Introduction
Disease course in ulcerative colitis (UC) is heterogeneous and difficult to predict at diagnosis. We aimed to develop and validate a microbiome-based classifier to identify newly diagnosed, treatment naïve UC patients at increased risk of an aggressive disease course.
Aims & Methods
Newly diagnosed, treatment naïve UC patients and healthy subjects were recruited from two independent cohorts: Swedish Inception Cohort (SIC-IBD)1 (n=167; clustering and model training) and Akershus University Hospital (AHUS), Norway (n=105; validation). Baseline fecal samples were collected prior to treatment initiation, and clinical data were recorded from inclusion through 12 months. Using the same criteria in both cohorts1. UC patients were classified at 12 months as having aggressive or indolent disease course; SIC-IBD and AHUS comprised 14 and 14 aggressive patients, 77 and 43 indolent patients, and 76 and 48 healthy subjects, respectively.
Baseline microbiome composition was assessed using the GA-map® Discovery panel comprising 174 bacteria-targeting DNA probes2 (Genetic Analysis AS, Oslo). Probes with consistently low signal intensity or a limited discriminatory value were excluded. A panel constituting 44 bacterial DNA probes were subsequently assembled, and all samples were reanalyzed.
Results
Multivariable PERMANOVA (Bray-Curtis) showed that the baseline microbiome composition was more strongly associated with the 12-months disease course than other clinical variables.
Prior to classifier training, clustering was applied separately within the indolent and healthy groups and the clinically aggressive group, generating three microbiome-derived risk classes: Low-Risk, Moderate-Risk, and High-Risk. Eight bacterial DNA probes showing the largest High- vs Low-Risk signal differences were selected to train a penalized multinomial logistic regression model (LASSO-GLM) using leave-one-out cross-validation. The model showed strong training performance, with sensitivities of 0.91, 0.80, and 1.00 for the Low-, Moderate-, and High-Risk classes, respectively. The model was subsequently frozen and applied to the validation dataset.
In the validation dataset, predicted risk classes were compared with the 12-months disease course outcome (Table 1). No patients with an aggressive disease course were classified as Low-Risk, corresponding to a false reassurance rate of 0%. A High-Risk microbiome profile was identified in 50% of patients with an aggressive disease course (7/14), while the remaining patients were classified as Moderate-Risk, indicating biological heterogeneity.
For clinical interpretation, ROC analyses were performed using the model’s continuous predicted probability for the High-Risk class. This yielded an AUC of 0.85 (95% CI 0.72–0.98) for aggressive versus indolent disease course and 0.94 (95% CI 0.84–1.00) for aggressive disease course versus healthy subjects.
Table 1. Microbiome-based classifier predicted risk class vs 12-month clinical outcome.
| | Microbiome based risk class |
| 12-months clinical outcome | Low-Risk | Moderate-Risk | High-Risk |
| Healthy (n=48) | 32 (67%) | 16 (33%) | 0 (0%) |
| Indolent course (n=43) | 12 (28%) | 29 (67%) | 2 (5%) |
| Aggressive course (n=14) | 0 (0%) | 7 (50%) | 7 (50%) |
Conclusion
We developed, and externally validated a microbiome classifier using baseline fecal microbial signatures that identifies a subgroup of treatment naïve UC patients at risk of developing an aggressive disease course. This highlights the potential clinical utility of this approach for early risk stratification and personalized treatment decisions at diagnosis.
References
1. Salomon B, Grännö O, Bergemalm D, Strid H, Carstens A, Hjortswang H, et al. Cohort profile: the Swedish Inception Cohort in inflammatory bowel disease (SIC-IBD). BMJ Open. 2025;15(5):e099218. doi:10.1136/bmjopen-2025-099218.
2. Casén C, Vebø HC, Sekelja M, Hegge FT, Karlsson MK, Ciemniejewska E, et al. Deviations in human gut microbiota: a novel diagnostic test for determining dysbiosis in patients with IBS or IBD. Aliment Pharmacol Ther. 2015;42(1):71-83. doi:10.1111/apt.13236.
Disclosure
G. T. Kirubakaran, K. Gravdal, T. Z. Hansen, and C. Casén are employees in Genetic Analysis. C. Casén owns shares in Genetic Analysis
