Drug‐induced shortening of the electromechanical window is an effective biomarker for in silico prediction of clinical risk of arrhythmias

Background and Purpose Early identification of drug‐induced cardiac adverse events is key in drug development. Human‐based computer models are emerging as an effective approach, complementary to in vitro and animal models. Drug‐induced shortening of the electromechanical window has been associated with increased risk of arrhythmias. This study investigates the potential of a cellular surrogate for the electromechanical window (EMw) for prediction of pro‐arrhythmic cardiotoxicity, and its underlying ionic mechanisms, using human‐based computer models. Experimental Approach In silico drug trials for 40 reference compounds were performed, testing up to 100‐fold the therapeutic concentrations (EFTPCmax) and using a control population of human ventricular action potential (AP) models, optimised to capture pro‐arrhythmic ionic profiles. EMw was calculated for each model in the population as the difference between AP and Ca2+ transient durations at 90%. Drug‐induced changes in the EMw and occurrence of repolarisation abnormalities (RA) were quantified. Key Results Drugs with clinical risk of Torsade de Pointes arrhythmias induced a concentration‐dependent EMw shortening, while safe drugs lead to increase or small change in EMw. Risk predictions based on EMw shortening achieved 90% accuracy at 10× EFTPCmax, whereas RA‐based predictions required 100× EFTPCmax to reach the same accuracy. As it is dependent on Ca2+ transient, the EMw was also more sensitive than AP prolongation in distinguishing between pure hERG blockers and multichannel compounds also blocking the calcium current. Conclusion and Implications The EMw is an effective biomarker for in silico predictions of drug‐induced clinical pro‐arrhythmic risk, particularly for compounds with multichannel blocking action.

Conclusion and Implications: The EMw is an effective biomarker for in silico predictions of drug-induced clinical pro-arrhythmic risk, particularly for compounds with multichannel blocking action.

| INTRODUCTION
Early prediction of drug-induced cardiotoxicity is key during drug development and still remains a major challenge (Laverty et al., 2011;Stevens & Baker, 2009). Animal models are widely used for preclinical in vitro, ex vivo, and in vivo studies, but safety findings do not always translate to humans (Berridge et al., 2013), and predictions of clinical risk of arrhythmias for large sets of compounds still show limited accuracy (Lawrence, Bridgland-Taylor, Pollard, Hammond, & Valentin, 2006;Valentin et al., 2009). The current ICH S7B/E14 guidelines (2005b, 2005a) focus on hERG channel block and QTc prolongation, as surrogate markers of pro-arrhythmia. Although this paradigm has been effective in preventing new pro-arrhythmic drugs from entering the market, it has important limitations and may have led to stopping the development of potentially valuable therapeutics (Sager, Gintant, Turner, Pettit, & Stockbridge, 2014). This suggests that new effective strategies and biomarkers, for a more comprehensive prediction of drug-induced cardiotoxicity in patients, are needed in the preclinical stages of drug development. In silico human-based methodologies are becoming increasingly established in pharmacology as a potential alternative to animal experiments in the early phases of drug development, using a variety of approaches (Abbasi, Small, Patel, Jamei, & Polak, 2017;Britton, Abi-Gerges, et al., 2017;Chang et al., 2017;Dutta et al., 2017;Krogh-Madsen, Jacobson, Ortega, & Christini, 2017;Lancaster & Sobie, 2016;Li et al., 2017Li et al., , 2019Paci, Passini, Severi, Hyttinen, & Rodriguez, 2017;Passini et al., 2017;Rodriguez et al., 2016). We recently demonstrated that human in silico drug trials can achieve high accuracy (close to 90%) in pro-arrhythmic cardiotoxicity prediction for more than 60 reference compounds . The occurrence of repolarisation abnormalities (RA), which are mechanistically linked with arrhythmias, was shown to be a sensitive biomarker to predict clinical drug-induced arrhythmic risk using populations of human ventricular in silico action potential (AP) models. However, the best prediction accuracy was achieved for drug concentrations up to 100-fold the maximal effective free therapeutic concentration (EFTPC max ). While investigating the effect of potential overdoses is important in drug safety assessment, testing drug concentrations much larger than the EFTPC max might lead to false positives (Krogh-Madsen et al., 2017).
The electromechanical window (EMw), defined as the difference between the duration of electrical and mechanical systole, has been suggested as a promising biomarker to predict clinical risk of Torsade de Pointes (TdP) arrhythmia in several preclinical animal models (Guns, Johnson, Van Op den Bosch, Weltens, & Lissens, 2012;Morissette et al., 2016;van der Linde et al., 2010). However, for dofetilide, Stams et al. (2014) showed that, in the in vivo atrioventricular block canine model, the EMw solely reflects changes in QT prolongation and therefore lacks specificity for prediction of drug-induced TdP.
The aim of this study was to investigate the potential of a cellular surrogate for the EMw as a biomarker for predictions of clinical druginduced arrhythmic risk using human in silico trials for 40 reference compounds. We hypothesised that the EMw was a more sensitive biomarker of arrhythmic risk than AP prolongation and RA, particularly for multichannel block involving both potassium and calcium currents, given its dependency with calcium transient duration.
2 | METHODS 2.1 | Design of the control population of models A control population of human ventricular AP models was constructed and optimised, blindly to the drug trials results. The population was constructed using the O'Hara-Rudy dynamic (ORd) model (O'Hara, Virág, Varró, & Rudy, 2011) as baseline and the experimentally calibrated population of models methodology (Britton et al., 2013;Muszkiewicz et al., 2016;Passini et al., 2017). The nine main ionic conductances were randomly varied: fast and late Na + (G Na and G NaL respectively), transient outward K + (G to ), rapid and slow delayed What is already known rectifier K + (G Kr and G Ks ), inward rectifier K + (G K1 ), Na + -Ca 2+ exchanger (G NCX ), Na + -K + pump (G NaK ), and the L-type Ca 2+ (G CaL ).
The ranges of variation of each conductance are shown in Table 1, and they were optimised based on results in Passini et al. (2017), to maximise the number of models accepted in the population while at the same time minimising the population size. In brief, models with severe G Na , G Kr, G NaK , or G K1 down-expression often fail to produce physiological APs, while models with low repolarisation reserve (increased G CaL , G NaL , and G NCX , reduced G NaK , G Kr , and G Ks ) are more prone to develop drug-induced RA. Using this process, an initial population of 150 models was produced.
This initial population was paced at 1 Hz for 500 beats (to allow the models to reach steady state), and the last AP trace for each model was used to compute a set of seven AP and two Ca 2+ transient biomarkers: AP duration at 40%, 50%, and 90% of repolarisation (APD 40, APD 50 , and  (Table 2) and were then used for the in silico drug trials.
The refinement of ionic conductance ranges significantly enhanced model acceptance in the final population to >70% compared to~40% in previous studies . It is worth noting that variability in the experimental data may be caused in part by experimental interventions, such as the isolation procedure, in addition to the original variability in ion channel density. However, with the population approach, we assume that different sources of variability can be globally modelled by varying ion channel conductances.
All the simulations presented in this study were conducted using Virtual Assay (v.2.4.800 © 2014 Oxford University Innovation Ltd. Oxford, UK), a user-friendly software package based on C++, and specifically designed for in silico drug assays in populations of computer models. The verification of Virtual Assay results against additional software packages has already been established .

| Human in silico drug trials
In silico drug trials were performed in the population of 107 human AP models for 40 reference compounds. Drug effects were simulated using a simple pore-block model (Brennan, Fink, & Rodriguez, 2009 Table S1. Each drug was assigned to a TdP risk category, based on the classification by CredibleMeds® (Woosley & Romer, 1999), available on www.crediblemeds.org (Accessed November 30, 2018): 1 (known risk), drugs which prolong the QT interval and are clearly associated with a known risk of TdP, even when taken as recommended; 2 Abbreviations: G Na /G NaL , fast/late Na + current conductance; G to , transient outward K + current conductance; G Kr /G Ks , rapid/slow delayed rectifier K + current conductance; G K1 , inward rectifier K + current conductance; G NCX , Na + -Ca 2+ exchanger; G NaK , Na + -K + ; G CaL , L-type Ca 2+ current conductance.  (Britton, Bueno-Orovio, et al., 2017;O'Hara et al., 2011;Passini et al., 2017) and Ca 2+ transient (Coppini et al., 2013;Passini et al., 2016) Passini et al. (2017). Representative traces of these drug-induced AP phenotypes are shown in Figure 1a,b. For models not displaying abnormalities, APD 90 and CTD 90 were computed as described above, together with the EMw, defined as the difference between CTD 90 and APD 90 , as shown in Figure 1c.

| TdP risk prediction and TdP score
In silico results were used to predict the TdP risk of each drug, based percentage changes of each model compared with the corresponding control was considered. The two above criteria will be hereafter referred to as "RA only" and "RA + EMw." It is worth noting that, following drug application, EMw distributions in the population vary up to four orders of magnitude: Therefore, EMw data are shown in a logarithmic scale, using the log-modulus transformation (L(x) = sign(x) * log(|x| + 1)), which spreads out the smaller data while preserving their sign (John & Draper, 1980).
By comparing in silico results against the TdP risk categories, drugrisk predictions were divided into true positives (TP, drug with TdP risk The scoring system based on RA occurrence proposed in Passini et al. (2017), in order to integrate the results obtained at different concentrations, was extended to include the information provided by the EMw, according to the following formulas: where nRA i is the number of models showing RA at the tested concentration i (C i ), nEMw i is the number of models for which EMw i < − 10%, w i = EFTPC max /C i is the weigth inversely related to the tested concentration C i , and n tot is the total number of models in the population. For each tested concentration C i , the score considers the fraction of models showing drug-induced RA (nRA i /n tot ) or shortening of the EMw (nEMw i / n tot ) beyond threshold. All contributions are multiplied for a weight inversely related to their testing concentration (e.g., 1/30 for nRA i +nEMw i observed at 30× EFTPC max ) and added together for metric normalisation. Therefore, the final TdP score varies between 0 and 1, where 0 corresponds to a drug that does not provoke RA nor EMw shortening beyond threshold in any of the models in the population, while 1 corresponds to a drug that shows either RA or EMw shortening beyond threshold in 100% of the models, at all tested concentrations. By using the proposed score, RA and EMw shortening are naturally considered more severe when occurring at lower concentrations and/or affecting a higher fraction of the population of models.

| Ion channel block sensitivity analysis
To investigate the difference in drug-induced TdP risk predictions across different biomarkers (RA, EMw, and APD 90 ), we performed a sensitivity analysis of multichannel drug blocks for the three ion channels which are known to play a major role in RA generation : I Kr , I CaL , and I NaL . Simulations were run for all the  Table 1; AP and Ca 2+ transient biomarker distributions (Figure 2c) and the corresponding human experimental ranges defined in   Figures S1-S2). For testing concentrations of 30× and 100× EFTPC max , some compounds categorised as safe (i.e., ivabradine, metoprolol, and verapamil) can induce EMw shortening beyond threshold ( Figure S2).
It is worth noting that, for all drugs, 75% of the models are either above or below the 10% EMw threshold, including the baseline ORd model (black dots, connected by a dashed black line in Figure 3). Even though predictions based on a single model would yield the same prediction accuracy as the population, for many drugs, such as ciprofloxacin, clarithromycin, clozapine, levofloxacin, ondansetron, pentamidine, and risperidone), the drug-induced EMw changes for the baseline model are very close to the threshold. In those cases, the population results provide additional evidence on the confidence and robustness of the simulations, thus reinforcing the results.
3.2 | Shortening of the EMw improves TdP risk prediction at low concentrations  . Compounds with a low TdP score are predicted to be safe, while compounds with a high TdP score are predicted to be risky. Drug TdP risk categories as in Figure 3, from green to red to using RA only (Figure 5a vs. 5b). Results shown in Figure 5a are also depicted in Figure S3 with a logarithmic scale, required to appreciate separation for low TdP values. The four drugs that are misclassified according to the TdP score plot correspond to the FN/FP in Figure 3. 3.3 | The EMw is strongly correlated with Ca 2+ transient biomarkers but not to APD 90 Figure 6 illustrates the relationship between EMw and APD 90 or CTD 90 in panels 6a and 6b, and 6c and 6d respectively. All traces and dots are colour-coded based on the corresponding EMw magnitude. Figure 6a,b shows a weak correlation between EMw and APD 90 , whereas a strong correlation is observed between EMw and CTD 90 (Figure 6c, This is particularly relevant when considering multichannel drug effects affecting ion channels with different block potencies. As an example, Figure S4 shows three different combinations of ion channel block, all inducing the same APD 90 prolongation (+44% with respect to control), but with very different outcomes in terms of CTD 90 , and in turn EMw: I Kr block; I Kr + I CaL block; and I Kr + I NaL block. When blocking I Kr only, or I Kr + I NaL , the APD 90 is prolonged, but there is almost no effect in CTD 90 , thus resulting in an overall shortening of the EMw. On the contrary, when I CaL block is also present, the APD 90 prolongation is accompanied by a CTD 90 prolongation, and therefore, no change is observed in the EMw. It is worth noting that both I CaL and I NaL blocks counteract the APD 90 prolongation induced by I Kr block: This is why a larger I Kr block is needed to obtain the same APD 90 prolongation when also including a 50% I CaL or I NaL blocks.
3.4 | The EMw is more sensitive to Ca 2+ current block than APD 90       Table S1). I NaL block contributes to decreasing the fraction of models involved for each scenario, by contrasting APD 90 prolongation. To better illustrate the effect of I NaL block, Figure S5 shows the results of the sensitivity analysis with I CaL and I NaL block swapped.

| DISCUSSION AND CONCLUSIONS
Human in silico drug trials were conducted for 40 reference compounds with multichannel blocking actions, to evaluate how druginduced EMw shortening can potentially improve predictions of clinical pro-arrhythmic risk. A population of human ventricular AP models was optimised to capture pro-arrhythmic ionic profiles, blindly to the drug trials results.
The main findings of this simulation study are the following: Compounds misclassified as FN (nicardipine, ivabradine, and amitriptyline) all belong to potential/conditional risk categories, which are usually associated with overdoses or interactions with other drugs, and often controversial. Two of them (nicardipine and ivabradine) have actually been considered as negative controls in previous studies (Champéroux et al., 2005;Morissette et al., 2016). By considering these two drugs as safe, our prediction accuracy would reach 95% (Sensitivity 96%, Specificity 92%). The last FN, mitriptyline, affects both I Kr and I CaL currents: The resulting safety profile is given by the balance between these two blocks, the former leading to QT prolongation and potentially TdP and the latter contributing to QT shortening and suppression of RA. Our input data could minimise amitriptyline risk by underestimating its effect on hERG trafficking (Dennis, Nassal, Deschenes, Thomas, & Ficker, 2011) and overestimating I CaL IC 50 , for which controversial values have been reported (Crumb, Vicente, Johannesen, & Strauss, 2016;Lancaster & Sobie, 2016;Mirams et al., 2011;Zahradnı et al., 2008).
The only FP in this study is mexiletine, also misclassified in Passini et al. (2017). Mexiletine is a multichannel blocker, affecting mainly fast and late I Na , as well as I Kr . Based on a literature review, we can speculate that the main reason for the misclassification relies in an overestimation of I Kr block when using a simple pore drug block model. Mexiletine binds preferably to the open state of the hERG channel (Gualdani et al., 2015), and predictions could potentially improve by using a dynamic hERG channel drug block model, as the one recently proposed by the FDA Li et al., 2017). A more detailed discussion on the FN/FP drugs is included in the Supporting Information.
Compared to the classification shown in Passini et al. (2017), based on RA occurrence alone, the inclusion of the EMw has the main advantage of decreasing the need to test very high concentrations (Krogh-Madsen et al., 2017), while still considering a wide concentration range, which allows exploration of the drug effect induced by potential overdoses, as well as the EFTPCmax variability across patients. Testing EFTPCmax multiples is also the methodology proposed by within the Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative (Li et al., 2019;Sager et al., 2014). Predictions based on EMw shortening achieved the same accuracy than those based on RA occurrence only, but at concentrations much closer to therapeutic doses (10× vs. 100× EFTPCmax). Reducing the range of tested concentrations also leads to a reduction of the simulation times required for each drug and, most importantly, limits the risk of inducing FP. While sensitivity increases with the tested concentration, specificity based on RA + EMw tends to decrease at high concentrations, when even safe compounds can cause a large EMw shortening ( Figure S3). FBs at high concentrations are less likely to occur for RA: A small degree of I CaL block is sufficient to inhibit drug-induced RA, even without counteracting APD 90 prolongation and EMw shortening ( Figure S6). The inclusion of the EMw also improves the separation between safe and risky drugs in the TdP score, originally presented in Passini et al. (2017). When considering the fraction of models displaying RA + EMw, the TdP score of risky drugs increases due to the EMw contribution, allowing for a clearer risk classification.
It is worth noting that all models displaying RA at a specific concentration also show EMw shortening at a lower one ( Figure S7).
Even if the shortening of the EMw has been presented as an effec-  Figure S4). However, our results suggest that the EMw is a better biomarker than APD 90 for drugs with a multichannel effect, as it also reflects changes in the Ca 2+ transient. It captures the balance between I Kr and I CaL blocks, which could result in a reduced likelihood of RA, while still causing APD 90 prolongation ( Figure S6). In addition, it is known that in the CAVB model, there is compensated hypertrophy and increased contractility, and numerous modifications in ion channels and Ca 2+ handling have been reported, confirming a reduced repolarisation reserve (Bourgonje et al., 2013;Oros, Beekman, & Vos, 2008;Sipido et al., 2000;Volders et al., 1998). This could mask the effects of test article inhibition on inward currents (I CaL and I Na ) which could lead to a greater effect of pure hERG blockers and to a reduced ability to assess multichannel block.
The results of this study are in overall agreement with experimental EMw measurements obtained in the in vivo guinea pig model for 26 compounds, and recently published by Morissette et al. (2016). In that paper, the authors classified compounds in four classes: proarrhythmic agents, rare cases of arrhythmias, negative controls, and positive inotropes. Compounds in the first two categories tend to induce EMw shortening, while compounds classified as negative controls show EMw increase or very little change following drug application. As mentioned above, this class also includes nicardipine and ivabradine. Results in the in vivo guinea pig are in qualitative agreement with our simulations, with the exception of the three positive inotropes: dobutamine, milrinone and levosimendan. These compounds are correctly classified as safe using our in silico trials, as they show a slight increase of the EMw following drug application. On the contrary, in the in vivo guinea pig model, they cause a substantial EMw shortening (>−10%), even though they have no effect on QTc interval and they are not classified as pro-arrhythmic agents. This is because the shortening of the EMw induced by these compounds is not driven by QT prolongation (Morissette et al., 2016), but rather by other mechanisms. Milrinone and dobutamine act by increasing cAMP which ultimately increases intracellular Ca 2+ , and in turn contractility, while levosimendan acts by increasing the affinity of troponin C for Ca 2+ , thus suggesting that the EMw can also shorten following an increase in contractility not necessarily related to an increase in intracellular Ca 2+ concentration. We did not include those mechanisms in our in silico models, and therefore, they did not yielded the EMw shortening observed experimentally. The EMw considered in this study (CTD 90 -APD 90 ) is a single cell surrogate measure for the in vivo EMw, defined as the difference between the end of the left ventricular pressure wave and the end of repolarisation (QT interval).
In our study, we evaluate the effects of ion channel block on the EMw, as IC 50 values are routinely evaluated during the drug development process. Therefore, the EMw only captures drug-induced changes on ionic currents, in turn affecting Ca 2+ transient. Extensions to this approach, such as the consideration of drugs affecting the contractile machinery directly, and not via electrophysiology, could be simulated by integrating cardiac contraction models available in literature (Land et al., 2017;Negroni & Lascano, 2008) and conducting electromechanical simulations, typically requiring supercomputing power, which would allow computation of biomarkers such as QT and left ventricular pressure. This would also improve EMw predictions for the positive inotropic compounds (i.e., dobutamine, milrinone, and levosimendan), bringing our simulation results closer to the experimental observations (Morissette et al., 2016). In addition, there are other Ca 2+ biomarkers described, which could provide similar or complementary information to the EMw, for example, the rate of Ca 2+ transient decay, shown to be slower in failing human ventricular myocytes (Piacentino et al., 2003). Finally, we considered drug effects on Na + , Ca 2+ , and K + ion channels as inputs, and the EMw proved to be an effective biomarker to predict arrhythmic risk based on those.
However, some drugs could affect other mechanisms, for example, SERCA pump, Na + -K + pump, or Na + -Ca 2+ exchanger, and should this information become available, it could be easily incorporated into our simulations.
To conclude, this study demonstrates that in silico drug trials using the EMw constitute a powerful methodology to predict clinical risk of arrhythmias based on ion channel information. Such information is frequently available at the early stages of lead compound identification, and the integration of computer models in the existing pipelines for drug safety assessment could lead to a major replacement of animal experiments in the preclinical stages of drug development.