Biomedical Engineering Reference
In-Depth Information
compounds against arrays of targets. Extensions of Bayesian classifiers have been
introduced for LBVS such as Bayesian (inference) networks [97,98]. In these net-
works, nodes are random Bayesian variables that are connected by edges if they are
involved in conditional dependency relationships. Such networks represent a more
complex form of Bayesian classification than naive classifiers, but follow essentially
the same principles. However, networks might not always offer advantages compared
to simpler similarity methods in LBVS (as also observed for other VS approaches
of increasing complexity, as discussed above). Furthermore, Bayesian screening has
been combined with concepts from information theory to predict recall rates of LBVS
calculations for a given system setup, that is, a chosen molecular representation (such
as a fingerprint), a set of active reference compounds, and a particular screening
database [99,100]. This type of analysis can help to identify preferred descriptors
for practical LBVS applications targeting a specific biological activity, which are
most likely to distinguish active compounds from other database molecules. Finally,
attempts have recently also been made to combine outcomes of different machine
learning approaches, such as neural nets, decision forests, SVMs, and naive Bayesian
classifiers into meta-classifiers for LBVS [101,102].
15.9 CONCLUSIONS
In this chapter, an overview over contemporary VS methods has been provided.
Initially, VS has been positioned relative to HTS in the context of drug discovery
research. Then, fundamental VS concepts have been introduced that should be taken
into consideration, regardless of the methods that might be used. The methodological
spectrum of VS approaches has been described. Furthermore, different requirements
for SBVS and LBVS have been discussed, and a brief account of ligand docking
as the primary SBVS approach has been provided. Moreover, the critically impor-
tant role of chemical reference space design and molecular similarity assessment for
LBVS has been rationalized and intrinsic limitations have been discussed. Emphasis
has been put on highlighting the conceptual diversity of LBVS that is methodologi-
cally much more heterogeneous than SBVS and covers approaches of greatly varying
design and complexity. The LBVS field can essentially be divided into similarity
search and compound classification methods. It has been pointed out that machine
learning plays an increasingly important role for LBVS. Taken together, pharma-
cophore searching, SVMs, and Bayesian methods currently dominate the LBVS field.
However, there is no single approach that is superior to others, and LBVS/SBVS
success strongly depends on the compound classes, biological activities, and tar-
gets that are under investigation. LBVS and SBVS calculations are often carried
out in a sequential (and sometimes parallel) manner, especially to limit the number
of test compounds for computationally demanding docking calculations. However,
truly integrated LBVS and SBVS approaches that incorporate molecular similarity
assessment into the computational analysis of ligand-target interactions are yet to
be reported, which provides interesting opportunities for future research and method
development activities.
