Article

pubs.acs.org/jcim

Prediction of Polypharmacological Proles of Drugs by the

Integration of Chemical, Side Eect, and Therapeutic Space
Feixiong Cheng, Weihua Li, Zengrui Wu, Xichuan Wang, Chen Zhang, Jie Li, Guixia Liu,
and Yun Tang*,

Shanghai Key Laboratory of New Drug Design, School of Pharmacy, East China University of Science and Technology, 130 Meilong
Road, Shanghai 200237, China

Department of Surgery, Shanghai MCC Hospital, 456 Chunlei Road, Shanghai 200941, China
S Supporting Information
*

ABSTRACT: Prediction of polypharmacological proles of

drugs enables us to investigate drug side eects and further
nd their new indications, i.e. drug repositioning, which could
reduce the costs while increase the productivity of drug
discovery. Here we describe a new computational framework
to predict polypharmacological proles of drugs by the
integration of chemical, side eect, and therapeutic space.
On the basis of our previous developed drug side eects
database, named MetaADEDB, a drug side eect similarity
inference (DSESI) method was developed for drugtarget interaction (DTI) prediction on a known DTI network connecting
621 approved drugs and 893 target proteins. The area under the receiver operating characteristic curve was 0.882 0.011
averaged from 100 simulated tests of 10-fold cross-validation for the DSESI method, which is comparative with drug structural
similarity inference and drug therapeutic similarity inference methods. Seven new predicted candidate target proteins for seven
approved drugs were conrmed by published experiments, with the successful hit rate more than 15.9%. Moreover, network
visualization of drugtarget interactions and o-target side eect associations provide new mechanism-of-action of three
approved antipsychotic drugs in a case study. The results indicated that the proposed methods could be helpful for prediction of
polypharmacological proles of drugs.

in about 100 000 deaths.5 In many case, ADEs are caused by

unintended activity at o-targets or drug promiscuity.6,7 Several
examples o-target toxicity included the drugdrug interactions
caused by CYP450 inhibition,8,9 cardiotoxicity caused by
inhibition of the human ether-a-go-go-related gene potassium
channel (hERG).10 Systemically identifying polypharmacological proles of drugs enable us to investigate drug side eects,7
provide new strategy for multitarget drug design,3 and further
nd their new indications, i.e. drug repositioning,11 which could
reduce the costs eectively while increase the productivity of
drug discovery. However, how to systemically identify
polypharmacological proles of drugs is a big challenge which
needs the development of a new methodology.
In the past decade, several computational methods have been
published to address the issues of predicting polypharmacological proles of drugs and drugtarget interactions
(DTIs).1220 Those methods were categorized into ligandbased, receptor-based, chemogenomics-based, and biological
network-based ones. Keiser et al. predicted new molecular
targets for known drugs using chemical two-dimensional (2D)
structural similarity method, namely similarity ensemble
approach (SEA).21,22 Twenty-three new DTI were conrmed,

INTRODUCTION
Over the past decade, the one gene, one drug, one disease
paradigm has become outdated in many cases and the concept
of network pharmacology or polypharmacology was hence
proposed for those drugs acting on multiple targets rather than
one target.1,2 Several published works demonstrated that a
single drug target could be therapeutically insucient,
especially for some complex diseases.3 For example, antipsychiatric drugs indicate multiple dierence activities against
four potential dopamine pathways in the human central
nervous systems, such as the nigrostriatal dopamine pathway,
mesolimbic dopamine pathway, mesocortical dopamine pathway, and tuberoinfundibular dopamine pathway.4 However,
multitarget drugs or drugs with polypharmacological proles
could have high risk of toxicity, such as undesirable proles of
o-target activity. Elucidation of undesirable proles of otarget activity could hinder or halt the development of
candidate drugs or even lead to market withdrawal by drug
side eects.2
Drug side eects, also known as adverse drug events (ADEs)
or adverse drug reactions, measure the harms associated with
uses of given medications at normal dosages. Every year,
thousands of people are reported to die from serious ADEs
around the world. For example, there are about two million
serious ADEs reported per year in the United States, resulting
2013 American Chemical Society

Received: January 4, 2013

Published: March 25, 2013
753

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Figure 1. Schematic diagram of computational framework in this study. (A) Construction of drugtarget interaction network from DrugBank and
therapeutic target database. (B) Development of drug side eects similarity inference (DSESI) method. (C) Application of DSESI, drug structural
similarity inference, and drug therapeutic similarity inference methods for predicting new potential drugtarget interactions.

ve of which were potent (Ki < 100 nM). Lounkine et al.

recently used the SEA approach to predict the activity of 656
marketed drugs on 73 unintended side eect targets.
Approximately half of the predictions were conrmed, either
from proprietary databases unknown to the method or by new
experimental assays. Anities for these new o-targets ranged
from 1 nM to 30 M.7 Wang et al. developed a computational
chemogenomic method and used it to discovery several new
ligands for four targets (i.e., GPR40, SIRT1, p38, and GSK-3)
by experimental assays.23 Recently, our group systemically
evaluated the performance of the multitarget quantitative
structureactivity relationships (mt-QSAR) and computational
chemogenomic methods.19 The results indicated that there is a
high false positive rate for the external validation set when using

chemogenomic methods. Meanwhile, our group reported three

supervised biological network inference methods, including
drug-based similarity inference, target-based similarity inference, and network-based inference (NBI) methods for DTI
prediction and drug repositioning. With the methods, ve
approved drugs were predicted and experimentally validated to
have novel polypharmacological proles on estrogen receptors
and dipeptidyl peptidase-IV.17 And two weighted biological
network inference methods, namely node-weighted NBI and
edge-weighted NBI, were further presented for chemical
protein interaction prediction. High performance were yielded
and a weak interaction hypothesis was found based on two
comprehensive chemicalprotein interaction databases targeting G-protein-coupled receptors (GPCRs) and kinases.18
754

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Recently, Besnard et al. described a new approach for the

automated design of ligands against proles of multiple drug
targets.3 The method is demonstrated by the evolution of an
approved acetylcholinesterase inhibitor drug into brainpenetrable ligands with either specic polypharmacology or
exquisite selectivity proles for GPCRs. Several reverse docking
strategies have been used for drug target binding anity
prediction, o-target prediction.2426 However, those methods
may not be suitable for targets whose three-dimensional (3D)
structures are unknown, such as most of GPCRs.
Interestingly, drug side eects can be seen as a clinical marker
or an important phenotypic resource, and we can take
advantage of this knowledge that, if two drugs are similar in
the side eects they elicit, they are more likely to share a
common drug target.27,28 Therefore, drug side eects provide
new dimensional space for predicting polypharmacological
proles of drugs recently.29 Campillos et al. successfully
predicted new DTI using drug side-eect similarity.27 They
tested 20 of those unexpected DTI and validated 13 implied
DTI by in vitro binding assays, of which 11 revealed inhibition
constants equal or less than 10 M. Yang and Agarwal
developed naive Bayes models to predict indications for 145
diseases using the drug side eects as features.30 They extracted
3175 side eect-disease relationships by combining the side
eectdrug relationships from drug labels and the drug
disease relationships from PharmGKB for naive Bayes model
building. The area under the receiver operating characteristic
curve (AUC) was above 0.8 in 92% of their models. A
bottleneck in Yangs work is the lack of the benchmark negative
side eectdisease relationship pairs. In our recent work,35 we
built a comprehensive drugADE association database so far,
entitled MetaADEDB [http://www.lmmd.org/online_services/
metaadedb/], which provides a comprehensive drug phenotypic resource for development of new computational methods
to predict DTI and drug side eects.
In this study, we developed a new computational framework
for predicting polypharmacological proles of drugs by the
integration of chemical, side eects, and therapeutic spaces
(Figure 1). First, a DTI network connecting 621 approved
drugs and 856 target proteins extracted from DrugBank
database31 and Therapeutic Target Database (TTD)32 was
built. The side eect similarity of each drugdrug pair was
calculated from MetaADEDB, and the drug sideeect
similarity inference (DSESI) method with high predictive
performance was developed, which is comparative with drug
structural similarity inference (DSSI) and drug therapeutic
similarity inference (DTSI) methods.

Each drug was grouped based on the Anatomical Therapeutic

Chemical Classication System (ATC) code. The structures in
SMILES format of each drug were collected from DrugBank.
Prediction of New Targets for Known Drugs. Denoting
the target protein set as T = {t1, t2, ..., tm} and the drug set as D
= {d1, d2, ..., dn}, the drugtarget (DT) binary interactions can
be described as a bipartite DT graph G(D, T, E), where E = {eij:
di D, tj T}. A link is drawn between di and tj only if the
drug di is interacted with the target tj. The DT bipartite network
can be presented by an n m adjacent matrix {aij}, where aij =
1 when di and tj interact with each other, otherwise aij = 0. For a
DT pair ditj, a linkage between di and tj is determined using
the collaborative ltering algorithm36,37 by the following
predicted score:
n

vij =

l = 1, l i S(di , dl)alj
n

l = 1, l i S(di , dl)

(1)

For dierent methods, including DSESI, DSSI, and DTSI, S(di,

dl) represents dierent similarities of drugdrug pairs as
follows:
Drug Side Eect Similarity. As shown in Figure 1B, each
drug was coded by 13 060 ADE bit vectors in MetaADEDB.
Each bit represents one ADE. If an ADE is associated with a
given drug in MetaADEDB, the corresponding bit is set to 1;
conversely, it is set to 0. Then, the side eect similarity SSE(di,
dl) between drugs di and dl was calculated by Tanimoto
coecient using the drugs ADE bit vectors.
Drug Structural Similarity. The 2D structural similarity
SD(di, dl) between drugs di and dl were calculated via the
Tanimoto similarity metric using MACCS keys, freely available
from OpenBabel v2.3.1.38
Drug Therapeutic Similarity. The ATC codes for all drugs
were collected from DrugBank database.31 The kth level drug
therapeutic similarity (Sk) between two drugs is dened via the
ATC codes as
Sk(di , dl) =

ATCk (di) ATCk (dl)

ATCk (di) ATCk (dl)

(2)

where ATCk(d) represents all ATC codes at the kth level of

drug d. A score TS(d1, d2) is used to dene the therapeutic
similarity between two drugs:
n

ST(di , dl) =

k = 1 Sk(di , dl)
n

(3)

where n represents the ve level ATC codes (ranges from 1 to

5).39 Note that some drugs have more than one ATC code, for
example, the classic drug of nicotine has four dierent ATC
codes (e.g., N07BA01, A11HA01, C04AC01, C10AD02). For a
drug with multiple ATC codes, the therapeutic similarity was
calculated for each ATC code, and then, the average therapeutic
similarity was used.
Consensus Similarity. In addition, we also investigated three
dierent consensus similarity inference methods: mean,
maximum, and geometric mean as follows:
(i) Mean

METHODS AND MATERIALS

Data Set of Drug Phenotypic Source. A comprehensive
drugADE association network, collected from our recently
developed drug adverse events database, namely MetaADEDB:
http://www.lmmd.org/online_services/metaadedb/. MetaADEDB was built using data integration from three drug side
eects associated databases, CTD,33 SIDER (version 2),34 and
OFFSIDES,28 and a text mining method.35 Only reported side
eect data with experimental evidence were used. All drugs and
ADE items in MetaADEDB were annotated using Medical
Subject Headings (MeSH) and Unied Medical Language
System (UMLS) vocabularies. And, duplicated drugADE
associations were excluded.
Data Set of DrugTarget Interactions. Drugtarget
interaction data were collected from DrugBank31 and TTD.32

SM(di , dl) =

SSE(di , dl) + SD(di , dl) + ST(di , dl)

3

(4)

(ii) Maximum
Smax(di , dl) = max{SSE(di , dl), SD(di , dl), ST(di , dl)}
755

(5)

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

(iii) Geometric Mean

SGM(di , dl) =

SSE(di , dl)SD(di , dl)ST(di , dl)

Table 1. Performance of Prioritizing New Targets for

Approved Drugs Using Dierent Methodsa
(6)

RE SD

method

Where SM(di, dl), Smax(di, dl), and SGM(di, dl) represent three
dierent combinational consensus similarities of a given drug
drug pair, SSE(di, dl), SD(di, dl), and ST(di, dl) represent side
eect similarity, structural similarity, and therapeutic similarity,
respectively, for a given drugdrug pair. The original denition
of consensus similarity is a similarity fusion method in the
ligand-based virtual screening, where a user-dened reference
structure is searched against a database using several dierent
similarity measures. Detailed descriptions about consensus
similarity were given in several related references.40,41
Performance Assessment. To test the performance of the
methods, a 10-fold cross-validation technique was used and
each result was repeated by 100 independent simulation tests.
In the process of 10-fold cross-validation, the entire links in the
DTI network were equally divided into 10 cross splits. In each
step of cross-validation, the model was trained on a set of nine
cross-validation splits together. The tenth subsample set was
used as an internal validation set (test set). At last, two
important metrics, namely the area under the receiver operating
characteristic curve (AUC) and recall enhancement (RE), were
calculated. The detailed descriptions of AUC and RE were
given in our previous published works.17,18

DSESI
DSSI
DTSI
mean
maximum
geometric mean

25.4
32.6
41.9
35.5
35.4
40.0

3.7
4.2
4.5
4.4
4.2
4.6

AUC SD
0.882
0.900
0.912
0.911
0.908
0.912

0.011
0.010
0.010
0.009
0.010
0.010

a
Recall enhancement (RE) was evaluated based on the top ve scored
lists. AUC: the area under the receiver operating characteristic curve.
DSESI: drug side eects similarity inference. DSSI: drug structural
similarity inference. DTSI: drug therapeutic similarity inference. The
standard deviation (SD) of the performance measured by 100 times
independent simulated test of 10-fold cross-validation. The detailed
performance is given in Table S2 of the Supporting Information.

Herein, the heat maps of side eect similarity, structural

similarity, therapeutic similarity, and three dierent consensus
similarities of drugdrug pairs for 621 approved drugs were
shown in Figure 2. The mean structural similarity of 621
approved drugs was 0.318, which indicated the structural
diversity of the training set. The overall therapeutic similarity of
drugdrug pairs is higher than side eects and chemical
structural similarities (Figure 2), which could explain that DTSI
is marginally better than DSSI and DSESI methods.
Our recently works demonstrated that consensus models
improve the performance of computational prediction.8,42 In
this work, three classic consensus similarity approaches,
including mean, maximum, and geometric mean, were
evaluated. As shown in Table 1, high performance was yielded
for three dierent consensus similarity-based inference
methods. And the geometric mean approach gives the best
performance, comparing with mean and maximum methods.
When comparing consensus similarity-based inference with
single similarity-based inference, consensus similarity-based
inference marginally outperforms single similarity-based inference, which could be caused by the redundancy in the
similarity reported by Keiser et al.21
Discovery New Drug Targets. The 621 known drugs
were computationally screened for their top 10 ranking scores
to bind with 893 target proteins using DSESI, DTSI, and DSSI
methods, respectively. Thousands of potential targets with high
ranking scores were prioritized for 621 drugs. In order to test
the feasibility of our methods, the top 10 predicted lists of each
drug with high ranking scores prioritized by three dierent
methods including DSESI, DTSI, and DSSI were overlapped.
We found that about 3000 new predicted DTIs were
overlapped (Supporting Information Table S3). Recently,
Lounkine et al. reported a large scale prediction and testing
of drug activity on side eect targets.7 The half-maximum
eective or inhibitory concentration (EC50 or IC50) values of
151 new predicted DTIs were validated with less than 30 M in
concentrationresponse curves.7 We mapped our predicted
results with 151 new validated DTI in Lounkines work. The
new prioritized DTI pairs of seven approved drugs were
mapped.7 We extracted all 44 new predicted DTI pairs
generated by DSESI, DTSI, and DSSI methods for those
seven approved drugs (Supporting Information Table S4).
Seven new predicted DTI pairs for seven approved drugs were
conrmed by published experiments.7 The range of IC50 or
EC50 value is from 0.21 to 23 M in concentrationresponse

RESULTS AND DISCUSSION

Performance of Proposed Methods. The 3195 drugtarget interaction pairs (Table S1) connected 621 approved
drugs and 893 target proteins were downloaded from
DrugBank31 and TTD.32 If a target connected with only one
drug (the degree of drug or target node is one) is just in the test
set and the corresponding links could not be predicted by the
collaborative ltering algorithm, because of no any information
for those target nodes or drug nodes in the training set.
Therefore, only drugs linked with two or more than two target
proteins were used. First, the DTI network was constructed as a
bipartite graph (Figure 1A). Then, the DSESI method was
developed for prediction of new candidate DTI (Figure 1). The
basic hypothesis of the DSESI method is that if two drugs are
similar in the side eects they elicit, they are more likely to
share a common drug target.27,28 As given in Table 1, the high
performance of the AUC value of 0.882 0.011 was yielded for
DSESI method measured by 100 independent simulation tests
of 10-fold cross-validation.
In addition, the DTSI and DSSI methods were also
developed for new DTI prediction. The DSSI method was
based on the hypothesis that two drugs with high structural
similarity may exhibit similar biological proles, which was
described in our previously published works.17,18 The DTSI
was based on the hypothesis that two drugs with high
therapeutic similarities may exhibit similar biological proles.39
As given in Table 1, the high performance of 0.900 0.011 and
0.912 0.010 were yielded for DTSI and DSSI methods
measured by 100 independent simulation tests of 10-fold crossvalidation, respectively. The detailed performance of 100
independent simulation tests of 10-fold cross-validation
prediction are given in Table S2 of the Supporting Information.
Comparing the performance of three dierent methods, DTSI
is marginally better than the DSSI and DSESI methods. High
diversity of molecules in the training set is an important issue
for the development and application of computational models.
756

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Figure 2. Heat maps of side eects similarity, structural similarity, therapeutic similarity, three dierent consensus (maximum, mean, and geometric
mean) similarities of drugdrug pairs for 621 approved drugs.

curves (Table 2) with the successful hit rate more than 15.9%.
In this study, 44 new predicted targets for seven known were
mapped with the published experimental data in Lounkines
work. Only seven new predicted targets were overlapped in the
published experimental data. If 44 new predicted targets were
experimentally assayed, there may be more targets validated.
Therefore, only a small part overlaps between our predicted
targets and the published experimental data, which could

explain why the success hit rate is much lower than the high
accurate rate in the cross-validation prediction.
The network graph analysis could help to visualize the new
and known DTI on a network level. The DTI network in
Figure 3 indicated that there are obvious polypharmacological
proles of several approved drugs. By searching in DrugBank31
and STITCH43 databases, the orally active antipsychotic drug
of pimozide binds with D2 dopamine receptor (DRD2), D3
757

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Table 2. Predicted and Published Bioassay Results of New Identied Indications for Seven Example Approved Drugs

a
Original primary pharmacological target information was extracted from DrugBank31 and STITCH43 databases. bThe IC50 or EC50 values were
extracted from Lounkines work.7 IC50: half maximal inhibitory concentration. EC50: half maximal eective concentration.

dopamine receptor (DRD3), potassium voltage-gated channel

subfamily H member 2 (KCNH2), and calmodulin (CALM1)
with high biological activities. As shown in Figure 3 and Table
2, the pimozide was predicted and conrmed to bind with
adrenoceptor alpha 1A (ADRA1A) with the IC50 value of 0.21
M. The antiasthma drug of fenoterol binds with beta-1, 2, and
3 adrenergic receptors with high agonist activities.31 In this
study, fenoterol was predicted and conrmed to bind with 5hydroxytryptamine receptor 2A (HTR2A) with the IC50 value
of 3.4 M (Table 2). The data demonstrated that our methods
can successfully identify polypharmacological proles of
approved drug in actually screening work.
Polypharmacological Proles of Approved Drugs.
Network visualization of drugtarget and targetADE
associations could provide a new mechanism-of-action
(MOA) for marketed drugs. To assess the potential clinical
relevance of predicted and validated targets systemically, we
built the o-target ADE association networks for several
example approved drugs. The 882 potential o-target ADE

pairs among 30 o-targets and 189 ADEs for seven approved

drugs (Supporting Information Table S5) were extracted from
Lounkines work.7 The clinical validation and predicted ADEs
for seven approved drugs were extracted from our previous
developed MetaADEDB.35 For an antiasthma drug of fenoterol,
the validated o-target of adrenoceptor alpha 1A (ADRA1A)
and two predicted o-targets of dopamine receptor D2
(DRD2) and adrenoceptor alpha 2A (ADRA2A) were linked
with several ADEs, e.g. parkinsonism, parkinsonism-like
phenotype (extrapyramidal disorder and hyperprolactinaemia),44 orthostatic hypotension, hypercholesterolaemia, etc.
(Supporting Information Table S6 and Figure 4A).
Pimozide is a conventional orally active antipsychotic drug
which shares with other antipsychotics the ability to blockade
dopaminergic receptors on neurons in the central nervous
system. Like most conventional antipsychotic drugs, pimozide
often generated propensity for or lack of motor side eects such
as extrapyramidal side eects (EPS), tardive dyskinesia,
endocrine disorder, galactorrhea, and amenorrhea (Figure 4B
758

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Figure 3. Discovered drugtarget interactions network produced with the software package Cytoscape (http://www.cytoscape.org/). The size of the
drug node is the fraction of the number of targets that the drug linked. The size of the target node is the fraction of the number of drugs that the
target linked. The Cytoscape format les of Figures 3 and 4 are available in the Supporting Information as Figure 3.cys, Figure 4A.cys, Figure 4B.cys,
and Figure 4C.cys.

subfamily H member 2 (KCNH2).46 The new predicted and

validated target of ADRA1A with IC50 value of 0.21 M was
linked with the side eect of hypertension. Interestingly, the dry
mouth side eect of pimozide could be explained by several
new predicted associated targets of muscarinic acetylcholine
receptor M1 and M2 (CHRM1, EF = 2.46, and CHRM2, EF =
2.42)7 in Supporting Information Table S5 and Figure 4B.
Comparing with conventional antipsychotic drugs, the
second generation atypical antipsychotic drug of loxapine not
only binds with D2 receptors but also binds with serotonin 5HT2 receptor (HTR2A) with high activities.47 As shown in the
o-target ADE network of loxapine (Figure 4C), several
common ADEs, such as EPS, tardive dyskinesia, galactorrhea,
and amenorrhea were also investigated for loxapien in clinical
use (Supporting Information Table S6). The molecular
hypothesis of ADEs was involved in the blockage of loxapine
on the nigrostriatal and tuberoinfundibular dopamine pathways.
The ADEs of tachycardia, urinary retention, and dry month
could be associated with the new validated target of CHRM2.
The known ADE of hypertension and new predicted anxiety
and bradycardia could be linked with new predicted target of
ADR1A1. Prediction of potential targets for old drugs could
help to explore the ADEs or the molecular mechanism of
known ADEs and postmarketing surveillance. Several common

and Supporting Information Table S6). The molecular

hypothesizes of pimozide involved in ADEs is that there are
four potential dierent dopamine pathways in human central
nervous systems, such as nigrostriatal dopamine pathway,
mesolimbic dopamine pathway, mesocortical dopamine pathway, and tuberoinfundibular dopamine pathway.4 For the
mesolimbic dopamine pathway, hyperactivity is thought to
cause psychosis and the positive symptoms of schizophrenia,
and blocking its hyperactivity could reduce psychosis or
eliminate positive symptoms. However, the nigrostriatal
dopamine pathway, as part of the extrapyramidal nervous
system, controls movements. When blocking the dopamine D2
receptor in this pathway by pimozide, it could lead to EPS and
tardive dyskinesia. The tuberoinfundibular dopamine pathway
inhibits prolactin release. If the normal function of tuberoinfundibular dopamine pathway was blocked by pimozide,
several side eects, such as endocrine, galactorrhea, and
amenorrhea will be generated. In addition to common ADEs
of antipsychotic drugs, several other ADEs such as hypertension, electrocardiogram QT prolonged,45 and dry mouth
were also investigated. The o-target ADE network of pimozide
was given in Figure 4B. The ADE of electrocardiogram QT
prolonged was associated with the high inhibitive activity of
IC50 value of 18 nM to potassium voltage-gated channel
759

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

Figure 4. Drug o-target ADE network prepared with the software package Cytoscape (http://www.cytoscape.org/). Known targets (green circle)
were collected from DrugBank and Therapeutic Target Database. Predicted o-target (gold circle) represents the overlap predicted results by drug
side eects similarity inference, drug structural similarity inference, and drug therapeutic similarity inference methods. Known adverse drug events
(ADEs) (purple square) represent the clinical validated data in MetaADEDB. Predicted ADEs (pink square) represent the new predicted data by
predictive phenotypic network inference model in our previous work.35 The size of the target node is the fraction of the number of ADEs that the
target linked. The size of the ADE node is the fraction of the number of targets that the ADE linked.

very helpful in prediction of drug polypharmacological proles

and provide a new perspective for network pharmacology.

ADEs with fuzzy molecular mechanisms were explained in this

work.

CONCLUSION

ASSOCIATED CONTENT

S Supporting Information
*

In this study, we systemically developed several dierent drugbased similarity inference methods, including drug side eect
similarity, drug structural similarity, drug therapeutic similarity,
and consensus similarity inference methods for predicting new
target proteins for approved drugs. High performance was
yielded when using our methodologies by cross-validation
prediction. Several new predicted target proteins for several
approved drugs were conrmed by published data. Moreover,
network visualization of drugtarget interactions and o-target
side eect associations provided a new mechanism-of-action for
three antipsychotic drugs in a case study. The results will be

Tables S1S6. This material is available free of charge via the

Internet at http://pubs.acs.org.

AUTHOR INFORMATION

Corresponding Author

*Tel.: +86-21-6425-1052. Fax: +86-21-6425-3651. E-mail:

[email protected].
Notes

The authors declare no competing nancial interest.

760

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

(17) Cheng, F.; Liu, C.; Jiang, J.; Lu, W.; Li, W.; Liu, G.; Zhou, W.;
Huang, J.; Tang, Y. Prediction of drug-target interactions and drug
repositioning via network-based inference. PLoS Comput. Biol. 2012, 8,
e1002503.
(18) Cheng, F.; Zhou, Y.; Li, W.; Liu, G.; Tang, Y. Prediction of
chemical-protein interactions network with weighted network-based
inference method. PLoS One 2012, 7, e41064.
(19) Cheng, F.; Zhou, Y.; Li, J.; Li, W.; Liu, G.; Tang, Y. Prediction of
chemical-protein interactions: multitarget-QSAR versus computational
chemogenomic methods. Mol. Biosyst. 2012, 8, 23732384.
(20) Cheng, F.; Li, W.; Zhou, Y.; Li, J.; Shen, J.; Lee, P. W.; Tang, Y.
Prediction of human genes and diseases targeted by xenobiotics using
predictive toxicogenomic-derived models (PTDMs). Mol. Biosyst.
2013, DOI: 10.1039/c3mb25309k.
(21) Keiser, M. J.; Setola, V.; Irwin, J. J.; Laggner, C.; Abbas, A. I.;
Hufeisen, S. J.; Jensen, N. H.; Kuijer, M. B.; Matos, R. C.; Tran, T. B.;
Whaley, R.; Glennon, R. A.; Hert, J.; Thomas, K. L.; Edwards, D. D.;
Shoichet, B. K.; Roth, B. L. Predicting new molecular targets for
known drugs. Nature 2009, 462, 175181.
(22) Keiser, M. J.; Roth, B. L.; Armbruster, B. N.; Ernsberger, P.;
Irwin, J. J.; Shoichet, B. K. Relating protein pharmacology by ligand
chemistry. Nat. Biotechnol. 2007, 25, 197206.
(23) Wang, F.; Liu, D.; Wang, H.; Luo, C.; Zheng, M.; Liu, H.; Zhu,
W.; Luo, X.; Zhang, J.; Jiang, H. Computational screening for active
compounds targeting protein sequences: methodology and experimental validation. J. Chem. Inf. Model. 2011, 51, 28212828.
(24) Li, H.; Gao, Z.; Kang, L.; Zhang, H.; Yang, K.; Yu, K.; Luo, X.;
Zhu, W.; Chen, K.; Shen, J.; Wang, X.; Jiang, H. TarFisDock: a web
server for identifying drug targets with docking approach. Nucleic Acids
Res. 2006, 34, W219224.
(25) Yang, L.; Wang, K.; Chen, J.; Jegga, A. G.; Luo, H.; Shi, L.; Wan,
C.; Guo, X.; Qin, S.; He, G.; Feng, G.; He, L. Exploring off-targets and
off-systems for adverse drug reactions via chemical-protein interactomeclozapine-induced agranulocytosis as a case study. PLoS
Comput. Biol. 2011, 7, e1002016.
(26) Xie, L.; Li, J.; Bourne, P. E. Drug discovery using chemical
systems biology: identification of the protein-ligand binding network
to explain the side effects of CETP inhibitors. PLoS Comput. Biol.
2009, 5, e1000387.
(27) Campillos, M.; Kuhn, M.; Gavin, A. C.; Jensen, L. J.; Bork, P.
Drug target identification using side-effect similarity. Science 2008, 321,
263266.
(28) Tatonetti, N. P.; Ye, P. P.; Daneshjou, R.; Altman, R. B. Datadriven prediction of drug effects and interactions. Sci. Transl. Med.
2012, 4, 125ra131.
(29) Tatonetti, N. P.; Liu, T.; Altman, R. B. Predicting drug sideeffects by chemical systems biology. Genome Biol. 2009, 10, 238.
(30) Yang, L.; Agarwal, P. Systematic drug repositioning based on
clinical side-effects. PLoS One 2011, 6, e28025.
(31) Knox, C.; Law, V.; Jewison, T.; Liu, P.; Ly, S.; Frolkis, A.; Pon,
A.; Banco, K.; Mak, C.; Neveu, V.; Djoumbou, Y.; Eisner, R.; Guo, A.
C.; Wishart, D. S. DrugBank 3.0: a comprehensive resource for
omics research on drugs. Nucleic Acids Res. 2011, 39, D1035D1041.
(32) Zhu, F.; Shi, Z.; Qin, C.; Tao, L.; Liu, X.; Xu, F.; Zhang, L.;
Song, Y.; Zhang, J.; Han, B.; Zhang, P.; Chen, Y. Therapeutic target
database update 2012: a resource for facilitating target-oriented drug
discovery. Nucleic Acids Res. 2012, 40, D11281136.
(33) Davis, A. P.; King, B. L.; Mockus, S.; Murphy, C. G.; SaraceniRichards, C.; Rosenstein, M.; Wiegers, T.; Mattingly, C. J. The
Comparative Toxicogenomics Database: update 2011. Nucleic Acids
Res. 2011, 39, D10671072.
(34) Kuhn, M.; Campillos, M.; Letunic, I.; Jensen, L. J.; Bork, P. A
side effect resource to capture phenotypic effects of drugs. Mol. Syst.
Biol. 2010, 6, 343.
(35) Cheng, F.; Li, W.; Wang, X.; Zhou, Y.; Wu, Z.; Shen, J.; Tang, Y.
Adverse Drug Events: Database Construction and in Silico Prediction.
J. Chem. Inf. Model. 2013, DOI: 10.1021/ci4000079.
(36) Sarwar, B.; Karypis, G.; Konstan, J.; Riedl, J. Item-Based
Collaborative Filtering Recommendation Algorithms. In Proceedings of

ACKNOWLEDGMENTS
This work was supported by the 863 Project (Grant
2012AA020308), the National Natural Science Foundation of
China (Grant 21072059), the Fundamental Research Funds for
the Central Universities (WY1113007), and the Shanghai
Committee of Science and Technology (11DZ2260600).

REFERENCES

(1) Hopkins, A. L. Network pharmacology: the next paradigm in

drug discovery. Nat. Chem. Biol. 2008, 4, 682690.
(2) Bowes, J.; Brown, A. J.; Hamon, J.; Jarolimek, W.; Sridhar, A.;
Waldron, G.; Whitebread, S. Reducing safety-related drug attrition: the
use of in vitro pharmacological profiling. Nat. Rev. Drug Discovery
2012, 11, 909922.
(3) Besnard, J.; Ruda, G. F.; Setola, V.; Abecassis, K.; Rodriguiz, R.
M.; Huang, X. P.; Norval, S.; Sassano, M. F.; Shin, A. I.; Webster, L. A.;
Simeons, F. R.; Stojanovski, L.; Prat, A.; Seidah, N. G.; Constam, D.
B.; Bickerton, G. R.; Read, K. D.; Wetsel, W. C.; Gilbert, I. H.; Roth, B.
L.; Hopkins, A. L. Automated design of ligands to polypharmacological
profiles. Nature 2012, 492, 215220.
(4) Stephen, M. S. Describing an Atypical Antipsychotic: Receptor
Binding and Its Role in Pathophysiology. J. Clin. Psychiat. 2003, 5, 9
13.
(5) Lazarou, J.; Pomeranz, B. H.; Corey, P. N. Incidence of adverse
drug reactions in hospitalized patients: a meta-analysis of prospective
studies. JAMA, J. Am. Med. Assoc. 1998, 279, 12001205.
(6) Cheng, F.; Yu, Y.; Zhou, Y.; Shen, Z.; Xiao, W.; Liu, G.; Li, W.;
Lee, P. W.; Tang, Y. Insights into molecular basis of cytochrome p450
inhibitory promiscuity of compounds. J. Chem. Inf. Model. 2011, 51,
24822495.
(7) Lounkine, E.; Keiser, M. J.; Whitebread, S.; Mikhailov, D.;
Hamon, J.; Jenkins, J. L.; Lavan, P.; Weber, E.; Doak, A. K.; Cote, S.;
Shoichet, B. K.; Urban, L. Large-scale prediction and testing of drug
activity on side-effect targets. Nature 2012, 486, 361367.
(8) Cheng, F.; Yu, Y.; Shen, J.; Yang, L.; Li, W.; Liu, G.; Lee, P. W.;
Tang, Y. Classification of Cytochrome P450 Inhibitors and nonInhibitors using Combined Classifiers. J. Chem. Inf. Model. 2011, 51,
9961011.
(9) Cheng, F.; Li, W.; Zhou, Y.; Shen, J.; Wu, Z.; Liu, G.; Lee, P. W.;
Tang, Y. admetSAR: A Comprehensive Source and Free Tool for
Assessment of Chemical ADMET Properties. J. Chem. Inf. Model.
2012, 52, 30993105.
(10) Curran, M. E.; Splawski, I.; Timothy, K. W.; Vincent, G. M.;
Green, E. D.; Keating, M. T. A molecular basis for cardiac arrhythmia:
HERG mutations cause long QT syndrome. Cell 1995, 80, 795803.
(11) Ashburn, T. T.; Thor, K. B. Drug repositioning: identifying and
developing new uses for existing drugs. Nat. Rev. Drug Discovery 2004,
3, 673683.
(12) Yamanishi, Y.; Araki, M.; Gutteridge, A.; Honda, W.; Kanehisa,
M. Prediction of drug-target interaction networks from the integration
of chemical and genomic spaces. Bioinformatics 2008, 24, i232240.
(13) Yabuuchi, H.; Niijima, S.; Takematsu, H.; Ida, T.; Hirokawa, T.;
Hara, T.; Ogawa, T.; Minowa, Y.; Tsujimoto, G.; Okuno, Y. Analysis of
multiple compound-protein interactions reveals novel bioactive
molecules. Mol. Syst. Biol. 2011, 7, 472.
(14) Gonzalez-Diaz, H.; Prado-Prado, F.; Garcia-Mera, X.; Alonso,
N.; Abeijon, P.; Caamano, O.; Yanez, M.; Munteanu, C. R.; Pazos, A.;
Dea-Ayuela, M. A.; Gomez-Munoz, M. T.; Garijo, M. M.; Sansano, J.;
Ubeira, F. M. MIND-BEST: Web Server for Drugs and Target
Discovery; Design, Synthesis, and Assay of MAO-B Inhibitors and
Theoretical-Experimental Study of G3PDH Protein from Trichomonas gallinae. J. Proteome Res. 2011, 10, 16981718.
(15) Bleakley, K.; Yamanishi, Y. Supervised prediction of drug-target
interactions using bipartite local models. Bioinformatics 2009, 25,
23972403.
(16) Yamanishi, Y.; Kotera, M.; Kanehisa, M.; Goto, S. Drug-target
interaction prediction from chemical, genomic and pharmacological
data in an integrated framework. Bioinformatics 2010, 26, i246254.
761

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762

Journal of Chemical Information and Modeling

Article

the World Wide Web Conference, Hong Kong, May 15, 2001; pp 285
295.
(37) Herlocker, J. L.; Konstan, J. A.; Terveen, K.; Riedl, J. T.
Evaluating collaborative filtering recommender systems. ACM T. Inf.
Syst. 2004, 22, 553.
(38) OBoyle, N. M; Banck, M.; James, C. A.; Morley, C.;
Vandermeersch, T.; Hutchison, G. R. Open Babel: an open chemical
toolbox. J. Cheminf. [Online] 2011, 3, Article 33; http://www.
jcheminf.com/content/3/1/33 (accessed Nov 19, 2011).
(39) Xu, K. J.; Song, J.; Zhao, X. M. The drug cocktail network. BMC
Syst. Biol. 2012, 6 (Suppl 1), S5.
(40) Willett, P. Similarity-based virtual screening using 2D
fingerprints. Drug Discovery Today 2006, 11, 10461053.
(41) Medina-Franco, J. L.; Yongye, A. B.; Perez-Villanueva, J.;
Houghten, R. A.; Martinez-Mayorga, K. Multitarget structure-activity
relationships characterized by activity-difference maps and consensus
similarity measure. J. Chem. Inf. Model. 2011, 51, 24272439.
(42) Cheng, F.; Ikenaga, Y.; Zhou, Y.; Yu, Y.; Li, W.; Shen, J.; Du, Z.;
Chen, L.; Xu, C.; Liu, G.; Lee, P. W.; Tang, Y. In silico assessment of
chemical biodegradability. J. Chem. Inf. Model. 2012, 52, 655669.
(43) Kuhn, M.; Szklarczyk, D.; Franceschini, A.; Campillos, M.; von
Mering, C.; Jensen, L. J.; Beyer, A.; Bork, P. STITCH 2: an interaction
network database for small molecules and proteins. Nucleic Acids Res.
2010, 38, D552D556.
(44) Farde, L.; Nordstrom, A. L.; Wiesel, F. A.; Pauli, S.; Halldin, C.;
Sedvall, G. Positron emission tomographic analysis of central D1 and
D2 dopamine receptor occupancy in patients treated with classical
neuroleptics and clozapine. Relation to extrapyramidal side effects.
Arch. Gen. Psychiat. 1992, 49, 538544.
(45) Aberg, K.; Adkins, D. E.; Liu, Y.; McClay, J. L.; Bukszar, J.; Jia,
P.; Zhao, Z.; Perkins, D.; Stroup, T. S.; Lieberman, J. A.; Sullivan, P. F.;
van den Oord, E. J. Genome-wide association study of antipsychoticinduced QTc interval prolongation. Pharmacogenomics J. 2012, 12,
165172.
(46) Kang, J.; Wang, L.; Cai, F.; Rampe, D. High affinity blockade of
the HERG cardiac K(+) channel by the neuroleptic pimozide. Eur. J.
Pharmacol. 2000, 392, 137140.
(47) Kapur, S.; Zipursky, R.; Remington, G.; Jones, C.; McKay, G.;
Houle, S. PET evidence that loxapine is an equipotent blocker of 5HT2 and D2 receptors: implications for the therapeutics of
schizophrenia. Am. J. Psychiat. 1997, 154, 15251529.

762

dx.doi.org/10.1021/ci400010x | J. Chem. Inf. Model. 2013, 53, 753762