Uncovering the latent structure in single-cell multi-omics data to study human diseases

抽象和简化是研究复杂系统的基本和强大的概念。在生物学中，我们将细胞建模为“RNA袋”，将组织建模为“细胞袋”。这些简化使科学家们能够做出重要的发现，有时甚至为癌症等致命疾病带来挽救生命的靶向治疗选择。有一些成功的案例使用药物靶向癌细胞中高度富集的特定蛋白质。不幸的是，这些都是罕见的例外事件，而不是一般规则。没有对组织细胞组织的深刻理解，就很难在开发广泛适用于组织疾病（如癌症，一种多细胞生物体疾病）的治疗方案方面取得重大进展。我的长期目标是开发计算和统计方法来整合单细胞基因组学数据和系统生物学数据，以获得对人体组织复杂细胞组织的机械理解。组织中的细胞组织与组织功能有关，而不仅仅是“细胞袋”。单细胞基因组学技术改变了我们研究细胞和组织生物学的方式。这些分析的持续发展将有助于我们了解人体组织的设计规则。我的短期目标将集中在揭示单细胞基因组数据中可解释的潜在结构。具体来说，我的实验室将专注于四个目标：1）使用神经网络搜索单细胞基因组学数据的统计模型空间; 2）通过考虑数据几何学来学习可解释的模型; 3）提供校准的不确定性输出; 4）使用空间分辨和多组学数据来研究表达异质性和细胞相互作用。除了发表我们的发现外，我们还将发表相关的软件和数据，以进行可重复的研究。学术界最令人愉快的活动之一是吸引和培养年轻科学家，激发他们的兴趣，看着他们取得进步，发展更高的推理能力，并获得独立性。该研究项目将培养两名博士。研究人员，5名硕士研究人员和10多名本科生研究人员。HQP将领导拟议的研究，具有挑战性但可行。我们的培训策略确保所有的HQP都有一个坚实的基础，并得到公平的培训。单细胞基因组学是一个快速发展的领域。不幸的是，我们看到许多研究误解了他们的数据，这可能导致不必要的昂贵和误导性的实验。拟议的研究将为生物学发现提供可解释的预测和校准的不确定性信息。我希望我们的研究也对一般生物数据分析和机器学习领域产生广泛的影响。此外，鉴于我们模型的可扩展性和灵活性，我们希望它们对人类细胞图谱等国际项目至关重要。

Abstractions and simplifications are essential and powerful concepts for studying complex systems. In biology, we model a cell as a 'bag of RNAs' and a tissue as a 'bag of cells'. These simplifications have enabled scientists to make important discoveries and sometimes even lead to life-saving targeted treatment options for deadly diseases such as cancer. There are successful cases of using drugs to target a specific protein that is highly enriched in cancer cells. Unfortunately, these are rare exceptional events instead of the general rule. Without a deep understanding of tissue cellular organization, it is hard to make significant progress in developing widely applicable treatment options for tissue diseases such as cancer, a disease of multi-cellular organisms. My long-term goal is to develop computational and statistical methods to integrate single-cell genomics data and systems biology data to gain a mechanistic understanding of the complex cellular organization of human tissues. The cellular organization in tissue is linked to tissue function and is not simply a 'bag of cells'. Single-cell genomics technologies have transformed our ways of studying cell and tissue biology. The continued developments of these assays will help us understand the design rules of human tissues. My short-term objectives will focus on uncovering the interpretable latent structure in single-cell genomic data. Specifically, my lab will focus on four objectives: 1) using neural networks to search the space of statistical models for single-cell genomics data; 2) learning interpretable models by considering data geometry; 3) providing calibrated uncertainty outputs; 4) using spatially resolved and multi-omics data to study expression heterogeneity and cellular interactions. In addition to publishing our discoveries, we will also publish the associated software and data for reproducible research. One of the most enjoyable activities in academia is attracting and nurturing young scientists, sparking their interests, watching them making progress, developing higher reasoning skills, and gaining independence. This research program will train two Ph.D. researchers, five M.Sc. researchers, and 10+ undergraduate researchers. HQP will lead the proposed research, challenging yet feasible. Our training strategies make sure all HQP receive a solid foundation and are equitably trained. Single-cell genomics is a rapidly evolving field. Unfortunately, we have seen many studies misinterpreting their data, which may lead to unnecessarily expensive and misleading experiments. The proposed research will provide interpretable predictions and calibrated uncertainty information for biological discoveries. I expect our studies also have a broad impact on general biological data analyses and the field of machine learning. Moreover, given the scalability and flexibility of our models, we hope they will be essential for international projects such as the Human Cell Atlas.