[Explore the potential diagnostic markers of primary Sjögren syndrome by bioinformatics combined with machine...

To develop a diagnostic model for primary Sjögren syndrome (pSS) utilizing transcriptomics-based machine learning techniques, and to explore potential therapeutic targets.

We selected patients diagnosed with pSS and control subjects with salivary gland samples from the Gene Expression Omnibus (GEO) database for transcriptome sequencing. This was complemented by methylation sequencing data, followed by weighted gene co-expression network analysis (WGCNA) in conjunction with clinical data to identify candidate diagnostic genes. Subsequently, we refined the gene list using least absolute shrinkage and selection operator (LASSO), random forest (RF), Gaussian mixed model (GMM), and support vector machine (SVM). Ultimately, we constructed a composite diagnostic model for pSS employing SVM, neural networks (NN), Logistic regression (LR), decision trees (DT), adaptive boosting (AdaBoost), bootstrap aggregating (Bagging), random forest, and Naive Bayes classifiers, totaling eight distinct machine learning algorithms, and predicted drug targets.

Through differential gene screening combined with WGCNA analysis of the GEO database, we identified 12 key genes. Further variable selection via machine learning techniques yielded four significant genes: VNN2, NAPSB, SAMSN1, and MS4A1. The area under curve (AUC) values were calculated for each gene: AUC_VNN2 =0.865, AUC_NAPSB=0.896, AUC_SAMSN1 =0.872, AUC_MS4A1 =0.914, all demonstrating diagnostic relevance. We then employed these machine learning methods to construct a composite diagnostic model, and F1 score (the harmonic mean of precision and recall rate) was used to evaluate the prediction performance of each model in the internal test set. The top-performing models included those based on SVM polynomial kernel models, SVM linear kernel models, random forest approaches, and Naive Bayes classification as determined through ranking criteria based on AUC comparisons; notably, both SVM polynomial kernel models and linear kernel models exhibited superior AUC values compared with individual genes alone. The external validation set, SVM polynomial kernel function model in GSE159574, AUC=0.865 (95%CI: 0.728-1.000), sensitivity 0.812, specificity 0.846; The AUC of SVM polynomial kernel model in GSE173808, AUC=0.896 (95%CI: 0.797-0.996), the sensitivity was 0.778, and the specificity was 1.000. In order to facilitate clinical diagnosis, we verified that the AUC of the SVM polynomial kernel function model in the peripheral blood GSE84844 dataset was 0.827 (95%CI: 0.719-0.934), the sensitivity was 0.867, and the specificity was 0.767. Thus, our proposed model demonstrated robust diagnostic capability for pSS. Additionally, our investigation into potential drug targets indicated that these four identified genes served as targets for glucocorticoid therapies while curcumin emerged as a promising candidate agent in treating pSS.

This research identifies VNN2, NAPSB, SAMSN1, and MS4A1 as specific biomarkers for diagnosing pSS, thereby enriching existing genetic markers relevant to diagnosis while establishing a reliable composite diagnostic framework which underscores the close association between methylation processes in the pathogenesis of pSS, offering novel insights into early diagnosis strategies along with prospective treatment avenues.