Abstract P2109: Elaborating Safety Margins To Predict Drug Proarrhythmia Using Deep Learning And Patient-derived IPSCs.

Introduction: Drug-induced arrhythmias are a common cause for drug attrition during development and for restricted use or withdrawal from the market. Cell based assays to assess arrhythmia risk typically rely on the quantification of waveform features - e.g. action potential prolongation and after depolarizations - in the cells’ action potential. However, the predictive power of these approaches is limited. Hypothesis: We hypothesize that deep learning can extract features relevant to discriminating input classes in a systematic and unbiased manner, effectively removing the need for human-defined metrics. This can lead to a new model to estimate torsadogenic risk of drugs and evaluate the influence of myopathic gene variants on drug-induced arrhythmia. Methods: We optically recorded action potentials optically recorded for 40 drugs - characterized high, intermediate, and low or no torsadogenic risk in patients- at 8 concentrations in hiPSC-CMs from 3 healthy donors and 5 hiPSC-CMs isogenic lines carrying 5 gene variants that cause dilated and hypertrophic cardiomyopathies. We designed a convolutional neural network (CNN) to classify non-arrhythmic, arrhythmic and asystolic traces in hiPSC-CMs. Using the class probabilities measured by the CNN, we created torsadogenic and asystolic safety margins for each drug and cell line. Results: The arrhythmic class probability computed by the CNN, provided a continuous, dose-dependent metric of the proarrhythmic risk of drugs in healthy and cardiomyopathic hiPSC-CMs. We used this metric to estimate safety margins for drug-induced arrhythmia and achieved a 0.942 AUC in classifying drugs of high-intermediate risk from safe ones. We used this approach to discern the contribution of putative genetic risk factors to arrhythmia susceptibility by comparing the risk profiles of the same drugs in healthy and isogenic hiPSC-CMs carrying causal HCM and DCM gene variants that are associated with arrhythmia in patients. Conclusions: We conclude that deep learning algorithms can effectively evaluate proarrhythmic risk of small molecules. Moreover, they can also be used to discern heightened arrhythmic risk caused by genetic mutations that increase the propensity for drug-induced arrhythmia in patients.