Machine Learning-Driven Raman Spectroscopy, Validated by Metabolomics, Identifies Novel Non-Invasive Biomarkers of Barrier Integrity and Outcomes in Inflammatory Bowel Disease

Background Intestinal barrier healing is a key therapeutic target in inflammatory bowel disease (IBD), yet its assessment remains challenging. Raman spectroscopy (RS) provides a label-free molecular fingerprint reflecting metabolic signatures linked to barrier status. We applied machine-learning (ML) models to plasma RS to identify non-invasive biomarkers of epithelial and vascular barrier integrity and major adverse outcomes (MAOs). RS-derived molecular profiles were confirmed at tissue level and validated by plasma metabolomics. Methods Seventy-eight individuals [42 Crohn’s disease (CD), 28 ulcerative colitis (UC), and 8 healthy controls] were prospectively enrolled. Epithelial and vascular barrier integrity were assessed using endocytoscopy (ECS) or probe-based confocal laser endomicroscopy (pCLE), according to our newly validated scores.1 Plasma RS captured non-invasive molecular fingerprints; tissue RS on advanced-endoscopy-guided biopsies provided confirmation. Untargeted plasma mass-spectrometry metabolomics provided biochemical validation, using Spearman correlation (Figure 1). Six-month MAOs were recorded. ML models based on LASSO logistic regression were applied to RS and metabolomics datasets to identify key predictors of barrier integrity and MAOs. Results Barrier healing was observed in 23/42 patients (55%) by ECS and in 9/36 (25%) by pCLE. Thirty-one of 78 (40%) experienced MAOs. ML applied to RS and metabolomics identified the most relevant predictors of barrier healing and MAOs with good accuracy (Table 1). RS peaks associated with protein and lipid metabolism were most predictive of epithelial impairment, significantly correlating (r≈0.3; p < 0.05) with L-ergothioneine, N-acetylornithine and LPE (22:6). Vascular barrier dysfunction demonstrated a complementary lipid–oxidative signature, showing significant correlation (r≈0.3-0.4; p < 0.05) with linoleamide, acyl-glycines and the endothelial stress mediator TMAO. Disease-stratified analyses revealed distinct biological signature: CD showed additional kynurenine pathway alterations alongside pronounced acyl-glycines disturbances, while UC displayed a clearer epithelial profile with indole-derived tryptophan metabolites. Notably, the same RS-defined molecular profiles -oxidative-stress/lipid-remodelling pathways in the overall cohort, kynurenine signature in CD and indole-derived pathways in UC- also predicted MAOs. Conclusion ML-enabled plasma RS identified non-invasive biomarkers predictive of intestinal barrier impairment and clinical outcomes in IBD, validated by tissue RS and plasma metabolomics. These multimodal fingerprints can stratify barrier endotypes and predict clinical trajectories, driving next-generation precision management in IBD.