0935: Targeted Synovial Tissue RNA-Seq Coupled with Artificial Intelligence Accurately Predicts Early Rheumatoid Arthritis Patients Likely to Respond to CsDMARDs, Enriching CsDMARDs Response Rates and Enabling Early Identification of Patients Requiring Subsequent Biological Therapy

Giorgio CASABURI, Tyler O'Malley, Todd Holscher and Ming-Chou Lee, Exagen, Inc., Vista, CA

Background/Purpose: Rheumatoid arthritis (RA) is a chronic, inflammatory, systemic autoimmune disease, affecting the joints with varying severity. RA affects approximately 0.5/1% of adults in the United States. Currently, conventional synthetic disease modifying anti-rheumatic drugs (csDMARDs) are the first line therapy for RA. However, about 40% of patients do not respond to csDMARDs and will require subsequent biological therapy. The identification at disease onset of patients who are likely to respond to csDMARDs remains a major unmet need, adding millions of dollars in medical expenses to the US healthcare system annually, as well as negatively impacting overall patients’ life. Here, we present a new generation of machine learning algorithms that can accurately identify RA treatment-naïve patients likely to respond to csDMARDs, informing initial RA treatment selection to maximize the potential of csDMARDs and promote early biologic intervention where appropriate.

Methods: A minimum of six synovial biopsies samples per patient were collected from 95 patients at baseline, followed by bulk RNA sequencing. Outcome data was collected at 3 months to classify patients who needed to switch to biological therapy and used for model training. After quality filtering, RNA-Seq annotation, and normalization, the resulting dataset went through an automated machine learning pipeline, capable of concurrently running hundreds of different cross-validated models. Models were trained to predict patients that would respond to csDMARDs. Feature selection and reduction techniques were applied to reduce the number of required predictive biomarkers, while maintaining elevated accuracy. A multi-fold generalizability test was performed on the final set of biomarkers by means of permutational hold out (HO) predictions, e.g., the portion of data that is always blind during training and validation modeling.

Results: A set of unique biomarkers were identified to be highly predictive in discriminating patients that responded or not to csDMARDs at 3 months after sample collection. The average area under the curve (AUC) for the blind HO portion of data was 0.85, with a sensitivity of 0.92 and a specificity of 0.69. Patient probabilities distributions derived from selected biomarkers showed a net separation of responder vs non-responder patients to csDMARDs.

Conclusion: The ability to identify the most appropriate treatment for early RA patients by targeted sequencing in the synovium offers a new opportunity to personalize therapeutic interventions to the patients most in need. This study reinforces the need to change current therapeutic procedures with the help of AI and Big Data to predict during therapeutical decisions the need of biological therapies at disease onset. This approach is likely to have a major impact on disease control, remission, overall healthcare cost, and especially patients’ quality of life.

Disclosures: G. CASABURI, Exagen Inc.; T. O'Malley, Exagen; T. Holscher, Exagen Inc.; M. Lee, None.