QSAR analysis of the inhibition of recombinant CYP 3A4 activity by structurally diverse compounds using a genetic algorithm-combined partial least squares method

Article abstract of DOI:10.1023/A:1025702009611

Purpose. To develop a quantitative structure/activity relationship (QSAR) model for predicting drug-CYP 3A4 interactions. Method. The inhibitory effect of 53 structurally diverse drugs on the metabolism of 7-benzyloxy-4-trifluoromethyl coumarin (BFC) by recombinant CYP 3A4 was evaluated using a rapid microtiter plate assay. For each drug, a total of 220 two-dimensional topological indices were calculated using Molconn-Z software. Using a genetic algorithm-based partial least squares (GA-PLS) method, the desired descriptors were automatically selected to maximize the predictability of the IC₅₀ values. Results. The IC₅₀ values of the drugs tested ranged from 9 nM to 2 mM. Based on the GA-PLS method, five principal components derived from 20 Molconn-Z descriptors were found to be effective for QSAR modeling. Interestingly, these descriptors suggested that the molecular size would be an important factor in determining drug-CYP 3A4 interactions). In the leave-one-out prediction, the r_pred and the standard error of prediction (s) were 0.754 and 0.787, respectively. Even in an external validation, the predictions were in good agreement with experimental values (r_pred = 0.744, s = 0.769, n = 9). Conclusions. The proposed model, in which two-dimensional topological descriptors were used as molecular descriptors, was able to predict drug-CYP 3A4 interactions with reasonable accuracy.

Full text of DOI:10.1023/A:1025702009611

Pharmaceutical Research, Vol. 20, No. 9, September 2003 (© 2003)
Research Paper
a reduction in the body clearance of the drugs and a rapid
and unexpected rise in their blood concentrations. Early
identification of potential CYP 3A4 inhibitors is therefore
needed to minimize the risk of clinically relevant inter-
actions.
In addition to in vitro high-throughput screening tech-
niques (4,5), in silico prediction of xenobiotic metabolism is
now attracting increasing attention (6,7). The advantage of in
silico technologies is that the properties of molecules can be
assessed from a knowledge of their chemical structures alone,
either two- or three-dimensionally. In silico filtering is ex-
pected to help identify and screen out compounds that are
unlikely to become useful drugs, thereby maximizing the out-
put of the drug discovery process.
QSAR Analysis of the Inhibition of
Recombinant CYP 3A4 Activity
by Structurally Diverse
Compounds Using a Genetic
Algorithm-Combined Partial Least
Squares Method
Suchada Wanchana,1 Fumiyoshi Yamashita,1 and
Mitsuru Hashida1,2
In parallel with the application of homology models,
where the active sites of human CYPs are predicted from
crystallized structures of bacterial soluble CYP enzymes (8,9),
three-dimensional quantitative structure/activity relationship
(3D-QSAR) methods are being used for modeling the CYP
enzymes (7,10–12). 3D-QSAR methods allow us to interpret
and understand enzyme active sites and receptors through the
superimposition of a set of different compounds; in other
words, these methods can provide graphic representations of
the binding sites even when no crystal structure is available.
In the case of CYP 3A4, Ekins et al. (10) built a 3D-QSAR
model using the program Catalyst௡ (Molecular Simulations,
San Diego, CA) from the Michaelis-Menten constants of 38
structurally diverse substrates. However, a 3D-QSAR model
assumes that the binding modes of the compounds are the
same. Because of the relatively large binding pocket of CYP
3A4 (9), it is clear that substrates can have a relatively large
conformational degree of freedom within the active site. The
binding pocket can also accept two molecules simultaneously
(13). Such a variety of binding modes for CYP 3A4 substrates
and inhibitors appears to limit the predictive power of a 3D-
QSAR model.
Received February 18, 2003; accepted May 12, 2003
Purpose. To develop a quantitative structure/activity relationship
(QSAR) model for predicting drug–CYP 3A4 interactions.
Method. The inhibitory effect of 53 structurally diverse drugs on the
metabolism of 7-benzyloxy-4-trifluoromethyl coumarin (BFC) by re-
combinant CYP 3A4 was evaluated using a rapid microtiter plate
assay. For each drug, a total of 220 two-dimensional topological in-
dices were calculated using Molconn-Z software. Using a genetic
algorithm-based partial least squares (GA-PLS) method, the desired
descriptors were automatically selected to maximize the predictabil-
ity of the IC50 values.
Results. The IC50 values of the drugs tested ranged from 9 nM to
2 mM. Based on the GA-PLS method, five principal components
derived from 20 Molconn-Z descriptors were found to be effective for
QSAR modeling. Interestingly, these descriptors suggested that the
molecular size would be an important factor in determining drug–
CYP 3A4 interactions. In the leave-one-out prediction, the rpred and
the standard error of prediction (s) were 0.754 and 0.787, respectively.
Even in an external validation, the predictions were in good agree-
ment with experimental values (rpred ס 0.744, s ס 0.769, n ס 9).
Conclusions. The proposed model, in which two-dimensional topo-
logical descriptors were used as molecular descriptors, was able to
predict drug–CYP 3A4 interactions with reasonable accuracy.
These limitations of 3D-QSAR techniques prompted us
to apply a 2D-QSAR model to predicting the interaction with
CYP 3A4. Topologic indices are attractive descriptors be-
cause they can be calculated easily and rapidly from any two-
dimensional structural formula. Various physicochemical
characteristics, e.g., partition coefficients (14) and boiling
points (15), can be described by topological indices. Recently,
we developed a genetic algorithm combined with the partial
least squares (GA-PLS) approach for feature selection in
QSAR (16,17), demonstrating that the model obtained per-
formed well as far as predicting the aqueous solubility and
membrane permeability of unknown compounds was con-
cerned.
In the present study, we applied the GA-PLS approach
to develop a 2D-QSAR model for predicting the interaction
with CYP 3A4. Molconn-Z-derived topological descriptors
(18) were used as structural descriptors for QSAR modeling,
and these included connectivity indices, shape indices, elec-
trotopological state (E-state) indices, and atom-type E-state
indices. The inhibitory effect of 53 structurally diverse drugs
on the metabolism of 7-benzyloxy-4-trifluoromethyl couma-
rin (BFC) by recombinant CYP 3A4 was evaluated using a
rapid microtiter plate assay, and then QSAR analysis was
performed.
KEY WORDS: genetic algorithm; partial least squares; CYP 3A4;
quantitative structure/activity relationship; Molconn-Z.
INTRODUCTION
The cytochrome P450s (CYPs) are a superfamily of
heme-containing mixed-function oxygenases that catalyze the
regio- and stereoselective oxidation of a wide variety of xe-
nobiotics, including drugs. CYP 3A4 is the most abundant
human hepatic CYP isoform and is responsible for the me-
tabolism of almost 50% of known drugs by humans (1). CYP
3A4 is also known to be functionally active in the intestinal
epithelium, which restricts the oral absorption of xenobiotics
(2). In addition, inhibition of CYP 3A4 by coadministered
drugs has been shown to result in adverse clinical drug–drug
interactions, some of which may be fatal (3), because of
1 Department of Drug Delivery Research, Graduate School of Phar-
maceutical Sciences, Kyoto University, Kyoto 606-8501, Japan.
2 To whom correspondence should be addressed. (e-mail: hashidam@
pharm.kyoto-u.ac.jp)
1401
0724-8741/03/0900-1401/0 © 2003 Plenum Publishing Corporation

1402
Wanchana et al.
MATERIALS AND METHODS
Materials
the first well in each row was then transferred to the second
well and serially diluted 1:3 through the eighth well. Wells 9
and 10 contained no test compound, and wells 11 and 12 were
used as controls for background fluorescence (enzyme and
substrate were added after the reaction was terminated). The
final concentration of the test compounds in the first well
varied between 1 nM and 10 mM, depending on their solu-
bility characteristics or potency. The plate was then preincu-
bated at 37°C for 20 min, and the reaction was initiated by the
addition of 100 ␮l 350 mM potassium phosphate buffer, pH
7.4, containing 10 pmole/ml insect cell-expressed CYP 3A4
and 100 ␮M BFC. The substrates were initially prepared in
acetonitrile. Here, the final concentration of BFC was set to
be 50 ␮M, in part because the metabolism of BFC is linear
with respect to BFC concentration up to 100 ␮M (4). The
reaction was terminated after 30 min by the addition of 75 ␮l
4:1 acetonitrile : 0.5 M Tris base solution. The fluorescence of
the BFC metabolite, 7-hydroxy-4-trifluoromethyl coumarin,
in each cell was measured using a Wallac multilabel counter
model 1420 fluorescence plate reader (Perkin Elmer, Fin-
land) at an excitation wavelength of 405 nm and an emission
wavelength of 535 nm. The IC50 values of each compound
were calculated through curve fitting of the Hill equation.
Midazolam was kindly supplied by Yamanouchi Pharma-
ceutical Co., Ltd. (Japan). Alprazolam, amiodarone, astemi-
zole, ␤-estradiol, carbamazepine, clonazepam, clotrimazole,
clozapine, cortisone, buspirone, dapsone, dexamethasone, dil-
tiazem, flutamide, hydrocortisone, ketoconazole, miconazole,
mifepristone, nicardipine, nimodipine, paclitaxol, quinidine,
quinine, sterigmatocystin, terfenadine, testosterone, tria-
zolam, and warfarin were purchased from Sigma Chemicals
(St. Louis, MO). Caffeine, 4-androstene-3,17,dione, benzo(a)
pyrene, chlorpheniramine, lidocaine, quercetin and verapamil
were from Nacalai Tesque (Kyoto, Japan). Progesterone was
from ICN Biomedical Inc. (Ohio, USA). Acetaminophen,
benzaldoxime, colchicine, cyclophosphamide, cyclosporin A,
dextromethorphan, diazepam, erythromycin, estriol, ethynyl-
estradiol, haloperidol, imipramine, nifedipine, tamoxifen,
troleandomycin, and vinblastine were obtained from Wako
Pure Chemicals Industries, Ltd. (Osaka, Japan). Baculovirus/
insect cell cDNA-expressed CYP 3A4 was purchased from
Gentest (Woburn, MA). Glucose-6-phosphate disodium salt,
glucose-6-phosphate dehydrogenase, ␤-nicotinamide adenine
dinucleotide phosphate sodium (␤-NADP), and 7-hydroxy-4-
trifluoromethylcoumarin were purchased from Sigma Chemi-
cals (St. Louis, MO).
Data Analysis
Calculated Molecular Descriptors
The topologic descriptors were calculated by Molconn-Z
software (Hall Associated Consulting, Quincy, MA) on the
basis of two-dimensional structures. A total of 220 connectiv-
ity, shape, and atom-type E-state indices were calculated from
the two-dimensional geometry. Molecular connectivity indi-
ces are nonempirical structure descriptions that contain infor-
mation on intermolecular accessibility (21), whereas E-state
indices contain information reflecting intermolecular accessi-
bility of atoms and groups in a molecule, specifically electron
accessibility (22).
Synthesis of 7-Benzyloxy-4-(Trifluoromethyl)-
Coumarin (BFC)
BFC was synthesized by the method of Bridge et al. (19).
Briefly, 7-hydroxy-4-trifluoromethylcoumarin (2.47 mmole)
was refluxed with benzyl bromide (4 mmole) and potassium
carbonate (5.06 mmole) in 10 ml acetone for 24 h. The crude
product was purified by recrystallization three times with di-
1
ethyl ether and stored at −20°C until use. H-NMR (CDCl3)
␦: 7.61–7.66 (m, 5H), 7.33–7.45 (m, −C6H5), 6.98–7.02 (q, 6H),
6.94–6.95 (d, 8H), 6.62 (s, 3H), 5.16 (s, -OCH2-); FAB-MS m/z
321 (M+). The recovery of BFC was 21.7%. Elemental analy-
sis gave C ס 63.68% (63.75%) H ס 3.33% (3.44%), F ס
17.61% (17.81%), and O ס 15.38% (15.00%).
Genetic Algorithm-Driven Optimization
A population of 100 random subsets of the structural
descriptors was generated. Each subset was encoded as a bi-
nary string of digits, with a length corresponding to the total
number of descriptors. A value of “1” implied that the de-
scriptor was regarded as being important, whereas a value of
“0” implied that the descriptor could be disregarded. The
predictive q2 (rpred2) value was used as a fitness function in
the genetic algorithm optimization:
Inhibition Studies
Fifty-three structurally diverse compounds that were
known or suspected to interact with CYP 3A4, as substrates
or inhibitors (20), were subjected to the assay. Their inhibi-
tory effect on the metabolism of BFC by recombinant CYP
3A4 was evaluated using the same rapid microtiter plate assay
as Stresser et al. reported (4). Incubation was conducted in a
volume of 200 ␮l in 96-well microtiter plates (Catalog 353072,
Becton Dickinson, NJ) according to a routine protocol. Serial
2
y − y
͒
͑
i
pred
͚
͚
rpred2 = 1 −
(1)
2
y − y͒
͑
i
dilutions were performed using a multichannel pipetter. A where yi and y are the observed dependent variables and their
cofactor/serial dilution (C/SD) buffer was prepared in 50 mM average, and ypred is the value predicted by the QSAR model.
potassium phosphate, pH 7.4. This buffer contained 2.6 mM The predictability of the model was evaluated using the
␤-NADP, 6.6 mM glucose-6-phosphate, and 0.8 U/ml glucose- “leave-one-out” procedure. This method systematically re-
6-phosphate dehydrogenase. Then 144 ␮l C/SD buffer that moved one data point at a time from the data set. A model
lacked test compound was added to the first well in each row, equation was then constructed on the basis of the reduced
and 100 ␮l of the same C/SD buffer to the second and all data set and subsequently used to predict the removed data
remaining wells. Six microliters of the test compounds was point. This procedure was repeated until a complete set of
added to the well in the first column. All test compounds were predicted values was obtained.
dissolved in acetonitrile. Fifty microliters of the solution from
In the genetic algorithm, two “parent” strings were se-

QSAR Analysis of Drug–CYP 3A4 Interaction
1403
lected randomly by a roulette wheel selection method accord- of BFC (testosterone) omitted, the log IC50 values of 44 com-
ing to the fitness values. A two-point crossover of the “par- pounds were subjected to QSAR analysis. Thirty-five com-
ent” strings was performed at a predefined probability (p) of pounds were selected randomly as training data, and the re-
0.8. One of the new strings was taken, subjected to random maining nine compounds were used for external validation of
mutation (p ס 0.01), and stored as an “offspring” in the next a prediction model. For all of these compounds, topological
generation. For each generation, the series of steps was re- indices were calculated using Molconn-Z software. The indi-
peated until the predefined population number (n ס 100) was ces with a path length of seven or higher were not used for
obtained, provided that the five best strings were kept for the modeling because their information is hard to interpret (18).
next generation (Elite ס 5). When the generation number In addition, the descriptors representing a heavily skewed
reached 500, the calculation was stopped, and the best string distribution (skewness greater than 3) were removed. This
in the generation was taken. A series of the computations method reduced the entire descriptor pool to 62 members.
were performed with an in-house program written in Micro-
Figure 2 shows the trajectory of the genetic algorithm-
soft Visual C++ 6.0 running on the Microsoft Windows plat- driven optimization, where the predictive q2 obtained in a
form.
“leave-one-out” procedure for 35 compounds was used as a
fitness function (Eq. 1). The average of the fitness values in
the population tends to increase with increasing generation
number and reaches a plateau at approximately 200 genera-
tions. The best solution at the 500th generation was taken for
modeling the interaction of compounds with CYP 3A4.
Table II summarizes 22 Molconn-Z descriptors selected
by genetic algorithm and their scaled PLS regression coeffi-
RESULTS
Inhibition Study Results
In this study BFC was selected as a model substrate be-
cause the metabolism of BFC was linear with respect to its
high concentration (∼100 ␮M), and the IC50 value for BFC
was in a good correlation with that for traditional compounds
(4). Figure 1 shows typical examples of the inhibitory effects
of eight concentrations of test compounds on recombinant
CYP 3A4 activity in the microtiter plate assay. Table I sum-
marizes the IC50 values of 53 structurally diverse compounds
for CYP 3A4. These IC50 values ranged from 9 nM to 2 mM,
covering the most potent inhibitors (IC50 < 1 ␮M, n ס 6),
potent inhibitors (IC50 ס 1–10 ␮M, n ס 11), fairly potent
inhibitors (IC50 ס 10–100 ␮M, n ס 13) and poor inhibitors
(IC50 > 100 ␮M, n ס 14). The IC50 value was not determined
when 50% inhibition did not occur at the highest concentra-
tion tested. The drugs that are known to cause significant
drug–drug interactions, such as ketoconazole (23) and eryth-
romycin (1), tend to inhibit the metabolism of BFC.
cients. Because the scaled PLS regression coefficients for 4␹
pc
and SdsCH were negligibly small, these descriptors could be
omitted. When the “leave-one-out” prediction was conducted
using the remaining 20 descriptors, the optimal number of
PLS principal components was found to be 5. The “leave-one-
out” prediction gave an rpred of 0.754 and a standard error of
prediction (s) of 0.787 (Fig. 3A). By taking five PLS principal
components of 20 descriptors from the entire data on the 35
compounds, the following linear equation for estimating the
IC50 values was obtained:
log IC50 = 3.3681 − 0.0160 nvx + 0.1627 nrings − 0.0013 fw
− 0.0203 0␹ − 0.0358 1␹ − 0.0301 2␹ + 0.0676 5␹
p
− 0.0225 0␹v + 0.1826 5␹vp + 0.2724 6␹
vp
− 6.6124 5␹ch − 5.7457 6␹ch − 6.0471 5␹
vch
− 0.0181 1␬␣ − 0.0695 SHsOH + 0.1848 SaaN
Genetic Algorithm-Driven Optimization in
QSAR Modeling
− 0.0136 SsOH − 0.0959 SHvin + 0.6432 NHBint5
+ 0.6136 NHBint6 r ס 0.88, s = 0.57, n = 35͒
͑
With the drugs that have limited solubility [benzo(a)py-
rene, caffeine, cortisone, dexamethasone, flutamide, hydro-
(2)
cortisone, taxol, and estriol] or that activate the metabolism As summarized in Table III, the five PLS principal compo-
Fig. 1. Inhibitory effects of eight concentrations of the representative test compounds
on the metabolism of 7-benzyloxy-4-trifluoromethyl coumarin (BFC) by recombinant
CYP 3A4.

1404
Wanchana et al.
Table I. The IC50 Values of Test Compounds for CYP 3A4 in the Microtiter Plate
Assay Calculated Using Three Parameters (the Hill Equation)
Drug
IC50 (␮M)
Drug
Flutamide
Haloperidol
Hydrocortisone
Imipramine
Ketoconazole
Lidocaine
Miconazole
Midazolam
Mifepristone
Nicardipine
Nifedipine
Nimodipine
Progesterone
Quercetin
IC50 (␮M)
4-Androstene-3,17,dione
Acetaminophen
Alprazolam
Amiodarone
Astemizole
Benzaldoxime
Benzo(a)pyrene
Buspirone
497
1496
423
1.79
3.23
361
>2000
58.9
>500
109
0.009
1451
0.477
3.54
0.682
0.387
16.6
>10
22.2
>2500
376
709
167
0.0341
30.5
194
Caffeine
Carbamazepine
Chlorpheniramine maleate
Clonazepam
Clotrimazole
Clozapine
2.26
46.0
19.9
76.0
Colchicines
Quinidine
Cortisone
>300
1976
4.19
80.3
>1000
99.2
115
Quinine
57.3
1.92
7.58
>100
3.70
Cyclophosphamide
Cyclosporin A
Dapsone
Dexamethasone
Dextromethophan
Diazepam
Sterigmatocystin
Tamoxifen citrate
Taxol
Terfenadine
Testosterone
Triazolam
Activation
163
Diltiazem
Erythromycin
␤-Estradiol
78.5
9.82
37.0
>50
Troleandomycin
Verapamil
Vinblastine
Warfarin
0.930
2.77
17.5
Estriol
314
Ethynylestradiol
2.31
nents account for 85.7% of the variation in the factors (mo-
lecular descriptors) and 77.4% of the variation in the re-
sponse (log IC50). The relationship between the observed log
IC50 values and those calculated from Eq. (1) is shown in Fig.
3B. In addition, an external validation of Eq. (2) was con-
ducted with a testing data set (Fig. 4). The predictions were in
good agreement with the observed values with an rpred of
0.744 and an s of 0.769. Thus, it was demonstrated that the
GA-PLS analysis gave a QSAR model of drug–CYP 3A4
interactions with a reasonable accuracy of prediction.
DISCUSSION
In the present study, a 2D-QSAR model for predicting
drug–CYP 3A4 interactions was constructed using the GA-
PLS method. In the PLS analysis (24), the matrix of explana-
tory variables is orthogonally decomposed with the inner re-
lation between the explanatory and response variables being
adhered to. Therefore, unlike MLR analysis, PLS analysis
avoids any multicollinearity problems in the explanatory vari-
ables. The PLS regression is sufficiently noise-free because a
minimal number of principal components are used for mod-
eling. Thus, the PLS analysis is widely used in many fields of
chemistry. However, it should be noted that the incorporation
of unnecessary explanatory variables adversely affects PLS
modeling. The use of a genetic algorithm is one way to elimi-
nate this problem (16,17,25,26). Genetic algorithms (27) are
search algorithms based on the mechanics of natural selection
and natural genetics, which exploit the idea of the survival of
the fittest and an interbreeding population. Genetic algo-
rithms differ from more traditional optimization techniques in
Fig. 2. Relationship between fitness and generation number in GA-
PLS QSAR modeling of the interaction with CYP 3A4. The fitness
function is defined as
2
y − y
͒
͑
i
pred
2
͚
rpred2 = 1 −
y − y͒
͑
͚
i
where yi and y were the observed dependent variables and their
average; ypred was the value predicted by the QSAR model. Dash-dot
(—и—) and solid (—) lines represent the maximum and average of
fitness values of individuals in the population. Each generation in-
cludes 100 individuals.

QSAR Analysis of Drug–CYP 3A4 Interaction
1405
Table II. Scaled PLS Regression Coefficients of the Subset of Molconn-Z Descriptors Selected
by a Genetic Algorithm in Combination with PLS Regression
Scaled PLS regression
Descriptor
Significant descriptors
Symbol
coefficienta
Number of nonhydrogen atoms in molecule
Number of rings in graph
Molecular weight
nvx
nrings
fw
−0.222
0.251
−0.273
−0.221
−0.229
−0.185
0.251
−0.210
0.357
0.389
−0.464
−0.414
−0.250
−0.242
−0.235
0.378
−0.175
−0.150
0.604
0
Path 0 simple connectivity index
Path 1 simple connectivity index
Path 2 simple connectivity index
Path 5 simple connectivity index
Path 0 valence connectivity index
Path 5 valence connectivity index
Path 6 valence connectivity index
Chain 5 simple connectivity index
Chain 6 simple connectivity index
Chain 5 valence connectivity index
Kappa simple index
Hydrogen E-state index value for atom type −OH
E-state index value for atom type --N--
E-state index value for atom type −OH
E-state of C-atom in the vinyl group, סCH−
Count of potential internal hydrogen bonders
Count of potential internal hydrogen bonders
Descriptors that can be removed
␹
1
␹
2
␹
5
␹
p
0
␹
v
5
␹
vp
6
␹
vp
5
␹
ch
6
␹
ch
5
␹
1
vch
␬
␣
SHsOH
SaaN
SsOH
SHvin
NHBint5
NHBint6
0.362
4
Path-cluster 4 simple connectivity index
E-state index value for atom type סCH−
␹
−0.072
−0.033
pc
SdsCH
a Scaled PLS coefficients were calculated by multiplying the PLS regression coefficient and the
standard deviation of the descriptors.
that they involve a search from a “population” of solutions, total number of combinations of molecular descriptors for
not from a single point. Therefore, genetic algorithms are modeling was 262 (∼4.6 × 1018), and the number of individuals
viewed as a global optimization method. It should be noted for each population was set to be only 100.
from Fig. 2 that genetic algorithms are able to find an optimal
The optimal number of PLS principal components was
solution very efficiently, taking into account the facts that the determined by evaluating the predictability of the model in a
Fig. 3. Leave-one-out cross-validated predictability (A) and goodness of fit (B) of the QSAR model obtained. The model
comprises the five PLS principal components of 20 descriptors listed in Table II. The correlation coefficient (r) between the
experimental and calculated log IC50 values is also given.

1406
Wanchana et al.
Table III. Contribution of PLS Principal Components to the Varia-
tions of Response Variable (Log IC50) and Explanatory Variables
(Molecular Descriptors)
the prediction for the same testing data was also within 1 log
unit (s ס 0.57, n ס 12). Thus, 2D-QSAR models appear to be
comparable with 3D-QSAR models in terms of predictability.
It is difficult to understand structural features involving
CYP 3A4 interaction directly from Eq. (2) because correla-
tions between the descriptors were relatively high. Because
the PLS score, which is a linear combination of the explana-
tory variables, is used for the regression with the response
variable, relationships of the PLS score in each principal com-
ponent with physicochemical parameters such as molecular
weight, hydrophobicity, and hydrogen-bonding ability were
investigated. As a result, the first principal component of the
20 descriptors selected correlated well with the molecular
weight, having an r value of −0.895 (Fig. 5). Taken together
with the finding that the first PLS principal component posi-
tively correlated with the IC50 value (data not shown), this
showed that the IC50 value of compounds in the CYP 3A4
interaction is smaller for larger compounds. In a CYP 3A4
homology model, erythromycin contacts more active site resi-
dues than progesterone because of its larger molecular size,
where both compounds are stabilized mainly through hydro-
phobic interactions (9). It has been reported that the hydro-
phobicity of compounds plays an important role in oxidation
by CYPs (28) and binding to liver microsomes (29). Because
hydrophobicity is highly dependent on molecular volume
(30), the molecular size would be an important factor deter-
mining any drug–CYP 3A4 interaction. However, it should be
noted that the contribution ratio of the first principal compo-
nent was 25.5%, which was not as high as the direct correla-
tion between log IC50 and molecular weight, was detectable.
In addition to the first principal component vs. molecular
weight relationship, unfortunately, no other significant rela-
tionships were observed.
Response variable
Explanatory variables
Number
of PCa
Variation
(%)
Cumulative
variation (%)
Variation
(%)
Cumulative
variation (%)
1
2
3
4
5
25.5
21.9
23.1
4.3
25.5
47.5
70.6
74.9
77.4
46.4
17.4
8.2
8.8
5.0
46.4
63.8
72.0
80.8
85.7
2.5
a PC, PLS principal components.
leave-one-out cross-validation procedure because extraction
of too many principal components resulted in overfitting to
the data. In this model, only five principal components were
used for modeling, a number that was small enough as com-
pared with the sample number (n ס 35). In addition, this
model could predict the IC50 values for external data as well.
The model obtained gave a reasonably accurate predic-
tion, with an rpred of 0.754 and 0.744 for the “leave-one-out”
internal validation and external validation, respectively. It
would be interesting to compare the predictability of the pre-
sent 2D-QSAR approach with that of the 3D-QSAR ap-
proach. Because Catalyst, a representative 3D-QSAR pro-
gram, was not available, we analyzed the data set of Ekins et
al. (10) using our approach and compared the results. For the
Km values of 38 substrates against human liver microsomal
CYP 3A4, the pharmacophore model built by Catalyst had an
r value of 0.67 (n ס 38), and the prediction for 12 testing data
was within 1 log unit for the residual (s ס 0.55, n ס 12). On
the other hand, our 2D-QSAR analysis gave an r value of 0.87
for the five PLS principal components of 14 descriptors, and
Riley et al. (31) measured inhibition of CYP 3A4 activity,
i.e., N-demethylation of erythromycin, by 30 compounds and
found that their IC50 values inversely correlated with lipo-
Fig. 5. Relationship between the molecular weight and the first PLS
principal component of the 20 descriptors selected. The score values
corresponding to the first PLS principal component were calculated
based on a loading vector (b1): b1 ס (0.335, 0.177, 0.326, 0.327, 0.338,
0.340, 0.351, 0.327, 0.313, −0.284, 0.159, −0.0168, 0.150, 0.310, 0.0940,
Fig. 4. External validation of the QSAR model obtained. IC50 values −0.0521, 0.112, 0.127, 0.227, 0.0416). The elements in the vector are
for nine compounds were predicted by using Eq. (2). coefficients of the molecular descriptors (Table II).

QSAR Analysis of Drug–CYP 3A4 Interaction
1407
methods for modeling cytochrome p450 active sites. Drug Metab.
Dispos. 29:936–944 (2001).
philicity (log D7.4), especially when N-containing heterocyclic
compounds and others are analyzed individually. We ana-
lyzed our data in the same manner for 31 compounds, of
which log D7.4 values were available from the literature (31–
34). Weak correlations were observed, with the r values of
0.67 and 0.65 for N-containing heterocyclic compounds (n ס
13) and others (n ס 18), respectively. However, these r values
were not so high as those from our model (r ס 0.88, n ס 35).
An advantage of QSAR models based on two dimen-
sional topological descriptors is that they eliminate the con-
formational and alignment ambiguities inherent within a 3D-
QSAR process. Additionally, the two-dimensional topologi-
cal descriptors are less computationally intensive, practically
completely automated, and have been used to produce highly
predictive models that are comparable to, or better than,
those obtained using 3D-QSAR approaches (26). The most
limiting feature of any 2D-QSAR approach is its insensitivity
to the stereochemistry of the members of the training and
prediction data sets and the lack of easily interpretable infor-
mation useful for the design of new highly active drugs. In
contrast, three-dimensional approaches provide graphic rep-
resentations of pharmacophores (35) or putative receptor
sites (36) and indicate the best directions for rational design.
In view of this and the fact that two-dimensional approaches
would be very helpful in screening a large number of virtual
compounds, it appears that 2D- and 3D-QSAR analyses
complement each other according to the purposes of the dif-
ferent screening stages.
8. D. F. V. Lewis and B. G Lake. Molecular modelling of CYP 1A
subfamily members based on an alignment with CYP 102: Ratio-
nalization of CYP 1A substrate specificity in terms of active site
amino acid residues. Xenobiotica 26:723–753 (1996).
9. G. D. Szklarz and J. R. Halpert. Molecular modeling of cyto-
chrome P450 3A4. J. Comp. Aided. Mol. Des. 11:265–272 (1997).
10. S. Ekins, G. Bravi, J. H. Wikel, and S. A. Wrighton. Three-
dimensional quantitative structure activity relationship analysis
of cytochrome P-450 3A4 substrates. J. Pharmacol. Exp. Ther.
291:424–433 (1999).
11. S. Ekins, G. Bravi, S. Binkley, J. S. Gillespie, B. J. Ring, J. H.
Wikel, and S. A. Wrighton. Three- and four-dimensional quanti-
tative structure activity relationship analyses of cytochrome
P-450 3A4 inhibitors. J. Pharmacol. Exp. Ther. 290:429–438
(1999).
12. D. F. Lewis, S. Modi, and M. Dickins. Structure-activity relation-
ship for human cytochrome P450 substrates and inhibitors. Drug
Metab. Rev. 34:69–82 (2002).
13. K. R. Korzekwa, N. Krishnamachary, M. Shou, A. Ogai, R. A.
Parise, A. E. Rettie, F. J. Gonzalez, and T. S. Tracy. Evaluation
of atypical cytochrome P450 kinetics with two-substrate models:
evidence that multiple substrates can simultaneously bind to cy-
tochrome P450 active sites. Biochemistry 37:4137–4147 (1998).
14. J. J. Huuskonen, A. E. P. Villa, and I. V. Tetko. Prediction of
partition coefficient based on atom-type electrotopological state
indices. J. Pharm. Sci. 88:229–233 (1999).
15. L. H. Hall and C. T. Story. Boiling point and critical temperature
of a heterogeneous data set: QSAR with atom type electrotopo-
logical state indices using artificial neural networks. J. Chem. Inf.
Comp. Sci. 36:1004–1014 (1996).
16. S. Wanchana, F. Yamashita, and M. Hashida. Quantitative struc-
ture/property relationship analysis on aqueous solubility using
genetic algorithm-combined partial least squares method.
Pharmazie 57:127–129 (2002).
17. F. Yamashita, S. Wanchana, and M. Hashida. Quantitative struc-
ture/property relationship analysis of Caco-2 permeability using a
genetic algorithm-based partial least squares method. J. Pharm.
Sci. 90:2230–2239 (2002).
18. L. B. Kier and L. H. Hall. Topological information. In L. B. Kier
and L. H. Hall (eds.), Molecular Connectivity in Structure-Activity
Analysis. Research Studies, Hertfordshire, 1986, pp. 43–68.
19. W. Bridge, A. J. Crocker, T. Cubin, and A. Robertson. Experi-
ments on the synthesis of rotenone and its derivatives. Part XIII.
J Chem. Soc. 1530–1535 (1937).
20. A. P. Li, D. L. Kaminski, and A. Rasmussen. Substrates of hu-
man hepatic cytochrome P450 3A4. Toxicology 104:1–8 (1995).
21. L. B. Kier and L. H. Hall. Intermolecular accessibility: the mean-
ing of molecular connectivity. J. Chem. Inf. Comput. Sci. 40:792–
795 (2000).
In conclusion, the proposed model, in which two-
dimensional topological descriptors are used as molecular de-
scriptors, is able to predict drug–CYP 3A4 interactions with
reasonable accuracy. Genetic algorithm-based approaches
would be useful in selecting a set of effective descriptors for
QSAR modeling.
ACKNOWLEDGMENTS
This research was supported in part by a Grant-in-Aid
for Scientific Research from the Ministry of Education, Cul-
ture, Sports, Science and Technology, Japan.
REFERENCES
22. L. H. Hall and L. B. Kier. The E-state as the basis for molecular
structure space definition and structure similarity. J. Chem. Inf.
Comput. Sci. 40:784–791 (2000).
23. A. Varhe, K. T. Olkkola, and P. J. Neuvonen. Oral triazolam is
potentially hazardous to patients receiving systemic antimycotics
ketoconazole or itraconazole. Clin. Pharmacol. Ther. 56:601–607
(1994).
1. F. P. Guengerich. Role of cytochrome P450 enzymes in drug-drug
interactions. Adv. Pharmacol. 43:7–35 (1997).
2. M. F. Paine, M. Khalighi, J. M. Fisher, D. D. Shen, K. L. Kunze,
C. L. Marsh, J. D. Perkins, and K. E. Thummel. Characterization
of interintestinal and intraintestinal variations in human CYP
3A-dependent metabolism. J. Pharmacol. Exp. Ther. 283:1552–
1562 (1997).
3. P. K. Honig, D. C. Wortham, K. Zamani, D. Conner, J. C. Mullin,
and L. R. Cantilena. Terfenadine–ketonazole interaction. JAMA
269:1513–1518 (1993).
4. D. M. Stresser, A. P. Blanchard, S. D. Turner, J. C. L. Erve, A. A.
Dandeneau, V. P. Miller, and C. L. Crespi. Substrate-dependent
modulation of CYP 3A4 catalytic activity: analysis of 27 test com-
pounds with four fluorometric substrates. Drug Metab. Dispos.
28:1440–1448 (2000).
24. P. Geladi and B. R. Kowalski. Partial least-squares regression: a
tutorial. Anal. Chim. Acta 185:1–17 (1986).
25. K. Hasegawa, Y. Miyashita, and K. Funatsu. GA strategy for
variable selection in QSAR studies: GA-based PLS analysis of
calcium channel antagonists. J. Chem. Inf. Comput. Sci. 37:306–
310 (1997).
26. B. Hoffman, S. J. Cho, W. Zheng, S. Wyrick, D. E. Nichols, R. B.
Mailman, and A. Tropsha. Quantitative structure activity rela-
tionship modeling of dopamine D1 antagonists using comparative
molecular field analysis, genetic algorithms–partial least squares,
and K nearest neighbor methods. J. Med. Chem. 42:3217–3226
(1999).
5. C. M. Masimirembwa, R. Thompson, and T. B. Andersson. In
vitro high throughput screening of compounds for favorable
metabolic properties in drug discovery. Comb. Chem. High
Throughput Screen. 4:245–263 (2001).
27. D. E. Goldberg. Genetic Algorithms in Search, Optimization, and
Machine Learning. Addison-Wesley, New York, 1989.
28. C. Hansch. Quantitative relationships between lipophilic charac-
ter and drug metabolism. Drug Metab. Rev. 1:1–14 (1972).
29. R. P. Austin, P. Barton, S. L. Cockroft, and M. C. Wenlock. The
6. J. Langowski and A. Long. Computer systems for the prediction
of xenobiotic metabolism. Adv. Drug Deliv. Rev. 54:407–415
(2002).
7. S. Ekins, M. J. de Groot, and J. P. Jones. Pharmacophore and
three-dimensional quantitative structure activity relationship

1408
influence of nonspecific microsomal binding on apparent intrinsic
Wanchana et al.
Basic and neutral compounds. J. Med. Chem. 44:2490–2497
(2001).
clearance, and its prediction from physicochemical properties.
Drug Metab. Dispos. 30:1497–1503 (2002).
34. C. Zhu, L. Jiang, T. M. Chen, and K. K. Hwang. A comparative
study of artificial membrane permeability assay for high through-
put profiling of drug absorption potential. Eur. J. Med. Chem.
37:399–407 (2002).
35. J. H. Van Drie, D. Weininger, and Y. C. Martin. ALADDIN: an
integrated tool for computer-assisted molecular design and phar-
macophore recognition from geometric, steric, and substructure
searching of three-dimensional molecular structures. J. Comput.
Aided Mol. Des. 3:225–251 (1989).
36. R. D. Cramer, D. E. Patterson, and J. D. Bunce. Comparative
molecular field analysis (CoMFA). 1. Effect of shape on binding
of steroids to carrier proteins. J. Am. Chem. Soc. 110:5959–5967
(1988).
30. A. Leo, C. Hansch, and P. Y. C. Jow. Dependence of hydropho-
bicity of apolar molecules on their molecular volume. J. Med.
Chem. 19:611–615 (1976).
31. R. J. Riley, A. J. Parker, S. Trigg, and C. N. Manners. Develop-
ment of a generalized, quantitative physicochemical model of
CYP 3A4 inhibition for use in early drug discovery. Pharm. Res.
18:652–655 (2001).
32. F. Yoshida and J. G. Topliss. QSAR model for drug human oral
bioavailability. J. Med. Chem. 43:2575–2585 (2000).
33. F. Lombardo, M. Y. Shalaeva, K. A. Tupper, and F. Gao. Elog-
Doct: a toot for lipophilicity determination in drug discovery. 2.