Design, synthesis and experimental validation of novel potential chemopreventive agents using random forest and support vector machine binary classifiers

Article abstract of DOI:10.1007/s10822-014-9748-9

Compared to the current knowledge on cancer chemotherapeutic agents, only limited information is available on the ability of organic compounds, such as drugs and/or natural products, to prevent or delay the onset of cancer. In order to evaluate chemical chemopreventive potentials and design novel chemopreventive agents with low to no toxicity, we developed predictive computational models for chemopreventive agents in this study. First, we curated a database containing over 400 organic compounds with known chemoprevention activities. Based on this database, various random forest and support vector machine binary classifiers were developed. All of the resulting models were validated by cross validation procedures. Then, the validated models were applied to virtually screen a chemical library containing around 23,000 natural products and derivatives. We selected a list of 148 novel chemopreventive compounds based on the consensus prediction of all validated models. We further analyzed the predicted active compounds by their ease of organic synthesis. Finally, 18 compounds were synthesized and experimentally validated for their chemopreventive activity. The experimental validation results paralleled the cross validation results, demonstrating the utility of the developed models. The predictive models developed in this study can be applied to virtually screen other chemical libraries to identify novel lead compounds for the chemoprevention of cancers.

Full text of DOI:10.1007/s10822-014-9748-9

J Comput Aided Mol Des (2014) 28:631–646
DOI 10.1007/s10822-014-9748-9
Design, synthesis and experimental validation of novel potential
chemopreventive agents using random forest and support vector
machine binary classifiers
•
Brienne Sprague Qian Shi Marlene T. Kim
•
•
•
Liying Zhang Alexander Sedykh Eiichiro Ichiishi
•
•
•
•
Harukuni Tokuda Kuo-Hsiung Lee Hao Zhu
Received: 12 December 2013 / Accepted: 5 May 2014 / Published online: 20 May 2014
Ó Springer International Publishing Switzerland 2014
Abstract Compared to the current knowledge on cancer
chemotherapeutic agents, only limited information is
available on the ability of organic compounds, such as
drugs and/or natural products, to prevent or delay the onset
of cancer. In order to evaluate chemical chemopreventive
potentials and design novel chemopreventive agents with
low to no toxicity, we developed predictive computational
models for chemopreventive agents in this study. First, we
curated a database containing over 400 organic compounds
with known chemoprevention activities. Based on this
database, various random forest and support vector
machine binary classifiers were developed. All of the
resulting models were validated by cross validation pro-
cedures. Then, the validated models were applied to vir-
tually screen a chemical library containing around 23,000
natural products and derivatives. We selected a list of 148
novel chemopreventive compounds based on the consensus
prediction of all validated models. We further analyzed the
predicted active compounds by their ease of organic syn-
thesis. Finally, 18 compounds were synthesized and
experimentally validated for their chemopreventive activ-
ity. The experimental validation results paralleled the cross
validation results, demonstrating the utility of the devel-
oped models. The predictive models developed in this
study can be applied to virtually screen other chemical
libraries to identify novel lead compounds for the chemo-
prevention of cancers.
Keywords Cheminformatics ꢀ Chemoprevention ꢀ
Natural products ꢀ QSAR ꢀ Virtual screening
Abbreviations
CCR
Correct classification rate
Electronic supplementary material The online version of this
article (doi:10.1007/s10822-014-9748-9) contains supplementary
material, which is available to authorized users.
Combi-QSAR Combinatorial quantitative structure–
activity relationship
B. Sprague ꢀ M. T. Kim ꢀ H. Zhu (&)
Department of Chemistry, Rutgers University, 315 Penn St.,
Camden, NJ 08102, USA
A. Sedykh
Multicase Inc, 23811 Chagrin Blvd, Suite 305, Beachwood,
OH 44122, USA
e-mail: hao.zhu99@rutgers.edu
E. Ichiishi
Department of Internal Medicine, International University of
Health and Welfare Hospital, Tochigi 329-2763, Japan
Q. Shi (&) ꢀ K.-H. Lee
Natural Products Research Laboratories, UNC Eshelman School
of Pharmacy, University of North Carolina at Chapel Hill, 310
Beard Hall, CB# 7568, Chapel Hill, NC 27599-7568, USA
e-mail: qshi1@email.unc.edu
H. Tokuda
Department of Complementary and Alternative Medicine
Clinical R&D, Kanazawa University of Graduate School of
Medical Science, Kanazawa 920-8640, Japan
M. T. Kim ꢀ H. Zhu
The Rutgers Center for Computational and Integrative Biology,
Camden, NJ 08102, USA
K.-H. Lee
Chinese Medicine Research and Development Center, China
Medical University and Hospital, Taichung, Taiwan
L. Zhang
Pfizer Worldwide Research and Development, Groton,
CT 06340, USA
123

632
J Comput Aided Mol Des (2014) 28:631–646
CPT
Consensus prediction thresholds
Epstein–Barr virus early activation
Multiple linear regression
major receptor for 12-O-tetradecanoylphorbol-13-acetate
(TPA) [19, 20]. In this test, the experiment is carried out in
Raji cells, a human lymphoma cell line with an Epstein–
Barr virus genome. The assay is less tumor type specific,
which limits its applicability in mechanistic studies of
chemoprevention; however, it can measure the chemopre-
ventive effects of a particular agent on the promotion and
progression phases of carcinogenesis, and usually results in
parallel outcomes to in vivo animal models. It is considered
to be a useful in vitro assay to initially screen chemopre-
ventive agents [12, 13, 18].
EBV-EA
MLR
MOE
PKC
Molecular operating environment
Protein Kinase C
PLS
Partial least square
PTLC
QSAR
Preparative thin layer chromatography
Quantitative
relationship
structure–activity
RF
Random forest
SVM
TLC
TMS
TPA
ZND
Support vector machines
Thin layer chromatography
Tetramethylsilane
Currently, only a few studies have employed computa-
tional modeling in the area of chemoprevention. Among
them, several Quantitative structure–activity relationship
(QSAR) studies used simple modeling approaches, such as
Multiple linear regression (MLR) or Partial least square
(PLS), to model a limited number of chemopreventive
agents. For example, Bertosa et al. [21] used PLS to
develop QSAR models for 59 amides and quinolones. The
antitumor activities of these compounds were tested against
MiaPaCa-2 (pancreatic carcinoma) and MCF-7 (breast
carcinoma) cells. In another study, Saeed et al. [22] used
linear regression to generate models for six curcumin
derivatives. In a recent published work, Aleksic et al. [23]
reported a three-dimensional (3D) QSAR study of substi-
tuted heterocyclic quinolones. Nineteen compounds with
similar quinolone scaffolds were synthesized and tested for
antitumor activity against multiple cell lines in this study.
The results were modeled with commercial modeling
software. Other receptor-based modeling studies, which
used molecular docking analysis, have focused on using
known cancer tumor targets and their interaction with
specific chemopreventive compounds. For example, a
recent study tested the binding affinities of curcumin
derivatives against several well-known cancer targets [24].
The results revealed a favorable correlation between the
two test compounds and their binding affinity for cancer
targets, thus, implying a functional role as enzyme inhibi-
tion activators. Although previous modeling studies
achieved certain successes (e.g., molecular docking studies
usually helped to explain the binding mechanisms of the
chemopreventive agents to the relevant receptors), they
have restricted predictive ability due to the limited number
of compounds used in the studies.
12-O-tetradecanoylphorbol-13-acetate
ZINC natural derivative
Introduction
Cancer is listed among the major causes of mortality in the
world. In 2013, cancer was ranked as the second leading
cause of death in the United States. A recent report from
the American Cancer Society revealed that, statistically,
the lifetime chances of developing cancers are as high as 1
in 3 for women and 1 in 2 for men [1]. Current strategies in
cancer patient treatments, such as chemotherapy, have met
with clinical success. However, most chemotherapeutic
agents have severe side effects, which negatively impact a
cancer patient’s quality of life [2]. For this reason, che-
moprevention, which normally uses either natural or syn-
thetic compounds with low toxicity, was employed to
impede, halt, or reverse the carcinogenesis process before a
tumor can develop [3], especially for patients at high risk.
The different stages of chemoprevention research have
been extensively reviewed [4–8].
As a relatively new area of cancer research, there is a
high demand for efficient methods to identify novel che-
moprevention agents. To date, a standardized method for
identifying chemopreventive compounds has not been
developed. But several in vitro methods have been used by
different research groups to evaluate potential chemopre-
ventive agents [9–11]. Most of these methods measure
chemopreventive activity on the basis of similar general
principles. These approaches involve measuring cellular
expression of a protein in a human cancer cell line with and
without the test compound. A positive impact is noted by a
decrease in expression, which indicates interference with a
potential cancer-inducing pathway [12–18]. The Epstein–
Barr virus early antigen (EBV-EA) activation assay is one
such test recognized as a primary screening method for
assessing antitumor promoting properties, through inhibi-
tion of Protein Kinase C (PKC) activity, which serves as a
In this study, we proposed to develop QSAR models
based on a large set of chemopreventive agents tested by
the EBV-EA assay and to apply the resulting models to
design new lead agents. To this end, we curated a che-
mopreventive database by collecting compounds tested by
the EBV-EA assay from published research articles [25–
51]. We used an in-house tool to automatically generate an
activity endpoint based on the original multi-dose chemo-
preventive response data. Various QSAR models were
123

J Comput Aided Mol Des (2014) 28:631–646
633
developed and validated. Then, these models were used to
virtually screen over 23,000 natural products and their
derivatives to identify novel chemopreventive agents.
Finally, 18 lead compounds resulting from the prediction
results were synthesized and experimentally validated
using the same EBV-EA assay.
of the number of positive cells compared with that of the
control without the test compound. Cell viability was
assayed by the Trypan Blue staining method. For the
determination of cytotoxicity, the cell viability was
required to be more than 60 % [53]. Each compound was
measured at four concentrations of 0.32, 3.2, 16, and
32 nmol, representing 10-, 100-, 500-, 1,000-fold of TPA
(32 pmol). Each measurement was repeated three times for
each test compound concentration and the average values
of the three readout data were used.
Methods
Chemopreventive agent data set
Chemical descriptors
The chemopreventive agent database was generated by
collecting data from different journal papers published
during the past decade [25–51]. All compounds in the
database were tested using the EBV-EA assay by the same
standard methodology. The original database contained
compounds that were reported in different papers. After
removing duplicated compounds by harmonizing their
activities, we had 405 unique compounds left in the final
curated chemopreventive agent database.
The chemical descriptors used in this study were obtained
from Dragon version 6.0 (Talete SRL, Milano, Italy) and
Molecular Operating Environment (MOE) version 2011.
The Dragon descriptors include E-state values and E-state
counts, constitutional descriptors, topological descriptors,
walk and path counts, connectivity and information indi-
ces, 2D autocorrelations, Burden eigenvalues, molecular
properties, Kappa, hydrogen bond acceptor/donor counts,
molecular distance edge, and molecular fragment counts.
The MOE descriptors include topological indices, struc-
tural keys, E-state indices, physical properties (i.e., LogP,
molecular weight, and molar refractivity), and topological
polar surface area. Over 4,000 Dragon descriptors were
initially generated, but most of them were redundant. We
removed redundant Dragon descriptors by using pairwise
comparisons between each of the two descriptor pairs. If
the correlation between two descriptors of our 405 com-
pounds was high (correlation coefficient [0.95), one of
them was randomly selected and removed. Eventually, 688
Dragon descriptors were left for this study. MOE generated
186 descriptors, all of which were used in the modeling
process.
The EBV-EA assay data were treated by the CurveP
algorithm to ensure monotonicity of each dose–response
curve and to convert it into a numeric value, LogCurveP
(a log10-transformed fingerprint), which was used as a
numeric indicator of activity. This algorithm was devel-
oped in our laboratory in a prior study [52]. Briefly, the
response at each of the various test concentrations was
represented by two bits (00, 01, 10, 11) for coding four
categories (0, 25, 50, and 75 %? relative inhibition). Then,
these bits were concatenated from lowest to highest test
concentrations and resulted in an eight bit value (CurveP).
If nonzero, the CurveP was then log10-transformed. The
chosen activity threshold of 1.25 corresponds to strong
inhibition at the two highest test concentrations. Based on
these factors, the chemopreventive agent database con-
tained 204 ‘‘actives’’ categorized as class-1 and 201
‘‘marginal actives’’ categorized as class-2. All 405 com-
pounds and their chemopreventive activity (both the ori-
ginal data and the LogCurvP results) are listed in a
supplemental file (Supplemental Table 1).
Modeling approaches
This study used the Random forest (RF) and Support vector
machine (SVM) algorithms available in R.2.15.1. These
two algorithms have been employed in several of our
previous modeling studies for various biological activities
[54, 55].
EBV-EA assay
The entire Combinatorial QSAR (Combi-QSAR) mod-
eling workflow is shown in Fig. 1. Individual models were
developed using a combination of Dragon or MOE
descriptors and RF or SVM algorithms. This technique
resulted in four different models: Dragon-RF, Dragon-SVM,
MOE-RF, and MOE-SVM. The results for each classifica-
tion model were averaged to generate consensus predictions,
which will be further referred to as a consensus model.
All models were validated using five-fold external cross
validation. Briefly, the original chemopreventive dataset
Raji cells (106 cells/mL) were incubated at 37 °C for 48 h
in RPMI-1640 medium with 10 % Fetal calf serum (FCS),
n-butyric acid (4 mmol), TPA (32 pmol), and test com-
pounds. Smears were made from the cell suspension, and
EBV-EA inducing cells were stained by an indirect
immunofluorescence technique. In each assay, at least 500
cells were counted and the number of stained cells (positive
cells) was recorded. The EBV-EA-inhibiting activity of the
test compound was estimated on the basis of the percentage
123

634
J Comput Aided Mol Des (2014) 28:631–646
Fig. 1 The Combi-QSAR
modeling workflow of this study
ꢀ
ꢁ
true positives
was randomly divided into five equal subsets. One subset
was used as the validation set (20 % of the original set) and
the other four subsets (80 % of the original set) were used
as the training sets. The training sets were used to develop
the models and the resulting models were used to predict
the left-out validation set. This procedure was repeated five
times, so that each compound was used for validation
purposes once.
% sensitivity ¼
ꢁ 100
true positives þ false negatives
ð1Þ
ꢀ
ꢁ
true negatives
% specificity ¼
ꢁ 100
true negatives þ false positives
ð2Þ
ð3Þ
ꢀ
Synthesis of selected compounds
ꢁ
sensitivity þ specificity
Robustness of QSAR models was verified using a
Y-randomization (randomization of response) approach as
described by Tropsha and coworkers [56–58]. We ran-
domly assigned the activities of the modeling set com-
pounds into class-1 or -2. Then, we developed QSAR
models using the same protocol as for compounds with
actual experimental results. The purpose of this procedure
was to see if statistically significant QSAR models could be
obtained for the original data, but could not be developed
with randomized activities. The Y-randomization tests for
each combination of modeling approach and descriptor
were repeated five times.
% CCR ¼
ꢁ 100
2
Eighteen compounds derived from the predicted results
were designated for chemical synthesis. Among them,
twelve compounds (1–12) resulted from the class-1 com-
pound list and six compounds (13–18) were derived from
the class-2 prediction set (Table 2). Compound selection
from more than 100 predicted leads was based on the
chemical capability of producing the molecules through
basic organic synthesis and the availability of chemical
reagents for making the compound. Compound 12 was
purchased from Aldrich. All final compounds were struc-
turally confirmed by mass spectrometry (Shimadzu LCMS-
2010 ESI–MS) and proton nuclear magnetic resonance
spectroscopy (1H NMR) [Varian 400 MHz with tetra-
methylsilane (TMS) as the internal standard]. Melting
points were determined on a Fisher-John melting point
apparatus and are uncorrected. CombiFlashÒ chromato-
graphic system (Isco Companion) with a Grace silica gel
cartridge was used for general separation and purification.
Preparative thin layer chromatography (PTLC) on silica gel
plates (Kieselgel 60, F254, 1.50 mm) was also used for
Universal criteria for model evaluation
Because various modeling approaches and different
descriptors were used in the modeling process, universal
statistical metrics were needed to evaluate the performance
of the models developed individually. The results were
harmonized by using sensitivity (percentage of class-1
compounds predicted correctly), specificity (percentage of
class-2 compounds predicted correctly), and CCR (correct
classification rate or balanced accuracy). These parameters
are defined as follows:
123

J Comput Aided Mol Des (2014) 28:631–646
635
Fig. 2 Chemical structure
space of chemopreventive agent
database (n = 405) using top 3
principal components of MOE
descriptors
separation and purification. Precoated silica gel plates
(Kieselgel 60, F254, 0.25 mm) were used for Thin layer
chromatography (TLC) analysis. All reagents and solvents
were purchased from Aldrich, Fisher, VWR, and other
vendors. Some chemicals were used after purification, and
others were used as purchased.
kept these outliers, since they were only a small portion
(*1 %) of the whole dataset. Furthermore, removing the
outliers did not improve the resulting models (data not
shown).
Chemopreventive activity endpoint
All 405 compounds were tested in the EBV-EA assay at
four different doses (10, 100, 500 and 1,000 times the TPA
dose) and the chemopreventive activity reported as relative
percentage induction of TPA-mediated EBV-EA activation
(Fig. 3a). Most of the compounds showed significant
activity at high dose levels, but half of the compounds
exhibited no activity at the lowest dose. We next applied a
method previously developed in our laboratory and suc-
cessfully applied in prior studies in which the multi-dose
response data for each compound was converted into a
meaningful endpoint that could be used for modeling
purposes [52]. Figure 3b shows the transformed Log-
CurveP results based on the original four dose induction
response data from Fig. 3a. Noticeably, a clear threshold
(LogCurveP = 1.25) was present (see middle of Fig. 3b),
which could be used to distinguish ‘‘actives’’ and ‘‘mar-
ginal actives’’. It should be emphasized that the definition
of these two categories is somewhat arbitrary, since most of
the compounds showed significant activity in the high dose
testing. However, this strategy gave us a criterion to dif-
ferentiate the compounds that are likely to have high effi-
cacy from the remaining compounds. On this basis, the
actual chemopreventive agent modeling database contained
204 ‘‘actives’’ and 201 ‘‘marginal actives’’.
Results and discussions
The overview of our chemopreventive agent database
We analyzed the structural similarities between the com-
pounds in the dataset by performing a Principal component
analysis (PCA) on the chemical descriptors. After gener-
ating the principal components using the 186 MOE
descriptors for all of the compounds in the database, we
selected the top three most important components to create
a 3D plot (Fig. 2) for all 405 compounds. Considering the
186 MOE descriptors that we used, these three principal
components captured around 30 % of the variance in our
database. In this way, we could visualize the chemical
similarity between modeling set compounds in this 3D plot
(Fig. 2). According to this analysis, not surprisingly, many
compounds were chemically similar, since they are deriv-
atives of several known chemopreventive agents (e.g.,
curcumin). But there were several structural outliers that
were dissimilar to the majority of the compounds. Some
previous studies showed that removing structural outliers
before the modeling process was beneficial to the results of
the QSAR models [55, 59, 60]. However, in our study, we
123

636
J Comput Aided Mol Des (2014) 28:631–646
models. The average CCR values obtained from five-fold
cross validation for all four individual models with ran-
domized classes were around 0.5, indicating that randomi-
zation of the classifications did not result in meaningful
models. In addition, we used Pearson’s Chi squared test to
calculate the v2 and p values for the prediction results
obtained using actual and randomized classes [61]. The
improvement achieved by our real QSAR models, compared
with those obtained by randomized classes, was statistically
significant (v2[30 and p \ 0.0001).
Figure 4 shows the Receiver Operating Characteristic
(ROC) of all four individual models. The Area under the
curve (AUC) is another metric to evaluate the performance
of each model. The RF models (AUC = 0.83 and 0.80),
either with Dragon and MOE descriptors, were superior to
the SVM models (AUC = 0.72 and 0.68).
[10]
[1000]
[100]
[500]
(a)
100
90
80
70
60
50
40
30
20
10
0
0
25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 425
Compound Reference Number (1-405)
(b)
2.5
Furthermore, we applied Consensus prediction thresh-
olds (CPT), as cited in one of our previous studies [62], to
the prediction results. Since all of the prediction values
from the individual models ranged from 0 to 1, we initially
used the 0.5 value as a single threshold to distinguish
compounds predicted as class-1 (CPT C0.5) or class-2
(CPT\0.5). However, as shown in Fig. 4, the use of more
restrictive thresholds improved the predictivities of all
models. Consequently, compounds with CPT values
around 0.5 should be considered ‘‘inconclusive’’. Based on
the results in Fig. 4, we removed these inconclusive pre-
dictions by using two arbitrary, but reasonable, CPT
thresholds to classify compounds as actives (CPT [0.7)
and marginal actives (CPT \0.3).
2
1.5
1
0.5
0
Compounds
Fig. 3 The chemopreventive data obtained from the EBV-EA assays:
a the original data shown as the distribution of relative induction of
TPA-mediated EBV-EA activations at four different doses [10 (blue),
100 (green), 500 (purple) and 1,000 (red) mol ratio per 32 pmol
TPA]; b the transformed LogCurveP results based on the four dose
testing data (red line shows the active/marginally active threshold at
LogCurvP = 1.25)
The application of CPT to define the prediction results
together with the removal of ‘‘inconclusive’’ compounds
clearly enhanced the predictivity of all models, including the
consensus model (Table 1). For example, the sensitivity,
specificity, and CCR metrics of the consensus model
increased to 83, 82, and 82 %, respectively (Table 1). How-
ever, the tradeoff was to decrease the coverage of this model
from 100 to 46 %. In addition, the coverage of the individual
models ranged from 45 to 58 % after applying CPT and
excluding inconclusive compounds. Since we expect that the
models developed in this project will be used to screen large
chemical libraries and prioritize a small portion of ‘‘hits’’ for
experimental validation, we feel that it is reasonable to sac-
rifice prediction coverage to increase predictivity.
Modeling results
We developed four individual and one consensus model for
the 405 compounds (204 actives and 201 marginal actives).
The fivefold external cross validation results for all of the
models are shown in Table 1. The sensitivity, specificity,
and CCR metrics for the four individual models ranged
from 57 to 75, 61 to 74, and 59 to 74 %, respectively. The
SVM-MOE model had the lowest predictivity
(CCR = 59 %), and the RF-DRG model had the highest
predictivity (CCR = 74 %). The consensus model showed
equivalent statistics with sensitivity, specificity, and CCR
all equal to 69 %.
Virtual screening
Once we developed and validated our predictive QSAR
models, they could then be used to screen new compounds
for chemopreventive activity. Since chemopreventive
agents usually must be administrated for a long period of
time, low toxicity and fewer side effects are essential
factors in the design of new agents. Therefore, we used the
Y-randomization tests were also performed for the mod-
eling set. After five time random assignments of class-1 or -2
to the 405 compounds, we developed four individual QSAR
123

J Comput Aided Mol Des (2014) 28:631–646
637
Table 1 The results of fivefold
cross validation
No CPT
CPT applied
Sensitivity
Sensitivity
Specificity
CCR
Specificity
CCR
RF-MOE
0.70
0.57
0.75
0.66
0.69
0.71
0.61
0.74
0.63
0.69
0.70
0.59
0.74
0.65
0.69
0.82
0.60
0.82
0.80
0.83
0.80
0.68
0.80
0.74
0.82
0.81
0.64
0.81
0.77
0.82
SVM-MOE
RF-DRG
SVM-MOE
Consensus
Fig. 4 ROC curves obtained as
a result of five-fold cross
validation: a two models using
Dragon descriptors; b two
models using MOE descriptors
ZINC natural derivatives (ZND) library that contains over
23,000 natural product molecules and their derivatives for
screening purposes [63]. The original ZND database was
curated to remove duplicates, including compounds that
overlapped with our existing dataset as well as compounds
that could not be handled by our program. This process
resulted in a total of over 23,385 unique compounds
available for virtual screening. Next, we evaluated these
compounds with all four individual models to prioritize and
choose hits. We prioritized those compounds that were
calculated as ‘‘active’’ and excluded those compounds that
were predicted as ‘‘inconclusive’’ (prediction values
between 0.3 and 0.7) based on the consensus predictions of
all four individual models. As another selection criterion,
we produced a combined score by summation of all indi-
vidual model prediction values. By applying both selection
criteria, we ultimately selected 148 compounds from the
ZND library. These compounds were predicted to be active
hits by all four models (individual prediction values were
all above 0.7), as well as had the highest combined scores
based on summations of all four predictions. For compar-
ison purposes, we intentionally selected 45 compounds that
were predicted to be ‘‘marginal active’’ by all models using
the same strategy as for the active hits.
from the active lead set and 6 compounds from the marginal
active set. Our selection rationale included chemical synthesis
capability and availability, as well as the SAR (structure–
activity relationship) profile from previously reported litera-
ture on chemopreventive agents. For instance, favored struc-
tural groups in most of the active predictions from our virtual
screening included phenolic groups and a biphenyl moiety
with a conjugated carbonyl system and/or a 4H-chromen-4-
one. These chemical structures are common features of some
known chemopreventive agents, such as curcumin derivatives
andflavonoids[64–66].Selected‘‘active’’compounds1–4 are
structural mimics of curcuminoids, while compounds 6 and
8–10 belong to the flavonoid chemical class.
We next synthesized the 18 chosen compounds. Com-
pounds 1, 3, and 13–16 were prepared by reaction of an
appropriately substituted benzoic acid (compound 16) or
cinnamic acid (compounds 1, 3, and 13–15) with an
appropriate amine in the presence of the coupling reagent
EDCI hydrochloride and the catalyst DAMP (Scheme 1).
Compounds 2 and 4 were obtained by subsequent
demethylation of 1 and 3 with BBr3 at low temperature.
Compound 5 was synthesized by reaction of 1-(2-hydroxy-
5-methoxyphenyl)ethanone with 2,3-dimethoxybenzalde-
hyde. Treatment of 5 with sodium acetate in aqueous eth-
anol and heating to reflux gave cyclized compound 5a,
which underwent demethylation with BBr3 in methylene
chloride yielding compound 6 (Scheme 2). Unexpectedly,
the ring-opened product 7 was also obtained during the
demethylation process, probably due to the instability of
the 2H-pyran-4(3H)-one moiety under the reaction
Chemical synthesis and experimental validation
by in vitro EBV-EA inhibition
From the 148 class-1 structures and 45 class-2 structures
predicted by virtual screening, we selected 12 compounds
123

638
J Comput Aided Mol Des (2014) 28:631–646
Scheme 1 Synthesis of acrylamide substituted compounds 1–4 and 13–16
conditions (Scheme 2). Compound 17 was prepared by
reaction of naphthalen-1-ol with methyl 2-chloroacetate in
the presence of potassium carbonate. Compound 18 was
afforded by demethylation of 17 with trimethylstannanol
(Scheme 3). Compounds 8-11 were synthesized by heating
1-(2-hydroxy-5-methoxyphenyl)ethanone and 3,4-dim-
ethoxybenzoyl chloride in pyridine (Scheme 4). After
treatment of the resulting compound with potassium
hydroxide followed by acidification, the cyclized com-
pound 8 was obtained. Selective demethylation of com-
pound 8 afforded compounds 9–11.
generally showed more potent inhibition of EBV-EA
expression than those derived from the class-2 set (13–18).
Based on our definition of class-1 and -2 as mentioned above,
all predicted ‘‘marginal active’’ compounds (compounds
13–18) were experimentally proved to be correctly predicted.
Among all actives (compounds 1–12), compounds 5–12 were
True positives (TP); however, compounds 1–4 were False
positives (FP). Structurally, compounds 1–4 are derived from
curcuminoids. However, in comparison with curcumin,
compounds 1–4 were weaker inhibitors in the validation
assay, especially at the higher concentration levels (Table 2).
Thus, the high activity of curcumin derivatives in our mod-
eling set was the major reason for these FP predictions. This
result also provided us with important information on revis-
ing/optimizing curcumin derivatives as chemoprevention
agents. In summary, the experimental validation showed
67 %sensitivity, 100 % specificity, and83 % CCR. Although
the number of experimentally tested compounds was not
great, theexperimentallyvalidatedresultsclearlydemonstrate
thatthenewlydevelopedcheminformaticsmodelscanbeused
All 18 synthesized compounds were evaluated for che-
moprevention activities measured as inhibition of TPA-
induced EBV-EA expression in Raji cells. These 18 com-
pounds and their relevant response data are shown in Table 2.
The value corresponding to each compound indicates a rela-
tiveratiotothepositive control TPA onactivationofEBV-EA
expression in Raji cells. Unsurprisingly, compounds derived
from the predicted class-1 set (compounds 1–12), especially
compounds 8–12 with a flavonoid structural scaffold,
123

J Comput Aided Mol Des (2014) 28:631–646
639
Scheme 2 Synthesis of flavonone 6 and its open rings compounds 5 and 7
Scheme 3 Synthesis of naphthalen-1-yloxyl compounds 17 and 18
Scheme 4 Synthesis of flavones 9–11
to screen new chemical libraries and prioritize novel hits for
future development.
and compared them to those existing in the modeling set.
To this end, chemical scaffolds were generated and com-
pared for the 18 new compounds against the 408-com-
pound dataset. All compounds were reduced to core
fragments (or ‘‘scaffolds’’) based on the method reported in
a previous study [67]. Eight unique scaffolds were gener-
ated out of the 18 compounds (Fig. 5), and 167 unique
scaffolds out of the 408 compounds. By comparison, 50 %
of the prior scaffolds (4 out of 8) were novel and did not
exist in the 408-compound dataset (Fig. 5).
Further Structure–Activity Relationship profiles result-
ing from this study indicated that phenolic substitution in
the molecule enhanced the EBV-EA inhibition ability.
Flavonoids 9–12 with multiple phenolic hydroxyl groups
displayed 100 % inhibition at the highest tested concen-
tration and 24–28 % inhibition even at 1 9 102 mol ratio
to TPA (32 pmol). Compound 9 was the most potent
analog among the tested compounds showing significant
inhibition even at concentrations as low as 1 9 10 mol
ratio to TPA. Interestingly, compound 7, a ring-opened
analog bearing four phenolic hydroxy groups, was less
potent than its ring-closed analog 9 against EBV-EA acti-
vation. These results suggested that the flavonoid skeleton
is essential for the inhibition activity and phenolic groups
enhance the inhibition potency.
In summary, we employed a Combi-QSAR workflow to
develop predictive models for a database consisting of 405
chemopreventive agents, which were all tested by EBV-EA
assay. We used our in-house tool to define a chemopreven-
tion activity endpoint that was suitable for modeling pur-
poses. The resulting four individual models were validated
by a five-fold cross validation procedure. The consensus
prediction showed superior performance compared with that
of the individual models. For this reason, we used all four
To evaluate the novelty of the 18 new compounds, we
analyzed the major chemical features of the compounds
123

640
J Comput Aided Mol Des (2014) 28:631–646
Table 2 Novel chemopreventive agents identified by virtual screening and their experimental EBV-EA inhibition activities
Compounds
Responses in different concentration (nM) (folds
of compound mol/TPA mol)a
LogCurvPb Pred. Act.
Exp. Act.
32 (1,000)
16 (500)
3.2 (100)
0.32 (10)
11.4 0.5 ([60)c 47.3 1.6 79.3 2.5 100 0.4 1.04
Active
Active
Mariginal
active
1
10.3 0.6 ([60) 45.3 1.6 78.1 2.4 100 0.4 1.04
Marginal
active
2
3
4
5
13.4 0.6 ([60) 48.9 1.6 80.3 2.5 100 0.3 1.04
11.6 0.5 ([60) 46.1 1.6 78.3 2.5 100 0.3 1.04
6.3 0.4 ([60) 42.5 1.4 74.6 2.3 100 0.5 1.43
Active
Active
Active
Marginal
active
Marginal
active
Active
Active
3.1 0.4 ([60) 41.3 1.3 72.8 2.3 98.7 0.6 1.43
Active
6
7
4.6 0.4 ([60) 40.3 1.4 73.1 2.4 98.6 0.5 1.43
Active
Active
Active
Active
3.2 0.4 ([60) 32.6 1.3 67.2 2.3 96.5 0.6 1.43
8
0
0.3 ([60) 26.4 1.3 62.5 2.3 89.1 0.7 1.43
Active
Active
9
0
0
0.3 ([60) 29.5 1.4 65.8 2.4 94.6 0.7 1.43
Active
Active
Active
Active
10
11
0.4 ([60) 28.9 1.4 64.3 2.5 92.4 0.6 1.43
123

J Comput Aided Mol Des (2014) 28:631–646
641
Table 2 continued
Compounds
Responses in different concentration (nM) (folds
of compound mol/TPA mol)a
LogCurvPb Pred. Act.
Exp. Act.
Active
32 (1,000)
16 (500)
3.2 (100)
0.32 (10)
0
0.4 ([60) 29.0 1.6 62.1 2.5 93.3 0.6 1.43
Active
12
13
13.1 0.6 ([60) 48.6 1.6
9.3 2.6 100 0.3 1.04
Marginal
active
Marginal
active
17.8 0.5 ([60) 51.6 1.5 83.0 2.4 100 0.4 0.85
Marginal
active
Marginal
active
14
15
16
15.9 0.5 ([60) 50.0 1.5 81.3 2.4 100 0.3 1.04
18.3 0.7 ([60) 53.5 1.4 84.3 2.6 100 0.2 0.85
11.3 0.5 ([60) 43.5 1.5 77.3 2.5 100 0.4 1.04
Marginal
active
Marginal
active
Marginal
active
Marginal
active
Marginal
active
Marginal
active
17
9.9 0.5 ([60) 41.3 1.5 75.1 2.5 100 0.4 1.04
Marginal
active
Marginal
active
18
0
0.5 ([60) 21.1 1.1 80.1 2.4 100 0.2 1.43
–
Active
Curcumin
a
TPA concentration is 32 pmol/mL
b
logCureP represents a log10 transformed response fingerprint values, as described in reference 45
c
Values in parentheses represent viability percentages of Raji cells. For the determination of cytotoxicity, the cell viability is required greater
than 60 %
individual models to virtually screen a large chemical
library. Eighteen ‘‘hits’’, 12 from a class-1 ‘‘active’’ set and 6
from a class-2 ‘‘marginal active’’ set, were finally selected,
synthesized, and then validated by the same EBV-EA assay.
The validation results indicated that the compounds derived
from the ‘‘active’’ prediction were more potent EBV-EA
activation inhibitors than the compounds derived from the
‘‘marginal active’’ prediction. Both the cross validation and
experimental validation results showed that our developed
models are suitable for designing novel chemopreventive
agents by prioritizing novel molecules for experimental
testing and further development.
Experiments
Chemical synthesis
General synthesis procedures for compounds 1–4
and 13–16
Substituted cinnamic acid (1 eq., for compounds 1,3,
13–15,) or benzoic acid (1 eq., for compound 16) was
dissolved in DMF. EDCI hydrochloride (1.5 eq.) and 10 %
(mol ratio) of DMAP were added. After being stirred at r.t.
for 30 min. an appropriate amine (1.5 eq.) was added. The
123

642
J Comput Aided Mol Des (2014) 28:631–646
Fig. 5 The major chemical scaffolds within the 18 new compounds. The four red ones are new scaffolds and the four black ones are within the
scaffolds of the modeling set compounds
(E)-N-(3-hydroxybenzyl)-3-(3-hydroxyphenyl)acrylamide (4)
resulting mixture was stirred at r.t. overnight. The solid was
removed by filtration and the filtrate was concentrated
under vacuum. The residue was partitioned in EtOAc and
water, and the organic portion was washed twice with
water. After drying over Na2SO4, the crude product was
purified by column chromatograph through a combiflash
system with hexanes/EtOAc as eluent.
Off-white crystalline solid; 1H NMR (400 MHz, CDCl3): d
7.47 (d, J = 15.6 Hz, 1H), 7.19–7.10 (m, aromatic-H, 2H),
7.01 (d, J = 7.2 Hz, 1H), 6.95 (s, aromatic-H, 1H),
6.78–6.73 (m, aromatic-H, 3H), 6.66–6.64 (m, aromatic-H,
1H), 6.57 (d, J = 15.6 Hz, 1H), 4.40 (s, PhCH2N–, 2H);
ESI MS m/z 270.21 (M ? H)?.
(E)-3-(benzo[d] [1, 3] dioxol-5-yl)-N-(3,4-
(E)-3-(3-hydroxyphenyl)-N-(3-methoxybenzyl)-N-
dimethoxyphenethyl)acrylamide (13)
methylacrylamide (1)
Off-white crystalline solid. 1H NMR (400 MHz, CDCl3): d
7.50 (d, J = 15.2 Hz, 1H), 6.95 (s, aromatic H, 1H), 6.93
(d, J = 1.6 Hz, aromatic H, 1H), 6.80–6.71 (m, aromatic
H, 4H), 6.11 (d, J = 15.2 Hz, 1H), 5.96 (s, methylene H,
2H), 5.52 (br. NH, 1H), 3.84 (s, OCH3, 6H), 3.60 (q,
NHCH2–, 2H), 2.78 (t, J = 7.2 Hz, 2H); ESI MS
m/z 356.20 (M ? H)?.
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
7.75 (d, J = 15.6 Hz, 1H), 7.29–7.02 (m, aromatic H, 4H),
6.92–6.73 (m, 4H), 6.55 (d, J = 17.2 Hz, 1H), 4.66
(d, J = 16.0 Hz, PhCH2N–, 2H), 3.82 (s, OCH3, 3H), 3.07
(s, NCH3, 3H); ESI MS m/z 298.20 (M ? H)?.
(E)-N-(3-hydroxybenzyl)-3-(3-hydroxyphenyl)-N-
methylacrylamide (2)
(E)-3-(benzo[d] [1, 3] dioxol-5-yl)-N-(4-
chlorophenethyl)acrylamide (14)
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
1
7.65 (d, J = 15.6 Hz, 1H), 7.24–6.69 (m, aromatic H, 7H),
6.02 (d, J = 17.2 Hz, 1H), 5.82 (d, J = 17.2 Hz, 1H), 4.62
(d, J = 13.6 Hz, PhCH2N–, 2H), 3.02 (s, NCH3, 3H); ESI
MS m/z 283.91 (M ? H)?.
White crystalline solid. H NMR (400 MHz, CDCl3): d
7.50 (d, J = 15.2 Hz, 1H), 7.25 (d, J = 8.4 Hz, aromatic
H, 2H), 7.12 (d, J = 8.4 Hz, aromatic H, 2H), 6.95
(s, aromatic H, 1H), 6.93 (d, J = 2.0 Hz, aromatic H, 1H),
6.76 (dd, J = 1.2, 7.6 Hz, aromatic H, 1H), 6.11 (d,
J = 15.2 Hz, 1H), 5.95 (s, methylene H, 2H), 5.52 (br. NH,
1H), 3.60 (q, J = 6.8 Hz, NHCH2–, 2H), 2.83
(t, J = 7.2 Hz, 2H); ESI MS m/z 330.15 (M ? H)?.
(E)-N-(3-methoxybenzyl)-3-(3-methoxyphenyl)acrylamide
(3)
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
(E)-3-(benzo[d] [1, 3] dioxol-5-yl)-N-(3-
7.61 (d, J = 15.6 Hz, 1H), 7.27–7.22 (m, aromatic-H, 2H),
7.07 (d, J = 7.6 Hz, 1H), 6.99 (s, aromatic-H, 1H),
6.99–6.79 (m, aromatic-H, 4H), 6.38 (d, J = 15.6 Hz, 1H),
5.96 (s, br, –NH–, 1H), 4.52 (d, J = 6.0 Hz, PhCH2N–,
2H), 3.79, 3.78 (s, OCH3 9 2, 3H each); ESI MS
m/z 298.20 (M ? H)?.
phenylpropyl)acrylamide (15)
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
7.47 (d, J = 15.6 Hz, 1H), 7.28–7.23 (m, aromatic H, 2H),
7.18–7.16 (m, aromatic H, 3H), 6.95 (d, J = 0.8 Hz, aro-
matic H, 1H), 6.93 (d, J = 1.2 Hz, aromatic H, 1H), 6.76
123

J Comput Aided Mol Des (2014) 28:631–646
643
2-(2,3-dihydroxyphenyl)-6-methoxychroman-4-one (6)
(d, J = 8.4 Hz, aromatic H, 1H), 6.11 (d, J = 15.6 Hz,
1H), 5.96 (s, methylene H, 2H), 5.52 (br. NH, 1H), 3.39 (q,
J = 6.8 Hz, NHCH2–, 2H), 2.70 (t, J = 7.6 Hz, 2H), 1.88
(pent, J = 7.2 Hz, 2H); ESI MS m/z 310.21 (M ? H)?.
Light yellow solid. 1H NMR (400 MHz, CDCl3): d 7.30 (d,
J = 3.2 Hz, aromatic H, 1H), 7.14 (dd, J = 3.2, 8.0 Hz,
aromatic H, 1H), 6.99 (d, J = 8.0 Hz, aromatic H, 1H),
6.97 (d, J = 5.6 Hz, aromatic H, 1H), 6.74 (td, J = 2.0,
8.0 Hz, aromatic H, 1H), 6.70 (t, J = 8.0 Hz,aromatic H,
1H), 5.73 (dd, J = 3.2, 12.0 Hz, 1H), 3.77 (s, OCH3, 3H),
2.97 (dd, J = 12.8, 30.0 Hz, 1H), 2.84 (dd, J = 3.2,
20.0 Hz, 1H); ESI MS m/z 285.21 (M ? H)?.
(E)-3-(benzo[d] [1, 3] dioxol-5-yl)-N-(3-
(dimethylamino)propyl)acrylamide (16)
1
Light yellow solid. H NMR (400 MHz, CDCl3): d 8.83
(br. NH, 1H), 6.93 (dd, J = 2.0, 9.2 Hz, aromatic H, 1H),
6.88 (s, aromatic H, 1H), 6.81 (dd, J = 2.0, 9.2 Hz, aro-
matic H, 1H), 5.99 (s, methylene H, 2H), 3.33 (q,
J = 6.8 Hz, NHCH2–, 2H), 2.42 (t, J = 7.6 Hz, 2H), 2.27
(s, N(CH3)2, 6H), 1.75 (pent, J = 7.2 Hz, 2H); ESI MS m/
z 251.24 (M ? H)?.
(E)-3-(2,3-dihydroxyphenyl)-1-(2,5-dihydroxyphenyl)prop-
2-en-1-one (7)
1
Yellow crystalline solid. H NMR (400 MHz, CDCl3): d
8.14 (d, J = 15.6 Hz, 1H), 7.84 (d, J = 15.6 Hz, 1H), 7.36
(d, J = 2.8 Hz, aromatic H, 1H), 7.11 (dd, J = 1.2, 8.0 Hz,
aromatic H, 1H), 6.99 (dd, J = 2.8, 8.0 Hz, aromatic H,
1H), 6.83 (dd, J = 1.6, 8.0 Hz, aromatic H, 1H), 6.79 (d,
J = 8.8 Hz, aromatic H, 1H), 6.70 (t, J = 8.0 Hz, aromatic
H, 1H); ESI MS m/z 273.21 (M ? H)?.
Syntheses of compounds 5–7
To a solution of 1-(2-hydroxy-5-methoxyphenyl)ethanone
(3 mmol) in 10 mL of EtOH was added 2,3-dimethoxy-
benzaldehyde (1.05 eq.). 20 % KOH (5 mL aq.) was added
dropwise. The resulting red mixture was stirred at r.t for
5 h with TLC monitoring, and then was poured into ice
water, acidified with 2 N HCl to pH 2, and extracted with
EtOAc. After purification through a combiflash column
chromatography system, compound 5 was obtained.
Compound 5 (0.6 mmol) was dissolved in 5 mL of EtOH
and NaOAc (10 eq.) was added. The mixture was heated to
reflux until the reaction was complete (approximately
24 h). The reaction mixture was poured into ice water and
extracted with EtOAc. The crude product was purified by
column chromatography on a combiflash system with
hexanes/EtOAc as eluent to yield 5a. Compound 5a
(0.545 mmol) was dissolved in 20 mL of anhydrous
methylene chloride (CH2Cl2) and cooled to -78 °C. BBr3
(1 M in CH2Cl2, 4.5 eq.) was added slowly. The resulting
mixture was stirred at -78 °C for 10 min, 0 °C for 10 min,
and r.t. for 3 h with TLC monitoring. The reaction mixture
cooled in an ice-bath then pured into ice water with stir-
ring. The mixture was extracted with ethyl ether (Et2O)
three times and dried over Na2SO4. The desired products 5
and 6 were obtained after purification over a combiflash
system with CH2Cl2/MeOH as eluent.
Syntheses of compounds 8–10
To a solution of 1-(2-hydroxy-5-methoxyphenyl)ethanone
(1.84 mmol) in pyridine (3 mL) was added 3,4-dim-
ethoxybenzoyl chloride (3 eq.). The resulting mixture was
heated to reflux for 1 h. After cooling to r.t., the reaction
mixture was pured into ice water with stirring, and solids
started to precipitate. After storage in a refrigerater over-
night, the off-white solid product (2-acetyl-4-methoxy-
phenyl 3,4-dimethoxybenzoate) was collected by filtration
and dried in vacuo (0.4 g). The resulting compound
(0.9 mmol) was further dissolved in 1 mL of pyridine, and
86 mg of KOH powder was added with stirring. The
mixture was then heated at 50 °C for 1 h, then poured into
10 % H2SO4 (8 mL). A light brown solid precipitated. The
solid was filtered and dissolved in 1.5 mL of EtOH con-
taining 0.1 mL of H2SO4. The solution was heated to reflux
for 1 h followed by alkalinization to pH 10 with 20 %
NaOH and refluxed for another 15 min. After cooling, the
solution was neutralized with 10 % H2SO4 to give a solid,
which was recrystallied from MeOH to afford compound 8
(0.121 g) as a dark gray crystalline solid. Compound 8
(0.096 mmol) was dissolved in 5 mL of anhydrous CH2Cl2
and cooled to -78 °C. Then, BBr3 (1 M in CH2Cl2, 4.5
eq.) was added slowly. The resulting mixture was stirred at
-78 °C for 10 min, 0 °C for 10 min, and r.t. for 2 h with
TLC monitoring. The reaction mixture then was cooled in
an ice-bath and poured into ice water. After stirring for
0.5 h, the mixture was extracted with Et2O and dried over
Na2SO4. The desired products 9–11 were obtained after
(E)-3-(2,3-dimethoxyphenyl)-1-(2-hydroxy-5-
methoxyphenyl)prop-2-en-1-one (5)
1
Yellow crystalline solid. H NMR (400 MHz, CDCl3): d
8.15 (d, J = 16.0 Hz, 1H), 7.68 (d, J = 16.0 Hz, 1H), 7.33
(d, J = 2.8 Hz, aromatic H, 1H), 7.24 (d, J = 8.0 Hz,
aromatic H, 1H), 7.13–7.07 (m, aromatic H, 2H), 6.98–6.94
(m, aromatic H, 2H), 3.89 (s, OCH3, 3H), 3.88 (s, OCH3,
3H), 3.80 (s, OCH3, 3H); ESI MS m/z 315.21 (M ? H)?.
123

644
J Comput Aided Mol Des (2014) 28:631–646
column chromatography over a combiflash system with
CH2Cl2/MeOH as eluent.
trimethylstannanol (10 eq.) in dichloroethane (1.5 mL). The
resulting mixture was heated at 80 °C for 3 h with TLC
monitoring. Upon completion, the solvent was evaporated
and the residue was diluted with EtOAc and washed with
5 % HCl followed by brine three times. After drying ove
Na2SO4, the crude product was purified through a combiflash
chromatography system with hexanes/EtOAc as eluent. The
desired product 18 was obtained as a off-white solid.
2-(3,4-dimethoxyphenyl)-6-methoxy-4H-chromen-4-one (8)
Yellow–brown crystalline solid. 1H NMR (400 MHz,
CDCl3): d 7.67-7.64 (m, aromatic-H 2H), 7.54–7.52 (m,
aromatic-H 2H), 7.37 (dd, J = 3.2, 9.2 Hz, aromatic H,
1H), 7.11 (d, J = 8.4 Hz, aromatic H, 1H), 6.82 (s, 1H),
3.92, 3.89, 3.88 (s, OCH3 9 3, 3H each); ESI MS
m/z 313.23 (M ? H)?.
Methyl 2-(naphthalen-1-yloxy)acetate (17)
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
2-(3,4-dihydroxyphenyl)-6-hydroxy-4H-chromen-4-one (9)
8.36-8.33 (m, aromatic-H 1H), 7.80–7.77 (m, aromatic-H,
1H), 7.50–7.45 (m, aromatic H, 3H), 7.32 (t, J = 7.6 Hz,
aromatic-H, 1H), 6.68 (d, J = 7.6 Hz, aromatic-H, 1H)
3.80 (s, OCH3, 3H); ESI MS m/z 217.21 (M ? H)?.
Yellow–brown crystalline solid. 1H NMR (400 MHz,
CDCl3): d 7.53 (d, J = 8.8 Hz, aromatic-H 1H), 7.41-7.37
(m, aromatic-H, 3H), 7.23 (dd, J = 2.8, 8.8 Hz, aromatic
H, 1H), 6.88 (d, J = 8.8 Hz, aromatic H, 1H), 6.65 (s, 1H);
ESI MS m/z 270.21 (M ? H)?.
2-(naphthalen-1-yloxy)acetic acid (18)
1
White crystalline solid. H NMR (400 MHz, CDCl3): d
2-(4-hydroxy-3-methoxyphenyl)-6-methoxy-4H-chromen-4-
8.30–8.27 (m, aromatic-H 1H), 7.77–7.75 (m, aromatic-H,
1H), 7.46–7.41 (m, aromatic H, 3H), 7.32 (t, J = 7.6 Hz,
aromatic-H, 1H), 6.79 (d, J = 7.6 Hz, aromatic-H, 1H)
3.80 (s, OCH3, 3H); ESI MS m/z 203.19 (M ? H)?.
one (10)
1
Yellow crystalline solid. H NMR (400 MHz, CDCl3): d
7.62 (d, J = 9.2 Hz, aromatic-H 1H), 7.52 (dd, J = 2.4,
8.4 Hz, aromatic-H, 2H), 7.43 (d, J = 2.4 Hz, aromatic H,
1H), 6.88 (dd, J = 2.8, 8.8 Hz, aromatic-H, 1H), 7.06 (d,
J = 8.8 Hz, aromatic-H, 1H), 6.74 (s, 1H), 3.91, 3.88 (s,
OCH3 9 2, 3H each); ESI MS m/z 299.21 (M ? H)?.
Acknowledgments We thank Kimberlee Moran, the Director of the
Center for Forensic Science Research & Education, for her help with
the manuscript preparation for the entire project. We also appreciate
the help of Dr. Susan Morris-Natschke, Natural Product Research
Laboratory, UNC Eshelman School of Pharmacy with proofreading
and editing the manuscript. This investigation was also supported in
part by NIH Grant CA 177584-01 from National Cancer Institute
awarded to K.H. Lee. This study was also supported in part by the
Taiwan Department of Health, China Medical University Hospital
Cancer Research Center of Excellence (DOH100-TD-C-111-005).
2-(3,4-dihydroxyphenyl)-6-methoxy-4H-chromen-4-one (11)
Yellow–brown crystalline solid. 1H NMR (400 MHz,
CDCl3): d 7.59 (d, J = 9.2 Hz, aromatic-H, 1H), 7.50 (d,
J = 2.8 Hz, aromatic-H, 1H), 7.41 (t, J = 2.4 Hz, aro-
matic-H, 1H), 7.39 (s, aromatic-H, 1H), 7.35 (dd, J = 3.2,
9.2 Hz, aromatic-H, 1H), 6.88 (dd, J = 1.2, 7.6 Hz, aro-
matic H, 1H), 6.69 (s, 1H), 3.87 (s, OCH3, 3H); ESI MS
m/z 285.19 (M ? H)?.
References
1. Amercian Cancer Society. Lifetime risk of developing or dying
from cancer. http://www.cancer.org/research/cancerfactsfigures/
cancerfactsfigures/cancer-facts-figures-2013. 11-14-2013. 2-1-
0014
Syntheses of compounds 17–18
2. Maduro JH, Pras E, Willemse PH, de Vries EG (2003) Acute and
long-term toxicity following radiotherapy alone or in combina-
tion with chemotherapy for locally advanced cervical cancer.
Cancer Treat Rev 29:471–488
3. Kelloff GJ, Sigman CC, Greenwald P (1999) Cancer chemopre-
vention: progress and promise. Eur J Cancer 35:2031–2038
4. Sharma D, Sukumar S (2013) Big punches come in nanosizes for
chemoprevention. Cancer Prev Res 6:1007–1010
5. Tsao AS, Kim ES, Hong WK (2004) Chemoprevention of cancer.
Ca-A Cancer J Clin 54:150–180
6. Takemura H, Sakakibara H, Yamazaki S, Shimoi K (2013) Breast
cancer and flavonoids—a role in prevention. Curr Pharm Des
19:6125–6132
7. Steward WP, Brown K (2013) Cancer chemoprevention: a rapidly
evolving field. Br J Cancer 109:1–7
To a solution of naphthalen-1-ol (1 mmol) in acetone
(10 mL) was added K2CO3 (3 eq.) followed by methyl
2-chloroacetate (1.5 eq.). The resulting mixture was heated
to reflux for 20 h with TLC monitoring. The solid was fil-
tered and the filtrate was concentrated to dryness. The resi-
due was diluted with EtOAc and washed twice with brine.
The organic portion was dried over Na2SO4, filtered, and
concentrated. The crude product was purified through a
combiflash chromatography system with hexanes/EtOAc as
eluent to afford the desired product 17. Compound 18 was
obtained by treatment of 17 (0.1 mmol) with
123

J Comput Aided Mol Des (2014) 28:631–646
645
8. Patterson SL, Maresso KC, Hawk E (2013) Cancer chemopre-
vention: successes and failures. Clin Chem 59:94–101
9. Malik M, Magnuson BA (2004) Rapid method for identification
of chemopreventive compounds using multiplex RT-PCR for
cyclooxygenase mRNA expression. Cancer Detect Prev
28:277–282
10. Heijink DM, Fehrmann RS, de Vries EG, Koornstra JJ, Oos-
terhuis D, van der Zee AG, Kleibeuker JH, de Jong S (2011) A
bioinformatical and functional approach to identify novel strate-
gies for chemoprevention of colorectal cancer. Oncogene
30:2026–2036
protein targets for procarcinogen activating enzyme inhibition.
J Proteom Bioinfor 3:200–203
25. Akihisa T, Tokuda H, Yasukawa K, Ukiya M, Kiyota A, Sa-
kamoto N, Suzuki T, Tanabe N, Nishino H (2005) Azaphilones,
furanoisophthalides, and amino acids from the extracts of Mon-
ascus pilosus-fermented rice (red-mold rice) and their chemo-
preventive effects. J Agric Food Chem 53:562–565
26. Nakagawa-Goto K, Yamada K, Taniguchi M, Tokuda H, Lee KH
(2009) Cancer preventive agents 9. Betulinic acid derivatives as
potent cancer chemopreventive agents. Bioorg Med Chem Lett
19:3378–3381
11. Gerhauser C, Klimo K, Heiss E, Neumann I, Gamal-Eldeen A,
Knauft J, Liu GY, Sitthimonchai S, Frank N (2003) Mechanism-
based in vitro screening of potential cancer chemopreventive
agents. Mutat Res 523–524:163–172
27. Akihisa T, Tokuda H, Hasegawa D, Ukiya M, Kimura Y, Enjo F,
Suzuki T, Nishino H (2006) Chalcones and other compounds
from the exudates of Angelica keiskei and their cancer chemo-
preventive effects. J Nat Prod 69:38–42
12. Tokuda H, Arai T, Suzuki R, Strong JM, Schneider A, Suzuzki N
(2012) Efficient evaluation of healthy tea, Gromwell seed against
tumor promoting stage. Planta Medica 78:1177
13. Perestelo NR, Jimenez IA, Tokuda H, Hayashi H, Bazzocchi IL
(2010) Sesquiterpenes from Maytenus jelskii as potential cancer
chemopreventive agents. J Nat Prod 73:127–132
14. Terazawa R, Garud DR, Hamada N, Fujita Y, Itoh T, Nozawa Y,
Nakane K, Deguchi T, Koketsu M, Ito M (2010) Identification of
organoselenium compounds that possess chemopreventive prop-
erties in human prostate cancer LNCaP cells. Bioorg Med Chem
18:7001–7008
15. Stan SD, Singh SV (2009) Transcriptional repression and inhi-
bition of nuclear translocation of androgen receptor by diallyl
trisulfide in human prostate cancer cells. Clin Cancer Res
15:4895–4903
16. Kim Y, Kim J, Lee SM, Lee HA, Park S, Kim Y, Kim JH (2012)
Chemopreventive effects of Rubus coreanus Miquel on prostate
cancer. Biosci Biotechnol Biochem 76:737–744
17. Johnson JJ, Syed DN, Suh Y, Heren CR, Saleem M, Siddiqui IA,
Mukhtar H (2010) Disruption of androgen and estrogen receptor
activity in prostate cancer by a novel dietary diterpene carnosol:
implications for chemoprevention. Cancer Prev Res (Phila)
3:1112–1123
18. Steele VE, Sharma S, Mehta R, Elmore E, Redpath L, Rudd C,
Bagheri D, Sigman CC, Kelloff GJ (1996) Use of in vitro assays
to predict the efficacy of chemopreventive agents in whole ani-
mals. J Cell Biochem Suppl 26:29–53
19. Ito Y, Kawanishi M, Harayama T, Takabayashi S (1981) Com-
bined effect of the extracts from Croton tiglium, Euphorbia
lathyris or Euphorbia tirucalli and n-butyrate on Epstein–Barr
virus expression in human lymphoblastoid P3HR-1 and Raji cells.
Cancer Lett 12:175–180
28. Akihisa T, Tokuda H, Ukiya M, Iizuka M, Schneider S, Ogasa-
wara K, Mukainaka T, Iwatsuki K, Suzuki T, Nishino H (2003)
Chalcones, coumarins, and flavanones from the exudate of
Angelica keiskei and their chemopreventive effects. Cancer Lett
201:133–137
29. Sakurai N, Kozuka M, Tokuda H, Mukainaka T, Enjo F, Nishino
H, Nagai M, Sakurai Y, Lee KH (2005) Cancer preventive agents.
Part 1: chemopreventive potential of cimigenol, cimigenol-3,15-
dione, and related compounds. Bioorg Med Chem 13:1403–1408
30. Ito C, Itoigawa M, Mishina Y, Filho VC, Enjo F, Tokuda H,
Nishino H, Furukawa H (2003) Chemical constituents of Calo-
phyllum brasiliense. 2. Structure of three new coumarins and
cancer chemopreventive activity of 4-substituted coumarins.
J Nat Prod 66:368–371
31. Suzuki M, Nakagawa-Goto K, Nakamura S, Tokuda H, Morris-
Natschke SL, Kozuka M, Nishino H, Lee KH (2006) Cancer
preventive agents. Part 5. Anti-tumor-promoting effects of cou-
marins and related compounds on Epstein–Barr virus activation
and two-stage mouse skin carcinogenesis. Pharm Biol
44:178–182
32. Akihisa T, Higo N, Tokuda H, Ukiya M, Akazawa H, Tochigi Y,
Kimura Y, Suzuki T, Nishino H (2007) Cucurbitane-type trit-
erpenoids from the fruits of Momordica charantia and their
cancer chemopreventive effects. J Nat Prod 70:1233–1239
33. Kikuchi T, Akihisa T, Tokuda H, Ukiya M, Watanabe K, Nishino
H (2007) Cancer chemopreventive effects of cycloartane-type
and related triterpenoids in in vitro and in vivo models. J Nat
Prod 70:918–922
34. Ito C, Itoigawa M, Mishina Y, Tomiyasu H, Litaudon M, Cosson
JP, Mukainaka T, Tokuda H, Nishino H, Furukawa H (2001)
Cancer chemopreventive agents. New depsidones from Garcinia
plants. J Nat Prod 64:147–150
20. Ito Y, Yanase S, Fujita J, Harayama T, Takashima M, Imanaka H
(1981) A short-term in vitro assay for promoter substances using
human lymphoblastoid cells latently infected with Epstein–Barr
virus. Cancer Lett 13:29–37
21. Bertosa B, Aleksic M, Karminiski-Zamola G, Tomic S (2010)
QSAR analysis of antitumor active amides and quinolones from
thiophene series. Int J Pharm 394:106–114
22. Saeed BA, Saour KY, Elias RS, Al-Masoudi NA, La Cola P
(2010) Antitumor and quantitative structure activity relationship
study for dihydropyridones derived from curcumin. Am J Immun
6:7–10
23. Aleksic M, Bertosa B, Nhili R, Uzelac L, Jarak I, Depauw S, vid-
Cordonnier MH, Kralj M, Tomic S, Karminski-Zamola G (2012)
Novel substituted benzothiophene and thienothiophene carbox-
anilides and quinolones: synthesis, photochemical synthesis,
DNA-binding properties, antitumor evaluation and 3D-derived
QSAR analysis. J Med Chem 55:5044–5060
35. Nakagawa-Goto K, Bastow KF, Wu JH, Tokuda H, Lee KH (2005)
Total synthesis and bioactivity of unique flavone desmosdumotin B
and its analogs. Bioorg Med Chem Lett 15:3016–3019
36. Lin AS, Shibano M, Nakagawa-Goto K, Tokuda H, Itokawa H,
Morris-Natschke SL, Lee KH (2007) Cancer preventive agents. 7.
Antitumor-promoting effects of seven active flavonolignans from
milk thistle (Silybum marianum) on Epstein–Barr virus activa-
tion. Pharm Biol 45:735–738
37. Wang XH, Nakagawa-Goto K, Kozuka M, Tokuda H, Nishino H,
Lee KH (2006) Cancer preventive agents. Part 6: chemopreven-
tive potential of furanocoumarins and related compounds. Pharm
Biol 44:116–120
38. Cui W, Iwasa K, Tokuda H, Kashihara A, Mitani Y, Hasegawa T,
Nishiyama Y, Moriyasu M, Nishino H, Hanaoka M, Mukai C,
Takeda K (2006) Potential cancer chemopreventive activity of
simple isoquinolines, 1-benzylisoquinolines, and protoberberines.
Phytochemistry 67:70–79
24. Girija CR, Karunakar P, Poojari CS, Begun NS, Syed AA (2013)
Molecular docking studies of curcumin derivatives with multiple
39. Akihisa T, Takahashi A, Kikuchi T, Takagi M, Watanabe K,
Fukatsu M, Fujita Y, Banno N, Tokuda H, Yasukawa K (2011)
123

646
J Comput Aided Mol Des (2014) 28:631–646
The melanogenesis-inhibitory, anti-inflammatory, and chemo-
preventive effects of limonoids in n-hexane extract of Aza-
dirachta indica A. Juss. (neem) seeds. J Oleo Sci 60:53–59
virus activation and two-stage carcinogenesis of skin tumors.
Cancer Lett 154:101–105
54. Solimeo R, Zhang J, Kim M, Sedykh A, Zhu H (2012) Predicting
chemical ocular toxicity using a combinatorial QSAR approach.
Chem Res Toxicol 25:2763–2769
55. Zhu H, Tropsha A, Fourches D, Varnek A, Papa E, Gramatica P,
Oberg T, Dao P, Cherkasov A, Tetko IV (2008) Combinatorial
QSAR modeling of chemical toxicants tested against Tetrahy-
mena pyriformis. J Chem Inf Model 48:766–784
56. Zhang S, Wei L, Bastow K, Zheng W, Brossi A, Lee KH,
Tropsha A (2007) Antitumor agents 252. Application of validated
QSAR models to database mining: discovery of novel tylophorine
derivatives as potential anticancer agents. J Comput Aided Mol
Des 21:97–112
40. Kapadia GJ, Azuine MA, Takayasu J, Konoshima T, Takasaki M,
Nishino H, Tokuda H (2000) Inhibition of Epstein–Barr virus
early antigen activation promoted by 12-O-tetradecanoylphorbol-
13-acetate by the non-steroidal anti-inflammatory drugs. Cancer
Lett 161:221–229
41. Itoigawa M, Ito C, Tan HT, Kuchide M, Tokuda H, Nishino H,
Furukawa H (2001) Cancer chemopreventive agents, 4-phen-
ylcoumarins from Calophyllum inophyllum. Cancer Lett
169:15–19
42. Itoigawa M, Ito C, Tokuda H, Enjo F, Nishino H, Furukawa H
(2004) Cancer chemopreventive activity of phenylpropanoids and
phytoquinoids from Illicium plants. Cancer Lett 214:165–169
43. Ito C, Itoigawa M, Miyamoto Y, Onoda S, Rao KS, Mukainaka T,
Tokuda H, Nishino H, Furukawa H (2003) Polyprenylated
benzophenones from Garcinia assigu and their potential cancer
chemopreventive activities. J Nat Prod 66:206–209
57. Medina-Franco JL, Golbraikh A, Oloff S, Castillo R, Tropsha A
(2005) Quantitative structure-activity relationship analysis of
pyridinone HIV-1 reverse transcriptase inhibitors using the k
nearest neighbor method and QSAR-based database mining.
J Comput Aided Mol Des 19:229–242
44. Ito C, Itoigawa M, Otsuka T, Tokuda H, Nishino H, Furukawa H
Nat Prod
58. Shen M, Beguin C, Golbraikh A, Stables JP, Kohn H, Tropsha A
(2004) Application of predictive QSAR models to database
mining: identification and experimental validation of novel anti-
convulsant compounds. J Med Chem 47:2356–2364
59. Zhang L, Zhu H, Oprea TI, Golbraikh A, Tropsha A (2008)
QSAR modeling of the blood-brain barrier permeability for
diverse organic compounds. Pharm Res 25:1902–1914
60. Zhu H, Martin TM, Ye L, Sedykh A, Young DM, Tropsha A
(2009) Quantitative structure-activity relationship modeling of rat
acute toxicity by oral exposure. Chem Res Toxicol 22:1913–1921
61. Greenwood PE, Nikulin MS (1996) A guide to chi squared test-
ing. Wiley, New York
(2000) Constituents of Boronia pinnata.
63:1344–1348
J
45. Nakamura S, Kozuka M, Bastow KF, Tokuda H, Nishino H,
Suzuki M, Tatsuzaki J, Morris Natschke SL, Kuo SC, Lee KH
(2005) Cancer preventive agents, part 2: synthesis and evaluation
of 2-phenyl-4-quinolone and 9-oxo-9,10-dihydroacridine deriva-
tives as novel antitumor promoters. Bioorg Med Chem
13:4396–4401
46. Ito C, Itoigawa M, Kojima N, Tan HT, Takayasu J, Tokuda H,
Nishino H, Furukawa H (2004) Cancer chemopreventive activity
of rotenoids from Derris trifoliata. Planta Med 70:585–588
47. Iranshahi M, Kalategi F, Rezaee R, Shahverdi AR, Ito C,
Furukawa H, Tokuda H, Itoigawa M (2008) Cancer chemopre-
ventive activity of terpenoid coumarins from Ferula species.
Planta Med 74:147–150
62. Kim M, Sedykh A, Chakravarti S K, Saiakhov R D, Zhu H (2014)
Critical evaluation of human oral bioavailability for pharmaceu-
tical drugs by using various cheminformatics approaches. Pharm
Res 31:1002–1014
48. Akihisa T, Tabata K, Banno N, Tokuda H, Nishimura R, Na-
kamura Y, Kimura Y, Yasukawa K, Suzuki T (2006) Cancer
chemopreventive effects and cytotoxic activities of the triterpene
acids from the resin of Boswellia carteri. Biol Pharm Bull
29:1976–1979
49. Akihisa T, Kojima N, Kikuchi T, Yasukawa K, Tokuda H,
Masters T, Manosroi A, Manosroi J (2010) Anti-inflammatory
and chemopreventive effects of triterpene cinnamates and ace-
tates from shea fat. J Oleo Sci 59:273–280
50. Takasaki M, Konoshima T, Tokuda H, Masuda K, Arai Y, Shi-
ojima K, Ageta H (1999) Anti-carcinogenic activity of Taraxa-
cum plant. II. Biol Pharm Bull 22:606–610
51. Takasaki M, Konoshima T, Tokuda H, Masuda K, Arai Y, Shi-
ojima K, Ageta H (1999) Anti-carcinogenic activity of Taraxa-
cum plant. I. Biol Pharm Bull 22:602–605
52. Sedykh A, Zhu H, Tang H, Zhang L, Richard A, Rusyn I, Tropsha
A (2011) Use of in vitro HTS-derived concentration-response
data as biological descriptors improves the accuracy of QSAR
models of in vivo toxicity. Environ Health Perspect 119:364–370
53. Iwase Y, Takemura Y, Ju-ichi M, Ito C, Furukawa H, Kawaii S,
Yano M, Mou XY, Takayasu J, Tokuda H, Nishino H (2000)
Inhibitory effect of flavonoids from citrus plants on Epstein–Barr
63. Irwin JJ, Shoichet BK (2005) ZINC—a free database of com-
mercially available compounds for virtual screening. J Chem Inf
Model 45:177–182
64. Gafner S, Lee SK, Cuendet M, Barthelemy S, Vergnes L, Lab-
idalle S, Mehta RG, Boone CW, Pezzuto JM (2004) Biologic
evaluation of curcumin and structural derivatives in cancer che-
moprevention model systems. Phytochemistry 65:2849–2859
65. Shehzad A, Wahid F, Lee YS (2010) Curcumin in cancer che-
moprevention: molecular targets, pharmacokinetics, bioavail-
ability, and clinical trials. Arch Pharm (Weinheim) 343:489–499
66. Meiyanto E, Hermawan A, Anindyajati A (2012) Natural pro-
ducts for cancer-targeted therapy: citrus flavonoids as potent
chemopreventive agents. Asian Pac J Cancer Prev 13:427–436
67. Guiguemde WA, Shelat AA, Bouck D, Duffy S, Crowther GJ,
Davis PH, Smithson DC, Connelly M, Clark J, Zhu F, Jimenez-
Diaz MB, Martinez MS, Wilson EB, Tripathi AK, Gut J, Sharlow
ER, Bathurst I, El MF, Fowble JW, Forquer I, McGinley PL,
Castro S, Ngulo-Barturen I, Ferrer S, Rosenthal PJ, Derisi JL,
Sullivan DJ, Lazo JS, Roos DS, Riscoe MK, Phillips MA, Rathod
PK, Van Voorhis WC, Avery VM, Guy RK (2010) Chemical
genetics of Plasmodium falciparum. Nature 465:311–315
123