Background: Tumor Necrosis Factor inhibitors (TNFi) are the first line biologic treatment for RA. However, about 40% of RA patients do not respond to TNFi, highlighting the need for predictive biomarkers.

Objectives: We integrated synovial and blood DNA methylation data with RNA-sequencing and genotype data to identify drivers of synovial gene expression, markers of disease activity and optimal pre-treatment predictors of TNFi response in RA.

Methods: Differential DNA methylation and gene expression analysis was performed on the pre-treatment baseline synovial biopsy samples (n=54) looking at differences between EULAR DAS28-ESR good responders/non-responder to etanercept and lymphoid/fibroid pathotypes classified based on synovial histology. Hierarchical clustering of gene expression and DNA methylation profiles covering key inflammatory pathways related to TNF, IL-6, B-cells (CD20), and Fibroblast Growth Factor (FGF) was performed. A machine learning model with 10x10-fold nested cross-validation (CV) with 25 repeats utilizing DNA methylation probe data as input was constructed to predict TNFi response.

Results: We identified 44 differential DNA methylation regions (DMRs) across 48 genes that differentiate good responders from non-responders. Notable genes included IL32 , linked to RA disease activity, and RA susceptibility genes ARID5B and RUNX3 , involved in immune regulation and T cell development. Response analysis revealed 45 significantly differentially expressed genes (DEGs), including TNFRSF13B and B cell signalling genes ( BLK, CD19 ). Lymphoid and fibroid pathotypes are linked to differential treatment response to TNF-inhibitors. Analysis showed distinct methylation and gene expression patterns between pathotypes, with 6389 differentially methylated probes (DMPs) and 3681 DEGs identified, of which 688 genes had 2680 DMPs located in their promoter regions, demonstrating that differential methylation influences differential gene expression. Network analysis highlighted 378 DEGs related to the immune system, 70 in lymphoid cell interactions, and 11 in the TNF-α signalling pathway ( TNFSF11, BIRC3, LTB ). Hierarchical clustering of pre-treatment DNA methylation revealed a subgroup (n=12, 83% good responders) with hypomethylation of TNF and IL6 module genes and FGF module hypermethylation. RNA-seq showed a similar subgroup (n=10, 90% good responders) with increased TNF/IL6/CD20 gene module expression and reduced FGF expression. These findings indicate that molecular patterns can distinguish responders at baseline, with specific subgroups showing varying response likelihoods to etanercept. Machine learning fitted using nested CV with a glmnet penalised regression model predicted a target DAS28<3.2 after 16 weeks of treatment with AUC of 0.81, suggesting that baseline DNA methylation data can accurately predict treatment response. Expression quantitative trait loci (eQTL) analysis highlighted 3954 genes, of which 8 eQTL driven genes, including B-cell specific genes BLK and FCLR1 , showed a significant association with response to treatment, underscoring the role of genotypic control over expression in affecting treatment responsiveness.

Conclusion: This study substantiates the predictive capacity of DNA methylation and multi-omic profiling in forecasting TNFi response in RA. The machine learning model demonstrates the robustness of epigenetic markers as predictors of treatment response, enhancing the prospects for personalized treatment strategies. Clinically, these findings have the potential to refine patient stratification and guide precision therapy, potentially transforming RA management by pre-emptively identifying individuals likely to benefit from anti-TNF therapy, thereby improving patient prognosis and optimizing healthcare resource allocation.

REFERENCES: NIL.

Acknowledgements: We thank all patients participating in the trial and the Patient Advisory Group. The STRAP trial was part of the Maximising Therapeutic Utility in RA (MATURA) programme, jointly funded by the UK Medical Research Council (MRC) and Versus Arthritis (grant reference MR/K015346/1). Infrastructure support was also provided by Versus Arthritis (Experimental Arthritis Treatment Centre grant, number 20022). We also acknowledge support from Barts Charity (grant number 523/819), NIHR (grant 131575) and MRC TRACT-RA (MR/V012509/1). This work acknowledges the support of the National Institute for Health Research Barts Biomedical Research Centre (NIHR 203330).

Disclosure of Interests: None declared.