The challenge the researchers set out to tackle is well known in rheumatology. Current biomarkers, including rheumatoid factor and anti-cyclic citrullinated peptide antibodies, demonstrate limited sensitivity and fail to reliably reflect fluctuations in disease activity. Because rheumatoid arthritis is highly heterogeneous, with synovial inflammation driven by coordinated action of many genes involved in recruiting various immune cells to the inflammatory site, single-marker approaches have repeatedly fallen short. Early diagnosis matters enormously, since timely intervention is needed to prevent irreversible structural joint damage and disability caused by pannus formation, in which hyperplastic synovial tissue invades cartilage and bone.

To build a more robust analytical foundation, the team downloaded three gene expression datasets from the Gene Expression Omnibus database: GSE55457, GSE12021 and GSE77298, all derived from human synovial tissue samples. Because the datasets came from different studies and microarray platforms, they first corrected for batch effects using the ComBat function of the sva package in R. Principal component analysis before and after correction showed that the three datasets were initially separated by study, but post-ComBat normalization achieved better sample integration while the biological distinction between rheumatoid arthritis and healthy control samples remained evident. This step is critical, as technical variability between batches can otherwise masquerade as disease signal.

With the integrated matrix in hand, the researchers identified differentially expressed genes using the limma package, applying linear models with empirical Bayes moderation and selecting genes with an absolute log2 fold change greater than 1 and an adjusted p-value below 0.05 according to the Benjamini-Hochberg method. The analysis yielded 440 differentially expressed genes, comprising 243 upregulated and 197 downregulated genes in rheumatoid arthritis synovium. Genes associated with immune response processes, such as chemokines and cytokine receptors, were mostly upregulated, while genes related to metabolism and structural proteins were mostly downregulated. Functional enrichment analysis using Gene Ontology, KEGG and gene set enrichment analysis showed the differentially expressed genes were involved in inflammatory response, T-cell activation, B-cell receptor signaling, cytokine-cytokine receptor interaction, hematopoietic cell lineage, primary immunodeficiency and rheumatoid arthritis pathways.

The next stage narrowed the field through network analysis and machine learning. The top 100 differentially expressed genes, ranked by absolute logFC, were submitted to the STRING database to construct a protein-protein interaction network, which was visualized and analyzed in Cytoscape. The maximum clique centrality algorithm in the cytoHubba plugin identified the top 20 genes as hub candidates, most associated with immune cells, chemokine signaling and inflammatory responses. Two complementary machine learning algorithms then filtered these candidates. LASSO logistic regression, fitted with the glmnet package and stratified 10-fold cross-validation, shrank the coefficients of most candidate genes to zero as the regularization parameter increased; under the lambda.1se criterion, only CXCL13 and CXCL10 retained non-zero coefficients. Support vector machine-recursive feature elimination, run with a linear SVM classifier and stratified five-fold cross-validation repeated 10 times, independently converged on the same two genes. The training dataset consisted of 41 rheumatoid arthritis and 26 healthy control samples, and all feature selection was performed exclusively on this data.

Diagnostic performance was assessed through receiver operating characteristic curve analysis. In the training set, CXCL13 produced an area under the curve of 0.954, with an accuracy of 89.6 percent, a sensitivity of 85.4 percent, a specificity of 96.2 percent and an F1 score of 90.9 percent. CXCL10 produced an AUC of 0.925, with an accuracy of 88.1 percent, a sensitivity of 87.8 percent, a specificity of 88.5 percent and an F1 score of 90.0 percent. The researchers then evaluated the biomarkers in an independent RNA-seq validation dataset, GSE89408, comprising 152 rheumatoid arthritis and 28 healthy control samples, which was not involved in any feature selection or modeling procedure. Although performance was somewhat lower in the validation cohort, the genes retained good diagnostic value: CXCL13 achieved an AUC of 0.901 with 81.7 percent accuracy, while CXCL10 achieved an AUC of 0.886 with 78.9 percent accuracy and a striking 96.4 percent specificity.

Beyond diagnosis, the study probed the mechanistic dimension of these biomarkers using CIBERSORTx, a digital cytometry tool that estimates the relative abundance of 22 immune cell subsets from bulk tissue expression data using the LM22 signature matrix. All 67 samples satisfied the deconvolution quality criterion. Compared with healthy subjects, rheumatoid arthritis synovium showed significantly increased infiltration of plasma cells, T follicular helper cells, gamma delta T cells, monocytes and M1 macrophages, along with decreased resting mast cells. Spearman correlation analysis, adjusted for multiple testing, revealed strong positive associations between the hub genes and infiltrated immune cells: CXCL10 expression correlated with M1 macrophages at r = 0.83, while CXCL13 correlated with M1 macrophages at r = 0.75 and with gamma delta T cells at r = 0.72. CXCL13 was associated with multiple immune cell populations, including plasma cells, T-cell subsets and M1 macrophages, whereas CXCL10 showed significant associations with M1 macrophages, plasma cells and selected T-cell subsets.

These correlation patterns carry biological weight. CXCL13 has previously been reported to be associated with B-cell migration and the formation of ectopic lymphoid structures in synovial tissue, while CXCL10 has been linked to T-cell migration and macrophage recruitment during inflammation. The researchers suggest that the corresponding receptor axes, CXCL13-CXCR5 and CXCL10-CXCR3, may represent potential target pathways warranting further investigation. Inhibiting CXCR3 and CXCR5 receptors on immune cells could block activation by other ligands within the multi-ligand, multi-receptor families of chemokines, offering a hypothesis for future therapeutic research. Notably, CXCL13 and CXCL10 have previously been examined in rheumatoid arthritis as baseline serum predictors of response to TNF-inhibitor therapy in a small prospective clinical cohort, but the present study differs in both aim and approach by establishing them as diagnostic biomarkers derived from synovial tissue transcriptomics.

The authors acknowledge limitations, noting that the analysis is based on microarray gene expression data, which may have inherent constraints, and that the unequal distribution of samples in the validation dataset means results were interpreted primarily based on ROC-AUC values as supportive rather than definitive evidence. Nevertheless, the integrative framework, combining batch-corrected multi-cohort gene expression data, dual machine learning feature selection and external validation in an independent RNA-seq dataset, provides an analytical platform that can be used to identify and validate potential biomarkers for heterogeneous autoimmune diseases. As synovial inflammation in rheumatoid arthritis is associated with the coordinated action of many genes involved in immune cell recruitment, CXCL13 and CXCL10 stand out as molecular signposts of the inflammatory storm within the joint, and their receptor pathways may guide the next generation of diagnostic and therapeutic exploration.

Subject of Research: Health

Article Title: Machine learning-based identification of CXCL13 and CXCL10 as key biomarkers in rheumatoid arthritis

Article References: M, S., S, S. A., M, S., P, N., Salau, A. O., Usman, A. G., & Yousfan, M. (2026). Machine learning-based identification of CXCL13 and CXCL10 as key biomarkers in rheumatoid arthritis. Frontiers in Bioinformatics, 6, Article 1925964. https://doi.org/10.3389/fbinf.2026.1925964

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1925964

Keywords: machine learning, rheumatoid arthritis, biomarkers, CXCL13, CXCL10, bioinformatics, immunology, Machine, learning-based, identification, rheumatoid, arthritis