4.4.2 MD Can Offer
Detailed Models at Shorter
Timescales
Despite the difficulty of accessing long timescales, analysis of rela-
tively short unbiased MD simulations is a very popular approach for
understanding cryptic allostery. The diversity of analysis approaches
in the literature is very broad, but all attempt to statistically corre-
late the dynamics of two remote regions of the protein target.
Mutual information has emerged as a powerful way to look at the
potentially subtle correlations in protein dynamics, as it makes
fewer assumptions than other common approaches like covariance
analysis. Mutinf applies mutual information to analyze sets of
repeated MD trajectories for correlations in dihedral motions
[155]. Mutinf compares trajectory replicas and applies bootstrap-
ping to remove predicted correlations which are considered noise.
This analysis was qualitatively compared to NMR chemical shifts
but has subsequently been used to understand allosteric pockets in
kinase systems [156, 157]. In one recent example a previously
unknown allosteric site was predicted, and subsequently active
compounds were discovered with virtual screening and experimen-
tally validated [158].
Researchers have now moved on from presenting pairwise
correlation matrices to considering network community analysis
of correlated motions [159]. Community analysis groups highly
correlated residues together and scores the size of correlations
between blocks. This is an important step, as it allows for simpler
understanding and comparison of these highly degenerate net-
works. A series of studies have applied a mutual information com-
munity approach to understand the dynamical correlation network
of protein kinase A (PKA) [160]. Calculated communities could be
functionally annotated but did not fit with motifs based on
sequence or secondary structure and were modified by conforma-
tional changes or ligand binding. This analysis of PKA also gave
new understanding to known functional mutations which was then
investigated with detailed biochemical experiments [161]. Also
some of the dynamic correlations of the important hydrophobic
spine regions have been validated with methyl-TROSY
NMR [162].
A similar approach applied to thrombin has been able to explain
how changes in the dynamics of the proteases surface loops on
active site occupation lower the entropic penalty for binding an
allosteric modulator [136, 137]. Interestingly, the slower dynamics
of thrombins surface loops required the application of aMD and the
frustratometer for the dynamic models to closely match the
NMR data.4.4.3 Combined
Platforms Maybe
an Optimal Way Forward
Analysis of atomistic MD clearly gives the quality models and detail
required to understand how ligand binding and mutations affect
the molecular switches which are often drug targets. However, as
exemplified by this last set of studies, combining approximate
methods like the frustratometer with detailed enhanced MD may354 Benjamin P. Cossins et al.