Machine Learning Models of Phase Contrast Images Predict Efficiency of Human Pluripotent Stem Cell Differentiation to Cardiomyocytes

Terminal cell types derived from human pluripotent stem cells (hPSCs) are at the forefront of emerging cell and gene therapy products. hPSC-derived cardiomyocytes (hPSC-CMs) are of particular interest in understanding and treating heart disease, which is highly prevalent worldwide; however, hPSC-CM manufacturing robustness is a roadblock to these applications. Non-destructive methods to monitor hPSC-CMs and predict differentiation efficiency throughout the process are needed. Here, we demonstrate a supervised machine learning approach utilizing random projection-based feature embedding and Linear Discriminant Analysis that can predict Day 12 CM purity from phase contrast images as early as 1 day after initiation of differentiation. In contrast, a heuristic based on a live-cell cTnT-GFP reporter provided limited information until Days 7-9. In summary, we provide evidence that machine learning analysis of phase contrast images is a promising approach for predicting hPSC-CM differentiation efficiency during early differentiation stages. Future efforts could use these models to inform the improvement of CM differentiation protocols and support CM biomanufacturing.