Biomedical Engineering Reference
In-Depth Information
largely incorrect assumption [17]. Rather, LBVS methods of very different sophis-
tication and computational complexity have been shown to be capable of identify-
ing structurally diverse active compounds, depending on the targets and compound
classes under investigation. In some instances, relatively simple molecular repre-
sentations and methods might be successful, whereas complex approaches fail; in
others, methods of varying complexity consistently identify hits or, alternatively,
no method succeeds [7,17]. The general compound class and target dependence of
VS approaches is currently not well understood. However, in LBVS, it is evident
that scaffold hopping ability does not correlate with methodological complexity. For
example, relatively simple similarity search tools, such as molecular fingerprints (i.e.,
bit-string representations of molecular structure and properties, as discussed further
below), have demonstrated scaffold hopping potential [22,23]. A recent investigation
has revealed mechanisms by which fingerprints recognize structurally diverse active
compounds [24]. In general, care must be taken to judge method performance. This
applies particularly to VS method evaluation in benchmark calculations. In this case,
known active compounds are added to screening databases as potential hits and their
retrieval (or recall) rates are calculated as a measure of VS performance [17]. In such
theoretical investigations, VS performance is typically overestimated and does not
scale with success rates in practical applications [1,8,9]. This is a direct consequence
of artificial system setups, where optimized active compounds are taken from the
literature as hits and added to background databases. Optimized active compounds
are generally larger and chemically or topologically more complex than screening
database compounds and thus easier to distinguish from them than are typical screen-
ing hits [25,26]. Thus, artificially high recall rates are often observed in benchmark
calculations in both LBVS and SBVS.
15.3 MOLECULAR SIMILARITY IN VIRTUAL SCREENING
All LBVS methods explore the similarity between active reference molecules and
candidate compounds. This is a fundamental aspect of the LBVS concept [27].
However, molecular similarity can be evaluated from different points of view.
15.3.1 Local vs. Global Similarity
Any similarity assessment is dependent on the molecular representations chosen,
such as molecular fingerprints or numerical property descriptors [28]. Importantly,
in LBVS, one needs to distinguish between local and global molecular similarity
comparisons.
Local similarity
refers to the presence of specific functional groups
or molecular features in a given topological environment or a specific geometric
arrangement. Here, only active parts (known or putative) of molecules are consid-
ered, and this local view of similarity is central to the pharmacophore concept [29],
as discussed below, and also to quantitative structure-activity relationship (QSAR)
modeling [30]. QSAR analysis is applied primarily to congeneric series of compounds
to predict chemical modifications that lead to improvements in potency. Accordingly,
