The generation of a conformational ensemble involves two
steps. First is to run an MD simulation on the target structure to
generate a trajectory. This atomic trajectory can be either a single
extensive trajectory, coming from a single simulation, or multiple
trajectories combined together to form a single long trajectory
[23]. The two approaches can lead to different outcomes, and it
seems that using multiple simulations can provide more improved
conformational sampling over the use of a single MD
trajectory [31].
The next step, involves the clustering of the long trajectory
obtained from the simulation. This clustering step not only identi-
fies representative conformations for the target protein, but is also
used to reduce the computational time in subsequent binding site
evaluation steps [38]. The most common way for clustering protein
conformations is using RMSD-based methods. Other methods
may also involve principal component analysis (PCA) [46], non-
negative matrix factorization (NMF) [47], and independent com-
ponent analysis (ICA) [46, 48].2.5 The Sampling
Problem
In many cases and depending on the plasticity of the target, a
binding site that is not obviously present in a crystal structure
may appear after a short simulation time. For example, Eyrisch
and Helms identified transient pockets on the protein surface
after 10 ns of MD simulation [20]. However, in most cases, con-
ventional MD simulations cannot access these sites, and a signifi-
cantly long MD simulation may be required to sample the
conformational space of the target (seeNotes 3and 4 ). This is
mainly due to the entrapment of the protein structure in a local
minimum within the energy surface and not being able to cross the
high-energy barriers separating these minima (Fig.5)[49].
In this case, one has two options: running a very long MD
simulation (in the 100 s of ns scale) or using other MD simulations
methods that are designed to solve this sampling problem such as
Replica Exchange MD (REMD) [50], accelerated Molecular
Dynamic (aMD) [49, 51], Free Energy Perturbation (FEP),
Metadynamics-Based Methods, and Steered Molecular Dynamic
(SMD) [21]. These methods can provide significant improvement
over conventional MD methods, although they require huge
computational resources. A good sampling of the protein structure
is important to identify binding sites. If the sampling is not suffi-
ciently done to fully explore all possible target conformations, one
may miss a very valuable binding site on the target. Many reviews
have focused on these different MD methods, and interested read-
ers are directed to these references [21, 22, 52–54].2.6 Cosolvent
Molecular Dynamics
Simulations
Flooding a protein structure with different probes during MD
simulations emerged as a new way to detect binding sites while
taking into account their flexibility. A promising example of such
methods is the cosolvent MD simulation approach. In this method,94 Tianhua Feng and Khaled Barakat