P537 Analysis of immunological VAriables in ex vivo Tofacitinib-treated human biopsies from Active ulcerative colitis patients to predict clinical Response (the AVATAR study)

Abstract

Abstract Background Ulcerative colitis (UC) is characterized by chronically relapsing inflammation of the colonic mucosa. Tofacitinib (TOFA) is a paninhibitor of the JAK kinases, a family of intracellular kinases in volved in the signal transduction pathway of multiple inflammatory signals. Despite the efficacy shown by TOFA in inducing and maintaining remission in moderately-to-severely active UC, rates of primary non response remains high. We propose a new method to identify predictors of responsiveness to TOFA based on "ex-vivo" challenge of organ cultures from biopsies collected from candidate patients. Methods Baseline biopsies from inflamed areas of UC patients with active disease and candidate to receive TOFA 10mg b.i.d. for 8 weeks, were used for "ex-vivo" organ culture and stimulated with TOFA 20μM and 200nM or left untreated. After 24 hours, proteomic analysis was performed and quantitative expression of 92 inflammation-related human proteins (Olink® Human Inflammation panel) was correlated with week 8 clinical response defined by Modified Mayo Score (MMS). Intestinal biopsies were also collected at baseline and at 8 weeks for proteomic analysis. To decipher the nature of TOFA's effects, both unsupervised and supervised machine learning (ML) methods were employed (sklearn module, Python). The K-means method was used for unsupervised learning followed by a supervised ML approach to classify week 8 responders and non-responders. Data were transformed in into a second-order polynomial format and logistic regression was used for modelling. Results Between Jannuary 2022 and January 2023, 17 UC patients were enrolled (Table 1). At week 8, 13/17 (76.5%) patients treated with TOFA showed clinical response. Unsupervised ML, utilizing proteomic data from organ cultures stimulated with TOFA 200nM, achieved a 90% success rate in predicting a positive clinical response. This prediction was based on forming two distinct clusters, with significant contributions from key variables such as CASP-8, CCL19, CCL20, CD5, CXCL9, EN-RAGE, GDNF, HGF, IL-18R1, MCP-1, MCP-2, MCP-4, MMP-10, OPG, PD-L1, STAMBP, TNFRSF9, TRANCE, TWEAK, and VEGFA. In our supervised ML approach, logistic regression was employed with 30 shuffle splits to refine the predictive model. We discovered that transforming the dataset into a second-order polynomial format significantly enhanced the model's performance. The most effective logistic regression model achieved a mean precision score of 0.82 (Fig.1). Conclusion Our predictive model demonstrates a high level of precision in identifying positive treatment responses, emphasizing the potential of further dataset expansion and the use of advanced machine-learning to predict response in TOFA-treated UC patients.