A Bayesian nonparametric approach to superresolved tracking of multiple molecules inside living cells

项目概要： 2014年诺贝尔化学奖被授予在荧光标记，仪器和分析方面的进展。 在过去的十年中，这些方法一起将粒子位置分辨到20-30 nm内。 也就是说，低于用于激发它们的光的衍射极限。超分辨率后来被用于 对与神经退行性疾病相关的β-淀粉样蛋白纤维进行成像，并直接观察衍射限制蛋白 与癌症表型相关的聚类。 虽然超分辨定位揭示了与健康直接相关的静态细胞结构， 不能提供对疾病动力学的直接洞察。直接观察单个分子的体内动力学 水平要求多粒子超分辨粒子跟踪。超分辨跟踪比 因为对于收集的相同数量的光子，跟踪需要移动的 粒子被定位在多个图像帧上。此外，多粒子超分辨跟踪重新， 要求这一切都要做到，同时考虑到不可避免的重叠粒子轨迹内的一个有限的 细胞体积在尺寸上是几个衍射极限体积。因此，到目前为止，还没有系统的方法来准确地 在一个E大小的体积内，追踪数百万蛋白质中的一种以上的蛋白质。coli的细胞质中。 因此，总体目标是：提供第一个原则性的多粒子超分辨轨道- 通过利用贝叶斯非参数化（BNP）的新工具， 在过去的十年里，影响了数据科学。BNP可以学习每帧中的粒子数量， 以计算上易处理的方式，以直接由 数据（每像素收集的光子）。所开发的跟踪方法将应用于多粒子问题 - 如丝氨酸化学感受器Tsr复合物在E.大肠杆菌内膜 - 以及涉及突然的动力学变化的问题--例如， RNA聚合酶-自然处理的原则跟踪框架提出。 提出了两个具体目标。具体目标I -开发第一个完全集成和无监督的， 多衍射极限粒子超分辨跟踪算法 以一个（未知的）扩散系数扩散。具体目标II -针对该病例重复具体目标1 其中粒子演化所依据的动力学模型是未知的或者甚至随时间变化（即， 解除了动力学受简单扩散控制的限制）。在每个目标中，我们将：确定 通过调整（非参数）伯努利过程，调整观测模型， 解释复杂的标签物理学和对快速移动粒子重要的混叠伪影;处理粒子 通过适应Dirichlet学习动力学模型时，小细菌细胞中粒子扩散的确认 过程;包含详细的相机噪声模型。

Project Summary: The 2014 Chemistry Nobel Prize was awarded for advances in fluorescent labeling, instrumentation and anal- ysis methods which together, over the last decade, have resolved particle positions to within ≈20-30 nm. That is, below the diffraction limit of light used to excite them. Superresolution has subsequently been used to image β-amyloid fibers tied to neurodegenerative disorders and directly observe diffraction limited protein clustering linked to cancer phenotypes. While superresolved localization reveals static cellular structures of immediate relevance to health, it does not provide direct insight into disease dynamics. Directly observing in vivo dynamics at the single molecule level demands multi-particle superresolved particle tracking. Superresolved tracking is more difficult than superresolved localization because – for the same number of photons collected – tracking requires mobile particles to be localized over multiple image frames. Furthermore, multi-particle superresolved tracking re- quires that this all be done while accounting for unavoidable overlapping particle trajectories within a confined cellular volume a few diffraction limited volumes in size. Thus, to date, there is no systematic way to accurately track more than one protein, of the millions of proteins, inside a volume the size of E. coli’s cytoplasm at once. The overarching goal is therefore: To provide the first principled multi-particle superresolved track- ing algorithm by exploiting the novel tools of Bayesian nonparametrics (BNPs) that have already deeply impacted Data Science over the last decade. BNPs can learn particle numbers in each frame and particle trajectories across all frames in a computationally tractable manner in a way that is directly informed by the data (photons collected per pixel). The tracking method developed will be applied to multi-particle problems – such as the assembly/disassembly of serine chemoreceptor, Tsr, complexes on E. coli’s inner membrane – and problems involving abrupt dynamical changes – such as transitions between bound/unbound states of RNA polymerases – naturally dealt with in the principled tracking framework proposed. Two Specific Aims are proposed. Specific Aim I – Develop the very first, fully-integrated and unsupervised, superresolved tracking algorithm for multiple diffraction-limited particles under the assumption that particles diffuse with a single (unknown) diffusion coefficient. Specific Aim II – Repeat Specific Aim 1 for the case where dynamical models according to which particles evolve are unknown or even changing in time (that is, the restriction that dynamics be governed by simple diffusion is lifted). Within each Aim, we will: determine particle numbers in each frame by adapting (nonparametric) Bernoulli processes; adapt observation models to account for complex label photophysics and aliasing artifacts important for fast-moving particle; treat particle confinement for particle diffusion in small bacterial cells while learning dynamical models by adapting Dirichlet processes; incorporate detailed camera noise models.