A BASELINE GENE EXPRESSION-BASED PROGNOSTIC FOR ANTI-TNFα THERAPY RESPONSE IN PATIENTS WITH INFLAMMATORY BOWEL DISEASE

Background Although anti-TNFα therapies have revolutionized the management and care of IBD, their administration and usage remain suboptimal because 1) over 50% of patients do not have a lasting therapeutic response, 2) they increase risk of infections, liver problems, arthritis, and lymphoma, and 3) they are expensive. With approximately 1.6 million people suffering from IBD in the US and global prevalence of IBD on the rise, a predictive test for anti-TNFα response would greatly improve the efficacy and cost-to-benefit ratio of these biologics. Methods We hypothesized that a multicohort analysis of the publicly available IBD gene expression datasets would yield a robust set of mRNAs for distinguishing anti-TNFα responders vs non-responders in the IBD patient population prior to treatment. We identified 5 datasets (n = 160) where whole-genome transcriptomic data was derived from colonic mucosal biopsies of IBD patients who were then subjected to anti-TNFα therapy and subsequently adjudicated for response. We used the MetaIntegrator framework which leverages a leave-one-study-out cross-validation technique in conjunction with effect size and FDR adjusted p-value to identify significant differentially expressed (DE) genes associated with a patient’s predisposition to a response outcome. DE genes were subjected to a greedy forward search to derive a parsimonious gene signature for a response score (geometric mean of the expression level for all positive mRNAs minus the geometric mean of the expression level of all negative mRNAs, multiplied by the ratio of counts of positive to negative genes). Area under the receiver operating characteristic curve (AUC) was subsequently calculated in a leave-one-study-out manner to assess discriminatory performance. Results We first identified 170 genes that were present in at least 40% of cohorts and significantly differentially expressed between responders and non-responders with effect size > 0.8 and q value < 0.1. A score based on these genes predicts responder vs non-responder across the 5 discovery cohorts with AUC of 0.82. Optimizing the variables with a greedy forward search algorithm allowed us to downselect to 7 genes from the set, and a score based on this parsimonious set of 7 genes improved the discriminatory performance to an AUC of 0.87. Choosing a high sensitivity (90%) for a rule-in scenario, the score had moderate specificity (60%); alternatively choosing a high specificity (90%) for a rule-out scenario, the score still had a good sensitivity (80%). Conclusions These initial findings suggest that there is a strong signal for predicting anti-TNFα response in colonic biopsies. In particular, we showed using the leave-one-study-out approach that a predictive signature using mRNA can be generalizable (works in independent cohorts). These initial results warrant further investigation.