The electron microscope is essential for resolving biological structures which are too complex for X-ray crystallographic methods and too small to be resolved with the light microscope. In contrast to the conventional instrument, the High-Voltage Electron Microscopes (HVEM) such as the instrument located at the National Center for Microscopy and Imaging Research (NCMIR) can image relatively thick specimens that contain substantial 3D structure. Tomographic methods can be applied to a set of images, acquired from different orientations by tilting the specimen, to derive a 3D representation of its biological structure. Tomography requires extensive computation and considerable processing time on conventional workstations in order to reconstruct typically large volumes from HVEM tilt series. Goals (1) Expedite tomographic processing by implementing electron microscope tomographic methods on parallel machines. (2) Provide transparent interactive access to these programs in a form easily used by structural biologists. 3) Investigate and implement alternative tomographic algorithms that may potentially provide improved reconstructions. Major achievements: We previously implemented the commonly used single axis tilt, R-weighted backprojection algorithm and two iterative reconstruction methods, algebraic reconstruction (ART) and simultaneous iterative reconstruction (SIRT) on the Intel Paragon. (1) During this last year we ported these programs to the Cray T3E and (2) completed parametric runs on the Paragon and the Cray T3E. With increasing numbers of nodes, Paragon performance is impaired by limited disk I/O bandwidth as more nodes attempt I/O at approximately at the same time. The I/O bottleneck is substantially less severe on the Cray. The Cray implementation is generally 2-3 times faster than the Paragon. Using only 16 nodes, the Cray is 10 times faster than a single processor SGI R10000 workstation. (3) We have examination of the parallel algorithms using specimens from NCMIR collaborative research projects involving changes in the structure of dendritic spines following loss of synaptic input, an analysis o f the complex three-dimensional structure of mitochondria, and changes in the 3D structure of cardiac muscle in a study of heart failure. These initial studies indicate that the iterative methods can produce superior reconstructions with fewer artifacts in comparison to the R-weighted algorithm. These evaluations are continuing. (4) To facilitate and encourage the use of these programs by biologists, we have constructed scripts to obviate the need to understand the NQS (Network Queuing System) facility, and to automatically handle data transfer to and from Cray local storage to the HPSS archival file storage at SDSC. The script and the high performance computing facilitate running multiple jobs on the same dataset to perform parameter manipulations for obtaining optimum reconstructions. We have also developed data format conversion programs to facilitate parallel processing of HDF file formatted data obtained using our telemicroscopy system. Collaboration and service: As described above, several biological projects associated with NCMIR are using the CrayT3E for tomographic reconstruction. We expect the number of users to increase as methods for access are further improved. The availability of these programs has been announced at an international meeting on electron microscope tomography and a paper based on this conference has been published (Perkins et al. 1997). Recently, we have provided our software and consultation to Drs. Carl Kessleman (Univ. of Southern California), Ian McNalty (Argonne National Laboratory), and Mark Rivers (U. Chicago) who plan to use our programs for tomographic reconstructions of data from advanced photon and other X-ray sources as part of their DOE Grand Challenge project """"""""Supercomputer Solution of Massive Crystallographic and Microtomographic Structural Problems."""""""" Future plans: 1) Increase the number of parallel platforms by porting the software to the IBM SP2. 2) Implement and evaluate new iterative algorithms for improved reconstructions, which will benefit from the performance gains achievable with parallel computing. We plan to parallelize an algorithm developed by Gabor Herman and Jose Carazo using """"""""blob"""""""" basis functions. In collaboration with these investigators, this algorithm has been implemented for single axis tilt tomography and is currently undergoing testing on conventional workstations. (3) Continue effort to interface these programs to a telemicroscopy system providing remote control of the HVEM for acquisition of tomographic data and transparent distribution of computationally intensive tasks to high performance computers on the network. (4) Develop web-based access.
Showing the most recent 10 out of 270 publications
