Parallelization and comparison of 3D iterative reconstruction algorithms

Abstract

High resolution structure determination of biological macromolecules by electron microscopy is central to understand their biological function. These structural analyses involve processing thousands projection images taken from the specimen at different orientations. Regularized iterative reconstruction methods are well suited to deal with the extremely noise conditions found in those studies, but they are computationally expensive. Parallel computing then emerges as a natural solution for those problems allowing huge jobs to be run in clusters of workstations. This work describes and analyzes the parallel implementations of five 3D iterative reconstruction algorithms, including simultaneous and block-iterative methods. The evaluation of the parallel approaches is carried put in terms of speedups and computation versus communication times. It is shown that there are specific iterative methods that are specially well suited for parallelization, with a great level of scalability and fast convergence rates. This work draws the conclusion that the use of those parallel reconstruction methods is going to be central to afford "grand challenge " problems currently unapproachable in structural biology, such as structure determination at close-to-atomic resolution by electron microscopy.