Quantitative structure-activity relationship modeling of dopamine D1 antagonists using comparative molecular field analysis, genetic algorithms- partial least-squares, and K nearest neighbor methods

Article abstract of DOI:10.1021/jm980415j

Several quantitative structure-activity relationship (QSAR) methods were applied to 29 chemically diverse D₁ dopamine antagonists. In addition to conventional 3D comparative molecular field analysis (CoMFA), cross-validated R² guided

Full text of DOI:10.1021/jm980415j

J . Med. Chem. 1999, 42, 3217-3226
3217
Qu a n tita tive Str u ctu r e-Activity Rela tion sh ip Mod elin g of Dop a m in e D1
An ta gon ists Usin g Com p a r a tive Molecu la r F ield An a lysis, Gen etic
Algor ith m s-P a r tia l Lea st-Squ a r es, a n d K Nea r est Neigh bor Meth od s
Brian Hoffman,† Sung J in Cho,† Weifan Zheng,† Steven Wyrick,† David E. Nichols,§ Richard B. Mailman,*,†,‡ and
Alexander Tropsha*,†
Division of Medicinal Chemistry and Natural Products, School of Pharmacy, and Departments of Psychiatry and
Pharmacology, University of North Carolina, Chapel Hill, North Carolina 27599, and Department of Medicinal Chemistry and
Molecular Pharmacology, School of Pharmacy, Purdue University, West Lafayette, Indiana 47907
Received J uly 13, 1998
Several quantitative structure-activity relationship (QSAR) methods were applied to 29
chemically diverse D1 dopamine antagonists. In addition to conventional 3D comparative
molecular field analysis (CoMFA), cross-validated R2 guided region selection (q2-GRS) CoMFA
(see ref 1) was employed, as were two novel variable selection QSAR methods recently developed
in one of our laboratories. These latter methods included genetic algorithm-partial least squares
(GA-PLS) and K nearest neighbor (KNN) procedures (see refs 2-4), which utilize 2D topological
descriptors of chemical structures. Each QSAR approach resulted in a highly predictive model,
with cross-validated R2 (q2) values of 0.57 for CoMFA, 0.54 for q2-GRS, 0.73 for GA-PLS, and
0.79 for KNN. The success of all of the QSAR methods indicates the presence of an intrinsic
structure-activity relationship in this group of compounds and affords more robust design
and prediction of biological activities of novel D1 ligands.
In tr od u ction
model of binding affinity. The first of these studies
(Charifson et al.11) detailed the conformational studies
and multiple linear regression (MLR) QSAR for a series
of seven 1-phenyltetrahydroisoquinolines, compounds
synthesized as ring-contracted analogues of the proto-
typical D1 antagonist SCH23390. These analogues,
although slightly lower in affinity, served as important
probes of the D1 antagonist pharmacophore, as they
were useful in confirming the stereochemistry of the
active site and provided variation in the 3D distribution
of the pharmacophoric elements. Using the calculated
dipole moment orientations of these ligands, a MLR
QSAR model with a r2 value of 0.95 was reported,11
whereas models utilizing pharmacophoric atom distance
comparisons yielded less favorable results.
The G protein-coupled receptor superfamily class
includes the D1-like and D2-like dopamine receptors.5
Because dopamine neurotransmission is important in
many aspects of CNS function (e.g., the modulation of
motor activity and emotional states6,7), it is not surpris-
ing that dopamine receptors play a role in the therapy
and, possibly, the etiology of disorders such as Parkin-
son’s disease and schizophrenia.8 For these reasons,
there has been a great interest in studies aimed at
mechanistic and functional understanding of dopamine
receptor pharmacology. Despite the pace of advances in
molecular and pharmacological knowledge of dopamine
receptors, their X-ray crystallographic structure is not
available. Thus, there is a particular need to develop
molecular models that could help to both understand
pharmacological data and predict novel biologically
active compounds. Ligand-based methods of analysis,
such as pharmacophore mapping and quantitative
structure-activity relationships (QSAR), remain the
major approach for developing predictive correlations
between ligand structure and activity.
The second literature report of D1 antagonist QSAR
(Minor et al.12) extended the work of Charifson et al.
via the incorporation of additional tetrahydroisoquino-
line (TIQ) analogues. Addition of five such TIQs to the
MLR QSAR model of Charifson et al. resulted in a
dramatic decrease in the correlation coefficient (r2
)
0.22), indicating the limitations of this QSAR method.
Inclusion of a second regressor, in the form of the torsion
angle of the accessory phenyl ring bond, significantly
enhanced the correlation, with an r2 value of 0.92. Minor
et al.12 also reported the application of a 3D QSAR
approach that made use of the field fit option of CoMFA
for the analysis of a data set composed of 11 compounds.
The resultant poor q2 value (0.16) from this alignment
method prompted the researchers to incorporate the
torsion angle of the accessory phenyl ring bond as a
regressor variable, as had been used in the development
of the conventional QSAR analysis. This multiple re-
gressor approach was more successful, with a q2 of
0.578.
Recent modeling studies of dopaminergic ligands by
us and other research groups have yielded several
pharmacophore and QSAR models for a limited number
of D1 agonists9,10 and antagonists11,12 (recently re-
viewed13). The two literature accounts that address the
QSAR of D1 antagonists11,12 report the use of conven-
tional regression analysis and CoMFA (comparative
molecular field analysis) to develop a D1 antagonist
* To whom correspondence should be addressed.
† Division of Medicinal Chemistry and Natural Products.
‡ Departments of Psychiatry and Pharmacology.
§ Department of Medicinal Chemistry and Molecular Pharmacology.
10.1021/jm980415j CCC: $18.00 © 1999 American Chemical Society
Published on Web 08/03/1999

3218 J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17
Hoffman et al.
The integrity and predictive power of such models
improve with the use of more extensive training sets
and the development of alternative QSAR models.
Ideally, the compounds comprising the model should be
numerous and of various structural classes. It is espe-
cially desirable that biological data used as the depend-
ent variable be extremely reliable, i.e., derived within
one laboratory utilizing sound experimental techniques.
Furthermore, various QSAR models have the common
purpose of establishing meaningful correlations between
activity and quantitative descriptors of chemical struc-
tures. Thus, the successful development of alternative
QSAR models confirms the existence of a structure-
activity relationship intrinsic to a data set.
indices, that have been developed on the basis of
chemical graph theory.34,35 Furthermore, we have imple-
mented the concept of variable selection in QSAR, a
process that has been explored recently by a number of
researchers36,37 using such optimization methods as
evolutionary algorithms,38,39 genetic algorithms,40,41 and
simulated annealing algorithms.42,43 These consider-
ations led us to develop two variable selection QSAR
algorithms: genetic algorithm-partial least-squares
(GA-PLS)3,4 and K nearest neighbor (KNN)2,4 analyses.
These approaches currently rely upon topological de-
scriptors of chemical structures that eliminate the
conformational and alignment ambiguities inherent
within the CoMFA process. Additionally, they are less
computationally intensive and are practically auto-
mated. They have been used to produce highly predic-
tive models2,3 that were comparable to, or better than,
those obtained with traditional CoMFA.
In this paper, we have applied several QSAR ap-
proaches, including CoMFA, q2-GRS/CoMFA, GA-PLS,
and KNN, to a data set of 29 structurally diverse D1
antagonists. We measured the biological activity of all
antagonists in rat striatal tissue. To the best of our
knowledge, this is the most extensive and diverse set
of D1 antagonists studied experimentally in one labora-
tory. The purpose of this multiple QSAR analysis was
both to test, in comparison, the efficiency of 2D vs 3D
QSAR methods and to establish robust QSAR models
for these ligands that can be used in the design of new
high-affinity antagonists. In the case of 3D QSAR
methods, i.e., CoMFA and q2-GRS, the Active Analogue
Approach was applied for identification of the active
pharmacophore in the alignment process. All four
methods led to comparable QSAR models with cross-
validated R2 (q2) values of 0.57 for CoMFA, 0.54 for q2-
GRS, 0.73 for GA-PLS, and 0.79 for KNN.
Several QSAR approaches have been developed over
the years. Rapid accumulation of experimental 3D
structural information about organic molecules of bio-
logical interest,14,15 paralleled by the more recent de-
velopment of fast and accurate methods for 3D structure
generation (e.g., CONCORD16,17) and alignment (e.g.,
Active Analogue Approach18,19), has led to the develop-
ment of 3D structural descriptors and associated 3D
QSAR methods. The examples of such descriptors
include 3D shape descriptors used in molecular shape
analysis,20,21 steric and electrostatic field sampling
implemented in CoMFA,22 and comparative molecular
similarity indices analysis (CoMSIA)23 (several recent
reviews on 3D QSAR have been published24). One
common characteristic of these methods, as opposed to
traditional QSAR, is a dramatic increase in the number
of descriptors. As this number increases, multiple
regression methods become inadequate. Advances in
chemometrics (e.g., principal component analysis25 and
partial least-squares22-26) and machine learning algo-
rithms (e.g., neural network27,28), however, have pro-
vided researchers with adequate statistical tools to deal
with this problem.
Meth od s
Reasonable simplicity, a high degree of automation,
and a clear physicochemical sense of steric and electro-
static descriptors have made CoMFA one of the most
popular methods for QSAR.29 However, despite many
successful applications, several problems have persisted
with this method. For instance, we have shown recently
that the results of conventional CoMFA may often be
nonreproducible due to a sometimes-strong dependence
of the CoMFA cross-validated correlation coefficient, q2,
on the orientation of rigidly aligned molecules on the
users’ terminal.1 We have offered a solution to this
problem1,30 by developing a q2-guided region selection
(q2-GRS) method that is based on rational selection of
only the most significant regions of steric and electro-
static fields of aligned molecules. Nevertheless, espe-
cially for structurally diverse molecules, unambiguous
3D alignment in order to initiate the CoMFA process
frequently remains a difficult task. In some reported
cases, only the knowledge of 3D receptor structure has
enabled the effective application of CoMFA.31,32
Biologica l Activity Da ta . For this work, we have selected
29 chemically diverse D1 dopamine antagonists (Tables 1-3)
whose receptor affinity was measured in our laboratory. This
data set includes a variety of tetrahydroisoquinolines (Tables
1 and 2), tetrahydrobenzazepines (Table 2), and phenylamino-
tetralins (Table 2), as well as representatives of several other
series. The competition binding activity of the compounds is
expressed as -log(K0.5) (Tables 1-3). The K0.5 for a compound
is an “apparent” affinity constant that is not affected by
experimental variables such as varied radioligand concentra-
tions. It may be considered a “corrected IC50”. In the case of
racemic mixtures, the affinity of the active enantiomer was
approximated by dividing the K0.5 by 2, making the assumption
that one enantiomer was significantly less active than the
other. This assumption is likely to be valid, based on the
available data with this receptor and compounds of the
structural classes we have utilized. Thus, for those compounds
for which an active isomer was not known, there was extensive
data on the absolute configuration of many close analogues.
For instance, the active configuration of every 1-phenyltetra-
hydrobenzazepine that has been resolved is identical.44-46
Similarly, the active configuration of every 1-phenyltetrahy-
droisoquinoline is also identical, although opposite from that
of the 1-phenyltetrahydrobenzazepines.11,12 Finally, the hy-
pothesis of a single active isomer was confirmed by the recent
analysis of the novel D1 ligand dinapsoline.47 Thus, modeling
these analogues using this assumption is not based on
conjecture alone.
The obvious difficulties with 3D alignment typical for
3D QSAR methods motivated us, as well as other
researchers,33 to consider possible alternatives that
would combine the simplicity of 2D QSAR with the
powerful statistical methods employed in 3D QSAR. To
this end, we have considered the use of multiple 2D
descriptors of chemical structures, such as topological
Ra d ior ecep tor Assa ys. Adult male Sprague-Dawley rats
were decapitated, and the corpus striatum was removed.

QSAR Modeling of Dopamine D1 Antagonists
J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17 3219
Ta ble 1. Comparison of Different QSAR Methods as Applied to Compounds 1-11 of the 29 D1 Antagonists Studied
Striatal tissue was frozen immediately on dry ice and stored
at -80 °C until needed. Homogenization of the tissue was
effected in 50 mM HEPES buffer at pH 7.4 (25 °C) by seven
manual strokes in a Wheaton Teflon-glass homogenizer. The
tissue suspension was centrifuged at 27000g for 10 min, and
the supernatant was discarded. This wash step was repeated
once. The final pellet was resuspended at 1.25 mg/mL buffer.
Binding was performed in 12- × 75-mm culture tubes at a
volume of 1.0 mL. Each tube contained 800 µL of tissue
homogenate suspension, 100 µL of the competing ligand
solution, and 100 µL of the radioligand solution. Competing
ligands were dissolved in methanol at a concentration of 1.0
mM and diluted with buffer to the appropriate concentrations.
[3H]SCH23390 was diluted from a methanol stock solution to
give a final incubation concentration of 0.25 nM. Nonspecific
binding was defined with unlabeled SCH23390 (1 µM). Binding
was terminated by rapid filtration with 15 mL of ice-cold buffer
on a Skatron cell harvester using glass fiber filter mats.
Following drying of the filters, 2-4 mL of Scintiverse E fluid
was added. After agitation for 30 min, radioactivity was
measured by scintillation analysis.
Most compounds (Tables 1-3) were either synthesized in
our laboratories or obtained from commercial sources. Other
compounds included SCH39166 (compound 26), a generous gift
of Schering-Plough, Inc., and the N-substituted analogues of
SCH23390 (compounds 13-16), provided by the Division of
Intramural Research, National Institute on Drug Abuse,
Baltimore, MD.
Syn th esis. In addition to existing antagonists, several novel
compounds were synthesized and included in the data set
(compounds 2-8). These compounds, 1-phenyltetrahydroiso-
quinolines, represent ring-contracted analogues of the 1-phen-
yltetrahydrobenzazepine SCH23390 and possess accessory
phenyl ring substituents. The synthesis of these ligands
followed the scheme of Charifson11 utilizing an appropriate
substituted benzoyl chloride. Scheme 1 outlines this general
synthesis, while Table 4 lists relevant analytical data for these
novel compounds.
field energies were calculated using sp3 carbon probe atoms
with a +1 charge. Low-energy conformers were obtained via
the SYBYL random search method. The CoMFA QSAR equa-
tions were calculated using the partial least-squares (PLS)
algorithm. The optimal number of components (ONC) in the
final PLS model was determined by the cross-validated R2 (q2)
and standard error of prediction (SDEP) values, as obtained
from the leave-one-out cross-validation technique. The q2 value
was calculated from the following standard equation:
(yi - yˇi)2
∑
q2 ) 1 -
(1)
(yi - yj)2
∑
in which yi and yˇi are the actual activity and the predicted
activity of the ith compound, respectively, and yj is the average
activity of all the compounds in the training set. Both sum-
mations are inclusive of all compounds in the training set. The
number of components with the lowest SDEP value was
selected as the ONC. All calculations were performed on a
Silicon Graphics Indigo2 workstation.
Str u ctu r e Align m en t. This section describes the method-
ology employed only in the cases of CoMFA and q2-GRS, since
these methods require rational 3D alignment of the database
molecules for generation of the descriptors. We have utilized
the protonated forms of the compounds, as it has been well-
established that the nitrogens would exist mainly in the
protonated form at physiological pH49 and all ligands that are
known to bind to dopamine receptors possess an amino
function with a pKa greater than 7.50 The geometry of each
antagonist was optimized individually with the Tripos force
field, with no constraints on the internal geometry of the
molecules. Compound 26 (SCH39166), a ring-constrained
analogue of compound 12 (SCH23390), was chosen as the
template molecule due to its rigidity and relatively high
affinity. The conformations of compound 12 and the template
antagonist 26 were based on their published51 X-ray crystal
structures. Additional X-ray crystal structures were obtained
from the Cambridge Structural Database, including those for
compounds 28 and 29. The conformation of compounds 1 and
24 that were utilized in our model have also been described
previously,11,12 as have the conformations of N-methyl-
substituted compounds 13-16.52 Compound 25 was modeled
in the half-chair conformation of tetrahydroisoquinolines
described by Charifson et al.; however, in the absence of
Con ven tion a l CoMF A. Structures were generated and
CoMFA was performed within the QSAR module of the SYBYL
molecular modeling software.48 Default Sybyl settings were
used except as otherwise noted. Molecular mechanics calcula-
tions were performed with the standard Tripos force field, with
a convergence criterion requiring a minimum energy change
of 0.05 kcal/mol. Charges were calculated using the Delre
method as implemented in SYBYL. The steric and electrostatic

3220 J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17
Hoffman et al.
Ta ble 2. Comparison of Different QSAR Methods as Applied to Compounds 12-23 of the 29 D1 Antagonists Studied
conformational data regarding the probable orientation of the
accessory benzyl ring (as exists for the accessory phenyl ring
of compounds 111 and 12),53,54 we modified the torsion angle
of the benzyl group to mimic the phenyl ring position of the
template molecule 26. For compounds 17-21, the chosen
conformation of compound 24 was utilized, with the deletion
and addition of appropriate atoms, followed by geometry
optimization as described above. This procedure was also used
in defining the conformation of compound 27. The appropriate
conformations of compounds 22 and 23 were determined by
deletion and addition of appropriate atoms of the chosen
conformation of compound 1, geometry optimization as de-
scribed above, and a subsequent random search using Sybyl
default values (except “check chirality”, set to preserve the
stereochemistry of the molecule). This random search yielded
two conformer candidates, with energy values of 9.88 and 10.53
kcal/mol, and the conformer with the lower energy value was
retained for alignment.
P h a r m a cop h or e Assign m en t. Structural alignment for
CoMFA and q2-GRS is dependent upon a rational identification
of the pharmacophore of each compound in the data set. The
alignment of our data set is illustrated in Figure 1. For this
process of pharmacophore assignment, we followed our previ-
ously developed model of the D1 receptor active site that was
obtained with the Active Analogue Approach.19 According to
that model, the pharmacophore of the tetrahydroisoquinolines
(compounds 1-11, 22-25) and tetrahydrobenzazepines (com-
pounds 12-16, 26) consists of the chlorine (or other halogen
in some molecules), the hydroxyl group oxygen (O.3), the
quaternary nitrogen (N.4), and the centroid of the accessory
phenyl ring. The following alignment rules were utilized for
antagonists that were neither tetrahydrobenzazepines nor
tetrahydroisoquinolines.
1. Compounds 17-21, phenylaminotetralin analogues, were
aligned using the oxygen (O.3) of their hydroxyl groups (fit to
the oxygen of the template molecule), their secondary, tertiary,
or quaternary nitrogen (fit to the quaternary nitrogen of the
template), the 6-position hydrogen (fit to the chlorine of the
template molecule), and the centroid of their accessory phenyl
rings (fit to the centroid of the accessory phenyl ring of the
template).
2. Compound 27, a fused-ring phenylaminotetralin ana-
logue, was aligned by its iodine substituent (fit to the chlorine
of the template), the hydroxyl group oxygen (O.3) (fit to the
oxygen of the hydroxyl group of the template), the quaternary
nitrogen (N.4) (fit to the quaternary nitrogen of the template),
and the centroid of the accessory phenyl ring (fit to the centroid
of the accessory phenyl ring of the template).
3. Compound 28, flupenthixol, was aligned by the centroid
of the substituted phenyl ring (fit to the fused ring of the
template), the centroid of the second phenyl ring (fit to the
centroid of the accessory phenyl ring of the template), and the
tertiary nitrogen (N.3) (fit to the quaternary nitrogen of the
template).

QSAR Modeling of Dopamine D1 Antagonists
J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17 3221
Ta ble 3. Comparison of Different QSAR Methods as Applied to Compounds 24-29 of the 29 D1 Antagonists Studied
4. Compound 29, butaclamol, was oriented using the cen-
troid of ring C (fit to the fused phenyl ring of the template
molecule), the centroid of ring A (fit to the centroid of the
accessory phenyl ring of the template), and the tertiary
nitrogen (N.3) (fit to the quaternary nitrogen of the template).
descriptors) and descriptors with zero variance are removed
(the atom id-dependent indices are eliminated because atom
ids are assigned arbitrarily and variations in these indices do
not reflect any structural changes). Step 3. All applicable
descriptors are numbered arbitrarily, and this enumeration
is maintained throughout the entire analysis. A population of
100 different random combinations of these descriptors is
generated. To apply GA methodology, each combination is
considered as a parent. Each parent represents a binary string
of digits, either “1” or “0”; the length of each string is the same
and is equal to the total number of descriptors (indices). The
value of “1” implies that the corresponding descriptor is
included for the parent, and “0” means that the descriptor is
excluded. Step 4. Using each parent combination of descrip-
tors, a QSAR equation is generated for the whole data set
using the PLS algorithm; thus for each parent an initial value
of q2 is obtained. The [1 - (n - 1)(1 - q2)/(n - c)] expression
(in which q2 is cross-validated R2, n is the number of com-
pounds, and c is the optimal number of components) then is
used as the fitting function to guide the GA optimization. Step
5. Two parents are selected randomly based on the roulette
wheel selection method. Step 6. The population is evolved by
performing a crossover between two randomly selected par-
ents, producing two offspring. Step 7. Each offspring is
subjected to a random single-point mutation, i.e., a randomly
q2-GRS Rou tin e. The q2-GRS process has been described
in detail elsewhere1,30 (see also recent review55). Unlike
conventional CoMFA, the q2-GRS method leads to reproducible
q2 values that do not depend on the orientation of a molecular
aggregate of aligned molecules on the user terminal.1 The q2-
GRS procedure is as follows: Step 1. Conventional CoMFA is
performed using the automatically generated region file (a
rectangular grid). Step 2. The rectangular grid, encompassing
the aligned molecules, is divided into 125 small boxes of equal
size. Step 3. For each of these newly generated subregion files,
a separate CoMFA is performed, with a step size of 1.0 Å. Step
4. The regions possessing a q2 value greater than the specified
threshold value are selected for further analysis. Step 5. The
selected regions are combined to generate a master region file.
Step 6. The final CoMFA is performed using the master region
file.
GA-P LS Rou tin e. The algorithm of the GA-PLS method3,4
is implemented as follows: Step 1. The Molconn-X program56
is applied to generate 462 variables (topological indices)
automatically. Step 2. All atom id-dependent descriptors (150

3222 J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17
Hoffman et al.
Sch em e 1. Synthesis of Novel Substituted 1-Phenyltetrahydroisoquinolines
Ta ble 4. Analytical Data for Novel Phenyl-Substituted Tetrahydroisoquinolines
selected “1” (or “0”) is changed to “0” (or “1”). Step 8. The
fitness of each offspring is evaluated as described above (cf.
step 4). Step 9. If the resulting offspring are characterized by
a higher value of the fitness function, then they replace less
fit parents; otherwise, the parents are kept. Step 10. Steps
5-9 are repeated until a predefined maximum number of
crossovers are reached.
CON-X program.56 The activity of each compound is predicted
as the average of two (or K in the general case) most similar
compounds in the data set; the similarity is determined by
the Euclidean distance between a pair of compounds in multi-
dimensional descriptor space. The search in the descriptor
space to select the best subset of variables (variable selection)
is performed using stochastic algorithms (simulated annealing
or genetic algorithms). The search continues until it converges
to the subset of descriptors that afford the highest q2 value.
The further details of this method are described elsewhere.2,4
KNN Rou tin e. The essential aspects of this method2,4 are
as follows: All compounds are initially described by multiple
descriptors, i.e., topological indices calculated with the MOL-

QSAR Modeling of Dopamine D1 Antagonists
J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17 3223
F igu r e 1. Stereoview of the 29 aligned D1 antagonists utilized in this study.
F igu r e 2. Stereoview of the electrostatic fields generated with CoMFA, with SCH23390 as a representative ligand: blue indicates
regions where binding is possibly improved by more positively charged substituents; red indicates regions where binding is possibly
improved by more negatively charged substituents.
Resu lts a n d Discu ssion
One or both of these possibilities may contribute to the
lack of improvement by the q2-GRS analysis.
Tables 1-3 list predicted vs actual biological values
for each compound, as determined by each type of QSAR
analysis. The use of biological data from one source is
of particular significance, preventing interlaboratory
variations in the values of dependent variable for the
QSAR analysis. We discuss the results of each analysis
separately vide infra.
Figures 2 and 3 illustrate steric and electrostatic
fields generated by CoMFA. These fields can be inter-
preted in term of structural modifications that might
improve binding affinity with changes in substituent
charges and/or sizes. In each figure, SCH23390 (com-
pound 12) is the representative ligand displayed within
the CoMFA fields. Figure 2 shows electrostatic fields,
with blue regions indicative of areas, where, as predicted
by CoMFA, the affinity might be improved by more
positively charged substituents. Red regions are indica-
tive of areas where the affinity might be improved by
more negatively charged substituents. The red regions
surrounding the 7-chloro substituent of SCH23390 may
be representative of the observation that the binding
affinities of D1 ligands generally improve with 7-position
substituents according to the following trend: I < H <
Br ∼ Cl, as observed in the 1-phenylbenzazepine series
by Iorio et al.57 and Tice et al.46 This trend in affinity
correlates with the electrostatic charges of the 7-position
substituents. The blue region extending from the ter-
tiary nitrogen of SCH23390 correlates to the position
of the extended chain substituents seen in compounds
13-16. While these compounds exhibited only modest
affinity when tested in our laboratories, they had higher
affinity than other extended N-substituted compounds,
such as analogues 9, 20, and 21.
QSAR Mod els Ba sed on CoMF A. The use of our
previously developed pharmacophore model19 for mo-
lecular alignment provided a rational foundation for the
3D QSAR analyses we report here. Compound 26
provided an appropriate template for alignment of the
members of the data set in CoMFA and q2-GRS, given
its high affinity and conformational rigidity. The results
for q2-GRS were practically the same as for conventional
CoMFA (cf. Tables 1-3). This may be due to fortuitously
optimal orientation of the database molecules: the
molecules may have been oriented on the user terminal
in a manner that yielded the highest q2 value possible.
Another possibility is that the molecules of the data set
possess structural features that translate into descrip-
tors that are not improved by the region-selection
procedure of the q2-GRS analysis. For example, although
the data set consisted of several molecular classes of
compounds, within each class there was little deviation
in the pharmacophoric arrangement. Were there more
variability within the data set with regard to the
pharmacophore of each ligand, the advantage of region
selection may have resulted in an improved q2 value.
Similarly, Figure 3 shows CoMFA steric fields. Yellow
regions indicate areas where bulkier substituents lead

3224 J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17
Hoffman et al.
F igu r e 3. Stereoview of the steric fields generated with CoMFA, with SCH23390 as a representative ligand: yellow indicates
regions of possible unfavorable steric interactions of bulky substituents with the receptor; green indicates regions where binding
is possibly improved by bulky substituents.
F igu r e 4. Trajectory of the SA-driven optimization of q2
values in developing the KNN QSAR model for D1 antagonists.
to the decrease in activity, and green regions are
indicative of areas predicted by CoMFA to be improved
by bulkier substituents. These fields are less pronounced
F igu r e 5. Trajectory of the GA-driven optimization of q2
values for the GA-PLS QSAR model of D1 antagonists.
than those illustrated for electrostatic fields in Figure
2. The yellow region at the bottom right corresponds to
steric factors exhibited by the low-affinity ligands 24
and 27. The green region proximal to the tertiary nitro-
gen appears to match van der Waals contributions from
the extended chain substituents of compounds 13-16.
QSAR Mod els Ba sed on Top ologica l Descr ip -
tor s: Com p a r ison w ith CoMF A. The trajectories of
the q2 values for KNN and GA-PLS QSAR methods in
the course of the model optimization are shown in
Figures 4 and 5, respectively. The final results obtained
with GA-PLS and KNN methods are given in Table 5
as well. These methods, which utilize only 2D topological
description of molecules, yielded q2 values of 0.73 and
0.79 for GA-PLS and KNN calculations, respectively.
Since all QSAR methods employed in this paper use the
same criteria of the quality of the QSAR model, i.e., q2,
the results of the methods can be compared directly.
Apparently, the q2 values obtained from GA-PLS and
KNN QSAR methods are better than those from the 3D
QSAR methods of CoMFA or q2-GRS (0.57 and 0.54,
respectively).
as applied to this data set. Both methods have certain
inherent advantages over 3D methods, such as circum-
vention of alignment issues as well as the elimination
of the possibility of inappropriate conformation selec-
tion. Furthermore, they employ variable selection pro-
cedures that improve the quality of QSAR analysis,
especially if the number of descriptors is large. The
success of alignment-free QSAR methods utilizing 2D
descriptors may also indicate that in the case of rela-
tively rigid compounds such as D1 antagonists, 2D
description of the pharmacophore is sufficient to estab-
lish adequate structure-activity correlations.
The greater success of 2D QSAR methods notwith-
standing, all four QSAR methods have yielded highly
predictive models. The success of these models is
noteworthy, since we have dealt with a large data set
comprised of a variety of structural classes. Structural
variations of the D1 ligands included differing substit-
uents of the accessory phenyl ring, differing positions
and substitutions of the nitrogen, and differing substit-
uents in the catechol ring.
We were curious to explore the contribution of the two
non-catechol-like antagonists (compounds 28 and 29) to
our QSAR models. It is interesting to note that dropping
These results demonstrate that both KNN and GA-
PLS are highly competitive QSAR analytical techniques

QSAR Modeling of Dopamine D1 Antagonists
J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17 3225
Ta ble 5. Statistical Data for QSAR Method Results
these two antagonists from the data set increases the
q2 value for CoMFA from 0.57 to 0.72. In contrast, no
appreciable difference in the q2 value was observed with
2D QSAR approaches. On the basis of these observa-
tions, we have proposed that these “nontraditional”
antagonists bind to the D1 receptor in a different
manner than traditional (benzazepine and isoquinoline)
antagonists. Our preliminary receptor-ligand docking
studies using a tentative 3D model of the D1 receptor
also support this hypothesis. We shall repeat CoMFA
analysis for the whole data set using receptor-based
alignment of all compounds once our receptor model has
been refined. The lack of sensitivity of both GA-PLS
and KNN methods to the presence or absence of these
two compounds in the data set is not surprising, since
neither 2D QSAR method relies upon compound align-
ment.
occupying part of the same binding region but probably
interacting with different residues. This hypothesis is
currently under investigation in our group using a 3D
model of the D1 receptor, and the results will be a
subject of separate publication.
The relative success of QSAR approaches as divergent
as those applied in this paper indicates the existence of
an inherent quantitative structure-activity relationship
for D1 receptor ligands, which can be expressed using a
variety of formal molecular descriptors and QSAR
models. The concurrent use of these different models
should significantly increase our confidence in the
prediction of biological activity for newly designed
compounds.
Refer en ces
(1) Cho, S. J .; Tropsha, A. Cross-Validated R2 Guided Region
Selection for Comparative Molecular Field Analysis (CoMFA):
A Simple Method to Achieve Consistent Results. J . Med. Chem.
1995, 38, 1060-1066.
Con clu sion s a n d P r osp ectu s
We have developed several QSAR models for 29
diverse antagonists of the D1 receptor. Although these
models are based on different QSAR protocols and
different types of descriptors, they are all statistically
valid as indicated by high q2 values. The CoMFA fields
generated from these studies predict improved affinity
for compounds possessing positively charged substitu-
ents extending from the tertiary nitrogen common to
dopaminergic antagonists. This prediction is in accord
with previously reported data regarding the extended
chain N-substituted compounds 13-16,52,58 although
these drugs possess lower affinity in our biological
assays. The prediction that negatively charged substit-
uents in the position analogous to the 7-position of
SCH23390 would improve affinity is probably based on
the rank order of affinity of various halogen-substituted
compounds, with more negatively charged chlorine-
substituted compounds more potent than iodinated or
unsubstituted compounds. This prediction is not sup-
ported by literature accounts of binding affinity of a
fluoro-substituted analogue of SCH23390, which has
weaker affinity than chloro- and bromo-substituted
analogues.57
An interesting outcome of these QSAR studies is the
observation that inclusion of two non-catechol antago-
nists, compounds 28 and 29, lowers the q2 values of the
3D QSAR analytical methods used here. When included
in data sets, analyzed by KNN and GA-PLS methods,
these compounds do not decrease q2 values, likely
because these QSAR methods do not rely upon align-
ment of members of the data sets. We have proposed
that these antagonists 28 and 29 bind to the D1 receptor
in a manner that differs from catechol-like antagonists,
(2) (a) Zheng, W.; Cho, S. J .; Tropsha, A. Novel computational tools
for rational drug design using combinatorial chemistry and
QSAR. Books of Abstracts. 213th National Meeting, Las Vegas,
NE, September 8-12, 1997; American Chemical Society: Wash-
ington, DC. (b) Zheng, W.; Tropsha, A. Application of the K
Nearest Neighbor with Variable Selection Method to QSAR. J .
Comput. Chem., accepted.
(3) (a) Cho, S. J .; Cummins, D.; Bentley, J .; Andrews, C. W.;
Tropsha, A. Application of Genetic Algorithms and Partial Least
Squares to Variable Selection of Topological Indices. J . Comput.-
Aided Mol. Des., submitted. (b) The method is described in detail
and can be executed on the QSAR Web server available at http://
mmlin1.pha.unc.edu/∼jin/QSAR.
(4) Tropsha, A.; Cho, S. J .; Zheng, W. “New Tricks for an Old Dog”:
Development and Application of Novel QSAR Methods for
Rational Design of Combinatorial Chemical Libraries and
Database Mining. Rational Drug Design: Novel Methodology
and Practical Applications; ACS Symp. Series 719; American
Chemical Society: Washington, DC, 1999, in press.
(5) Kebabian, J .; Calne, D. B. Multiple receptors for dopamine.
Nature 1979, 277, 93-96.
(6) Strange, P. G. Brain Biochemistry and Brain Disorders; Oxford
University Press: New York, 1993.
(7) Waddington, J . D1:D2 Dopamine Receptor Interactions; Academic
Press: New York, 1993.
(8) Seeman, P.; Bzowej, N. H.; Guan, H. C.; Bergeon, C.; Reynolds,
G. P.; Bird, E. D.; Riederer, P.; J ellinger, K.; Tortellotte, W. W.
Human brain D1 and D2 dopamine receptors in schizophrenia,
Alzheimer’s, Parkinson’s, and Huntington’s diseases. Neuro-
psychopharmacology 1987, 1, 5-15.
(9) Mottola, D. M.; Laiter, S.; Watts, V. J .; Tropsha, A.; Wyrick, S.
W.; Nichols, D. E.; Mailman, R. B. Conformational analysis of
D
1 dopamine receptor agonists: pharmacophore assessment and
receptor mapping. J . Med. Chem. 1996, 39, 285-296.
(10) Brusnniak, M. K.; Pearlman, R. S.; Neve, K. A.; Wilcox, R. E.
Comparative molecular Field Analysis-based prediction of drug
affinities at recombinant D1A dopamine receptors. J . Med. Chem.
1996, 39, 850-859.
(11) Charifson, P. S.; Bowen, J . P.; Wyrick, S. D.; Hoffman, A. J .;
Cory, M.; McPhail, A. T.; Mailman, R. B. Conformational
analysis and molecular modeling of 1-phenyl-, 4-phenyl-, and
1-benzyl-1,2,3,4-tetrahydroisoquinolines as D1 dopamine recep-
tor ligands. J . Med. Chem. 1989, 32, 2050-2058.

3226 J ournal of Medicinal Chemistry, 1999, Vol. 42, No. 17
Hoffman et al.
(12) Minor, D. L.; Wyrick, S. D.; Charifson, P. S.; Watts, V. J .; Nichols,
D. E.; Mailman, R. B. Synthesis and molecular modeling of
1-phenyl-1,2,3,4-tetrahydroisoquinolines and related 5,6,8,9-
tetrahydro-13bH-dibenzo[a,h]quinolizines as D1 dopamine an-
tagonists. J . Med. Chem. 1994, 37, 4317-4328.
(13) Mailman, R. B.; Nichols, D. E.; Tropsha, A. Molecular Drug
Design and Dopamine Receptors. In Dopamine Receptors, Neve,
K., Neve, R., Eds.; Humana Press: Totowa, NJ , 1996; pp 105-
133.
(14) Allen, F. H.; Davies, J . E.; Galloy, J . J .; J ohnson, O.; Kennard,
O.; Macrae, C. F.; Mitchell, E. M.; Mitchell, G. F.; Smith, J . M.;
Watson, D. G. The development of versions 3 and 4 of the
Cambridge Structural Database System. J . Chem. Inf. Comput.
Sci. 1991, 31, 187-204.
(15) Allen, F. H.; Bellard, S.; Brice, M. D.; Cartwright, B. A.;
Doubleday, A.; Higgs, H.; Hummelink, T.; Hummelink-Peters,
B. G.; Kennard, O.; Motherwell, W. D. S.; Rodgers, J . R.; Watson,
D. G. The Cambridge Crystal Data Centre: computer-based
search, retrieval, analysis, and display of information. Acta-
Crystallogr. Sect B: Struct. Crystallogr. Cryst. Chem. 1979, B35,
2331-2339.
(35) Hansen, P. J .; J urs, P. C. Chemical Applications of Graph Theory
II: Isomer Enumeration. J . Chem. Educ. 1988, 65, 574.
(36) Kubinyi, H. Variable Selection in QSAR Studies. I. An Evolu-
tionary Algorithm. Quant. Struct.-Act. Relat. 1994, 13, 285-294.
(37) Sutter, J . M.; Dixon, S. L.; J urs, P. C. Automated Descriptor
Selection for Quantitative Structure-Activity Relationships
Using Generalized Simulated Annealing. J . Chem. Inf. Comput.
Sci. 1995, 35, 77-84.
(38) Fogel, D. B. Applying Evolutionary Programming to Selected
Traveling Salesman Problems. Cybern. Syst. (USA) 1993, 24,
27-36.
(39) Fogel, D. B.; Fogel, L. J .; Porto, V. W. Evolutionary Methods for
Training Neural Networks; IEEE Conference on Neural Net-
works for Ocean Engineering (Catal. No. 91CH3064-3); IEEE,
1991; pp 317-327.
(40) Goldberg, D. E. Genetic Algorithm in Search, Optimization, and
Machine Learning; Addison-Wesley: Reading, MA, 1989.
(41) Forrest, S. Genetic Algorithms: Principles of Natural Selection
Applied to Computation. Science 1993, 261, 872-878.
(42) Bohachevsky, I. O.; J ohnson, M. E.; Stein, M. L. Generalized
Simulated Annealing for Function Optimization. Technometrics
1986, 28, 209-217.
(43) Kalivas, J . H., Generalized Simulated Annealing for Calibration
Sample Selection from an Existing Set and Orthoginization of
Undesigned Experiments. J . Chemomet. 1991, 5, 37-48.
(44) Iorio, L. C.; Barnett, A.; Leitz, F. H.; Houser, V. P.; Korduba, C.
A. SCH23390, a potential benzazepine antipsychotic with unique
interactions on dopaminergic systems. J . Pharmacol. Exp. Ther.
1983, 226, 462-468.
(45) Iorio, L. C.; Barnett, A.; Billard, W.; Gold, E. H. Benzazepines:
structure-activity relationships between D1 receptor blockade
and selected pharmacological effects. In Neurobiology of Central
D1-Dopamine Receptors; Creese, I., Breese, G. R., Eds.; Plenum
Press: New York, 1986; pp 1-14.
(46) Tice, M. A.; Hashemi, T.; Taylor, L. A.; Duffy, R. A.; McQuade,
R. D. Characterization of the binding of SCH39166 to the five
cloned dopamine receptor subtypes. Pharmacol. Biochem. Behav.
1994, 49, 567-571.
(47) Ghosh, D.; Snyder, S. E.; Watts, V. J .; Mailman, R. B.; Nichols,
D. E. 8,9-Dihydroxy-2,3,7,11b-tetrahydro-1H-naph[1,2,3-de]iso-
quinoline: a potent full dopamine D1 agonist containing a rigid
â-phenyl dopamine pharmacophore. J . Med. Chem. 1996, 39,
549-555.
(48) Sybyl User’s Manual Version 6.2; Tripos, Inc.: St. Louis, MO,
1995.
(49) Miller, D. D. Steric aspects of dopaminergic drugs. Fed. Proc.
1978, 37, 2392-2395.
(50) Strange, P. G. The binding of agonists and antagonists to
dopamine receptors. Biochem. Soc. Trans. 1996, 24, 188-192.
(51) Berger, J . G.; Chang, W. K.; Clader, J . W.; Hou, D.; Chipkin, R.
E.; McPhail, A. T. Synthesis and receptor affinities of some
conformationally restricted analogues of the dopamine D1 selec-
tive ligand (5R)-8-chloro-2,3,4,5-tetrahydro-3-methyl-5-phenyl-
1H-3-benzazepin-7-ol. J . Med. Chem. 1989, 32, 1913-1921.
(52) Shah, J . H.; Kline, R. H.; Geter-Douglass, B.; Izenwasser, S.;
Witkin, J . M.; Newman, A. H. (()-3-[4′-(N,N-dimethylamino)-
cinnamyl]benzazepine analogues: novel dopamine D1 receptor
antagonists. J . Med. Chem. 1996, 39, 3423-3428.
(53) Pettersson, I.; Gundertofte, K.; Palm, J .; Liljefors, T. A study
on the contribution of the 1-phenyl substituent to the molecular
electrostatic potentials of some benzazepines in relation to
selective dopamine D-1 receptor activity. J . Med. Chem. 1992,
35, 502-507.
(16) Rusinko, A., III; Skell, J . M.; Balducci, R.; McGarity, C. M.;
Pearlman, R. S. Concord, A Program for the Rapid Generation
of High Quality Approximate 3-Dimensional Molecular Struc-
tures; The University of Texas at Austin and Tripos Associates:
St. Louis, MO, 1988.
(17) Pearlman, R. S. Rapid generation of high quality approximate
3D molecular structures. Chem. Des. Aut. News 1987, 2, 1-6.
(18) Marshall, G. R.; Barry, C. D.; Bosshard, H. E.; Dammkoehler,
R. A.; Dunn, D. A. The Conformational Parameter in Drug
Design: the Active Analogue Approach. In Computer-Assisted
Drug Design; Olson, E. C., Christoffersen, R. E., Eds.; American
Chemical Society, Washington, DC, 1979; Vol. 112, pp 205-226.
(19) Marshall, G. R.; Cramer, R. D., III. Three-Dimensional Structure-
Activity Relationships. TIPS Rev. 1988, 9, 285-289.
(20) Simon, Z.; Badileuscu, I.; Racovitan, T. Mapping of dihydrofolate-
reductase receptor site by correlation with minimal topological
(steric) differences. J . Theor. Biol. 1977, 66, 485-495.
(21) Rhyu, K. B.; Patel, H. C.; Hopfinger, A. J . A 3D-QSAR Study of
Anticoccidial Triazines using molecular shape analysis J . Chem.
Inf. Comput. Sci. 1995, 35, 771-778.
(22) Cramer, R. D., III; Patterson, D. E.; Bunce, J . D. Comparative
Molecular Field Analysis (CoMFA). 1. Effect of Shape on Binding
of Steroids to Carrier Proteins. J . Am. Chem. Soc. 1988, 110,
5959-5967.
(23) Klebe, G.; Abraham, U.; Mietzner, T. Molecular Similarity
Indices in a Comparative Analysis (CoMSIA) of Drug Molecules
to Correlate and Predict Their Biological Activity. J . Med. Chem.
1994, 37, 4130-4146.
(24) Kubinyi, H.; Folkers, G.; Martyn, Y. 3D QSAR in Drug Design;
Kluwer/ESCOM: Leiden, 1998; Vol. 2-3.
(25) Camilleri, P.; Livingstone, D. J .; Murphy, J . A.; Manallack, D.
T. Chiral Chromatography and Multivariate Quantitative Struc-
ture-Property Relationships of Benzimidazole Sulphoxides. J .
Comput.-Aided Mol. Des. 1993, 7, 61-69.
(26) Geladi, P.; Kowalski, B. R. Partial Least Squares Regression:
A Tutorial. Anal. Chim. Acta 1986, 185, 1-17.
(27) J ansson, P. A. Neural Networks: An Overview. Anal. Chem.
1991, 63, 357A-362A.
(28) So, S. S.; Richards, W. G. Application of Neural Networks:
Quantitative Structure-Activity Relationships of the Deriva-
tives of 2,4-Diamino-5-(substituted-benzyl)pyrimidines as DHFR
Inhibitors. J . Med. Chem. 1992, 35, 3201-3207.
(29) Cramer, R. D., III; DePriest, S. A.; Patterson, D. E.; Hecht, P.
The Developing Practice of Comparative Molecular Field Analy-
sis. In 3D QSAR in Drug Design: Theory, Methods, and
Applications; Kubinyi, H., Ed.; ESCOM: Leiden, 1993; pp 443-
485.
(30) Cho, S. J .; Tropsha, A.; Suffness, M.; Cheng, Y. C.; Lee, K. H.
Antitumor Agents. 163. Three-Dimensional Quantitative Struc-
ture-Activity Relationship Study of 4′-O-Demethylpipodophyl-
lotoxin Analogues Using the Modified CoMFA/q2-GRS Approach.
J . Med. Chem. 1996, 39, 1383-1395.
(31) Waller, C. L.; Oprea, T. I.; Giolitti, A.; Marshall, G. R. Three-
Dimensional QSAR of Human Immunodeficiency Virus (I)
Protease Inhibitors. 1. A CoMFA Study Employing Experimen-
tally-Determined Alignment Rules. J . Med. Chem. 1993, 36,
4152-4160.
(54) Pettersson, I.; Liljefors, T.; Bogeso, K. Conformational analysis
and structure-activity relationships of selective dopamine D-1
receptor agonists and antagonists of the benzazepine series. J .
Med. Chem. 1990, 33, 2197-2204.
(55) Tropsha, A.; Cho, S. J . Cross-Validated R2 Guided Region
Selection for CoMFA Studies. In Kubinyi, H., Folkers, G.,
Martin, Y. C., Eds.; 3D QSAR in Drug Design; Kluwer Academic
Publishers: Dordrecht, The Netherlands, 1998; Vol. III, pp 57-
69.
(56) MOLCONN-X version 2.0, Hall Associates Consulting: Quincy,
MA.
(57) Iorio, L. C.; Barnett, A.; Billard, W.; Gold, E. H. Benzazepines:
structure-activity relationships between D1 receptor blockade
and selected pharmacological effects. Adv. Exp. Med. Biol. 1986,
204, 1-14.
(58) Shah, J . H.; Izenwasser, S.; Geter-Douglass, B.; Witkin, J . M.;
Newman, A. H. (()-(N-alkylamino)benzazepine analogues: novel
dopamine D1 receptor antagonists. J . Med. Chem. 1995, 38,
4284-4293.
(32) Cho, S. J .; Garsia, M. L. S.; Bier, J .; Tropsha, A. Structure Based
Alignment and Comparative Molecular Field Analysis of Acetyl-
cholinesterase Inhibitors. J . Med. Chem. 1996, 39, 5064-5071.
(33) Rogers, D.; Hopfinger, A. J . Application of Genetic Function
Approximation to Quantitative Structure-Activity Relationships
and Quantitative Structure-Property Relationships. J . Chem.
Inf. Comput. Sci. 1994, 34, 854-866.
(34) Trinajstic, N. Chemical Graph Theory; CRC Press: Boca Raton,
FL, 1983; Vol. I, II.
J M980415J