Biomedical Engineering Reference
In-Depth Information
utilized for compound database searching. The LBVS field is methodologically more
heterogeneous, and alternative classifications of LBVS methods are possible. Herein,
LBVS approaches are divided broadly into similarity search and compound classifica-
tion methods. As discussed below, both areas include many different methodological
concepts and algorithms. For example, the similarity search category covers two-
dimensional fingerprints, pharmacophore models, three-dimensional pharmacophore
fingerprints, and shape queries. Regardless of their algorithmic differences, most sim-
ilarity search methods produce rankings of database compounds with respect to refer-
ence molecules. Herein, compound classification approaches are divided further into
basic classification methods such as clustering and partitioning (for which many dif-
ferent algorithms exist) and machine learning approaches that are becoming increas-
ingly popular in LBVS. Compound classification methods typically produce a binary
readout by labeling test compounds as “active” or “inactive.” However, database
rankings can also be produced by some machine learning approaches. Essentially all
SBVS and LBVS approaches assign the likelihoods of activity to test compounds (on
the basis of rankings) or yield binary active/inactive class label predictions. How-
ever, none of these approaches predicts actual potency values. Exceptions to this rule
are QSAR models adapted for VS. Apart from QSAR, very few approaches have
considered compound potency as a VS search parameter.
15.5 DOCKING
The discussion of specific VS approaches begins with a brief account of molecular
docking. For further details concerning docking algorithms, the interested reader is
referred to a number of very informative reviews that are cited in this section. Ligand
docking is the major SBVS approach. In fact, if we view the VS field as a whole,
docking is the single most popular VS technique on the basis of published applications
[1]. In general, ligand docking aims at predicting the binding conformation and
orientation, i.e., the so-called pose, of a test compound within a predefined active-site
region of the target structure. Scoring functions are applied to guide the posing process
and, ultimately, to rank database compounds according to their likelihood of activity.
Currently, the state of the art in this field is fully flexible ligand docking into rigid
template structures (or template structures with limited side-chain flexibility in the
active-site region) [9,37]. Avariety of different docking algorithms and programs have
been introduced [38-42]. Despite many algorithmic differences in conformational
search and pose finding, and also differences in force field-based or empirical scoring
functions, all methods have in common that the docking process is divided into energy
calculation-assisted posing and final scoring and ranking. Force field-based scoring
functions treat intra- and intermolecular interactions using molecular mechanics-
type potentials, whereas empirical scoring functions contain energy terms that are
specifically adjusted to reproduce experimental structures of ligand-target complexes
or fit binding data [43]. In addition, potentials of mean force are derived via statistical
analysis of distributions of atomic ligand-protein interactions and favor types of
interactions that are frequently observed in complex x-ray structures [44].
