A SUBSET OF ALGORITHMIC BIOMARKERS DERIVED USING QUANTITATIVE AI BASED ON EVOLUTIONARY COMPUTATION THAT PREDICT TNF RESPONSE IN RHEUMATOID ARTHRITIS ALSO HAVE 100% ACCURACY PREDICTING TNF RESPONSE IN ULCERATIVE COLITIS, CROHN’S ILEITIS AND CROHN’S CO

Abstract Biomarkers that accurately predict responses to IBD therapy in individual patients would significantly optimize clinical outcomes. Biology and disease are complex systems from which actionable insights can be derived using algorithms. We hypothesized that algorithmic gene expression biomarkers, highly predictive of individual responses to TNF therapy in rheumatoid arthritis (RA), would also enable derivation of algorithmic biomarkers predictive of individual responses to TNF therapy in both ulcerative colitis (UC) and Crohn’s disease (CD). We assumed that similar molecular networks underlie the response to TNF therapy in different diseases, and these could be characterized algorithmically as predictive biomarkers. Applying quantitative AI, based on evolutionary computation to baseline peripheral blood gene expression data, we previously identified 8 variables that yielded algorithms using subsets of 4 variables that had 100% accuracy in predicting individual responses to TNF therapy in RA patients. Following discovery from a placebo-controlled clinical trial, these 8 variables were validated in 6 independent data sets and are in development as a clinical test: https://doi.org/10.12688/f1000research.164015.1. We then applied the 8 RA variables to two previously published data sets of baseline mucosal gene expression obtained prior to initiation of infliximab therapy in IBD patients. The first cohort included 46 UC patients (20 responders and 26 non-responders) and 6 polypectomy controls. The second cohort included 61 patients: 24 UC (8 responders and 16 non-responders), 19 CD colitis (12 responders and 7 non-responders), and 18 CD ileitis (8 responders and 10 non-responders), with 12 polypectomy controls: PMID: 19700435 and PMID: 19956723. A subset of 2 variables from the 8 previously identified in RA yielded predictive algorithms of TNF response in UC, CD colitis and CD ileitis. The software-derived algorithms, containing these 2 variables (C1orf105 and PTPRC) plus dose, correctly predicted responders versus non-responders for all patients in both data sets, giving 100% accuracy, sensitivity and specificity. This accuracy surpassed that reported in the original publications. We present illustrative algorithms to provide transparency and to enable independent verification. These findings suggest that the RA test in development will also have clinical utility in IBD. This approach may be useful for developing response predictors for other therapies, thereby realizing the vision of precision medicine to optimize outcomes. Algorithmic biomarkers might also be the basis for dynamic product labels accurately predicting therapeutic responses for individual patients. The two components of the algorithms, C1orf105 and PTPRC, may have important roles in pathogenesis of IBD that have not been previously apparent, including consideration as novel therapeutic targets.