Differentiating Engineered Tissue Images and Experimental Factors to Classify Cardiomyocyte Content.

Abstract

In this study, we used machine learning (ML) to classify the cardiomyocyte (CM) content on day 10 of the differentiation of human-induced pluripotent stem cell (hiPSC)-laden microspheroids using easily acquirable nondestructive phase-contrast images taken in the middle of differentiation and tunable experimental parameters. Scale-up suspension culture, use of engineered tissues to support stem cell differentiation, and CM production for improved control over cellular microenvironment in the suspension system need nondestructive methods to track engineered tissue development. The ability to couple images that capture experimenter perceived "good" or "bad" batches based on visualization at early differentiation time points with actual experimental outcomes in an unbiased way is a step toward building these methods. In recent years, ML techniques have been successfully applied to identify critical process parameters and use this information to build models that describe process outcomes in cell production and hiPSC differentiation. Building upon these successes, here, we utilize convolutional neural networks (CNNs) to build a binary classifier model for CM content on differentiation day 10 (dd10) for hiPSC-CMs. We consider two separate data sets as potential input features for the classification models. The first set includes phase-contrast images of microspheroid tissues taken on days 3 and 5 of the differentiation batches at different experimental conditions. The second set supplements the images with tunable experimental differentiation parameters, such as cell concentration and microspheroids' size. The CM content classes were sufficient and insufficient. The accuracy of the CNN classifier using images only was 63%. The addition of experimental features increased the accuracy to 85%, indicating the importance of tunable parameters in predicting CM content. Impact statement Machine learning approaches were used to predict the final cardiomyocyte (CM) content class (sufficient vs. insufficient) of engineered cardiac tissue microspheroids produced through suspension-based cardiac differentiation of human-induced pluripotent stem cell-laden engineered tissue microspheroids. The models used specified experimental features and data collected using nondestructive inexpensive methods, specifically phase-contrast images taken during the initial days of differentiation as inputs. The best model was a convolutional neural network trained using experimental features and differentiation day 5 images. It classified the CM content with 85% accuracy and replicated and formalized experimenter's visual intuition about differentiation outcomes by incorporating images from early time points.