Uncovering CD248, MMP28, and SLC16A10 in Sjögren's disease: a machine learning-driven SHAP approach for CD4+ T...

Sjögren's disease (SjD) is a highly heterogeneous autoimmune disease with substantial challenges in early diagnosis and therapeutic intervention. We developed an integrated approach combining machine learning algorithms, SHAP interpretable modelling, molecular docking, and single-cell analysis to facilitate early diagnosis and treatment of SjD.

Transcriptomic data and 12 machine learning algorithms were employed to identify diagnostic signature genes. SHAP (Shapley Additive exPlanations) analysis further prioritised hub genes, followed by functional annotation using CIBERSORT, GSVA, and GSEA. Validation was performed using clinical cohorts, single-cell RNA sequencing (scRNA-seq), and molecular docking.

The training cohort comprised 382 samples (61 healthy controls, 321 SjD patients) and 10,015 genes. Machine learning and SHAP analysis identified three hub genes (CD248, MMP28, SLC16A10), validated in external datasets with significant differential expression (p<0.05) and robust diagnostic performance (AUC >0.7). Immune infiltration analysis revealed positive correlations between CD248/SLC16A10 and naive CD4+ T cells (p<0.05), and between SLC16A10/MMP28 and memory resting CD4+ T cells (p<0.05). Single-cell profiling localised CD248 predominantly in naive CD4+ T cells, while SLC16A10 and MMP28 were expressed in both naive and memory CD4+ T cells subsets. Molecular docking demonstrated stable targeting of CD248, MMP28, and SLC16A10 by azathioprine, leflunomide, methotrexate, hydroxychloroquine, iguratimod, pilocarpine, and cevimeline.

Our bioinformatic study identifies CD248, MMP28 and SLC16A10 as candidate biomarkers and therapeutic targets for SjD, with their dysregulation specifically enriched in CD4+ T cell subsets, unveiling a previously underappreciated mechanism in SjD pathogenesis. naive and memory CD4+ T cells emerge as key contributors to inflammatory cascades, with azathioprine, leflunomide, methotrexate, hydroxychloroquine, iguratimod, pilocarpine, and cevimeline predicted to bind potently to these targets. This integrative multi-omics framework, combining machine learning, SHAP, and molecular docking, presents a promising approach for autoimmune disease diagnostics and early therapeutic intervention, although future experimental validation is essential to confirm its translational potential.