The existing receptor–ligand information is explored and exploited
to generate frequency database of each possible atom pairs and
models are trained using advanced statistical methods such as sup-
port vector machine and deep learning.
In the case of LBDND, the information is derived from the
structure–activity relationship data of known biologically active
molecules. The 3D-pharmacophore models are used to extract
the features of active molecules, responsible for the biological
activity. Using such pharmacophore, a pseudoreceptor model is
feasible to exploit to come up with a scoring function as an alterna-
tive to the SBDND [36, 37].2.4 Structural
Sampling
To start the process of de novo ligand design, either a single atom
or a fragment or the library of fragments can work as building
block. Atom-based approach generates more diverse chemical
space as compared to the fragment based approach but also
increases the possible solutions and makes harder to choose the
right compounds. Therefore, increasing the chemical space can be
advantageous in the absence of diverse fragment library, otherwise,
increasing the chemical search space may require higher computa-
tion power. Fragment based design strategies, on the other hand,
limit the chemical search space and provide easier opportunity to
select the potential candidates. Such a reduction is called “mean-
ingful reduction,” if fragment library is diverse enough to cover
most of the fragments that occur in drug molecules. However, an
atom is a subset of fragment and there is no limit defined for the
size of a fragment. The general relationship is that smaller the
fragment size, bigger the combinatorial space problem. The struc-
tural sampling deals with the expansion of building blocks. There
are several approaches for the growth and expansion of the initial
seed fragment, such as linking, growing, random structure muta-
tions, lattice-based sampling, and graph-based sampling
[38–40]. Linking, growing, and random structure mutations are
used for the SBDND while in the absence of interactions site for
LBDND, graph-based sampling particular are of high significance
[41]. The most of LBDND depends on topological molecular
graph and evolutionary algorithms [42].2.5 Combinatorial
Search Strategies
As discussed above, the smaller the fragment size, the bigger the
combinatorial explosion issue. Though larger fragment libraries
limit the search space, they also have their own limitation, as a
fragment library must be diverse enough. Under such contradic-
tory advantages, de novo design experiments have to tackle combi-
natorial explosion related problems. There is no optimal solution
available with right blend to fulfill both the demands. Combinato-
rial search algorithms are designed to offer practical solution by
giving up at least one of the advantages. There are three types
of algorithms used based on the input design: ‘Breath-first and128 Shashank P. Katiyar et al.