(2). The Grandi model does have a distinct fast Ito current, and so its conductance is altered directly. Models that have separate Ito components may be better for predictions based on screening Kv4.3 in future. We performed the simulation study three times in parallel, based on the following datasets: Quattro 5 channel (Q); Barracuda & Quattro 4 channel (B&Q2); and a third variant using the Quattro 5 channel screen but with hERG manual patch clamp IC50 values replacing the Quattro screening data. The manual data are taken from ICH-S7B Good Laboratory

Practice (GLP) studies featured in regulatory submission documents, and gathered by Gintant (2011). We refer to the third dataset as the Manual & Quattro (M&Q) dataset. Note that QTc MAPK Inhibitor Library chemical structure is designed to be equal to QT at 1 Hz, so in the simulations we pace cells at 1 Hz (using the square wave stimulus current

with magnitude Trichostatin A datasheet and duration as defined in the models’ CellML implementations, see below). We begin with a control simulation, pacing the model until it reaches a pseudo-steady state (see Supplementary Material S1.3 for details on steady state detection). Compound concentration is then increased from 1 nM to 100 μM, taking 20 increments equally spaced on a log10 scale. At each concentration, the data shown in Table 1 is used with Eqs. (1) and (2) to impose a new maximal conductance value for each of the screened ion currents. We then continue pacing until a new steady state is reached, and evaluate the action potential duration at 90% repolarisation

(APD90). The process is repeated with all permutations of mathematical model and dataset, giving a total of nine concentration–APD curves per compound. We use Non-specific serine/threonine protein kinase the method outlined in Elkins et al. (2013) to quantify the uncertainty on our APD90 predictions due to assay variability. In brief, we characterise variability associated with ion channel screens by examining the pIC50 distribution from the relevant control assays. A Bayesian inference scheme then produces a probability distribution for the mean of a large number of independent repeats. pIC50 values are then sampled from this distribution at random, and simulations are repeated with these values to build up a distribution of possible outcomes (as displayed in e.g. Fig. 3 and Fig. 4). The resulting intervals show where there is 95% probability that the simulation prediction lies, based on the variability we measured in the control screens for each channel. CellML is a machine-readable XML-based markup language used to describe models’ ordinary differential equations, initial conditions and parameters (Lloyd, Lawson, Hunter, & Nielsen, 2008). The ten Tusscher and Panfilov (2006), Grandi et al. (2010), and O’Hara et al. (2011) models were downloaded from the Physiome Project repository (https://models.physiomeproject.org/electrophysiology).