Inflammatory bowel disease (IBD), encompassing Crohn’s disease (CD) and ulcerative colitis (UC), is a chronic immune-mediated condition with a heterogeneous clinical course that is increasingly being explored through advances in omics, particularly metabolomics. This study investigates whether untargeted serum metabolite profiles, acquired via LC-MS processing, are associated with IBD severity in 183 patients sampled two years before to one year after diagnosis. Unsupervised clustering was applied to non-linear metabolite projections, enabling a metabolite-based approach followed by exploratory machine learning on the resulting clusters. More specifically, five machine learning models were evaluated to identify associations between metabolite clusters and disease severity, with Random Forest and Support Vector Classifier showing the best performance for CD and UC, respectively. Notably, clusters formed around chemically homogeneous groups, with molecules such as hippurate and indole-propionic acid linked to non-severe CD cases, whereas glycerophosphocholines were associated with severe CD but non-severe UC. Our findings underscore the need for subtype-specific modeling approaches and demonstrate that well-designed metabolite-based analyses in IBD can serve as valuable tools for uncovering or validating severity-related patterns, particularly when combined with mechanistic and clinical interpretation.

Co-authors: Malene Revsbech Christiansen, Marie Vibeke Vestergaard, Tine Jess, Filip Ottosson