517
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
Reducing the number of snapshots resembles the stochastic experiment of picking individuals out of a larger collection: The possible combinations to select r
different snapshots out of N different snapshots registered in a trajectory scales
with an N over r relation leading to an explosion of possibilities with N [87, 88].
For ensembles of X-ray structures an extensive analysis over potential combinations was performed. Indeed, the performance of the individual combinations is not
equal [87]. In our test case we observed that reducing 300 snapshots to 100 results
in a similar score profile if the snapshots are selected continuously. The further
reduction by increasing the incremental interval of extraction to 3 ns yielding 10
snapshots, which would be a rational number of structures for virtual screening, is
not sufficient to reproduce the original scoring profile (data not shown). This shows
that this large extraction interval is not sufficient as one might lose crucial structural
diversity.
More appropriate is the selection of structures representing specific regions of
conformational space to ensure coverage of the structural variation in the trajectory.
Such a clustering approach needs a metric to compare the binding site characteristics of the snapshots and groups them by similarity or distance. Subsequently,
single structures are selected to represent the individual cluster groups representing
a specific fraction of a trajectory (population). A plethora of clustering algorithms
[89] is available for conformations from MD simulations, e.g., as implemented in
the analysis tools of widely used simulation packages such as AmberTools PTRAJ
[34]. For the case of the mutant W191G of cytochrome c peroxidase [84], Amaro
et al. discuss the application of root mean square deviation (RMSD) of atom positions between pairs of conformations as distance criterion and subsequent clustering. The applied exclusion sphere clustering based on a similarity criterion of 1 Å
RMSD between backbone atoms was especially developed to analyze the conformational space sampled during an MD simulation [90]. As most clustering algorithms are based on atom positions, a proper alignment of the overall structures is
needed. Similarly, the tool “QR factorization” uses an RMSD criterion between the
snapshots to perform a hierarchical clustering. This procedure allows to select a distance threshold after clustering, and hence, selection of snapshots by cluster level.
It was shown that similar results regarding the scoring profile could be produced
for the RNA editing ligase when docking to an ensemble of 400 snapshots extracted
periodically or to reduced set of 33 representative structures selected by a certain
RMSD threshold [84, 91].
Clustering based on active site atoms RMSD was selected as method of ensemble representation in a study to rationalize the binding mode of the natural product
katsumadain A to influenza neuraminidase [27]. A similar strategy was followed
for the same target in a lead discovery study [26]. For virtual screening of the database the authors first used representative structures of the three most populated
clusters for two simulations with and without ligand and the two crystal structures
resulting in 8 parallel docking runs. Subsequently, top ranked compounds were rescored using the mean of scores from docking to all 27 cluster representatives of
the holo simulation following the RCS protocol [84]. Discrimination of active and
inactive compounds by docking could be improved for serine racemase with cluster
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
Reducing the number of snapshots resembles the stochastic experiment of picking individuals out of a larger collection: The possible combinations to select r
different snapshots out of N different snapshots registered in a trajectory scales
with an N over r relation leading to an explosion of possibilities with N [87, 88].
For ensembles of X-ray structures an extensive analysis over potential combinations was performed. Indeed, the performance of the individual combinations is not
equal [87]. In our test case we observed that reducing 300 snapshots to 100 results
in a similar score profile if the snapshots are selected continuously. The further
reduction by increasing the incremental interval of extraction to 3 ns yielding 10
snapshots, which would be a rational number of structures for virtual screening, is
not sufficient to reproduce the original scoring profile (data not shown). This shows
that this large extraction interval is not sufficient as one might lose crucial structural
diversity.
More appropriate is the selection of structures representing specific regions of
conformational space to ensure coverage of the structural variation in the trajectory.
Such a clustering approach needs a metric to compare the binding site characteristics of the snapshots and groups them by similarity or distance. Subsequently,
single structures are selected to represent the individual cluster groups representing
a specific fraction of a trajectory (population). A plethora of clustering algorithms
[89] is available for conformations from MD simulations, e.g., as implemented in
the analysis tools of widely used simulation packages such as AmberTools PTRAJ
[34]. For the case of the mutant W191G of cytochrome c peroxidase [84], Amaro
et al. discuss the application of root mean square deviation (RMSD) of atom positions between pairs of conformations as distance criterion and subsequent clustering. The applied exclusion sphere clustering based on a similarity criterion of 1 Å
RMSD between backbone atoms was especially developed to analyze the conformational space sampled during an MD simulation [90]. As most clustering algorithms are based on atom positions, a proper alignment of the overall structures is
needed. Similarly, the tool “QR factorization” uses an RMSD criterion between the
snapshots to perform a hierarchical clustering. This procedure allows to select a distance threshold after clustering, and hence, selection of snapshots by cluster level.
It was shown that similar results regarding the scoring profile could be produced
for the RNA editing ligase when docking to an ensemble of 400 snapshots extracted
periodically or to reduced set of 33 representative structures selected by a certain
RMSD threshold [84, 91].
Clustering based on active site atoms RMSD was selected as method of ensemble representation in a study to rationalize the binding mode of the natural product
katsumadain A to influenza neuraminidase [27]. A similar strategy was followed
for the same target in a lead discovery study [26]. For virtual screening of the database the authors first used representative structures of the three most populated
clusters for two simulations with and without ligand and the two crystal structures
resulting in 8 parallel docking runs. Subsequently, top ranked compounds were rescored using the mean of scores from docking to all 27 cluster representatives of
the holo simulation following the RCS protocol [84]. Discrimination of active and
inactive compounds by docking could be improved for serine racemase with cluster