A Deep Learning Platform to Assess Drug Proarrhythmia Risk

Summary

Drug safety initiatives have endorsed human iPSC-derived cardiomyocytes (hiPSC-CMs) as an in vitro model for predicting drug-induced cardiac arrhythmia. However, the extent to which human-defined features of in vitro arrhythmia predict actual clinical risk has been much debated. Here, we trained a convolutional neural network classifier (CNN) to learn features of in vitro action potential recordings of hiPSC-CMs that are associated with lethal Torsade de Pointes arrhythmia. The CNN classifier accurately predicted the risk of drug-induced arrhythmia in people. The risk profile of the test drugs was similar across hiPSC-CMs derived from different healthy donors. In contrast, pathogenic mutations that cause arrhythmogenic cardiomyopathies in patients significantly increased the proarrhythmic propensity to certain intermediate and high-risk drugs in the hiPSC-CMs. Thus, deep learning can identify in vitro arrhythmic features that correlate with clinical arrhythmia and discern the influence of patient genetics on the risk of drug-induced arrhythmia.

Graphical Abstract

ETOC:

Serrano et al. used human iPSC-derived cardiomyocytes and deep learning data analysis to establish an In vitro safety margin that predicts clinical proarrhythmic effects of drugs. Their platform shows high accuracy in identifying risky drugs as well as genotypes associated with increased risk of arrhythmia in people.

Introduction

Drug-induced arrhythmias are a common cause of drug attrition during development and for restricted use or withdrawal from the market^1–3. As people vary in their predisposition to drug-induced arrhythmia ^4–6, there is a widely accepted need to assess risk in susceptible populations ^7,8. For ethical reasons and practical limitations, susceptible individuals, including carriers of rare predisposing gene variants, are not generally included in clinical trials ⁹. hiPSC-CMs retain an individual’s genetic makeup and enable scalable production of cardiac cells for in vitro testing and, therefore, are a breakthrough technology for risk assessment ^10,11. This notion is supported by the findings that several mutations that cause electrophysiological or myopathic heart disease predispose cardiomyocytes to drug-induced arrhythmia ¹².

Cell-based assays assess arrhythmia risk by quantifying waveform features in the cells’ action potential. Typically, these features are quantified using human-defined metrics such as the action potential duration at 90% amplitude (APD90) or incidence of after depolarizations ^6,13,14. However, these human-defined features do not accurately predict clinical arrhythmia ^15–17. Altogether, the complex manifestations of arrhythmia, the uncertain correspondence between in vitro action potential waveform features and actual clinical arrhythmia, and the influence of disease susceptibility loci present significant challenges for risk prediction.

To address this problem, we developed a deep learning approach to discriminate the in vitro electrophysiological features induced in hiPSC-CMs by reference drugs with well-established (high to low) risk of eliciting the life-threatening ventricular tachyarrhythmia (Torsades de Pointes, TdP). Deep learning is a type of artificial intelligence (AI) that uses multiple computational layers in a deep neural network (DNN). DNNs extract features relevant to discriminating input classes in a systematic and unbiased manner, effectively removing the need for human-defined metrics ¹⁸. Among the different types of DNNs, convolutional neural networks (CNNs), which learn complex features from input data by assigning weights to the result of convolutional operations, are showing tremendous success in various biomedical applications such as medical image analysis ¹⁹ and physiological signal analysis ²⁰. Recently, CNNs have been used to automate the detection and classification of arrhythmias both in vitro²¹ and in vivo ²².

We trained a CNN to discriminate high versus low-risk drugs based on intrinsic drug-induced electrophysiological waveforms in hiPSC-CMs rather than human expectations. The CNN more accurately classified actual drug risk in patients than did human-defined metrics. Moreover, the trained CNN successfully quantified the increase risk of drug induced arrhythmia caused by cardiomyopathic gene variants, which pose a clinically significant risk factor that has been challenging to quantify ²³. In summary, deep learning out-performed human-defined methods for drug risk assessment and detected an influence of patient genetics on susceptibility to drug-induced arrhythmia.

Results

High throughput screening of electrophysiological effects of drugs in healthy and disease hiPSC-CMs models

This study aimed to develop a new paradigm to accurately predict the risk of drugs and genetics on TdP arrhythmia without imposing human bias (Fig. 1). Genotype-phenotype relationships were modeled in a cohort of hiPSC-CMs generated from 8 hiPSC lines. Three (3) lines were derived from healthy donors (HD.113, HD.273, and HD.15S1), while five (5) lines harbored pathogenic mutations that cause HCM [HCM.MYBPC3 p.R943X ²⁴], left ventricular noncompaction (LVNC), and HCM [HCM.TPM1 p.K37E ²⁵] and DCM [DCM.PLN p.R14del ²⁶, DCM.RBM20 p.R634Q ²⁷, DCM.TNNT2 p.R183W ²⁸]. These mutations were introduced in the same genetic background (HD.15S1 iPSCs) by CRISPR-mediated genome editing, minimizing the potentially confounding effects of the genetic background. All these mutations, except for the TNNT2 p.R183W mutation, are associated with elevated arrhythmic risk in patients, including ventricular arrhythmias and sudden cardiac death.^29–34

Figure 1. Strategy to determine the influence of myopathic gene variants on the proarrhythmic effect of drugs.

hiPSC-CMs were generated from three donor patients without risk-associated genetics. DCM and HCM causing mutations were introduced to one of the healthy background and multiple batches of hiPSC-CMs were generated. The hiPSC-CMs were treated with 37 drugs of characterized high, intermediate and low/no arrhythmic risk. Each drug was tested at 8 different concentrations and a voltage sensitive dye was used to obtain membrane potential recordings. A CNN was trained to classify voltage traces based on the drug’s risk of inducing arrhythmia in patients. Class probabilities from the CNN were used to rank the proarrhythmic effects of drugs and evaluate the influence of myopathic gene variants.

Experiments were carried out using a panel of drugs with well-characterized proarrhythmia risk profiles, including the 28 drugs proposed by Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative ³⁵ and 9 additional drugs to facilitate generalization across studies. For training and evaluation purposes, we used the CiPA collection categorized as high, intermediate, and low/no risk for developing TdP arrhythmia at clinical exposures based on published reports, the FDA adverse effect database, and expert opinion ³⁶.

The uniformity and quality of the cardiomyocyte batches were assessed based on the baseline morphology of the action potential traces from 28 differentiation batches from 8 hiPSC lines displayed typical electrophysiological characteristics at baseline (see Methods). Differentiation batches that presented with abnormal baseline characteristics, such as baseline arrhythmias and cessation in certain wells, were excluded from further analysis (Supplementary Fig. S2).

Deep learning features from voltage traces

A CNN was designed to classify voltage traces as non-arrhythmic, arrhythmic, or asystolic (Fig. 2A, see Supplementary Fig. S1A and Methods for details). For training and testing the CNN, sets of annotated traces are required. First, traces with no spontaneous action potential activity were manually labeled as asystolic (Fig. 2B). To remove human bias regarding the classification of non-arrhythmic versus arrhythmic traces, we employed the following criteria: 1) Traces from wells treated with drugs carrying high risk of inducing TdP arrhythmia [classified according to CiPA ³⁵] and treated at a concentration greater or equal than maximum free plasma concentration (free C_max) were annotated as arrhythmic. 2) Traces from wells treated with drugs with a no or low-risk CiPA classification and the concentration was less or equal to free C_max were annotated as non-arrhythmic. Traces that did not follow any of these criteria – i.e., wells treated with drugs of intermediate CiPA risk, wells treated with low doses of high-risk compounds, or high doses of low-risk compounds – were not included in either training or test datasets.

Figure 2. A deep learning neural network for classification of voltage traces.

A) Schematic of CNN classifier for voltage traces into the classes: non-arrhythmic, asystolic, and arrhythmic.

B) Steps for trace annotation

C) Splitting data for training and test datasets.

D) Examples of each trace category and values of class probabilities outputted by the CNN.

E) Confusion matrices for the training and test datasets.

The training dataset comprised traces from healthy donor lines treated with the training compound library proposed by CiPA (Fig. 2C, Supplementary Fig. S1B). The training dataset was used to refine the convolutional layers’ number and size of filters and dropout percentage until an accuracy of 88.6% was achieved. The confusion matrix of the training set (Fig. 2E) revealed that most errors involved traces classified as ‘non-arrhythmic’ by the CNN but were derived from drugs annotated as ‘arrhythmic.’ Further analysis showed that 127 of these 142 misclassified traces were the result of treatment with bepridil, for which previous studies had shown consistently classified as producing non-arrhythmic responses in hiPSC-derived cardiomyocytes despite being tested at concentrations exceeding 100 times the free C_max ^13,14,37.

The accuracy of the trace classifier was verified against different sets of unseen data (Fig 2C, E). The test set 1 contained data from compounds that differed from the training library but in the same cell lines used for training (yielding an accuracy of 98.5%). The test set 2 contained data from different cell lines than those used for training but the same compounds (yielding an accuracy of 86.2%). Lastly, test set 3 contained data from both cell lines and compounds that were not used for training (yielding an accuracy of 97.5%). Note that test set 2 contained unseen data from bepridil, which explains the lower accuracy relative to test sets 1 and 3.

Representative traces and the CNN class probabilities output are shown in Fig. 2D. Briefly, the CNN interpreted whole voltage waveforms from non-arrhythmic, cessation, and arrhythmic classes and generated their respective probabilities (Fig. 2D). These probabilities can be used as a unified set of metrics applicable for all action potential waveforms, overcoming the problems that human-defined metrics suffer when quantifying specific phenotypes (e.g., in asystole APD₉₀ loses its meaning as no action potential exists; or the detection of early after depolarizations (EADs) or delayed after depolarizations (DADs) which often requires human inspection of the trace). Most importantly, the training trace annotations were based on the risk of clinical arrhythmia. This approach allowed the CNN to learn features of the in vitro traces that corresponded to the effect of clinically risky drugs, circumventing the concern that human intuition might not recognize the in vitro features that are predictive of clinical arrhythmia.

Use of the CNN to determine the probability of drug-induced arrhythmia

The trained CNN reflected the dose-dependent proarrhythmic effects in the voltage waveforms of cardiomyocytes treated with high-risk drugs as increases in the probability of arrhythmic or asystolic phenotypes. To illustrate this result, Fig. 3A shows traces from ibutilide treated hiPSC-CMs showing progressively arrhythmic traces with increasing dose [manifested by action potential prolongation and EADs]. The trained CNN generated a continuous, dose-dependent increase in arrhythmia probability based on the waveform input (Fig. 3B). Note that the calculated 50% arrhythmia class probability (EC₅₀, triangle) was lower than the free C_max value (dashed line), indicating that arrhythmia is detected at therapeutic concentrations of the drug. Consistent with the dependence of contractility on intracellular [Ca²⁺], the calcium channel blocker nifedipine showed a dose-dependent induction of asystole (Fig. 3C) that corresponded with continuous increase in asystole probability determined by the CNN (Fig. 3D). Similarly, dose-response curves can be calculated for any compound regardless of whether arrhythmic risk or even pharmacokinetic data are known.

Figure 3. Development of the Torsadogenic Safety Margin.

A) Representative traces of different concentrations of the torsadogenic drug ibutilide.

B) Proarrhythmic dose-response curve (solid line) plotting the probability of arrhythmic class from the traces shown in A. 50% probability of arrhythmic value EC₅₀ (triangle). Clinical maximum free plasma concentration, C_max (dashed line).

C) Representative traces of different concentrations of the calcium channel blocker nifedipine.

D) Proarrhythmic dose-response curve (solid line) plotting the probability of asystole class from the traces shown in A. 50% probability of asystole value EC₅₀ (triangle). Clinical maximum free plasma concentration, C_max (dashed line). Average of 3 differentiation batches per cell line with 3 technical repeats per dose in each batch.

E) Dose-response curves of for other drugs and their CiPA risk classification.

F) Color-coded dose-response curves where probability of arrhythmic 0-1 is encoded as green-red and the dose is normalized by C_max.

G) Torsadogenic safety margin for each drug and healthy donor line of the screen. Drug names are color-coded based on CiPA classification.

H) ROC curves for using the Torsadogenic Safety Margin as predictor for a model to identify high-intermediate risk drugs vs. no-low, high vs. no-low, and intermediate vs. no-low.

Quantitative metrics of drug safety

The preceding analysis was applied to the entire dataset (37 drugs, 8 hiPSC lines). Examples of CiPA-classified high (ibutilide), intermediate (ondansetron), and low (verapamil) arrhythmic risk compounds are shown with 3 drugs (flecainide, citalopram, and aspirin) that are not among the CiPA set but showed effects consistent with published clinical data (Fig. 3E, Supplementary Table S1, the entire dataset is shown in Supplementary Data. S1). In general, compounds classified by CiPA as having higher TdP risk displayed lower EC₅₀ values when normalized by C_max (Fig. 3F). To quantify this observation, we elaborated the Torsadogenic Safety Margin, defined as the log₁₀(EC₅₀/C_max).

To visualize the relationship between the Torsadogenic safety margin and clinical risk, we rank-ordered the compounds by their Torsadogenic safety margin (Fig. 3G). 8 out of the top 9 places were occupied by drugs classified by CiPA as carrying high risk (Supplementary Table S1). The safety margin generally correlated with the CiPA-assigned clinical risk, although some exceptions were noted. Droperidol, classified as having intermediate risk by CiPA, ranked 7^th highest, whereas bepridil, classified as high-risk by CiPA, was ranked 17^th, placing it among intermediate-risk drugs. Mexiletine and ranolazine showed safety margin values comparable to intermediate-class drugs, as reported in other hiPSC-CMs studies ^13,14,38, despite being considered a low-no risk by CiPA classification.

The Torsadogenic safety margin to classify risk can be used as a predictor in a logistic regression model to evaluate discretized CiPA risk. To compare to existing literature, we created a model to assign drugs to high-intermediate and low-no-risk discrete categories. This model resulted in an AUC value of 0.95 (Fig. 3H), higher than previously proposed models that use human-defined predictors applied to our dataset (Supplementary Fig. S3) ¹³. In addition, we trained models to distinguish high and low-risk compounds (AUC = 0.97) and intermediate and low-risk (AUC = 0.8) (Fig. 3H).

We computed the Torsadogenic safety margin of 9 additional drugs not contained in the CiPA collection (amiodarone, amitriptyline, aspirin, citalopram, digoxin, epinephrine, erythromycin, flecainide, and fluoxetine). We determined their risk by comparing their rank position against the CiPA reference compounds (Fig. 3G). Flecainide ranked among high-risk drugs, whereas citalopram, fluoxetine, and amiodarone ranked as intermediate-risk, and erythromycin was at the border between high and intermediate-risk. The probability of arrhythmia EC₅₀ values could not be determined for amitriptyline, aspirin, digoxin, and epinephrine, indicating that they are no or low-risk drugs in agreement with the clinical literature on these drugs (Supplementary Table 1). We conclude that the continuous nature of the Torsadogenic safety margin allows it to stratify the risk of drugs at a finer scale than discrete categorization, in which all drugs within the same category (e.g., high risk) are considered equal.

DCM and HCM mutations influence responses to proarrhythmic drugs

Next, we asked whether the CNN could discern the influence of patient genetics on drug-induced arrhythmia. To test this idea, we focused on variants that cause familial DCM and HCM. Certain DCM and HCM-causing variants place patients at risk for ventricular arrhythmias, and current treatment guidelines call for caution in treating familial DCM and HCM patients with Torsadogenic drugs ^39,40. The mechanisms responsible for arrhythmia in familial DCM and HCM involve altered Ca²⁺ and Na⁺ flux as well as tissue remodeling (e.g., elevated fibrosis) in addition to K⁺ current inhibition ^41–43 that is responsible for most drug-induced TdP ⁴⁴. Despite these clinical and mechanistic associations, the contribution of DCM and HCM to drug-induced TdP risk has not been addressed quantitatively. Furthermore, an in vitro system to quantify TdP probability would aid in elucidating risk mechanisms.

We used the trained CNN to determine if DCM and HCM mutations increase TdP probabilities in response to drug treatment. To remove the influence of background genetics, we created a panel of isogenic DCM and HCM lines by introducing disease-causing mutations into the healthy donor hiPSC line HD.15S1 (Fig. 4A). The DCM-causing RBM20 p.R634Q and PLN p.R14del variants and the HCM-causing MYBPC3 p.R943X and TPM1 p.K37E variants are recognized as “at-risk” for fatal arrhythmia ^29,30,32,33. In contrast, the DCM TNNT2 p.R183W is reported to confer less arrhythmic risk ³⁴.

Figure 4. Influence of DCM and HCM gene variants.

A) Myopathic gene variants were introduced by CRISPR/Cas9 gene editing onto a common healthy donor hiPSC line.

B) Torsadogenic safety margin for each drug and healthy donor line of the screen for the HCM and DCM cell lines and the healthy donor isogenic line. Drug names are color-coded based on CiPA classification.

C) Asystolic safety margin for each drug and healthy donor line of the screen for the HCM and DCM cell lines and the healthy donor isogenic line. Drug names are color-coded based on CiPA classification.

D) Hypothesized model for increased sensitivity of arrhythmogenic cell lines to TdP and asystole (see Discussion).

E) Dose-response curves of probability of arrhythmic in HCM and DCM lines and the isogenic control cell lines, treated with ibutilide, droperidol, vandetanib, nifedipine. Point markers indicate EC₅₀ with error bars at a 95% confidence interval. Shaded region signifies all traces were asystolic at that concentration range. Average of 3 differentiation batches per cell line with 3 technical repeats per dose in each batch.

F) Dose-response curves of probability of asystolic in HCM and DCM lines and the isogenic control cell lines, treated with ibutilide, droperidol, vandetanib, nifedipine. Point markers indicate EC₅₀ with error bars at a 95% confidence interval. Average of 3 differentiation batches per cell line with 3 technical repeats per dose in each batch.

Action potential waveforms in hiPSC-CMs derived from each line were recorded at baseline and upon dose-escalation treatment with the 37 drugs. Results in the mutant lines were compared to the isogenic control to determine the contribution of the gene variants. Most drugs showed similar Torsadogenic safety margins in the DCM and HCM hiPSC-CMs as for the isogenic control, although some trended towards increased risk (e.g., ibutilide, droperidol) (Fig. 4B, E). In contrast, the asystole safety margin (calculated analogously using the EC₅₀ values of the probability of asystole) revealed a strong influence of DCM and HCM genotype on drug effects (Fig 4C). For example, ibutilide, dofetilide, and droperidol showed a heightened propensity to cause asystole in the DCM variants RBM20 p.R634Q and PLN p.R14del, and the HCM causal variants MYBPC3 p.R943X, and TPM1 p.K37E relative to the isogenic control hiPSC-CMs (Fig. 4C, F). Interestingly, hiPSC-CMs carrying the DCM TNNT2 p.R183W variant (that is less arrhythmogenic in patients) had a similar profile to the isogenic healthy donor control (Fig. 4C, F).

Discussion

Current in vitro proarrhythmia assays rely on the measurement of human-defined features of cardiac electrophysiology, such as APD and beat rate ^6,13,45. However, arrhythmias present complex geometries, such as EADs, that are challenging to quantify by conventional metrics. In practice, arrhythmic phenotypes are typically classified based on categorical descriptors (e.g., sustained ventricular tachycardia or after depolarizations) or binary variables (e.g., presence or absence of EADs) that require human interpretation of the action potential ^13,14,38. Treating arrhythmic phenotypes as yes/no parameters imposes an artificial threshold for an arrhythmia that depends on human intuition and cannot quantify progression from normal to arrhythmic waveforms as a function of drug dose, chemical modification, or hiPSC genetics. More fundamentally, human intuition is not based on an established ground truth regarding the in vitro effect of drugs that cause arrhythmia in patients. Thus, the motivation for this study was that human-defined metrics might not capture features of the in vitro waveforms that correspond with the actual arrhythmic risk of the drugs in people.

We developed our deep learning algorithm to overcome the limitations of human-defined metrics by recognizing features uniquely associated with drug-induced arrhythmia in hiPSC-CMs. Training a CNN to discriminate features was based on a dataset of 3 classes of action potential traces: 1) traces generated by treating hiPSC-CMs with high doses of drugs that cause ventricular arrhythmia in people (class: arrhythmic), 2) traces generated with low doses of safe compounds (class: non-arrhythmic), and 3) traces in which drugs induced asystole, which is characteristic of high doses of Ca²⁺ channel blockers (class: asystole) but also resulted from treatment with very high doses of proarrhythmic drugs (Fig. 2). The probability of classification as arrhythmic and asystolic was a continuous dose-dependent metric that quantified the behavior of a drug. Relating the EC₅₀ values for these probabilities to the free plasma concentration of a drug used in clinical practice (free C_max) defined a Torsadogenic safety margin for each drug (Fig. 3). In hiPSC-CMs from healthy donors, the Torsadogenic safety margin accurately predicted the clinical risk of drugs with an AUC of 0.95 (Fig. 3H), representing an improvement over previously published multiparametric methods based on human classification (Supplementary Fig. 3).

Patients with structural heart diseases such as DCM and HCM are at risk for drug-induced arrhythmia. They should be carefully monitored using electrocardiographic and other modalities when treated with drugs at risk for inducing arrhythmia ⁴⁰. We applied the Torsadogenic and asystolic safety margins to traces generated by treating isogenic hiPSC-CMs carrying gene variants that cause DCM and HCM in patients. The CNN probabilities and the calculated safety margins revealed that pathological gene variants associated with arrhythmic cardiomyopathies in patients sensitized hiPSC-CMs to adverse proarrhythmic and asystolic effects of high-risk drugs. In particular, hiPSC-CMs carrying pathogenic variants in PLN, RBM20, TPM1, and MYBPC3 associated with increased arrhythmic risk in patients ^29,30,32,33 trended towards increased probabilities of being classified as arrhythmic (Fig. 4B, D) and showed highly significant increases in the probabilities of asystole (Fig. 4C, E) compared to isogenic controls.

The finding that arrhythmogenic DCM and HCM gene variants increase the risk of Torsadogenic drugs has implications for understanding the electrophysiological substrates for arrhythmia. The genetic lesions examined here affect sarcomeric, RNA splicing, and Ca²⁺ handling proteins. Electrophysiological remodeling in familial DCM and HCM includes reduced repolarizing K⁺ currents (I_K) and increased intracellular diastolic Ca²⁺ and late Na⁺ (I_NaL) current, as reviewed ^42,43. For example, these current changes have been reported in isolated cardiomyocytes from mice, human myectomy samples, and hiPSCs carrying MYBPC3 mutations that are functionally equivalent to the p.R943X truncation mutation used in this study ^41,46–48. A downstream circuit involving CaMKII sustains electrophysiological remodeling ^49,50 and decreases repolarizing I_K ⁵¹. Decreased I_K in the hiPSC-CMs, which are more depolarized relative to adult cardiomyocytes (at least in monolayer culture) ^52,53, would enhance the effect of Torsadogenic drugs ⁴⁴ and is a possible mechanism for their asystolic effect in hiPSC-CMs carrying the DCM and HCM mutations (Fig. 4D).

In conclusion, the deep learning algorithm recognized in vitro arrhythmic features in the hiPSC-CMs based on drug effects. It did not rely on human adjudication nor human-defined in vitro hiPSC-CM arrhythmic phenotypes; instead, it focused the CNN on recognizing hiPSC-CM phenotypes associated with patient drug responses. We derived safety margins from the relationship between the machine-generated class probabilities and the free plasma concentrations of each drug. Unlike categorical descriptors of arrhythmia (such as EADs, action potential prolongation, and triangulation), the CNN class probabilities map each trace to a continuous spectrum spanning non-arrhythmic to arrhythmic and asystolic phenotypes. Overall, the calculated safety margins more accurately discriminated high, intermediate, and low-risk drugs than prior methods based on human-defined features, and revealed arrhythmogenic HCM and DCM variants increase sensitivity to drug-induced arrhythmia. Thus, the recognition of hiPSC-CM features of arrhythmia by deep learning should improve the detection of risky compounds during development as well as assign risk to gene variants of unknown significance.

Limitations of the study

Deep learning has the inherent limitation that the machine-learned features typically lack human-intelligible meaning (the so-called ‘black box’ problem). In other words, the mathematical operations that the machine has optimized to calculate the probability of arrhythmia are too abstract to be interpretable by humans and cannot be ascribed to action potential features. Although we cannot query the CNN to understand the basis for classification, visual inspection of the traces revealed that the dose-dependent increases in arrhythmia probability corresponded to AP prolongation, EADs, and alternans. The CiPA reference collection causes a limited range of arrhythmogenic phenotypes (typically AP prolongation and EADs). We have found that the CNN (RS, unpublished data) can be applied to a broader range of phenotypes [e.g., DADs, sustained and non-sustained ventricular tachycardia [Sustained VT (SVT), Non-sustained VT (NSVT)] associated with cardiac electrophysiological disorders ^54–56] by increasing the number of possible classifications (i.e., the number nodes in the last layer of the CNN).

Certain drugs in this study (droperidol, ranolazine and mexiletine) presented as “riskier” than in actual clinical practice while bepridil appeared less “risky” as discussed above. Previous studies using hiPSC-CMs similarly misclassified these drugs ^13,37. This might reflect a limitation in using hiPS-CM datasets to classify these drugs. Concerning the deep learning methodology, we speculate that the features needed to accurately classify these drugs might have been masked by variability in the waveforms in the current dataset, and that a larger dataset might be adequately powered to train the neural network to correctly define the risk of these drugs.

STAR METHODS

RESOURCE AVAILABILITY

Lead Contact

Further information and request for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Mark Mercola (mmercola@stanford.edu).

Materials Availability

All unique reagents generated in this study are available from the Lead Contact with a completed Materials Transfer Agreement.

Data and Code Availability

The datasets and codes generated during this study are available in the Supplementary Information and publicly available from GitHub (https://github.com/rikserrano/DeepLearning_TdP_Risk).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

hiPSCs culture and differentiation

The protocols used in this study were approved by the Stanford University Institutional Review Board. Sendai viral vector was used for the reprogramming of human PBMCs. The de-identified healthy donor hiPSC lines HD.113 (male), HD.15S1 (male), and HD.273 (female) were obtained from the Stanford Cardiovascular Institute Biobank. The hiPSC clones were cultured in E8 cell culture media (Thermo Fisher Scientific, A1517001) in plates coated with growth factor-reduced Matrigel (Corning, 356231). The hiPSC were differentiated into CMs utilizing a chemically defined cardiomyocyte differentiation protocol ⁵⁷ with modifications. Briefly, hiPSCs were treated with CHIR99021 (Tocris, 4423) for 3 days in RPMI 1640 (Thermo Fisher Scientific, 11875119) with B27-insulin (Thermo Fisher Scientific, A1895601). Subsequently, the cells were treated with the Wnt inhibitor C59 (Tocris, 5148) in RPMI supplemented with B27-insulin (RPMI/B27-) for another 2 days. Between 5–11 days of differentiation, RPMI/B27-media refreshed every other day and switched to RPMI supplemented with 1X B27(RPMI/B27+) after beating was observed. On day 11, cells were cultured in RPMI/B27+ without glucose, supplemented with 5 mM sodium L-lactate (Sigma Aldrich, 71718) for 3 days, to improve the CMs purity. On day 14, the cells were cultured on RPMI/B27+. On day 15, the hiPSC-CMs were dissociated with TrypLE 10x (Thermo Fisher Scientific, A1217703) and seeded in 6-well Matrigel-coated plates at a density of 3x10⁶ per well in RPMI/B27+ containing 10% KOSR and ROCK inhibitor Y-27632 (Tocris, 1253) (replating media). After 2 days, cells were cultured in RPMI/B27+ without glucose for 3 days and then switched to maturation media ⁵⁸ and cultured for an additional 4-5 weeks prior to replating in 384-well plates for high throughput analysis.

Genome Editing in hiPSCs

The MYBPC3 p. R943X, TPM1 p. K37E, PLN p.R14del, RBM20 p. R634Q, and TNNT2 p.R183W mutation were introduced to the HD.15s1 iPSC line as previously described ^{24,25,26,27,28} Briefly, CRISPR guide RNA (gRNA) sequences were designed using an online tool (http://crispr.mit.edu/), and we selected the gRNAs with the highest score (specificity) for cloning into into the BbsI site of the pSpCas9(BB)-2A-GFP (PX458; a gift from Feng Zhang; Addgene plasmid #48138) by reverse complementary guide DNA oligos. Briefly, oligos were annealed in T4 ligation buffer (NEB) and phosphorylated with T4 PNK (NEB). The annealed and phosphorylated oligos were cloned into the BbsI sites of the pSpCas9(BB)-2A-GFP plasmid ⁵⁹ and transformed in STBL3 E. coli cells. The Cas9-gRNA vector (1.0μg) was co-transfected with single-stranded oligodeoxynucleotide (ssODN, 4.0μg) into hiPSCs cultured on 6-well plate using 10 μL Lipofectamine Stemfect (ThermoFisher Scientific). At 24 hours after transfection, the cells were dissociated with TrypLE Express (ThermoFisher Scientific) and GFP+ cells were sorted and plated in 6-well plates at a density of 2,000 cells per well in E8 media supplemented with Rock inhibitor (Y-27632, 5 μM). Single cell clones were manually picked and expanded. To verify the on-target HDR, the clones were genotyped by direct PCR and Sanger sequencing.

METHOD DETAILS

High throughout optical voltage assay

hiPSC-CMs were plated on Matrigel at 25,000 cells per well of a 384-well plate (Greiner Bio-One) in 50 μl replating media. The subsequent day an additional 50 μl maturation media was added and cells were grown for a minimum of 4 days prior to analysis. For the analysis, hiPSC-CMs were removed from a 37°C 5% CO₂ incubator and placed immediately into a tissue culture cabinet on a dry heat block set to 37°C to prevent temperature fluctuation during the subsequent washing and dye loading steps. Cells were washed by removing 50 μl media and replacing it with 50 μl FluoroBrite DMEM (Gibco) solution five times. After the last wash step, 50 μL of the 2x VF2.1.Cl dye with Hoechst 33258 (Life Technologies) in FluoroBrite was added to each well ⁴⁵. The plate was returned to the 37°C 5% CO₂ incubator for 50 min. After incubation, cells were washed four times in FluoroBrite as described above. After a 20min recovery in the incubator, 2x reference drug was added to the respective well and incubated at 37°C and 5% CO₂ for 15 min before image acquisition. Time series images were acquired automatically using the IC200 KIC instrument (Vala Sciences) at an acquisition frequency of 33 Hz for a duration of 10 seconds. The voltage image analysis and physiological parameter calculation were conducted using commercially available Cyteseer (Vala Sciences) as previously described ⁴⁵.

Convolutional Neural Network Classifier

The convolutional neural network classifier was implemented using R interface to Keras ⁶⁰ and RStudio (PBC, Boston, MA http://www.rstudio.com/). A schematic view of our architecture is shown in Supplementary Fig. S1. Briefly, an input layer with 330 nodes (one per data point of the voltage trace) receives each normalized voltage trace. The following formula was used to normalize the traces $\Delta F_n=\frac{F\left(t\right)-F_{min}^{plate}}{F_{max}^{plate}-F_{min}^{plate}}$, where $F_{max}^{plate}$ and $F_{min}^{plate}$ are the maximum and minimum fluorescence intensity signal of all traces recorded from wells within a multiwell plate. Then two convolutional layers with 33 filters and kernel size five. Next, a max pool 1D layer with a pool size of three. After, two more convolutional layers with 66 filters and kernel size of ten. Later, a global average pooling. A dropout of 20% was added to decrease overfitting. Finally, a dense layer with 3 nodes and softmax activation for the output of the classifier. The experimental design, deep neural network classifier architecture and representative traces are described in Fig. 2A–C.

The network was trained on a training set comprising 645 “Non-arrhythmic”, 550 “Arrhythmic”, and 150 “Asystolic” traces (see examples in Fig. 2D), from the three healthy donor lines. The traces labelled as arrhythmic were from wells treated with CiPA high risk drugs at doses higher or equal than maximum free plasma concentration C_max. The traces labelled as non-arrhythmic were from wells treated with CiPA no-low risk drugs at doses lower or equal than maximum free plasma concentration C_max. The asystolic traces were manually identified in the dataset. The hyperparameters of the network were manually tuned to achieve 88.6% accuracy using Adam optimizer with categorical cross-entropy as loss function. After hyperparameter tuning, the trained network was tested on three different datasets to check performance against effects of unseen drugs in the same cell lines (test set 1), the effects of training drugs in the HCM/DCM lines (test set 2), and the effects of unseen drugs in HCM/DCM lines (test set 3). Accuracies and confusion matrices are shown in Fig. 2E.

QUANTIFICATION AND STATISTICAL ANALYSIS

28 independent differentiation batches of hiPSC were utilized in this study, some of them were rejected during quality control (Supplementary Fig. S1). Cell line name: included batches-rejected batches. HD.113: 3-1; rejected; HD.15S1: 3-0; HD.273: 5-0; HCM.TPM1: 4-2; HCM.MYBPC3: 3-2; DCM.TNNT2: 3-1; DCM.RBM20: 3-1; DCM.PLN: 3-0. All compounds of the library were tested in an 8-dose response curve with 3 technical replicates.

Probability of arrhythmic and asystolic dose-response curves

Dose-response curves were fitted using the binomial fit from the generalized linear model function in R ‘glm’. The logarithm of the dose was provided as ‘x-data’ and the probability of arrhythmic from the CNN classifier was provided as ‘y-data’.

Calculation of AUC of discretized TdP safety margin

The accuracy of using the Torsadogenic safety margin (Fig. 3G) against the discretized proarrhythmic classification of drugs was reported as the area under the curve of a binomial model with the Torsadogenic safety margin as its sole predictor.

Logit(P) = (Torsadogenic safety margin) ,where $\text{logit}\left(\text{P}\right)=\log \left(\frac{p}{1-p}\right)$ and P is the probability of a drug of being High-intermediate, High, or Intermediate (Fig. 3H). The models were trained following k-fold cross-validation with k=10, in R using the ‘glm.fit’ function. Receiver Operating Characteristic (ROC) curves were generated in R using the ‘plotROC’ package.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, Peptides, and Recombinant Proteins
CHIR99021	Tocris	Cat#4423
Wnt-C59	Tocris	Cat#5148
Y-27632 dihydrochloride	Tocris	Cat#1254
Amiodarone	Tocris	Cat#4095
Amitriptyline	Tocris	Cat#4456
Aspirin	Sigma-Aldrich	Cat#A5376-100G
Astemizole	Santa Cruz Biotechnology	Cat#sc-201088
Azimilide	Fisher Scientific	Cat#501151950
Bepridil	Santa Cruz Biotechnology	Cat#sc-202974
Chlorpromazine	Santa Cruz Biotechnology	Cat#sc-202537
Cisapride	Abcam	Cat#ab120548
Citalopram	Tocris	Cat#1427
Clarithromycin	Abcam	Cat#ab141202
Clozapine	Santa Cruz Biotechnology	Cat#sc-200402
Digoxin	Tocris	Cat#4583
Diltiazem	Santa Cruz Biotechnology	Cat#sc-200199
Disopyramide	Sigma-Aldrich	Cat#D6035-1G
Dofetilide	Tocris	Cat#3757
Domperidone	Santa Cruz Biotechnology	Cat#sc-203032
Droperidol	Fisher Scientific	Cat#585850
Epinephrine	Sigma-Aldrich	Cat#E4250-1G
Erythromycin	Sigma-Aldrich	Cat#E5389-1G
Flecainide	Tocris	Cat#1470
Fluoxetine	Tocris	Cat#0927
Ibutilide	Fisher Scientific	Cat#501362076
Loratadine	Fisher Scientific	Cat#501361172
Metoprolol	Abcam	Cat#Ab120711
Mexiletine	Tocris	Cat#2596
Nifedipine	Abcam	Cat#ab120135
Nitrendipine	Fisher Scientific	Cat#507407
Ondansetron	Fisher Scientific	Cat#501361204
Pimozide	Santa Cruz Biotechnology	Cat#sc-203662
Quinidine	Fisher Scientific	Cat#AAJ6042606
Ranolazine	Tocris	Cat#3118
Risperidone	Fisher Scientific	Cat#505885
Sotalol	Tocris	Cat#0952
Tamoxifen	Abcam	Cat#Ab120656
Terfenadine	Fisher Scientific	Cat#394850
Vandetanib	Selleckchem	Cat#S1046
Verapamil	Tocris	Cat#0654
Experimental Models: Cell Lines
Human iPSC line (healthy donor)	SCVI Biobank	SCVI-273
Human iPSC line (healthy donor)	SCVI Biobank	SCVI-15S1
Human iPSC line (healthy donor)	SCVI Biobank	SCVI-113
Human iPSC line DCM.TNNT2	(Perea-Gil et al., 2022)	TNNT2 R173W
Human iPSC line DCM.RBM20	(Briganti et al., 2020)	RBM20 R634Q
Human iPSC line DCM.PLN	(Feyen et al., 2021)	PLN R14del C9
Human iPSC line HCM.TPM1	(Perea-Gil et al. 2020)	TPM1 K37E 8
Human iPSC line HCM.MYBPC3	(Seeger et al., 2019)	MYBPC3 R943x
Oligonucleotides
Single guide RNA sequence for DCM.TNNT2: GGACAAAGCCTTCTTCTTCC	Integrated DNA Technologies
Single guide RNA sequence for HCM.TPM1: GGCGGAAGACAGGAGCAAGC	Integrated DNA Technologies
HDR sequence DCM.TNNT2: TAATTTGCTTTCTTCCTCCCTGCTGTAAATCAG GAAGAGAGGGCTCGACGAGAGGAGGAGGAGA ACAGGAGGAAGGCTGAGGATGAGGCATGGAA GAAGAAGGCTTTGTCCAACATGATGCATTTTG	ThermoFisher Scientific
HDR sequence HCM.TPM1: CCGAGTCCTCGGGGTGGGGATCCAGCCGGGG GTGCCAGGCTCGAGTCCCGGCGCTCTGGGCG CGGGCGCAGGGCCGGGGAGGCGCAGACCTG CTCGCTCCTGTCTTCCGCCGCCTTCTTGTCGG CCT	ThermoFisher Scientific
Recombinant DNA
Plasmid: pSpCas9(BB)-2A-GFP	Addgene	RRID:Addgene_48138
Software and Algorithms
CyteSeer	Vala Sciences	RRID:SCR_012232
RStudio	PBC	RRID:SCR_000432
Codes and Datasets	This study and Github	http://doi.org/10.5281/zenodo.7348974

Highlights:

• Deep learning detects in vitro features that correlate with clinical arrhythmia

• Drug safety margin from AI identifies drug risk with AUC = 0.95

• Cardiomyopathic genotypes increase sensitivity to drug proarrhythmia risk