Background/Purpose: Current rheumatoid arthritis (RA) treatments, including TNFa inhibitors (TNFi) and JAK inhibitors (JAKi), have transformed the management of RA by controlling symptoms and slowing disease progression. However, not all patients will respond to a specific therapy, resulting in patients failing multiple therapies before finding an effective treatment.  Here, we present the development of a machine learning (ML) classifier leveraging active chromatin cell-free DNA (cfDNA_ac) signals in the blood to differentiate RA patients likely to respond to TNFi or JAKi therapy from those who will not.  

Methods: A total of 48 RA participants undergoing a switch in therapy to either TNFi (n=19) or JAKi (n=29) were selected from the CorEvitas BIO-100 registry based on response class. The participants, aged 20-80 years (median=57), consisted of 89.5% females, with 89.6% identifying as White. The cohort included 21 responders and 27 non-responders based on Clinical Disease Activity Index (CDAI) scores collected at baseline and six months post-therapy initiation. Response was defined according to minimally clinically important differences established for CDAI, requiring improvement >12 units for patients starting in high disease activity, and >6 units for patients starting in moderate disease activity. Baseline plasma samples were processed using a novel active chromatin capture assay that enables non-invasive access to tissue- and disease-specific molecular signatures. Two separate ML classifiers were developed using a 5-repeat 10-fold cross-validation approach for training and testing— an overall therapy response classifier and a drug class-specific therapy response classifier.

Results: Approximately 650 features were identified during feature selection and subsequently used to develop a ML model that differentiates responders from non-responders. In a combined cohort of patients receiving either TNFi or JAKi treatments, the therapy response classifier achieved a sensitivity of 80% at a specificity of 85%. When comparing features that discriminated responders from non-responders, cell-type enrichment analysis identified signatures of cell types known to be associated with the joint synovium including fibroblasts and immune cells such as antigen presenting cells, macrophages, and monocytes. Furthermore, a separate response classifier developed specifically for JAKi exhibited a sensitivity of 80% and a specificity of 83% in identifying JAKi responders.

Conclusion: Using this novel active chromatin capture method, we developed two classifiers— one capable of predicting responders from non-responders, and another capable of predicting outcomes within a therapy class. Cell-type enrichment analysis established a link between the classifier features and tissue-specific RA pathobiology, offering insights into the underlying cellular mechanisms driving treatment response. Further efforts to expand the cohort are underway to develop a more robust classifier that captures the inherent heterogeneity of RA and delivers a positive prediction of drug-class specific therapy response for multiple drug classes.

« Back to ACR Convergence 2024

ACR Meeting Abstracts - https://acrabstracts.org/abstract/a-novel-blood-based-assay-that-predicts-clinical-response-to-tnf%ce%b1-inhibitors-or-janus-kinase-inhibitors-in-patients-with-rheumatoid-arthritis/