Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2023 Jan 5;30(1):86-95.e4.
doi: 10.1016/j.stem.2022.12.002. Epub 2022 Dec 22.

A deep learning platform to assess drug proarrhythmia risk

Ricardo Serrano  1 , Dries A M Feyen  1 , Arne A N Bruyneel  1 , Anna P Hnatiuk  1 , Michelle M Vu  1 , Prashila L Amatya  1 , Isaac Perea-Gil  2 , Maricela Prado  2 , Timon Seeger  1 , Joseph C Wu  1 , Ioannis Karakikes  2 , Mark Mercola  3
Affiliations

A deep learning platform to assess drug proarrhythmia risk

Ricardo Serrano et al. Cell Stem Cell. .

Abstract

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.

Keywords: AI; CiPA; artificial; cardiomyocytes; deep learning; drug screening; drug-induced arrhythmia; iPSC; induced pluripotent stem cells; intelligence; safety pharmacology.

PubMed Disclaimer

Conflict of interest statement

Declaration of interests R.S. is a paid consultant of Vala Sciences, which manufactures a high content instrument used in these studies. M.M. serves on the scientific advisory board of Vala Sciences. J.C.W. is co-founder and scientific advisory board member of Greenstone Biosciences.

Figures

Figure 1.
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.
Figure 2.
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.
Figure 3.
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, Cmax (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, Cmax (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 Cmax. 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.
Figure 4.
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.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

See all "Cited by" articles

References

    1. Valentin J-P, and Delaunois A (2018). Developing solutions to detect and avoid cardiovascular toxicity in the clinic. Toxicology Letters 295, S48–S49.
    1. Onakpoya IJ, Heneghan CJ, and Aronson JK (2016). Post-marketing withdrawal of 462 medicinal products because of adverse drug reactions: a systematic review of the world literature. BMC Med 14, 10. 10.1186/s12916-016-0553-2. - DOI - PMC - PubMed
    1. Waring MJ, Arrowsmith J, Leach AR, Leeson PD, Mandrell S, Owen RM, Pairaudeau G, Pennie WD, Pickett SD, Wang J, et al. (2015). An analysis of the attrition of drug candidates from four major pharmaceutical companies. Nat Rev Drug Discov 14, 475–486. 10.1038/nrd4609. - DOI - PubMed
    1. Baguley WA, Hay WT, Mackie KP, Cheney FW, and Cullen BF (1997). Cardiac dysrhythmias associated with the intravenous administration of ondansetron and metoclopramide. Anesth Analg 84, 1380–1381. 10.1097/00000539-199706000-00038. - DOI - PubMed
    1. Frommeyer G, and Eckardt L (2016). Drug-induced proarrhythmia: risk factors and electrophysiological mechanisms. Nat Rev Cardiol 13, 36–47. 10.1038/nrcardio.2015.110. - DOI - PubMed

Publication types

MeSH terms