Search in the Abstract Database

Abstracts Search 2019

P836 The predictive role of gut microbiota in treatment response to vedolizumab and ustekinumab in inflammatory bowel disease

C. Caenepeel*1, S. Vieira-Silva2, B. Verstockt1,3, M. Ferrante1,3, J. Raes2, S. Vermeire1,3

1KU Leuven, TARGID, Leuven, Belgium, 2Rega Institute for Medical Research, Microbiology and Immunology, Leuven, Belgium, 3University hospitals Leuven, Gastroenterology and Hepatology, Leuven, Belgium

Background

The faecal microbiota is evolving as a useful predictive and diagnostic biomarker for IBD in the development of personalised medicine. We here investigated whether the faecal microbiota aids in predicting therapeutic response to vedolizumab (VDZ) or ustekinumab (UST) in Crohn’s disease (CD) and ulcerative colitis (UC).

Methods

Faecal samples of 116 patients with IBD, treated with UST (n = 68 CD) or VDZ (n = 30 for CD, 18 for UC) with endoscopic active disease were collected prior to biological therapy. Quantitative microbiota phylogenetic profiling was conducted by combining 16S rRNA gene sequencing and microbial loads determination by flow cytometry.

Endoscopic response in the UST cohort was defined as a 50% decrease in SES-CD score at Week 24. Remission in the VDZ cohort was defined as an endoscopic Mayo-subscore of ≤1 at Week 14 in UC and absence of endoscopic ulcera at Week 24 in CD.

Multi-variate hyperbolic tangent neural network models (JMP) were trained to predict treatment response based on features describing the baseline faecal microbiota, clinical data (age, sex, BMI, diagnosis, disease duration and smoking) and biomarkers (CRP, albumin, haemoglobin and faecal calprotectin) or the combination. Microbiota features comprised enterotypes and quantitative abundances of taxa significantly (p < 0.1) correlated with outcome. The cohorts were split into training (2/3) and validation sets (1/3).

Results

Ten (14.7%) UST and 27 (56.2%) VDZ patients showed endoscopic response (UST) or remission (VDZ). 13 genera correlated with treatment outcome in the VDZ cohort and 14 in the UST cohort, with 3 overlapping. Neural networks were trained to predict treatment response in VDZ and UST (Figure 1) , based on clinical features and biomarkers, microbiota features, or both.

Figure 1: Receiver-operating characteristic curves of the different neural network trained for treatment response prediction for VDZ and UST

For VDZ treatment response prediction, all models had reliable training (AUC=[0.71−0.87]; sensitivity=[0.62–0.88], specificity=[0.55–0.85]), but the combined model had the best validation performance (misclassification rate=31%, N = 17). Similarly, UST response prediction was best with the combined model(training AUC=0.86, sensitivity=0.88, specificity=0.33, with a validation misclassification rate of 4% (N = 23).

Conclusion

Our analyses do show that quantitative faecal microbiota profiling is helpful in predicting therapeutic outcome and provides valuable additional information beyond clinical features and biomarkers. Nevertheless, these predictive models were trained on still relatively small cohorts, and therefore further validation in preferably large prospective randomised cohorts is needed.