Machine-learning-based quality control of contractility of cultured human-induced pluripotent stem-cell-derived cardiomyocytes

Abstract

The precise and early assessment of cardiotoxicity is fundamental to bring forward novel drug candidates to the pharmaceutical market and to avoid their withdrawal from the market. Recent preclinical studies have attempted to use human-induced pluripotent stem-cell-derived cardiomyocytes (hiPSC-CMs) to predict clinical cardiotoxicity, but the heterogeneity and inconsistency in the functional qualities of the spontaneous contractility of hiPSC-CMs across cell culture wells and product lots still matter. To rapidly assess the functional qualities of hiPSC-CMs without histological labeling, we optically detected the contractility of confluently cultured hiPSC-CMs using bright-field microscopy. Using a method that consisted of data preprocessing, data augmentation, dimensionality reduction, and supervised learning, we succeeded in precisely discriminating between functionally normal and abnormal contractions of hiPSC-CMs.