AI based on evolutionary computation yields algorithmic biomarker summary of a randomized rheumatoid arthritis clinical trial, accurately predicting individual patient outcomes, enabling precision medicine

Abstract

Background Producing transparent interpretable algorithms summarizing clinical trial outcomes to accurately predict individual patient’s responses would be a significant advance. We hypothesized that software designed to analyze biomedical data, based on evolutionary computation (EC), could produce summary algorithmic biomarkers from a clinical trial, predictive of individual responses to therapy. Methods and Findings A previously published randomized double-blind placebo controlled clinical trial was analyzed. Patients with active rheumatoid arthritis on a stable dose of methotrexate and naive to anti-tumor necrosis factor biologic therapy, were randomized to receive infliximab or placebo. The primary endpoint was synovial disease activity assessed by magnetic resonance imaging. Secondary endpoints included the Disease Activity Score 28 (DAS28). Baseline peripheral blood gene expression variable data were available for 59 patients, plus the treatment variable, infliximab or placebo, yielding a total of 52,379 baseline variables. The binary dependent variable for analysis was DAS28 response, defined by a decrease in DAS28 score of ≥1.2, at 14 weeks. At 14 weeks, 20 of the 30 patients receiving infliximab had responded, and ten of the 29 patients receiving placebo had responded. The software derived an algorithm, with 4 gene expression variables plus treatment assignment and 12 mathematical operations, that correctly predicted responders versus non-responders for all 59 patients with available gene expression data, giving 100% accuracy, 100% sensitivity and 100% specificity. We present the algorithm to provide transparency and to enable verification. Excluding the 4 gene expression variables, we then derived similarly predictive algorithms with 4 other gene expression variables. We hypothesized that the software could derive algorithms as predictors of treatment response to anti-tumor necrosis factor biologic therapy using just these 8 gene expression variables using previously published independent datasets from 6 rheumatoid arthritis studies. In each validation analysis the accuracy of the predictors we derived surpassed those previously reported by the original study authors. Conclusions and Relevance Software based on EC summarized the outcome of a clinical trial, with transparent biomarker algorithms correctly predicted the clinical outcome for all 59 RA patients. The biomarker variables were validated in 6 independent RA cohorts. This approach simplifies and expedites the development of algorithmic biomarkers accurately predicting individual treatment response, thereby enabling the deployment of precision medicine, and, in the future, providing a basis for dynamic labeling of prescription drugs. Original Trial Registration used for analysis: ClinicalTrials.gov registration:NCT01313520