Related Experiment Videos
Integration of Single-Cell Transcriptomics and Network-Based Machine Learning Identifies Bone Marrow Mesenchymal Stem
Hanyu Wang1, Zhikang Chen2, Shikai Chen1
1Spine Institute, Longhua Hospital, Shanghai University of Traditional Chinese Medicine, Shanghai, 200032, China.
Introduction:
Osteoarthritis (OA) is a whole-joint disease involving coordinated changes in subchondral bone, Bone Marrow Mesenchymal Stem cells (BM-MSCs), and synovium. This study aimed to identify BM-MSC-associated and OA-related candidate biomarkers and evaluate their potential diagnostic and immune-related relevance.
Methods:
The scRNA-seq dataset GSE147287 was used to construct an OA subchondral bonemarrow atlas, identify BM-MSC populations, perform BM-MSC sub-clustering, and derive BMMSC co-expression modules and hub genes using hdWGCNA. Three synovial transcriptomic datasets (GSE55457, GSE55235, and GSE55584) were integrated as the training cohort after datasetspecific log2 (expression + 1) transformation and ComBat correction, whereas GSE12021 was processed independently as an external validation cohort. BM-MSC hub genes were intersected with OA-related synovial DEGs and further prioritized using LASSO, SVM-RFE, and XGBoost. A combined four-gene model was constructed using Firth's penalized logistic regression. Diagnostic performance was evaluated by ROC analysis, cellular expression was validated in IL-1β-stimulated hBMSCs by qRT-PCR, and immune-cell and Hallmark pathway correlations were assessed using ssGSEA and Spearman analysis with Benjamini-Hochberg correction.
Results:
The single-cell atlas contained nine major cell types, and BM-MSC sub-clustering identified four transcriptionally distinct subclusters. Cross-tissue integration yielded 35 BM-MSCassociated and OA-related candidate genes, from which EFEMP2, CTSO, SPRY1, and NFIC were selected. In the training cohort, the AUCs were 0.890 (95% CI, 0.792-0.989) for EFEMP2, 0.896 (95% CI, 0.802-0.990) for CTSO, 0.892 (95% CI, 0.797-0.987) for SPRY1, and 0.769 (95% CI, 0.630-0.908) for NFIC. In GSE12021, the corresponding AUCs were 0.867 (95% CI, 0.698-1.000), 0.922 (95% CI, 0.801-1.000), 0.922 (95% CI, 0.797-1.000), and 0.856 (95% CI, 0.674-1.000), respectively. The combined four-gene model achieved AUCs of 0.998 (0.993-1.000) in the training cohort and 1.000 (1.000-1.000) in GSE12021. qRT-PCR confirmed the upregulation of EFEMP2 and CTSO and the downregulation of SPRY1 and NFIC in IL-1β-stimulated hBMSCs. After Benjamini- Hochberg correction, no immune-cell or Hallmark pathway correlations remained statistically significant, although several moderate exploratory trends were observed.
Discussion:
EFEMP2, CTSO, SPRY1, and NFIC represent BM-MSC-associated and OA-related candidate biomarkers. These findings provide a hypothesis-generating framework for characterizing OA-related stromal states and developing future tissue-fitness or patient-stratification tools, but further clinical and tissue-specific validation is required.
Conclusion:
EFEMP2, CTSO, SPRY1, and NFIC were identified as BM-MSC-associated and OArelated candidate biomarkers. These findings provide a hypothesis-generating framework for characterizing OA-related stromal states and developing future tissue-state assessment or patientstratification tools, although further tissue-specific and clinical validation is required.
