Computational Methods for Single Cell Biology

生物科学正在经历一场由高通量技术驱动的革命，该技术允许测量单细胞。一个突出的例子是使用基于测序的方法快速采用单细胞RNA分析来探索生物学的各个方面。最近，已经开发了其他单细胞测序方法，允许测量细胞的DNA和表观遗传特征。迄今为止，大多数方法都需要将细胞置于液体悬浮液中进行测序。这需要将组织分开，并丢弃空间信息，例如哪些细胞是相邻的。这使得研究细胞如何相互作用变得困难。单细胞生物学的前沿技术正在通过将高通量测量与成像或空间条形码相结合来消除这一限制。空间单细胞技术在促进我们对生物系统的理解方面具有巨大的潜力。然而，这些技术产生的数据正在产生现有计算工具无法充分解决的分析挑战。这些挑战的范围从信号处理中的低水平问题，如执行自动图像分析以识别细胞及其表达谱，到更高水平的问题，如从数据中提取生物学上感兴趣的量的可解释的估计。现成的机器学习方法并不能充分解决其中的许多问题。这一领域的两个最重要的特征挑战了现有的工具，这两个特征是缺乏良好注释的数据集来训练监督学习方法，以及生物学家希望拥有可解释的计算方法。此外，利用领域知识对问题施加额外结构的能力可以显著提高性能。我将开发新的计算和统计模型来分析多模态单细胞数据集，以提供生物系统的综合视图。我们的方法将使我们能够深入探索单个细胞的特征，同时测量它们存在的空间背景和环境。我开发的方法将通过允许肿瘤等复杂系统的空间单细胞分析来推动新的生物学发现。

The biological sciences are undergoing a revolution driven by high throughput technologies which allow for the measurement of single cells. A prominent example is the rapid adoption of single cell RNA profiling using sequencing based approaches to explore all facets of biology. More recently other single cell sequencing approaches have been developed, allowing for the measurement of the DNA and epigenetic features of cells. The majority of approaches to date have required that cells be put into a liquid suspension for sequencing. This requires that tissue be broken apart and spatial information such as which cells were neighbouring be discarded. This makes it difficult to study how cells interact with each other. The cutting edge of single cell biology is removing this limitation by combining high throughput measurements with imaging or spatial barcoding. Spatial single cell technologies have enormous potential for advancing our understanding of biological systems. However, the data generated by these technologies are creating analytical challenges that are not adequately addressed by existing computational tools. These challenges range from low level problems in signal processing such as performing automated image analysis to identify cells and their expression profiles, to higher level problems such as extracting interpretable estimates of biologically interesting quantities from the data. Off the shelf machine learning approaches do not adequately address many of these problems. The two most important characteristics of this domain that challenge existing tools are the lack of well annotated datasets to train supervised learning approaches and the desire by biologists to have interpretable computational methods. In addition, the ability to leverage domain knowledge to impose additional structure on problems can significantly improve performance. I will develop novel computational and statistical models to analyze multi-modal single cell datasets to provide an integrated view of biological systems. Our approach will allow us to deeply probe the features of individual cells, while also measuring the spatial context and environment they exist in. The methods I develop will drive new biological discoveries by allowing spatial single cell profiling of complex systems such as tumours.