Project E aimes to develop computational methods for finding
an optimun three-dimensional structure give a finite set of noisy
two-dimentional projections with unknown orientation parameters. It is
proposed to determine limits of currently used computational methods,
such as orientation refinement procedures, and to develop methods that
are capable of systematic exploration of the space of possible
solutions. To make comparisions among different viable structures
possible, we will develop a self-consistency measure and we will relate
the quality of the reconstruction to limits imposed by the quality of
the data through dedicated statistical tests. The software
developed will be ported within the framework of SPARX and SPIDER
systems. We will use a vast collection of amassed experimental
data and previously solved structures to verify the software. The
methods developed will be immediately tested on cryo-electron
microscopy data collected within the framework of this proposal.
We aim to develop generalized measures of the quality of 3-D
reconstructuions that employ amplitude as well as phase information,
which can be used to compare candidate solution to the
particle-alignment problem. Also, we want to design a method for
determining optimal orientations for particle projections, recognizing
the fact that the task will be carried out with a highly parallel
rather than serial machine. In one approach, the currently used
bootstrap method will be employed independently, and alternative
solutions will be found by using sets of orientation parameters taht
have been modified by random perturbationa. In a second approach,
a generic algorithm will be used to seek a global minimum by processing
a large set of possible solutions sumultaneously.
Besides those goals mentioned above, we also aim to use multivariate
statistical tests and signal/noise characteristics to assess the
quality of 3-D reconstructions relative to the quality of the
images. These tests will measure whether results from different
algorithms/solutions differ significantly in the statistical
sense. Furthermore, we want to implement all algorithms in a way
that distributes the CPU-intensive part of the work among the nodes of
a PC-based cluster or among the nodes of massively parallel computer
such as IBM SP. This will be achieved using standard programming
tools in order to assure full protability across platforms.
.
