Deciphering Genetic and Epigenetic Regulatory Logic of Germ Layer Differentiation with Manifold Learning

项目概要： 深入了解控制人类早期发育的遗传和表观遗传调控逻辑 mans 对于揭示发育性疾病的机制和设计新方案至关重要 用于再生医学应用。尽管多年来许多对发育很重要的基因， 目前还没有系统地了解这些基因如何动态相互作用以产生细胞 和有机体表型。为此，我们建议将实验和计算相结合 开发人类胚胎干早期胚层发育预测模型的方法 细胞（hESC）。在我们的前期工作中，我们生成了一个单细胞 RNA 测序 (scRNA-seq) 数据集 31,000 个 hESC 在 27 天的时间内以类胚体 (EB) 的形式生长，以观察分化为 不同的细胞谱系。我们开发并应用了一种新的降维和可视化方法 将此系统称为 PHATE，发现 PHATE 生成全面且可解释的 差异化的图片。它捕获早期发育的所有分支，包括 ESC、神经嵴细胞 及其衍生物、神经祖细胞以及中胚层和内胚层细胞。建立在 根据这些发现，我们建议将这项研究扩展到 60 天的时间过程，并使 PHATE 更具规模化 能够捕捉到更成熟谱系的分化。然后我们建议整合 scRNA-seq 和 表观遗传数据，通过将排序群体的批量 CHIP-seq 测量值插值到伪单 细胞分辨率。最后，为了理解指导分化的基因调控逻辑 针对特定谱系，我们将训练一种新的神经网络架构，称为 DyMon（动态建模） 网络），通过数据流形来学习胚层解码的预测计算模型 其隐藏层的发展。因此，我们将基因调控逻辑重新布线与发育联系起来 细胞表型并提供在此过程中重编程的见解。

Project Summary: A deep understanding of the genetic and epigenetic regulatory logic that controls early development in hu- mans is essential for uncovering the mechanisms of developmental diseases and designing new protocols for regenerative medicine applications. Although over the years many developmentally important genes, there has not been a systematic understanding of how these genes interact dynamically to create cellular and organismal phenotype. For this purpose, we propose to combine experimental and computational approaches to develop predictive models of early germ layer development from human embryonic stem cell (hESC). In our preliminary work, we generated a single-cell RNA-sequencing (scRNA-seq) dataset of 31,000 hESCs, grown as embryoid bodies (EBs) over a period of 27 days to observe differentiation into diverse cell lineages. We developed and applied a new dimensionality reduction and visualization method called PHATE to this system and discovered that PHATE generates a comprehensive and interpretable picture of differentiation. It captures all branches of early development, including ESCs, neural crest cells and their derivatives, neural progenitors, and cells of the mesoderm and endoderm layers. Building upon these findings, we propose to extend this study to a 60-day time course and rendering PHATE more scal- able to capture differentiation to more mature lineages. Then we propose to integrate scRNA-seq and epigenetic data, by interpolating bulk CHIP-seq measurements on sorted populations to a pseudo single- cell resolution. Finally, in order to understand the gene regulatory logic that guides differentiation along specific lineages, we will train a new neural network architecture known as DyMon (dynamics modeling network), to walk through the data-manifold to learn a predictive computational model of germ layer de- velopment in its hidden layers. Thus we will connect gene regulatory logic rewiring with developmental cellular phenotypes and offer insights into reprogramming during this process.