Vol. 29 no.

2 2013, pages 238–245

BIOINFORMATICS ORIGINAL PAPER doi:10.1093/bioinformatics/bts670

Systems biology Advance Access publication November 17, 2012

Drug–target interaction prediction by learning from local

information and neighbors
Jian-Ping Mei1,*, Chee-Keong Kwoh1, Peng Yang1, Xiao-Li Li1,2 and Jie Zheng1
1
Bioinformatics Research Centre, School of Computer Engineering, Nanyang Technological University, 50 Nanyang
Avenue, Singapore 639798, Singapore and 2Institute for Infocomm Research, A*Star, 1 Fusionopolis Way #21-01
Connexis, Singapore 138632, Singapore
Associate Editor: Trey Ideker

ABSTRACT potential complement that provides useful information in an

Motivation: In silico methods provide efficient ways to predict pos- efficient way.
sible interactions between drugs and targets. Supervised learning Generally, the prediction performance is decided by both the
approach, bipartite local model (BLM), has recently been shown to data used and the particular analysis method that is applied to.
be effective in prediction of drug–target interactions. However, for An intuitive and straightforward way to identify new targets for
drug-candidate compounds or target-candidate proteins that currently a drug is to compare the candidate proteins with those existing
have no known interactions available, its pure ‘local’ model is not able targets of that drug. Different results may be obtained depending
to be learned and hence BLM may fail to make correct prediction on which perspective the comparison is made with respect to.
when involving such kind of new candidates. Keiser et al. (2009) compare targets based on the chemical struc-
Results: We present a simple procedure called neighbor-based ture of ligands that bind to them. As reviewed in Haupt and
interaction-profile inferring (NII) and integrate it into the existing BLM Schroeder (2011), the structure of binding sites is another import-
method to handle the new candidate problem. Specifically, the ant way to compare proteins or to measure the similarity be-
inferred interaction profile is treated as label information and is used tween proteins. Although binding site is an effective measure
for model learning of new candidates. This functionality is particularly for identification of new targets, the structures of binding site
important in practice to find targets for new drug-candidate com- are only available for a small set of proteins, of which the 3D
pounds and identify targeting drugs for new target-candidate proteins. structures are known. To be able to consider more proteins,
Consistent good performance of the new BLM–NII approach has been amino acid sequence may be used as it is available for most
observed in the experiment for the prediction of interactions between proteins. Similarly, to identify new targeting compounds for a
drugs and four categories of target proteins. Especially for nuclear specific target, comparison is made on the compound side or
receptors, BLM–NII achieves the most significant improvement as drug side with respect to chemical structures (Laggner et al.,
this dataset contains many drugs/targets with no interactions in the 2012; Martin et al., 2002), side effects (Campillos et al., 2008)
cross-validation. This demonstrates the effectiveness of the NII strat- or other possible measurements of drug.
egy and also shows the great potential of BLM–NII for prediction of More sophisticated statistical and machine learning methods
compound–protein interactions. have been developed recently for prediction of genome-wide
Contact: [email protected] drug–target interactions. In He et al. (2010) and Perlman et al.
Supplementary information: Supplementary data are available at (2011), multiple groups of drug-related features and
Bioinformatics online. protein-related features have been extracted to describe each
Received on July 2, 2012; revised on October 15, 2012; accepted on drug–target pair. After feature selection, a certain classifier is
November 12, 2012 used to predict whether a given pair is interacting or not.
Yamanishi et al. (2008) proposed a supervised bipartite graph
learning approach. In this approach, the chemical space and the
1 INTRODUCTION geometric space are mapped into a unified space so that those
Identification of interactions between drugs/compounds and interacting drugs and targets are close to each other while those
protein targets is an important part of the drug discovery pipe- non-interacting drugs and targets are far away from each other.
line. The great advances in molecular medicine and the human By mapping the query pair of drug and target to that space with
genome project provide more opportunities to discover unknown the learned mapping function, the probability of interaction
associations in the compound–protein interaction network. The between them is then calculated as their closeness in the
newly discovered interactions are helpful for discovering new mapped space. Another method called the weighted profile
drugs by screening candidate compounds and also may help method was also given in Yamanishi et al. (2008). For a query
understand the causes of side effects of existing drugs. Since drug, the weighted profile method assigns a probability of
experimental way to determine drug–target interactions is interaction to the query target based on how the neighbors of
costly and time-consuming, in silico prediction becomes a this drug interact with this target. Basically, weighted profile is a
nearest-neighbor approach and it is called drug-based/target-
*To whom correspondence should be addressed. based similarity inference in Cheng et al. (2012). Other than

238 ß The Author 2012. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: [email protected]
 Bipartite model for learning from local information and neighbors

inferring interactions from the drug similarity or target profile method, which is a nearest-neighbor approach. Our ex-
similarity, network-based inference was also studied in perimental results show that BLM–NII performs much better
Cheng et al. (2012), which infers or predicts drug–target inter- than the weighted profile method.
actions based on the topology of the known interaction network. Systematic experiments are conducted to simulate the task of
Different from the work in Cheng et al. (2012), which makes use drug–target interactions prediction cross four datasets.
of the drug similarity, target similarity and network-based simi- Compared with state-of-the-art approaches, our proposed
larity separately, Chen et al. (2012) apply random walk on a approach achieves consistent improvement in terms of area
heterogeneous network constructed with these three types of under ROC (AUC) curve and area under precision versus
similarities. Another promising approach is the bipartite local recall (AUPR) curve. As these four datasets contain different
model (BLM) approach. Bleakley and Yamanishi (2009) portions of new drug candidates and target candidates in the
showed that the ensemble of independent drug-based prediction simulation, the improvements of BLM–NII compared with
and target-based prediction with supervised learning performs BLM are also different for the four datasets. The most significant
much better than only using each single type of prediction. The improvement is achieved on the nuclear receptor dataset, which
BLM method has been further studied and improved in Xia et al. contains the largest portion of new candidates. This shows that
(2010) and Laarhoven et al. (2011). The main differences of these the NII strategy, i.e. to infer label information or training
three methods include the drug–drug and target–target similari- data from neighbors when there is no training data readily avail-
ties, the classifiers and the way used to combine the drug-based able from the query compound/protein itself, is feasible and ef-
and target-based interaction probabilities. In Xia et al. (2010), fective for dealing with the new candidate problem of the original
semi-supervised approach is used instead of supervised approach BLM.
for local model learning; while Laarhoven et al. (2011) found
that using only the kernel based on the topology of the known
interaction network is able to obtain a very good performance. 2 METHODS
In the existing framework of BLM, the model for the query
drug or target is learned based on local information, i.e. its own
2.1 Problem formalization
interaction profile. Despite a good performance, BLM has limi- Assume that the bipartite interaction network N1 illustrated in Figure 1
tations. It is unable to learn without training data and hence is involves md drugs/compounds and mt targets, which are referred to as
existing drug candidates and target candidates, respectively. We use
not able to provide a reasonable prediction for drug/target can-
matrix A to represent this network, i.e. aij 2 A ¼ 1 if the i-th compound
didates that are currently new. Here, a drug-candidate com-
di is known to interact with the j-th target tj. All other entries of A are 0.
pound is new if it does not have any known targets, and a The problem under consideration is how to make use of the
target-candidate protein is new if it is not targeted by any known interactions together with the compound similarities and protein
drugs/compounds. We call this the new candidate problem of similarities to predict new interactions between nd drug-candidate com-
BLM. Since a large number of compounds and proteins, which pounds and nt target-candidate proteins, where nd4md and nt4mt. This
are possible drug candidates and target candidates, respectively, means there are md ¼ nd  md new drug candidates and mt ¼ nt  mt
are new, in this study, we focus on handling the new candidate new target candidates, which have no interactions currently known.
problem by proposing an improved version of BLM called BLM The whole network involving nd compounds and nt proteins can be rep-
with neighbor-based interaction-profile inferring (BLM–NII). resented as
" #  
The NII procedure is developed to incorporate the capacity of ðN1 Þmd mt ðN2 Þmd mt A 0
learning from neighbors into the original BLM method. More Nnd  nt ¼ ¼ , ð1Þ
ðN3 Þmd md ðN4 Þmd mt 0 0
specifically, when the query involves a new drug/target candi-
date, we first derive the initial weighted interactions for the where known interactions correspond to non-zero entries of A. Now, we
new candidate from its neighbors’ interaction profiles, and then want to predict possible interactions in N1 between existing drug candi-
use the inferred interactions as label information to train the dates and target candidates, as well as in other three subnetworks N2, N3
model. In general, neighbors refer to compounds/proteins that and N4, where the interactions at least involve one type of new
have large similarities to the query compound/protein.
The presented NII idea happen to be similar to the weighted
profile method in some sense. However, our BLM–NII method
is substantially different from the weighted profile method in the
following aspects. In BLM–NII, the derived interaction profile is
used as label information to train the local model or the classi-
fier, while in the weighted profile method, the derived weighted
interaction is directly used as the final predicted interaction prob-
ability. Moreover, in BLM–NII, the NII procedure is integrated
into the BLM framework where a certain classifier plays the
main role in model learning, and NII is activated only for new
drug/target candidates; while in the weighted profile method,
there is no other classifier and the procedure of deriving the Fig. 1. Bipartite interaction network: a network consists of two types of
weighted profile acts as a classification process, which is applied nodes, where edges only connect different types of nodes. The drug–target
for any drug/target candidates. To sum up, the BLM–NII is an interaction network is a bipartite network, where drug and target are two
enhanced BLM method, and it is different from the weighted types of nodes and the interactions between them are the edges

239
J.-P.Mei et al.

candidates, i.e. the target candidate is new, the drug candidate is new or similarity which encodes the topology information of the interaction net-
both are new. work has been shown to provide good results. With the Gaussian kernel,
the network-based drug similarity Sdn and network-based target similarity
Stn are calculated as:
2.2 Bipartite local model
 ka0  a0 k2 
To predict pij, the probability that a drug di and a target tj interact, i j
Sdn ði; jÞ ¼ exp  ; ð11Þ
the basic BLM proposed by Bleakley and Yamanishi (2009) is described 
as follows. A local model for di denoted as Modd(i) is first learned
based on its interaction profile a0i and the similarities between targets  ka  a k2 
i j
Stn ði; jÞ ¼ exp  ; ð12Þ
St, i.e. 
P
Modd ðiÞ ¼ trainðSt ; a0 i Þ: ð2Þ where the bandwidth  ¼ 0  1n ni¼1 a2ij , and different bandwidths may
be used for drug and target, respectively. However, the result with
Here, train represents the learning process of a certain classifier, e.g.
network-based similarity may not remain good when the information
support vector machine or (Kernel) regularized least squares (RLS), the
contained in the interaction network is not sufficient enough. Rather
similarity matrix St is used as the observed data of target candidates, and
than considering one type of similarity, a more general way is to combine
the interaction profile a0i , i.e. the i-th row vector of A, serves as label
several types of similarities. Here, we use both the network-based simi-
information to label each target candidate whether interacting with this
larity and chemical similarity for drug similarity Sd, and the
drug. Once the model Modd(i) is learned, it is used to predict pdij , the
network-based similarity and sequence similarity for target similarity St
probability of interaction between di and the query target candidate tj:
through linear combination:
pdij ¼ testðModd ðiÞ; stj Þ, ð3Þ
Sd ¼ Sdc þ ð1  ÞSdn ; ð13Þ
t
stj
where is the j-th column of S recording the similarities between tj and
other targets. The similar model learning and prediction process are St ¼ Sts þ ð1  ÞStn ; ð14Þ
performed independently from the query-target side to get ptij , i.e.
Sdc Sts
where is the chemical structure similarity for drug, is the amino acid
Modt ðjÞ ¼ trainðSd ; aj Þ; ð4Þ sequence similarity for protein and  is the combination weight set by
user. Although more sophisticated ways such as Kronecker product
ptij ¼ testðModt ðjÞ; sdi Þ; ð5Þ may be used to combine two types of similarity matrices or kernel
matrices, experimental results in (Laarhoven et al. 2011) show that the
where aj is the j-th column vector of A or the interaction profile of target linear combination gives comparable performance with a much lower
tj. Once both pdij and ptij have been calculated, they are combined to get computational complexity.
probability pij:
pij ¼ gðpdij ; ptij Þ; ð6Þ 2.3 Neighbor-based interaction-profile inferring
Good performance of supervised learning is largely dependent on the
where g is a function that combines or integrates pdij and ptij . Examples
amount and quality of labeled training data. When a drug/target candi-
include pij ¼ maxfpdij ; ptij g and pij ¼ 0:5ðpdij þ ptij Þ, where g is the max or
date is new, it has no existing interactions that can be used as label
average function.
information and the model for this candidate thus can not be learned.
After pij is calculated for each pair of compound i and protein j, the
As shown in (7), interactions between new drug candidates and new
output network of BLM may be represented as
" # target candidates remain unpredicted in BLM. To extend the application
NBLM
1 NBLM
2
domain of BLM to new drug/target candidates, we propose to derive
NBLM ¼ ; ð7Þ training data from their neighbors. Based on the assumption that
NBLM
3 0
drugs/compounds which are similar to each other interact with the
with same targets, interaction profile for new drug-candidate compounds
could be possibly inferred from their neighbors’ interactions.
NBLM
1 ¼ N1 þ P1 ðModd ; Modt Þ; ð8Þ
Compounds with large similarities to the new drug-candidate compound
are said to be its neighbors. Since new drug-candidate compounds have
NBLM
2 ¼ P2 ðModd Þ; ð9Þ no interactions, or all the elements of its current interaction profile vector
are 0, it is not suitable to consider network-based similarity here, so only
NBLM
3 ¼ P3 ðModt Þ: ð10Þ chemical structure similarity is used to define the neighbors of a
drug-candidate compound. Formally, for a compound di which is a
where P1 gives the predicted interactions between existing drug candi-
new drug-candidate, we infer the j-th dimension of its interaction profile
dates and existing target candidates, P2 are predicted interactions between
ld(i) with
existing drug candidates and new target candidates and P3 gives
predicted interactions between new drug candidates and existing target X
md

candidates. ldj ðiÞ ¼ sih ahj ; ð15Þ

h¼1
For any classifier that is used, the known targets of di corresponding to
non-zero elements of a0i and the pairwise target similarity St are critical to where sih is the chemical similarity between two compounds di and dh. The
the final prediction of pdij . The model learned for di describes how this above formula shows that the interaction weight of this drug with respect
drug selects targets. Once the model is learned, the similarities between to the j-th target is the collection of its neighbors’ interactions to this
the query target and those known targets of di largely decide pdij . Similarly, target. For a given new drug-candidate compound, the simple formula
known targeting drugs of tj or non-zero elements of tj ’s interaction profile given in Equation (15) defines that the inferred weight of interaction
aj and the pairwise drug similarity Sd are critical to the final prediction of between this compound and a target is high if many of its neighbors
ptij . Under the same BLM framework, different results are produced due interact with this target, and also it is decided more by neighbors with
to the differences in Sd, St, the classifier and the combination function g. large similarities than those with small similarities. Since new
According to the study of Laarhoven et al. (2011), network-based target-candidate proteins have no interactions with any compound, the

240
 Bipartite model for learning from local information and neighbors

inferred interactions for di are only with existing target candidates. To be NBLMNII ¼ P2 ðModd ; Mod0t Þ; ð20Þ
2
more specific, ldj ðiÞ > 0 if the j-th target candidate is an existing one, i.e.
ahj40 for at least one h, and ldj ðiÞ ¼ 0 if the j-th target candidate is
NBLMNII
3 ¼ P3 ðMod0d ; Modt Þ; ð21Þ
new, i.e. ahj ¼ 0 for all h. To ensure the value of each ldj ðiÞ is in the
range of [0, 1], linear scale is performed subsequently, i.e.
NBLMNII ¼ P4 ðMod0d ; Mod0t Þ: ð22Þ
ldj ðiÞ ¼ ðldj ðiÞ  minh ldh ðiÞÞ=ðmaxh ldh ðiÞ  minh ldh ðiÞÞ. After we obtained 4

the inferred interaction profile, we can use it as label information to Comparing NBLMNII and NBLM, it is observed that the interactions be-
learn the model of di: tween existing drug candiates and target candiates are the same for the
two approaches, while the interactions in the other three cases in BLM–
Mod0d ðiÞ ¼ trainðSt ; ld ðiÞÞ: ð16Þ
NII are different from those in BLM. First, BLM–NII is able to predict
In the same way, this procedure is applied to a new target-candidate P4, the interactions between drug candidates and target candidates that
protein tj to obtain its inferred interaction profile lt(j), where its neighbors are both new. Second, P2 and P3 in BLM–NII are predicted from both
are defined based on sequence similarity. The model of tj can then be the drug side and the target side, while in BLM are predicted only from
learned with lt(j): one side.
Learning from neighbors allows drug/target candidates to obtain
Mod0t ðjÞ ¼ trainðSd ; lt ðjÞÞ: ð17Þ labeled data when themselves do not have or have insufficient labeled
This interaction profile inferring technique is particularly useful for data for training. This procedure actually introduces some degree of glo-
those new drug/target candidates, for which existing supervised methods balization into the original local model to provide more chances of learn-
(e.g. BLM) fail to produce reasonable predictions. It can also be useful to ing from known knowledge. However, too much globalization is not
enhance the classification models for any compounds/proteins without desired as it could eliminate the local characteristics and make the
enough training data or label information. models of individual candidates less discriminative. Moreover, the low
quality of neighbors due to imprecise similarity measure may cause nega-
tive impact when the learning process replies on too much neighbors’
2.4 BLM with NII information. In other words, the inferred interaction profile, although
By integrating the above presented NII strategy into the BLM frame- is helpful, may introduce a certain amount of noise. Therefore, in this
work, we have the BLM with NII (BLM–NII). The detailed steps of study, we only activate the neighbor-based learning for totally new can-
BLM–NII to predict the probability pij between any compound i and didates. For other cases, we still train the model locally with its own
any protein j is described in Algorithms 1 and 2. known interactions.

Algorithm 1: BLM–NII
3 MATERIALS
input : A, Sdc , Sts
output: pij To facilitate comparison with published approaches, we used the
get pdij ¼ NII-integrated Learning and Prediction ( A, Sdc , Sts ) from di; same groups of four datasets which are first analyzed by
get ptij ¼ NII-integrated Learning and Prediction ( A, Sts , Sdc ) from tj; Yamanishi et al. (2008) and then later by Bleakley and
Combine pdij and ptij to get the final result pij ¼ gðpdij ; ptij Þ Yamanishi (2009), Xia et al. (2010), Laarhoven et al. (2011)
and Cheng et al. (2012). These four datasets correspond to
drug–target interactions of four important categories of protein
Algorithm 2: NII-integrated learning and prediction targets, namely enzyme, ion channel, G-protein-coupled receptor
input : A, Sdc , Sts (GPCR) and nuclear receptor, respectively. The datasets were
output: pdij downloaded from http://web.kuicr.kyoto-u.ac.jp/supp/yoshi/
if di is new then drugtarget/.
| obtain ld(i) with Eq. (15) with Sdc Table 1 gives some statistics of each dataset including the total
else number of drugs (nd), the total number of targets (nt), the total
| ld(i) is the i-th row of A number of interactions (E), the average number of targets for
end each drug (D d ), the average number of targeting drugs for each
Compute Stn with Eq. (12) and St with Eq. (14);
target (D t ), the percentage of drugs that have only one target
Learn a local model for di, i.e., Modd(i) ¼ train(St, ld(i)) ;
(Dd ¼ 1) and the percentage of targets that have one targeting
if tj is new then
| predict pdij with Modd(i) and Sts drug (Dt ¼ 1). It is shown from this table that among the four
else drug–target interaction networks, on average, each drug and
| predict pdij with Modd(i) and St target in ion channel and enzyme have more interactions than
end those in GPCR and nuclear receptor. It is also worthy noting
that in the leave-one-out cross–validation (LOOCV), drugs and
targets with one interaction are ‘new candidates’ as the only one
interaction is covered over to leave no recorded interaction, e.g.
The output network of BLM–NII is expressed as 72% drugs in the nuclear receptor are ‘new candidates’ in the
" # simulation.
NBLMNII
1 NBLMNII
2
NBLMNII ¼ ; ð18Þ Each dataset is described by three types of information in the
NBLMNII NBLMNII
3 4 form of three matrices: (i) the drug–target interaction matrix; (ii)
with the drug–drug similarity matrix and (iii) the target–target simi-
larity matrix. The interaction networks were retrieved from the
NBLMNII
1 ¼ NBLM
1 ; ð19Þ
KEGG BRITE (Kanehisa et al., 2006), BRENDA (Schomburg

241
J.-P.Mei et al.

Table 1. Some statistics of the four datasets Table 2. Comparison with existing approaches for the four datasets

Dataset Enzyme Ion channel GPCR Nuclear Dataset Method AUC AUPR
receptor
Enzyme Weighted profile 86.4 6.30
nd 445 210 223 54 BY(2009) 97.6 83.3
nt 664 204 95 26 Laarhoven et al. (2011) 97.8 91.5
E 2926 1476 635 90 BLM–NII 98.8 92.9
Dd 6.58 7.03 2.85 1.67 Ion channel Weighted profile 81.9 17.2
D t 4.41 7.24 6.68 3.46 BY(2009) 97.3 78.1
Dd ¼ 1(%) 39.78 38.57 47.53 72.22 Laarhoven et al. (2011) 98.4 94.3
Dt ¼ 1(%) 43.37 11.27 35.79 30.77 BLM–NII 99.0 95.0
GPCR Weighted profile 76.5 10.9
BY(2009) 95.5 66.7
Laarhoven et al. (2011) 95.4 79.0
et al., 2004), SuperTarget (Gnther et al., 2008) and DrugBank BLM–NII 98.4 86.5
(Wishart et al., 2008). The drug–drug similarity is measured Nuclear receptor Weighted profile 74.9 17.1
based on chemical structures from the DRUG and BY(2009) 88.1 61.2
COMPOUND sections in the KEGG LIGAND database Laarhoven et al. (2011) 92.2 68.4
(Kanehisa et al., 2006). The chemical structure similarities be- BLM–NII 98.1 86.6
tween drugs are computed with SIMCOMP (Hattori et al.,
2003), which uses a graph alignment algorithm to get a global
similarity score based on the size of the common substructures three for all the datasets. Since the results of weighted profile are
between two compounds. The target–target similarity is mea- much worse than those of the three BLM-based methods namely
sured based on the amino acid sequences retrieved from the BY (2009), Laarhoven et al. (2011) and BLM–NII, we now focus
KEGG GENES database (Kanehisa et al., 2006). The sequence on the comparison of these three approaches. As been discussed
similarities between proteins are computed with a normalized in Laarhoven et al. (2011), by incorporating the network-based
version of Smith–Waterman score. More details on how the similarity, the performance of BLM can be improved, i.e. the
data have been collected and calculated are given in Yamanishi results of Laarhoven et al. (2011) in terms of AUPR are much
et al. (2008). better than those of BY (2009). It is also shown that the per-
formance of BLM can further be improved by integrating the
NII procedure, i.e. the results of BLM–NII is consistently better
4 EVALUATION than those of Laarhoven et al. (2011).
Systematic experiments are performed to evaluate the perform- It is interesting to observe that different levels of improve-
ance of the presented approach with datasets summarized in ments have been achieved for different datasets. Comparing
Table 1. As in Laarhoven et al. (2011), LOOCV is performed. Laarhoven et al. (2011) and BY (2009), the improvement is the
Since the real interaction to be predicted is left out, compounds most significant on ion channel and the least significant on nu-
and proteins with one interaction (i.e. Dd ¼ 1 or Dt ¼ 1) turn out clear receptor. Differently, comparing BLM–NII and Laarhoven
to have no training data and thus they are treated as ‘new can- et al. (2011), the improvement is the largest for nuclear receptor
didates’ in the cross-validation. To test the robustness of the and the least for ion channel. Such kind of differences are ex-
presented approach, we also performed 10-fold cross-validation. pected due to the differences in the structure of the datasets.
The results of 10 trials 10-fold cross-validation can be found in From Table 1, it is shown that among the four datasets, the
Tables S5–S8 of the Supplementary Material. average numbers of interactions of each drug and target are
the largest for ion channel and the smallest for nuclear receptor.
This means that the interaction network of ion channel contains
4.1 Compare with state-of-the-art approaches more information than nuclear receptor and thus the
First, we compare the performance of BLM–NII (g ¼ max, network-based similarity of ion channel is more robust and in-
 ¼ 0.5) with the weighted profile method (Yamanishi et al., formative than that of nuclear receptor. Therefore, incorporating
2008) and two other state-of-the-art approaches (Bleakley and the network-based similarity results in larger improvement for
Yamanishi, 2009) and (Laarhoven et al., 2011) denoted as BY ion channel. Since drugs or targets with one interaction are ‘new
(2009) and Laarhoven et al. (2011), respectively. The same RLS candidates’ in the simulation, it is also shown from Table 1 that
classifier is used for BLM–NII as Laarhoven et al. (2011). We the nuclear receptor contains the largest portion of ‘new candi-
measure the quality of the predicted interactions in terms of dates’ while the ion channel contains the least. Thus, by applying
AUC curve (or true-positive rate versus false-positive rate the NII procedure, BLM–NII has more chances to improve the
curve) and AUPR curve. results for nuclear receptor than for Ion Channel.
Table 2 gives the AUC and AUPR scores of the four
approaches for the four datasets. The results of BY (2009) and
Laarhoven et al. (2011) are the best ones reported in Bleakley 4.2 Comparison between BLM and BLM–NII
and Yamanishi (2009) and Laarhoven et al. (2011), respectively. To directly show the improvements attributed to the NII strat-
From this table, it is clear that BLM–NII outperforms the other egy, we now compare BLM–NII and BLM, i.e. the results of

242
 Bipartite model for learning from local information and neighbors

(a) (b) (c)

Fig. 2. AUPR of BLM and BLM–NII for nuclear receptor with different types of similarities: (a)  ¼ 1, (b)  ¼ 0 and (c)  ¼ 0.5

(a) (b) (c)

Fig. 3. Precision–recall curve of BLM and BLM–NII for nuclear receptor with different similarities: (a)  ¼ 1, (b)  ¼ 0 and (c)  ¼ 0.5

Table 3. Compare 1% and 3% top ranked pairs of BLM and BLM–NII for nuclear receptor

Top 1% Top 3%

 Method Sensitivity PPV MCC Sensitivity PPV MCC

1 BLM 13.3 85.7 32.5 35.6 76.2 50.0

BLM–NII 15.6 100.0 38.3 44.4 95.2 63.7
0 BLM 16.7 93.8 38.3 32.2 67.4 44.3
BLM–NII 18.9 100.0 42.3 45.6 97.6 65.4
0.5 BLM 15.6 100.0 38.3 40.0 85.7 56.9
BLM–NII 15.6 100.0 38.3 45.6 97.6 65.4

BLM–NII where new candidates are treated as existing ones. We the precision–recall curve of BLM and BLM–NII. Table 3 shows
applied both BLM and BLM–NII with three different groups of the sensitivity (or recall), PPV (positive predictive value or pre-
inputs by setting  in Equations (13) and (14) to 1, 0 and 0.5. cision) and MCC (Matthews correlation coefficient). The two
We obtained the AUC and AUPR scores of both methods groups of results in Table 3 are calculated by considering the
with g ¼ max. The results of both with g ¼ average or mean 1% and 3% pairs with the highest pij values as positive, respect-
have also been produced, which can be found in Tables S1–S4 ively. It is clearly shown from these results that with NII being
of the Supplementary Material. Since the same conclusion can be integrated, the performance of BLM has been improved.
drawn with respect to either of the two metrics, we put the AUC
scores in the Supplementary Material and plot the AUPR scores
of BLM and BLM–NII for the four datasets with three different
types of similarities in Figure 2. It is shown that for any type of 4.3 Detailed analysis of the effectiveness of NII
similarities, BLM–NII performs better than BLM for all the To take a close look at the difference in the results attributed to
datasets. Again, the improvements made by BLM–NII are the NII strategy, we now compare those top ranked interactions
more significant for nuclear receptor and GPCR than for the of the nuclear receptor dataset produced by BLM–NII and
other two datasets. BLM. Since this dataset has 90 known interactions, we inspect
Now using nuclear receptor, we make further comparison of the 90 interactions with the highest probabilities predicted by
the performance between BLM and BLM–NII. Figure 3 plots each algorithm.

243
J.-P.Mei et al.

As summarized in Table 4 (More detailed results are in Table (nuclear receptor subfamily 1, group I, member 3) and D05341
S11 of the Supplementary Material), among the top 90 predicted [Palmitic acid (NF)] – hsa3174 (hepatocyte nuclear factor 4,
interactions, BLM only correctly detected 58 known interactions gamma), which are assigned extremely low ranks by BLM are
while BLM–NII detected 71, and 57 known interactions are successfully detected by BLM–NII as shown in Figure 4. After
ranked within 90 by both. Although one interaction detected checking, we find that the query drug D00163 of the first pair
by BLM is missed by BLM–NII, this one ranks 104 in BLM– only has one target which happens to be the query target
NII, which indicates that this pair is still recognized to be inter- hsa9971, and the query target is known to be only interacting
acting with a highly possibility by BLM–NII. Nevertheless, 14 with the query drug. The other two pairs have the same situation
interactions detected by BLM–NII are missed by BLM. The as this pair. As we left out the true interaction in our simulation,
average rank of these 14 interactions produced by BLM is 388 the testing for these three pairs becomes to predict interaction
as some of them ranks very low. between new drug-candidate compound and new target-
Among these 14 drug–target pairs, three pairs namely D00163 candidate protein. Since training data are absent for both the
(Chenodeoxycholic acid) – hsa9971 (nuclear receptor subfamily query drug and query target, BLM fails to detect interactions
1, group H, member 4), D00506 (Phenobarbital) – hsa9970 for those three pairs. Although difficulty is presented for such
kind of cases, BLM–NII successfully detected these three pairs to
be interacting. This shows the effectiveness of NII for prediction
of interaction involving new candidates.
Table 4. Performance of BLM and BLM–NII on nuclear receptor Now using D00163 and hsa9971 as an example, we give inter-
mediate results to illustrate how NII helps detect the interactions
Total known interactions: 90 between new drug-candidate compounds and new target-
candidate proteins. Figure 5 shows the local model learned for
Interactions detected by BLM 58 D00163 with the help of inferred training data. Specifically,
Interactions detected by BLM–NII 71 Figure 5a shows the inferred interaction profile of D00163, i.e.
Interactions detected by both BLM and BLM-–NII 57 the weighted interactions between D00163 and 25 non-query tar-
gets calculated with Equation (15). It shows that the associations
between D00163 and several targets such as hsa2099 are large.
This is because many of D00163’s neighbors or similar drugs,
(a) (b) (c) such as D00066, interact with this target as seen from Figure 5b.
Using this inferred interaction profile as label information,
Figure 5c shows the learned local model of D00163, or the
pair1 weight of each of the targets learned with the classifier with re-
spect to D00163. With this learned model, BLM–NII successfully
pair2
detected the interaction between D00163 and hsa9971 based on
the similarities between the query target hsa9971 and other tar-
gets especially those with large weights in the model of D00163.
In the same manner, the local model of the query target which is
a ‘new’ candidate can be learned with NII. This example illus-
pair3 trates the feasibility and effectiveness of the presented approach
to infer training data or label information from the interaction
Fig. 4. Drug–target interaction matrix of nuclear receptor. (a) Predicted
profiles of neighbors.
by BLM, (b) predicted by BLM–NII, (c) real interaction matrix. Each
entry of the interaction matrix is plotted as a pixel. The brightness of a
pixel represents the interaction possibility of the corresponding pair, i.e. the
5 CONCLUSION AND DISCUSSION
brighter the more possible that the pair interacts. Three pairs are circled in
(b). These three pairs which consist of drug candidate and protein candi- We proposed an intuitive solution to the new candidate problem
date that are both ‘new’ in Loo validation are detected by BLM–NII of BLM by integrating a NII procedure, i.e. infer training data

(a) (b) (c) (d)

Fig. 5. Local model learning for D00163 with BLM–NII. (a) inferred interaction profile ld of D00163, (b) weighted interaction of D00163’s neighbors to
hsa2099 calculated with s(D00163, i)  a(i, hsa2099) for each drug i, (c) learned model of D00163 by the RLS classifier, (d) similarities between hsa9971
and other proteins, i.e. sthas9971

244
 Bipartite model for learning from local information and neighbors

from neighbors’ interaction profiles. Through systematic REFERENCES

experiments with benchmark datasets, we demonstrated the Bleakley,K. and Yamanishi,Y. (2009) Supervised prediction of drug–target inter-
effectiveness of BLM–NII for predicting interactions between actions using bipartite local models. Bioinformatics, 25, 2397–2403.
new drug-candidate compounds and new target-candidate Campillos,M. et al. (2008) Drug target identification using side-effect similarity.
proteins. Science, 321, 263–266.
Chen,X. et al. (2012) Drug–target interaction prediction by random walk on the
In the presented approach, we allow all the neighbors to par-
heterogeneous network. Mol. BioSyst, 8, 1970–1978.
ticipate in training data inferring. To allow only neighbors with Cheng,F. et al. (2012) Prediction of drug–target interactions and drug repositioning
large similarities to contribute, a threshold may be used to reduce via network-based inference. PLoS Comput. Biol., 8, e1002503.
the impact of those non-important neighbors to 0. Alternately, a Gunther,S. et al. (2008) Supertarget and matador: resources for exploring drug–
Gaussian function may be introduced to gradually decrease the target relationships. Nucleic Acids Res., 36, D919–D922.
Hattori,M. et al. (2003) Development of a chemical structure comparison method
influence of neighbors based on their distances to the new drug/
for integrated analysis of chemical and genomic information in the metabolic
target candidate in query. pathways. J. Am. Chem Soc., 125, 11853–11865.
In the current work, we only apply the NII procedure for those Haupt,V.J. and Schroeder,M. (2011) Old friends in new guise: repositioning of
completely new candidates that have no existing training data at known drugs with structural bioinformatics. Breif. Bioinform, 12, 312–326.
all, and we find that the results are already good enough to show He,Z. et al. (2010) Predicting drug–target interaction networks based on functional
groups and biological features. PLoS One, 5, e9603.
the usefulness of NII. Since it is quite common that drugs only
Kanehisa,M. et al. (2006) From genomics to chemical genomics: new developments
activate or inhibit a small number of targets and targets are only in kegg. Nucleic Acids Res., 34, D354–D357.
activated or inhibited by very limited drugs, the NII procedure Keiser,M.J. et al. (2009) Predicting new molecular targets for known drugs. Nature,
may be applied to drugs and targets which do not have sufficient 462, 175–181.
training data. We expect that more accurate prediction models Laarhoven,T.V. et al. (2011) Gaussian interaction profile kernels for predicting
drug–target interaction. Bioinformatics, 27, 3036–3043.
may be build by using neighbors’ information to enhance the
Laggner,C. et al. (2012) Chemical informatics and target identification in a zebrafish
limited training examples. However, too much emphasis on phenotypic screen. Nat. Chem. Biol., 8, 144–146.
neighbors tends to eliminate the local characteristics of each Martin,Y.C. et al. (2002) Do structurally similar molecules have similar biological
drug and target and could cause deterioration in the prediction activity? J. Med. Chem., 45, 4350–4358.
performance. Nevertheless, it would be an interesting future Perlman,L. et al. (2011) Combining durg and gene similarity measures for drug–
target elucidation. J. Comput. Biol., 18, 133–145.
work to explore the balance between local information and
Schomburg,I. et al. (2004) Brenda, the enzyme database: updates and major new
global information in model learning. developments. Nucleic Acids Res., 32, D431–D433.
Wishart,D.S. et al. (2008) Drugbank: a knowledgebase for drugs, drug actions and
Funding: This research was supported by Singapore MOE AcRF drug targets. Nucleic Acids Res., 36, D901–D906.
(MOE2008-T2-1-074) and Startup (M4080108.020) from Xia,Z. et al. (2010) Semi-supervised drug–protein interaction prediction from het-
Nanyang Technological University, Singapore. erogeneous biological spaces. BMC Syst. Biol., 4 (Suppl. 2), S6.
Yamanishi,Y. et al. (2008) Prediction of drug–target interaction networks from the in-
Conflict of Interest: none declared. tegration of chemical and genomic spaces. Bioinformatics, 24, i232–i240.

245