Machine learning methods for interpreting spatial multi-omics data

项目总结 拟议的研究计划旨在开发创新的计算工具，用于分析和 整合来自新兴的空间分辨基因组技术的数据，这些技术有可能揭示 与环境的相互作用在正常发育和疾病中的作用。现有的分析工具用于 分析空间组学数据的可解释性有限，并且不能集成多模式 数据。 利用我们在单元格数据计算建模方面的丰富经验 高维基因组数据类型，我们将以概率的形式设计机器学习框架 和深度生成模型来解决空间分辨基因组数据的分析挑战和 重要的是，集成了多种数据模式。该框架将使邻域识别成为可能 模式定义为具有唯一细胞状态组成的区域，来自空间剖面图的整合 MRNAs、蛋白质和组织成像(目标1)。我们将建立在基因调控建模的基础上 网络将开发第一个计算工具，用于从集成的 空间ATAC-seq和RNA-seq(目标2)。此外，我们将开发一个计算工具来推断 具有不同拷贝数轮廓的细胞及其相关基因程序的空间分布(目标3)。 我们通过应用我们的技术来强调我们的计算方法的通用性和通用性 与我们的合作者一起构建多种生物系统。这些应用将在以下方面提供新的见解 了解人类和小鼠胚胎发育的空间模式基础，脑器官 模型，以及神经精神障碍、胶质母细胞瘤和乳腺癌等疾病系统。我们的 我们的目标是将我们的计算工具箱作为开源软件传播给更广泛的基因组学社区 以揭示关于细胞类型的空间组织、它们的相互作用和机制的新见解 各种生物系统。

PROJECT SUMMARY The proposed research program aims to develop innovative computational tools for the analysis and integration of data from emerging spatially-resolved genomic technologies, which have the potential to uncover the role of interactions with the environment in normal development and disease. Existing analytical tools for analyzing spatial omics data are limited in their interpretability and are not capable of integrating multi-modal data. Leveraging our extensive experience in computational modeling of single-cell data as another high-dimensional genomic data type, we will design machine learning frameworks in the form of probabilistic and deep generative models to tackle the analytical challenges of spatially-resolved genomic data and importantly integrate multiple data modalities. This framework will enable the identification of neighborhood patterns defined as regions with a unique composition of cell states, from the integration of spatial profiling of mRNAs, proteins, and histological imaging (Aim 1). We will build on a foundation of modeling gene regulatory networks to develop the first computational tool for inferring spatially-varying regulation from the integration of spatial ATAC-seq and RNA-seq (Aim 2). Additionally, we will develop a computational tool for inferring the spatial distribution of cells with distinct copy number profiles, and their associated gene programs (Aim 3). We highlight the versatility and generalizability of our computational methods by applying our techniques in multiple biological systems with our collaborators. These applications will provide novel insights in understanding the basis of spatial patterns in human and mouse embryonic development, brain organoid models, as well as disease systems such as neuropsychiatric disorders, glioblastoma and breast cancer. Our goal is to disseminate our computational toolbox as open-source software to the broader genomics community to unlock novel insights about the spatial organization of cell types, their interactions, and mechanisms in various biological systems.