1990: Artificial Intelligence Applied to Transcriptomics Profiling of Synovial Tissue Biopsies Accurately Predicts Rheumatoid Arthritis Patients Who Will Respond or Be Refractory to Standard Biological Treatments

Giorgio CASABURI, Todd Holscher and Ming-Chou Lee, Exagen, Inc., Vista, CA

Background/Purpose: In recent years, biological therapies have revolutionized treatments of Rheumatoid Arthritis (RA). However, about 40% of RA patients do not respond to given biologics (based on ACR50 response), with an alarming 15% showing resistant phenotypes to all available medications. To date, the biological mechanisms characterizing the patient population who will be resistant to standard biological treatments remain unclear, leaving healthcare providers with limited tools that are often not sufficient to anticipate a successful therapy. Synovial biopsies (SB) carry unique histopathological signatures associated with treatment response, prompting investigators to further study the complex heterogeneity of RA from gene expression extracted from this source. Here, we coupled targeted RNA-Seq sequencing derived from SB with artificial intelligence (AI). We developed three distinct machine learning algorithms to identify a set of biomarkers capable of predicting patients who will likely respond to standard biological treatments, including Tocilizumab (TOC) and Rituximab (RTX), or those who will be refractory to both.

Methods: A minimum of 6 SB samples per patient were collected from 112 difficult to treat RA patients who had previously failed conventional disease-modifying antirheumatic drugs, as well as subsequent anti-TNFα therapy. SB underwent full RNA sequencing. Outcome data was collected at 16 weeks to classify patients who failed TOC, RTX or both treatments. 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 who will respond to either TOC, RTX or be refractory to both based on CDAI50. Features ensemble techniques were applied to reduce the number of predictive biomarkers to an essential minimum without compromising overall performances.

Results: A set of unique biomarkers were identified to be highly predictive in discriminating patients that will respond to TOC, RTX , or be refractory to both. The average area under the ROC curve (AUC) for the blind hold-out (HO) portion of data was 0.86, 0.96, 0.92 for TOC, RTX, or refractory response statuses, respectively. The average HO Specificity was 0.82, 0.91 and 0.83, while the average HO Sensitivity was 0.86, 1, 0.93 for TOC, RTX and refractory statuses, respectively. Patient probability distributions derived from the identified biomarkers showed a clear separation of responder vs non-responder patients.

Conclusion: Presently, the clinical response to treatment in RA is still difficult to predict. The advent of meta-omics coupled with advanced AI methodologies is establishing a new patient-oriented therapeutic approach increasingly leading to more sophisticated personalized medicine. Combining a test harnessing the information gained from SB, meta-omics profiling, and large-scale AI model deployment to achieve high predictive value for RA biologic response can improve effectiveness of standard treatment protocols, as well as dramatically reduce healthcare costs for RA.

Disclosures: G. CASABURI, Exagen Inc.; T. Holscher, Exagen Inc.; M. Lee, None.