Deep Learning for 3-D reconstruction of heterogeneous molecular structures from Cryo-EM data

确定生物分子的结构是结构生物学的目标，对于理解生命和药物发现的生物机制是必不可少的。生物大分子可以被认为是在活细胞中执行复杂操作的复杂机器。这些动态机器在它们的行动过程中经历了不同的构象，完全理解它们的工作需要确定多重构象。单粒子电子冷冻显微镜(Cryo-EM)已经成为一种在近原子分辨率下确定分子结构的独特方法。要实现对动态蛋白质复合体结构的高分辨率估计，需要大量的图像和计算密集型算法。这样的重建问题通常是通过设计使用关于成像过程和需要重建的对象的属性的信息的方法来解决的，并且多年来已经取得了许多显着的突破。这些计算方法被称为“基于模型的”方法，具有可预测性和稳定性的优点。然而，它们的计算代价可能很高，而且并不总是从复杂数据中获得最大价值。特别是，它们往往无法解析复杂的异质结构。相比之下，像深度神经网络这样的“数据驱动”方法在其他情况下证明了提高生物医学图像质量的非凡能力。许多深度学习方法的问题在于它们是不可预测的，因为即使是输入数据中的微小偏差也会导致输出的巨大偏差，这在生物成像应用中可能会产生毁灭性的影响。此外，通常很难解释深度网络机器真正优化的是什么。这个项目将提出一种新的深度神经网络家族，用于从低温电子数据中三维重建动态蛋白质复合体。通过与结构生物学家的密切合作，我们将提出系统地将关于信号和数据形成过程的物理的先验知识和约束嵌入到深层神经网络结构中的方法。我们还将与瑞士EPFL的M.Unser教授和他的团队合作。他们将提供深度神经网络稳定性分析领域的专业知识，并将分享他们在开发基于人工智能的方法以从Cryo-EM数据重建分子方面的经验。我们希望这种方法和这些合作将使我们能够引入稳定和可解释的神经网络，能够以当前方法无法解决的分辨率解析不同的生物结构。产生的方法将是开源格式，集成在现有的计算套件中，如CCP-EM，并向最有可能的社区提供。

Determining the structure of biomolecules is the goal of structuralbiology and is essential to understand biological mechanismsresponsible for life and in drug discovery. Biological macromoleculescan be thought of as complex machines that perform complexoperations in living cells. These dynamic machines pass throughvarious conformations in the course of their actions and a completeunderstanding of their working requires the determination of multipleconformations.Single particle Electron Cryo-Microscopy (Cryo-EM) has emerged as a unique method to determine molecular structures at near-atomic resolution. Achieving high-resolution estimation of structures of dynamic protein complexes requires large numbers of images and computationally intensive algorithms. Such reconstruction problems have been often approached by devising methods that use information about the imaging procedure and the properties of the object that needs to be reconstructed and many remarkable breakthroughs have been achieved over the years. These computational approaches are called "model-based" methods andhave the advantage to be predictable and stable. However, they can be computationally expensive and do not always derive maximum value from complex data. In particular, they are often unable to resolve complex heterogeneous structures. In contrast "data-driven" methods like deep neural networks have demonstrated, in other contexts, a remarkable ability to improve the quality of biomedical images. The problem with many deep learning approaches is that they are not predictable in the sense that often even small deviations in the inputdata can result in a huge deviation of the output, which can have devastating effects in bio-imaging applications. Moreover, it is often very difficult to interpret what a deep network machine is really optimizing.This project will advance a new family of deep neural networks for the3-D reconstruction of dynamic protein complexes from cryo-electrondata. By working closely with structural biologists, we will put forwardapproaches that systematically embed prior knowledge and constraintsabout the signal and the physics of the data formation process into thedeep neural network architectures. We will also collaborate with Prof.M. Unser and his team from EPFL Switzerland. They will provideexpertise in the area of analysis of stability of deep neural networksand will share their experience in developing AI-based methods formolecule reconstruction from Cryo-EM data. We expect that thisapproach and these collaborations will allow us to introduce stable andinterpretable neural networks able to resolve heterogeneous biologicalstructures at a resolution that current methods cannot. The methodsproduced will be in open-source format, integrated in existingcomputational suites like CCP-EM and made available to the broadestpossible community.