Fast 3D Reconstruction Algorithms for Cryo-EM

描述(申请人提供)：这项研究建议开发快速期望最大化(EM)算法，用于低温电子显微镜(Cryo-EM)中高分辨率小颗粒重建的3D重建。该算法结合了显微镜的对比度传递函数(CTF)，可以与任何基函数一起使用。此外，还提出了两种策略，将算法的速度提高了1000倍。这两种策略是(1)域缩减，加速算法的E步；(2)在商用显卡上实现，其中图形处理器的巨大并行性加速了整个算法。基于核糖体和IP3R钙通道蛋白结构的模拟的初步结果表明，所提出的算法的E和M步骤是快速和准确的。 该研究通过(1)将区域归约策略进一步发展为多分辨率策略，(2)系统地并行化M步，以及(3)实现和验证完整的算法，扩展了前人的工作。此外，为在图形处理器单元上实现该算法，计划进行系统的软件设计。该软件将在万维网上供其他研究人员使用。该算法将使用模拟和真实世界的Cryo-EM图像进行验证。重建精度用重建时的平均均方误差来衡量，分辨率用傅里叶壳相关来衡量。并将其精度与传统算法的重建精度进行比较。算法的速度将作为图像数量和信噪比的函数进行测量。

DESCRIPTION (provided by applicant): This research proposes to develop fast expectation-maximization (EM) algorithms for 3D reconstruction of high resolution reconstruction of small particles in Cryogenic Electron Microscopy (Cryo-EM). The proposed algorithm incorporates the contrast transfer function (CTF) of the microscope and can be used with any basis functions. Further, two strategies are proposed to speed up the algorithm by a factor of 1000. The two strategies are (1) domain reduction where the E step of the algorithm is sped up, and (2) implementation on commodity graphics cards, where the massive parallelism of the graphics processors speeds up the entire algorithm. Preliminary results using simulations based on ribosome and the IP3R Ca+ channel protein structure show that the E and M steps of the proposed algorithm are fast and accurate. The proposed research extends previous work by (1) further developing the domain reduction strategy into a multi-resolution strategy, (2) systematically parallelizing the M step, and (3) implementing and validating the complete algorithm. Further, systematic software design is planned for implementing the algorithm on graphics processor units. This software will be made available on the world-wide web for other researchers. The algorithm will be validated using simulations and real-world Cryo-EM images. The reconstruction accuracy will be measured by the average mean squared error in reconstruction, and the resolution will be measured by Fourier shell correlation. The accuracy will also be compared to the reconstruction accuracy of conventional algorithms. The speed of the algorithm will be measured as a function of number of images and signal to noise ratio.