Tu223 - Anti-tnf Treatment Response Classifier for Inflammatory Bowel Disease Patients

Hamid Mattoo – Immune Mediated Disease Lead, Precision Medicine and Computational Biology, Sanofi; Annie Kruger – Clinical Research Director, Early Clinical Development Inflammation & Immunology, Sanofi; Lu Zhang – Principal Scientist, Precision Medicine and Computational Biology, Sanofi; George Fu – Principal Scientist, Precision Medicine and Computational Biology, Sanofi; Crystal Rawlings – Senior Scientist, Precision Medicine and Computational Biology, Sanofi; Heming Xing – Senior Principal Scientist, Precision Medicine and Computational Biology, Sanofi; Franck Rapaport – Director of Precision Medicine and Computational Biology Data Science, Precision Medicine and Computational Biology, Sanofi; Corneliu Bodea – Associate Director, Precision Medicine and Computational Biology, Sanofi; Emanuele de Rinaldis – Global Head of Precision Medicine and Computational Biology, Precision Medicine and Computational Biology, Sanofi

Abstract Text:

Background: Inflammatory bowel disease (IBD) is a chronic inflammatory autoimmune disease which includes ulcerative colitis (UC) and Crohn’s disease (CD). It is characterized by chronic inflammation of the gastrointestinal tract and rectal bleeding. Anti-TNF-α Ab therapy is a frontline targeted therapy for IBD but it is not clear why some patients fail to respond or develop resistance after a short time. Clinical biomarkers capable of predicting anti-TNF-α Ab response at baseline can help guide therapy decisions, reduce treatment costs and lead to better treatment outcomes. Here we developed a robust predictive model, characterized by a gene signature, capable of discriminating anti-TNF-α response in IBD patients at baseline.

Methods: We collected public IBD transcriptomic datasets containing Mayo score-based response labels. Following normalization, baseline colonic UC samples were used to train and evaluate an Elastic Net classification model employing a leave-one-out, nested cross-validation strategy. The model was evaluated in independent cohorts consisting of UC or CD baseline colonic samples from different platforms. Finally, the resulting gene classifier was compared to a previously published classifier.

Results: Based on our gene signature, cross-validated predictive model accuracy on the UC training set was 84%. Our model also classified independent colonic baseline samples, interrogated on different platforms, with an accuracy of 86% (UC) and 83% (CD). Lastly, we used principal component analysis to compare our gene-signature with a published signature to evaluate their ability to separate responders from non-responders. In various testing datasets, our signature separated samples as well, or better, than the previously published signature.