chapter 1 - Bentham Science

Multivariate Image Analysis Applied to QSAR Chemoinformatics: Directions Toward Combating Neglected Diseases 155
aegypti and An. stephensi mosquito larvae, using physicochemical descriptors, mainly hydrophobic () and
Hammett electronic ( + ) parameters of the substituents to correlate chemical structures and bioactivities.
In addition to classical physicochemical descriptors, such as those applied by Hansch and Verma [13],
multidimensional (nD) QSAR, in which descriptors are generated in a grid cell for a given molecule (3D)
and may account for ensemble averaging, receptor dependency and/or solvent effects (4D, 5D and 6D) [14-
18], has shown to be of widespread use. However, specific programs, exhaustive conformational screening
and/or difficult alignment rules are often required to obtain good models by means of these methodologies.
On the other hand, an image-based QSAR approach, in which images are two-dimensional chemical
structures of a pharmacophore, has given valuable and predictive QSAR models [19-22]. The so called
multivariate image analysis applied to QSAR (MIA-QSAR) is easily accessible and simple to manipulate;
its procedure is scrutinized here and applied to the series of organotin compounds evaluated by Hansch and
Verma [13]. The regression parameters of the MIA-QSAR model may then be used to estimate the activity
of novel, eventually more potent mosquito controllers.
2. MIA-QSAR PROCEDURE
In classical QSAR, physicochemical descriptors are usually calculated or measured for a set of compounds,
and then correlated with the respective bioactivity through a given statistical tool. In MIA-QSAR,
descriptors are images, i.e., the 2D chemical structures of the series of compounds; they should be
transformed in numerical values in order to allow correlation with the biological activities. Each pixel of an
image is a grey value; different 2D chemical structures are shapes with pixels distributed at specific
coordinates, and this variance along the series of compounds (e.g., different substituent positions in an
aromatic ring) explains the variance in the activities block. Although no physicochemical meaning, like
group electronegativity and volume, nor conformation aspects are supposed to be considered in MIA-
QSAR, it seems that structural shapes play a significant role in deriving a QSAR model. The steps required
to achieve a MIA-QSAR model are depicted as follows:
1) Building of the 2D chemical structures: this step requires drawing chemical structures using an
appropriate program for this purpose. There are numerous commercially and freely available
programs used to this end, but an important task is to represent compounds systematically, for
example: use the same font type and size for chemical elements of different molecules in a given
series; use the same bond length (usually represented as sticks) and ring sizes for all compounds;
use the same connectivity rules for similar substituent groups, in order to allow maximum
similarity when aligning, etc. 3D information, like representation of stereogenic centers, may be
schematically designated as hashed or wedge lines (bonds) in relation to the chiral center. An
example of how a chemical structure can be draw is:
O
O
2) Alignment of the 2D chemical structures: each image (chemical structure) should be saved in
a workspace with well defined m×n dimension, e.g., using the Paint applicative of Windows.
The result will be independent of the file extension (.bmp, .jpeg, .png, .tiff, etc.) [23]. Also, a
common pixel among the whole series of compounds should be manually adjusted in a given
coordinate of the workspace, since all images will be superposed later and the similarity
moieties of all congeneric compounds must be congruent; the variable substructures
correspond to variance in the activities block – the basis of a structure-based QSAR. For the
series of organotin compounds used as mosquito (Ae. aegypti and An. stephensi) controllers,
Sn