Biomedical Engineering Reference
In-Depth Information
replaced by NMR-invisible 2 H. 13-16 This drastically reduces the amount of
structural information that can be gained from NOESY experiments, however.
Solid-state NMR removes the dependency on molecular tumbling but suffers
from reduced sensitivity at the current state of the technology. 17
Extending the experimental frontiers to large systems or lowly populated
states usually limits the number, resolution and fidelity of structural restraints
that can be obtained. Therefore it seems both attractive and timely to combine
structure prediction efforts with experimental data. 18-21 In particular, the
ROSETTA structure-prediction software has successfully demonstrated its
power in conjunction with various experimental data including NMR chemical
shifts, 22,23
data, 24,25
un-phased
crystallographic
and
NMR
data
from
perdeuterated proteins, such as backbone NOE, RDC and PCS data. 26
This chapter aims to aid both, the NMR spectroscopist seeking to extend his
knowledge about computation and the computational structural biologist
eager to contribute to the field by developing novel structure calculation
algorithms. To meet the requirements of these diverse backgrounds the chapter
contains a thorough introduction to computational models, i.e., force-fields,
and to NMR experiments that yield structurally informative data.
Subsequently, we will review the current approaches to structure calculations.
As discussed below, structure calculations using structure-prediction force-
fields and sparse data are currently strongly limited by their ability to sample
the conformational space sufficiently thoroughly to detect the lowest-energy
structures. Thus, there is a strong need for on-going development of
optimisation methods. Accordingly, we inform the reader about novel
computational approaches, and hope that this chapter provides a good entry
point
for
readers
interested
in
developing
novel
structure
determination
protocols.
It is helpful to discuss NMR structure determination in the 'language' of
Bayesian inference. 27 A structural model or 'belief' is equated with a posterior
probability density which is inferred from a prior (general) knowledge about
protein structures and experimental data. Inference is straightforward if the
data speaks for itself, i.e., the prior information hardly influences the posterior
distribution. High-quality data merits a non-informative prior that is diffuse on
the scale of the posterior distribution. 28 In X-ray or NMR structure
calculations it became customary to employ a minimal protein model with
potential terms for bond lengths, angles, chirality, planarity and non-bonded
repulsion. 29 This model is non-informative with regard to torsion angles, and
overall shape of the protein and does not differentiate in any way between
structures that are fully extended, bury charges, or don't have a hydrophobic
core on one side and well-folded structures with tightly packed hydrophobic
cores on the other side. Clearly, it is known that protein structures are
generally of the latter kind, and hence these general physical and chemical
properties of natively folded proteins can also be used in the model to enhance
the resolution of the prior distribution when experimental data is sparse.
Indeed, it was shown that short MD simulations in explicit solvent 30,31
or
Search WWH ::




Custom Search