Reverse engineering cell competition using automated microscopy and recurrent neural networks

这个项目的目的是使用最先进的机器学习(ML)、自动延时显微镜和蛋白质组学来了解细胞竞争。细胞竞争是一种导致组织中不太适合的细胞被清除的现象--这是发育、体内平衡和疾病的关键过程。失败细胞的生存能力强烈依赖于环境：当它们单独培养时，它们会茁壮成长，但当它们处于混合种群中时，它们会被更适合的细胞淘汰。在发展中，竞争既是一种质量控制机制，也参与了格局的形成。在衰老过程中，竞争可能会消除组织中的衰老细胞，以防止与年龄相关的病理。在干细胞的缝隙中，竞争可能决定哪些细胞分化，哪些保持多能性。到目前为止，已经确定了许多细胞竞争的机制，涉及生化竞争(例如，通过竞争有利于生存的生长因子)或机械竞争(例如，快速生长的克隆压缩生长缓慢的克隆中的细胞，这导致现在密度更高的缓慢生长的克隆的细胞被挤出)。虽然竞争最初被认为只发生在细胞谱系之间的界面上，但机械竞争的发现揭示了情况并不一定是这样，并且挤出可能发生在距离该界面几个细胞直径的地方。迄今为止，绝大多数研究都研究了单细胞竞争的生化机制和种群水平的竞争，然而，越来越清楚的是，组织的拓扑结构在决定竞争结果方面发挥着核心作用。尽管如此，细胞竞争仍然知之甚少--我们不知道决定每个细胞命运的相互作用“规则”。这在很大程度上是因为大多数研究只量化了很少的时间点和少数细胞的整个种群转移。理解单个细胞行为如何导致种群转移的一个主要障碍是，它需要在数百个时间点上跟踪数千个细胞。为了应对这一挑战，我们最近建立了第一个深度学习和自动化单细胞显微镜系统来分析细胞竞争。我们使用深度卷积神经网络分析了数百万个单细胞在机械竞争中的细胞周期状态，包括细胞分裂和死亡。在这个项目中，我们将使用我们的时移数据中包含的全部信息来确定控制细胞竞争的物理和拓扑参数。我们将开发一种深度学习方法来提取单细胞环境中预测其在生化和机械竞争中的命运的时间依赖特征。我们将使用ML模型来确定控制细胞竞争的物理和拓扑特征。我们将结合ML和蛋白质组学来确定参与承诺途径的蛋白质，确定它们在信号级联中的层次结构，并确定收敛途径。

The aim of this project is to use state-of-the-art machine learning (ML), automated time-lapse microscopy, and proteomics to understand cell competition. Cell competition is a phenomenon that results in the elimination of less fit cells from a tissue - a critical process in development, homeostasis and disease. The viability of loser cells depends strongly on context: when they are cultured alone, they thrive, but when in a mixed population, they are eliminated by cells with greater fitness. In development, competition acts as a quality control mechanism and also participates in pattern formation. In ageing, competition may eliminate senescent cells from tissues to prevent age-related pathologies. In stem cell niches, competition may determine which cells differentiate and which remain pluripotent. A number of mechanisms of cell competition have been identified to date involving either biochemical competition (for example through competition for pro-survival growth factors) or mechanical competition (for example a fast growing clone compresses cells in a slow growing clone, which results in cell extrusion for the now denser slow growing clone).While competition was initially thought to take place only at the interface between cell lineages, the discovery of mechanical competition revealed that this is not necessarily the case and that extrusion may take place several cell diameters away from this interface.To date, the vast majority of studies have examined the biochemical mechanisms of competition in single cells and competition at the population level, however it is becoming increasingly clear that the topology of the tissue plays a central role in determining the outcome of competition. Despite this, cell competition remains poorly understood -- we do not know the interaction "rules" that determine each cell's fate. This is largely because most studies only quantify whole population shifts for very few time points and for few cells. One major obstacle to understanding how population shifts occur as a result of single cell behaviours is that it requires thousands of cells to be tracked over hundreds of time points. To address this challenge, we recently built the first deep learning and automated single-cell microscopy system to analyse cell competition. We used deep convolutional neural networks to analyse the cell cycle state of millions of single cells in mechanical competition, including cell division and death.In this project, we will use the full scope of the information contained in our time-lapse data to determine the physical and topological parameters that govern cell competition. We will develop a deep learning approach to extract time-dependent features of a single-cell's environment that predicts its fate in biochemical and mechanical competition. We will use the ML model to determine what physical and topological features govern cell competition. We will combine ML and proteomics to identify proteins involved in the commitment pathway, determine their hierarchy in the signalling cascade, and identify convergent pathways.