Unexpected Actors in Inflammatory Bowel Disease Revealed by Machine Learning from Whole-Blood Transcriptomic Data

Although big data from transcriptomic analyses have helped transform our understanding of inflammatory bowel disease (IBD), they remain underexploited. We hypothesized that the application of machine learning using lasso regression to transcriptomic data from IBD patients and controls can help identify previously overlooked genes. Transcriptomic data provided by Ostrowski et al. (ENA PRJEB28822) were subjected to a two-stage process of feature selection to discriminate between IBD and controls. First, a principal component analysis was used for dimensionality reduction. Second, the least absolute shrinkage and selection operator (lasso) regression was employed to identify genes potentially involved in the pathobiology of IBD. The study included data from 294 participants: 100 with ulcerative colitis (48 adults and 52 children), 99 with Crohn's disease (45 adults and 54 children), and 95 controls (46 adults and 49 children). IBD patients presented a wide range of disease severity. Lasso regression preceded by principal component analysis successfully selected interesting features in the IBD transcriptomic data and yielded 12 models. The models achieved high discriminatory value (range of the area under the receiver operating characteristic curve 0.61-0.95) and identified over 100 genes as potentially associated with IBD. PURA, GALNT14, and FCGR1A were the most consistently selected, highlighting the role of the cell cycle, glycosylation, and immunoglobulin binding. Several known IBD-related genes were among the results. The results included genes involved in the TGF-beta pathway, expressed in NK cells, and they were enriched in ontology terms related to immunity. Future IBD research should emphasize the TGF-beta pathway, immunoglobulins, NK cells, and the role of glycosylation.