Discovering new drug-drug interactions using data science

Commonly prescribed small molecule drugs can have net-positive and well-understood safety profiles when prescribed individually, but unexpected consequences when taken at the same time. Detection of these drug-drug interactions (DDIs) continues to be a critical and unmet area of translational research. The Centers for Disease Control and Prevention (CDC) estimate that one third of Americans are concurrently taking two or more prescription drugs, and DDIs are estimated to be responsible for 17% of all drug adverse events. The consequences of DDIs can be relatively minor (headache, skin rash) or much more severe (bleeding, liver toxicity). At a cellular level, DDIs can occur as a result of both drugs competing for metabolism (known as pharmacokinetic interactions) or targeting the same protein target or biological pathway (pharmacodynamic interactions). Clinical trials typically focus on the effects of individual drugs, leaving DDIs to usually be discovered only after the drugs have been approved. One of the most carefully studied drug adverse events is long QT syndrome (LQTS), an unexpected change in the heart's electrical activity that can lead to a potentially fatal ventricular tachycardia known as torsades de pointes (TdP). Some patients have genetic mutations that lead to congenital forms of LQTS, while drug-induced LQTS typically occurs via block of the hERG potassium channel (KCNH2) responsible for ventricular repolarization. After a number of high profile drugs were withdrawn from the market due to discovered risk of TdP, the FDA issued guidelines so that pharmaceutical companies could anticipate and test for this side effect before a new drug is approved. These recommendations have helped prevent new QT-prolonging drugs from entering the market, but nonetheless over 180 approved drugs have been associated with drug-induced LQTS. While information on individual QT-prolonging drugs is thus readily available to clinicians, little has remained known about DDIs (QT-DDIs). There are many more commonly prescribed drugs that are safe when given individually but could increase TdP risk when administered together. This troubling situation is compounded by the fact that traditional post-market surveillance algorithms are poorly equipped to sensitively and specifically detect DDIs. Data science – the application of rigorous analytical methods to large datasets – offers an opportunity for predicting previously unknown QT-DDIs. Some biomedical datasets (such as drug-target binding affinities and experiments to determine protein-protein interactions) have been collected explicitly for research, while other valuable datasets (such as electronic health records) were initially recorded for billing purposes. Each data modality has its own important set of advantages and disadvantages, and integrative data science approaches can incorporate multiple types of data to help account for these limitations. In this thesis we develop new data sciences techniques that combine clinical, biological, chemical, and genetic data. These approaches are explicitly designed to be robust to biased and missing data. We apply these new methodologies to (1) predict new QT-DDIs, (2) validate them experimentally, and (3) investigate their molecular and genetic mechanisms. We exemplify this approach in the discovery of a previously unknown QT-DDI between ceftriaxone (cephalosporin antibiotic) and lansoprazole (proton pump inhibitor); importantly, both drugs have no cardiac indications and are safe when given individually. The clinical data mining, drug target prediction, biological network analysis, genetic ancestry prediction, and experimental validation methods described in this thesis form the basis for a comprehensive pipeline to predict QT-DDIs rapidly and robustly. They also provide an opportunity for further enriching our understanding of LQTS biology and ultimately enabling the design of safer drugs.