Objective This study aimed to develop a robust transcriptomic diagnostic signature for Sjögren’s disease (SjD; formerly Sjögren’s syndrome) and elucidate key biomarker functions by integrating machine learning and single-cell analysis. Methods Three SjD peripheral blood datasets (GSE143153, GSE51092, GSE66795; n = 414) were integrated for training, with GSE84844 and GSE40611 for external validation. Candidate biomarkers were identified through differential expression analysis and WGCNA. A total of 113 machine learning algorithm combinations were evaluated. The top biomarker was characterized through SHAP interpretability analysis, immune infiltration profiling, pathway enrichment, genetic colocalization, ceRNA network construction, and RT-qPCR validation. Single-cell RNA sequencing, CellChat, and virtual gene knockout analyses were performed to investigate cell-type expression and regulatory networks. Results Eighty-six core candidate genes were identified. The optimal model (plsRglm + rf) selected 14 diagnostic genes, achieving AUC values of 0.896 (training), 0.871 (GSE84844), and 0.874 (GSE40611). EPSTI1 showed the highest single-gene performance (AUC = 0.844) and significant correlations with IgG (R = 0.64, P = 0.00012) and ANA (R = 0.49, P = 0.0066). SHAP analysis ranked EPSTI1 as the top feature. RT-qPCR in 65 SjD patients and 48 controls confirmed significant EPSTI1 upregulation. Single-cell analysis localized EPSTI1 to monocytes and dendritic cells. CellChat identified enhanced MIF-CD74/CXCR4 signaling, and virtual knockout demonstrated EPSTI1 as a specific downstream effector within the interferon cascade. Conclusion This study established an integrated machine learning framework identifying EPSTI1 as a robust SjD diagnostic biomarker predominantly expressed in myeloid cells. Multi-dimensional validation supports its clinical potential for precision diagnosis of SjD, pending prospective confirmation in larger cohorts. <Table Float="No" ID="Taba"> <tgroup cols="2"> <colspec align="left" colname="c1" colnum="1" /> <colspec align="left" colname="c2" colnum="2" /> <tbody> <row> <entry align="left" nameend="c2" namest="c1"> Key Points • A systematic evaluation of 113 machine learning algorithm combinations identified a 14-gene diagnostic signature for Sjögren’s disease with robust performance across multiple independent cohorts (AUC > 0.87). • EPSTI1 emerged as the top-ranked diagnostic biomarker through convergent evidence from machine learning feature selection, SHAP interpretability analysis, and RT-qPCR validation in clinical samples. • Single-cell RNA sequencing localized EPSTI1 expression predominantly to monocytes and dendritic cells, linking its diagnostic utility to myeloid-mediated immune dysregulation. • Virtual gene knockout analysis positioned EPSTI1 as a specific downstream effector within the interferon signaling cascade, distinguishing it from broad upstream regulators like STAT1. </entry> </row> </tbody> </tgroup> </Table>

Integrated machine learning framework identifies EPSTI1 as a key diagnostic biomarker for Sjögren’s disease: multi-cohort transcriptomic validation and single-cell characterization

期刊论文
发表日期：

2026年7月10日

查看全文

Jianbin Li,
Renhe Li,
Wenwen Wang,
Yuzhen Gesang,
Wei Liu

摘要

Objective

This study aimed to develop a robust transcriptomic diagnostic signature for Sjögren’s disease (SjD; formerly Sjögren’s syndrome) and elucidate key biomarker functions by integrating machine learning and single-cell analysis.

Methods

Three SjD peripheral blood datasets (GSE143153, GSE51092, GSE66795; n = 414) were integrated for training, with GSE84844 and GSE40611 for external validation. Candidate biomarkers were identified through differential expression analysis and WGCNA. A total of 113 machine learning algorithm combinations were evaluated. The top biomarker was characterized through SHAP interpretability analysis, immune infiltration profiling, pathway enrichment, genetic colocalization, ceRNA network construction, and RT-qPCR validation. Single-cell RNA sequencing, CellChat, and virtual gene knockout analyses were performed to investigate cell-type expression and regulatory networks.

Results

Eighty-six core candidate genes were identified. The optimal model (plsRglm + rf) selected 14 diagnostic genes, achieving AUC values of 0.896 (training), 0.871 (GSE84844), and 0.874 (GSE40611). EPSTI1 showed the highest single-gene performance (AUC = 0.844) and significant correlations with IgG (R = 0.64, P = 0.00012) and ANA (R = 0.49, P = 0.0066). SHAP analysis ranked EPSTI1 as the top feature. RT-qPCR in 65 SjD patients and 48 controls confirmed significant EPSTI1 upregulation. Single-cell analysis localized EPSTI1 to monocytes and dendritic cells. CellChat identified enhanced MIF-CD74/CXCR4 signaling, and virtual knockout demonstrated EPSTI1 as a specific downstream effector within the interferon cascade.

Conclusion

This study established an integrated machine learning framework identifying EPSTI1 as a robust SjD diagnostic biomarker predominantly expressed in myeloid cells. Multi-dimensional validation supports its clinical potential for precision diagnosis of SjD, pending prospective confirmation in larger cohorts.

Key Points

• A systematic evaluation of 113 machine learning algorithm combinations identified a 14-gene diagnostic signature for Sjögren’s disease with robust performance across multiple independent cohorts (AUC > 0.87).

• EPSTI1 emerged as the top-ranked diagnostic biomarker through convergent evidence from machine learning feature selection, SHAP interpretability analysis, and RT-qPCR validation in clinical samples.

• Single-cell RNA sequencing localized EPSTI1 expression predominantly to monocytes and dendritic cells, linking its diagnostic utility to myeloid-mediated immune dysregulation.

• Virtual gene knockout analysis positioned EPSTI1 as a specific downstream effector within the interferon signaling cascade, distinguishing it from broad upstream regulators like STAT1.