IIBR Informatics: Reducing the training data annotation cost for learning-based macromolecule identification in cellular electron cryo-tomography

细胞是所有已知生物体的基本结构和功能单位。了解单个大分子及其在细胞内的复合体的结构是生物学研究领域的基础。然而，由于缺乏数据采集技术，这样的结构信息一直非常难以获得。细胞电子冷冻断层成像(CECT)的最新进展使亚分子分辨率和近自然状态下的细胞结构和组织的三维成像成为可能。CECT图像包含数以千万计的结构高度多样化的大分子，这给后续计算分析的吞吐量带来了重大挑战。如何有效和准确地处理图像以识别每种不同类型的大分子复合体一直是一个具有挑战性的问题。试图对每种类型的复合体进行比对和分类的传统方法速度太慢，无法处理正在获取的大量数据。该项目将专注于使用有监督的深度学习来降低数据标注成本，这是高分子结构识别的主要瓶颈。该项目将对研究生和本科生进行计算生物学、生物信息学和生物图像分析方面的培训，并将研究成果整合到大学课程中。该项目将开发新的方法，显著降低培训数据注释成本。这将通过(1)通过随机化模拟成像条件来提高训练的分类模型的泛化能力；(2)通过半监督自动编码器从未标记的子图中学习结构表示来增强分类模型的训练；以及(3)通过选择最重要的子图的最小集合来减少注释，以使用主动学习来改进分类。通过将这三个目标集成到一个统一的系统中，该项目将为用户提供基于学习的亚断层图像分类模型，提高精度，并将带注释的训练数据减少到最少。该项目的成功是在由不同种类的CECT数据集捕获的单细胞中实现系统的原位高分子结构分类完全自动化的关键一步。在这个项目中开发的算法和软件将直接造福于结构生物学社区。为了促进从该项目开发的方法的广泛使用，将开发一个开放源码的CECT分析软件AITOM(https://github.com/xulabs/aitom)，。这项研究的结果将在徐氏实验室网站(http://xulabs.org)和GitHub网站)上提供(https://github.com/xulabs)This奖项反映了美国国家科学基金会的法定使命，并已通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Cells are basic structural and functional units of all known living organisms. Understanding the structures of large individual macromolecules and their complexes inside cells is fundamental to biological research community. However, such structural information has been extremely difficult to obtain due to the lack of data acquisition techniques. Recent advances in cellular electron cryo-tomography (CECT) have enabled 3D imaging of cellular structure and organization at sub-molecular resolution and in near-native state. CECT images contain tens of millions of structurally highly diverse macromolecules, which introduces a major challenge in the throughput of subsequent computational analysis. How to efficiently and accurately process the images to identify each distinct type of macromolecular complex has been challenging. Conventional methods that try to align and classify each type of complex are too slow to process the large volumes of data being acquired. This project will focus on reducing data annotation costs using supervised deep learning, which is a major bottleneck for macromolecule structural recognition. This project will train graduate and undergraduate students in computational biology, bioinformatics, and bioimage analysis, as well integrate research results into university curricula.The project will develop novel approaches to significantly reduce the training data annotation cost. This will be achieved by (1) improving the generalization ability of classification model trained by randomizing the simulated imaging conditions; (2) enhancing training of classification model using structural representation learned from unlabeled subtomograms via a semi-supervised autoencoder; and (3) reducing annotation by selecting the minimal set of most important subtomograms for improving the classification using active learning. By integrating the three objectives into a unified system, this project will provide users learning-based subtomogram classification models with improved accuracy and a minimal amount of annotated training data. The success of the project is a critical step towards fully automating the systematic in situ macromolecule structural classification in single cells captured by heterogeneous CECT datasets. The algorithms and software developed in this project will have direct benefit to the structural biology community. To facilitate broad use of the methods developed from this project, an open-source CECT analysis software, AITom (https://github.com/xulabs/aitom), will be developed. Results of this research will be provided on the Xu Lab website (http://xulabs.org) and GitHub site (https://github.com/xulabs)This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.