Drug target discovery using knowledge graph embeddings

Abstract

The field of drug discovery has entered a plateau stage lately. It is increasingly more expensive and time-demanding to introduce new drugs into the market. One of the main reasons is the slow progress in finding novel targets for drug candidates and the lack of insight in terms of the associated mechanisms of action. Current works in this area mainly utilise different chemical, genetic and proteomic methods, which are limited in terms of the scalability of experimentation and the scope of studied drugs and targets per experiment. This is mainly due to their dependency on laboratory experiments and available physical resource. This has led to an increasing importance of computational methods for the identification of candidate drug targets. In this work, we introduce a novel computational approach for predicting drug target proteins. We approach the problem as a link prediction task on knowledge graphs. We process drug and target information as a knowledge graph of interconnected drugs, proteins, disease, pathways and other relevant entities. We then apply knowledge graph embedding (KGE) models over this data to enable scoring drug-target associations, where we employ a customised version of state-of-the-art KGE model ComplEx. We generate a benchmarking dataset based on KEGG database to train and evaluate our method. Our experiments show that our method achieves best results in comparison to other traditional KGE models. Specifically, the method predicts drug target links with mean reciprocal rank (MRR) of 0.78 and [email protected] of 0.88. This provides a promising basis for further experimentation and comparisons with domain-specific predictive models.

Full text

Provided by the author(s) and NUI Galway in accordance with publisher policies. Please cite the published
version when available.
Downloaded 2019-06-13T11:25:52Z
Some rights reserved. For more information, please see the item record link above.
Title Drug target discovery using knowledge graph embeddings
Author(s) Mohamed, Sameh K.; Nováek, Vít; Nounu, Aayah
Publication
Date 2019-04-08
Publication
Information
Mohamed, Sameh K., Nováek, Vít, & Nounu, Aayah. (2019).
Drug target discovery using knowledge graph embeddings.
Paper presented at the 34th ACM/SIGAPP Symposium on
Applied Computing (SAC ’19), Limassol, Cyprus, 08-12 April.
Publisher Association for Computing Machinery
Link to
publisher's
version https://doi.org/10. 1145/3297280.3297282
Item record http://hdl.handle.net/10379/15065
DOI http://dx.doi.org/10. 1145/3297280.3297282

Drug Target Discovery Using Knowledge Graph Embeddings
Sameh K. Mohamed
Data Science Institue
Insight Centre for Data Analytics
National University of Ireland Galway
sameh.kamal@insight-centre.org
Aayah Nounu
MRC Integrative Epidemiology Unit
University of Bristol
An0435@bristol.ac.uk
Vít Nováček
Data Science Institue
Insight Centre for Data Analytics
National University of Ireland Galway
vit.novacek@insight-centre.org
ABSTRACT
The eld of drug discovery has entered a plateau stage lately. It
is increasingly more expensive and time-demanding to introduce
new drugs into the market. One of the main reasons is the slow
progress in nding novel targets for drug candidates and the lack
of insight in terms of the associated mechanisms of action. Current
works in this area mainly utilise dierent chemical, genetic and
proteomic methods, which are limited in terms of the scalability
of experimentation and the scope of studied drugs and targets per
experiment. This is mainly due to their dependency on laboratory
experiments and available physical resource. This has led to an
increasing importance of computational methods for the identica-
tion of candidate drug targets. In this work, we introduce a novel
computational approach for predicting drug target proteins. We ap-
proach the problem as a link prediction task on knowledge graphs.
We process drug and target information as a knowledge graph of
interconnected drugs, proteins, disease, pathways and other rele-
vant entities. We then apply knowledge graph embedding (KGE)
models over this data to enable scoring drug-target associations,
where we employ a customised version of state-of-the-art KGE
model ComplEx. We generate a benchmarking dataset based on
KEGG database to train and evaluate our method. Our experiments
show that our method achieves best results in comparison to other
traditional KGE models. Specically, the method predicts drug tar-
get links with mean reciprocal rank (MRR) of 0.78 and Hits@10 of
0.88. This provides a promising basis for further experimentation
and comparisons with domain-specic predictive models.
CCS CONCEPTS
•Semantic networks
;
•Machine Learning
;
•Machine learned
ranking;
KEYWORDS
Drug Target Discovery, Knowledge Graph Embeddings, Link Pre-
diction
ACM Reference Format:
Sameh K. Mohamed, Aayah Nounu, and Vít Nováček. 2019. Drug Target
Discovery Using Knowledge Graph Embeddings. In The 34th ACM/SIGAPP
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
SAC ’19, April 8–12, 2019, Limassol, Cyprus
©2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-5933-7/19/04.. .$15.00
https://doi.org/10.1145/3297280.3297282
Symposium on Applied Computing (SAC ’19), April 8–12, 2019, Limassol,
Cyprus. ACM, New York, NY, USA, Article 4, 8 pages. https://doi.org/10.
1145/3297280.3297282
1 INTRODUCTION
The development of drugs has a long history [
6
]. Until quite recently,
pharmacological eects were often discovered using primitive trial
and error procedure, applying plant extracts on living system and
observing the outcomes. Later, drug development evolved to eluci-
dating mechanisms-of-actions of drug substances and their eects
on phenotype. The ability to pharmacologically isolate active sub-
stances was a key step towards modern drug discovery [
34
,
35
].
More recently, advances in molecular biology and biochemistry
allowed for more complex analysis of drugs, their targets and their
mechanisms of action. The study of drug targets has become very
popular, where studies utilise dierent chemical genetic [
35
] and
proteomic methods [
33
] such as anity chromatography and ex-
pression cloning approaches. These, however, can only process
limited number of possible drugs and targets due to dependency
on laboratory experiments and available physical resource. Com-
putational approaches have therefore been extensively studied
lately [17, 18, 42].
In this work we introduce a specic computational approach for
predicting drug target proteins. Our objective is to score possible
associations between drugs and proteins according to the proba-
bility of the association holding true. The ultimate goal is to assist
lab experimentation in narrowing the scope of possible new drug
targets investigated. In the current drug target knowledge bases
like DrugBank [
40
] and KEGG [
12
], information about drugs con-
tains their relationship with target proteins (or their genes), action
pathways and targeted diseases. These components are represented
as graphs form of interconnected entities and relations. Such data
can naturally be interpreted as a knowledge graph, where the task
of nding new associations between drugs and their targets can
be formulated as a link prediction task. In this context, knowledge
graph embedding (KGE) models are a t natural application, where
they are known to provide state-of-the-art results in link prediction
on knowledge graphs [
23
]. Despite the growing body of computer
simulation based drug target prediction frameworks [
39
,
42
,
43
],
none of these works utilise knowledge graph embedding models in
their predictive pipelines.
The objective of this work is to demonstrate the usefulness of
knowledge graph embedding models in the area of drug target
prediction. We also identify KGE techniques that can provide the
best accuracy in predicting drug targets. This is presented as a
stepping stone towards a domain-specic KGE-based drug target
prediction model and its extensive comparison with existing related
models such as the DDR [25] and the DNILMF [11] models.

SAC ’19, April 8–12, 2019, Limassol, Cyprus Sameh K. Mohamed, Aayah Nounu, and Vít Nováček
−
x
x′
−Corruptor
Entity or relation
Embedding
ηScoring function
≈Activation
LLoss function
▽Gradiant Update
NEmbbeding Norm
Exη
Ex′
η
≈
≈
f(x)
f(x′)
L
∇E
▽
E
NE
(2) Lookup(1) Corruption (3) Scoring (4) Loss (5) Update (6) Norm
Input
Figure 1: Phases of one epoch training of a KGE model over one training instance
Our knowledge graph embedding approach, ComplEx-SE (Complex
embeddings with squared error loss), is a customisation of the state-
of-the-art KGE model ComplEx [
38
] with a square error-based
training objective. We build a drug-target centred dataset from
KEGG database to train and evaluate our model, and we show by
experiments that it achieves best results in terms of predicting
new drug target links. To the best of our knowledge, there are
currently no models for discovering new drug targets using KGE
model for link prediction on biomedical knowledge graph, so we
evaluate our method compared to other state-of-the-art KGE models
like the Translating Embeddings (TransE) model [
2
], the DistMult
model [44] and the Complex Embedding (ComplEx) model [38].
The rest of this paper is structured as follows: section 2 discuss the
problem of drug target discovery and presents fundamental back-
ground concepts about KGE models and their evaluation metrics.
Section 4 present the subset of KEGG that we use, and discusses its
component. Section 5 presents our KGE approach and its base work
the ComplEx model. Section 6 present the experimental setup, the
evaluation protocol, section 3 discusses similar related works that
uses computational-based approaches for nding drug targets and
section 7 discusses the results of our experiments and lesson learnt.
We nally present our conclusions and possible future works in
section 9.
2 BACKGROUND
In this section, we discuss the advantages and implications of nd-
ing drug targets that are not yet known. We also discuss mod-
elling information in knowledge graphs, the underlying concepts
of knowledge graph embedding models and their evaluation tech-
niques.
2.1 Drug Target Discovery
The process of discovering and developing drugs with one gene
target requires time and money. Rarely does a drug only bind to its
intended target, but rather o-target eects are common [
41
], and
this may lead to unwanted side-eects [
3
]. Conversely, the o-target
eects may be useful for drug-repurposing reasons. Drug repur-
posing is dened as the use of approved drugs for new diseases [
4
].
It is believed to take around 10-17 years from the conception of a
drug to when it becomes a licensed treatment for disease, with a
success rate of less than 10% [
1
]. Drug repurposing is advantageous
as the safety prole of the drug is already known and reduces the
time and cost required to bring a new drug into the clinic [4].
The identication of new protein targets also allows the develop-
ment of drugs that specically target the protein of interest. For
example, aspirin is currently being considered for use as a chemo-
preventative agent [
27
,
30
,
31
] but there are concerns with regards
to side-eects caused by its long-term use, such as upper gastroin-
testinal bleeding [
16
]. By identifying the exact protein targets of
aspirin, new drugs can be developed specically for these proteins
to avoid the unwanted side-eects.
The use of computational approaches is useful as they are free from
bias and are therefore not inuenced by prior knowledge and opin-
ions, unlike laboratory-based methods. These approaches bypasses
the need to spend a long amount of time in the laboratory and can
be used to provide guidance on the direction of research within
the laboratory, therefore saving both time and money. Follow-up
experiments can then be carried out in the laboratory for conr-
mation of the new proteins targeted by the drug, allowing direct
conclusions to be made of treatment eect.
Overall, computational approaches are considered useful methods
to identify o-target interactions and can also be used for the pos-
sibility of drug repurposing. As they reduce the time required to
manually discover other unintended protein targets and may reduce
the large costs required in doing so.
2.2 Knowledge Graphs
Knowledge graphs are a data representation that model relational
information as a graph, where the graph nodes represent knowl-
edge entities and its edges represent relations between them. They
model facts as (subject, predicate, object) (SPO) triples e.g. (Aspirin,
Drug-Target,COX-1), where a subject entity is connected to an ob-
ject entity through a predicate relation.
In recent years, knowledge graphs have become a popular means
for data representation in the semantic web community to create
the "web of data", which is a network of interconnected entities

Drug Target Discovery Using Knowledge Graph Embeddings SAC ’19, April 8–12, 2019, Limassol, Cyprus
that can be easily interpreted by both humans and machines [
36
],
where knowledge graphs are used to model linked data. They have
also been used as convenient means for modelling information in
many dierent domains, including general human knowledge [
15
],
biomedical information [
7
] and language lexical information [
19
].
Knowledge graphs are now used in dierent applications such as
enhancing semantics of search engine results [
26
,
32
], biomedical
discoveries [
21
], or powering question answering and decision
support systems [8].
2.3 Knowledge Graph Embedding
Knowledge graph embedding models learn a low rank vector rep-
resentation of knowledge entities and relations that can be used
to rank knowledge assertions according to their factuality. KGE
models are trained in a multi-phase procedure as shown in Fig. 1,
where their objective is to eectively learn a vector representation
of entities and relations that can be used to score and rank possible
knowledge facts.
First, a KGE model initialises all embedding vectors using random
noise values. It then uses these embeddings to score the set of true
and false training facts using a model-dependent scoring function.
The output scores are then passed to the training loss function
to compute training error as shown in Fig. 1. These errors are
used by optimisers like AMSGrad [
28
] to generate gradients and
update the initial embeddings, where the updated embeddings give
higher scores for true facts and less for false facts. This procedure
is performed iteratively for a set of iterations i.e. epochs in order to
reach a state where embeddings provide best possible scoring for
both true and false possible facts.
2.4 Ranking Metrics
In the following, we present the metrics that we use in the evalua-
tion of our approach.
(1) Mean reciprocal rank (MRR):
This is the harmonic mean
of the rank position of the rst relevant element, and it is dened
as follows:
MRR =1
|Q|
|Q|
X
i=1
1
ranki
where
ranki
refers to the rank position of the rst relevant element
for the
i
-th query. The output values of mean reciprocal rank are
normalised from 0 to 1, where 1 represents perfect ranking and
decaying values towards 0 represent decreasing accuracy.
(2) Hits@k:
This is the number of correct elements predicted
among the top-
k
elements, where we use Hits@1, Hits@3 and
Hits@10. This metric indicates that the model’s probability of rank-
ing the relevant (true) fact in the top kelement scores in the rank.
3 RELATED WORK
In this section we discuss related works where we target two kinds
of activities. Firstly, other computer based approaches for predicting
drug targets. Secondly, relation link prediction approaches and
state-of-the-art knowledge graph embedding models.
3.1 Computer Based Drug Target Prediction
Yamanishi et al
. [42]
developed one of the early computational ap-
proaches to predict drug targets, where their approach utilised a
statistical model that infers drug targets based on a bipartite graph
of both chemical and genomic information. More recent works like
COSINE [
29
] and NRLMF [
17
] approaches introduced the use of
drug-drug and target-target similarity measures to infer possible
drug targets. These approaches enabled new drugs and drug tar-
gets with limited or no information about their interaction data
since they depend on the drug-drug and target-target similarities.
However, these methods only utilised a single measure to model
components similarity.
Other drug target prediction models like KronRLS-MKL [
22
] and
BLM-NII [
18
] integrated dierent similarity measures to model
the similarity between drugs and targets. These approaches use
both linear and non-linear combinations of similarity measures
to encode the similarity between drugs and their targets, where
non-linear combinations provided better predicting drug-target
predictions [18].
Recently, Hao et al
. [11]
proposed a model that uses matrix fac-
torisation to predict drug targets over drug information networks.
Their model, DNILMF, operates in a four-step procedure. First, it
infers dierent proles for both drugs and targets and constructs
kernel matrices for these proles. It then diuses drug proles ker-
nel matrices with their structure kernel matrices. It then diuses
target proles kernel matrices with their sequence kernel matrices.
Finally, the DNILFM model uses the outputs of the previous steps
to predict drug targets based on their network neighbours. This
approach showed signicant predictive accuracy improvements
over other methods on standard benchmarking datasets [11, 42].
In most recent times, the current state-of-the-art work on compu-
tation drug target discovery is the DDR model [
25
], which predicts
drug targets using heterogeneous graphs that contain drug target
interactions in a multiple phases procedure. First, it computes simi-
larity indices for drugs and their targets. It then selects a subset of
these similarities in a heuristic process to obtain optimal combina-
tions of similarities. Finally, it combines selected similarities using
a non-linear fusion technique, and combines diusion output with
random walk features from the heterogeneous graphs to predict
drug targets. Despite the complexity of the the DDR model, it cur-
rently provides state-of-the-art results in predicting drug targets
using computational approaches [25].
3.2 Link Prediction in Knowledge Graph
In recent years, various predictive frameworks were developed
to predict new links in knowledge graphs, where these frame-
works serve in various applications such as semantic search en-
gines [
26
,
32
], biomedical discoveries [
21
], and question answering
systems [
8
]. Link prediction models can be categorised into two cat-
egories: graph-feature based models and latent-feature based mode.
Graph-feature based models utilise graph features like paths and
graph patterns to predict possible connecting links between graph
entities. For example, the path ranking algorithm (PRA) [
14
] uses

SAC ’19, April 8–12, 2019, Limassol, Cyprus Sameh K. Mohamed, Aayah Nounu, and Vít Nováček
P
DG
SN
Drug Target Gene
Drug Pathway
Drug Disease
Disease Gene
Disease Pathway
Disease Network
Gene Pathway Gene Network
Network Pathway
Figure 2: Knowledge graph about drugs, their target genes,
pathways, diseases and gene variant networks extracted
from KEGG.
connecting paths between entities generated by random walks to
infer possible links between them, where other models like the sub-
graph feature extraction model (SFE) [
9
] and the distinct subgraph
path (DSP) [
20
] employ a combination of connecting path and sub-
graph paths of two entities to predict their possible associations. On
the other hand, latent-feature based models i.e. knowledge graph
embedding models, use a generative approach to learn low-rank em-
beddings for knowledge entities and relations in order to score their
possible associations. These approaches use multiple techniques
like tensor factorisation as in the DistMult model [
2
] and latent dis-
tance similarity as in the TransE model [
44
] to model possible inter-
actions between graph embeddings and provide scores for possible
graph links. For further information on both approaches, Nickel
et al
. [23]
provides an extended review for both graph-feature based
and latent-feature based models in the task of link prediction in
knowledge graphs.
4 DATA
In this section we discuss the KEGG database [
12
,
13
] with focus
on the components that we use to train and evaluate our approach.
KEGG is a knowledge base that contains information about bio-
logical systems like cells and organisms at the molecular level. It
contains dierent types of biological information entities like genes,
pathways, drugs, disease, etc. The data in KEGG is structured as a
network of inter-connected entities that resembles the biological
eco-system at the molecular level. In our study, we focus on infor-
mation related drugs and their targets, where we only the following
KEGG components are considered:
(1) Drugs
: The KEGG drug database
1
is a comprehensive drug in-
formation resource for approved drugs. It contains multiple types of
information about drugs such as the chemical structure, associated
1https://www.genome.jp/kegg/drug/
Table 1: Statistics of objects and their inter-connections in
the subset of KEGG dataset that we use.
Object Count
Drug
Gene
Pathway
Disease
Network
Drug 4670 •12004 7910 2160 0
Gene 8881 12004 •497 239 4534
Pathway 329 7910 497 •1803 524
Disease 1873 2160 239 1803 •441
Network 448 0 4534 524 441 •
targets, action pathways and targeted diseases. In our study, we
only consider drug associations to the elements specied in Table 1.
(2) Genes
: The KEGG Gene database
2
contains information about
genes, their sequences and their associations with other biological
entities. While drug targets in the living systems are usually pro-
teins, the KEGG database uses genes as a representation of drug
targets, where genes represent their product proteins. In the rest of
this study, we use the KEGG genes to represent product proteins
as drug targets in the knowledge graph we have created.
(3) Pathways
: The KEGG database also contains information about
biological pathways associated with manually curated maps of their
reactions. The pathway database
3
in KEGG includes pathways of
dierent activities such metabolism, environmental information
processing, human disease, etc. Each pathway is associated to its
related entities e.g. genes, drugs and diseases, where we use such
associations to construct our knowledge about pathways.
(4) Diseases
: The KEGG disease database
4
is structured in a similar
form as in the previously mentioned entity databases, where our
main interest is to utilise the associations between disease and our
other investigated entities. However, the disease database contains
associations to other entity types such as carcinogens that can be
helpful to extend the knowledge about specic cancerous disease
in future studies.
(5) Networks
: The KEGG network database
5
contains information
on the perturbation of human genes, where it encodes knowledge
about the dierent variants and other perturbants of human genes
that are involved in the perturbed molecular reaction networks.
Similarly, instances of the network database are linked to their
related entities in other KEGG databases.
In our study, we gather all the possible associations between the
previously mentioned KEGG entities to generate a biological knowl-
edge graph that is centred around drugs and their target genes as
shown in Fig. 2. The statistics of the counts of each entity type and
the inter-connecting links between entities are provided in Table 1.
2https://www.genome.jp/kegg/genes.html
3https://www.genome.jp/kegg/pathway.html
4https://www.genome.jp/kegg/disease/
5https://www.kegg.jp/kegg/network.html

Drug Target Discovery Using Knowledge Graph Embeddings SAC ’19, April 8–12, 2019, Limassol, Cyprus
-  
-


  
- 
-


 
Negatives Positives   
Figure 3: Plots of loss growth of squared error and logistic
pointwise loss functions compared to scores of positive and
negative instances.
5 OUR APPROACH
In this section, we present the technical details of our approach,
which is a modication upon the the ComplEx knowledge graph
embedding model [
38
]. We discuss the ComplEx model, its scoring
and loss functions and our modications.
5.1 Complex Scoring Function
The ComplEx model is a tensor factorisation based knowledge
graph embedding model. It represents knowledge entities and re-
lation using complex vector embeddings, where each embedding
is represented by two vectors (real and imaginary). Fact assertions
in the ComplEx model are evaluated using a factorisation based
scoring function that is dened as follows:
fComplEx(s,p,o)=
K
X
k=1
Re(<es k ,erk ,eok >)
where
es
,
er
and
eo
are the embeddings of the subject, the relation
and the object respectively,
esk
is the
k
-th component of the em-
bedding
es
,
K
is the embedding size (vector length),
Re (x)
is the
real part of complex value
x
and
x
is the complex conjugate of
x
such that
x=a−ib
if
x=a+ib
. This formulation can be further
relaxed as follow:
fComplEx(s,p,o)=
K
X
k=1
er
sk er
r k er
ok +ei
sk er
r k ei
ok +er
sk ei
r k ei
ok −ei
sk ei
r k er
ok
where
xr
and
xi
are the real and imaginary parts of complex value
x
respectively. The use of the complex conjugate in the ComplEx
model allows it to encode embedding interactions in an asymmetric
operation, which enables it to model facts with both symmetric and
asymmetric predicates unlike other factorisation based models like
the DistMult model [44].
5.2 Training Objective
In the task of link prediction, knowledge graph embedding models
are considered learning to rank models, where they employ tradi-
tional ranking functions e.g. pointwise and pairwise ranking losses
to model their training loss. The ComplEx model by default uses a
pointwise ranking loss with a negative-logistic transformation to
model its training loss, which is dened as follows:
L
logisticP t
=X
x∈T
log(1+exp(−l(x)·f(x))),(1)
where
x
is an (s, p, o) fact and
T
is the set of all training facts with
negative samples and
l(x)
is the true label of fact
x
such that
x
is
equal to 1 when true and -1 otherwise. This allows the models to
eectively update the embeddings of both entities and relations
to give high scores to true facts and lower scores to false facts as
shown in Fig. 3.
5.3 New Loss Objective for The ComplEx model
Trouillon and Nickel
[37]
have shown that the choice of objective
loss in KGE models has a huge impact on their predictive accuracy.
They showed that despite the equivalence of both the ComplEx and
the Holographic embedding (HolE) [
24
] models, they vary in accu-
racy due to their dependency on dierent training loss objectives.
This dierence is caused by the fact that the HolE model uses a
max-margin loss while the ComplEx model uses a log-likelihood
loss. Following their remarks, in this work, we propose a new loss
objective to the ComplEx model, and we show that it suit the lim-
ited size and number of predicates in our dataset. We propose a new
loss function based on the square error of the dierence between
ComplEx scores assertions and their true labels using a 0 and 1
labelling such that 0 represents false fact assertions and 1 represents
true fact assertions. The new square error based loss is dened as
follows:
L
SEP t
=X
x∈T
1
2(l(x)−f(x))2(2)
where
l(x)
is the label of fact
x
with
l(x)=
0 if
x
is false and 1
otherwise. This also allows the square error loss to force embedding
updates that produce normalised scores around 0 and 1 unlike the
logistic loss with an open range of scores.
We prove that our new representation of ComplEx model training
loss outperforms its default version using an empirical evaluation
framework described in Section 6. In the following, we discuss some
properties of both the logistic and square error based losses.
Fig. 3 shows the dierences between the growth of both the
square error based loss and ComplEx’s default logistic loss, where
both functions show dierent loss growth rates which aect the
growth rate of the values of their output gradients. The logistic
loss has a linear growth rate and its gradient per single instance is
dened as follows:
∆x=exp(f(x))/[1 +exp(f(x)) ] (3)
where this form grows in a sub-linear sigmoid fashion. This limits
the output gradients for each training instance in the range of [0
,
1].
On the other hand, the square growth of square error loss yields
linearly growing gradients dened as follows:
∆x=









0 for f(x)−l(x)=0
f(x)for f(x)−l(x)<0
−f(x)for f(x)−l(x)>0
(4)
which enables the ComplEx model to produce gradients within the
range [0,∞).

SAC ’19, April 8–12, 2019, Limassol, Cyprus Sameh K. Mohamed, Aayah Nounu, and Vít Nováček
spo
Test triples
.
.
.
.
.
.
.
.
.
Corruptions
|E| × 3
corrupt
|E| × 3
corrupt
scoring
Scores
scoring
ltering
Filtered scores Filtered metrics
ltering
compute
metrics
R RR H@n
compute
metrics
R RR H@n
DrugsTargets
corruptionscorruptions
Figure 4: Evaluation protocol of KGE models for a single drug target testing instance. Note: R represents rank, RR represents
reciprocal rank, and H@n represents Hits@n
Table 2: Statistics of entities, relations, facts and drug-target
fact count per split of KEGG50k dataset
Dataset Entities Relations Facts DT-Facts
KEGG50k-full 16201 9 63080 12004
KEGG50k-train 16201 9 57080 10769
KEGG50k-valid 16201 9 3000 585
KEGG50k-test 16201 9 3000 650
6 EXPERIMENTS
In this section we describe the setup of our experiments and the
evaluation pipeline.
6.1 Dataset
In our experiments, we divide the KEGG dataset subset into train-
ing, validation and testing splits with ratios of 90%, 5% and 5%
respectively. All the splits contain facts that describe all entities
and relations, where the drug-target links are distributed among
the three splits with the same ratio as the splits sizes. Table 2 shows
statistics about the dataset and its splits in terms of number of en-
tities, relations, facts and drug-target facts. The KEGG50k dataset
can be downloaded from gshare 6.
6.2 Implementation
We use Tensorow framework (GPU) along with Python 3.5 to
perform our experiments. All experiments were executed on a
Linux machine with processor Intel(R) Core(TM) i70.4790K CPU @
4.00GHz, 32 GB RAM, and an nVidia Titan Xp GPU.
6.3 Evaluation protocol
KGE models are evaluated using a unied protocol that assesses
their performance in the task of link prediction. Let
X
be the set
of facts, i.e. triples,
ΘE
be the embeddings of the set of all entities
E
, and
ΘR
be the embeddings of the set of all relations
R
. The KGE
6KEGG50k dataset is found at: https://gshare.com/s/bbfc7b82d17e0b8b6a43
evaluation protocol works in four steps (Fig. 4 shows a visual ow
of the evaluation process steps for a single triple instance):
(1) Corruption
Let
x=(s,p,o)∈X
, then for each
x
, it is corrupted
2
|E| −
1 times by replacing its subject and object entities with all
the other entities in
E
. The corrupted triples can be dened as:
xcorr =Ss′∈E(s′,p,o)∪So′∈E(s,p,o′)
, where
s′,s
and
o′,o
.
These corruptions eectively provide negative examples for the
supervised training and testing processes due to the Local Closed
World Assumption [23].
(2) Scoring
: Both original triples and their corrupted instances
are evaluated using a model-dependent scoring function. This pro-
cess involves looking up embeddings of entities and relations, and
computing scores depending on these embeddings using the model-
dependent scoring function.
(3) Filtering
: It is possible that corruptions of triples may contain
positive instances that exist among training or validation triples.
This problem is alleviated by ltering out scores of positive in-
stances in the triple corruptions.
(4) Computing metrics
: Each triple and its corresponding subject
and object corruption triples produce two sets of ltered scores
following previous evaluation steps. Then, for each set of ltered
scores, the KGE model computes rank, reciprocal rank, and hits@n
metrics.
6.4 Experimental Setup
In the experiments, we use state-of-the-art KGE models the TransE [
2
],
the DistMult [
44
], and the ComplEx [
38
] models compared to our
customised version of the ComplEx model to perform link predic-
tion over KEGG50k dataset in two settings. First, the general link
prediction setting, where the objective is to learn rank all the link
in the testing set according to their factuality compared to their all
other possible corrupted assertions. Second, the drug target link
prediction setting, where the same previous procedure is applied
only to drug target links.

Drug Target Discovery Using Knowledge Graph Embeddings SAC ’19, April 8–12, 2019, Limassol, Cyprus
Table 3: Link prediction results over KEGG50k dataset on both general links and drug target links only. For all the metrics
except for mean rank (Rank), the higher the value the better.
Model General dataset links Drug target links
Rank MRR Hits@1 Hits@3 Hits@10 Rank MRR Hits@1 Hits@3 Hits@10
TransE [2] 192 0.46 0.38 0.50 0.63 81 0.75 0.69 0.79 0.86
DistMult [44] 430 0.37 0.27 0.42 0.57 186 0.61 0.50 0.69 0.81
ComplEx [38] 506 0.39 0.31 0.43 0.57 208 0.68 0.61 0.71 0.82
ComplEx-SE (This work) 534 0.52 0.45 0.56 0.68 145 0.78 0.73 0.81 0.88
We run all the models over the previously mentioned benchmark-
ing dataset KEGG50k. A grid search is performed to obtain best
hyper parameters for each model, where the set of investigated
parameters are: embeddings size
K∈ {
50
,
100
,
150
,
200
}
, margin
λ∈ {
1
,
2
,
3
,
4
,
5
}
for the TransE and the DistMult models, and num-
ber of negative samples
n∈ {
2
,
4
,
6
,
10
}
. All embeddings vectors of
our models are initialised using the uniform Xavier random initial-
izer [
10
]. For all the experiments, we use batches of size 5000, with
a maximum of 1000 training iterations i.e. epochs. The gradient
update procedure is performed using the AMSGrad optimiser [
28
]
with a xed 0.01 learning rate.
7 RESULTS AND DISCUSSION
Table 3 shows the output results of our experiments, where the
experiments are executed in two congurations: general link pre-
diction over all association types and link prediction over drug
targets associations only.
The results show that our approach, Complex-SE, outperforms
other state-of-the-art models in terms of MRR, Hits@1, Hits@3 and
Hits@10 on both experiment congurations, where it achieve a
better MRR with a 6% margin in predicting general links and a 3%
MRR margin over other models in predicting drug-target links.
The link prediction task is executed such that for each investi-
gated possible drug-target association such as (Aspirin,Drug-Target,
COX2) each model is required to answer two questions: (1) Which
drug targets COX2? and (2) Which target does the drug Aspirin
target? A model has to choose an answer from the set of all vocab-
ulary entities, where all correct answers except for drug Aspirin
and target COX2 are removed. The answers to these questions are
formatted as a rank, where the model is required to position the cor-
rect answer in the rst place in its rank to achieve perfect accuracy
as shown in Fig. 4, which presents the ow of the link prediction
evaluation pipeline for one test instance. In this setting, a random
baseline model would choose the right answers in the rst position
of the rank with a probability of
1
|E|
, where
|E|
is the size of the set
of all entities vocabulary, this is equal to 16201 in our experiments.
Our knowledge graph embedding model, ComplEx-SE, is able to
identify the correct answers for both of the previous questions
with a mean reciprocal rank of 0.78, where it identies the correct
answer within the rank with probabilities of 0.73, 0.81, 0.88 at the
rst, the third, the tenth positions respectively.
The use of knowledge graph embedding approaches such as ComplEx-
SE enables predicting new associations for both new drugs and new
targets since it does not depend on their interaction proles. KGE
models are also capable of learning dierent types of associations
between entities of dierent types with no extra congurations as
shown in the general link prediction conguration in table 3. For
example, they can be used to identify the relation between proteins
and pathways, or the relation between drugs and pathways. This
can lead to discovering further unknown activities for both drugs
and proteins with no extra computational cost.
8 ACKNOWLEDGEMENTS
This work has been supported by Insight Centre for Data Analytics
at National University of Ireland Galway, Ireland (supported by
the Science Foundation Ireland grant 12/RC/2289). The GPU card
used in our experiments is granted to us by the Nvidia GPU Grant
Program.
9 CONCLUSIONS AND FUTURE WORK
In this work, we introduced the use of knowledge graph embedding
models for predicting drug targets using currently available drug
knowledge bases, where we formulated the problem as a link pre-
diction task over drug targets centred knowledge graphs. We have
created a knowledge graph dataset, KEGG50k, from KEGG database,
which is centred around drugs and their targeted genes, disease
and reaction pathways. We then used this dataset to evaluate the
predictive accuracy of knowledge graph embedding models.
We proposed a knowledge graph embedding approach, ComplEx-
SE that is a customised version of the ComplEx model with a square
error based loss function, and we showed by empirical evaluation
that our approach outperforms other state-of-the-art knowledge
embedding models in the task of predicting drug target links over
KEGG database. Our results showed that the ComplEx-SE approach
is able to identify drug target link with a mean reciprocal rank of
0.78 with a 3% margin better than other state-of-the-art knowledge
graph embedding models. Results also showed that our approach
is able to provide a rank of 16201 possible drug-target association
statements with only one true statement, where it identies the
true state with probabilities of 0.73, 0.81 and 0.88 at the rst, the
third and the tenth positions of the rank.

SAC ’19, April 8–12, 2019, Limassol, Cyprus Sameh K. Mohamed, Aayah Nounu, and Vít Nováček
Despite the growing body of research on computer based approaches
for predicting drug targets, our objective in this work was limited to
evaluating knowledge graph embedding approaches and identifying
their optimal techniques for predicting drug-targets. However, in
future work, we aim to perform a comparison between our knowl-
edge graph embedding technique and other state-of-the-art drug
target discovery computer based approaches based on the bench-
marking dataset and evaluation metrics. We also aim to perform
in-lab experimental evaluation for the top predicted drug target
associations that is not in the currently public available knowledge
bases to validate possible new undiscovered drug targets.
REFERENCES
[1]
Ted T Ashburn and Karl B Thor. 2004. Drug repositioning: identifying and
developing new uses for existing drugs. Nature reviews Drug discovery 3, 8 (2004),
673.
[2]
Antoine Bordes, Nicolas Usunier, Alberto García-Durán, Jason Weston, and Ok-
sana Yakhnenko. 2013. Translating Embeddings for Modeling Multi-relational
Data. In NIPS. 2787–2795.
[3]
Joanne Bowes, Andrew J Brown, Jacques Hamon, Wolfgang Jarolimek, Arun
Sridhar, Gareth Waldron, and Steven Whitebread. 2012. Reducing safety-related
drug attrition: the use of in vitro pharmacological proling. Nature reviews Drug
discovery 11, 12 (2012), 909.
[4]
Anne Corbett, James Pickett, Alistair Burns, Jonathan Corcoran, Stephen B Dun-
nett, Paul Edison, Jim J Hagan, Clive Holmes, Emma Jones, Cornelius Katona,
et al
.
2012. Drug repositioning for Alzheimer’s disease. Nature Reviews Drug
Discovery 11, 11 (2012), 833.
[5]
Michael Dickson and Jean Paul Gagnon. 2009. The cost of new drug discovery
and development. Discovery medicine 4, 22 (2009), 172–179.
[6]
Jürgen Drews. 2000. Drug Discovery: A Historical Perspective. Science 287, 5460
(2000), 1960–1964. https://doi.org/10.1126/science.287.5460.1960
[7]
Michel Dumontier, Alison Callahan, Jose Cruz-Toledo, Peter Ansell, Vincent
Emonet, François Belleau, and Arnaud Droit. 2014. Bio2RDF Release 3: A larger,
more connected network of Linked Data for the Life Sciences. In Proceedings of
the ISWC 2014 Posters & Demonstrations Track a track within the 13th International
Semantic Web Conference, ISWC 2014, Riva del Garda, Italy, October 21, 2014.
401–404.
[8]
David A. Ferrucci, Eric W.Brown, Jennifer Chu-Carroll, James Fan, David Gondek,
Aditya Kalyanpur, Adam Lally, J. William Murdock, Eric Nyberg, John M. Prager,
Nico Schlaefer, and Christopher A. Welty. 2010. Building Watson: An Overview
of the DeepQA Project. AI Magazine 31, 3 (2010), 59–79.
[9]
Matt Gardner and Tom M. Mitchell. 2015. Ecient and Expressive Knowledge
Base Completion Using Subgraph Feature Extraction. In EMNLP. The Association
for Computational Linguistics, 1488–1498.
[10]
Xavier Glorot and Yoshua Bengio. 2010. Understanding the diculty of train-
ing deep feedforward neural networks. In AISTATS (JMLR Proceedings), Vol. 9.
JMLR.org, 249–256.
[11]
Ming Hao, Stephen H Bryant, and Yanli Wang. 2017. Predicting drug-target
interactions by dual-network integrated logistic matrix factorization. Scientic
reports 7 (2017), 40376.
[12]
Minoru Kanehisa, Miho Furumichi, Mao Tanabe, Yoko Sato, and Kanae Morishima.
2017. KEGG: new perspectives on genomes, pathways, diseases and drugs. Nucleic
Acids Research 45, D1 (2017), D353–D361.
[13]
Minoru Kanehisa, Yoko Sato, Masayuki Kawashima, Miho Furumichi, and Mao
Tanabe. 2016. KEGG as a reference resource for gene and protein annotation.
Nucleic Acids Research 44, D1 (2016), D457–D462.
[14]
Ni Lao and William W. Cohen. 2010. Relational retrieval using a combination of
path-constrained random walks. Machine Learning 81, 1 (2010), 53–67.
[15]
Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas,
Pablo Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick van Kleef, Sören
Auer, and Chris Bizer. 2014. DBpedia - A Large-scale, Multilingual Knowledge
Base Extracted from Wikipedia. Semantic Web Journal (2014).
[16]
Linxin Li, Olivia C Geraghty, Ziyah Mehta, Peter M Rothwell, and Oxford Vascular
Study. 2017. Age-specic risks, severity, time course, and outcome of bleeding
on long-term antiplatelet treatment after vascular events: a population-based
cohort study. The Lancet 390, 10093 (2017), 490–499.
[17]
Hui Liu, Jianjiang Sun, Jihong Guan, Jie Zheng, and Shuigeng Zhou. 2015. Im-
proving compound–protein interaction prediction by building up highly credible
negative samples. Bioinformatics 31, 12 (2015), i221–i229.
[18]
Jian-Ping Mei, Chee-Keong Kwoh, Peng Yang, Xiao-Li Li, and Jie Zheng. 2012.
Drug–target interaction prediction by learning from local information and neigh-
bors. Bioinformatics 29, 2 (2012), 238–245.
[19]
George A. Miller. 1995. WordNet: A Lexical Database for English. Commun. ACM
38, 11 (Nov. 1995), 39–41. https://doi.org/10.1145/219717.219748
[20]
Sameh K. Mohamed, Vít Novácek, and Pierre-Yves Vandenbussche. 2018. Knowl-
edge base completion using distinct subgraph paths. In SAC. ACM, 1992–1999.
[21]
Emir Muñoz, Vít Novácek, and Pierre-Yves Vandenbussche. 2016. Using Drug
Similarities for Discovery of Possible Adverse Reactions. In AMIA 2016, American
Medical Informatics Association Annual Symposium, Chicago, IL, USA, November
12-16, 2016. AMIA. http://knowledge.amia.org/amia- 63300-1.3360278/t004-1.
3364525/f004-1.3364526/2499657- 1.3364713/2500122-1.3364708
[22]
André CA Nascimento, Ricardo BC Prudêncio, and Ivan G Costa. 2016. A mul-
tiple kernel learning algorithm for drug-target interaction prediction. BMC
bioinformatics 17, 1 (2016), 46.
[23]
Maximilian Nickel, Kevin Murphy, Volker Tresp, and Evgeniy Gabrilovich. 2016.
A review of relational machine learning for knowledge graphs. Proc. IEEE 104, 1
(2016), 11–33.
[24]
Maximilian Nickel, Lorenzo Rosasco, and Tomaso A. Poggio. 2016. Holographic
Embeddings of Knowledge Graphs. In AAAI. AAAI Press, 1955–1961.
[25]
Rawan S Olayan, Haitham Ashoor, and Vladimir B Bajic. 2017. DDR: ecient
computational method to predict drug–target interactions using graph mining
and machine learning approaches. Bioinformatics 34, 7 (2017), 1164–1173.
[26]
Richard Qian. 2013. Understand Your World with Bing. http://blogs.bing.com/
search/2013/03/21/understand-your-world-with- bing/ Bing Blogs.
[27]
Yan Qiao, Tingting Yang, Yong Gan, Wenzhen Li, Chao Wang, Yanhong Gong,
and Zuxun Lu. 2018. Associations between aspirin use and the risk of cancers: a
meta-analysis of observational studies. BMC cancer 18, 1 (2018), 288.
[28]
Sashank Reddi, Satyen Kale, and Sanjiv Kumar. 2018. On the Convergence of
Adam and Beyond. In ICLR.
[29]
Ayeshah A Rosdah, Jessica K. Holien, Lea MD Delbridge, Gregory J Dusting, and
Shiang Y Lim. 2016. Mitochondrial ssion–a drug target for cytoprotection or
cytodestruction? Pharmacology research & perspectives 4, 3 (2016), e00235.
[30]
Peter M Rothwell, F Gerald R Fowkes, Jill FF Belch, Hisao Ogawa, Charles P
Warlow, and Tom W Meade. 2011. Eect of daily aspirin on long-term risk of
death due to cancer: analysis of individual patient data from randomised trials.
The Lancet 377, 9759 (2011), 31–41.
[31]
Peter M Rothwell, Michelle Wilson, Carl-Eric Elwin, Bo Norrving, Ale Algra,
Charles P Warlow, and Tom W Meade. 2010. Long-term eect of aspirin on
colorectal cancer incidence and mortality: 20-year follow-up of ve randomised
trials. The Lancet 376, 9754 (2010), 1741–1750.
[32]
Amit Singhal. 2012. Introducing the Knowledge Graph: things, not strings. "https:
//googleblog.blogspot.ie/2012/05/introducing-knowledge- graph-things-not.
html" Google Ocial Blog.
[33]
Lekha Sleno and Andrew Emili. 2008. Proteomic methods for drug target discov-
ery. Current opinion in chemical biology 12, 1 (2008), 46–54.
[34] Walter Sneader. 2005. Drug discovery: a history. John Wiley & Sons.
[35]
Georg C Terstappen, Christina Schlüpen, Roberto Raggiaschi, and Giovanni
Gaviraghi. 2007. Target deconvolution strategies in drug discovery. Nature
Reviews Drug Discovery 6, 11 (2007), 891.
[36]
James Hendler Tim Berners-Lee and Ora Lassila. 2001. The Semantic Web, A new
form of Web content that is meaningful to computers will unleash a revolution
of new possibilities. Scientic American: https://www.scienticamerican.com/
article/the-semantic- web/. Retrieved: 2017-04-21.
[37]
Théo Trouillon and Maximilian Nickel. 2017. Complex and Holographic Embed-
dings of Knowledge Graphs: A Comparison. CoRR abs/1707.01475 (2017).
[38]
Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume
Bouchard. 2016. Complex Embeddings for Simple Link Prediction. In ICML (JMLR
Workshop and Conference Proceedings), Vol. 48. JMLR.org, 2071–2080.
[39]
Twan van Laarhoven, Sander B Nabuurs, and Elena Marchiori. 2011. Gaussian
interaction prole kernels for predicting drug–target interaction. Bioinformatics
27, 21 (2011), 3036–3043.
[40]
David S Wishart, Craig Knox, An Chi Guo, Savita Shrivastava, Murtaza Has-
sanali, Paul Stothard, Zhan Chang, and Jennifer Woolsey. 2006. DrugBank: a
comprehensive resource for in silico drug discovery and exploration. Nucleic
acids research 34, suppl_1 (2006), D668–D672.
[41]
Lei Xie, Li Xie, Sarah L Kinnings, and Philip E Bourne. 2012. Novel computational
approaches to polypharmacology as a means to dene responses to individual
drugs. Annual review of pharmacology and toxicology 52 (2012), 361–379.
[42]
Yoshihiro Yamanishi, Michihiro Araki, Alex Gutteridge, Wataru Honda, and
Minoru Kanehisa. 2008. Prediction of drug–target interaction networks from
the integration of chemical and genomic spaces. Bioinformatics 24, 13 (2008),
i232–i240.
[43]
Yoshihiro Yamanishi, Masaaki Kotera, Minoru Kanehisa, and Susumu Goto. 2010.
Drug-target interaction prediction from chemical, genomic and pharmacological
data in an integrated framework. Bioinformatics 26, 12 (2010), i246–i254.
[44]
Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Em-
bedding Entities and Relations for Learning and Inference in Knowledge Bases.
In ICLR.