Generative AI to Profile Structure-Function Correlation of Human iPSC-Derived Cardiomyocytes

Introduction: With cardiovascular disease being a leading cause of death nationwide, it is imperative to develop new models capable of investigating the fundamental cardiomyocyte contractile physiology at the single cell level. Classic molecular models based on Frank-Starling law have been developed to understand the cardiac contractile apparatus based on myofilament length dependent activation. It is critical to establish a deep understanding of how the sarcomere structure of a cardiomyocyte can translate into complex contractile behaviors. Generative artificial intelligence (AI) has been explored to generate synthetic images of cells based on training data, which allows for identifying cellular structures, tracking changes over time, or detect abnormalities. In this study, we aim to develop a U-Net style Generative Adversarial Networks (GAN) deep learning model that allows us to gain deeper insight into the structure-function relationship of single cardiomyocytes. This GAN-based model learns to capture and replicate the contractile behaviors of individual cardiomyocytes from a single image of cell morphology or sarcomere structure. Our deep learning model is built in Pytorch utilizing pix2pix architecture comprised of a U-Net generator paired with a patch-GAN discriminator that learns a loss-function to train the mapping of the cell morphology into a quantifiable heatmap of contractile motion. This model creates a quantifiable and rapid methodology for investigating the structure and functioning of cardiomyocytes in an animal free model.

Materials and

Methods: Human induced pluripotent stem cells (hiPSCs) with ACTN2-GFP reporter for sarcomere structure were used to differentiate cardiomyocytes (hiPSC-CMs), which were then purified and seeded upon fibronectin micropatterned surfaces. Brightfield videos of contracting hiPSC-CMs were recorded at 50 frames per second for ten seconds and analyzed using video-based motion tracking software to compute the motion vectors based on pixel movement. This software generates contractile motion heatmaps that visually represent each cell’s total contractile behaviors. Meanwhile, fluorescent images based on ACTN2-GFP were taken from the same cells. Therefore, each hiPSC-CMs had a color heatmap image of contractile motion intensity and a single fluorescent image of sarcomere structure. The resulting dataset was randomly split into 75%-20%-5% training-validation-test subsets. The U-Net generator encodes the cellular images (brightfield morphology or fluorescent sarcomeres) through multiple layers into a latent space. The decoder translates the latent space of the original cells into generated heatmap images. A patch-GAN discriminator is utilized to penalize the patches of the generated images to focus the training of the generator to retain the true structure of the image (fig 1a). Thus, this U-Net-GAN deep learning model generates contractile heatmaps that visually represented each cardiomyocyte’s total contractile behaviors in an overlay of the cellular or molecular structure. Thus, we can generate contractile heatmaps that visually represents each cell’s total contractile behavior in an overlay of the cell’s cellular or molecular structure.

Results, Conclusions, and Discussions: We first built the model using traditional U-Net autoencoder network, and then added the patch GAN to the U-Net network, to predict the contraction heatmap from a brightfield cell morphology image or fluorescent sarcomere structure image (fig. 1b). We compared AI-predicted heatmaps to the ground truth heatmaps that were generated by the beating video analysis. Training for the predicted image was based upon the mean absolute error, L1 loss, of the predicted image compared to the ground true contractile heatmap for the cell. Furthermore, structural comparison of the images was achieved by using Structural Similarity Index Measure (SSIM) as a more qualitative measure in the final testing set. By comparing two deep learning models, we confirmed that U-Net-GAN model improved the predictivity of heatmap images with the patch GAN discriminator by boosting the accuracy of predicted heatmaps from 0.80 & 0.79 SSI (U-Net only) to 0.81 & 0.82 SSI (U-Net-GAN) for our testing set based on the cellular and molecular structures respectively. Furthermore, we can extract and isolate key features from the cell’s morphology and sarcomere structure by investigating the structures that result in high contraction within the cells. Comparing to traditional analysis that requires labor-intensive microscope recording and computing-intensive video analysis, this AI-based approach only needs a quick single snapshot of brightfield or fluorescent images to make the prediction of contractile behaviors of the cardiomyocytes.

Examining the fundamental contractile units of a single hiPSC-CM through generative AI is not only feasible but also yields functional outputs akin to traditional methods, significantly faster. This innovative approach could be potentially used for rapid high-throughput investigation into contractile dysfunction in human cardiomyopathies with genetic mutations.

Acknowledgements (Optional):