Biomedical Engineering Reference
In-Depth Information
Currently, regardless of the specifics of docking algorithms and scoring functions,
posing is more reliable than scoring. This means that binding modes of active ligands
can often be modeled with reasonable to very good accuracy but that it is rather diffi-
cult to distinguish between active and inactive compounds on the basis of calculated
scores. The major reason for this complication is that it is not possible to compute free
energies of binding accurately for ligand-target complexes in a consistent manner
(although successful case studies are occasionally reported in the literature). In this
context, it must also be taken into account that the vast majority of database com-
pounds are inactive but that many of these inactives might yield a “science fiction”
pose that displays reasonable chemical and shape complementarity to the active site.
In fact, even based on visual inspection, taking chemical knowledge and intuition into
account, it is usually difficult to distinguish between “true” active compounds and
false positives. Hence, compound selection from database rankings is far from being
trivial. Given the well-known limitations of current scoring schemes, essentially any
successful docking application reported in the literature involves visual inspection
and subjective analysis of docked complexes in order to select candidate compounds
for experimental testing, a situation that has not changed over the past decade [9,45].
Because the computational requirements of flexible ligand docking are generally
high, it is often attempted to reduce the size of the screening database prior to
docking. For this purpose, LBVS and SBVS are often carried out in a sequential
manner (although truly integrated LBVS/SBVS approaches are yet to be introduced).
Typically, if at least a few active compounds are known for a docking target,
similarity search calculations are carried out to preselect database compounds for
docking that display at least some remote chemical similarity to the known actives.
This preselection scheme might work against the identification of structurally
completely novel hits but reduces the size of a source database by at least an order
of magnitude, or even more. Reducing 1 million database compounds to a subset
containing about 100,000 or fewer compounds enables fully flexible ligand docking
within a reasonable time frame.
15.6 SIMILARITY SEARCHING
Similarity searching is a traditional LBVS approach. Regardless of the specifics of
similarity search methods, they generally produce rankings of database compounds
relative to reference molecules in the order of decreasing similarity and thus decreas-
ing probability of activity. It is also common practice in LBVS to prefilter screening
databases and omit compounds with potential toxicity and other liabilities or non-
compliance with preformulated property range rules such as Lipinski's rule-of-5 [46].
However, such filtering techniques are not considered a part of the spectrum of simi-
larity search approaches. This also applies to substructure searching that is an “all or
nothing” matching technique.
Pharmacophore models [47,48] and fingerprints represents the major similarity
search approaches. Pharmacophore searching currently is the most popular LBVS
approach; the majority of prospective LBVS applications reported in the literature
