441
13 Continuous Molecular Fields Approach Applied to Structure-Activity Modeling
possible to deduce fields of model coefficients not only for 3D-QSAR regression,
but also for: (a) classification (in the case of Support Vector Machines they describe
angular orientation of the hyperplane separating active from inactive compounds in
the infinite-dimensional feature space); (b) novelty detection (or one-class classification [31], which can be used for virtual screening); (c) clustering; (d) dimensionality reduction; etc. We hope that numerous methods of analyzing and visualizing
chemical databases and SAR/QSAR/QSPR models offered by the use of continuous
molecular fields would deepen insights into the nature of structure–activity relationships and facilitate drug design. This is one of the main directions of our current
studies in this direction and a topic of future publications.
13.3 Modelling Biological Activity
We validated the CMF approach in two case studies and obtained preliminary results, which have been published as short communications [7, 13]. The first one
dealt with the use of the CMF to build 3D-QSAR regression models [7]. In the
second case study [13], the performance of a new method for virtual screening of
organic compounds based on the combination of the CMF methodology with the
one-class SVM method (1-SVM) has been assessed. In both cases the CMF has not
only proven its efficiency, but has also demonstrated some advantages compared to
state-of-the-art approaches in chemoinformatics.
13.3.1 Building 3D-QSAR Regression Models
We have tested the performance of the CMF approach in building 3D-QSAR regression models by using eight data sets. These sets were selected as expanded benchmark for 3D-QSAR methods as well as examples of various types of biologycal
activity of organic ligands. We have studied 114 angiotensin converting enzyme
(ACE) inhibitors [32], 111 acetylcholinesterase (AChE) inhibitors [33], 163 ligands
for benzodiazepine receptors (BZR) [33], 322 cyclooxygenase-2 (COX-2) inhibitors [33], 397 dihydrofolatereductase (DHFR) inhibitors [33], 66 glycogen phosphorylase b (GPB) inhibitors [34], 76 thermolysin (THER) inhibitors [34], and 88
thrombine (THR) inhibitors [35].
All data on these sets were taken from the supplementary materials to Sutherland’s paper [33]. They included chemical structures, activity values, splitting into
the training and test sets, ionization states and conformations for all molecules, their
spatial alignment and partial charges on atoms. As indicated in [25], ionization states
of molecules had been prepared by deprotonating carbocyclic acids and phosphates
and protonating non-aryl basic amines (except NH 2 groups that coordinate Zn in
the ACE set), energy-minimizing the aligned molecules with MMFF94S force field
in Sybyl was used for determination of atomic coordinates, scaled MNDO ESP-fit
13 Continuous Molecular Fields Approach Applied to Structure-Activity Modeling
possible to deduce fields of model coefficients not only for 3D-QSAR regression,
but also for: (a) classification (in the case of Support Vector Machines they describe
angular orientation of the hyperplane separating active from inactive compounds in
the infinite-dimensional feature space); (b) novelty detection (or one-class classification [31], which can be used for virtual screening); (c) clustering; (d) dimensionality reduction; etc. We hope that numerous methods of analyzing and visualizing
chemical databases and SAR/QSAR/QSPR models offered by the use of continuous
molecular fields would deepen insights into the nature of structure–activity relationships and facilitate drug design. This is one of the main directions of our current
studies in this direction and a topic of future publications.
13.3 Modelling Biological Activity
We validated the CMF approach in two case studies and obtained preliminary results, which have been published as short communications [7, 13]. The first one
dealt with the use of the CMF to build 3D-QSAR regression models [7]. In the
second case study [13], the performance of a new method for virtual screening of
organic compounds based on the combination of the CMF methodology with the
one-class SVM method (1-SVM) has been assessed. In both cases the CMF has not
only proven its efficiency, but has also demonstrated some advantages compared to
state-of-the-art approaches in chemoinformatics.
13.3.1 Building 3D-QSAR Regression Models
We have tested the performance of the CMF approach in building 3D-QSAR regression models by using eight data sets. These sets were selected as expanded benchmark for 3D-QSAR methods as well as examples of various types of biologycal
activity of organic ligands. We have studied 114 angiotensin converting enzyme
(ACE) inhibitors [32], 111 acetylcholinesterase (AChE) inhibitors [33], 163 ligands
for benzodiazepine receptors (BZR) [33], 322 cyclooxygenase-2 (COX-2) inhibitors [33], 397 dihydrofolatereductase (DHFR) inhibitors [33], 66 glycogen phosphorylase b (GPB) inhibitors [34], 76 thermolysin (THER) inhibitors [34], and 88
thrombine (THR) inhibitors [35].
All data on these sets were taken from the supplementary materials to Sutherland’s paper [33]. They included chemical structures, activity values, splitting into
the training and test sets, ionization states and conformations for all molecules, their
spatial alignment and partial charges on atoms. As indicated in [25], ionization states
of molecules had been prepared by deprotonating carbocyclic acids and phosphates
and protonating non-aryl basic amines (except NH 2 groups that coordinate Zn in
the ACE set), energy-minimizing the aligned molecules with MMFF94S force field
in Sybyl was used for determination of atomic coordinates, scaled MNDO ESP-fit