Deep Learning Platform to Better Determine Risk of Drug-Induced Arrhythmias
by Amanda Chase, PhD
January 18, 2023
The technology to develop new medicines is ever changing, adapting, and growing to facilitate making more effective and safe treatments. Drug development is lengthy and expensive - costing upwards of $2.5 billion per new drug. One of the drivers of the high cost is the high failure rate during development or even after market launch. In many cases, this is due to unanticipated side-effects. Drugs that cause arrhythmias (irregular heartbeat) are one such cause of drug attrition during development, restricted use, or withdrawal from the market. Therefore, there is a real need to understand and/or know how drugs can affect patients, especially for those who are susceptible to arrhythmias. To add further complications, people vary in their predisposition to drug-induced arrhythmia, making it necessary to determine the risk in these susceptible populations.
Induced pluripotent stem cell derived cardiomyocytes (iPSC-CMs; heart cells) have the amazing ability to retain an individual’s genetic make-up and are widely heralded as a possible model for predicting drug-induced arrhythmia. However, predictivity of the human response to drugs has been much debated. In a recent publication in Cell Stem Cell, first author Ricardo Serrano and senior author Mark Mercola, along with several other Stanford Cardiovascular Institute researchers, built on the iPSC-CM platform to develop a deep learning approach that accurately predicted the clinical risk of drug-induced arrhythmia.
Development of the deep learning platform. iPSC-CMs from healthy donors pr with introduced mutations to be more at-risk for arrhythmias were used. Deep learning algorithm was developed to determine risk of drug-induced arrhythmias, training with established panel of drugs.
Currently, prediction of drug arrhythmia risk from iPSC-CM data rely on human-defined metrics – basically human interpretation of an arrhythmic signature. Unfortunately, these features are not always correct in predicting clinical arrhythmia. To address this problem, the researchers developed a sophisticated deep learning approach to remove human influence. They used a combination of iPSC-CMs derived from healthy patients and from lines with mutations known to cause heart problems (hypertophic cardiomyopathy, HCM; left ventricular noncompaction, LVNC; and dilated cardiomyopathy, DCM).
Patients with these kinds of structural heart disease have a higher risk for drug-induced arrhythmia. A panel of drugs known to have risk of causing arrhythmias was used to train and evaluate the deep learning platform. Impressively, the researchers could conclude that their platform could allow better categorization of risk of the drug into high-, medium-, or low-risk. Further, they could show that DCM and HCM genetic variants increased the risk of certain drugs causing arrhythmias.
The deep learning algorithm was shown to overcome limitations of human-defined metrics, which include uncertainty about what features the in vitro drug response predict actual clinical arrhythmia. This platform should improve the accuracy of detecting risky compounds during development before drugs are used in people. It also makes it possible to determine how gene variants contribute to the risk of developing arrhythmia after being treated with drugs. Overall, the deep learning platform provides a critically needed tool for both improving the drug development pipeline and determining the influence of patient genetics.
Other Stanford Cardiovascular Institute researchers are Dries Feyen, Arne Bruyneel, Anna Hnatiuk, Michelle Vu, Prashila Amatya, Isaac Perea-Gil, Maricela Prado, Timon Seeger, Joseph Wu, and Ioannis Karakikes.
Mark Mercola, PhD
Ricardo Serrano, PhD