Chemistry Reference
In-Depth Information
matrix approaches and an atomic model is subsequently fitted to this 'proton density'.
So far, this approach has only been demonstrated for very small proteins for which high-
resolution, high-sensitivity data can be obtained. Another example of a top-down approach
is the program AUREMOL,
[
53
]
wherein a trial structure is iteratively refined until a good
match to the experimental data is obtained. Both CLOUDS and AUREMOL are focused
on protein structure determination, not on ligand binding pose determinations.
As discussed elsewhere,
15
NOE matching is primarily a specialized top-down approach,
focused on ligand binding pose evaluation, that can also readily incorporate information
derived from bottom-up approaches. The results presented in this chapter demonstrate
that NOE matching is applicable to fragments and lead-like/drug-like compounds bound to
relatively small proteins. The main limitation to applying the method to larger proteins is the
increased difficulty of observing protein-ligand NOEs in such systems. Our initial forays
into approaches aimed at dealing with larger systems have been described in this chapter
and our initial results are very promising. In addition to the protein stability enhancement
and selective labeling strategies that we have utilized, other technologies that have yet to
be explored have the potential to be major, permitting breakthroughs with respect to the
application of NOE matching to large systems. These include the use of SAIL (stereo-
arrayed isotope labeling) amino acids
[
57, 58
]
and the use of reverse-micelle encapsulation
technologies.
[
49, 50
]
Two key points regarding NOEmatching are worth reiterating: (1) if the ensemble of trial
poses contains some that are very similar to the true pose, NOEmatchingwill generally score
these poses with a low COST relative to most of the decoy poses; and (2) to ensure that one
obtains a correct pose in the ensemble of trial poses, one needs to do extensive, systematic
sampling of 'pose space'. Even with extensive sampling, one still may detect gaps in the
RMSD space of the sampled poses with respect to the target pose (e.g. see Figures 5.4, 5.6
and 5.13); these gaps likely result from RMSD ranges for which no acceptable poses could
be found, presumably due to steric hindrance, etc. As we have discussed, many additional
improvements to NOE matching are possible. Some methods for evaluating the results of
NOEmatching in the absence of a known pose have been demonstrated in this chapter. Other
ways to evaluate the results are also being explored. For the correct pose, most experimental
NOE peaks should be assigned and the assignments should be plausible. As the bipartite
graph matching algorithm requires predefined edge costs that cannot be adjusted during the
search for an optimal match, it is difficult to incorporate explicitly connectivity information
into the matching procedure. However, one could check the resulting assignments from
NOEmatching for consistency with known connectivity information. For example, we may
know from TOCSY or COSY data that several experimental groups arise from the same
(unassigned) residue - the assignments produced by NOE matching could be checked
for consistency with this information. Another potential area of improvement involves
ranking poses with low NOE matching COSTs by molecular mechanics energies and/or
knowledge-based scoring potentials. Pose scoring based on observed and predicted ligand
1
H chemical shift changes
[
32
]
could also be used to rank a small subset of poses. More
generally, NOE matching could be readily combined with other pose ranking procedures
such as MM-GBSA
[
59
]
or MM-PBSA
[
60
]
as part of a consensus scoring approach (e.g. see
ref. 61). Finally, as mentioned previously,
[
15
]
much of the NOE matching procedure may be
recast in terms of Bayesian probabilities, e.g. a Bayesian analysis of chemical shifts can be
used to predict the probability of a spin system arising from a specific amino acid type.
[
62
]
