Biomedical Engineering Reference
In-Depth Information
cost of computing and thus 100 000 CPU hours are likely to become viable for
routine calculations within the next decade.
4.4.2 Top-Down vs. Bottom-Up
The most popular optimisation method in structure calculation is a top-down
method called simulated annealing which is usually applied in combination
with Metropolis Monte Carlo (MC) or Molecular Dynamics (MD) as a
sampling method. This method works very reliably and efficiently if the
structural restraints yield sufficient guidance.
100
Long-range restraints
modulate the energy landscape such that a deep and wide depression is
centered at the native conformation. Short-range restraints, in contrast, yield a
more golf-course like landscape, i.e., a flat hypersurface without guidance
towards the native structure. Such a golf-course energy landscape leads toa
large entropic barrier for the optimisation process, which cannot be overcome
by merely heating the system as in simulated annealing. Similarly grand-
canonical
methods,
like
replica
exchange,
do
not
help
with
such
search
problems.
101-103
The search problem in flat energy landscapes is significantly reduced in
bottom-up approaches that partition the conformational space into over-
lapping subspaces within which exhaustive sampling is more readily
achieved.
104
Fragment assembly, a widely used method for structure
calculation from chemical shift and sparse RDC data, employs this bottom-
up strategy idea by partitioning the protein backbone into overlapping
stretches of 3-20 residues. A particular conformation defined by the backbone
torsions within such a window is called a fragment. Fragments can be
generated from scratch,
105
but more often they are drawn from a large
database of high-resolution protein structures using chemical shifts, residual
dipolar couplings and sequence homology. If this procedure yields precisely
and accurately defined backbone conformations for a sufficient number of
windows, the resulting reduction in accessible conformational space for the
full-length protein will suffice for a simulated annealing procedure to succeed
in sampling structures within 1-2
˚
from the native. These in turn can be
identified using high-resolution but short-range energy terms such as all-atom
force-fields or chemical shift based restraint energies.
Fragment assembly was first suggested with short-range NOE data,
106
and
was later applied to RDC data.
107,108
Simons et al. developed ROSETTA
which assembles fragments using a simulated annealing MC search within an
empirical low-resolution force-field.
109
Bowers et al. combined ROSETTA
with sparse NOE data
20
and Rohl et al. with RDC and CS data.
21
Combined
with full-atom refinement
110
and improved chemical shift scoring functions
83,87
fragment assembly has been shown to yield consistently accurate structures in
atomic resolution for small protein domains (,100 residues) using backbone
CS data alone
22,23
and supplemented by sparse RDC data for proteins up to