SCH: Graph-based Spatial Transcriptomics Computational Methods in Kidney Diseases

慢性肾脏疾病（CKD）和急性肾损伤（AKI）是两种常见的交叉肾脏疾病。慢性肾病已被公认为全球主要的公共卫生问题，影响着全球约15%的人口。AKI可导致慢性肾病，每年在美国影响超过20万人，并在远端器官（如大脑、心脏和肺部）留下后遗症。为了更好地了解肾脏疾病的发病机制并潜在地预防AKI向CKD的转变，有必要确定细胞类型和状态的异质性、它们相关的分子特征以及微环境中复杂的相互作用。新兴的空间转录组技术（如10X Genomics Visium）产生了高通量的空间转录组数据，为了解肾脏健康和疾病中的异质细胞类型提供了见解。然而，与大脑等具有更大结构特征的器官相比，肾脏由100多万个肾单位组成，这些肾单位来自100多种细胞类型，排列得很近。在确定共定位细胞类型和阐明纤维化、免疫相互作用和上皮修复的机制方面，仍然存在巨大的计算挑战。为了填补这一空白，我们建议基于空间转录组数据开发基于人工智能的肾脏疾病研究计算方法。首先，我们将建立一个深度学习框架，以解决肾损伤中的异构、稀疏和镶嵌样细胞类型分布，并通过自监督学习训练风格的图神经网络进行授权。其次，我们将比较健康和损伤的细胞状态，以阐明CKD和AKI损伤的内在机制。第三，我们将通过可解释的生成过程预测肾损伤的影响。我们将利用人类细胞图谱（HCA）、人类生物分子图谱计划（HuBMAP）和肾脏精准医学项目（KPMP）中健康和损伤肾脏细胞图谱的多组学数据来评估这些方法的性能。我们的长期目标是创建一个生态社区，为肾脏研究中的医生和生物信息学家分析、共享和传播空间转录组学数据。

Chronic Kidney Disease (CKD) and Acute Kidney Injury (AKI) are two common intersecting kidney diseases. CKD has been recognized as a leading public health problem worldwide, affecting about 15% of the global population. AKI can lead to CKD and affects more than 200,000 individuals across the US annually, with sequelae in distant organs such as the brain, heart, and lungs. To better understand the pathogenesis of kidney disease and potentially prevent the transition of AKI into CKD, it is necessary to define the heterogeneity of cell types and states, their associated molecular signatures, and complex interactions within the microenvironment. Emerging spatial transcriptomic technologies (e.g., 10X Genomics Visium) generate high-throughput spatial transcriptome data, which provides insights into the heterogeneous cell types within kidney health and disease. However, in contrast to organs with larger structural features like the brain, the kidney is organized into over a million nephrons with representation from more than 100 cell types arranged in close proximity. There are still tremendous computational challenges in identifying the colocalizing cell types and elucidating mechanisms in fibrosis, immune interactions, and epithelial repair. To fill such gaps, we propose to develop AI-based computational methods for studying kidney diseases based on spatial transcriptome data. First, we will build a deep learning framework to address heterogeneous, sparse, and mosaic-like cell type distribution in kidney injury, empowered by graph neural networks in a self-supervised learning training style. Second, we will compare the healthy and injured cell states to illustrate the inherent mechanism beneath the injury of CKD and AKI. Third, we will predict the effects of kidney injury with an interpretable generative process. We will evaluate the methods’ performances using the multi-omics data from the cell atlas of the healthy and injured kidneys in the Human Cell Atlas (HCA), Human Biomolecular Atlas Program (HuBMAP), and Kidney Precision Medicine project (KPMP). Our long-term goal is to create an eco-community for analyzing, sharing, and disseminating spatial transcriptomics data for physicians and bioinformaticians in kidney research.