Using Deep Learning to Predict Induced Pluripotent Stem Cell-Derived Cardiomyocyte (iPSC-CM) Differentiation Outcomes

摘要 人类诱导多能干细胞衍生的心肌细胞（iPSC-CM）提供了变革性的新途径 来对抗心脏病它们已被广泛用于疾病机制建模和药物预测。 应答尽管iPSC-CM分化取得了重大进展，但分化结果仍各不相同 不同批次、细胞系和方案之间的差异，导致显著的实验变异性。因此，在本发明中， 区分结果的特征至关重要。表征iPSC-CM的现行标准 分化结果涉及在分化期间监测心肌细胞的功能或遗传属性。 分化然而，这些过程非常耗时、不精确或昂贵。发现 更早的时间点和可扩展的，准确的标记物，可以确定分化结果是至关重要的， 消除了心肌细胞分化的严重瓶颈，并创造了更好的iPSC-CM材料。给定 随着iPSC-CM在心血管研究中的重要性日益增加，iPSC-CM分化的进展将广泛 加快寻找治愈方法。 在这里，我建议利用人工智能（AI），深度学习和计算机视觉来开发可扩展的， 用于表征和预测iPSC-CM分化结果的准确方法。为此，我将首先 使用深度学习模型来识别可用于预测和确定的标记和时间点 差异化结果。我已经区分了iPSC-CM，并获得了每天的图像数据集。 这些分化标记有其最终分化结果。我将使用一个深度 学习模型-图像分类器-以确定可用于预测的最早时间点 差异化结果。然后，我将把结果与转录组学数据相关联，以获得机制性的见解。到 评估深度学习方法是否是预测分化的可扩展和准确模型 结果，我将区分遗传多样性的iPSC系，以创建额外的数据集来微调模型 然后评估模型的可扩展性潜力。为了验证模型的准确性，我将首先验证 该模型的预测与分化心肌细胞的传统功能和遗传标记一致。 然后，我将从功能上、形态上和遗传上比较模型做出的预测与现有的预测。 评估差异化结果的方法。 该提案的完成将消除iPSC-CM差异化的主要瓶颈;创建广泛的iPSC- CM分化“形态图谱”;加速AI和深度学习在iPSC-CM中的应用。 此外，概述的培训将为我提供计算和再生医学专业知识 后来成为一名独立的研究者和医生科学家。

ABSTRACT Human induced pluripotent stem cell derived cardiomyocytes (iPSC-CMs) provide transformative new avenues to combat heart diseases. They have been used extensively to model disease mechanisms and predict drug responses. Despite significant advancements in iPSC-CMs differentiation, differentiation outcomes still vary across batches, cell lines, and protocols, resulting in significant experimental variability. As a result, characterizing differentiation outcomes is essential. The current standards to characterize iPSC-CM differentiation outcomes involve monitoring for functional or genetic attributes of cardiomyocytes during the differentiation. However these processes are prohibitively time consuming, imprecise, or expensive. Discovering earlier time points and scalable, accurate markers that can determine differentiation outcomes is critical for eliminating a severe bottleneck in cardiomyocyte differentiation and creating better iPSC-CM materials. Given the rising importance of iPSC-CMs in cardiovascular research, advances in iPSC-CM differentiation would widely accelerate the search for cures. Here, I propose leveraging Artificial Intelligence (AI), deep learning and computer vision to develop scalable, accurate methods for characterizing and predicting iPSC-CM differentiation outcomes. To achieve this, I will first use deep learning models to identify markers and time points that can be used to predict and determine differentiation outcomes. I have differentiated iPSC-CMs and obtained a dataset of images at each day of differentiation that are labeled with their final differentiation outcome. I will analyze the dataset using a deep learning model—an image classifier—to determine the earliest time point that can be used to predict differentiation outcomes. I will then correlate results with transcriptomic data to gain mechanistic insight. To evaluate whether deep learning methods are scalable and accurate models for predicting differentiation outcomes, I will differentiate genetically diverse iPSCs lines to create an additional dataset to fine tune the model and then evaluate the model’s potential for scalability. To validate the accuracy of the model, I will first verify that the model’s predictions align with conventional functional and genetic markers of differentiated cardiomyocytes. Then, I will functionally, morphologically, and genetically compare predictions made by the model against existing methods for evaluating differentiating outcomes. Completion of this proposal will eliminate a main bottleneck in iPSC-CM differentiation; create an extensive iPSC- CM differentiation ‘morphology atlas’; and accelerate the application of AI and deep learning to iPSC-CMs. Additionally, the outlined training will provide me with the computational and regenerative medicine expertise required to later succeed as an independent investigator and physician scientist.