Uncertainty characterisation in action potential modelling for cardiac drug safety

Drugs can interact with cardiac cells to produce dangerous effects on the heart's natural rhythm. Pharmaceutical companies and regulators are interested in predicting these unintended side effects during the drug development and safety testing process. One clinical trial is the Thorough QT (TQT) study, where a drug is administered to healthy patients and particular changes in their cardiac electrical behaviour are monitored. Clinical trials such as the TQT study are expensive and time-consuming. Following a workshop in July, 2013 at the US Food and Drug Administration (FDA), the Comprehensive in-vitro Proarrhythmia Assay (CiPA) initiative was proposed to improve safety testing, led by a number of drug safety and regulatory bodies around the world. The CiPA initiative aims to replace the TQT study with a combination of in-vitro ion channel screening and mathematical mechanistic models of single-cell cardiac electrophysiology to predict a drug's proarrhythmic risk. Typically, mathematical models of ion channel screening and of single-cell electrophysiology are published with single best-fit parameter values. These point estimates do not account for any uncertainty in the models, which can come from many different sources. Therefore, predictions from these models will provide single point estimates. While such a prediction may suggest that the 'most likely' outcome given the available data is safe, it will not describe the other likely outcomes which may be dangerous. We adopt a Bayesian statistical framework to characterise some of the uncertainty in both ion channel screening data and in action potential model parameters. We then propagate this uncertainty into predictions of how a drug will affect a cardiac cell's electrical behaviour. We obtain a probability distribution describing possible effects of a particular concentration of a drug, instead of a point estimate. Such probability distributions contain more information than single point estimates: for example, the probability of crossing some dangerous threshold, which may be omitted from the point estimate if the most likely output is evaluated as safe. This process of uncertainty characterisation and propagation allows for more-informed decision-making when assessing a pharmaceutical compound's potential proarrhythmic risk.