Article

KDC YJMBI-65714; No. of pages: 12; 4C:

Annotating Diseases Using Human

Phenotype Ontology Improves
Prediction of Disease-Associated
Long Non-coding RNAs

Duc-Hau Le 1, 2 and Lan T.M. Dao 2

1 - School of Computer Science and Engineering, Thuyloi University, 175 Tay Son, Dong Da, Hanoi, Vietnam
2 - Vinmec Research Institute of Stem Cell and Gene Technology, 458 Minh Khai, Hai Ba Trung, Hanoi, Vietnam

Correspondence to Duc-Hau Le: School of Computer Science and Engineering, Thuyloi University, 175 Tay Son, Dong Da,
Hanoi, Vietnam. hauldhut@gmail.com
https://doi.org/10.1016/j.jmb.2018.05.006
Edited by Nir Yosef

Abstract
Recently, many long non-coding RNAs (lncRNAs) have been identified and their biological function has been
characterized; however, our understanding of their underlying molecular mechanisms related to disease is still
limited. To overcome the limitation in experimentally identifying disease–lncRNA associations, computational
methods have been proposed as a powerful tool to predict such associations. These methods are usually based
on the similarities between diseases or lncRNAs since it was reported that similar diseases are associated with
functionally similar lncRNAs. Therefore, prediction performance is highly dependent on how well the similarities
can be captured. Previous studies have calculated the similarity between two diseases by mapping exactly each
disease to a single Disease Ontology (DO) term, and then use a semantic similarity measure to calculate the
similarity between them. However, the problem of this approach is that a disease can be described by more than
one DO terms. Until now, there is no annotation database of DO terms for diseases except for genes. In contrast,
Human Phenotype Ontology (HPO) is designed to fully annotate human disease phenotypes. Therefore, in this
study, we constructed disease similarity networks/matrices using HPO instead of DO. Then, we used these
networks/matrices as inputs of two representative machine learning-based and network-based ranking
algorithms, that is, regularized least square and heterogeneous graph-based inference, respectively. The results
showed that the prediction performance of the two algorithms on HPO-based is better than that on DO-based
networks/matrices. In addition, our method can predict 11 novel cancer-associated lncRNAs, which are supported
by literature evidence.
© 2018 Elsevier Ltd. All rights reserved.

Introduction number of identified lncRNAs, many related databases

Recent genome-wide studies have revealed that
two-thirds of the genome is being transcribed but only a
minority of the transcriptional output encodes for
proteins [1–4]. One class of non-coding RNAs is
termed long non-coding RNAs (lncRNAs), described
as non-coding RNA transcripts longer than 200
nucleotides. LncRNAs are characterized by being
transcribed by RNA polymerase II but at low levels,
and to exhibit alternative splicing as well as being multi-
exonic, polyadenylated, and generally exhibit low
coding potential [4,5]. Studies of lncRNA at a
genome-wide scale have revealed that a large
proportion of long non-coding transcripts can be
functionally important for normal and pathological
developmental processes [6]. With the increasing
have been established including GENCODE [7],
lncRNAdb [8], lncRBase [9], lnRNA2Function [10]
and LncRNA2Target [11]. More specifically, focusing
on their roles in the common diseases and cancer,
there are the lncRNADisease [12] and Lnc2Cancer
[13] databases, respectively. Even if a certain number
of disease–lncRNA associations are experimentally
confirmed, the obvious fact is that the vast majority is
still unknown. Indeed, approximately 1000 lncRNA–
disease entries have been reported in lncRNA2Di-
sease [12]. Therefore, a number of computational
methods have been recently proposed to predict novel
disease–lncRNA associations [14].
Computational methods to predict novel disease–
lncRNA association fall into two main categories:
machine learning-based and network-based methods.

0022-2836/© 2018 Elsevier Ltd. All rights reserved. J Mol Biol (2018) xx, xxx–xxx

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
2 Using HPO for Better Prediction of Disease-Associated lncRNA
The machine learning-based methods usually use
known disease–lncRNA association as well as un-
known ones to train learning models, then the learned
models are used to predict novel associations. This
approach also integrates various biological information
to annotate lncRNAs. For example, a naïve Bayesian
model was used to integrate genome, regulome and
transcriptome features to identify novel cancer-related
lncRNAs [15]. However, this method required negative
training samples (i.e., lncRNAs which are not associ-
ated with diseases) to train the model. Given that there
is no experimental confirmation of such negative
associations, unknown lncRNA–disease associations
were used in that study. Recently, a semi-supervised
model, for example, regularized least square (RLS)
[16], was used to overcome this limitation. The model
does not require negative training samples, since it
trains the model based on the known and the unknown
disease–lncRNA associations. In contrast to machine
learning-based studies, where only a few have been
used, many network-based methods have been
proposed to predict novel disease-associated
lncRNAs. The network-based methods usually rank
candidate lncRNAs based on their relevance to a
disease of interest. The most commonly used algo-
rithms are label propagation algorithms such as
random walk with restart (RWR) [17–21] and KATZ
[22]. The main difference between these studies is the
construction of underlying networks on which the
propagation algorithms were applied. For example,
Sun et al. [17] applied RWR on a constructed lncRNA
functional similarity network (abbreviated as lncRNA
FSN). Liu et al. [18] constructed a protein-coding gene–
lncRNA bipartite network based on lncRNAs and
protein-coding gene expression profiles, and then
used RWR to predict cancer-related lncRNAs. Mean-
while, Zhou et al. [19] and Ganegoda et al. [20] built
heterogeneous networks of lncRNAs and diseases by
combining a lncRNA similarity network, a disease
similarity network and known disease–lncRNA associ-
ations, and then applied the RWR algorithm on the
heterogeneous networks to predict novel disease-
associated lncRNAs. These network-based methods
were proposed based on an observation that function-
ally similar lncRNAs are usually associated with the
same or similar diseases, known as “disease module”
principle. By additionally integrating disease similarity
networks, the “disease module” principle was used
more effectively for methods based on the heteroge-
neous networks compared to those solely based on the
lncRNA similarity network [19,20,22].
Although network- and machine learning-based
methods are different in terms of data representation
(i.e., machine learning-based methods use vector-
based representation of lncRNAs and interactions, and
similarity matrices of lncRNAs and diseases; mean-
while, network-based method are based on similarity
networks of lncRNAs and diseases), they are both
based on the similarity networks/matrices of lncRNAs
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
and diseases. Therefore, the construction of exact
similarity networks/matrices is the most important step
to correctly identify novel disease-associated lncRNAs.
Indeed, many efforts have been made to build effective
lncRNA FSNs such as studies based on disease
association [23–25], shared microRNAs [19,26], or a
study that combine both lncRNA-related transcriptional
and post-transcriptional information [27]. However, few
efforts were made for constructing disease similarity
networks based on sharing of associated lncRNAs [28]
and Disease Ontology (DO) [19]. DO [29] terms
describe phenotype characteristics, related medical
vocabulary and disease concepts; thus, a certain
disease can be annotated with several DO terms. In
other words, a disease cannot be fully described by a
single DO term. For example, breast cancer (OMIM ID:
114480) was characterized by only one DO term,
mammary cancer (DO ID: 1612). Nevertheless, previ-
ous studies directly mapped each disease to only one
DO term [17,19]. More specifically, to calculate a
similarity between two diseases, each disease was
directly mapped with only one DO term, then a
semantic similarity measure was used to calculate the
similarity between two corresponding DO terms
[17,19]. In addition, until now, there is no database
that stores annotation of DO terms for diseases except
for genes [30]. Therefore, to measure the similarity
between two DO terms, a directed acyclic graph (DAG)
for each DO term was used instead of annotation
information. In contrast, the Human Phenotype Ontol-
ogy (HPO) [31] (i.e., a controlled vocabulary database)
is designed to fully annotate human disease pheno-
types. For example, breast cancer is annotated with
three HPO terms, that is, autosomal dominant inheri-
tance (HP:0000006), heterogeneous (HP:0001425)
and breast carcinoma (HP:0003002).
In this study, diseases were first mapped to OMIM
records [32,33], which were annotated by HPO [31].
Then, we calculated the similarities between any pair
of diseases to construct disease similarity networks/
matrices using the HPO annotation database (i.e.,
HPO-based method). In addition, we constructed other
disease similarity networks/matrices based on DO, as
in other previous methods [17,19] (i.e., DO-based
method). For fair comparison, we collected a set of
diseases that have at least one known lncRNA and
were exactly mapped to one OMIM record as well as
one DO term. Comparing our HPO-based method with
the DO-based one on the same ranking algorithms, we
found that our method achieved better prediction
performances, in terms of area under the curve
(AUC) values, for two representative machine
learning-based and network-based algorithms, that
is, RLS and heterogeneous graph-based inference
(HGBI), respectively. Furthermore, we tested our
method in identifying cancer-associated lncRNAs
and found that 11 out of highly ranked lncRNAs were
annotated in Lnc2cancer as associated to 10 cancer
diseases [13].
Using HPO for Better Prediction of Disease-Associated lncRNA 3

Results and Discussion a

Overall comparison of prediction performance

To compare the prediction performance of our

method (i.e., HPO-based) with the DO-based method
on two representative ranking algorithms (i.e., RLS-
based and HGBI-based algorithms), we used cross-
validation methods for each disease in a set of 42
diseases (see Materials and Methods). For the RLS-
based algorithm, we set the best RLS parameter
settings (ηL = ηD = 1 and w = 0.9), as recommended in
[34,35]. Meanwhile, for the HGBI-based algorithm, we
set decay factor α = 0.4 as in the original study [36].
First, we compared the two methods using the
unweighted lncRNA FSN collected from Zhou et al.
[19] on all of 42 diseases using the leave-one-out cross-
validation (LOOCV) method. The result shows that the
prediction performance of the HPO-based method is
better than the DO-based in predicting disease- b
associated lncRNAs in terms of AUC value (Fig. 1).
More specifically, the HPO-based method achieved
AUC = 0.946, better than the DO-based method
(AUC = 0.870), with the RLS-based ranking algorithm
(Fig. 1(a)). Similarly, with the HGBI-based ranking
algorithm, the HPO-based method (AUC = 0.837) is
superior to the DO-based one (AUC = 0.592) (Fig. 1(b)).
Second, we compared the two methods on 10 diseases
having no less than three known associated lncRNAs
using a 3-fold cross-validation. The result shows that
the HPO-based method (AUC = 0.975) outperformed
the DO-based method (AUC = 0.904) with the RLS-
based ranking algorithm (Fig. S1(a)). Similarly, with
the HGBI-based algorithm, the HPO-based method
(AUC = 0.906) was also superior to the DO-based
method (AUC = 0.747) (Fig. S1(b)). In addition to the
unweighted lncRNA FSN, we also compared the two
methods on the weighted lncRNA FSN using LOOCV
method. With the RLS-based ranking algorithm, the
HPO-based method achieved an AUC = 0.984, which
was superior to the DO-based method (AUC = 0.919) Fig. 1. Performance comparison between HPO-based
and DO-based methods. The prediction performance of the
(Fig. S1(c)). Meanwhile, with the HGBI-based algo-
two methods was compared on the unweighted lncRNA FSN
rithm, the HPO-based and the DO-based methods by LOOCV method using RLS-based (a) and HGBI-based
achieved AUC = 0.412 and AUC = 0.382, respectively (b) ranking algorithms.
(Fig. S1(d)). In this case, both of the methods
performed badly since AUC values are less than 0.5.
Taken together, these results indicate that the HPO- phenotypes, see Materials and Methods) results
based method performed better than the DO-based in different sizes of the disease similarity networks
method irrespective of the used ranking algorithms. (i.e., number of diseases and number of associations
This could be due to the HPO-based method better between them), thus we further investigated whether
reflected the similarity between two diseases compared the size of the networks affects the prediction
to the DO-based method. performance. It was reported that ranking of nodes in
a network generated by ranking algorithms is highly
Performance comparison with respect to degree dependent on its degree of connectivity (i.e., number of
of connectivity its neighbors in the network) [37]. Therefore, for a fair
performance comparison, we constructed different
The difference in number of DO terms and annotated networks based on the original DO-based and HPO-
OMIM records (i.e., 2152 DO terms and 6521 disease based disease similarity networks (see Materials and

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
4 Using HPO for Better Prediction of Disease-Associated lncRNA

Methods) with a fixed degree of connectivity. More
specifically, for each disease in the original disease
similarity networks, we selected only a fixed number of
neighboring diseases (e.g., 10) having the largest
similarities with it. These networks were also repre-
sented by adjacency matrices as inputs for the ranking
algorithms. Then, we applied the RLS-based ranking
algorithm on these networks to predict disease-
associated lncRNAs. First, we combined these disease
networks with the unweighted lncRNA FSN and
assessed the prediction performance of each network
using LOOCV method on the set of 42 diseases. As
shown in Fig. 2, regardless of the degree of connec-
tivity, the HPO-based method always performed better
than the DO-based one. We also achieved similar
results with the 3-fold cross-validation method on the
set of 10 valid diseases (Fig. S2(a)). In addition, we
combined these disease similarity networks with the
weighted lncRNA FSN and compared the prediction
performance of the two methods on these networks
using the LOOCV method on the set of 42 diseases.
Figure S2(b) also shows that the HPO-based method
achieved higher performance compared to that of the
DO-based method. Once again, these results indicate
that annotating disease with HPO terms better reflects
the similarity between two diseases than directly
mapping them to DO terms, and thus, the HPO-
based method was superior to the DO-based method
irrespective of degree of connectivity of constructed
disease similarity networks.
Prediction of novel cancer-associated lncRNAs
In this experiment, we tried to predict novel associ-
ations between lncRNA and cancer. Cancer is a group
of diseases involving out-of-control cell growth with the
potential to invade or spread to other parts of the body.
Many intensive studies have been carried out but little is
known about the underlying molecular mechanisms.
Recently, a cancer-associated lncRNA database
Lnc2Cancer was introduced [13]. This is a manually
curated database that provides comprehensive exper-
imentally supported associations between lncRNA and
human cancers. The current version of Lnc2Cancer
documents 1488 entries of associations between 666
human lncRNAs and 97 human cancers through the
revision of more than 2000 published papers. There are
10 cancers in common between Lnc2Cancer and
lncRNADisease [12]. For each of the 10 cancers, we
used known disease–lncRNA associations in lncRNA-
Disease as input of the RLS-based ranking algorithm to
rank candidate lncRNAs (remaining ones in the
unweighted lncRNA FSN, which are not known to be
associated with the cancer of interest) and then
selected the top 100 ranked candidates. Table 1 lists
11 lncRNAs in the top which are reported in

Fig. 2. Performance comparison between HPO-based and DO-based methods based on degree of connectivity. The
unweighted lncRNA FSN and LOOCV method were used. By setting the same degree of connectivity for each disease in
HPO-based and DO-based networks, the prediction performance using RLS-based ranking method on HPO-based
networks is higher than that on DO-based networks. Note: 375 is the ratio between the number of associations and the
number of diseases in the original DO-based disease similarity network. Using this ratio, we constructed a network from
the original HPO-based disease similarity matrix.

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
Using HPO for Better Prediction of Disease-Associated lncRNA 5

Table 1. List of evidenced disease–lncRNA associations in top 100 ranked candidates for 10 cancers
OMIM ID Disease lncRNA PubMed ID
MIM109800 Bladder cancer PVT1 26517688
MIM114500 Colorectal cancer H19 11120891, 19926638, 22427002, 26068968, 26989025
PVT1 24196785, 26990997
MIM114550 Hepatocellular carcinoma GAS5 26109807, 25120813, 26163879, 26404135, 26404135
H19 15736456, 17786216, 24063685, 24761865, 24939300, 23222811
HOTAIR 24663081, 23292722, 21327457, 26024833, 22289527, 27301338
MIM151400 Chronic lymphocytic leukemia DLEU2 19347735, 9395242, 11161783, 19591824
NEAT1 25971364
MIM211980 Lung cancer H19 16707459, 24063685
ZNRD1-AS1 27166266
MIM275355 Oral squamous cell carcinoma HOTAIR 23292713, 25901533
MALAT1 26522444
MEG3 23292713
MIM603956 Cervical cancer PVT1 27272214, 27232880
MIM607107 Nasopharyngeal carcinoma H19 27040767
NEAT1 27020592
MIM607174 Meningioma H19 10738131
MEG3 20179190
MIM613659 Gastric cancer CCAT1 25561974, 25674211, 23143645, 25755774
HOXA-AS2 26384350
PVT1 25258543, 26096073, 25956062, 26925791

Lnc2Cancer for the 10 cancers. Among these
lncRNAs, six of them (i.e., GAS5, H19, HOTAIR,
MALAT1, MEG3 and PVT1) are known to be associ-
ated with at least one disease in lncRNADisease
database (excluding the disease of interest), mean-
while five remaining ones (i.e., DLEU2, NEAT1,
ZNRD1-AS1, CCAT1 and HOXA-AS2) are not known
to be associated with any disease in lncRNADisease
database. More importantly, newly predicted associa-
tions between these 11 lncRNAs and the 10 cancers
were supported by literature evidence. For instance,
PVT1 was upregulated in bladder cancer tissues and
further experiments revealed that PVT1 promoted cell
proliferation and suppressed cell apoptosis [38]. H19
was found highly expressed in mesenchymal-like
cancer cells and primary colorectal cancer tissues
[39]. HOTAIR expression was detected in primary
hepatocellular carcinoma in 13 out of 64 patients [40].
The lncRNAs NEAT1 and lincRNA-p21 were detected
as novel elements of the p53-dependent DNA damage
response machinery in chronic lymphocytic leukemia
and lymphoma [41]. Higher expression of ZNRD1-AS1
was shown in lung cancer tissues [42]. MALAT1 was
overexpressed in oral squamous cell carcinoma
tissues compared to normal oral mucosa by real-time
PCR [43]. PVT1 was upregulated in cervical cancer
tissues [44]. NEAT1 lncRNA was significantly upregu-
lated in nasopharyngeal carcinoma cell lines and
tissues [45]. MEG3 was not expressed in the majority
of human meningiomas or the human meningioma cell
lines IOMM-Lee and CH157-MN [46]. Expression
levels of the CCAT2 lncRNA in gastric cancer tissues
were significantly higher than those in adjacent non-
tumor tissues [47]. Taken together, these results
indicate that our HPO-based method can not only
achieve high prediction performance in terms of AUC
values on known disease–lncRNA associations, but
also predict novel disease-associated lncRNAs.
Conclusions
Recent studies have shown the important roles of
lncRNAs in the development of a number of diseases.
In addition, they can be potential targets for drug
discovery. However, many potential disease–lncRNA
associations have not yet been revealed experimen-
tally. Therefore, computational methods have been
proposed as an alternative approach to reduce the
cost and time of such laborious tasks. A large number
of computational methods have been developed for
predicting associations between other non-coding
RNAs (e.g., miRNAs) and diseases [48–54]. Compu-
tational methods often relied on the similarity between
diseases and lncRNAs using network- and machine
learning-based techniques. Therefore, careful con-
struction of similarity matrices/networks among dis-
eases and lncRNAs is an important step to identify
novel disease-associated lncRNAs. For constructing
disease similarity networks, a previous study built a
disease similarity matrix based on the interaction
profile of known disease–lncRNA associations [28].
Obviously, this method limits the size of the disease
similarity networks/matrices to the number of diseases
which are known to be associated with at least one
lncRNAs. Therefore, other studies constructed the
disease similarity networks/matrices by exactly map-
ping each disease with one DO term, then the similarity
of two diseases was calculated based on the similarity
between two corresponding DO terms using a

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
6 Using HPO for Better Prediction of Disease-Associated lncRNA
semantic similarity measure. Although DO was recent-
ly designed to describe phenotype characteristics and
related medical vocabulary disease concepts, only the
DGA database [30] has annotations that relate DO
terms to genes but not to diseases. More importantly,
only a small set of DO terms have been used for such
annotations (i.e., approximately 2161 DO terms out of
8839 were used). This also limits the number of
diseases possible to be investigated. In contrast, HPO
is designed to be supported for a decade [55], with
continued updating [56] of disease phenotype anno-
tation. In addition, most disease phenotypes are
annotated by HPO terms [i.e., 6521 disease pheno-
types out of 6875 (including 3739 phenotypes with
known molecular basis, 1597 with unknown molecular
basis, and mendelian phenotypes or loci with unknown
molecular basis) from OMIM were annotated]. There-
fore, using HPO to calculate disease similarity could
yield more accurate results than using DO. Indeed,
experimental results show that HPO-based networks/
matrices are better than the DO-based ones in the
prediction of disease-associated lncRNAs. Our pro-
posed method can also predict 11 novel lncRNAs
associated with 10 cancers which were not yet
reported in lncRNADisease [12]. In summary, using
HPO for the calculation of disease similarity could be a
promising approach for the prediction of disease-
associated lncRNAs.
Materials and Methods
Known disease–lncRNA associations
We used lncRNADisease database [12] for known
disease–lncRNA associations, which contain experi-
mentally validated disease–lncRNA associations.
Currently, there are 1028 associations between 321
lncRNAs and 221 diseases collected from ~ 500
publications.
DO and annotation databases
DO was obtained from http://disease-ontology.org/
[29]. This database describes phenotype characteris-
tics and related medical vocabulary disease concepts.
The DO semantically integrates disease and medical
vocabularies through extensive cross-mapping of DO
terms to MeSH [57,58], ICD [59], NCI's thesaurus [60],
SNOMED [61,62] and OMIM [32,33]. DO currently
contains approximately 10,878 terms. The DO is
presented as a DAG with the root term as “disease”
(DO ID: 4). Unfortunately, there is no database
containing DO term annotations to any disease.
However, there are 2161 DO terms used to annotate
genes in DGA [30]. Therefore, in this study, this
annotation information was used to calculate the
similarity between DO terms.
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
HPO and annotation databases
HPO and annotation databases were obtained from
http://www.human-phenotype-ontology.org [31]. This
database provides a standardized vocabulary of
phenotypic abnormalities encountered in human
disease. Each term in the HPO describes a pheno-
typic abnormality. The HPO is currently being
developed using the medical literature, Orphanet
[63], DECIPHER [64] and OMIM [32,33]. HPO
currently contains approximately 11,000 terms and
over 115,000 annotations to hereditary diseases for
6521 phenotypes (including approximately 4000
disease phenotypes). The HPO is presented as a
DAG with the root term as “All” (HP:0000001). Using
this graph, ancestors and descendants of a term can
be specified.
Construction of disease similarity
networks/matrices
To construct a disease similarity matrix, previous
studies [17,19] mapped directly each disease to
one DO term [29]. Therefore, similarity between
two diseases was calculated based on similarity
between two mapped DO terms. In this study, we
first mapped each disease to one OMIM record
and then annotated the OMIM with HPO terms.
After that, we calculated the similarity between two
diseases based on the annotated HPO terms. In
the next section, we will introduce the two
methods including the one used by previous
studies [17,19] (abbreviated as “DO-based
method”) and our method (abbreviated as “HPO-
based method”).
DO-based method
To construct DO-based disease similarity matrix,
we mapped each disease to one DO term as in
previous studies [17,19], and then calculated the
similarity between any pair of mapped DO terms in
the set of 2161 DO terms and annotations in DGA
database [30]. The similarity between two ontology
terms was calculated based on information content
(IC) of each term, which is defined as follows:
ICðt Þ ¼ − logðp ðt ÞÞ
where p(t) is the probability of term t occurring in a
corpus (i.e., an annotation database, e.g., DGA for
f ðt Þ
DO). More specifically, that is, pðt Þ ¼ f ðrootÞ such that
f ðt Þ ¼ Annotðt Þ þ ∑c∈Childrenðt Þ f ðcÞ . In this formula,
Annot(t) means the number of phenotypes annotated
with t in the corpus and Children(t) represents the set
of children terms of t in the DO graph. “root” is root term
of the DO graph. Then, the semantic similarity
between the two DO terms, ti and tj, based on the
Using HPO for Better Prediction of Disease-Associated lncRNA
most informative common ancestor approach Resnik
[65], is calculated as follows:


simTerm t i ; t j ¼ max ðICðc ÞÞ
c ∈ P ðt i ;t j Þ
where P(ti, tj) is the set of shared ancestors of ti and tj.
For a pair of diseases, di and dj are directly mapped
to ti and tj, respectively. The similarity between them is
defined as follows:



w ij ¼ simDis d i ; d j ¼ simTerm t i ; t j
We calculated the similarity for every pair of DO
terms in the total of 2161 DO terms to construct a DO-
based disease similarity matrix. By selecting pairs
having simDis(di, dj) N 0, we also constructed a DO-
based disease similarity network containing 806,505
interactions (also known as an original DO-based
similarity network). Both the disease similarity
network and the matrix can be represented as an
adjacency matrix WD, where its element (WD)i, j was
set to wij representing the similarity between disease
di and dj.
HPO-based method
To construct HPO-based disease similarity net-
work, we first mapped each disease to one OMIM
record, then annotated the OMIM with HPO terms
using the HPO annotation database [31]. Then, the
similarity of every pair of term was calculated using
the same formulas as for a pair of DO terms. More
specifically, the similarity between a pair of disease
di and dj is calculated as the maximum of simTerm
values between all possible pairs of terms as follows:




simDis d i ; d j ¼ max simTerm t i ; t j
t i ∈T ðd i Þ;t j ∈T ðd j Þ
where T(di) and T(dj) represent the set of terms
annotating to di and dj, respectively.
This value is normalized in range [0, 1] to account
for an unequal number of HPO terms for both
disease phenotypes as follows:


2  simDis d i ; d j
w ij ¼

simDisðd i ; d i Þ þ simDis d j ; d j
We calculated the similarity for every pair of diseases
in the total of 6521 disease phenotypes to construct an
HPO-based disease similarity matrix. By selecting
pairs having simDis(di, dj) N 0, we also constructed an
HPO-based disease similarity network containing
21,258,460 interactions (also called as an original
HPO-based disease similarity network). Similarly, both
HPO-based similarity matrix and network can be
represented as an adjacency matrix WD, where its
element (WD)i, j was set to wij representing the similarity
between disease di and dj.
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
7
Construction of lncRNA functional similarity
network/matrix
In this study, we first collected an unweighted
lncRNA FSN data from Zhou et al. [19]. Each
interaction in this network was constructed by signif-
icant co-occurrence of shared miRNA response
elements on lncRNA transcripts. This network includes
13,640 interactions between 697 lncRNAs. This
network can be represented as an adjacency matrix
WL, where its element (WL)i, j was set to 1 or 0 with
respect to whether an interaction between lncRNA i
and j exists or not. Second, we collected original data of
lncRNA–miRNA associations and then repeated the
same procedure introduced in that study to calculate
degree of significance of co-occurrence of shared
miRNAs between lncRNAs. After setting the threshold
of significance to 0.0001 as in that study, only
significant lncRNA interactions were kept. However,
instead of setting the same value (i.e., 1) to those
interactions, we transformed significant values using a
logarithm of base 10 and normalized them into a range
(0, 1) to form a weighted lncRNA FSN (see Supple-
mentary Table S2). Similarly, this network can be
represented as an adjacency matrix WL, where its
element (WL)i, j was set to a corresponding weight or 0
with respect to whether an interaction between lncRNA
i and j exists or not.
Construction of adjacency network/matrix of
known disease–lncRNA associations
As aforementioned, known disease–lncRNA asso-
ciations were collected from lncRNADisease data-
base [12]. These diseases and lncRNAs were
mapped onto disease and lncRNA similarity net-
 works/matrices. After mapping, a total of 42 diseases
and 151 associated lncRNAs remained for assessing
the prediction performance, in which each disease
was mapped exactly to one OMIM record and one DO
term (see Supplementary Table S1). These associa-
tions can be represented as a bipartite network with an
adjacency matrix WDL, where (WDL)i, j = 1 if disease i is
known to be associated with lncRNA j, otherwise
(WDL)i, j = 0.
Ranking algorithms
In this study, we proposed a novel method (i.e., HPO-
based method) to construct disease similarity network/
matrix using HPO for prediction of disease-associated
lncRNAs. To compare prediction performance of our
method (i.e., HPO-based method) with that of previous
one (i.e., DO-based method), we tested it with different
ranking algorithms and showed that the HPO-based
method is superior to the DO-based one, irrespective of
the used ranking algorithms. Many ranking algorithms
including machine learning-based and network-based
algorithms working on the heterogeneous networks
8 Using HPO for Better Prediction of Disease-Associated lncRNA

have been proposed. Here, we used two representa-
tive algorithms for predicting disease-associated
lncRNAs. Figure 3 shows an Illustration of HPO-
based and DO-based methods and overall framework
of prediction of disease-associated lncRNAs using
ranking algorithms.
RLS-based ranking algorithm
RLS is a semi-supervised and global learning
method since it can rank disease–lncRNA associa-
tions for all the diseases simultaneously, without using
negative samples. This method was designed to
construct a continuous classification function which
can determine the association probability between
each lncRNA and a given disease (i.e., the higher this
probability is, the more each lncRNA is related to a
given disease). To this end, a RLS classifier was
constructed by defining and minimizing a cost function.
This cost function was trained in the lncRNA FSN and
the disease similarity network, and then it was used to
optimize the classification function [34]. Formally, the

Fig. 3. Illustration of HPO-based and DO-based methods and overall framework for the prediction of disease-
associated lncRNAs. (a) Similarity between two diseases was calculated using ontology data. In HPO-based method
(upper panel), each disease was mapped to one OMIM record, and then they were annotated with HPO terms. Finally, the
similarity between two diseases was calculated based on the similarity of every pair of the annotating HPO terms
annotating to each disease. Meanwhile, in DO-based method (lower panel), each disease was directly mapped to one DO
term, and then the similarity between two diseases was calculated as the similarity of the two annotating DO terms.
(b) Overall framework of prediction of disease-associated lncRNAs. A disease of interest and its known associated
lncRNAs (if any) as well as candidate lncRNAs were mapped onto the similarity networks/matrices. Then, a ranking
algorithm was applied to rank all the candidate lncRNAs.

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
Using HPO for Better Prediction of Disease-Associated lncRNA
optimal classifier in these two spaces was defined as
follows:
F  ¼ w F T 
L þ ð1−w ÞF D
where
- FL∗ and FD∗ are optimal classification functions in
the lncRNA and disease phenotype spaces,
respectively, as follows:
F L ¼ W L ðW L þ ηL I L ÞW DL
F D ¼ W D ðW D þ ηD I D ÞW DL
- w is the weight between these two spaces. ηM
and ηD are trade-off parameters in the lncRNA
and disease phenotype spaces, respectively.
- IL and ID are identity matrices with the same
size as matrices WL and WD, respectively.
HGBI-based ranking algorithm
In addition to the machine learning-based algorithm
in previous section (i.e., RLS-based ranking algorithm),
in this section, we introduce an HGBI algorithm. This
algorithm was first proposed to infer novel drug–target
interactions [36], and then it was applied to predict
disease-associated miRNAs [66]. HGBI is based on the
guilt-by-association principle on a heterogeneous
network of diseases and lncRNAs and predicts new
disease–lncRNA associations by iteratively updating
the measure of strength between unlinked disease–
lncRNA pairs by taking all the paths in the network into
account.
The potential association probability between a
disease in the disease similarity network and a lncRNA
in the lncRNA network can be defined as follows:
t þ1
F ¼ ∝W 0D F t W 0L
where α is a decay factor.
According to the original study [36], association
probability matrix F will converge when WD′ and WL′
are normalized as follows:

 ðW D Þi; j
W 0D i; j ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm
k ¼1 ð W Þ
D i;k

 ðW L Þi; j
W 0L i; j
¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn
k ¼1 ðW L Þi;k
where m and n are number of diseases and lncRNAs in
the disease similarity networks and the lncRNA FSN,
respectively.
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
9
All disease–lncRNA pairs in the heterogeneous
networks are eventually ranked according to the
steady-state probability vector F ∞, which is obtained
by repeating the iterations until convergence is
reached (in this study, the number of iterations is
set to 10).
Performance assessment
To assess the prediction performance of ranking
algorithms (i.e., RLS-based and HGBI-based
T
algorithms) on different DO- and HPO-based disease
similarity networks, we used the LOOCV method for
each disease in the set of 42 diseases. More
specifically, for each disease (d) with known associated
lncRNAs (S), in each round of LOOCV, we held out one
known d-associated lncRNA. The held-out lncRNA (s)
and remaining lncRNAs (C) in the lncRNA network/
matrix, which were not known to be associated to d,
were then ranked by the method. After that, we plotted
the receiver operating characteristic (ROC) curve and
calculated the AUC to compare the performance of
the methods. This curve represents the relationship
between sensitivity and (1 − specificity), where
sensitivity refers to the percentage of known d-
associated lncRNAs that were ranked above a
particular threshold and specificity refers to the
percentage of lncRNAs which were not known to be
associated top ranked below this threshold. More
specifically, given a threshold τ, we counted TP (true
positives), FN (false negatives), FP (false positives)
and TN (true negatives), which were formally defined
as follows:
X X
P¼ I ð rankðs Þ ≤ τ Þ FN ¼ I ð rankðsÞ Nτ Þ
s∈S s∈S
X X
FP ¼ I ð rankðc Þ ≤ τ Þ TN ¼ I ð rankðc ÞN τ Þ
þ ð1−∝ÞW DL c ∈C c ∈C
where rank(s), rank(c) and I(∙) denote the rank of s, the
rank of a lncRNA c out of the set C and the indicator
function, respectively. Then, we defined sensitivity and
(1 − specificity) as follows:
TP FP
sensitivity ¼ 1−specificity ¼
TP þ FN FP þ TN
Pm ffi
k ¼1 ð W Þ
D k; j By varying τ from one to the number of lncRNAs in
the set C ∪ {s}, the relationship between sensitivity and
(1 − specificity) was plotted. The ROC curve is the
Pn
curve constructed based on those pairs of values, and
the AUC is the area under the ROC curve.
k ¼1 ðW L Þk ; j
In addition to the LOOCV, we assessed the
prediction performance of the methods using 3-fold
cross validation method. This is because there were
10 out of the set of 42 diseases having no less than 3
known associated lncRNAs which are available on
10 Using HPO for Better Prediction of Disease-Associated lncRNA
the lncRNA FSN. Figure S3 shows the distribution of
known disease-associated lncRNAs.
Acknowledgment
This research is funded by Vietnam National
Foundation for Science and Technology Development
(NAFOSTED) under grant number 102.01-2017.14.
Appendix A. Supplementary data
Supplementary data to this article can be found
online at https://doi.org/10.1016/j.jmb.2018.05.006.
Received 2 November 2017;
Received in revised form 28 April 2018;
Accepted 5 May 2018
Available online xxxx
Keywords:
disease-associated lncRNA;
Human Phenotype Ontology;
Disease Ontology;
semantic similarity;
ranking algorithms
Abbreviations used:
lncRNAs, long non-coding RNAs; DAG, directed acyclic
graph; HPO, Human Phenotype Ontology; RLS, regularized
least square; HGBI, heterogeneous graph-based inference;
LOOCV, leave-one-out cross-validation; ROC, receiver
operating characteristic; AUC, area under the curve.
References
[1] P. Bertone, V. Stolc, T.E. Royce, J.S. Rozowsky, A.E. Urban, X.
Zhu, J.L. Rinn, W. Tongprasit, M. Samanta, S. Weissman, et al.,
Global identification of human transcribed sequences with
genome tiling arrays, Science 306 (5705) (2004) 2242–2246.
[2] P. Carninci, T. Kasukawa, S. Katayama, J. Gough, M.C.
Frith, N. Maeda, R. Oyama, T. Ravasi, B. Lenhard, C. Wells,
et al., The transcriptional landscape of the mammalian
genome, Science 309 (5740) (2005) 1559–1563.
[3] E. Birney, J.A. Stamatoyannopoulos, A. Dutta, R. Guigó, T.R.
Gingeras, E.H. Margulies, Z. Weng, M. Snyder, E.T. Dermitzakis,
R.E. Thurman, Identification and analysis of functional elements
in 1% of the human genome by the ENCODE pilot project,
Nature 447 (7146) (2007) 799–816.
[4] P. Kapranov, J. Cheng, S. Dike, D.A. Nix, R. Duttagupta, A.T.
Willingham, P.F. Stadler, J. Hertel, J. Hackermüller, I.L.
Hofacker, et al., RNA maps reveal new rna classes and a
possible function for pervasive transcription, Science 316 (5830)
(2007) 1484–1488.
[5] M. Guttman, I. Amit, M. Garber, C. French, M.F. Lin, D.
Feldser, M. Huarte, O. Zuk, B.W. Carey, J.P. Cassady, et al.,
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
Chromatin signature reveals over a thousand highly con-
served large non-coding RNAs in mammals, Nature 458
(7235) (2009) 223–227.
[6] È.-L. Mathieu, M. Belhocine, L.T.M. Dao, D. Puthier, S. Spicuglia,
Rôle des longs ARN non codants dans le développement normal
et pathologique, Med. Sci. (Paris) 30 (8-9) (2014) 790–796.
[7] T. Derrien, R. Johnson, G. Bussotti, A. Tanzer, S. Djebali, H.
Tilgner, G. Guernec, D. Martin, A. Merkel, D.G. Knowles,
et al., The GENCODE v7 catalog of human long noncoding
RNAs: analysis of their gene structure, evolution, and
expression, Genome Res. 22 (9) (2012) 1775–1789.
[8] X.C. Quek, D.W. Thomson, L.V. Maag Jesper, N. Bartonicek,
B. Signal, M.B. Clark, B.S. Gloss, M.E. Dinger, lncRNAdb v2.0:
expanding the reference database for functional long noncod-
ing RNAs, Nucleic Acids Res. 43 (D1) (2015) D168–D173.
[9] S. Chakraborty, A. Deb, R.K. Maji, S. Saha, Z. Ghosh,
LncRBase: an enriched resource for lncRNA information,
PLoS One 9 (9) (2014), e108010.
[10] Q. Jiang, R. Ma, J. Wang, X. Wu, S. Jin, J. Peng, R. Tan, T.
Zhang, Y. Li, Y. Wang, LncRNA2Function: a comprehensive
resource for functional investigation of human lncRNAs
based on RNA-seq data, BMC Genomics 16 (3) (2015) S2.
[11] Q. Jiang, J. Wang, X. Wu, R. Ma, T. Zhang, S. Jin, Z. Han, R. Tan,
J. Peng, G. Liu, et al., LncRNA2Target: a database for
differentially expressed genes after lncRNA knockdown or
overexpression, Nucleic Acids Res. 43 (D1) (2015) D193–D196.
[12] G. Chen, Z. Wang, D. Wang, C. Qiu, M. Liu, X. Chen, Q.
Zhang, G. Yan, Q. Cui, LncRNADisease: a database for long-
non-coding RNA-associated diseases, Nucleic Acids Res. 41
(D1) (2013) D983–D986.
[13] S. Ning, J. Zhang, P. Wang, H. Zhi, J. Wang, Y. Liu, Y. Gao,
M. Guo, M. Yue, L. Wang, et al., Lnc2Cancer: a manually
curated database of experimentally supported lncRNAs
associated with various human cancers, Nucleic Acids Res.
44 (D1) (2016) D980–D985.
[14] X. Chen, C.C. Yan, X. Zhang, Z.-H. You, Long non-coding RNAs
and complex diseases: from experimental results to computa-
tional models, Brief. Bioinform. 18 (4) (2017) 558–576.
[15] T. Zhao, J. Xu, L. Liu, J. Bai, C. Xu, Y. Xiao, X. Li, L. Zhang,
Identification of cancer-related lncRNAs through integrating
genome, regulome and transcriptome features, Mol. BioSyst.
11 (1) (2015) 126–136.
[16] X. Chen, G.-Y. Yan, Novel human lncRNA–disease associ-
ation inference based on lncRNA expression profiles,
Bioinformatics 29 (20) (2013) 2617–2624.
[17] J. Sun, H. Shi, Z. Wang, C. Zhang, L. Liu, L. Wang, W. He, D.
Hao, S. Liu, M. Zhou, Inferring novel lncRNA–disease associ-
ations based on a random walk model of a lncRNA functional
similarity network, Mol. BioSyst. 10 (8) (2014) 2074–2081.
[18] Y. Liu, R. Zhang, F. Qiu, K. Li, Y. Zhou, D. Shang, Y. Xu,
Construction of a lncRNA-PCG bipartite network and
identification of cancer-related lncRNAs: a case study in
prostate cancer, Mol. BioSyst. 11 (2) (2015) 384–393.
[19] M. Zhou, X. Wang, J. Li, D. Hao, Z. Wang, H. Shi, L. Han, H.
Zhou, J. Sun, Prioritizing candidate disease-related long non-
coding RNAs by walking on the heterogeneous lncRNA and
disease network, Mol. BioSyst. 11 (3) (2015) 760–769.
[20] G.U. Ganegoda, M. Li, W. Wang, Q. Feng, Heterogeneous
network model to infer human disease-long intergenic non-
coding RNA associations, IEEE Trans. Nanobiosci. 14 (2)
(2015) 175–183.
[21] X. Chen, Z.-H. You, G.-Y. Yan, D.-W. Gong, IRWRLDA:
improved random walk with restart for lncRNA–disease associ-
ation prediction, Oncotarget 7 (36) (2016) 57919.
Using HPO for Better Prediction of Disease-Associated lncRNA
[22] X. Chen, KATZLDA: KATZ measure for the lncRNA–disease
association prediction, Sci. Rep. 5 (2015) 16840.
[23] X. Chen, C. Clarence Yan, C. Luo, W. Ji, Y. Zhang, Q. Dai,
Constructing lncRNA functional similarity network based on
lncRNA–disease associations and disease semantic similar-
ity, 5 (2015) 11338.
[24] Y.-A. Huang, X. Chen, Z.-H. You, D.-S. Huang, K. Chan,
ILNCSIM: improved lncRNA functional similarity calculation
model, Oncotarget 7 (18) (2016) 25902–25914.
[25] X. Chen, Y.-A. Huang, X.-S. Wang, Z.-H. You, K. Chan,
FMLNCSIM: fuzzy measure-based lncRNA functional similarity
calculation model, Oncotarget 7 (29) (2016) 45948–45958.
[26] X. Chen, Predicting lncRNA–disease associations and
constructing lncRNA functional similarity network based on
the information of miRNA, Sci. Rep. 5 (2015) 13186.
[27] L. Cheng, H. Shi, Z. Wang, Y. Hu, H. Yang, C. Zhou, J. Sun,
M. Zhou, IntNetLncSim: an integrative network analysis
method to infer human lncRNA functional similarity, Onco-
target 7 (30) (2016) 47864.
[28] X. Yang, L. Gao, X. Guo, X. Shi, H. Wu, F. Song, B. Wang, A
network based method for analysis of lncRNA–disease
associations and prediction of lncRNAs implicated in
diseases, PLoS One 9 (1) (2014), e87797.
[29] W.A. Kibbe, C. Arze, V. Felix, E. Mitraka, E. Bolton, G. Fu, C.J.
Mungall, J.X. Binder, J. Malone, D. Vasant, et al., Disease
Ontology 2015 update: an expanded and updated database of
human diseases for linking biomedical knowledge through
disease data, Nucleic Acids Res. 43 (D1) (2015)
D1071–D1078.
[30] K. Peng, W. Xu, J. Zheng, K. Huang, H. Wang, J. Tong, Z.
Lin, J. Liu, W. Cheng, D. Fu, et al., The disease and gene
annotations (DGA): an annotation resource for human
disease, Nucleic Acids Res. 41 (D1) (2013) D553–D560.
[31] S. Köhler, S.C. Doelken, C.J. Mungall, S. Bauer, H.V. Firth, I.
Bailleul-Forestier, G.C.M. Black, D.L. Brown, M. Brudno, J.
Campbell, et al., The Human Phenotype Ontology project:
linking molecular biology and disease through phenotype
data, Nucleic Acids Res. 42 (D1) (2014) D966–D974.
[32] A. Hamosh, A.F. Scott, J.S. Amberger, C.A. Bocchini, V.A.
McKusick, Online Mendelian Inheritance in Man (OMIM), a
knowledgebase of human genes and genetic disorders,
Nucleic Acids Res. 33 (Suppl. 1) (2005) D514–517.
[33] J. Amberger, C.A. Bocchini, A.F. Scott, A. Hamosh,
McKusick's Online Mendelian Inheritance in Man (OMIM®),
Nucleic Acids Res. 37 (Suppl. 1) (2009) D793–D796.
[34] X. Chen, G.-Y. Yan, Semi-supervised learning for potential
human microRNA-disease associations inference, Sci. Rep.
4 (2014) 5501.
[35] T. van Laarhoven, S.B. Nabuurs, E. Marchiori, Gaussian
interaction profile kernels for predicting drug–target interac-
tion, Bioinformatics 27 (21) (2011) 3036–3043.
[36] W. Wang, S. Yang, J. Li, Drug target predictions based on
heterogeneous graph inference, Biocomputing 2013, World
Scientific 2013, pp. 53–64.
[37] S. Erten, G. Bebek, R. Ewing, M. Koyuturk, DADA: degree-
aware algorithms for network-based disease gene prioritiza-
tion, BioData Min. 4 (1) (2011) 19.
[38] C. Zhuang, J. Li, Y. Liu, M. Chen, J. Yuan, X. Fu, Y. Zhan, L. Liu,
J. Lin, Q. Zhou, Tetracycline-inducible shRNA targeting long non-
coding RNA PVT1 inhibits cell growth and induces apoptosis in
bladder cancer cells, Oncotarget 6 (38) (2015) 41194.
[39] W.-C. Liang, W.-M. Fu, C.-W. Wong, Y. Wang, W.-M. Wang,
G.-X. Hu, L. Zhang, L.-J. Xiao, D.C.-C. Wan, J.-F. Zhang, The
Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006
11
lncRNA H19 promotes epithelial to mesenchymal transition
by functioning as miRNA sponges in colorectal cancer,
Oncotarget 6 (26) (2015) 22513.
[40] M. Ishibashi, R. Kogo, K. Shibata, G. Sawada, Y. Takahashi,
J. Kurashige, S. Akiyoshi, S. Sasaki, T. Iwaya, T. Sudo,
Clinical significance of the expression of long non-coding
RNA HOTAIR in primary hepatocellular carcinoma, Oncol.
Rep. 29 (3) (2013) 946–950.
[41] C. Blume, A. Hotz-Wagenblatt, J. Hüllein, L. Sellner, A.
Jethwa, T. Stolz, M. Slabicki, K. Lee, A. Sharathchandra, A.
Benner, p53-dependent non-coding RNA networks in chronic
lymphocytic leukemia, Leukemia 29 (10) (2015).
[42] D. Li, L. Song, Z. Wen, X. Li, J. Jie, Y. Wang, L. Peng, Strong
evidence for LncRNA ZNRD1-AS1, and its functional Cis-
eQTL locus contributing more to the susceptibility of lung
cancer, Oncotarget 7 (24) (2016) 35813.
[43] X. Zhou, S. Liu, G. Cai, L. Kong, T. Zhang, Y. Ren, Y. Wu, M. Mei,
L. Zhang, X. Wang, Long non-coding RNA MALAT1 promotes
tumor growth and metastasis by inducing epithelial–mesenchy-
mal transition in oral squamous cell carcinoma, 5 (2015) 15972.
[44] S. Zhang, G. Zhang, J. Liu, Long noncoding RNA PVT1
promotes cervical cancer progression through epigenetically
silencing miR-200b, APMIS 124 (8) (2016) 649–658.
[45] Y. Lu, T. Li, G. Wei, L. Liu, Q. Chen, L. Xu, K. Zhang, D. Zeng,
R. Liao, The long non-coding RNA NEAT1 regulates
epithelial to mesenchymal transition and radioresistance in
through miR-204/ZEB1 axis in nasopharyngeal carcinoma,
Tumor Biol. 37 (9) (2016) 11733–11741.
[46] X. Zhang, R. Gejman, A. Mahta, Y. Zhong, K.A. Rice, Y.
Zhou, P. Cheunsuchon, D.N. Louis, A. Klibanski, Maternally
expressed gene 3, an imprinted noncoding RNA gene, is
associated with meningioma pathogenesis and progression,
Cancer Res. 70 (6) (2010) 2350–2358.
[47] C.-Y. Wang, L. Hua, K.-H. Yao, J.-T. Chen, J.-J. Zhang, J.-H.
Hu, Long non-coding RNA CCAT2 is up-regulated in gastric
cancer and associated with poor prognosis, Int. J. Clin. Exp.
Pathol. 8 (1) (2015) 779.
[48] D.-H. Le, L. Verbeke, L.H. Son, D.-T. Chu, V.-H. Pham,
Random walks on mutual microRNA–target gene interaction
network improve the prediction of disease-associated micro-
RNAs, BMC Bioinforma. 18 (1) (2017) 479.
[49] D.-H. Le, Network-based ranking methods for prediction of
novel disease associated microRNAs, Comput. Biol. Chem.
58 (2015) 139–148.
[50] D.H. Le, V.H. Pham, T.T. Nguyen, An ensemble learning-
based method for prediction of novel disease-microRNA
associations, 2017 9th International Conference on Knowl-
edge and Systems Engineering (KSE): 19-21 Oct. 2017
2017, pp. 7–12.
[51] D.-H. Le, Disease phenotype similarity improves the prediction of
novel disease-associated microRNAs, Information and Comput-
er Science (NICS), 2015 2nd National Foundation for Science
and Technology Development Conference on: 16–18 Sept. 2015
2015, pp. 76–81.
[52] X. Chen, D. Xie, Q. Zhao, Z.-H. You, MicroRNAs and
complex diseases: from experimental results to computa-
tional models, Brief. Bioinform. (2017) bbx130–bbx130.
[53] Z.-H. You, Z.-A. Huang, Z. Zhu, G.-Y. Yan, Z.-W. Li, Z. Wen, X.
Chen, PBMDA: a novel and effective path-based computational
model for miRNA–disease association prediction, PLoS Comput.
Biol. 13 (3) (2017), e1005455.
[54] X. Chen, C.C. Yan, X. Zhang, Z.-H. You, L. Deng, Y. Liu, Y.
Zhang, Q. Dai, WBSMDA: Within and Between Score for
12 Using HPO for Better Prediction of Disease-Associated lncRNA

MiRNA–disease Association prediction, Sci. Rep. 6 (2016)
21106.
[55] P.N. Robinson, S. Köhler, S. Bauer, D. Seelow, D. Horn, S.
Mundlos, The Human Phenotype Ontology: a tool for annotating
and analyzing human hereditary disease, Am. J. Hum. Genet. 83
(5) (2008) 610–615.
[56] S. Köhler, N.A. Vasilevsky, M. Engelstad, E. Foster, J. McMurry,
S. Aymé, G. Baynam, S.M. Bello, C.F. Boerkoel, K.M. Boycott,
et al., The Human Phenotype Ontology in 2017, Nucleic Acids
Res. 45 (D1) (2017) D865–D876.
[57] H.J. Lowe, G.O. Barnett, Understanding and using the Medical
Subject Headings (MeSH) vocabulary to perform literature
searches, JAMA 271 (14) (1994) 1103–1108.
[58] C.E. Lipscomb, Medical Subject Headings (MeSH), Bull.
Med. Libr. Assoc. 88 (3) (2000) 265–266.
[59] Organization WH, The ICD-10 Classification of Mental and
Behavioural Disorders: Clinical Descriptions and Diagnos-
tic Guidelines, vol. 1, World Health Organization, 1992.
[60] N. Sioutos, Sd Coronado, M.W. Haber, F.W. Hartel, W.-L. Shaiu,
L.W. Wright, NCI Thesaurus: a semantic model integrating
cancer-related clinical and molecular information, J. Biomed.
Inform. 40 (1) (2007) 30–43.
[61] M.Q. Stearns, C. Price, K.A. Spackman, A.Y. Wang, SNOMED
clinical terms: overview of the development process and project
status, Proceedings of the AMIA Symposium 2001,
pp. 662–666.
[62] R. Cornet, N. de Keizer, Forty years of SNOMED: a literature
review, BMC Med. Inform. Decis. Mak. 8 (1) (2008) S2.
[63] S.S. Weinreich, R. Mangon, J. Sikkens, M. Teeuw, M. Cornel,
Orphanet: a European database for rare diseases, Ned.
Tijdschr. Geneeskd. 152 (9) (2008) 518–519.
[64] H.V. Firth, S.M. Richards, A.P. Bevan, S. Clayton, M. Corpas, D.
Rajan, S.V. Vooren, Y. Moreau, R.M. Pettett, N.P. Carter,
DECIPHER: Database of Chromosomal Imbalance and Pheno-
type in Humans Using Ensembl Resources, Am. J. Hum. Genet.
84 (4) (2009) 524–533.
[65] P. Resnik, Using information content to evaluate semantic
similarity in a taxonomy, Proceedings of the 14th International
Joint Conference on Artificial Intelligence, vol. 1, Morgan
Kaufmann Publishers Inc., Montreal, Quebec, Canada, 1995.
[66] X. Chen, C.C. Yan, X. Zhang, Z.-H. You, Y.-A. Huang, G.-Y.
Yan, HGIMDA: heterogeneous graph inference for miRNA–
disease association prediction, Oncotarget 7 (40) (2016)
65257–65269.

Please cite this article as: D.-H. Le, L. T.M. Dao, Annotating Diseases Using Human Phenotype Ontology Improves Prediction of
Disease-Associated Long Non-coding RNAs, J. Mol. Biol. (2018), https://doi.org/10.1016/j.jmb.2018.05.006