OS9. The direction and number of significant genes, the authors note, were not attributable solely to an unadjusted comparison, although residual confounding by the non-healthy comparator could not be excluded.

Pathway-level enrichment told a clearer story. In the age- and sex-adjusted preranked analysis, the KEGG pathway for extracellular matrix-receptor interaction showed a normalized enrichment score of 2.299 and remained strongly significant after false-discovery correction, with a q-value of 0.001. Over-representation analysis of the upregulated genes was dominated by extracellular-structure terms: extracellular matrix organization contained 38 of 176 input genes, collagen fibril organization contained 10 of 42, and extracellular matrix disassembly contained 10 of 44. Leading-edge genes for the enriched pathway included ITGA1, THBS1, COL4A2, COL4A1, and AGRN, all components of the matrix-receptor interface. This is consistent with decades of histopathology showing myofibroblast-rich capsular fibrosis, type-III collagen deposition, and dysregulated matrix metalloproteinases in frozen-shoulder tissue, and with recent single-cell work identifying DKK3-positive and POSTN-positive fibroblast populations alongside inflammatory macrophages in the diseased capsule.

The regulated cell-death results were far more equivocal. The team constructed signature scores from curated gene sets: a nine-gene pyroptosis core including GSDMD, CASP1, NLRP3, IL1B, and IL18; a ferroptosis set including GPX4, SLC7A11, ACSL4, and HMOX1; and a necroptosis set including RIPK1, RIPK3, MLKL, and CASP8. Each score was calculated as the mean of gene-wise z-scores across samples, a summary of relative transcript abundance rather than a direct measurement of gasdermin pore formation, lipid peroxidation, or RIPK3/MLKL-dependent death execution. After adjustment, the pyroptosis score showed no difference between cases and controls, with a delta of -0.105 and a q-value of 0.708. Ferroptosis was lower in adhesive capsulitis but only borderline after correction, with a delta of -0.426 and a q-value of 0.056. Necroptosis was significantly lower, with a delta of -0.320 and a q-value of 0.035, and the core necroptosis gene set was also negatively enriched in the preranked analysis.

Across the three cohorts, the picture grew murkier still. The adjusted and unadjusted point estimates for pyroptosis were -0.10, +0.99, and -0.68 in GSE140731, GSE190023, and GSE238053 respectively, showing clearly different directions across datasets. Ferroptosis and necroptosis point estimates were consistently non-positive rather than directionally conflicting, but the effects were close to zero and imprecise in the small validation cohorts, and formal significance did not replicate. In GSE190023, severity trends across control, mild, and severe adhesive capsulitis for genes such as RIPK3, CASP1, IL1B, and NLRP3 all failed to survive multiplicity correction, with the smallest adjusted value at 0.592. Collectively, the score analyses did not support a claim of uniform regulated cell-death activation in adhesive capsulitis, a conclusion that directly challenges the assumption, common in analogous bioinformatic studies of rheumatoid arthritis and osteoarthritis, that these programs are uniformly switched on in fibroinflammatory joint disease.

Immune marker-gene scoring produced a similar pattern of discovery-cohort promise and validation-cohort failure. Fifteen curated marker panels, assembled from commonly used LM22- and Bindea-derived gene lists, summarized B cells, plasma cells, CD8 and CD4 T cells, regulatory T cells, natural killer cells, monocytes, macrophage-like states, dendritic cells, mast cells, neutrophils, endothelial cells, fibroblasts, and smooth muscle cells. After adjustment, six of the 15 scores met false-discovery significance in the discovery cohort: smooth muscle, endothelial cell, natural killer cell, CD8 T cell, and B cell scores were higher in adhesive capsulitis, while the monocyte score was lower, with the largest positive coefficients at 1.421 for endothelial cells and 1.168 for natural killer cells. But when the same panels were evaluated in the validation cohorts, their directions varied and no panel reached cohort-wise false-discovery significance in either dataset. The authors are careful to note that these scores are marker-gene summaries, not deconvolution-derived cell fractions, and cannot be translated into percentages or absolute numbers of immune or stromal cells.

Cross-cohort concordance statistics quantified the heterogeneity. Gene-wide effects in the adjusted discovery cohort correlated only modestly with GSE238053, with a Spearman rho of 0.383 and 68 genes significant in both cohorts, but were essentially uncorrelated with GSE190023, with a rho of -0.047 and only four shared significant genes. Same-direction proportions across all matched genes were 0.637 for GSE238053 and 0.493 for GSE190023. The near-zero correlation with GSE190023 indicated lack of replication in that small cohort rather than identifying a technical bias or proving either dataset erroneous. Only two regulated cell-death-related genes, PLA2G4A, downregulated, and MAPK12, upregulated, were significant with the same direction in two cohorts. By contrast, MMP9 showed the most stable large positive effect across all three datasets, with log2 fold changes of +8.3, +7.0, and +5.0, while POSTN, SPON2, and PITPNC1 retained generally positive directions.

The team also built an exploratory three-gene classifier from an L1-penalized logistic regression, selecting COL11A1, SPON2, and PITPNC1 as the retained genes. Using pooled out-of-fold predictions, the discovery-cohort area under the curve was 0.820, with a 95 percent confidence interval of 0.684 to 0.930. External-cohort AUCs ranged from 0.750 to 0.812 but carried very wide confidence intervals because of the limited sample sizes, and because the candidate pool was defined before cross-validation, the entire classifier analysis remained non-nested and exploratory. The authors explicitly state that the model provides an exploratory summary of separability in these datasets rather than evidence of clinical utility, and that prospective evaluation in larger, independently recruited populations with prespecified features and fully nested model development would be required before any diagnostic claim could be considered.

A laboratory experiment added a partial, model-specific layer of support. The team stimulated SW982 human synovial sarcoma cells with 10 nanograms per milliliter of recombinant human TGF-beta1 for 24 hours, an operationally defined profibrotic stimulus, and measured four genes by quantitative real-time PCR with GAPDH as the reference. COL11A1 expression rose from a mean relative value of 1.036 in vehicle controls to 1.888 under stimulation, with a P-value of 0.0006. SPON2 rose from 1.016 to 1.518, with a P-value of 0.0040. PLA2G4A fell from 1.013 to 0.700, with a P-value of 0.0065. PITPNC1 showed a smaller numerical increase that was not statistically significant, with a P-value of 0.272. The non-significant PITPNC1 result illustrates that computational feature selection does not guarantee experimental replication, and the authors interpret the qPCR findings as model-specific expression validation rather than proof of regulated cell-death pathway execution, causal relevance to adhesive capsulitis, or clinical validity.

The broader significance of the study lies in its discipline about what the data can and cannot show. Bulk RNA-sequencing cannot resolve whether observed expression differences reflect transcriptional regulation within cells or shifts in cellular composition, and transcript abundance alone should never be read as evidence that a cell-death program has been executed. The discovery comparator was a younger, more male shoulder-instability surgical population rather than a healthy control group, and age and sex adjustment reduces but cannot eliminate confounding. Both validation cohorts were small, limiting statistical power. Future work, the authors argue, should combine independent clinical cohorts with spatial or histologic validation and functional experiments in primary adhesive capsulitis fibroblasts and macrophages using pathway-specific perturbation and rescue assays. Until such evidence is available, the regulated cell-death and immune findings should be regarded as hypothesis-generating associations, while extracellular-matrix remodeling stands as the most reproducible molecular feature of frozen shoulder in the present data, reinforcing an ECM-centered, cohort-aware framework for a condition that affects millions worldwide.

Subject of Research: A three-cohort transcriptomic reanalysis of adhesive capsulitis identifies extracellular-matrix remodeling as the most reproducible molecular feature, while regulated cell-death and immune signatures show cohort-dependent, non-replicating signals.

Article Title: Extracellular matrix remodeling and regulated cell-death-immune associations in adhesive capsulitis: a three-cohort transcriptomic analysis

Article References: Huang, L., Cheng, L., He, J., Zhang, R., Li, X., & Xia, Q. (2026). Extracellular matrix remodeling and regulated cell-death-immune associations in adhesive capsulitis: a three-cohort transcriptomic analysis. Frontiers in Cell and Developmental Biology, 14, Article 1951682. https://doi.org/10.3389/fcell.2026.1951682

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1951682

Keywords: adhesive capsulitis, frozen shoulder, extracellular matrix, transcriptomics, pyroptosis, ferroptosis, necroptosis, fibrosis, Extracellular, matrix, remodeling, regulated