IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 19, NO.
3, JULY 2020 571
Similar Disease Prediction With Heterogeneous
Disease Information Networks
Jianliang Gao , Ling Tian , Jianxin Wang, Yibo Chen, Bo Song , and Xiaohua Hu
Abstract — Studying the similarity of diseases can help
us to explore the pathological characteristics of complex
diseases, and help provide reliable reference information
for inferring the relationship between new diseases and
known diseases, so as to develop effective treatment plans.
To obtain the similarity of the disease, most previous meth-
ods either use a single similarity metric such as semantic
score, functional score from single data source, or utilize
weighting coefficients to simply combine multiple metrics
with different dimensions. In this paper, we proposes a
method to predict the similarity of diseases by node repre-
sentation learning. We first integrate the semantic score and
topological score between diseases by combining multiple
data sources. Then for each disease, its integrated scores
with all other diseases are utilized to map it into a vector
of the same spatial dimension, and the vectors are used
to measure and comprehensively analyze the similarity
between diseases. Lastly, we conduct comparative experi-
ment based on benchmark set and other disease nodes out-
side the benchmark set. Using the statistics such as aver-
age, variance, and coefficient of variation in the benchmark
set to evaluate multiple methods demonstrates the effective-
ness of our approach in the prediction of similar diseases.
Index Terms — Disease similarity, disease prediction, rep-
resentation learning, graph.
I. I NTRODUCTION
A. Disease Similarity
S TUDYING the similarities between diseases can help us
predict disease genes and infer disease associated ncRNAs
[1], [2]. It has been playing an important role in understanding
the pathogenesis of complex diseases, early prevention and
diagnosis of major diseases [3]–[5], providing reliable refer-
ence information in the development and safety assessment
of new drugs [6]–[8]. In addition, since similar diseases might
Manuscript received April 27, 2020; accepted May 5, 2020. Date of
publication May 25, 2020; date of current version July 1, 2020. This work
was supported in part by NSFC under Grant 61873288, in part by
the National Science Foundation (NSF) under Grant 1815256 and
Grant 1744661, and in part by the Hunan Key Laboratory for Internet
of Things in Electricity under Grant 2019TP1016. This article was
presented in part at the 2019 IEEE International Conference on Bioin-
formatics and Biomedicine. (Corresponding author: Yibo Chen.)
Jianliang Gao, Ling Tian, and Jianxin Wang are with the School of
Computer Science and Engineering, Central South University, Changsha
410083, China.
Yibo Chen is with the Information and Communication Branch, State
Grid Hunan Electric Power Company Ltd., Changsha 410000, China
(e-mail: chenyibo8224@gmail.com).
Bo Song and Xiaohua Hu are with the College of Computing &
Informatics, Drexel University, Philadelphia, PA 19104 USA.
Digital Object Identifier 10.1109/TNB.2020.2994983
1536-1241 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
have similar drug targets, searching similar diseases could help
drug repositioning [9], which is instrumental in reducing the
cost of drug development and clinical trial.
Similar diseases are mainly manifested in the following
aspects: clinically, they have significantly similar phenotypes
such as symptoms and signs; genetically, the same gene causes
different mutations; and at the molecular level, they have the
same molecular pathway [10]. Therefore, the existence of
similar pathogenic mechanisms, similar treatment schemes and
the same specific targeted signaling pathways can be used as a
theoretical basis to determine the similarities in diseases [11].
In addition to the disease characteristics, the calculation
of disease similarity also relies heavily on disease-related
biomedical data, which has greatly promoted the development
of new bioinformatics technologies. Most of the previous
research mainly chose rich biological ontology data as the
source of disease terminology for the calculation of dis-
ease similarity, such as the Disease Ontology (DO) [12],
the Human Phenotype Ontology (HPO), and the Gene Ontol-
ogy (GO) [11], [11]. DO provides biomedical information
related to human diseases, including the concepts of ratio-
nal descriptions, phenotypic characteristics, and related med-
ical vocabularies. HPO provides phenotypes and standardized
vocabularies that can also reflect human diseases, and its
similarity features can help define criteria for the classification
of similar diseases. For example, some semantic-based meth-
ods [13]–[15] usually utilize the structure and the semantic
grammar of disease terms (such as DO and MeSH [16])to
calculate the similarity between diseases; Many function-based
methods [17]–[19] take advantage of gene functional asso-
ciations [20], [21] to enhance the measurement of disease
similarity, for instance the BOG [22] proposed by Mathur
and Dinakarpandian is based on the similarity of overlapping
gene sets between DO diseases. However, these methods rely
too much on ontologies when calculating disease similarity.
Although bio-ontology terms can provide accurate descriptions
of disease concepts and their semantic relationships, it is not
applicable to all situations, as not all diseases have char-
acteristics of the ontological structure. For diseases without
ontology, methods mentioned previously become impossible.
Using different data sources unrelated to the ontology to
calculate disease similarity can solve this problem.
However, using different data sources for the task is difficult
and uncommon. There exist quite many challenges including
the lack of comprehensive consideration and evaluation of the
similarity measurement. For instance, most methods for the
572
measurement of disease similarity leverage either associations
of ontological disease concepts, interconnections of gene
related to each disease, or protein and protein interaction
network [23], [24]. The measurements between a disease with
all other diseases must be done in the same spatial dimension
or the same metric to ensure the consistency and accuracy of
the disease prediction. In order to solve the above problems
and achieve the goal of making the results of disease prediction
more accurate, we propose in this paper a method to calculate
disease similarity by considering multiple data sources related
to diseases at the same time. It maps all diseases in the same
spatial dimension to construct a disease similarity network and
predict similar diseases.
B. Motivation
How to predict similar diseases to a great extent is of
great significance to the discovery of the interrelationship
between diseases and the further study of diseases with similar
symptoms in the field of biomedicine. Although there exist
many studies that work on similarity diseases, there are still
two important issues need to be addressed:
1) How to Utilize More Available Data Sources to Increase the
Robustness of Disease Similarity Calculation: It is considered
that in addition to the symptoms, pathology, characteristics of
disease, biomedical data related to the disease can also provide
reference for the measurement of similarity between the dis-
eases. Some methods mainly use disease ontology to calculate
disease semantic similarity, or calculate the functional simi-
larity of disease based on the interaction relationship between
disease and disease-related genes or gene ontology. However,
not all biological entities have ontologies, and the formation
of many diseases is not related to genes, resulting in a lack
of interaction information between diseases and genes; Other
research works, while using ontology-independent biomedical
data, only use a single data source to measure disease similar-
ity. Unlike most previous methods, we integrate semantic sim-
ilarity and network-based structural features by utilizing mul-
tiple biomedical data sources irrelevant to the ontology to pre-
dict similar diseases and conduct comprehensive evaluation.
2) How to Measure Similarity of Diseases More Objectively:
The calculation of the similarity between each disease and all
other diseases must be done in the same spatial dimension to
ensure the accuracy of the disease prediction. Most previous
methods either use a single similarity metric to measure
the similarity between diseases such as semantic similarity
score and functional similarity score; or utilize weighting
coefficients to simply combine multiple metrics with different
nature or dimensions to obtain the similarity of the disease.
This leads to a lack of objectivity in the prediction of similar
diseases.
In order to solve this problem, our proposed method maps
multiple metrics of all diseases to the same spatial dimension
at the same time, using a multi-dimensional vector to represent
each disease node. The value of each dimension in the
vector has practical significance derived from the correlation
between diseases, and the similarity between the vectors of
multi-dimensional real values represents the similarity between
diseases.
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 19, NO. 3, JULY 2020
C. Contribution
To solve the above mentioned problems, we propose a novel
method that has following main contributions:
• We take into account multiple ontology-independent
biomedical data sources and combine the relationship
between disease and different biomedical entities for the
calculation of disease similarity.
• Our method applys different similarity metrics and
includes two calculation strategies. One is based on
semantic similarity calculation in heterogeneous networks
where different meta paths can be expressed as different
semantic relationships between diseases. Another is to
exploit the topological characteristics in the networks
of disease and biomedical entities while considering the
multi-hop neighbor nodes.
• We transform disease nodes into multi-dimensional
real-valued vectors, which allows more effective com-
putation of similarities between different diseases in the
same vector space. The value of each dimension is of
practical significance and is derived from the correlation
between the diseases.
• We design various sets of experiments based on bench-
mark set and other disease nodes outside the bench-
mark set, by adjusting parameters between different data
sources and similarity scores obtained from different
calculation metrics. Our extensive experiments and eval-
uations fully demonstrate the effectiveness of the method
we proposed in predicting similar diseases.
II. R ELATED W ORK
Prediction of similar diseases has received much attention in
medical communities. The goal of predicting similar diseases
is to obtain the diseases with the best similarity score from
networks of multiple diseases and biomedical entities. In order
to calculate the best similarity score between diseases, many
previous methods proposed various algorithms. In this section,
we review the related researches in calculating the similarity
of diseases. Existing methods can be classified into four
categories.
Methods based on semantic similarity measure the relevance
of the ontology terms related to the disease. DOSE [25] is an
R package providing semantic similarity computations doSi m
among DO terms and genes which allows biologists to explore
the similarities of diseases and of gene functions in disease
perspective. Wang [26] presented a method M I S I M for mea-
suring miRNA functional similarity based on disease semantic
similarity and construction of miRNA functional networks,
and the semantic values of diseases were calculated based on
the Directed Acyclic Graph (DAG) of corresponding diseases.
Resnic [13] proposed a method based on classification of
disease concept assessment. The method also utilizes proba-
bilistic knowledge to construct a measure of disease semantic
similarity calculations, which combines disease semantics with
empirical probabilities to achieve higher accuracy. Schlicker
proposed a method for comparing sets of GO terms and
for assessing the functional similarity of gene products [27].
The method mainly relied on two measure metrics of disease
GAO et al.: SIMILAR DISEASE PREDICTION WITH HETEROGENEOUS DISEASE INFORMATION NETWORKS
similarity: si m Rel and f unSi m. si m Rel is used to compare the
biological processes of the evolution of different substances,
and f unSi m is used to measure the functionally related gene
products in the relevant genome. Although bio-ontology terms
can provide accurate descriptions of disease concepts and their
semantic relationships for the calculation of disease similarity,
it should be considered that not all diseases have an ontology.
For diseases that don’t have an ontology, these methods cannot
predict similar diseases.
Methods based on function similarity [20] compare the
disease-related gene sets. Zhang [21] described the construc-
tion of the expanded human disease network (eH D N) by
combining the available disease gene information in GAD with
protein and protein interaction network. Process-similarity
based (PSB) [17] involves the associations based on GO terms
for calculating disease similarity. Mathur proposed a method
named BOG [22] which computes disease similarity using
Meta Map to map the disease annotation of Swissprot protein
entries to DO terms, and estimates the similarity between
diseases using co-annotation and the DO semantic hierarchy.
However, existing relationships between disease and genes
are rare, and the formation of many diseases is not even
related to genes. Moreover, above mentioned methods only
use single similarity metric to calculate similarity between
diseases, whereas combining various data sources and multiple
metrics could obtain more comprehensive and meaningful
results.
For the semantic- and function-based method, Cheng [18]
proposed Sem Funsi m method that combines the semantic
similarity and the functional similarity. The Sem Si m (Seman-
tic Similarity) is obtained by utilizing semantic association
from DO, while the FunSi m (Functional Similarity) is calcu-
lated from a weighted network of the human gene function
association. The Sem FunSi m is obtained in the end by
combining both Sem Si m and FunSi m. However, the values
of different dimensions are simply weighted by coefficients
for their combination as the similarity of the diseases, which
could result in a lack of objectivity in the prediction of similar
diseases.
For the methods based on topology, I n f oFlowSi m [19]
and Net Si m [24] take into account the protein and pro-
tein interaction networks (PPIN) and disease-gene network
to discover disease-disease associations. They apply infor-
mation Flow from disease to PPIN and random walk
with restart (RWR) respectively. However, since PPIN and
disease-related genes are not fully available, there is still room
for improvement in topology-based methods. RADAR [10]
proposed a framework to learn representations of diseases
that captures both of their semantics and structural identities
for the calculation of disease similarity. Although the method
combines semantics and topological information, it can not
calculate disease similarity in multiple disease network with
real-time update when a new data source is added.
In order to solve the above mentioned problems, we propose
a new approach for measuring disease similarity that combines
semantic similarity and topological similarity, and constructs
disease similarity network by mapping multiple metrics of all
diseases to the same spatial dimension.
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
573
III. M ETHOD AND A LGORITHM
A. System Overview
In order to achieve the optimal prediction of similar diseases
with enhanced supports from multiple data source, we estab-
lish a similarity scoring function that could reflect com-
prehensive information from both semantical and structural
aspects of various disease and biomedical entity interaction
networks. Semantical characteristics and topological features
are well quantified and integrated in our overall scoring
function to guide the calculation process and facilitate the final
predictions.
The overall framework of our approach is shown in Fig. 1.
The first step is to construct multiple Heterogeneous Disease
information networks (HeDINs) from raw data sources, such
as the relationship between disease and chemical or pathway;
The second step is from the constructed HeDINs to obtain
the semantic score according to meta path between diseases
as well as the topological characteristics of each disease, and
combines both of which to jointly build a Homogeneous Dis-
ease Similarity Network (HoDSN) composed of only disease
nodes; The third step is to convert the disease nodes in the
HoDSN to be dense vectors with same spatial dimension
by representation learning [28]; Finally, in the vector space
of the same dimension, the vector representation of each
node is used to calculate and predict the similarities among
diseases.
B. Transforming HeDINs Into HoDSN
A heterogeneous information network contains multiple
types of objects and links, where the links represent different
types of connections and relationships between objects [29].
From various disease-related data sources that contain multiple
types of objects (e.g. disease, pathway, and chemicals) and
their relationships, we are able to construct the HeDINs
accordingly. The HeDINs can be regarded as undirected
graphs, in which diseases and biomedical entities (chemical,
pathway, etc.) are represented by different types of nodes,
and the relationships between diseases and different biomed-
ical entities represent different semantic paths. For example,
the disease-chemical network GB = {V B , V B , E B } includes a
node set of disease V B and a node set of chemical V B , and
E B represents a set of relationships between the disease node
and the chemical node. The more detailed definition of het-
erogeneous networks and its construction can be found in the
path S I M [30], which are used for reference as we construct
the HeDINs from disease-chemical and disease-pathway data
sources.
For multiple HeDINs, we obtain the set of total diseases that
appear in all heterogeneous networks by applying the intersect
operation on all disease node sets to filter and extract common
diseases that help construct a HoDSN.
In order to transform multiple HeDINs to a HoDSN, we first
define two kinds of scores: the semantic score M(x, y) and
the topology score T (x, y), where x and y represent any two
different disease nodes throughout this paper. For the semantic
score M(x, y), we calculate it through meta path, which has
following definition:
574 IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 19, NO. 3, JULY 2020

Fig. 1. The framework of predicting similar diseases. Firstly, heterogeneous disease information networks (HeDINs) (e.g.  denotes disease,
 denotes pathway, and  denotes chemical) are constructed from raw data resources. In this example, there are two HeDINs GA and GB . The
corresponding subgraphs of these HeDINs are obtained by filtering vertices. Then topological scores and semantics scores are calculated in these
heterogenous subgraphs using Dynamic Time Warping (DTW) algorithm and meta path method respectively. In this way, we transform multiple
HeDINs to a Homogeneous Disease Similarity Network (HoDSN) with different weight values on the edges. Finally, the disease nodes can be
embedded according to the weight values and the similarity between diseases can then calculated using these n-dimensional vectors.

Definition 1 (Meta Path.): A meta path p is a path defined
by the path length and the types of nodes and edges on the
graph.
For example, (“di sease” → “chemi cal” → “di sease”) is
a meta path with length of three. Intuitively, the semantics
underneath different paths imply different similarities. The
semantic score M(x, y) is defined as:
2 ∗ |{ p x→y | px→y ∈ P}|
M(x, y) =
|{ p x→x | p x→x ∈ P}| + |{ p y→y | p y→y ∈ P}|
(1)
where px→y is a meta path instance between disease x and
disease y, px→x is a meta path instance between disease x
and disease x, P represents the set of pre-defined meta paths
between disease nodes. M(x, y) refers to a similarity score
between diseases based on semantic.
Secondly, topological features in the graph are usually
measured by the degree sequence of the nodes. Traditional
approaches of sequence similarity calculation usually apply
the algorithm such as Euclidean distance, but cannot solve the
complex case of sequences with different lengths. Dynamic
Time Warping [31] is a very effective algorithm that can mea-
sure similarity between two sequences with different lengths.
We define the topology score T (x, y) as follows:
τ
β i ∗DT W (i (x),i (y)))
T (x, y) = e−( i=0 , (2)
where i (·) refers to the degree sequence of the i t h hop
neighbor nodes, and i = 0 represents the node itself. β is
a parameter indicates the weight of the neighbor nodes of
different hops. DT W (i (x), i (y)) represents the distance
between the sequence of disease node degrees using Dynamic
Time Warping algorithm.
Then, we integrate both semantic and topology scores
together to be the integrated score, which can be defined as
follows:
S(x, y) = α ∗ M(x, y) + (1 − α) ∗ T (x, y), (3)
where α∈[0,1] is a parameter to adjust the contribution of each
similarity score M and T towards the integrated similarity
score S.
Finally, we transform multiple incomplete, non-weighted
HeDINs into a HoDSN with different weights on the edges.
The disease similarity network is an information network with
only disease type nodes and can be thought of as a complete
undirected graph, where the weight of the edge between each
disease node is generated by the integrated score we calculated

Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
GAO et al.: SIMILAR DISEASE PREDICTION WITH HETEROGENEOUS DISEASE INFORMATION NETWORKS
in the previous step.
|G|

W (x, y) = wk ∗ SGk (x, y), (4)
k=1
where SGk (x, y) denotes a weight on the edge of two disease
nodes in the k t h G, G ∈ G, and wk represents the proportion
of the k t h G network contributing to the weight of the link.
C. Node Embedding for Measuring Disease Similarity
The network representation learning algorithm can learn and
generate a vector representation for each node by analyzing
the connections between nodes in the complex information
network, while effectively integrate the network structure and
the external information of the nodes. Upon our constructed
networks, we further obtain the vector representation of each
node which can be used for various network application tasks
such as node similarity measurement.
In the process of vectorization of each node, our method
orderly encodes disease nodes in the network. The purpose is
to get the vector representation of each node according to the
same permutation criterion, so that the value of each dimension
in the multi-dimensional vector generated by each node can
be compared under the same metric and get better similarity
between any disease nodes in the same vector space. Thus,
we embed each disease node as a n-dimensional vector, where
n is the number of nodes in the homogeneous disease network.
For any disease node v i , its vector representation is:
v i = (W (i, 1), W (i, 2), . . . , W (i, n)), (5)
where W (i, j ) denotes the weigh value from disease node i
to j . The value of each dimension in the vector is generated
by the weight of the edge in the HoDSN.
Since each node can be represented by a vector in the same
dimensional space, the similarity between disease nodes can be
calculated by similarity measurement between vectors, such as
Cosine distance, Euclidean distance, and so on. The similarity
score between disease nodes hence can be referred as the
distance between two vectors in n-dimensional space.. Our
method adopt the Euclidean distance on vectors to calculate
the similarity of diseases with following equation:
1 benchmark set, so these disease pairs have been proved as
Si m(x, y) =  (6)
1+ (v x − v y ) · (v x − v y )T
where v x , v y are the vector representations of disease nodes
x and y respectively. The normalizing operation is to make
similarity scores of all diseases fall in the scope of [0,1] for
fair comparison. The larger Si m(x, y) of two diseases means
the higher similarity between them. The whole framework is
summarized in Algorithm 1.
IV. E XPERIMENT E VALUATION
A. Dataset Preparation
1) The Datasets: The datasets about associations of dis-
eases and other entities related diseases is obtained from
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
575
Algorithm 1 Disease Similarity
Input:
The number of HeDINs, N;
The set of HeDINs, NG = {Gk = (Vk , Vk , E)|k =
1, . . . , N};
Output:
Similarity matrix of HoDSN, Si m;
1: Initialize the set of disease nodes Node_di se = Φ;
2: for each Gk ∈ NG do
3: Node_di se = Node_di se ∩ Vk ;
4: end for
5: Vk = Node_di se of each HeDIN;
6: for each Gk ∈ NG do
7: Calculate semantic score M and topological score T for
each pair of diseases in Node_di se;
8: Combine two kinds of score, integrated score SGk = α ∗
M + (1 − α) ∗ T ;
9: end for
N
10: W = k=1 ωk ∗ SGk ;
11: Connect any two disease nodes in Node_di se, the weight
of edge is Wx y ;
12: Transform multiple HeDINs into HoDSN;
13: Learning node embeddings;
14: Si m ← similarity between all disease nodes;
15: return Si m;
CTDbase.1 In our experiments of disease similarity, the two
datasets as Table I include: disease-chemical association and
disease-pathway association. The disease-chemical association
includes 6,206 disease nodes, 4,180 chemical nodes, and
1,048,547 disease-chemical relationships; the disease-pathway
association includes 4,997 disease nodes, 2,338 pathway
nodes, and 569,716 disease-pathway relationships. How-
ever, after filtering and screening the two raw datasets,
the dataset used as the experiment of our method includes
4,986 disease nodes, 4,159 chemical nodes, 2,336 path-
way nodes, 1,042,765 disease-chemical relationships, and
569,642 disease-pathway relationships.
2) The Benchmark Set: The benchmark set is composed
of 65 disease pairs. Two manually checked datasets of
disease pairs with high similarity were integrated into the
similar [32], [33].
B. Experiment Setup
The integrated score of each disease node pair is obtained
by combining the semantic similarity score and the topological
similarity score together on a customizable coefficient α.
We adjusted the parameter α from 0.1 to 0.9 to explore the
influences of different contributions of two kinds of scores on
the experimental results.
In order to investigate the effects of different disease-related
data sources on disease similarity results, we also perform
experimental comparisons of multiple values of parameter w
1 Datasets are available at http://ctdbase.org/.
576 IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 19, NO. 3, JULY 2020

TABLE I
D ATASETS

Fig. 3. Parameter α from multiply similarity score.

Fig. 2. Parameter w from multiply data sources.

from 0.1 to 0.9. To discuss the effectiveness of the proposed
method, we utilize the disease pairs in the benchmark to
evaluate our method, and we set the range of prediction as 1%,
5%, 10% and 20%, then statistics the correct rate of disease
prediction in the different ranges.
C. Evaluation Results
1) Parameter w for Multiple Disease Information Networks:
We define parameter w and 1 − w to indicate the effected
weights of the chemical and pathway on diseases respectively.
We conduct the experiments based on the 65 diseases pairs in
the benchmark. Receiver operating characteristic (ROC) curves
are then drawn with the benchmark set against 50 random sets.
Each random set contains 650 randomly selected pairs. The
experimental results are shown in Fig. 2.
In Fig. 2(a), each color represents a parameter value
of w that varies between [0.1, 0.9]. Each column value in
Fig. 2(b) is obtained from the area under the ROC Curve
(AUC). From Fig. 2, we can see that: with the parameter value
changing from 0.1 to 0.9, the AUC value tends to decrease
gradually, but the overall AUC value can reach 72% or more.
When the value of parameter w1 equal to 0.1 and w2 equal
to 0.9, the correct rate on the benchmark calculated by our
proposed method of disease similarity can reach 85%. In our
experiment of comparing each single data source, the dif-
ference of similarity results obtained by single data source
of chemical-disease and pathway-disease is relatively large.
Therefore, we can conclude from Fig. 2 that by combining
two or more data sources, the proposed method tend to have
more stable results.
2) Parameter α for Integrated Similarity Score: In the
research, the contributions of semantic and topological scores
are defined by the parameter α and 1 − α for the integrated
similarity score. We set the parameter α to change from 0.1 to
0.9, and the experimental results are shown in Fig. 3.
In Fig. 3, each color represents a parameter value that
varies between [0.1, 0.9]. Each column value in Fig. 3(b) is
obtained from the area under the ROC Curve (AUC). From
Fig. 3, we can see that in the results of combining two

Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
GAO et al.: SIMILAR DISEASE PREDICTION WITH HETEROGENEOUS DISEASE INFORMATION NETWORKS
TABLE II
C OMPARISON R ESULT OF D ISEASE S IMILARITY FOR E ACH M ETHOD
similarity score based on two data sources, our proposed
method get the best when α = 0.8, and the accuracy reached
86.8%. However, the difference in the accuracy obtained by
the other parameter values is not obvious, so this result proves
that the weight parameters between the two similar score
calculation indicators are not sensitive in general.
3) Comparison With Other Methods: In order to compare the
effectiveness and superiority of different methods in the task of
measuring disease similarity, we use three metrics to evaluate
the experimental results. These three metrics are average,
standard deviation, and coefficient of variation. The average is
used to measure the average similarity score of disease pairs
in the benchmark set, and the standard deviation reflects the
degree of dispersion of similarity scores between disease pairs.
We apply the coefficient of variation that reflects the degree of
dispersion on the unit mean to compare the relative degree of
dispersion among data sets with different mean values, which
is more accurate than using only the standard deviation. The
coefficient of variation is the ratio of the standard deviation
to the average. This indicator is used to explain the balance,
rhythm, and stability of things in the process of development
and change. The smaller the coefficient of variation, the higher
the stability of the method.
Table II shows the evaluation results of comparative experi-
ments of six methods, where the bold numbers indicate the top
three methods in each metric. We can clearly see that although
the averages of the Resnik [13] and Wang [34] methods are
high, their standard deviations are more than double to other
methods; the PSB [17] and SemFunSim [18] methods perform
stable but their averages of similarity scores are lower than our
method. For comprehensive consideration, our method has the
lowest coefficient of variation, which means that the degree of
dispersion on the unit mean is the lowest. In the calculation
of similarity scores for all disease pairs, our method has better
balance and stability than other methods.
4) Disease Prediction: The similar disease pair set (A − B)
in the benchmark is set of one-to-one correspondence between
nodes in the disease set A and nodes in the disease set B. In the
section, our experiment take disease set A as the set of nodes to
be predicted, and take disease set B as the reference set. Using
our proposed method to calculate the similarity score between
any pair of diseases, with reference to the corresponding
disease nodes in the reference set B, the accuracy of similar
disease predictions in different ranges is counted. Table III
compares the results of similar disease predictions based on a
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
577
TABLE III
ACCURACY OF D ISEASE P REDICTION B ASED ON B ENCHMARK
chemical-disease data source, a pathway-disease data source,
and a dataset that combines the two data sources.
From Table III, we can see that: As the scope of predic-
tion expands, the accuracy increases. Moreover, our proposed
method performs better than a single data source. As the range
of predictions expands, the proposed method is better than the
disease prediction of only chemical-disease data source, and
it is better than or equal to the prediction result of the only
pathway-disease data source. In summary, the prediction result
of disease similarity with combining multiply data sources
performs better than a single data source.
V. C ONCLUSION
To measure the similarity between diseases, we propose
a new approach by utilizing multiple heterogeneous dis-
ease information networks. In each network, we integrate
the semantic similarity and the topological similarity. Then,
the vector representation learning method is applied to trans-
form the disease nodes into multi-dimension vectors in the
same spatial dimension, so that the vectors of the nodes
have ability to represent and reason the relationship between
diseases. The similarity calculation between vectors is used to
express the similarity between diseases, which is more effec-
tive for finding similar diseases. And our method propose to
apply coefficient of variation for the first time in the evaluation
of the experimental results for more balanced comparison of
different methods on the benchmark set, which better explain
the balance and stability of our method in the task of predicting
similar diseases.
R EFERENCES
[1] W. Lan, J. Wang, M. Li, J. Liu, F.-X. Wu, and Y. Pan, “Predicting
MicroRNA-disease associations based on improved MicroRNA and
disease similarities,” IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 15,
no. 6, pp. 1774–1782, Dec. 2018.
[2] W. Lan et al., “LDAP: A Web server for lncRNA-disease association
prediction,” Bioinformatics, vol. 33, no. 3, pp. 458–460, Feb. 2017.
[3] O. Vanunu, O. Magger, E. Ruppin, T. Shlomi, and R. Sharan, “Associat-
ing genes and protein complexes with disease via network propagation,”
PLOS Comput. Biol., vol. 6, no. 1, pp. 1–9, 2010.
[4] M. Li, R. Zheng, Q. Li, J. Wang, F.-X. Wu, and Z. Zhang, “Prioritizing
disease genes by using search engine algorithm,” Current Bioinf., vol. 11,
no. 2, pp. 195–202, Apr. 2016.
[5] J. Ni, M. Koyuturk, H. Tong, J. Haines, R. Xu, and X. Zhang, “Disease
gene prioritization by integrating tissue-specific molecular networks
using a robust multi-network model,” BMC Bioinf., vol. 17, no. 1,
pp. 453–466, Dec. 2016.
[6] L. Perlman, A. Gottlieb, N. Atias, E. Ruppin, and R. Sharan, “Com-
bining drug and gene similarity measures for drug-target elucidation,”
J. Comput. Biol., vol. 18, no. 2, pp. 133–145, Feb. 2011.
[7] A. Gottlieb, G. Y. Stein, E. Ruppin, and R. Sharan, “PREDICT:
A method for inferring novel drug indications with application to
personalized medicine,” Mol. Syst. Biol., vol. 7, no. 1, pp. 496–505,
2011.
578
[8] L. Cheng et al., “DisSim: An online system for exploring significant
similar diseases and exhibiting potential therapeutic drugs,” Sci. Rep.,
vol. 6, no. 1, pp. 30024–30030, Jul. 2016.
[9] P. Ni, J. Wang, P. Zhong, Y. Li, F. Wu, and Y. Pan, “Constructing disease
similarity networks based on disease module theory,” IEEE/ACM Trans.
Comput. Biol. Bioinf., vol. 17, no. 3, pp. 906–915, May/Jun. 2020.
[10] R. Qin, L. Duan, H. Zheng, J. Li-Ling, K. Song, and X. Lan, “RADAR:
Representation learning across disease information networks for similar
disease detection,” in Proc. IEEE Int. Conf. Bioinf. Biomed. (BIBM),
Dec. 2018, pp. 482–487.
[11] M. Ashburner et al., “Gene ontology: Tool for the unification of biology,”
Nature Genet., vol. 25, no. 1, pp. 25–29, 2000.
[12] L. M. Schriml et al., “Disease ontology: A backbone for disease seman-
tic integration,” Nucleic Acids Res., vol. 40, no. D1, pp. D940–D946,
Jan. 2012.
[13] P. Resnik, “Using information content to evaluate semantic similarity
in a taxonomy,” in Proc. 14th Int. Joint Conf. Artif. Intell., 1995,
pp. 448–453.
[14] D. Lin et al., “An information-theoretic definition of similarity,” in Proc.
ICML. Los Alamitos, CA, USA: Citeseer, 1998, pp. 296–304.
[15] S. Bandyopadhyay and K. Mallick, “A new path based hybrid measure
for gene ontology similarity,” IEEE/ACM Trans. Comput. Biol. Bioinf.,
vol. 11, no. 1, pp. 116–127, Jan. 2014.
[16] C. E. Lipscomb, “Medical subject headings (MeSH),” Bull. Med. Library
Assoc., vol. 88, no. 3, pp. 265–266, Jul. 2000.
[17] S. Mathur and D. Dinakarpandian, “Finding disease similarity based
on implicit semantic similarity,” J. Biomed. Informat., vol. 45, no. 2,
pp. 363–371, Apr. 2012.
[18] L. Cheng, J. Li, P. Ju, J. Peng, and Y. Wang, “SemFunSim: A new
method for measuring disease similarity by integrating semantic and
gene functional association,” PLoS ONE, vol. 9, no. 6, pp. 1–11, 2014.
[19] M. B. Hamaneh and Y.-K. Yu, “Relating diseases by integrating gene
associations and information flow through protein interaction network,”
PLoS ONE, vol. 9, no. 10, pp. 1–14, 2014.
[20] K. Goh, M. E. Cusick, D. Valle, B. Childs, M. Vidal, and A. L. Barabási,
“The human disease network,” Proc. Nat. Acad. Sci. USA, vol. 104,
no. 21, pp. 8685–8690, 2007.
[21] X. Zhang et al., “The expanded human disease network combining
protein–protein interaction information,” Eur. J. Hum. Genet., vol. 19,
no. 7, pp. 783–788, Jul. 2011.
[22] S. Mathur and D. Dinakarpandian, “Automated ontological gene annota-
tion for computing disease similarity,” Summit Transl. Bioinf., vol. 2010,
pp. 12–16, Aug. 2010.
Authorized licensed use limited to: Vasavi College of Engineering. Downloaded on September 21,2023 at 13:18:29 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 19, NO. 3, JULY 2020
[23] J. Gao, L. Tian, T. Lv, J. Wang, B. Song, and X. Hu, “Pro-
tein2Vec: Aligning multiple PPI networks with representation learning,”
IEEE/ACM Trans. Comput. Biol. Bioinf., early access, Aug. 27, 2019,
doi: 10.1109/TCBB.2019.2937771.
[24] P. Li, Y. Nie, and J. Yu, “Fusing literature and full network data
improves disease similarity computation,” BMC Bioinf., vol. 17, no. 1,
pp. 326–339, Dec. 2016.
[25] G. Yu, L.-G. Wang, G.-R. Yan, and Q.-Y. He, “DOSE:
An R/Bioconductor package for disease ontology semantic and
enrichment analysis,” Bioinformatics, vol. 31, no. 4, pp. 608–609,
Feb. 2015.
[26] D. Wang, J. Wang, M. Lu, F. Song, and Q. Cui, “Inferring the
human microRNA functional similarity and functional network based
on microRNA-associated diseases,” Bioinformatics, vol. 26, no. 13,
pp. 1644–1650, Jul. 2010.
[27] A. Schlicker, F. S. Domingues, J. Rahnenführer, and T. Lengauer,
“A new measure for functional similarity of gene products based
on gene ontology,” BMC Bioinf., vol. 7, no. 1, pp. 302–318,
Dec. 2006.
[28] Y. Wang, Y. Yao, H. Tong, F. Xu, and J. Lu, “A brief review of network
embedding,” Big Data Mining Analytics, vol. 2, no. 1, pp. 35–47,
Mar. 2019.
[29] C. Sun, Q. Li, L. Cui, H. Li, and Y. Shi, “Heterogeneous network-based
chronic disease progression mining,” Big Data Mining Analytics, vol. 2,
no. 1, pp. 25–34, Mar. 2019.
[30] Y. Sun, J. Han, X. Yan, P. S. Yu, and T. Wu, “PathSim: Meta path-based
top-K similarity search in heterogeneous information networks,” Proc.
VLDB Endowment, vol. 4, no. 11, pp. 992–1003, Aug. 2011.
[31] S. Salvador and P. Chan, “Toward accurate dynamic time warping in
linear time and space,” Intell. Data Anal., vol. 11, no. 5, pp. 561–580,
Oct. 2007.
[32] S. Suthram, J. T. Dudley, A. P. Chiang, R. Chen, T. J. Hastie,
and A. J. Butte, “Network-based elucidation of human disease sim-
ilarities reveals common functional modules enriched for pluripo-
tent drug targets,” PLOS Comput. Biol., vol. 6, no. 2, pp. 1–10,
2010.
[33] S. Pakhomov, B. McInnes, T. Adam, Y. Liu, T. Pedersen, and
G. B. Melton, “Semantic similarity and relatedness between clinical
terms: An experimental study,” in Proc. AMIA Annu. Symp., 2010,
pp. 572–576.
[34] J. Z. Wang, Z. Du, R. Payattakool, P. S. Yu, and C.-F. Chen, “A new
method to measure the semantic similarity of GO terms,” Bioinformatics,
vol. 23, no. 10, pp. 1274–1281, May 2007.