Molecular Basis of Hair Cell Stereocilia Bundle Morphology

项目简介：内耳的感觉细胞，毛细胞，携带着一束精确组织的 立体纤毛束在它们的表面。立体纤毛是探测亚纳米振动的微绒毛状突起 使我们能够感知声音。这种难以置信的敏感性需要一系列蛋白质，其中一些 在立体纤毛表面形成机械连接。长期以来，这些联系都是用电子可视化的。 然而，显微镜下一些细胞的具体功能仍然难以捉摸。这项建议侧重于一部具体的小说。 蛋白质、多囊肾和肝病-1样蛋白(PKHD1L1)，被认为是主要成分 体毛纤毛的表面，并假设促进体毛纤毛束的凝聚力。要了解 蛋白质在整个发育过程中的时空分布及其引起的机制 耳聋时的突变，多次使用不同的高分辨率光学和电子显微镜实验 正在进行成像，产生了包含大量数据的大量2D和3D图像 不幸的是，对它们进行整体分析的成本很高。通常为回答一个问题而生成，这样的数据 成套设备可以与科学界共享，供其他团体重复使用，以满足他们的研究需求。作为数据 收集变得越来越昂贵和耗时，越来越多的组加入到共享其原始数据的行列中 集与科学界，在各种格式往往不适合容易分析。机械方面的进展 计算机视觉的学习使自动化这些手动分析任务成为可能，然而 这种模型的开发严重受制于训练数据的可用性，这些数据已经被适当地 准备和注解。本补充提案旨在通过以下方式增加此类培训数据的可得性 以便于训练使用的方式对大量成像数据进行注释和存储 机器学习模型。我们将对我们从野生型和Ko小鼠收集的三种类型的成像数据进行注释 研究PKHD1L1在毛细胞中的作用：1)耳蜗毛细胞在3D共聚焦Z-STACKS上，2)毛细胞立体纤毛 和线粒体在我们的3D聚焦离子束扫描电子显微镜体积中，最后是立体纤毛 (2D)在我们的扫描电子显微镜上的耳蜗毛细胞束。将对这些注释进行评估 由我们专门从事机器学习的合作者进行的体积图像分割，帮助我们评估数据 偏向、评估我们的数据注释的通用性，并展示数据在AI/ML中的可用性 通过微型AI/ML应用程序作为概念证明。我们将特别注意 通过对开源机器学习算法进行再培训来记录我们的数据并展示其可用性 无论是在我们合作者的帮助下，还是在内部。数据、注释、文档和示例使用 案例将公开托管在开放存储库中。目前在机器学习方面的工作还很匮乏 听音场。通过公开发布支持机器学习的图像数据，我们的工作将鼓励发展 这些模型有可能显著提高未来图像分析的效率和吞吐量。

Project Summary: The sensory cells of the inner ear, the hair cells, carry a bundle of precisely organized stereocilia bundles on their surface. Stereocilia are microvilli-like protrusions detecting sub-nanometer vibrations enabling our perception of sound. This incredible sensitivity requires an ensemble of proteins, some of which form mechanical linkages at the surface of stereocilia. These linkages have long been visualized with electron microscopy, however, the specific function of some remain elusive. This proposal focuses on a specific novel protein, Polycystic Kidney and Hepatic Disease 1-Like 1 (PKHD1L1), which is thought to be a major component of the stereocilia surface coat and is hypothesized to facilitate stereocilia bundle cohesion. To understand the spatiotemporal distribution of the protein throughout development, as well as the mechanism by which it causes deafness when mutated, multiple high-resolution light and electron microscopy experiments using different imaging modalities are being performed, resulting in a plethora of 2D and 3D images containing a wealth of data which unfortunately are costly to analyze in their entirety. Often generated to answer one question, such data sets could be shared with the scientific community for other groups to reuse for their research needs. As data collection becomes increasingly more expensive and time consuming, more groups join in sharing their raw data sets with the scientific community, in various formats often unsuitable for easy analysis. Advances in machine learning for computer vision has made it feasible to automate these manual analysis tasks, however the development of such models is heavily bottlenecked by the availability of training data which has been properly prepared and annotated. This supplemental proposal aims to increase the availability of such training data by annotating and depositing a wealth of imaging data in such a way that they may be readily used for the training of machine learning models. We will annotate three types of imaging data we collect from wild-type and ko mice to study the role of PKHD1L1 in hair cells: 1) cochlear hair cells on 3D confocal Z-stacks, 2) hair cell stereocilia and mitochondria in our 3D focused ion beam scanning electron micrograph volumes, and finally 3) stereocilia (in 2D) on our scanning electron micrographs of cochlear hair cell bundles. These annotations will be evaluated by our collaborators specializing in machine learning volumetric image segmentation to help us assess data biases, assess generalizability of our data annotation, and demonstrate the usability of the data in AI/ML applications through mini-AI/ML applications as a proof of concept. A special attention will be given to documenting our data and demonstrating their usability by retraining open-source machine learning algorithms both, with the help of our collaborators, and in-house. The data, annotations, documentation, and example use cases will be publicly hosted in open repositories. There is currently a dearth of machine learning work in the hearing field. By publicly releasing machine learning ready image data, our work will encourage the development of models that have the potential to dramatically increase the efficiency and throughput of future image analysis.