519
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
in efficiency by the reduction of time per docked compound makes clustering attractive compared to the periodic selection of snapshots. The major assumption with
this strategy is that a suitable selection of representatives should allow to cover the
important structural variability. By now, a solid statistical evaluation of a standard
recipe is out of reach considering the variety of clustering criteria and algorithms.
A systematic reduction of structures from the MD simulation can also be
achieved via a performance assessment of individual snapshots, e.g., selected by
periodic extraction. Subsequent applications such as virtual screening of large databases would only proceed with the snapshot(s) performing best. Docking performance can be evaluated in, e.g., three categories: reproduction of native binding
poses, ranking of active compounds, and discrimination of active and inactive compounds [88, 92]. Dependent on the purpose of docking one might choose between
different performance measures. Various binding site properties such as volume,
hydrophobicity and opening state were evaluated for ensembles of X-ray structures
in respect to their performance in discriminating active from inactive compounds
[93]. The authors could develop some rules depending on the inherent characteristics of the protein binding pocket which might be useful to select a specific Xray structure prone to outperform alternative ones prior of extensive evaluation.
By now, similar rules are not available for MD-derived structural ensembles [43].
Therefore, evaluation strategies are necessary and their dependence on experimental data might be a general limit of this structure selection strategy. When evaluating
the performance of periodically extracted snapshots from MD simulations, Nichols et al. could identify individual conformations that outperformed even the best
available X-ray structure in terms of discrimination between known ligands and
decoys for all investigated cases [43]. Xu et al. suggest a similar strategy to identify
suitable snapshots generated via implicit ligand simulations (LIMOC) according to
their ability to score ligands with more favorable scores than decoys [38]. Besides
lack of experimental data, the limited chemical diversity of experimentally verified
ligands hinders performance-based selection of structures. Performance can only
be evaluated within the chemical space covered by active and inactive compounds.
However, a conformation recovering known active ligands is not necessarily the
best conformation to identify novel ligand scaffolds. Searching for novel chemical
scaffolds by virtual screening one intentionally leaves the applicability area of the
underlying performance metric.
The restriction on specific conformations selected either by clustering or by performance reduces the complexity of ensemble docking by neglecting that individual
ligands might have preference for individual snapshots sampled during the MD. To
overcome this limitation a last strategy attempts to allow a selection of the conformation for each ligand individually. The challenge is the implementation of this
additional degree of freedom without exhaustively sampling every ligand versus
snapshot combination. Some ensemble docking strategies integrate the selection of
the conformation from the ensemble within the docking procedure. The selection
of the optimal conformation is also subject of the docking algorithm in analogy to
the translational and rotational degrees of freedom of the ligand optimized during
posing. In this context the ensemble is not be pre-generated. Zaccharias proposed
15 (How to) Profit from Molecular Dynamics-based Ensemble Docking
in efficiency by the reduction of time per docked compound makes clustering attractive compared to the periodic selection of snapshots. The major assumption with
this strategy is that a suitable selection of representatives should allow to cover the
important structural variability. By now, a solid statistical evaluation of a standard
recipe is out of reach considering the variety of clustering criteria and algorithms.
A systematic reduction of structures from the MD simulation can also be
achieved via a performance assessment of individual snapshots, e.g., selected by
periodic extraction. Subsequent applications such as virtual screening of large databases would only proceed with the snapshot(s) performing best. Docking performance can be evaluated in, e.g., three categories: reproduction of native binding
poses, ranking of active compounds, and discrimination of active and inactive compounds [88, 92]. Dependent on the purpose of docking one might choose between
different performance measures. Various binding site properties such as volume,
hydrophobicity and opening state were evaluated for ensembles of X-ray structures
in respect to their performance in discriminating active from inactive compounds
[93]. The authors could develop some rules depending on the inherent characteristics of the protein binding pocket which might be useful to select a specific Xray structure prone to outperform alternative ones prior of extensive evaluation.
By now, similar rules are not available for MD-derived structural ensembles [43].
Therefore, evaluation strategies are necessary and their dependence on experimental data might be a general limit of this structure selection strategy. When evaluating
the performance of periodically extracted snapshots from MD simulations, Nichols et al. could identify individual conformations that outperformed even the best
available X-ray structure in terms of discrimination between known ligands and
decoys for all investigated cases [43]. Xu et al. suggest a similar strategy to identify
suitable snapshots generated via implicit ligand simulations (LIMOC) according to
their ability to score ligands with more favorable scores than decoys [38]. Besides
lack of experimental data, the limited chemical diversity of experimentally verified
ligands hinders performance-based selection of structures. Performance can only
be evaluated within the chemical space covered by active and inactive compounds.
However, a conformation recovering known active ligands is not necessarily the
best conformation to identify novel ligand scaffolds. Searching for novel chemical
scaffolds by virtual screening one intentionally leaves the applicability area of the
underlying performance metric.
The restriction on specific conformations selected either by clustering or by performance reduces the complexity of ensemble docking by neglecting that individual
ligands might have preference for individual snapshots sampled during the MD. To
overcome this limitation a last strategy attempts to allow a selection of the conformation for each ligand individually. The challenge is the implementation of this
additional degree of freedom without exhaustively sampling every ligand versus
snapshot combination. Some ensemble docking strategies integrate the selection of
the conformation from the ensemble within the docking procedure. The selection
of the optimal conformation is also subject of the docking algorithm in analogy to
the translational and rotational degrees of freedom of the ligand optimized during
posing. In this context the ensemble is not be pre-generated. Zaccharias proposed