Cancer precision medicine through spatially informative single cell image and transcriptomics data analysis

摘要 人类癌症是高度异质性的，由遗传、表观遗传、基因组和基因环境引起 交互.研究单细胞癌异质性对于有效诊断、预后和预后至关重要。 个性化抗癌治疗的发展。然而，单细胞水平的肿瘤异质性原位与 直到最近，随着新的前沿技术的发展， 空间转录组学（ST）和单细胞成像质谱细胞术（IMC）等平台。由于复杂性 在这些新的空间数据类型中，计算是使这些技术达到精度的主要瓶颈 在临床空间的治疗干预。在这个项目中，我们采取三管齐下的方法，提出一个 一系列新的计算方法，将利用空间信息组学和成像的力量 数据，用于药物治疗和癌症患者预后预测。建立在以前高生产力的基础上， R01项目，我们的目标是继续调查单细胞研究，新的重点是单细胞空间 数据分析首先，我们将开发一种新的个性化药物再利用算法，称为STADS， 癌症空间转录组学数据。接下来，构建新的计算模型STimpute以估算空间 转录组学数据从容易获得的组织病理学图像数据，使用迁移学习和图形神经 GNN模型。ST插补将允许通过使用插补的ST，根据组织病理学数据预测药物 数据作为STADS的代理输入。最后，我们将建立一个新的计算框架scImageProg来预测 从单细胞图像细胞术数据中获得群体水平的患者存活率，实现多尺度 建模将单细胞数据与人群健康联系起来。这项工作预计将产生变革性的临床效果 各种尖端的空间信息和复杂的基因组学和成像数据类型的影响。

Abstract Human cancers are highly heterogeneous, arising from genetic, epigenetic, genomic, and gene environment interactions. Studying single-cell cancer heterogeneity is essential for effective diagnosis, prognosis and development of personalized anti-cancer therapy. However, single-cell level tumor heterogeneity in situ with the original spatial context is not addressed until very recently, with development of new frontier technological platforms such as spatial transcriptomics (ST) and single cell imaging mass cytometry (IMC). Due to complexity of these new spatial data types, computation is a major bottle neck to bring these technologies to precision therapeutic interventions in the clinical space. In this project, we take a three-pronged approach to propose a series of novel computational methods that will harness the power of spatially informative omics and imaging data, for drug treatment and cancer patient prognosis predictions. Building upon a previously highly productive R01 project, we aim to continue investigations in single cell research, with a new focus on single-cell spatial data analysis. First, we will develop a novel personalized drug repurposing algorithm called STADS using cancer spatial transcriptomics data. Next, build a new computational model STimpute to impute spatial transcriptomics data from easily accessible histopathology image data, using transfer learning and graph neural network (GNN) models. STimpute will allow predictions of drugs from histopathology data, by using imputed ST data as the proxy input of STADS. Lastly, we will build a new computational framework scImageProg to predict patient survival at the population level from the single-cell image cytometry data, accomplishing multi-scale modeling to link single-cell data to population health. The work will be expected to have transformative clinical impacts from various cutting-edge spatially informative and complex genomics and imaging data types.