Causal Representation Learning for the Spatial Analysis of Transcriptomic and Imaging Data in Tissue Contexts

美国国立卫生研究院新创新者奖 摘要 通过融合成像和基因组学，现在有可能获得空间分辨的转录组 数据集；然而，用于分析此类数据集的计算方法已经滞后 实验发展。为了充分发挥空间转录(ST)数据的潜力，我们 不能依赖已经开发的用于分析离婚的单细胞数据的方法 从微环境中分离出来的细胞。就像看到ST 通过融合成像和测序的突破，我们认为同样的道理也将适用于 计算领域，因此提出了一个分析这些数据的框架，该框架 将成像和测序与因果关系相结合，以推断潜在的调控机制 空间驱动的过程。 我们建议通过创新地统一机器中两个充满活力的领域来实现这一点 学习(ML)、表征学习和因果推理。这是一项重大任务，因为 表征学习虽然在像推荐系统这样的预测性任务中取得了成功， 一般不会阐明因果关系。为了克服这一点，我们将使用表示法 学习识别在ST中可用的所有数据形态中存在的相关性，以及 从而利用不变性原理从因果关联中辨别出伪关联。在……里面 此外，我们还将在ML中构建三个基本概念： -图像修复：识别组织结构中的主题以及异常组织块 -最佳传输：从快照中及时推断组织谱系 -因果结构发现：识别监管模块并预测扰动的影响 这种统一将产生一个将空间、时间和表达集成到一起的ML框架 确定空间过程背后的生物学机制。尽管这个框架将是 广泛适用，它以三种疾病背景为中心，这将作为前景 为了测试和改进我们的方法，并且已经获得了ST数据： -肠道炎症/纤维化；研究细胞募集、基质沉积和清除； -阿尔茨海默病；研究分泌物和蛋白质聚集问题；以及 -经典霍奇金淋巴瘤；研究肿瘤-免疫细胞相互作用和免疫侵袭。 了解这些疾病背景下细胞间通讯的调节机制 有可能产生新的治疗靶点，可以在合作伙伴关系中得到验证 与我们的实验合作者合作，并造福于患者的生活。

NIH New Innovators Award Abstract By melding imaging and genomics it is now possible to obtain spatially resolved transcriptomic datasets; however, computational methods for analyzing such datasets have lagged behind experimental developments. To realize the full potential of spatial transcriptomic (ST) data, we cannot rely on the methods that have been developed for analyzing single cell data that divorce cells from their microenvironment. As with experimental developments that saw ST breakthroughs by melding imaging and sequencing, we argue that the same will hold true in the computational domain, and, therefore, propose a framework for the analysis of this data that integrates imaging and sequencing with causality to infer regulatory mechanisms underlying spatially driven processes. We propose to achieve this through an innovative unification of two vibrant areas in machine learning (ML); representation learning and causal inference. This is a momentous task since representation learning, although successful in predictive tasks like recommender systems, does not generally elucidate causal relationships. To overcome this, we will use representation learning to identify correlations that are present in all data modalities available in ST, and thereby discern spurious correlations from causal ones using the principle of invariance. In addition, we will build on three fundamental concepts in ML: - Image inpainting: to identify motifs in tissue architecture as well as anomalous tissue patches - Optimal transport: to infer tissue lineages from snapshots in time - Causal structure discovery: to identify regulatory modules & predict the effect of perturbations This unification will result in an ML framework that integrates space, time, and expression to identify biological mechanisms underlying spatial processes. Although this framework will be broadly applicable, it is centered on three disease contexts, which will serve as the foreground to test and refine our methods and for which ST data have already been obtained: - Inflammation/fibrosis in the gut; to study cell recruitment, matrix deposition, and clearance; - Alzheimer's disease; to study questions of secretion and protein aggregation; and - Classic Hodgkin lymphoma; to study tumor-immune cell interactions & immunological invasion. Understanding the regulatory mechanisms of cell-cell communication in these disease contexts has the potential to give rise to new therapeutic targets that could be validated in partnership with our experimental collaborators and benefit patients' lives.