Abstract Background We aimed to build a serum proteomics-based model to predict primary nonresponse (PNR) to infliximab (IFX) in pediatric colonic inflammatory bowel disease, with early proactive therapeutic drug monitoring. Methods Children in the prospective Canadian Children IBD Network with ulcerative colitis (UC), inflammatory bowel disease unclassified (IBD-U), or colonic Crohn’s disease (CD) with serum pre-IFX were eligible. We defined PNR as IFX cessation plus surgery/drug switch within 6 months. We compared clinical features between groups (Mann Whitney U, chi-square test). We measured serum proteins with Olink Inflammation/Immune Response panels. We built a regularized regression (generalized linear model GLM) machine learning model and compared its performance with other models with 10-fold cross-validation repeated 10 times (receiver-operating characteristic/precision-recall curves, predictive score separation). We ranked proteomic features by SHAP (SHapley Additive exPlanations) analysis. We hypothesized that treatment-naïve serum would be more informative than treatment-exposed serum. Results We included 96 patients: 71 UC/IBD-U (23 nonresponders), 42 treatment-naïve (12 nonresponders); and 25 CD, 19 treatment-naïve. Pre–third and pre–fourth dose serum infliximab levels were similar and robust (10 µg/mL) in primary nonresponders and responders. Predictive performance was superior for diagnostic, treatment-naïve samples; the GLM showed good ability to separate primary nonresponders and responders. The GLM model on treatment-naïve serum (area under the curve ∼0.75) had better specificity to predict responders and included 21 proteins, with CSF1 and ITM2A top ranked. UC/IBD-U responders more often were steroid refractory and received infliximab as first maintenance. Conclusions A serum proteomics linear model on treatment-naïve serum best predicted PNR. Findings require external validation but suggest that the diagnostic/pretreatment window may be key to understanding biology central to effective drug sequencing.
Sathya et al. (Thu,) studied this question.