Background/Purpose: The histopathologic features of synovial tissue vary widely among patients with rheumatoid arthritis (RA) undergoing arthroplasty and the clinical significance of this variability is unknown. Improved methods for assessing synovial histopathology could help characterize subtypes of RA and predict clinical outcomes.

Methods: 149 RA patients undergoing total hip or knee arthroplasty were enrolled. 130 synovial tissue samples were stained with hematoxylin and eosin and assessed for 20 histologic features such as lymphocytic infiltrates, neutrophils, mucoid change, vascularity, synovial lining hyperplasia, detritus, fibrin, plasma cell inflammation, binucleate plasma cells, germinal centers, Russell bodies, mast cells, giant cells and synovial giant cells. 19 of the synovial samples were disaggregated and their synovicoytes analyzed by RNA sequencing. We used a standard unsupervised hierarchical clustering approach to group samples and genes with similar expression profiles. Using the resultant RNAseq sample clusters, we applied a machine learning algorithm (standard Support Vector Machine) using binarized histology features and a leave-one-out cross-validation approach. The model was able to accurately predict which RNAseq cluster a given sample would most likely belong to, suggesting good harmony between the genomic data and the histology. We then used our histology SVM model derived on the 19 samples to cast cluster predictions on the broader set of 130 synovial samples, for which we had complete histology data but no RNAseq.

Results: Unsupervised hierarchical clustering of 19 synoviocyte samples based on genome-wide RNA expression levels identified 2 well-separated subtypes of synovial tissue composition that distinguished samples based on the level of inflammatory cell presence, and a third subgroup within the low-inflammatory cell subset that suggests further subclassification of RA synovial tissue based on non-immune cell types. Confirming the validity of this approach, histology scoring of lymphocytic infiltrates correlated significantly with the RNAseq clusters containing the highest level of hematopoietic lineage transcripts. Furthermore, clinical features such as ESR and obesity correlated with the subsets defined by the RNAseq analyses. Gene set enrichment analysis of differentially expressed genes between the clusters showed variable levels of immune and stromal cell subsets, noting particular differences in IFNg responses. Application of the histology SVM models to the remaining ~100 samples generated predictions of RA synovial gene expression subtypes, which together with ongoing studies connecting clinical parameters to this model suggest an approach to subclassify the disease with greater specificity.

Conclusion: Gene expression clusters derived from RNAseq were used to train a histology scoring system to distinguish between inflammatory and noninflammatory RA synovial tissue. SVM weights indicate that lymphocytic infiltrates, plasma cells, fibrin and synovial lining hyperplasia are predicted to be associated with inflammatory gene expression, while mast cells, vascularity, detritus and mucoid changes are predicted to associate with non-inflammatory samples.

« Back to 2016 ACR/ARHP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/identifying-rheumatoid-arthritis-subtypes-using-synovial-tissue-gene-expression-profiling-histologic-scoring-and-clinical-correlates/