A systematic prediction of adverse drug reactions using pre-clinical drug characteristics and spontaneous reports

Adverse drug reactions (ADRs) are a major global health concern accounting for more than two million injuries, hospitalization and deaths each year in the U.S. Alone. A reduction in both the harm to patients and cost can be archived if at prescription time, effective and accurate methods are available to predict the likelihood for a patient to develop known or potentially new adverse reactions. This can be based on known properties of drugs and patient characteristics. The detection or assessment of potential ADRs is traditionally done during the early stages of drug development. However, despite the methodological rigor of clinical trials, it is generally not possible to identify all ADRs of a drug primarily due to cost and efficiency. The size and characteristics of patient population, drug doses and duration of use, and other realistic variables frequently observed at the post-marketing phase can be impossible to model at the clinical trial phase. Thus, it is important to incorporate information on drugs observed at the post-marketing phase for more accurate identification of ADRs. This work presents a systematic and structured predictive model for ADRs generated from pre-clinical characteristics of drugs and spontaneous reports of ADRs in a distributed high performance computing (HPC) framework. The presented framework improves predictive accuracy by making use of a recent computationally efficient Bayesian graphical ensemble learning technique that incorporates hidden information transferred from distributed heterogeneous spontaneous reporting databases to improve accuracy. Implemented on HPC cloud machines, the graphical ensemble method outperformed other compared methods on a total of 800 known side-effects in terms of AUC and G-mean.