Esmond G Ng | PubMed Health

PubMed2006

The parallelization of SPIDER on distributed-memory computers using MPI.

We describe the strategies and implementation details we employed to parallelize the SPIDER software package on distributed-memory parallel computers using the message passing interface (MPI). The MPI-enabled SPIDER preserves the interactive command line and batch interface used in the sequential version of SPIDER, thus does not require users to modify their existing batch programs. We show the excellent performance of the MPI-enabled SPIDER when it is used to perform multi-reference alignment and 3-D reconstruction operations on a number of different computing platforms. We point out some performance issues when the MPI-enabled SPIDER is used for a complete 3-D projection matching refinement run, and propose several ways to further improve the parallel performance of SPIDER on distributed-memory machines.

Algorithms↗

PubMed2005

Unified 3-D structure and projection orientation refinement using quasi-Newton algorithm.

We describe an algorithm for simultaneous refinement of a three-dimensional (3-D) density map and of the orientation parameters of two-dimensional (2-D) projections that are used to reconstruct this map. The application is in electron microscopy, where the 3-D structure of a protein has to be determined from a set of 2-D projections collected at random but initially unknown angles. The design of the algorithm is based on the assumption that initial low resolution approximation of the density map and reasonable guesses for orientation parameters are available. Thus, the algorithm is applicable in final stages of the structure refinement, when the quality of the results is of main concern. We define the objective function to be minimized in real space and solve the resulting nonlinear optimization problem using a Quasi-Newton algorithm. We calculate analytical derivatives with respect to density distribution and the finite difference approximations of derivatives with respect to orientation parameters. We demonstrate that calculation of derivatives is robust with respect to noise in the data. This is due to the fact that noise is annihilated by the back-projection operations. Our algorithm is distinguished from other orientation refinement methods (i) by the simultaneous update of the density map and orientation parameters resulting in a highly efficient computational scheme and (ii) by the high quality of the results produced by a direct minimization of the discrepancy between the 2-D data and the projected views of the reconstructed 3-D structure. We demonstrate the speed and accuracy of our method by using simulated data.

Algorithms↗