P471 A Transcriptome-dependent Prognostic Model of Response in Patients with Ulcerative Colitis

Johnson, T.(1);Reinisch, W.(2);Steere, B.(3);Milch, C.(4);Harris, C.(3);Morris, N.(3);D'Haens, G.(5);Huang, K.(1);Krishnan, V.(3)*;

(1)Indiana University, Clinical and Translational Sciences Institute, Indianapolis, United States;(2)Medical University of Vienna, Clinical Department for Gastroenterology and Hepatology, Vienna, Austria;(3)Eli Lilly and Company, Immunology, Indianapolis, United States;(4)Eli Lilly and Company, Former affiliation, Indianapolis, United States;(5)Amsterdam University Medical Centers, Department of Gastroenterology, Amsterdam, The Netherlands;


Mirikizumab (miri), a p19-directed IL-23 antibody, showed efficacy in ulcerative colitis (UC)1. Miri modulates transcripts associated with disease activity and anti-TNF resistance in UC and increases expression of transcripts associated with healthy mucosa2. Here, we evaluate the transcription profile of patients with UC who respond to placebo (PBO, 20.6%) or miri 200mg treatment (59.7%) versus non-responders and build a model that can predict response in patients with UC.


In the phase 2 AMAC trial (NCT02589665), patients were randomized 1:1:1:1 to receive intravenous PBO (N=63), 50mg miri (N=63), or 200mg miri (n=62) with possible exposure-based dose increases or fixed 600mg miri (n=61) at Week (W)0, 4 and 8. Colonic biopsies were collected at baseline, W12, and W52. An Affymetrix HTA2.0 microarray measured gene expression. Limma comparing values at baseline and W12 determined differentially expressed genes (DEGs). DEGs maintaining their W12 expression level through W52 among responders in either the PBO and miri 200mg arms were designated as similarly expressed genes (DEGSEGs).

XGBoost models3 were trained to predict clinical response (decrease in 9-point Mayo subscore [rectal bleeding, stool frequency, endoscopy] of ≥2 points and ≥35% from baseline [BL] with either a decrease of RB subscore of ≥1 or an RB subscore of 0 or 1) using baseline levels of DEGSEGs and TNF biomarkers. Leave-one-out cross validation of the models using AMAC data and validation experiments by training on the PROTECT data4 then predicting in AMAC were used to evaluate our models.


Miri 200mg responders were associated with 84 DEGSEGs, PBO responders were associated with 26 DEGSEGs, and 21 DEGSEGs were common to both. When using the PROTECT data as the training data and AMAC data as the validation cohort (Figure A), we were able to achieve high ROC-AUC and PR-AUCs using the top 36 (FC > 1.5) miri DEGSEGs (PR-AUC=0.78), TNF biomarkers (PR-AUC=0.91), or combined (PR-AUC=0.93) gene sets as the features (Figure B-E) and found that the response prediction was significantly higher in AMAC responders vs. non-responders (Figure F).


Baseline transcript features were predictive of clinical response in patients with UC treated with miri 200mg and our model distinguished responders from non-responders. Future studies may further validate this model.

1 Sandborn WJ, et al. Gastroenterology. 2020;158(3):537-549.e10
2 Steere B, et al. J Crohns Colitis. 2020;S103-S104
3 Chen T, et al. Proceedings of the 22nd ACM SIGKDD International Conference. 2016;7285-794
4 Hyams JS, et al. Lancet Gastroenterol Hepatol. 2017;2(12):855-868