Till Örebro universitet

Background: The disease course of patients with newly diagnosed ulcerative colitis (UC) is highly uncertain, and there is a lack of validated prognostic biomarkers that could aid in clinical decision making.

Methods: Newly diagnosed, mainly treatment naïve patients with UC from three large inception cohorts were used to develop and validate a serum proteomics-based risk score for prognostication of disease course during the first year from diagnosis. In the discovery cohort (n = 161) and validation cohort 1 (n = 186) an aggressive disease course was defined as the presence of any IBD-related surgery, hospital admission for active disease, treatment refractoriness towards targeted therapies (i.e. biologics, JAK-inhibitors or S1P receptor modulators), and >2 courses or high cumulative doses of systemic corticosteroids. In validation cohort 2 (n = 120), an aggressive disease course was defined as the need for a biologic, ciclosporin or surgery. 178 proteins were measured on Olink platforms, and a machine learning algorithm (i.e. regularised regression) was applied to the discovery cohort to develop an UC risk score comprising 23 proteins. The performance of the UC risk score was assessed in the two external validation cohorts. For validation cohort 2, a condensed version of the UC risk score was applied, as only 14 of the original 23 proteins were available. Cox regression estimated hazard ratios (HR) for the association between the UC risk score at diagnosis, time to escalation to targeted therapy (validation cohort I) and time to the defining episode of an aggressive disease course (validation cohort II).

Results: Based on univariate analyses, we identified 59 proteins associated with an aggressive disease course in the discovery cohort (PFDR <0.10; Figure 1A). Twenty could be validated in validation cohort 1, and nine remained in validation cohort 2 (PFDR <0.10; Figure 1B-C).

In the discovery cohort, the machine learning model showed a high predictive capacity, with an area under the curve (AUC) of 0.81 (Figure 1D) and was numerically superior compared with a clinical model comprising sex, age and CRP (AUC = 0.72). Next, the performance of the UC risk score was confirmed in the two external validation cohorts, displaying AUC:s of 0.77 (Figure 1E-F). Patients with a higher UC risk score at diagnosis had increased risk of initiating targeted therapy (HR 4.26, 95% CI 1.91–9.49; Figure 2A). The HR for having an aggressive disease course was 9.12 (95% CI 3.04–27.3; Figure 2B).

Conclusion: We develop a UC risk score for prognostication of disease course and confirmed its high predictive performance by external validation in two independent cohorts. The risk score is a promising tool for quantifying the risk of having an aggressive disease course in UC.