CardioMotion: identification of functional and structural cardiotoxic liabilities in small molecules through brightfield kinetic imaging

Abstract

Cardiovascular toxicity is an important cause of drug failures in the later stages of drug development, early clinical safety assessment, and even postmarket withdrawals. Early-stage in vitro assessment of potential cardiovascular liabilities in the pharmaceutical industry involves assessment of interactions with cardiac ion channels, as well as induced pluripotent stem cell-derived cardiomyocyte-based functional assays, such as calcium flux and multielectrode-array assays. These methods are appropriate for the identification of acute functional cardiotoxicity but structural cardiotoxicity, which manifests effects after chronic exposure, is often only captured in vivo. CardioMotion is a novel, label-free, high throughput, in vitro assay and analysis pipeline which records and assesses the spontaneous beating of cardiomyocytes and identifies compounds which impact beating. This is achieved through the acquisition of brightfield images at a high framerate, combined with an optical flow-based python analysis pipeline which transforms the images into waveform data which are then parameterized. Validation of this assay with a large dataset showed that cardioactive compounds with diverse known direct functional and structural mechanisms-of-action on cardiomyocytes are identified (sensitivity = 72.9%), importantly, known structural cardiotoxins also disrupt cardiomyocyte beating (sensitivity = 86%) in this method. Furthermore, the CardioMotion method presents a high specificity of 82.5%.

Over the past decades, a sustained increase in attrition rate across the drug development pipeline has occurred. The extent of these challenges is reflected in an overall 90% drug failure rate during clinical development, with 75%–80% of failures attributed to safety or efficacy (Cook et al., 2014; Kola and Landis, 2004). Cardiovascular toxicity accounts for around one-third of failures during drug development, often in late stage (Abassi et al., 2012; Kola and Landis, 2004; Laverty et al., 2011) and also accounts for up to one-third of postmarketing drug withdrawals (Cavero and Holzgrefe, 2014; Colatsky et al., 2016). Given that this results in rising drug development costs, timeline delays, increased animal use, and additional clinical monitoring, there is a clear need to find improved ways of assessing cardiac safety liabilities.

To reduce attrition in this field, industry wide efforts, such as the comprehensive in vitro proarrhythmia assay (CiPA) initiative have shown success in reducing the clinical risk of potentially torsadogenic compounds, ie, capable of inducing Torsade de pointes (TdP) (Blinova et al., 2018; Cavero and Holzgrefe, 2014; Colatsky et al., 2016; Kanda et al., 2018; Saleem et al., 2020). As part of the evolving CiPA paradigm, current in vitro methods have been tailored toward acute arrythmia-inducing compounds, which represent only a portion of cardiotoxicants.

Drug-induced cardiotoxicity can manifest in hemodynamic (ie, heart rate, contractility, blood pressure), functional, and/or structural perturbations of the cardiac muscle and vasculature (Mamoshina et al., 2021), which are undetectable via acute proarrhythmia evaluation. Structural cardiotoxicity as defined here includes degenerative and/or reparative changes in myocardial structure over time, which ultimately compromise function. This risk is particularly relevant in drugs whose administration is chronic, such as chemotherapeutic agents, eg, tyrosine kinase inhibitors. These compounds are considered structural cardiotoxicants which may not show a functional cardiac effect until up to 6 months of treatment (Takasuna et al., 2017).

The current in vitro strategy to reduce cardiovascular liabilities has evolved in similar ways within different pharmaceutical companies; this involves assessment of compound interactions with cardiac ion channels (including human ether a go-go; hERG) at an early stage during lead molecule optimization, utilizing high throughput screening strategies coupled with in silico modeling (Beattie et al., 2013). Frequently, later in development, compounds are assessed using multielectrode array (MEA) assays with human-induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) (Harris, 2015). These hiPSC-CMs present an attractive model for assessing potential cardiac liabilities and have been shown in multiple publications to be useful in a screening modality (Blinova et al., 2018; Harris et al., 2013; Pointon et al., 2015, 2017; Saleem et al., 2020; Takasuna et al., 2017). This is often the first stage at which putative drugs will have been assessed for potential cardiac liabilities using nonrecombinant human cells.

At present, identification of structural liabilities is captured through multiple doses, animal in vivo studies ranging from days to months in duration. Assessing these chronic effects earlier in the drug discovery process through the introduction of human-based in vitro studies has the potential to reduce the use of animals and enables the prioritization of compounds with a higher chance of success earlier in the discovery process, reducing time to market and cost of drug development.

For high throughput, in vitro assessment of functional cardiotoxicity, another widely used technique is measurement of calcium flux (Dempsey et al., 2016; Li et al., 2016; Sirenko et al., 2013). This technique is based on the use of a fluorescent calcium sensitive dye (eg, Cal-520 AM), which monitors transient changes in intracellular Ca²⁺, as a proxy for beating. This assay has well established and robust data analysis workflows and is simple to perform however, despite these benefits, there are some drawbacks: (1) intracellular calcium changes do not correlate with beating in all cases (eg, compounds which act directly on actin/myosin); (2) the use of a dye means that this assay can be run in the same set of cells at a single timepoint only, as the dye is cytotoxic with prolonged exposure; (3) many instruments, such as the commonly used FLIPR Tetra, have a low data acquisition rate (8 Hz for FLIPR Tetra), which can impede some important measurements such as rise time, which is the time taken from fully relaxed to fully contracted, a proxy measurement of inotropy.

Direct measurement of cardiomyocyte beating is an attractive option to overcome the constraints of both MEA and calcium flux assays by simply acquiring brightfield images at a high framerate (≥25 Hz). Analysis methods to assess the aggregate movement of cells have been developed, such as MuscleMotion (Sala et al., 2018) and PULSE (Maddah et al., 2015), but these are either not currently suitable for high throughput use or require the use of proprietary software.

Here, we describe CardioMotion: A novel in vitro assay and data analysis pipeline to predict cardiovascular activity of small molecule compounds using label-free 384-well brightfield kinetic imaging of spontaneously beating hiPSC-CMs. This pipeline is aimed to be implemented at the early stage of the drug development pipeline to identify both functional and structural liabilities in putative small molecule therapeutics. Therefore, the pipeline was assessed using a large and diverse compound set including both functional and structural cardiotoxicants and compared against the calcium flux assay.

Materials and methods

Cell culture

Frozen human-induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CM; Cardiomyocytes², obtained from Fujifilm Cellular Dynamics International, Madison, Wisconsin) were thawed and resuspended in cardiomyocyte plating media as per manufacturer’s instructions, and viability was confirmed at over 80% using a NucleoCounter NC-200 (ChemoMetec, Denmark). Cells were plated on fibronectin coated Greiner screenstar 384-well plates (781866, Greiner, Austria) at 12 000 cells/well and incubated at room temperature for 1 h, then 37°C, 5% CO₂ for 4 h. Media were then replaced with cardiomyocyte maintenance media as per manufacturer’s instructions. Cells were maintained at 37°C, 5% CO₂ with feeding every 48 or 72 h for 7 days, at which point a visual inspection was performed to confirm the formation of a synchronous syncytium prior to compound addition.

Compound preparation and treatment

Compounds were initially formulated in dimethyl sulfoxide (DMSO) with a top concentration of 10 mM. Insolubility or precipitation was not observed with the compounds from visual inspection. For treatment, cardiomyocyte maintenance media were removed and replaced with fresh media containing compounds diluted by a factor of 200 to give 8 concentrations ranging from 0.02 to 50 µM. Of note, compound information was blinded.

A total of 136 compounds were screened to a minimum of n = 3 biological replicates, using a new vial of cardiomyocytes for each replicate. Table 1 presents summary data of the compounds used, see Supplementary Table 1 for a full list of the compounds assessed. Compound classification was determined for all non-GSK compounds using the Chembl target gene and grouping using the HUGO ontology groups.

Brightfield kinetic imaging

Following 72 h of compound exposure, brightfield images were recorded using the CellVoyager 8000 (Yokogawa, Tokyo, Japan). Prior to use, the stage temperature was set to 37°C and CO₂ level to 5% in a humidified atmosphere. A 20× (0.9 NA) objective lens was used, with 250 frames over 10 s (25 frames/s, 27 ms exposure, 2 × 2 binning) recorded per well. This took approximately 70 min per 384 well plate, and example images can be seen in Figure 1A. Please see Supplementary File 2 for examples videos of the data recorded.

To translate the observable beating in the recordings from the brightfield kinetic imaging into parameters for curve fitting, a custom-made Python script was used (see Supplementary File 3 for the full script).

Calcium flux imaging

Following live imaging, the cell media were replaced with iCELL Cardiomyocyte Maintenance medium containing a 1:2 dilution of EarlyTox Cardiotoxicity Dye (Molecular Devices, San Jose, California). The plates were then incubated at 37°C, CO₂ 5% for 1 h before being read on the prewarmed FLIPR Tetra (Molecular Devices). Fluorescence (relative fluorescence units [RFU]) was measured (600 measurements at 10 s each, laser excitation 470–495 nm, and a CCD camera exposure of 0.08 s). Various parameters of calcium flux from the resulting time courses for each well (Peak frequency (BPM), Peak Width, Peak amplitude, Peak Spacing) were quantified using Molecular Devices ScreenWorks 4.2.1 Peak Pro software.

Secondary data analysis

The output from the cardio-motion analysis and calcium flux were then imported into Genedata Screener (Genedata AG, Basel, Switzerland), whereupon the 8 concentrations of each compound were normalized to the vehicle control values (for each parameter) and fitted to a 4-parameter logistic regression curve. Automatic masking of outliers through model driven and data driven modes were used and manual masking was kept to a minimum.

The pIC50s calculated for each of the parameters were assessed and given a binary active/inactive annotation based on whether the pIC50 was greater that 4.4 (ie, IC50 < approximately 40 µM).

To assess the assay, 4 metrics were used to compare the activity in the assay to the Positive/Negative control annotation described above:

• Sensitivity, or true positive rate—the proportion of true positives correctly identified

  • $Sensitivity=TP(TP+FN)$⁠, where TP = true positives and FN = false negatives

• Specificity, or true negative rate—the proportion of true negatives correctly identified

  • $Specificity= TN(TN+FP)$⁠, where TN = true negatives and FP = false positives

• Positive predictive value (PPV), or precision—the proportion of positive results in assay that are true positives

  $PPV= TP(TP+FP)$

• Negative predictive value (NPV) – the proportion of negative results in assay that are true negatives

  $NPV= TN(TN+FN).$

Results

Rationale for selecting compounds

To validate the CardioMotion assay, a compound set was collated including 3 classes:

• Negative controls (N = 40) were defined as compounds with no FDA label warning for cardiovascular toxicity nor any incidences in the FDA Adverse Events Reporting System’s (FAERs) database. Additionally, substances were included which have no known cardiovascular liability nor reported effect on cardiomyocytes in vitro.

• Positive controls (N = 82) were defined by the presence of a black box warning on the FDA label for marketed molecules with known cardiovascular liabilities

• GSK internal structural cardiotoxins (N = 14) consist of compounds from internal GSK drug discovery programs where a positive histopathological finding for structural cardiotoxicity was found in vivo. The histopathological findings for all compounds in this class included either myocardial degenerations, necrosis, cardiomyocyte vacuolation or fibrosis in vivo.

For the purposes of sensitivity and specificity calculation, both positive controls and GSK internal structural cardiotoxicants were annotated as “cardio-active,” and therefore expected to be active in the in vitro assays performed.

The compounds in this study, were selected based on: (1) a diverse portfolio of positive and negative controls, positive inotropes, negative inotropes, and noninotropes; (2) availability of structural and functional in vivo and/or in vitro data with respect to contractility (3) a variety of structurally and mechanistically different compound classes including, ion channel blockers, proarrhythmic compounds, metabolic modulators and those with poorly understood mechanisms (eg, antineoplastic compounds). Table 1 broadly categorizes the compounds into subsets based on their primary mechanism of action and/or inotropic effect. Note that some compounds are included in multiple subsets (eg, verapamil is both a negative inotrope and a calcium channel blocker).

The GSK internal structural cardiotoxins were included both to demonstrate the real-world value of this assay to drug discovery programs, and to ensure that CardioMotion is predictive with regards to structural cardiotoxicity. These represent compounds for which early detection would have had a significant impact on decision-making.

CardioMotion platform

A high throughput assay was established in which spontaneously beating cardiomyocytes are routinely used to screen compounds at an earlier stage of the drug development pipeline for potential cardiotoxic liabilities (Figure 1). Compounds are tested over a range of concentrations of 0.02–50 µM (8 concentrations; enabling the calculation of pIC50 [negative log of the IC50 value when converted to molar] values in the range of 7.7–4.3) for 72 h, permitting the detection of both functional and structural cardio toxicants, through the effect of these perturbagens on cardiomyocyte contraction. Following compound treatment, 25 frames/s movies (label-free) were recorded from cardiomyocytes and images were processed in Python. The response of 4 contractility metrics of average peak spacing, peak width, peak amplitude, and peak frequency were fitted to a 4-parameter logistic regression curve whereby, a pIC50 value was determined. If no significant deviation from the negative control (DMSO) was observed, then a pIC50 of 4.30 was recorded, which was also the negative log10 of the maximum concentration tested of any compound. To understand the performance of the contractility metrics as a whole, the pIC50s were aggregated by taking the maximum across all the contractility metrics. Data for an assay run was only considered valid if the calculated Verapamil and Doxorubicin Log IC50s were within 0.5 log molar units of the average value across all runs and the vehicle control wells presented a robust coefficient of variance under 20%.

CardioMotion algorithm development

This analysis workflow requires a single command line input to run through a 384-well plate. During development, the analysis took roughly 10 s per well to run on an HP Z4 G4 workstation with a 12 core i9 9920X CPU. To attenuate I/O bottlenecks, 10 Gbp connections between the instrument, centralized file-shares in the data center and the analysis PC were used and SSDs with >500 Mbp read/write speeds were installed on the instrument and analysis PCs.

The Python script underlying CardioMotion uses Gunnar Farnebäck’s algorithm (Bigun and Gustavsson, 2003) to compare the movement of objects between 2 images, a process known as optical flow (Bradski, 2000). This approach is known as a dense optical flow technique and approximates groups of pixels by quadratic polynomials and compares these across 2 images to estimate the global displacement.

From the sequence of brightfield images, a reference frame is constructed as the pixel-wise median intensity across all frames, approximately depicting the cells in their relaxed state. For each frame, Gunnar Farneback’s algorithm is used to compare it to the reference frame, which produces, for each pixel, a vector describing the local displacement from the reference frame. These vectors are then shifted to have mean zero across the whole frame to mitigate any noise due to camera shake. The signal is then computed as the mean of the magnitudes of the vectors, and considering the magnification and binning used, this length in pixels can be converted into nanometers, interpreted as the average cardiomyocyte contractile movement across the image at this frame.

This method of extracting a signal from the sequence of brightfield images can be summarized as follows:

Require: Video frames f₁, f₂, …, f_n

   1: m ← median(f1, f₂, …, f_n)

   2: for i ← 1 to n do

   3:  d ← optical flow(m, f_i)

   4:  d ← d − mean(d)

   5:  s_i ← mean(abs(d))

   6: end for

   7: return s₁, s₂, …, s_n

These data were then plotted to create graphs as represented in Figure 2. These waveform data were then parameterized to yield commonly used descriptors of such data, with the average peak amplitude, frequency, spacing, and width all calculated as indicated in Figure 2C. As shown, monolayer iPSC-CMs exhibited synchronous contraction throughout the whole field yielding accurate measures of contraction. See Supplementary File 2 for representative videos corresponding to the data presented in this figure.

Evaluation of cardiac activity prediction with CardioMotion

Having shown that CardioMotion is fit-for-purpose for detecting changes in contraction, we next sought to demonstrate its capability of detecting the effects of known inotropes.

Examples are shown in Figure 3 for the correct prediction of aspirin, a negative control, verapamil, a known negative inotrope and Sunitinib malate, a known structural cardiovascular toxicant with a chronic exposure-dependent negative inotropic effect. As expected, exposure to aspirin had no effect on beating parameters, meanwhile, verapamil and sunitinib decreased the peak frequency and peak amplitude in a concentration dependent manner with pIC50 values of 5.87 and 6.34, respectively.

CardioMotion assay is sensitive and specific at identifying cardio-active compounds including structural cardiotoxicants

Given that there is a growing interest within pharmaceutical companies and regulatory authorities in developing hiPSC-CM-based strategies for target validation and/or safety pharmacology (Sala et al., 2017), we next sought to validate CardioMotion sensitivity (ability to detect true positives) and specificity (ability to detect true negatives) at scale with the calcium flux assay, pharmacologically challenging hiPSC-CMs with a 136 compound set, including the 14 GSK structural cardiotoxins.

The maximum pIC50s for CardioMotion were compared to the maximum pIC50s previously obtained using calcium flux data in a similar manner. Thresholds were defined during validation-based predictivity and internal GSK experience:

• Low/Green (pIC50 <4.4, IC50 >40 μM): Compounds without a defined pIC50 ≥4.4 were considered inactive in the assay and unlikely to be of concern.

• Medium/Amber (4.4 ≤ pIC50, IC50 ≤3 μM): Threshold at which compounds are considered active in the assay. This threshold was set based on the level which yielded the best overall predictivity.

• High/Red (pIC50 ≥ 5.5, IC50 ≤3 μM): Threshold at which active compounds are of highest concern. This threshold was set based on the fact that all compounds with a pIC50≥ 5.5 had been previously annotated as cardiotoxic from in vivo or clinical data, ie, specificity = 100%.

Summary data for the results from these screens can be found in Table 2 (see Supplementary Table 1 for average pIC50 values for all compounds). All 4 parameters showed a minimum correlation (R2) between replicate runs of at least 0.8 for all parameters. As can be observed in Table 2, all parameters in CardioMotion had a greater specificity than Calcium flux, and all but one (Peak amplitude) also had a greater sensitivity when comparing the average pIC50s with the 4.4 threshold mentioned above.

As the intent of this assay is to flag compounds with potential cardiac liabilities, any perturbation to the waveform could be informative. Therefore, we assessed the specificity and sensitivity by taking the maximum pIC50 of the 4 parameters per compound and comparing that against the threshold at which compounds are considered active in the assay (ie, pIC50 = 4.4). This maximum pIC50 value leads to the highest sensitivity, when compared to any of the parameters individually, whilst maintaining a specificity above 80%.

To further assess these data, the performance of the 2 assays was assessed, using the maximum pIC50 metric, on the compound groups outlined in Table 1. Table 3 shows the proportion of compounds within each group which was flagged as active in the assay, using the maximum pIC50 metric.

As can be observed in the below table, The 2 techniques are comparable in the total number of compounds flagged, as well as many of the compound groups generally associated with functional cardiotoxicity, including inotropic compounds (positive or negative), calcium, sodium, or HERG blockers.

The groups in which the assays differ the most is in the GSK internal structural cardiotoxicants, with 85.7% (12 of 14) and 57.1% (8 of 14) of the compounds flagged in the CardioMotion and Calcium Flux assays, respectively. The results from this group, in part, led to the increased number of positive controls flagged in CardioMotion when compared with Calcium flux, leading to a greater overall sensitivity.

The table below also highlights the disparity in the proportion of negative control compounds flagged in both assays, with the CardioMotion assay flagging 7 of the 40 total compounds (17.5%) and the Calcium Flux assay flagging 17 (42.5%), reflecting the decreased specificity in the latter.

Discussion

The data presented above show that (1) CardioMotion can derive waveform data representing the beating of the cardiomyocyte monolayer; (2) peak statistics can be derived from the waveform; (3) these data can be used to screen compounds in a serial concentration and high throughput format; (4) this assay can accurately predict cardioactivity, with potential consequent cardiac liabilities with a sensitivity of 72.9% and specificity of 82.5%; and (5) this assay can also identify structural cardiotoxicants, with 85.6% of the GSK internal structural cardiotoxins identified in this assay.

The sensitivity and specificity of the CardioMotion assay was measured using a 136-compound set containing compounds with varied mechanisms of action and including compounds with varied inotropic effect as well as internal GSK structural cardiotoxins. This set was considered sufficiently varied to represent the broad range of mechanisms of cardiotoxicity likely to be observed in early drug development screening. Using this compound set, CardioMotion outperformed Calcium flux with regards to both sensitivity (72.9% vs 66.7%, respectively) and specificity (82.5% vs 57.5%, respectively).

Table 3 shows that whilst both assays perform comparably across most groups, the CardioMotion assay far outperformed the Calcium Flux assay with regards to the GSK internal structural cardiotoxins group (85.7% vs 57.1%, respectively, for CardioMotion and Calcium Flux). These 14 compounds had previously only been identified as cardiotoxicants in animal models, presenting with myocardial findings including degeneration, necrosis, cardiomyocyte vacuolation, and/or fibrosis, having failed to be flagged during early acute in vitro screening studies. These data highlight the utility of the CardioMotion assay, utilizing a 72-h exposure paradigm, as an early screening tool for both functional and structural cardiotoxicity.

Furthermore, Table 3 illustrates that the Topoisomerase/DNA interactor category had the lowest percentage positive detected by CardioMotion (45.5%). Deeper analysis of this category shows that those compounds known to have effects on nonproliferating cells including cardiomyocytes (eg, on cellular energetics, oxidative stress, RNA expression) such as the anthracyclines (doxorubicin, daunorubicin, amsacrine, mitoxantrone [Fox, 2004; Russo et al., 2021]) are active in CardioMotion. Conversely, nucleoside analogues or other compounds which specifically affect cell cycle/cell division (eg, cytarabine, fluorouracil, gemcitabine) and fluoroquinolone antibiotics (levofloxacin, sparfloxacin, moxifloxacin) which are inactive in CardioMotion cause clinical cardiotoxicity through either (i) cardiomyocyte-independent pathways (eg, inhibiting vascular endothelial cell division), affecting remodeling and function over a prolonged time period, or (ii) QT-prolongation, respectively, with secondary effects on function (Pai and Nahata, 2000; Plunkett et al., 1995; Rubinstein and Camm, 2002); these would not be expected to be positive in a homogeneous cardiomyocyte-based assay, a limitation of the cell substrate highlighted below. Mechanistic variability is also considered to underpin the performance of CardioMotion in the “Other mechanism/Idiopathic cardiotoxic MoA” category of compounds tested; this set includes compounds with known cardiotoxic mechanism, but for which there are only 1 or 2 exemplars, for instance Omecamtiv mecarbil, which is a cardiac specific myosin activator (active in CardioMotion) and rotenone, a mitochondrial electron transfer chain inhibitor (active in CardioMotion). This group also includes compounds for which the cardiotoxic mechanism of action is unknown, which may include compounds whose clinical cardiotoxic effect is through cardiomyocyte independent mechanisms, which would not be expected to be active in this assay.

When comparing with literature data, the specificity and sensitivity noted for the calcium flux data set in this study are lower than the specificity and sensitivity reported in published studies for this assay, whereas the CardioMotion assay presents lower sensitivity, but higher specificity, with the phenotypic analysis allowing for the detection of mechanisms which do not primarily affect calcium flux, such as myosin activation (omecamtiv mecarbil).

For example, a study conducted in 2014 (Pointon et al., 2015), achieved a sensitivity and specificity of 87% and 70%, respectively. However, their study was focused on the use of calcium flux to specifically detect inotropic compounds through evaluation of a data set which consisted of 31 inotropic compounds as positive controls. Whereas the data set used in this study included a total of 136 compounds with a wide array of cardiotoxic effects which included but was not limited to 5 inotropic compounds. Therefore, a direct comparison of the sensitivity and specificity across literature collated data sets that do not include similarly matching data sets cannot accurately be performed, and we rely on the direct comparison of calcium flux with CardioMotion using the same set of 136 compounds.

The parameterization of the waveform data was performed by spot checking random waveforms in the validation set and performing the parameterization manually. Although the data is not shown here, other parameters including rise time (time from fully relaxed to fully contracted) and decay time (time from fully contracted to fully relaxed) were also assessed and found to not add to the predictivity of the assay.

The curve fits on the resulting parameters were assessed in Genedata Screener and were set to only fit to a decrease in any of the parameters. Although this means that no compounds with an increasing effect on these parameters would be found to be active, we found that the vast majority of compounds showed a decreasing effect on the parameters and even those which showed an increase, showed either a mild (>30%) increase, or a bell-shaped curve. This includes those compounds which, acutely, would be expected to induce an increase in parameters, such as positive inotropes, which would be expected to increase both amplitude and frequency. This is largely due to the timepoint used; compounds such as isoproterenol, a β1 and β2 adrenergic receptor agonist, increased peak frequency with the CardioMotion assay at an acute timepoint (data not shown), however at the 72-h timepoint, a decreasing curve was noted for all parameters. It is possible that this is due to receptor desensitization and perhaps cellular exhaustion as it is well known with compounds such as caffeine, that the initial increase in beating frequency in cardiomyocytes upon exposure to this compound is only transient, with the cells ceasing to beat after a relatively short period of time (Itzhaki et al., 2011). Furthermore, the use of a longer incubation time allows the expression of cardiotoxicity which is missed in more acute assays, such as the effects of doxorubicin.

The calculated sensitivity and specificity for the 4 parameters varied and the most predictive way of identifying perturbations to the waveform data was by using the maximum pIC50 for any parameter. This is unsurprising, as compounds can have a myriad of effects on cardiomyocyte beating, and not all 4 parameters will be affected equally in all cases. This comes at a small cost to specificity, however, this assay is designed to be run at an early stage in the drug discovery pipeline, where projects have multiple compound series to prioritize and select from and therefore sensitivity is prioritized.

The data generated using the CardioMotion assay suggest added value in early-stage screening which would be conducted on all putative medicines, prior to progressing through to animal models; including a reduction in animal use, increased translation due to its phenotypic nature and reduced downstream development costs.

This assay can be implemented in any laboratory with a Yokogawa CV8000 with ease, and in any laboratory with a high content imaging instrument with appropriate frame rate and environmental control with a small amount of effort to change the code as described.

Being label free, the CardioMotion assay also has the advantage of enabling data acquisition at multiple time points, for instance an acute readout would be more akin to the MEA and calcium flux readouts, as well as time points more relevant to the in vivo measurements (up to 6 weeks). This may address, through further work, the current limitation of the use of the 72-h timepoint, as both acute acting compounds and compounds with an effect at later timepoints may not be active in the assay.

However, at the timepoint used in this study and to an even greater extent in longer term studies, the use of a mono-culture of cardiomyocytes has inherent limitations with regards to identifying compounds with potential cardiotoxic liabilities. Cardiomyocytes account for approximately 30% of all cells in the human heart, whereas the remaining majority consists of cardiac fibroblasts, endothelial cells and immune cells (Nag, 1980; Zhou and Pu, 2016). Studies have shown that inclusion of these cell types promotes maturity of cardiomyocytes (Giacomelli et al., 2020; Ravenscroft et al., 2016) and they may themselves be a site for a cardiotoxic effect, for example endothelial cell effects of VEGF inhibition of sunitinib, however other effects of this tyrosine kinase inhibitor, such as energetic/mitochondrial fusion-fission homeostasis via inhibition of AMPK signaling, directly affect cardiomyocytes (Force and Kolaja, 2011); sunitinib was classified as highly active in the CardioMotion assay. Employing a more comprehensive model may enhance the detection of cardiotoxic compounds that impact cardiac components beyond just the cardiomyocyte, such as fibrosis inducing compounds, compounds with hypertrophic effects or those with hemodynamic effects which may indirectly affect the myocardium. Moving beyond cardiomyocyte only models, complex in vitro models with multiple cell-types, and also for which loading conditions could be varied, could prove to be the solution to this challenge resulting in increased translatability of this model system. However, these improvements must be balanced against pragmatic considerations around throughput and cost, when intended to be used as an early-stage screen.

CardioMotion was developed using iCell Cardiomyocytes², obtained from Fujifilm Cellular Dynamics International. Although these cells have been used extensively in the literature for drug discovery, including cardiovascular safety (Blinova et al., 2018; Feric et al., 2019; Palmer et al., 2020; Patel et al., 2019) it corresponds to cells derived from a single donor, an apparently healthy normal Caucasian female. This is a limitation of the assay, as the evolving area of pharmacogenomics has highlighted the importance of the genotype in determining the susceptibility of individuals to drug-induced cardiotoxicity (Li et al., 2022). Further work should assess the value add of this “personalized medicine” approach, using donors with genetic predispositions such as those with mutations associated with hypertrophic cardiomyopathy.

Overall, the CardioMotion assay provides a means for the identification of both structural and functional CV toxicants across a wide range of compound classes that could go undetected until chronic in vivo administration in animals and humans. Thus, incorporation of the assay into the drug development pipeline could lead to an overall reduction in later stage attrition as well as a reduction in animal use.

Supplementary data

Supplementary data are available at Toxicological Sciences online.

Funding

This work was supported by GSK.

Declaration of conflicting interests

All authors were employees of GSK when the work was conducted.

Code availability

All software source code is available at https://github.com/GSK-Screening/CardioMotion.

Data availability

The CardioMotion dataset with the pIC50 values for all parameters in all 136 compounds has been deposited in Supplementary File 1.

References