Nikola Kotur1*, Sanja Dragašević Vučićević2, Uršula Prosenc Zmrzljak3, Saša Šterpin3, Katarina Krstajić1, Marina Jelovac1, Bojan Ristivojevic1, Branka Zukic1 and Biljana Stankovic1
1Institute of Molecular Genetics and Genetic Engineering, University of Belgrade, Serbia
2Medical Faculty, University of Belgrade
3Labena company
nikola.kotur [at] imgge.bg.ac.rs
Abstract
Glucocorticoids (GCs) are widely used for the induction of remission in inflammatory bowel disease (IBD), but many patients exhibit resistance or dependency. This highlights the need for biomarkers that can predict therapeutic response. Cytokines are central mediators of intestinal inflammation and may contribute to variability in GC responsiveness. Machine learning modeling such as penalized linear models that use multi-variable biological data could be particularly useful for selecting predictive features and identifying novel biomarkers of GC response.
Adult patients with active Crohn’s disease or ulcerative colitis initiating systemic prednisone therapy were prospectively recruited. Plasma samples collected before therapy were analyzed using a Bio-Plex Pro Human Cytokine 48-Plex panel. After quality control, 27 cytokines were retained for analysis. Patients were classified as GC good responders or GC poor responders, which include GC-resistant and GC-dependent patients.Penalized logistic regression models were applied to evaluate the predictive potential of cytokine and clinical variables.
Thirty-seven patients were included in the final analysis. Elevated baseline levels of macrophage migration inhibitory factor (MIF), IL-9, and GRO-α were significantly associated with poor GC response, while TNF-β, MIP-1β, SCF, FGF, and IL-16 showed borderline associations. Increased inflammatory cytokine burden, assessed as the sum of standardized cytokine values, was also associated with poor GC response. Principal component analysis showed cytokine profile differences between response groups. Penalized logistic regression models achieved the best predictive performance using a limited number of variables, particularly combinations including MIF and baseline clinical activity, with mean test accuracies up to approximately 73%.
Elevated pro-inflammatory cytokine profiles, particularly increased MIF, IL-9, and GRO-α, are associated with the poor GC response in IBD. Combining cytokine markers with the baseline clinical activity may support the development of predictive models for personalized GC therapy in IBD.
Keywords: IBD, cytokines, glucoocrticoids, machine learning
Acknowledgement: The authors acknowledge Labena d.o.o. (Ljubljana, Slovenia) for providing a grant and expertise to perform cytokine measurements.