508
S. von Grafenstein et al.
trajectories for subsequent ensemble-based docking studies [41]. Systematic analysis indicate that conformations similar to the ligand-bound state can be sampled
in trajectories where no ligand was present during sampling. For our test case, the
active site C-α atoms of the ligand-free simulation remain close to the starting conformation of the ligand-bound simulations reflected in RMSD values below 1 Å in
Fig. 15.1d.
However, the conformational space of the protein–ligand complex is different
from the ligand-free state and the overlap is not necessarily large. Upon ligand
binding a shift in population and accessible conformations is expected [24]. Examples where the crystal structures of bound and ligand-free state differ in several
Å in RMSD are known, e.g., the backbone between the L-leucine binding protein
differs in 7.1 Å RMSD in free state versus phenylalanine-bound state [42]. In such
cases sophisticated protocols are necessary to generate a conformation resembling
a ligand-bound state [42]. However, there are cases where a ligand-bound receptor
state is never sampled in an unbiased ligand-free simulation.
Conformations sampled in the ligand-bound trajectory are adapted to the ligand.
This is reflected when docking performance is compared for ligand-bound simulation and ligand-free simulation. Enrichment of known ligands within sets of decoys
is higher for snapshots extracted from trajectories where the ligand is present [43].
Ligand-bound X-ray structures representing the holo state are prepared to accommodate a ligand. The presence of a ligand in the MD simulations also influences
conformational sampling. In our test case influenza neuraminidase, the side chain of
Arg371 stays in an stretched conformation forming a salt bridge with the carboxylate of zanamivir (see Fig. 15.1b), whereas it explores alternative states in ligand-free
simulations (see Fig. 15.1a, c). However, the presence of a specific ligand biases
the sampled conformations of the receptor towards a distinct chemotype. This leads
to restricted recovery of the correct binding mode of alternative scaffolds [44]. For
X-ray structures it was shown that ligand-bound structures might be considered as
overspecialized [44]. Especially, in virtual screening, when the search for novel inhibitor scaffolds motivates a docking project, this overspecialization needs to be
avoided to allow an enrichment of different chemotypes [45]. Therefore, a simulation starting from a ligand-free simulation might be more appropriate. In order to
allow the binding site to be adapted for ligand binding without bias toward a specific
ligand, Xu et al. developed an implicit ligand-model [38]. The ligand–model concept
(LIMOC) uses small molecular probes covering the accessible surface of the binding
site dependent on the local environment. Subsequent sampling follows an algorithm
which allows the probes to explore different sites as well as an adaption of the binding site to the fragment-like structures. Application of the concept on thrombin and
acetylcholinesterase shows that the recovery of the native binding pose is between
the one for the simulation with the native ligand and a ligand-free simulation [38].
With focus on subsequent investigation of the protein–ligand interactions in a
defined binding site within docking the flexibility of regions away from this binding site might be negligible. Therefore, flexibility is frequently limited on specific
regions during the MD simulations. The influence of restricted flexibility was analyzed by Bolstad et al. for dihydrofolate reductase of different organisms and by
Armen et al. for the case of p38α mitogen-activated protein kinase [46, 47]. Bolstad
S. von Grafenstein et al.
trajectories for subsequent ensemble-based docking studies [41]. Systematic analysis indicate that conformations similar to the ligand-bound state can be sampled
in trajectories where no ligand was present during sampling. For our test case, the
active site C-α atoms of the ligand-free simulation remain close to the starting conformation of the ligand-bound simulations reflected in RMSD values below 1 Å in
Fig. 15.1d.
However, the conformational space of the protein–ligand complex is different
from the ligand-free state and the overlap is not necessarily large. Upon ligand
binding a shift in population and accessible conformations is expected [24]. Examples where the crystal structures of bound and ligand-free state differ in several
Å in RMSD are known, e.g., the backbone between the L-leucine binding protein
differs in 7.1 Å RMSD in free state versus phenylalanine-bound state [42]. In such
cases sophisticated protocols are necessary to generate a conformation resembling
a ligand-bound state [42]. However, there are cases where a ligand-bound receptor
state is never sampled in an unbiased ligand-free simulation.
Conformations sampled in the ligand-bound trajectory are adapted to the ligand.
This is reflected when docking performance is compared for ligand-bound simulation and ligand-free simulation. Enrichment of known ligands within sets of decoys
is higher for snapshots extracted from trajectories where the ligand is present [43].
Ligand-bound X-ray structures representing the holo state are prepared to accommodate a ligand. The presence of a ligand in the MD simulations also influences
conformational sampling. In our test case influenza neuraminidase, the side chain of
Arg371 stays in an stretched conformation forming a salt bridge with the carboxylate of zanamivir (see Fig. 15.1b), whereas it explores alternative states in ligand-free
simulations (see Fig. 15.1a, c). However, the presence of a specific ligand biases
the sampled conformations of the receptor towards a distinct chemotype. This leads
to restricted recovery of the correct binding mode of alternative scaffolds [44]. For
X-ray structures it was shown that ligand-bound structures might be considered as
overspecialized [44]. Especially, in virtual screening, when the search for novel inhibitor scaffolds motivates a docking project, this overspecialization needs to be
avoided to allow an enrichment of different chemotypes [45]. Therefore, a simulation starting from a ligand-free simulation might be more appropriate. In order to
allow the binding site to be adapted for ligand binding without bias toward a specific
ligand, Xu et al. developed an implicit ligand-model [38]. The ligand–model concept
(LIMOC) uses small molecular probes covering the accessible surface of the binding
site dependent on the local environment. Subsequent sampling follows an algorithm
which allows the probes to explore different sites as well as an adaption of the binding site to the fragment-like structures. Application of the concept on thrombin and
acetylcholinesterase shows that the recovery of the native binding pose is between
the one for the simulation with the native ligand and a ligand-free simulation [38].
With focus on subsequent investigation of the protein–ligand interactions in a
defined binding site within docking the flexibility of regions away from this binding site might be negligible. Therefore, flexibility is frequently limited on specific
regions during the MD simulations. The influence of restricted flexibility was analyzed by Bolstad et al. for dihydrofolate reductase of different organisms and by
Armen et al. for the case of p38α mitogen-activated protein kinase [46, 47]. Bolstad