### Introduction

*Torsade de Pointes*(TdP).[1] Terfenadine and astemizole were withdrawn from the market in 1998 and 1999, respectively, because of the risk of cardiac arrhythmia.[2] Thereafter, the drug-induced QT prolongation has become important in the drug development industry. This issue has been identified as a considerable public health problem and has received attention from the drug regulatory authorities. Regulatory authorities have focused on identifying the risk of TdP during nonclinical and clinical development of a drug.

*in vitro*Proarrhythmia Assay (CiPA) initiative, a partnership between Food and Drug Administration (FDA) and several agencies and consortia, including Health Canada, the European Medicines Agency, and Japan's National Institute of Health Sciences, is an effort to overcome limitations of the current assessment model.[8] The objective of CiPA initiative is to facilitate the adoption of a new paradigm for assessment of the clinical potential of TdP that is not measured exclusively by the potency of hERG block and not at all by QT prolongation. Because of the complexity of the task proposed, four work streams that focus on each component of the project, ion channel,

*in silico*, myocyte and early phase human ECG, have been established.[9] For the ion channel work stream, seven important ionic currents playing important roles in generation of action potential were selected - the rapid inward sodium current (INa), the late sodium current (INaL), L-type calcium channel (ICaL), multiple outward potassium currents comprising the transient outward current (Ito), the slow (IKs) and rapid (IKr) components of the delayed rectifier potassium channels, and the inward rectifier channel (IK1)[9] – and voltage clamp protocols for the core set of cardiac ion channel types have been developing. For the

*in silico*workstream, a consensus

*in silico*model has been developed to reconstruct electrophysiologic activity within a heart cell. For the myocyte work stream, capabilities of human stem-cell derived cardiomyocyte assays have been investigated to confirm findings from

*in vitro*and

*in silico*assays. For the early phase human ECG work stream, phase I ECG working group investigated new ECG biomarkers to determine if there are unexpected ion channel effects in humans compared to the preclinical ion channel data.[710]

*in vitro*assays coupled to

*in silico*reconstructions of cellular cardiac electrophysiologic activity, with verification of completeness through comparison of predicted and observed responses in human-derived cardiomyocytes and early phase human ECG data.[10] Hence, the component of CiPA is to develop a standardized and reliable

*in silico*model and a metric that can quantitatively evaluate cellular cardiac electrophysiologic activity and ultimately assess the risk of cardiotoxicity.

### Basic principles of CiPA *in silico* study

*in silico*models, but they were limited to simulating drug effects using the half-maximal blocking concentration (IC50) for different drugs, which assumes simple pore block of the ion channels and neglects any intricacies of drugs ion channel interactions that may be important factors in predicting relative TdP risk.[1112] In CiPA,

*in silico*ventricular action potential (AP) model and a mechanism-based metric for TdP risk stratification were investigated for a more physiologic and pharmacodynamic assessment model. Twenty-eight drugs with well-known characteristics were selected and divided into 12 training drugs and 16 validation drugs for development and validation of an

*in silico*model (Table 1).

^{*}and O

^{*}) and trapping (C

^{*}) components (Fig. 2). The hERG/IKr model was then incorporated into the ORd AP model to produce the IKr-dynamic ORd model. Optimization of the model was conducted by scaling ionic current conductances to better reflect changes in AP duration observed in human ventricular myocytes when ionic currents were blocked. The optimized IKr-dynamic ORd model was adopted as a CiPAORdv1.0 model for validation.[14] From this model,

*in silico*biomarker for TdP risk, qNet metric, was derived.[15] As CiPA

*in silico*model aims to assess the integrated effects of multiple ion channel block on TdP risk, uncertainty in drug effects on ion channels to account for variations in experiments is characterized by incorporating uncertainty quantification into modeling predictions.[1516]

### hERG fitting step

_{max}(maximum drug effect at saturating concentrations), K

_{u}(rate of drug unbinding), n (Hill coefficient of drug binding), halfmax (EC50

^{n}, nth power of the half-maximal drug concentration), and V

_{half-trap}(membrane voltage at which half of the drug-bound channels are open) (Table 2).[17] The hERG model is defined by ordinary differential equations and solved with lsoda solver that selects automatically between stiff and non-stiff methods to solve problems. CMA-ES (Covariance Matrix Adaptation-Evolutionary Strategy) algorithm, which is a stochastic and derivative-free global optimization algorithm for non-linear or non-convex continuous optimization problems, was selected for fitting the hERG model. Model optimization was performed with a population size of 80 and 10

^{−3}of stopping tolerance to minimize objective function value. All parameters to be estimated are encoded logarithmically from their selected ranges [a, b] to the range [0, 10], with the equation below:

*in vitro*data used to the model fitting should have columns of time, time in milliseconds (ms) during the sweep; frac, fractional current; conc, drug concentration in nM; exp, experiment (cell) number; sweep, the sweep number. A sweep is equivalent of a single pulse of heart and it also equates to one episode in modeling. The data is resampled by non-parametric bootstrapping in order to characterize uncertainty to the parameters of hERG model. In the software developed by Kelly Chang et al., the number of bootstrap samples can be manually set. Prior to performing model fitting for the bootstrap datasets, optimal parameters must be predicted using original data by running the hERG fitting code without any options. The optimal parameter is used as initial values of model parameters for the bootstrap samples.

*N*: the total number of data points used across all doses and all episodes*conc*: the total number of drug concentrationsn: the number of drug concentrations where at least 50% block were achieved

rel(peak1, peak2): relative reduction between peak currents in the episodes $\left(\frac{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}1-\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}2}{\mathrm{p}\mathrm{e}\mathrm{a}\mathrm{k}1}\right)$

*N*is the total number of selected data points across all of doses and episodes.

*y*are mean values of bootstrap samples at the specific time points at the same drug concentration and episode, and

_{obs}*y*are predicted values at the same time points of the model fitted by

_{Pred}*y*. Small

_{obs}*n*is the number of examined drug concentrations, and the second and third components in the formula describe the relative reduction of ion current throughout the sweeping procedure to add trapping errors in the function.[17]

*y*in the second term are relative reductions of the first peak currents between the observed mean values of 1st episode and 10th episode at the specific drug concentrations, and

_{Obs,rel(1,10)}*y*are relative reductions of the first peak currents between corresponding simulated values of 1st episode and 10th episode.

_{Pred,rel(1,10)}*y*in the third term are relative reductions of the first peak currents between the observed mean values of the 1st episode and 2nd episode at the specific drug concentrations, and

_{Obs,rel(1,2)}*y*are relative reductions of the first peak currents between corresponding simulated values of 1st episode and 2nd episode. The two components weigh 0.2*N/conc. Negative error term imposes a penalty to the negative ion current for the negative constraint violation at the same drug concentration.

_{Pred,rel(1,2)}### Hill fitting step

*in vitro*data includes variables of drug names, conc (drug concentration in nM), channel (name of ionic current tested) and block (amount of block (%)). Two parameters IC50 and h (Hill coefficient) are estimated by Hill fitting.

^{2}(a nuisance parameter) was set to the variance of the fitted residuals, and the prior accuracy parameter was set to give equal weight to the prior and current error variance. By default, the first 10,000 of simulations are discarded as burn-in, and simulation results are saved every 10th iteration over the next 20,000 iterations. The number of burn-in iterations saved simulation results and the intervals can be changed manually. Convergence of parameters is evaluated with the Geweke diagnostic test using version 0.18–1 of coda R package. If the Geweke test comparing the first 10% and the last 50% of the saved iterations indicated a lack of convergence at the 95% confidence interval, then the burn-in for the adaptation period is increased by additional burn-in iterations and the entire MCMC simulation is rerun. This process is repeated until the absolute value of z is greater than 1.96. Optimal parameters and simulated parameters with joint distributions are generated as the result data of Hill fitting.[15]

### AP simulation step

*in vitro*data. After cross-validations to assess TdP risk stratification performance using 12 CiPA training drugs, CiPA

*in silico*working group redefined the optimal metric to assess the TdP risk, as

*torsade metric score*which is an averaged qNet value across 1–4 × free C

_{max}for each drug with drug effect on the four essential currents – IKr (rapidly activating delayed rectifier potassium current), INaL (late sodium current), ICaL (L-type calcium current), and INa (peak sodium current).[1520]. The measures for model prediction performance were defined as thresholds calculated using data sets of 12 training drugs with ordinal logistic regression. As one of the objectives of CiPA is to use the high throughput patch clamp systems (HTS), “hybrid” data sets collected in both manual and automated were used to validation.[520] Thresholds for the “hybrid” data set were defined by 0.0671 µC/µF for separating low from intermediate/low risk and 0.0581 µC/µF for separating high from intermediate/high risk (Fig. 3).[20]

### Current CiPA studies

*In vitro*studies with human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CM) and clinical phase I ECG studies were performed to compare whether

*in silico*model integrated with

*in vitro*data well-reflects the response of actual human ventricular cardiomyocytes. In the hiPSC-CM study, sites variability of the assessment of electrophysiology effects of the drugs using either microelectrode array or voltage-sensing optical techniques was characterized, and important hiPSC-CM assay endpoints that had been used to predict drugs for high, intermediate, and low TdP risk categories using linear regression models were identified. The

*in vitro*study was conducted with 28 drugs classified by proarrhythmic risk under the CiPA initiative, using two commercial human cardiomyocyte lines and 5 devices, across the 10 sites. The study showed fairly consistent results across the sites despite the variations in the experimental protocols, and two regression models with the three most useful predictors were constructed predicting the clinical TdP risk well with area under the receiver operating characteristic (ROC) curve values greater than 0.8.[21] The three useful predictors identified in the study were that: 1) ability of a drug to induce “mild” or “severe” arrhythmia-like events at any concentration, 2) the extent of drug-induced repolarization prolongation at any concentration, and 3) the extent of drug-induced prolongation at the clinical C

_{max}.

_{peak}(J-T

_{peakc}) interval, identified at the prior study to differentiate predominant hERG blockers from balanced blockers. For balanced blockers, which is predicted to be low risk by qNet, the CiPA initiative proposes to use ECG analysis in early phase I clinical trials to determine if there is evidence of unexpected ion channel effects in humans compared to the preclinical data. The study showed that concentration-response analysis of QTc and J-T

_{peakc}can differentiate QTc prolonging drugs with predominant hERG block from drugs that have balanced ion channel block.[22] Results of the latest validation studies suggested that the hiPSC-CM assays can be useful when combined with other CiPA nonclinical assessment strategies, and the ECG data should be interpreted with the nonclinical ion channel data and

*in silico torsade metric score*for more confident assessment. To the extent of the successful validations, the steering team of CiPA is preparing to amend the regulatory requirements for proarrhythmia assessment of the International Council for Harmonization (ICH) S7B and E14 guidelines in compliance with the proposition from ICH S7B/E14 working group.[102324]

### Summary

*In vitro*data of seven ionic currents – IKr, IKs, ICaL, INaL, Ito, IK1, and INa – is used as model inputs. With the data,

*in silico*model yields a qNet metric which is calculated by six important ionic currents – IKr, IKs, ICaL, INaL, Ito, IK1, except INa. The metric optimized for the assessment of TdP risk is defined as the torsade metric score, which is a mean qNet value averaged across 1–4 × C

_{max}for each drug with

*in vitro*data of drug effects on the four essential currents – IKr, INa, INaL, and ICaL – as model inputs. Comparisons with the hiPSC-CM assays and the clinical ECG study were conducted for validation of the CiPA

*in silico*model. Eventually, CiPA endeavor will modify the ICH guidelines on assessments of drugs' proarrhythmic risk for more precise and quantitative evaluations.