O08 Developing a DNA methylation signature for predicting rheumatoid arthritis using a machine learning pipeline

Background/Aims Early diagnosis and intervention improves outcomes of immune mediated rheumatic and musculoskeletal diseases (RMDs) but may be hampered by diagnostic uncertainty. The extent to which rationally selected molecular parameters add value to clinical characteristics for diagnostic prediction in undifferentiated disease states warrants investigation. B lymphocytes play an increasingly recognised role in rheumatoid arthritis (RA) pathogenesis, and cell-specific methylation patterns link environmental exposures to genetic risk. We derived and tested the practical utility of a B lymphocyte-derived DNA methylation signature for predicting RA in an early arthritis clinic cohort. Methods CD19+ B cell and peripheral blood mononuclear cell (PBMC) whole genome DNA methylation array data were available, respectively, from 109 inflammatory arthritis patients naïve to immunomodulatory drugs (Newcastle, UK; 38% confirmed to have a diagnosis of RA within 1 year) and 50 untreated undifferentiated arthritis (UA) patients (Leiden, The Netherlands; 68% classifiable RA within 1 year by 1987 ACR criteria versus alternate diagnoses). A bespoke machine learning pipeline employed a sequential model-based optimisation (SMBO) procedure for selecting, tuning and applying methods amongst ten feature-selection, six data-sampling and two classification algorithms in the Newcastle “training cohort.” The predictive performance of the resultant optimised molecular classifier was assessed in the independent Leiden “test cohort” alongside a previously described clinical prediction rule, using comparative area under receiver operating characteristic (AUROC) curves. A modification to the clinical prediction rule that incorporated a single parameter to reflect molecular classification was also assessed. The pipeline was implemented using the R machine learning package mlr. Results Using the SMBO approach, 27 CpGs maximally discriminatory for RA were selected from B lymphocyte DNA methylome training data, and a molecular classifier was derived using the random forest algorithm. Applied to the independent PBMC methylome in UA patients, the classifier and the validated Leiden prediction rule performed similarly in predicting RA (AUROC [95% CI] = 0.8 [0.66-0.94] versus 0.78 [0.64-0.92]). Interestingly, incorporating a molecular risk score based on the 27-CpG signature into the validated Leiden clinical prediction rule significantly improved its performance (AUROC [95% CI] = 0.89 [0.79-0.98] versus 0.78 [0.64-0.92]; p = 0.048). When applied to the sub-cohort of 25 patients in the Leiden cohort who were negative for anti-citrullinated peptide autoantibodies (ACPA), enhanced performance of the modified over the un-modified clinical prediction rule was maintained (AUROC [95% CI] = 0.82 [0.65-1] versus 0.70 [0.45-0.95], respectively), although the difference did not reach statistical significance in this smaller cohort. Conclusion We provide a proof of principle for the application of a B lymphocyte-derived epigenetic signature to enhance prediction of RA in UA patients using stored PBMCs. Further refinement of our pipeline represents a plausible means to expedite the diagnosis in undifferentiated RMDs and could offer pathophysiological insight. Disclosure N. Naamane: None. E. Niemantsverdriet: None. N. Thalayasingam: None. N. Nair: None. A.D. Clark: None. K. Murray: None. B. Hargreaves: None. L.N. Reynard: None. S. Eyre: None. A. Barton: None. A.H.M. van der Helm-van Mil: None. A.G. Pratt: None.