522
S. von Grafenstein et al.
been focused during the set-up of the MD simulation topology (see Sect. 15.2.1.2).
Nevertheless, structure preparation before docking could allow an adaptation of
the protonation state depending on the conformation; an aspect which was not yet
investigated to our knowledge. A greater issue might be the treatment of solvent
molecules [49]. Valuable information on the fluctuation, positional and rotational
degrees of freedom of water molecules in the binding site is generated when one
uses an explicit solvent model in the MD simulation. This information is available
from the trajectory stored for conformer generation. It might be worth to perform
an in-depth analysis to extract this information as it will guide the decision on how
to proceed with water molecules in receptor preparation. Fixed tightly bound water
molecules are unlikely to be displaced by ligand molecules and can be treated as
part of the receptor [49]. If a water position from the X-ray structure is not characterized by high density during the MD simulation, this would support to remove
the water molecule before docking. Displacement of “unhappy” water molecules is
expected to contribute favorable to the binding of a ligand [111]. Attempts to integrate MD-derived information on water behavior in scoring are made for example
in the scoring function Wscore. This scoring function is currently under development by the group of Richard Friesner using the WaterMap concept and is likely to
be implemented in the docking program Glide [108, 109].
Depending on the scoring function one might also consider minimization of
snapshots or representative structures [48, 86, 112]. For several targets Marelius
et al. applied an empirical scoring function to evaluate poses sampled during an
MD simulation. Paulsen et al. used minimized structures from an MD sampling
of several picoseconds of Candida albicans dihydrofolate reductase and influenza
neuraminidase to dock and rank active ligands. Ranking accuracy increased for the
empirical scoring function Surflex-Dock when minimizing the conformations sequentially taken from an MD simulation. Additionally, they discuss that an adaption
of the subsequent pooling strategy of snapshots should be considered using minimized structures [86]. Especially the hydrogen bond term showed high sensitivity
towards structural variations sampled during the trajectory. Energy minimization of
structures selected from the MD trajectory consistently led to binding free energy
predictions which were lower and closer to the experimental value [48]. For scoring
functions fitted on minimized protein structures this step might improve scoring,
as observed for an empirical function [113]. Fitting to crystal structures, which per
se should represent a local energy minimum conformation, might in general limit
the applicability of empirical scoring functions on MD snapshots. During an MD
simulation performed at a given temperature, e.g., at 300 K, the conformations are
explicitly allowed to sample non-minimum configurations. A non-optimal hydrogen position at an amino acid will result in an unfavorable scoring of a hydrogen
bond, when the scoring function judges hydrogen bonds via geometry. In contrast to
these observations in literature, we observe that minimization does not affect docking scores or they show even a decrease indicating less favorable poses in GOLD
(see Fig. 15.5). The minor effect of minimization of hydrogen positions is not unexpected as GOLD optimizes hydrogen orientations during docking (Fig. 15.5a). Minimization of side chain atoms with the Amber force field FF99SB-ILDN especially
S. von Grafenstein et al.
been focused during the set-up of the MD simulation topology (see Sect. 15.2.1.2).
Nevertheless, structure preparation before docking could allow an adaptation of
the protonation state depending on the conformation; an aspect which was not yet
investigated to our knowledge. A greater issue might be the treatment of solvent
molecules [49]. Valuable information on the fluctuation, positional and rotational
degrees of freedom of water molecules in the binding site is generated when one
uses an explicit solvent model in the MD simulation. This information is available
from the trajectory stored for conformer generation. It might be worth to perform
an in-depth analysis to extract this information as it will guide the decision on how
to proceed with water molecules in receptor preparation. Fixed tightly bound water
molecules are unlikely to be displaced by ligand molecules and can be treated as
part of the receptor [49]. If a water position from the X-ray structure is not characterized by high density during the MD simulation, this would support to remove
the water molecule before docking. Displacement of “unhappy” water molecules is
expected to contribute favorable to the binding of a ligand [111]. Attempts to integrate MD-derived information on water behavior in scoring are made for example
in the scoring function Wscore. This scoring function is currently under development by the group of Richard Friesner using the WaterMap concept and is likely to
be implemented in the docking program Glide [108, 109].
Depending on the scoring function one might also consider minimization of
snapshots or representative structures [48, 86, 112]. For several targets Marelius
et al. applied an empirical scoring function to evaluate poses sampled during an
MD simulation. Paulsen et al. used minimized structures from an MD sampling
of several picoseconds of Candida albicans dihydrofolate reductase and influenza
neuraminidase to dock and rank active ligands. Ranking accuracy increased for the
empirical scoring function Surflex-Dock when minimizing the conformations sequentially taken from an MD simulation. Additionally, they discuss that an adaption
of the subsequent pooling strategy of snapshots should be considered using minimized structures [86]. Especially the hydrogen bond term showed high sensitivity
towards structural variations sampled during the trajectory. Energy minimization of
structures selected from the MD trajectory consistently led to binding free energy
predictions which were lower and closer to the experimental value [48]. For scoring
functions fitted on minimized protein structures this step might improve scoring,
as observed for an empirical function [113]. Fitting to crystal structures, which per
se should represent a local energy minimum conformation, might in general limit
the applicability of empirical scoring functions on MD snapshots. During an MD
simulation performed at a given temperature, e.g., at 300 K, the conformations are
explicitly allowed to sample non-minimum configurations. A non-optimal hydrogen position at an amino acid will result in an unfavorable scoring of a hydrogen
bond, when the scoring function judges hydrogen bonds via geometry. In contrast to
these observations in literature, we observe that minimization does not affect docking scores or they show even a decrease indicating less favorable poses in GOLD
(see Fig. 15.5). The minor effect of minimization of hydrogen positions is not unexpected as GOLD optimizes hydrogen orientations during docking (Fig. 15.5a). Minimization of side chain atoms with the Amber force field FF99SB-ILDN especially