Application of Predictive QSAR Models to Database Mining: Identification and Experimental Validation of Novel Anticonvulsant Compounds

Article abstract of DOI:10.1021/jm030584q

We have developed a drug discovery strategy that employs variable selection quantitative structure-activity relationship (QSAR) models for chemical database mining. The approach starts with the development of rigorously validated QSAR models obtained with

Full text of DOI:10.1021/jm030584q

2356
J . Med. Chem. 2004, 47, 2356-2364
Ap p lica tion of P r ed ictive QSAR Mod els to Da ta ba se Min in g: Id en tifica tion a n d
Exp er im en ta l Va lid a tion of Novel An ticon vu lsa n t Com p ou n d s
Min Shen,† Ce´cile Be´guin,‡ Alexander Golbraikh,† J ames P. Stables,§ Harold Kohn,*,† and Alexander Tropsha*,†
Division of Medicinal Chemistry and Natural Products, School of Pharmacy, CB# 7360, University of North Carolina at
Chapel Hill, Chapel Hill, North Carolina 27599-7360, Department of Chemistry, University of Houston,
Houston, Texas 77204-5641, and Epilepsy Branch, National Institute of Neurological Disorders and Stroke,
National Institutes of Health, 6001 Executive Boulevard, Suite 2106, Rockville, Maryland 20852
Received November 20, 2003
We have developed a drug discovery strategy that employs variable selection quantitative
structure-activity relationship (QSAR) models for chemical database mining. The approach
starts with the development of rigorously validated QSAR models obtained with the variable
selection k nearest neighbor (kNN) method (or, in principle, with any other robust model-
building technique). Model validation is based on several statistical criteria, including the
randomization of the target property (Y-randomization), independent assessment of the training
set model’s predictive power using external test sets, and the establishment of the model’s
applicability domain. All successful models are employed in database mining concurrently; in
each case, only variables selected as a result of model building (termed descriptor pharma-
cophore) are used in chemical similarity searches comparing active compounds of the training
set (queries) with those in chemical databases. Specific biological activity (characteristic of
the training set compounds) of external database entries found to be within a predefined
similarity threshold of the training set molecules is predicted on the basis of the validated
QSAR models using the applicability domain criteria. Compounds judged to have high predicted
activities by all or the majority of all models are considered as consensus hits. We report on
the application of this computational strategy for the first time for the discovery of anticon-
vulsant agents in the Maybridge and National Cancer Institute (NCI) databases containing
ca. 250 000 compounds combined. Forty-eight anticonvulsant agents of the functionalized amino
acid (FAA) series were used to build kNN variable selection QSAR models. The 10 best models
were applied to mining chemical databases, and 22 compounds were selected as consensus
hits. Nine compounds were synthesized and tested at the NIH Epilepsy Branch, Rockville,
MD using the same biological test that was employed to assess the anticonvulsant activity of
the training set compounds; of these nine, four were exact database hits and five were derived
from the hits by minor chemical modifications. Seven of these nine compounds were confirmed
to be active, indicating an exceptionally high hit rate. The approach described in this report
can be used as a general rational drug discovery tool.
In tr od u ction
Over the years, we have advanced a series of anti-
convulsant agents termed functionalized amino acids (1,
Epilepsy is a chronic disorder, characterized by recur-
rent unprovoked seizures. In the United States, some
2 million people, including 340 000 children, suffer from
epilepsy and its sequelae.1 Currently the main treat-
ment for epileptic disorder is the long-term and consist-
ent administration of anticonvulsant drugs.2,3 Unfortu-
nately, current medications are ineffective for more than
a third of the patients with epilepsy.4 Many continue
to have seizures, while others experience disturbing side
effects (e.g., drowsiness, dizziness, nausea, liver dam-
age).5 Current therapies have failed to adequately
control this disorder, documenting the need for new
agents with different mechanisms of action.
FAA).6-8 The molecular target(s) of FAA compounds has
not been identified. Nearly 250 FAAs have been pre-
pared and evaluated in animal seizure models. Of these,
12 conformed to general structure 2 and provided
protection against maximal electroshock (MES) induced
seizures,9,10 at doses comparable to or better than those
for phenytoin.11 The MES test is a proven method for
the identification of new drug candidates for partial
and generalized seizures. While these studies have
provided structural patterns beneficial for FAA activity,
it has become increasingly difficult to formulate a useful
SAR, as the diversity of molecular structures has
increased.
* To whom correspondence should be addressed. Tel. 919-966-2955
(A.T.), 919-966-2680 (H.K.). Fax: 919-966-0204 (A.T.), 919-843-7835
(H.K.). E-mail: alex_tropsha@unc.edu (A.T.), harold_kohn@unc.edu,
(H.K.).
† University of North Carolina at Chapel Hill.
‡ University of Houston.
§ National Institutes of Health.
10.1021/jm030584q CCC: $27.50 © 2004 American Chemical Society
Published on Web 03/20/2004

Application of QSAR Models to Database Mining
J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9 2357
Molecular modeling can facilitate the understanding
of the pharmacological data and permit the development
of novel anticonvulsants. Since the structure of macro-
molecular targets of FAA action remains unknown,
ligand-based methods of analysis, such as pharmaco-
phore mapping and quantitative structure-activity
relationships (QSAR), represent the most efficient ap-
proaches to rationalize available experimental data and
improve the design of new FAA.
Several QSAR models for FAA compounds have been
developed over the years, including our most recent
variable selection k nearest neighbor (kNN) QSAR
models for 48 chemically diverse FAA anticonvulsants
(see ref 12 and references therein). These models could
be employed for further design and discovery of novel
anticonvulsant agents.
F igu r e 1. Flowchart of database mining that employs predic-
tive QSAR models.
One popular computational approach to rational drug
discovery is database mining, which relies on the
structures of known active molecules as queries.13 Most
often, the queries are derived from three-dimensional
(3D) pharmacophores, which entail essential chemical
structural features in the particular mutual orientation
responsible for the biological activity of a series of
compounds.14 Three-dimensional database mining seeks
to identify novel compounds that match the query. If
the repository of compound samples as well as relevant
biological assays are available, the computational hits
can be validated experimentally, ultimately yielding
useful lead compounds.15
among the 22 computational hits and five were designed
on the basis of two of the hits. Seven out of nine tested
compounds were confirmed to be active, indicating an
exceptionally high experimental hit rate. The success
of this study suggests that our approach, which com-
bines QSAR modeling and database mining, can be
explored as a general rational drug discovery tool and
applied to a large variety of available datasets of
biologically active compounds.
Com p u ta tion a l Deta ils
Ap p lica tion of QSAR Mod els to Da ta ba se Min -
in g. Our database-mining strategy makes use of vali-
dated QSAR models developed for the available series
of biologically active compounds. The general flowchart
of the database-mining procedure is shown in Figure 1
and includes the following major steps.
1. Develop validated variable selection QSAR models
for a dataset of compounds with known structures and
activities; define descriptor pharmacophores16 and ap-
plicability domains19 for all models.
2. Compute chemical descriptors used in QSAR model
development for all compounds in the available chemical
database(s); in our studies we have used molecular
topological indices calculated with the MolConnZ pro-
gram.23
3. Calculate chemical similarity values (we use the
Euclidean distance in the descriptor pharmacophore
space) between all active probes (i.e., molecules used for
QSAR model development) and every structure in the
database.
4. Rank all database structures by their similarity to
a probe(s), and select M structures within certain
similarity threshold.
5. Predict biological activity values for these M
structures based on preconstructed QSAR models using
applicability domain.
6. Select structures predicted by all (or a majority of)
QSAR models to have high values of biological activity
as computational hits.
We now discuss individual steps outlined above.
QSAR Mod elin g a n d Descr ip tor P h a r m a cop h or e
Id en tifica tion . In our previous study of anticonvul-
sants,12 we applied two variable selection QSAR algo-
rithms, genetic algorithm (GA) or simulated annealing
(SA) partial least squares (GA/SA-PLS)17,24,25 and kNN
analyses,26 to build predictive models for the FAA
An obvious parallel can be established between the
selection of characteristic pharmacophoric features re-
sponsible for specific biological activity and the selection
of a subset of chemical descriptors implicated in the
most statistically significant variable selection QSAR
models. This analogy led us earlier to define the latter
collection of descriptors as a “descriptor pharmaco-
phore”.16 We proposed that descriptor pharmacophores
can be employed in database mining and demonstrated
that variable selection QSAR models can be used for
both finding molecular structures with the same biologi-
cal activity as the probe molecules in chemical data-
bases17 or virtual chemical libraries18 and predicting the
value of the activity. In this paper, we have refined the
QSAR-based database-mining approach and employed
it for anticonvulsant lead compound discovery.
In a previous publication,12 we described the develop-
ment of multiple kNN QSAR models for 48 FAA anti-
convulsants; the models were extensively validated
using several criteria of robustness and accuracy.19,20
Herein, we describe the application of these models to
mining two publicly available chemical databases, i.e.,
the Maybridge21 and National Cancer Institute (NCI)22
collections containing ca. 250 000 compounds combined.
Our objective was to identify new lead compounds and
experimentally confirm their predicted anticonvulsant
activity. The 10 best models with the highest validated
predictive power from our previous study12 have been
used for database mining. Twenty-two compounds,
which were predicted to have high activities by all or
the majority of all models, were chosen as consensus
hits. Nine compounds were synthesized and sent to the
Anticonvulsant Screening Program (ASP) at the Na-
tional Institutes of Health (NIH) to test for the anti-
convulsant activity; four of these compounds were

2358 J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9
Shen et al.
dataset. Both of these methods have been shown to
produce predictive models for other datasets as well24-26
and can be easily automatized and adapted to the task
of database searching, or virtual screening.17,18 Both
methods typically employ multiple descriptors derived
from 2D molecular topology (e.g., molecular connectivity
indices or atom pair descriptors), which eliminates the
conformational and alignment ambiguities inherent to
most 3D QSAR methods. Stochastic optimization algo-
rithms such as GA or SA are used to build robust QSAR
models characterized by high values of cross-validated
R2 (q2). All models derived in our previous studies12 were
subjected to extensive validation using several impor-
tant criteria for predictive QSAR modeling. These
criteria are discussed extensively in our recent paper;19
they include the randomization of the target property
(Y-randomization), independent assessment of the train-
ing set model predictive power using external test sets,
and the establishment of the model applicability do-
main. All specific details of the calculations can be found
in an earlier publication.12 For this paper, we have used
only kNN models developed with the MolConnZ descrip-
tors.
lated. The training set models are developed in kNN
QSAR modeling by interpolating activities of the k
nearest neighbors of each compound to predict the
activity of that compound.26 This procedure allowed us
to derive a special applicability domain (i.e., similarity
threshold) specific to each particular kNN QSAR model
to avoid making predictions for compounds that differ
substantially from the training set molecules.19 In our
studies, this threshold, DT is calculated from the train-
ing set models as follows
DT ) yj + Zσ
(2)
Here, yj is the average Euclidean distance between each
compound and its k nearest neighbors (where k is the
parameter optimized in the course of QSAR modeling),
σ is the standard deviation of these Euclidean distances,
and Z is an arbitrary parameter to control the signifi-
cance level. We set the default value of this parameter
Z at 0.5, which formally places the allowed distance
threshold at one-half of the standard deviation (assum-
ing a Boltzman distribution of distances between k
nearest neighbor compounds in the training set). Thus,
if the distance of the external compound from at least
one of its nearest neighbors in the training set exceeds
this threshold, the prediction is considered unreliable.
Da ta ba ses. Two publicly available databases have
been explored: the Maybridge21 and the NCI22 data-
bases, including 55 273 and 237 771 chemical struc-
tures, respectively. The NCI database was curated; thus,
metal-containing compounds as well as compounds with
incomplete chemical structures (which could not be
processed by MolConnZ) were excluded, leaving 194 736
compounds for database mining. The total collection of
compounds subjected to database mining included
250 009 molecules.
As discussed in the Introduction, we have defined
descriptor pharmacophores as a subset of descriptor
types implicated in successful variable selection QSAR
models.16,17 The similarity searches between known
active FAA molecules and compounds in chemical
databases were conducted using descriptor pharma-
cophores found in those variable selection QSAR models
that satisfied our rigorous validation criteria.
Selection of Sim ila r ity P r obes. Forty-four com-
pounds out of 48 FAAs with the anticonvulsant activity
with ED50 less than 100 mg/kg, which is considered
promising by the NIH standard, were selected as the
similarity probes for database mining. The structures
and activity of these 44 compounds were reported in
Table 1 of our previous paper;12 this table is also
included in the Supporting Information for reference.
Molecu la r Descr ip tor s a n d Sim ila r ity Ca lcu la -
tion s. Molecular topological descriptors were calculated
with the MolConnZ program23 for both probes and data-
base molecules. Since absolute scales for the various
MolConnZ descriptors can differ by orders of magnitude,
range scaling was used to avoid giving descriptors with
significantly higher ranges a disproportionably higher
weight upon the molecular similarity calculations.
Euclidean distance was used as the measure of similar-
ity in the multidimensional descriptor space. The latter
distance dij between any two compounds i and j in
N-dimensional descriptor space was calculated as
Resu lts a n d Discu ssion
QSAR Mod elin g of F AA Com p ou n d s. The devel-
opment of rigorously validated QSAR models for 48
chemically diverse FAA anticonvulsants was reported
earlier.12 Model building included multiple divisions of
the original dataset into a training and test set (see ref
27 for details). Multiple QSAR models (760 total) were
generated independently for all training sets and vali-
dated using the test sets. Generally, we accept models
with leave-one-out cross-validation (q2) values for the
training set greater than 0.5 and R2 values for predicted
vs actual activities of the test set compounds greater
than 0.6 as the most significant criteria.19 Only the 10
best models obtained from multiple kNN-QSAR analy-
ses using MolConnZ descriptors in our previous study12
were used for database mining. The training and test
set sizes were 43 and 5, 39 and 9, and 38 and 10
compounds, respectively, and the optimal number of
descriptors was in the range between 12 and 20 (see
Table 6 in our previous paper;12 this table is also
included as Table 2 in the Supporting Information for
reference).
N
dij )
(X - Xjn)2
(1)
∑
in
x
n)1
where Xin and Xjn are the values of nth descriptor for
compounds i and j, respectively, and the summation is
over all descriptors. Compounds with the smallest
distance (highest similarity) from the active probe were
considered as hits and subjected to the prediction of
their activity based on QSAR models.
Da ta ba se Min in g. As illustrated in Figure 2, vali-
dated QSAR models and descriptor pharmacophores
derived from these models were used to mine chemical
databases for novel lead anticonvulsant agents.
All compounds in the databases within the chosen
similarity cutoff value (0.5 Euclidean distance units) of
Ap p lica bility Dom a in of QSAR Mod els. Formally,
a QSAR model can predict the target property for any
compound for which chemical descriptors can be calcu-

Application of QSAR Models to Database Mining
J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9 2359
F igu r e 2. Consensus database-mining workflow based on QSAR modeling.
any of the probes were selected, resulting in 4334 initial
mining hits total. Because of the differences in the
descriptor pharmacophores, the resulting 10 hit lists
were not identical. Thus, the initial list was additionally
refined by selecting consensus hits, i.e., molecules found
in all 10 individual hit lists, reducing the initial list to
50 compounds only.
pendent observations by Geurts et al.,29 who showed
that compounds incorporating the Cbz moiety indeed
displayed anticonvulsant activity. Finally, we purpo-
sively opted not to prepare compounds that contained
the PhCH2N(H) moiety (compounds 8 and 12 in Table
1) since it has been already shown in FAAs (e.g., 2) to
be beneficial for anticonvulsant activity.6-8 Several
additional compounds (25-29), which were close ana-
logues of computational hits 3 and 5, were designed de
novo. Where applicable, we prepared the racemic com-
pound mixtures, since we did not know if either isomer
would exhibit preferential activity. In total, seven
compounds were synthesized and submitted to the
NIH’s ASP for evaluation using the MES test (a
standard test for the anticonvulsant activity, which was
used for the training set compounds as well). The results
of testing available at this time are shown in Table 2.
The anticonvulsant activity of these 50 consensus hits
was predicted using 10 QSAR models with applicability
domains specific to each model. This procedure resulted
in the final prediction of biological activity for only 22
compounds found within all 10 individual applicability
domains. Table 1 shows the structures and averaged
predicted activity of these 22 compounds. Four of these
molecules were chosen for synthesis and experimental
testing directly, and five compounds were designed as
analogues of two of these hits (see below). Table 2 lists
the chemical structures and predicted anticonvulsant
activities for these nine compounds. Additional details
of mining the individual databases and selecting hits
for the experimental validation are described below.
Hits fr om th e NCI Da ta ba se a n d Th eir Exp er i-
m en ta l Va lid a tion . Application of the 10 best QSAR
models to mining the NCI database resulted in 432, 175,
309, 766, 410, 202, 341, 443, 219, and 174 hits, respec-
tively. Only 27 compounds were identified in all 10
searches, and their activities were predicted by all 10
QSAR models. Eleven compounds with consistently high
predicted MES values (ED50 less than 100 mg/kg) were
selected as consensus hits.
The experimentally confirmed anticonvulsant activi-
ties for 3, 5, and 25-29 demonstrate the ability of our
database-mining approach to identify molecules with
chemical substructures that differed from those ob-
served in the FAA training set. All seven compounds
were tested in mice (ip) and rats (po). Each compound
was tested at least at three doses, 300, 100, and 30 mg/
kg in mice, and one dose, 30 mg/kg, in rats. The
biological testing results (Table 2) indicate that upon
biological screening in mice, five out of seven tested
compounds showed anticonvulsant activity with MES
ED50 less than 100 mg/kg, which is considered promis-
ing by the NIH standard. Interestingly, all seven
compounds, when evaluated in the MES test in rats,
were active (MES ED50 < 52 mg/kg) (experimental data
on rats for the entire training set were not available,
and therefore, no QSAR models were built).
The two compounds initially chosen for experimental
evaluation featured a fully substituted terminal amide
group, a functional unit not anticipated to provide an
active compound based on prior experience (cf. com-
pounds 3 and 5 in Table 1).28 For example, we showed
that the amide 30 displayed excellent activity in MES-
Hits fr om th e Ma ybr id ge Da ta ba se a n d Th eir
Exp er im en ta l Va lid a tion . Mining of the Maybridge
database using the same 10 best QSAR models that
were employed in mining the NCI database yielded 156,
89, 127, 104, 93, 170, 138, 182, 113, and 141 hits,
respectively. Only 23 compounds were identified in all
10 searches, and their activities were predicted by all
10 QSAR models using the respective applicability
domain criteria. Eleven compounds with consistently
high predicted MES values (ED50 less than 100 mg/kg)
were selected as consensus hits.
induced seizure test in mice (MES ED50 ) 8.3 mg/kg),
while the corresponding fully N-substituted amide 31
exhibited little activity (MES ED50 > 100, <300 mg/kg).
Another interesting feature of these compounds was the
presence of the carbobenzyloxy (Cbz) group, which was
absent in any of the training set compounds.12 The
importance of this group was corroborated by inde-
Two out of 11 consensus hits were synthesized and
tested by the NIH, and both compounds were found to
be active. Again, we elected not to prepare compounds
that contained either PhCH2N(H) or a structurally
related moiety (Table 1: 16, 17, 19, 23, 24), because of

2360 J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9
Shen et al.
Ta ble 1. Consensus Hits of Database Mining Found within All Individual Applicability Domains with High Predicted Average
Activity Based on the 10 Best kNN-QSAR Models
a
b
Predicted mice MES (maximal electroshock seizure test) ED50 value in mg/kg. Database mining hits that were synthesized and sent
to NIH for experimental test.
their apparent structural similarity to known FAA
anticonvulsants such as 2. Rather, we chose 14 and 18
as our test compounds, since the earlier SAR study on
2 showed that deletion of the methylene group within
the PhCH2N(H) moiety resulting in a corresponding
FAA arylamide led to a loss in anticonvulsant activity.30
One of the compounds, 14, exhibited significant anti-
convulsant activity, providing MES ED50 values be-
tween 30 and 100 mg/kg (in mice), and another, 18, was
found to be a highly potent anticonvulsant agent with
a MES ED50 of 18 mg/kg in mice (ip) that was compa-
rable to the activity of phenobarbital (MES ED50 ) 22
mg/kg).31 Both compounds were found to be very active
antiarrhythmic.33 In vitro metabolism of lidocaine to 18
is mediated by CYP3A4.34 Our results indicate that
besides known anesthetic activity, 18 is likely to have
an anticonvulsant activity as well.
Con clu sion s
Our computational strategies afforded the identifica-
tion of only 22 potentially active anticonvulsant agents
that were selected rationally after computational screen-
ing of more than 250 000 compounds in two chemical
databases. Nine compounds, including four computa-
tional hits and five molecules derived from two ad-
ditional hits by minor chemical modifications, were
selected for synthesis and evaluation. Most of these
compounds (3, 5, 14, 18, 27-29) contained structural
units (e.g., fully N-substituted amide, N-arylamide) that
were previously shown to reduce activity in the MES-
induced seizure test.28,30 Seven of the nine compounds
in rats (po) as well, and the potency of 18 (MES ED50
)
11 mg/kg) exceeded that of phenytoin (MES ED50 ) 30
mg/kg)11 and was comparable to that of phenobarbital
(MES ED50 ) 9.1 mg/kg)31 (Table 2). An interesting find-
ing is that 18 is a metabolite of lidocaine (lignocaine),
which is widely used as a local anesthetic32 and cardiac

Application of QSAR Models to Database Mining
J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9 2361
Ta ble 2. Results of Anticonvulsant Activity Testing from the Anticonvulsant Screening Project at the National Institutes of Healthg
a
The compounds were administered intraperitoneally. ED50 and TD50 values are in mg/kg. Numbers in parentheses are 95% confidence
intervals. The dose effect data was obtained at the “time of peak effect” (indicated in hours in the brackets). Data for most promising
b
d
compounds are shown in bold. The compounds were administered orally. c Melting points (°C) are uncorrected. MES ) maximal
electroshock seizure test. e Tox ) neurologic toxicity determined from rotorod test. f Hit compounds selected from NCI and Maybridge
databases.
were confirmed to be active by experimental studies,
indicating an exceptionally high experimental hit rate.
Two compounds, 3 and 26, predicted to have significant
MES-induced anticonvulsant activity in mice (ip), were
found to be only moderately active (ED50 > 100, <300
mg/kg). At this stage, we cannot provide an explanation
for this finding but do note that both compounds when
tested in rats (po) displayed appreciable anticonvulsant
activity (Table 2).
The experimental findings reported in this paper are
exciting in light of our long-term studies of FAA anti-
convulsants. In 1985, the Kohn laboratory discovered
the lead compound N-acetyl-D,L-alanine benzylamide (2,
R2 ) CH3)35 with a MES ED50 of 77 mg/kg in mice (ip),
which is comparable with the activity of 28 (Table 2).
The structure of this compound was later optimized in
the Kohn laboratory using synthetic lead optimization
strategies providing (R)-N-benzyl-2-acetamido-3-meth-
oxypropionamide (2, R2 ) CH2OCH3).8 This compound
exhibited a MES ED50 of 4.5 mg/kg in mice (ip) and is
currently under phase II clinical studies for the treat-
ment of epilepsy and neuropathic pain.36 The work is
now in progress to utilize comparable synthetic strate-
gies in tandem with computational methods to optimize
the structure of compound 28. A similar approach is
being employed for 18 as well.
repertoire of anticonvulsant agents. Our search meth-
odologies were based on chemical similarity estimated
by Euclidean distance in the descriptor pharmacophore
subspace of the whole descriptor space as well as on
quantitative predictions from existing QSAR models.
Due to the nature of the descriptors (i.e., connectivity
indices, which are derived from chemical structures but
cannot be easily used in the inverse design experiments
to design structures), such searches are more likely to
result in the identification of novel compounds than
traditional fragment-based search methodologies. The
results reported in this paper demonstrate that our
chemical similarity searches indeed afford the identi-
fication of potent anticonvulsants with chemical sub-
structures that differ from those observed in the train-
ing set, such as Cbz, the fully N-substituted amide
moiety, and the N-arylamide group.
The success of the computational and experimental
studies reported in this paper lends firm support for all
aspects of our computational drug discovery strategies.
These strategies have been refined over several years
of QSAR modeling research, including methods for
model development,26 training and test set selection,27
validation,19,20 and database mining.16,17 The integration
of these techniques and approaches into a comprehen-
sive workflow reported for the first time in this paper
(Figures 1 and 2) afforded a very high experimental hit
rate of a very small collection of computational hits
selected rationally from a very large compound collec-
tion. We believe that this enormous reduction of the
experimental drug discovery search space is attributed
to several important aspects of the workflow and its
individual components. First, we place particular em-
The traditional approaches to database mining are
based on chemical fragment or subfragment similarity
searches. While this has been an efficient approach that
has enjoyed some successes,16-18 it limits the chemical
diversity of selected compounds to those similar to
existing ligands. One of the major goals of our database-
mining experiments was to diversify the chemical

2362 J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9
Shen et al.
23.90 mmol), 4-methylmorpholine (2.9 mL, 26.29 mmol),
isobutyl chloroformate (3.4 mL, 26.29 mmol), dimethylamine
(2.0 M in THF, 13.2 mL, 26.40 mmol), and THF (150 mL) gave
crude 27. The product was purified by column chromatography
(SiO2; 1:49, MeOH:CHCl3) to obtain 4.24 g (75%) of pure 27
as a pale green solid: mp 57-58 °C; Rf 0.36 (1:19, MeOH:
CHCl3); IR (KBr) 3261, 3052, 2929, 1725, 1643, 1546, 1407,
1248, 1170, 1045, 990, 741, 647 cm-1; 1H NMR (CDCl3) δ 2.96
(s, CH3), 2.98 (s, CH3), 4.01 (d, J ) 4.5 Hz, CH2C(O)), 5.13 (s,
OCH2), 5.93 (br s, NH), 7.30-7.38 (m, PhH); 13C NMR (CDCl3)
35.4 (CH3), 35.6 (CH3), 42.5 (CH2C(O)), 66.6 (OCH2), 127.8 (C4′),
127.9 (2C2′ or 2C3′), 128.3 (2C2′ or 2C3′), 136.4 (C1′), 156.1
(OC(O)NH), 167.7 (C(O)NH) ppm; MS (+CI) (rel intensity) 238
(10), 237 (M+ + 1, 75), 194 (14), 193 (100), 154 (31), 129 (67);
Mr (+CI) 237.123 88 [M+ + 1] (calcd for C12H17N2O3 237.123 92).
Anal. (C12H16N2O3) C, H, N.
phasis on the rigorous validation of QSAR models as
well as conservative extrapolation limited to the ap-
plicability domain, which allows us to achieve the
highest possible accuracy in predicted biological activity
of compounds external to the training set. Second, the
similarity metric used in database mining is defined by
the Euclidean distances in the descriptor pharmaco-
phore space as opposed to the full descriptor space.
Third, we select only consensus hits obtained with
multiple validated QSAR models as opposed to the
predictions based on a single best model.
Future studies should provide additional enhance-
ments of the workflow employed in this paper. Thus,
we are currently enlarging the selection of data model-
ing algorithms and descriptor sets in the context of
combinatorial QSAR modeling that we began to explore
recently.37 We suggest that the combined predictive
QSAR modeling and database-mining approach devel-
oped in this paper can be applied to a variety of
experimental datasets and exploited as a general tool
for the design and discovery of novel, biologically active
compounds.
Syn th esis of N-(Ben zyloxyca r bon yl)glycin e-N,N-d i-
eth yla m id e (5). N-(Benzyloxycarbonyl)glycine (5.00 g, 23.90
mmol), 4-methylmorpholine (2.9 mL, 26.29 mmol), isobutyl
chloroformate (3.4 mL, 26.29 mmol), diethylamine (3.0 mL,
26.29 mmol), and THF (150 mL) gave crude 5. The product
was purified by column chromatography (SiO2; 3:2, EtOAc:
hexanes) to obtain 5.68 g (90%) of pure 5 as a translucent
solid: mp 40-41 °C; Rf 0.31 (3:2, EtOAc:hexanes); IR (KBr)
3268, 3066, 2986, 2940, 1726, 1642, 1550, 1454, 1255, 1167,
1049, 1006, 743, 653 cm-1; 1H NMR (CDCl3) δ 1.13 (t, J ) 7.2
Hz, CH2CH3), 1.19 (t, J ) 7.2 Hz, CH2CH3), 3.26 (q, J ) 7.2
Hz, CH2CH3), 3.40 (q, J ) 7.2 Hz, CH2CH3), 4.02 (d, J ) 4.2
Hz, CH2C(O)), 5.13 (s, OCH2), 5.85 (br s, NH), 7.30-7.37 (m,
PhH); 13C NMR (CDCl3) 12.7 (CH2CH3), 13.7 (CH2CH3), 40.2
(CH2CH3), 40.7 (CH2CH3), 42.3 (CH2C(O)), 66.5 (OCH2), 127.7
(C4′), 127.8 (2C2′ or 2C3′), 128.2 (2C2′ or 2C3′), 136.4 (C1′), 156.0
(OC(O)NH), 166.7 (C(O)NH) ppm; MS (+CI) (rel intensity) 266
(17), 265 (M+ + 1, 100), 221 (32), 157 (16), 154 (11); Mr (+CI)
265.154 61 [M+ + 1] (calcd for C14H21N2O3 265.155 22). Anal.
(C14H20N2O3) C, H, N.
Exp er im en ta l Meth od s
Gen er a l P r oced u r e. To a cold (-78 °C) THF solution
(∼0.1-0.2 mmol N-(benzyloxycarbonyl) amino acid/mL of
solvent) containing the N-(benzyloxycarbonyl)amino acid, the
4-methylmorpholine (1.1 equiv) was slowly added. After stir-
ring (2 min), isobutyl chloroformate (1.1 equiv) was added. The
reaction was stirred (2 min) and then the amine (1.1 equiv)
was added dropwise. The reaction was stirred at -78 °C (15
min) and allowed to warm to room temperature (1 h). The
reaction mixture was filtered and the filtrate evaporated in
vacuo. The residue was purified by chromatography on SiO2
gel to obtain the desired product. Using this general procedure,
the following compounds were prepared.
Syn th esis of N-(Ben zyloxyca r bon yl)glycin e-N-m eth yl-
a m id e (25).29 N-(Benzyloxycarbonyl)glycine (5.00 g, 23.90
mmol), 4-methylmorpholine (2.9 mL, 26.29 mmol), isobutyl
chloroformate (3.4 mL, 26.29 mmol), methylamine (2.0 M in
THF, 13.2 mL, 26.40 mmol), and THF (150 mL) gave crude
25. The product was purified by column chromatography (SiO2;
1:19, MeOH:CHCl3) to obtain 5.03 g (95%) of pure 25 as a white
solid: mp 105-106 °C (lit.29 mp 105-107 °C); Rf 0.31 (1:19,
MeOH:CHCl3); 1H NMR (CDCl3) δ 2.76 (d, J ) 4.5 Hz, NCH3),
3.81 (d, J ) 5.7 Hz, CH2C(O)), 5.10 (s, OCH2), 5.86 (br t, J )
5.7 Hz, NH), 6.48 (br s, NH), 7.27-7.33 (m, PhH); 13C NMR
(CDCl3) 26.0 (NCH3), 44.5 (CH2C(O)), 67.0 (OCH2), 127.9 (C4′),
128.1 (2C2′ or 2C3′), 128.4 (2C2′ or 2C3′), 136.1 (C1′), 156.6
(OC(O)NH), 169.7 (C(O)NH) ppm.
Syn t h esis of N-(Ben zyloxyca r b on yl)glycin e-N-et h yl-
a m id e (26). N-(Benzyloxycarbonyl)glycine (5.00 g, 23.90
mmol), 4-methylmorpholine (2.9 mL, 26.29 mmol), isobutyl
chloroformate (3.4 mL, 26.29 mmol), ethylamine (2.0 M in
THF, 13.2 mL, 26.40 mmol), and THF (150 mL) gave crude
26. The product was purified by column chromatography (SiO2;
1:49, MeOH:CHCl3 to 1:19, MeOH:CHCl3) to obtain 5.07 g
(90%) of pure 26 as a white solid: mp 99-100 °C; Rf 0.41
(1:19, MeOH:CHCl3); IR (KBr) 3318, 3051, 2971, 1702, 1655,
1545, 1453, 1373, 1252, 1155, 732, 698 cm-1; 1H NMR (CDCl3)
δ 1.01 (t, J ) 7.2 Hz, CH2CH3), 3.19-3.28 (m, CH2CH3), 3.81
(d, J ) 5.7 Hz, CH2C(O)), 5.09 (s, OCH2), 6.16 (t, J ) 5.7 Hz,
NH), 6.76 (br s, NH), 7.28-7.32 (m, PhH); 13C NMR (CDCl3)
14.3 (CH2CH3), 34.1 (CH2CH3), 44.2 (CH2C(O)), 66.7 (OCH2),
127.7 (C4′), 127.9 (2C2′ or 2C3′), 128.3 (2C2′ or 2C3′), 136.0 (C1′),
156.6 (OC(O)NH), 168.9 (C(O)NH) ppm; MS (+CI) (rel inten-
sity) 238 (12), 237 (M+ + 1, 83), 194 (12), 193 (100); Mr (+CI)
237.123 88 [M+ + 1] (calcd for C12H17N2O3 237.123 92). Anal.
(C12H16N2O3) C, H, N.
Syn th esis of N-(Ben zyloxyca r bon yl)glycin e-N-m eth -
oxy-N-m eth yla m id e (28).38 N,O-Dimethylhydroxylamine was
prepared by stirring N,O-dimethylhydroxylamine hydrochlo-
ride salt (1.57 g, 16.13 mmol), K2CO3 (4.46 g, 32.2 mmol), THF
(30 mL), and H2O (3 mL) at room temperature (3 h). N-
(Benzyloxycarbonyl)glycine (2.81 g, 13.44 mmol), 4-methyl-
morpholine (1.6 mL, 14.78 mmol), isobutyl chloroformate (1.9
mL, 14.78 mmol), the previously prepared N,O-dimethyl-
hydroxylamine solution (33 mL), and THF (70 mL) gave crude
28. The product was purified by column chromatography (SiO2;
3:2, EtOAc:hexanes) and then crystallized from EtOAc (15 mL)
to obtain 2.41 g (71%) of pure 28 as a translucent solid: mp
76-77 °C (lit.38 mp 76-77 °C); Rf 0.31 (3:2, EtOAc:hexanes);
IR (KBr) 3291, 1725, 1661, 1545, 1469, 1252, 1169, 977, 741
1
cm-1; H NMR (CDCl3) δ 3.20 (s, NCH3), 3.71 (s, OCH3), 4.14
(d, J ) 4.8 Hz, CH2C(O)), 5.12 (s, OCH2), 5.61 (br s, NH), 7.30-
7.38 (m, PhH); 13C NMR (CDCl3) 32.0 (NCH3), 41.7 (CH2C(O)),
61.0 (OCH3), 66.4 (OCH2), 127.6 (C4′), 127.7 (2C2′ or 2C3′), 128.1
(2C2′ or 2C3′), 136.2 (C1′), 156.1 (OC(O)NH), 169.5 (C(O)NH)
ppm; MS (+CI) (rel intensity) 253 (M+ + 1, 62), 210 (12), 209
(100), 192 (14), 145 (37), 119 (12); Mr (+CI) 253.119 59 [M+
+
1] (calcd for C12H17N2O4 253.118 83). Anal. (C12H16N2O4) C, H,
N.
Syn th esis of (R,S)-N-(Ben zyloxyca r bon yl)a la n in e-N-
m eth oxy-N-m eth yla m id e ((R,S)-3). N,O-Dimethylhydroxyl-
amine was prepared by stirring N,O-dimethylhydroxylamine
hydrochloride salt (1.57 g, 16.13 mmol), K2CO3 (4.46 g, 32.2
mmol), THF (30 mL), and H2O (3 mL) at room temperature (3
h). (R,S)-N-(Benzyloxycarbonyl)alanine (3.00 g, 13.44 mmol),
4-methylmorpholine (1.6 mL, 14.78 mmol), isobutyl chloro-
formate (1.9 mL, 14.78 mmol), the previously prepared N,O-
dimethylhydroxylamine solution (33 mL), and THF (70 mL)
gave crude 3. The product was purified by column chroma-
tography (SiO2; 1:49, MeOH:CHCl3) to obtain 3.16 g (88%) of
pure 3 as an off-white solid: mp 63-64 °C; Rf 0.42 (1:49,
MeOH:CHCl3); IR (KBr) 3267, 1709, 1650, 1532, 1254, 1071,
981, 740 cm-1; 1H NMR (CDCl3) δ 1.32 (d, J ) 6.9 Hz, CH3CH),
3.18 (s, NCH3), 3.74 (s, OCH3), 4.73 (dq, J ) 8.1, 6.9 Hz,
Syn t h esis of N-(Ben zyloxyca r b on yl)glycin e-N,N-d i-
m et h yla m id e (27). N-(Benzyloxycarbonyl)glycine (5.00 g,

Application of QSAR Models to Database Mining
J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9 2363
CH3CH), 5.03-5.13 (m, OCH2), 5.84 (br d, J ) 8.1 Hz, NH),
7.28-7.34 (m, PhH); 13C NMR (CDCl3) 18.2 (CH3CH), 31.9
(NCH3), 46.9 (CH3CH), 61.3 (OCH3), 66.4 (OCH2), 127.7 (C4′),
127.8 (2C2′ or 2C3′), 128.2 (2C2′ or 2C3′), 136.3 (C1′), 155.5
(OC(O)NH), 173.0 (C(O)NH) ppm; MS (+CI) (rel intensity) 268
(14), 267 (M+ + 1, 100), 223 (42), 206 (38), 159 (26); Mr (+CI)
267.135 15 [M+ + 1] (calcd for C13H19N2O4 267.134 48). Anal.
(C13H18N2O4) C, H, N.
Su p p or tin g In for m a tion Ava ila ble: Tables listing the
anticonvulsant activities of the 48 FAA compounds and the
statistical characteristics of the 10 best QSAR models. This
material is available free of charge via the Internet at
http://pubs.acs.org.
Refer en ces
(1) Begley, C. E.; Annegers, J . F.; Lairson, D. R.; Reynolds, T. F.;
Hauser, W. A. Cost of epilepsy in the United States: A model
based on incidence and prognosis. Epilepsia 1994, 35, 1230-
1243.
(2) Rogawski, M. A.; Porter, R. J . Antiepileptic drugs: Pharmaco-
logical mechanisms and clinical efficacy with consideration of
promising development stage compounds. Pharmacol. Rev. 1997,
42, 223-286.
(3) Unverferth, K.; Rundfeldt, C. Antiepileptic drugs. Pharmaceu-
ticals 2000, 2, 469-488.
(4) Mattson, R. H.; Cramer, J . A.; Collins, J . F.; Smith, D. B.
Comparison of carbamazepine, phenobarbital, phenytoin, and
primidone in partial and secondarily genealized tonic-clonic
seizures. New Eng. J . Med. 1985, 313, 145-151.
(5) Pellock, J . M.; Willmore, L. J . A rational guide to monitoring in
patients receiving anticonvulsants. Neurology 1991, 41, 961-
964.
(6) Kohn, H.; Sawhney, K. N.; LeGall, P.; Conley, J . D.; Robertson,
D. W.; Leander, J . D. Preparation and anticonvulsant activity
of a series of functionalized R-aromatic and R-heteroaromatic
amino acids. J . Med. Chem. 1990, 33, 919-926.
(7) Kohn, H.; Sawhney, K. H.; Bardel, P.; Robertson, D. W.; Leander,
J . D. Synthesis and anticonvulsant activities of R-hetero-
aromatic-R-acetamido-N-benzylacetamide derivatives. J . Med.
Chem. 1993, 36, 3350-3360.
Syn th esis of (R,S)-N-(Ben zyloxyca r bon yl)a la n in e-N,N-
d ieth yla m id e ((R,S)-29). (R,S)-N-(Benzyloxycarbonyl)alanine
(2.80 g, 12.54 mmol), 4-methylmorpholine (1.5 mL, 13.79
mmol), isobutyl chloroformate (1.8 mL, 13.79 mmol), diethyl-
amine (1.4 mL, 13.79 mmol), and THF (100 mL) gave crude
29. The product was purified by column chromatography (SiO2;
1:49, MeOH:CHCl3) to obtain 2.33 g (67%) of pure 29 as a clear
oil: Rf 0.26 (1:49, MeOH:CHCl3); IR (KBr) 3410, 3282 (br),
2978, 1715, 1639, 1528, 1455, 1247, 1061, 745 cm-1; 1H NMR
(CDCl3) δ 1.12 (t, J ) 7.2 Hz, CH2CH3), 1.24 (t, J ) 7.2 Hz,
CH2CH3), 1.33 (d, J ) 6.9 Hz, CH3CH), 3.20-3.56 (m, 2
CH2CH3), 4.62 (dq, J ) 7.8, 6.9 Hz, CH3CH), 5.05-5.14 (m,
OCH2), 5.78 (br d, J ) 7.8 Hz, NH), 7.30-7.36 (m, PhH); 13C
NMR (CDCl3) 12.7 (CH2CH3), 14.3 (CH2CH3), 19.4 (CH3CH),
40.1 (CH2CH3), 41.5 (CH2CH3), 46.6 (CH3CH), 66.4 (OCH2),
127.7 (C4′), 127.8 (2C2′ or 2C3′), 128.3 (2C2′ or 2C3′), 136.4 (C1′),
155.4 (OC(O)NH), 171.5 (C(O)NH) ppm; MS (+CI) (rel inten-
sity) 280 (18), 279 (M+ + 1, 100), 117 (17); Mr (+CI) 279.170 73
[M+ + 1] (calcd for C15H23N2O3 279.170 87). Anal. (C15H22N2O3·
0.3H2O) C, H, N.
Syn t h esis of N-(E t h yl)glycin e-N-2,6-d im et h yla n ilid e
(18).39 2-Chloro-N-(2,6-dimethylphenyl)acetamide40 (3.00 g,
15.19 mmol) was added to a THF solution containing EtNH2
(2.0 M in THF, 20 mL, 40 mmol), and then the solution was
stirred at 50 °C (18 h). The solvent was evaporated and the
residue was purified by crystallization (hexanes) to obtain 2.20
g (70%) of pure 18 as a white solid: mp 48-49 °C (lit.39 mp
(8) Choi, D.; Stables, J . P.; Kohn, H. Synthesis and anticonvulsant
activities of N-benzyl-2-acetamidopropionamide derivatives. J .
Med. Chem. 1996, 39, 1907-1916.
(9) Stables, J . P.; Kupferberg, H. J . In Molecular and Cellular
Targets for Antiepileptic Drugs; Avanzini, G., Tanganelli, P.,
Avoli, M., Eds.; J ohn Libbey: London, 1997; pp 191-198.
(10) Swinyard, E. A.; Woodhead, J . H.; White, H. S.; Franklin, M. R.
In Antiepileptic Drugs, 3rd Edn.; Levy, R. H., Driefuss, F. E.,
Mattson, R. H., Meldrum, B. S., Perry, J . K., Eds.; Raven Press:
New York, 1989; pp 85-102.
1
49-50 °C); H NMR (CDCl3) δ 1.41 (t, J ) 7.1 Hz, CH2CH3),
1.68 (br s, NH), 2.13 (s, 2CH3), 2.73 (q, J ) 7.1 Hz, CH2CH3),
3.40 (s, CH2), 7.06 (s, 3 PhH), 8.84 (br s, NH); 13C NMR (CDCl3)
15.3 (CH2CH3), 18.3 (2CH3), 44.6 (CH2CH3 or CH2), 52.4
(CH2CH3 or CH3), 126.8, 128.0, 133.8, 134.9 (Ph), 170.1
(11) Levy, R. H.; Mattson, R.; Meldrum, B. Antiepileptic Drugs, 4th
Edn.; Raven Press: New York, 1995; Chapter 6.
(12) Shen, M.; LeTiran, A.; Xiao, Y.; Golbraikh, A.; Kohn, H.; Tropsha,
A. Quantitative structure-activity relationship analysis of func-
tionalized amino acid anticonvulsant agents using k nearest
neighbor and simulated annealing PLS methods. J . Med. Chem.
2002, 45, 2811-2823.
(13) Gu¨ner, O. F. History and evolution of the pharmacophore concept
in computer-aided drug design. Curr. Top. Med. Chem. 2002, 2,
1321-1332.
(14) Kurogi, Y.; Gu¨ner, O. F. Pharmacophore modeling and three-
dimensional database searching for drug design using Catalyst.
Curr. Med. Chem. 2001, 8, 1035-1055.
(15) Martin, Y. C. 3D database searching in drug design. J . Med.
Chem. 1992, 35, 2145-2154.
(16) Tropsha, A.; Zheng, W. Identification of the descriptor pharma-
cophores using variable selection QSAR: Applications to data-
base mining. Curr. Pharm. Design 2001, 7, 599-612.
(17) 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. In Rational Drug Design: Novel Methodology and
Practical Applications; Parrill, A. L., Reddy, M. R., Eds.; ACS
Symposium Series 719; American Chemical Society: Washing-
ton, DC, 1999; pp 198-211.
(C(O)NH) ppm; MS (+CI) (rel intensity) 208 (11), 207 (M+
+
1, 100); Mr (+CI) 207.149 83 [M+ + 1] (calcd for C12H19N2O
207.149 74). Anal. (C12H18N2O) C, H, N.
Syn t h esis of N,N-Dim et h yl-N′-(3-m et h ylp h en yl)u r ea
(14).41 A THF (10 mL) solution of m-tosyl isocyanate (2.0 mL,
15.52 mmol) and dimethylamine (2.0 M in THF, 7.8 mL, 15.6
mmol) was stirred at room temperature (30 min). The solvent
was evaporated and the solid residue was purified by crystal-
lization (EtOH) to obtain 2.40 g (89%) of pure 14 as a white
1
crystalline solid: mp 128-129 °C (lit.41 mp 129-131 °C); H
NMR (CDCl3) δ 2.32 (s, PhCH3), 3.01 (s, NCH3(CH3)), 3.02 (s,
NCH3(CH3)), 6.31 (br s, NH), 6.83-6.86 (m, 1 H, PhH), 7.11-
7.19 (m, 2 H, PhH), 7.26-7.27 (m, 1 H, PhH); 13C NMR (CDCl3)
21.4 (PhCH3), 36.4 (N(CH3)2), 116.8, 120.5, 123.7, 128.6, 138.6,
139.1 (Ph), 155.7 (C(O)NH2) ppm; MS (+CI) (rel intensity) 179
(M+ + 1, 100); Mr (+CI) 179.117 94 [M+ + 1] (calcd for
C10H15N2O 179.118 44). Anal. (C10H14N2O) C, H, N.
P h a r m a cology. Procedures identical to those described in
ref 12 were employed in this study.
(18) Tropsha, A. Rational combinatorial library design and database
mining using inverse QSAR approach. In Combinatorial Library
Design and Evaluation for Drug Discovery: Principles, Methods,
Software Tools and Applications. Ghose, A. K., Viswanadhan,
V. N., Eds.; Marcel Dekker, Inc., New York, 2001; Chapter 12,
pp 363-378.
Ack n ow led gm en t. We dedicate this paper to the
memory of Dr. David W. Robertson, who contributed to
the development of functionalized amino acid anticon-
vulsants while at the Lilly Research Laboratories
(Indianapolis, IN). The authors thank the Anticon-
vulsant Screening Project at the National Institutes of
Health, for kindly performing the pharmacological stud-
ies via the ASP’s contract site at the University of Utah
with Drs. H. Wolf, S. White, and K. Wilcox. This
research was supported in part by the NIH research
grant GM66940 awarded to A.T. The authors also
acknowledge Tripos, Inc for the software grant.
(19) Tropsha, A.; Gramatica, P.; Gombar, V. K. The importance of
being earnest: Validation is the absolute essential for successful
application and interpretation of QSPR models. Quant. Struct.
Act. Relat. Comb. Sci. 2003, 22, 69-77.
(20) Golbraikh, A.; Tropsha, A. Beware of q2! J . Mol. Graph. Model.
2002, 20, 269-276.
(21) http://www.maybridge.com.
(22) http://dtp.nci.nih.gov/docs/3d_database/structural_information/
smiles_strings.html.
(23) MolConn-Z, version 3.5; Hall Associates Consulting: Quincy,
MA.

2364 J ournal of Medicinal Chemistry, 2004, Vol. 47, No. 9
Shen et al.
(24) Hoffman, B. T.; Cho, S. J .; Zheng, W.; Wyrick, S.; Nichols, D.
E.; Mailman, R. B.; Tropsha, A. QSAR modeling of dopamine
D1 agonists using comparative molecular field analysis, genetic
algorithms-partial least squares, and K nearest neighbor meth-
ods. J . Med. Chem. 1999, 42, 3217-3226.
(25) Cho, S. J .; Zheng, W.; Tropsha, A. Rational combinatorial library
design. 2. Rational design of targeted combinatorial peptide
libraries using chemical similarity probe and the inverse QSAR
approaches. J . Chem. Inf. Comput. Sci. 1998, 38, 259-268.
(26) Zheng, W.; Tropsha, A. A novel variable selection QSAR ap-
proach based on the K-nearest neighbor principle. J . Chem. Inf.
Comput. Sci. 2000, 40, 185-194.
(27) Golbraikh, A.; Tropsha, A. Predictive QSAR modeling based on
rational division of experimental datasets into diverse training
and test sets. J . Comput. Aid. Mol. Des. 2002, 16, 357-369.
(28) LeTiran, A.; Stables, J . P.; Kohn, H. Functionalized amino acid
anticonvulsants: Synthesis and pharmacological evaluation of
conformationally restricted analogues. Bioorg. Med. Chem. 2001,
9, 2693-2708.
(29) Geurts, M.; Poupaert, J . H.; Scriba, G. K. E.; Lambert, D. M.
N-(Benzyloxycarbonyl)glycine Esters and Amides as New Anti-
convulsants. J . Med. Chem. 1998, 41, 24-30.
of lignocaine and coumarin in patients with liver and kidney
diseases and epileptic patients. Br. J . Clin. Pharmacol. 1995,
39, 71-76.
(35) Cortes, S.; Liao, Z-K.; Watson, D.; Kohn, H. Effect of structural
modification of the hydantoin ring on anticonvulsant activity.
J . Med. Chem. 1985, 28, 601-606.
(36) Bialer, M.; J ohannessen, S.; Kupferberg, H.; Levy, R.; Loiseau,
P.; Perucca, E. Progress report on new antiepileptic drugs: A
summary of the Fifth Eilat Conference (EILAT V). Epilepsy Res.
2001, 43, 11-58.
(37) Kovatcheva, A.; Golbraikh, A.; Oloff, S.; Xiao, Y.; Zheng, W.;
Wolschann, P.; Buchbauer, G.; Tropsha, A. Combinatorial QSAR
of ambergris fragrance compounds. J . Chem. Inf. Comput. Sci.
2004. (jn press).
(38) Sawamura, M.; Nakayama, Y.; Kato, T.; Ito, Y. Gold(I)-catalyzed
asymmetric aldol reaction of N-methoxy-N-methyl-R-isocyano-
acetamide (R-isocyano weinreb amide). An efficient synthesis of
optically active â-hydroxy R-amino aldehydes and ketones. J .
Org. Chem. 1995, 60, 1727-1732.
(39) Nelson, S. D.; Breck, G. D.; Trager, W. F. In vivo metabolite
condensations. Formation of N1-ethyl-2-methyl-N3-(2,6-dimethyl-
phenyl)-4-imidazolidinone from the reaction of a metabolite of
alcohol with a metabolite of lidocaine. J . Med. Chem. 1973, 16,
1106-1112.
(40) Lazar, J .; Bernath, G. Saturated heterocycles, Part 172. Syn-
thesis of 2,6-disubstituted-5,6,7,8-tetrahydropyrido[4,3-d]pyri-
midine Derivatives. J . Heterocycl. Chem. 1990, 27, 1885-1892.
(41) Martin, H.; J aniak, S.; Toepfl, W.; Duerr, D.; Rohr, O.; Pissiotas,
G. Urea derivatives as selective herbicides. CIBA, DE 1905598,
1969.
(30) Robertson, D. W.; Leander, J . D. Private communication.
(31) Krall, R. L.; Perry, J . K.; White, B. G.; Kupferberg, H. J .;
Swinyard, E. A. Antiepileptic drug development: II Anticonvul-
sant drug screening. Epilepsia 1978, 19, 409-428.
(32) Tucher, G. T. Pharmcokinetics of local anesthetics. Brit. J .
Anaesth. 1986, 58, 717-731.
(33) Burney, R. G.; Difazio, C. A.; Peach, M. J .; Petrie, K. A.; Silvester,
M. J . Antiarrhythmic effects of lidocaine metabolites. Am. Heart
J . 1974, 88, 765-769.
(34) Sotaniemi, E. A.; Rautio, A.; Backstrom, M.; Arvela, P.; Pelkonen,
O. CYP3A4 and CYP2A6 activities marked by the metabolism
J M030584Q