Chemistry Reference
In-Depth Information
special atom label and bond type for atoms and bonds in aliphatic five- and six-membered
rings or aromatic rings, including atom type specifiers. Planar atoms and bonds are shown
in chemical representation
II
in Figure 8.9 All other atoms were labeled with the atom
type. An additional atom specifier for the atom type was connected to heteroatoms and
halogens and a specifier for implicit hydrogens was connected to heteroatom. Standard and
elaborate chemical representations were used to extract substructures from mutagenicity
data, both with and without considering nonlinear fragments. The dataset consisted of 4069
compounds from the Chemical Carcinogenesis Research Information System database.
[
41
]
Compounds were categorized as nonmutagens if all mutagenicity tests had a negative
outcome. This resulted in 2294 compounds classified as mutagens.
Fragments from all methods were used together to find nonredundant substructures that
are discriminative for mutagenicity. Only those substructures that occurred in more than 70
mutagens were considered. Adecision list was constructed (Figure 8.10) by using the frag-
ment with lowest
p
-value to split the set into two subsets (one that contained the fragment
and one that did not). The
p
-value of a fragment was defined as the probability of finding
a statistical association with mutagenicity based on chance alone. It was calculated from
the amount of mutagens versus nonmutagens that are detected using that fragment. For the
subset that did not contain the fragment,
p
-values were recomputed and the next most muta-
genic fragment was used to split this set. In the case of multiple fragments with the lowest
p
-value, the largest fragment was used. The process was repeated as long as the new set
had more than 60% mutagenic compounds. If the best-selected fragment had a
p
-value of
more than 10
−
20
, no further splits were made. From all methods, the use of elaborate chem-
ical representation combined with detection of nonlinear fragments proved best: mutagens
were detected with a sensitivity of 84%. The resulting decision list (Figure 8.10) consisted
of six nonredundant discriminating substructures, starting with a polycyclic planar system
that described at least three rings and consisted of 11 planar atoms connected by planar
bonds. The next most discriminating fragment was a nitrogen atom double-bonded to a
nitrogen or oxygen, followed by a three-membered heterocycle (aliphatic epoxides and
aziridines) and then an aliphatic halogen (chlorine, bromine and iodine). The second-last
fragment was an aromatic primary amine and the list ended with a heteroatom bonded
to a heteroatom fragment. Some of these substructures proved to be very similar to the
general toxicophores derived previously by the authors in a laborious approach.
[
42
]
These
results emphasize the benefit of elaborate chemical representation. For instance, the most
discriminative fragment for mutagenicity would not have been detected by other methods,
since the planar atom notation proved essential. Moreover, the importance of wildcards is
underlined by their presence in all six substructures. Since the list contained two branched
and one cyclic substructure, all possible graphs must be considered in substructure mining.
8.3.5 Ligand Design
The ring-linker frameworks approach described by Bemis and Murcko
[
28
]
was used to
design new scaffold classes based on experimental structural information and to guide the
optimization of modestly active ligands.
[
43
]
A set of 119 kinase inhibitors for at least 18
different targets was fragmented into ring systems and linkers and frequencies of occurrence
were analyzed. Since bi- and tricyclic ring systems were relatively rare in the fragmented
set, only monocyclic rings were considered. The authors found that the four rings benzene,
