706 Predictive modeling of patient response to JAK/STAT inhibitors and dynamic patient-matching

Abstract

Alopecia areata (AA) is an autoimmune disease characterized by the immune-mediated destruction of hair follicles in scalp skin, resulting in partial or total hair loss. Ongoing research has identified JAK/STAT and the T cell costimulatory signaling pathways as key druggable targets, and several clinical trials have been conducted. Although efficacy signals have been impressive, there are subsets of patients in each trial that do not respond to treatment. This variable response correlates with distinct molecular mechanisms of action (MoA) that define each tested JAK/STAT inhibitor. Using reverse-engineered regulatory networks (ARACNe networks) and machine learning algorithms, we defined molecular predictors of drug response for four compounds used in AA clinical trials: tofacitinib (pan-JAK inhibitor), ruxolitinib (JAK1/2 inhibitor), abatacept (CTLA4-Ig), and intralesional triamcinolone, through the transcriptional activity of ten candidate master regulators including RREB1, CAMKK2, DACH1, and HLF. The activity of these candidate regulators produced four non-overlapping (FDR<0.05) MoA gene signatures. In patients, non-responder status was defined as a lack of improvement in SALT score (delta<20%) at the end of a corresponding clinical trial in combination with lack of resolution of the ALADIN score. Patients were binned into nonresponder and nonresponder status based on these criteria for analysis. For each nonresponder patient, these signatures were percentile-ranked for overlap with each of the drug MoAs to predict the overall efficacy of each compound, producing a preliminary tool for assessing overall treatment efficacy prior to administering treatment.